European Polytechnical Institute Kunovice
Transkript
European Polytechnical Institute Kunovice
EUROPEAN POLYTECHNICAL INSTITUTE KUNOVICE PROCEEDINGS SECOND INTERNATIONAL CONFERENCE ON SOFT COMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 January 29 - 30, Kunovice, Czech Republic 1 Edited by: Prof. Ing. Imrich Rukovanský, CSc, and Doc. Ing. Pavel Ošmera, CSc Papers authorized under supervisoring of the committee: Prof. PhDr. Karel Lacina, DrSc, Prof.Ing.Vladimír Mikula,CSc Doc. Ing. Jaroslav Ďaďo, PhD Prepared to print by: Bc. Andrea Šimonová, DiS, Bc. Pavel Kubala, DiS and Bc. Evžen Němec, DiS Printed by: © European Polytechnical Institute Kunovice, 2004 ISBN : 80–7314– 025- X 2 SECOND INTERNATIONAL CONFERENCE ON SOFT COMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 Organized by EUROPEAN POLYTECHNICAL INSTITUTE, KUNOVICE, CZECH REPUBLIC and PLEHANOV RUSSIAN ACADEMY OF ECONOMICS, MOSCOW, RUSSIA in cooperation with ASSOCIATION OF SMALL AND MEDIUM-SIZED ENTERPRISES AND CRAFTS CZ Conference Chairman Ing. Oldřich Kratochvíl, rector Conference Co-Chairmen Prof. Ing. Imrich Rukovanský, CSc, vice-rector Assoc. Prof. Ing.Pavel Ošmera, CSc 3 INTERNATIONAL PROGRAMME COMMITEE O. M. D. V. O. J. R. T. J. U. B. V. V. F. E. B. K. I. V. Y. J. P. B. J. J. J. K. I. N. G. J. N. G. V. J. A. W. Kratochvíl – Chairman (CZ) Baraňski (Poland) Bartoněk (Czech Republic) P. Beljanskij (Russia) Biro (Austria) Ďaďo (Slovak Republic) Hampel (Germany) Hlačina (Czech Republic) M. Honzík (Czech Republic) K. Chakraborthy (USA) Katalic (Austria) Kebo (Czech Republic) G. Knjazev (Russia) Koliba (Czech Republic) Krikavskij (Ukraine) Kulcsár (Hungary) Lacina (Czech Republic) Matušíková (Czech Republic) Mikula (Czech Republic) A. Mylopoulos (Greece) Nesvadba (Czech Republic) Ošmera (Czech Republic) Ošťádal (Czech Republic) Pavo (Hungary) Petrucha (Czech Republic) Radová (Czech Republic) Rais (Czech Republic) Rukovanský (Czech Republic) Sano (Japan) N. Smirnov (Russia) Strišš (Slovak Republic) P. Tichomirov (Russia) Vértesi (Hungary) I. Vidjapin (Russia) Wiktor (Poland) M. S. Zalzala (U.K.) Zamojski (Poland) ORGANIZING COMMITEE I. P. M. A. P. E. J. L. I. P. Rukovanský (Chairman) Ošmera Vinklárik Šimonová Kubala Němec Kavka Šimíček Gosiorovský A. Makarenko P. S. M. M. I. M. J. T. Matušík Amoutzas Balus Brázda Polášková Měska Šáchová Chmela Session 1: Marketing and Globalization Chairman: Prof. PhDr. Karel Lacina, DrSc, Co-Chairman : Assoc.Prof. J.Radová, PhD Session 2: Financial systems Chairman: Assoc.Prof..Ing.J.Ďaďo, PhD Co-Chairman: Assoc.Prof.Ing.J.Strišš,CSc Session 3: Soft Computing in Computer Environments Chairman: Prof.Ing.V.Mikula,CSc Co-Chairman: Assoc.Prof.Ing. P.Ošmera,CSc 4 DODÁVKA A MONTÁŽ PODLAHOVÝCH KRYTIN A INTÉRIÉROVÝCH DVEŘÍ Českomoravská 12a/2255, 190 00, Praha 9 Tel: 283 083 111, 333 fax: 283 083 534, 536 ELEKTROMECHANIKA Úvaly 4 CONTENTS: A Message from the General Chairman ICSC Conference...................................................... 9 MARKETING AND GLOBALIZATION THE SOLUTION OF LOGISTICS AND TRANSPORT PROBLEMS BY MEANS OF GENETIC ALGORITHM Dostál Pavel, Rais Karel (Czech Rep.)................................................................................. 11 EVOLUTIONARY ALGORITHMS FOR MARKETING AND MANAGEMENT Ošmera Pavel, Kratochvíl Oldřich ........................................................................................ 15 DIAGNOSTIC APPROACH TO THE HUMAN CAPITAL DEVELOPMENT Kucharčíková Alžbeta, Vodák Jozef (Slovak Rep.).............................................................. 21 MARKETING ACTIVITIES OF SLOVAK BUSINESSES AFTER THE ENTRY TO EU Strišš Jozef, Rešková Monika (Slovak Rep. )...................................................................... 27 THE DECISION - MAKING PROCESS MODELLING Hittmár Štefan. (Slovak Rep.)............................................................................................... 31 MARKETING AND ITS GLOBAL ORIENTATION ON TURN OF THE CENTURIES Jedlička Milan (Slovak Rep.)............................................................................................... 35 METHODS USED IN MARKETING SEGMENTATION PROCESS Rešková Monika ( Slovak Rep.) ......................................................................................... 41 ADVERTISING AND GLOBALISATION INFLUENCES Achalova Larisa Vladislavovna (Russia) ............................................................................. 47 CUSTOMER BEHAVIOUR Jankal Radoslav ( Slovak Rep. ) .......................................................................................... 53 CONSUMER BEHAVIOUR IN GLOBAL MARKET Seifoullaeva Elvira (Russia) ................................................................................................ 57 SALES PROMOTION Berchik Helen (Russia)........................................................................................................ 61 STRATEGIES OF AN INTERNATIONAL MARKET ENTRY Ponomareva Maria Alexandrovna (Russia)........................................................................... 65 SCOPES OF MARKETING STRATEGY FOR 3G MOBILE SERVICES Tokarčíková Emese ( Slovak Rep.)..................................................................................... 69 5 CULTURE IN INTERNATIONAL MARKETING AND BUYER BEHAVIOUR Kalugin Evgeniy (Russia).................................................................................................... 73 EFFECTIVE COMMUNICATION AND PROMOTION STRATEGY Oreshkin Alexey Gennadievich (Russia) ............................................................................. 77 MARKETING PREPAREDNESS OF SMALL AND MIDDLE-SIZE ENTERPRISES FOR THE ENTRY INTO THE EU Jaroslav Ďaďo ( Slovak Rep.) ............................................................................................... 83 INTERCULTURAL DIFFERENCES BETWEEN SLOVAK AND EU COUNTRIES Vladimír Laššák, Jaroslav Ďaďo ( Slovak Rep.) ................................................................... 89 FINANCIAL SYSTEMS FUZZY LOGIC AND FINANCIAL TIME SERIES Dostál Pavel, Žák Libor. (Czech Rep.) .....................................................................................93 DETERMINATION OF TAX SHIELD AND ITS INFLUENCE IN FINANCIAL DECISIONS Radová Jarmila, Marek Petr, Hlačina Tibor (Czech Rep.) ........................................................99 SOFT COMPUTING IN COMPUTER ENVIRONMENTS MODELLING OF INFLATION FUZZY TIME SERIES USING THE COMPETITIVE NEURAL NETWORK Marček Dušan (Slovak Rep.)............................................................................................. 107 USE OF NEW METHOD FOR LEARNING OF NEURAL NETWORKS Petrucha Jindřich, Mikula Vladimír (Czech Rep.)............................................................... 113 NEURAL NETWORK AND JOINT TRANSFORM CORRELATOR APPLIED FOR THERMAL IMAGERY-BASED RECOGNITION Kościuszkiewicz Krzysztof, Kobel Joanna, Mazurkiewicz Jacek, (Poland) ..................... 119 NEURAL NETWORK IN ADAPTIVE CONTROL Veleba Václav, Petr Pivoňka (Czech Rep.) ........................................................................ 125 6 FUZZY INFERENCE SYSTEM AND PREDICTION Žák Libor (Czech Rep.) ...................................................................................................... 131 EVOLVING CONTROLLERS Ošmera Pavel, Chakraborthy K.Uday, Rukovanský Imrich (Czech Rep./USA)................... 137 AN IMPROVED ALGORITHM BASED ON STOCHASTIC HEURISTIC METHODS FOR SOLVING STEINER TREE PROBLEM IN GRAPHS Smutek Daniel (Czech Rep.) ............................................................................................. 145 DETECTING PARETO OPTIMAL SOLUTIONS WITH PSO Baumgartner,U.,Magele,Ch.,Renhart,W (Austria) ............................................................ 155 THE RISK ANALYSIS OF SOFT COMPUTING PROJECTS Lacko Branislav ( Czech Rep.) ........................................................................................... 163 NETWORK VIRTUAL LABORATORY – INDUSTRIAL ROBOT Baranski Wlodzimierz M., Walkowiak Tomasz (Poland)................................................... 169 SIMD APPROACH TO RETRIEVING ALGOROTHM OF MULTILAYER PERCEPTRONd Mazurkiewicz Jacek (Poland) ............................................................................................ 177 ACOUSTIC LOCALIZATION OF MULTIPLE VEHICLES Zamojski Wojciech, Walkowiak Tomasz (Poland) ............................................................ 185 THE FUZZY RATIONAL DATABASE SYSTEM FSEARCH 2.0 Bednář Josef (Czech Rep.) ................................................................................................. 193 A SPECIAL DATABASE FOR THE MEASURING IN FIELD Bartoněk Dalibor (Czech Rep.).......................................................................................... 199 INTERNET – A MODERN E-LEARNING MEDIUM Woda Marek, Walkowiak Tomasz (Poland)....................................................................... 205 USING THE RUP FOR SMALL WEB INFORMATION SYSTEMS Švec Petr (Czech Rep.)....................................................................................................... 215 SHAPE OPTIMISATION OF THE FERROMAGNETIC CORE OF FLUXSET MAGNETOMETER Vértesy Gábor, Pávó,J ( Hungary) ..................................................................................... 221 RECONSTRUCTION AND REPRESENTATION OF EGYPTIAN MONUMENTS FROM POINT CLOUDS USING NEURAL NETWORK BASED METHOD Ashraf S.Hussein, Mohammed F.Tolba (Egypt) ............................................................... 229 EVALUATION OF DEFECT OF PIPELINE ISOLATION DETECTED BY PEARSON METHOD Bartoněk Dalibor,Rukovanský Imrich (Czech Rep.) .......................................................... 239 7 AN ENHANCED WWW-BASED SCIENTIFIC DATA VISUALIZATION SERVICE USING VRML Dina Reda Khattab, Amr Hassan Abdel Aziz, and Ashraf Saad Hussein (Egypt) ............. 245 SYSTEMS FOR EVALUATION OF ANTICORROSIVE PROTECTION OF PIPELINE Bartoněk Dalibor, Nesvadba Jaroslav (Czech Republic) ................................................... 255 COURSE TIMETABLING BY AN ARTIFICIAL IMMUNE SYSTEM Čechman J. (Czech Rep.) ................................................................................................... 261 MANAGING A HIGH SPEED LAN USING DISTRIBUTED ARTIFICIAL INTELLIGENCE Ibrahiem M. M-El Emary (Jordan) ................................................................................... 269 MARKOV CHAINS AND EXAMPLE OF THEIR USE Ševčík Vítězslav (Czech Rep.) ........................................................................................... 277 APPLICATION OF THE TWO-DIMENSIONAL HELLINGER AND SHANNON QUASI-NORM Jurák Petr, Karpíšek Zdeněk (Czech rep.)........................................................................... 285 THE NEW MODEL OF SYSTEM FUZZY RELIABILITY Jelínek Pavel, Karpíšek Zdeněk (Czech Rep.).................................................................... 291 ON THE APPLICATION OF INTELLIGENT AUTONOMOUS AGENTS FOR MANAGING THE REAL TIME COMMUNICATION NETWORK Ibrahiem M. M-El Emary, (Jordan) ................................................................................... 297 HOLOGRAPHIC REDUCED REPRESENTATION IN ARTIFICIAL INTELLIGENCE AND COGNITIVE SCIENCE Kvasnička Vladimír, Pospíchal Jiří (Slovak Rep.).............................................................. 305 A NOTE ABOUT NON-LINEAR FUNCTION APPROXIMATION WITH FUNCTION BOUNDARY Radomil Matoušek (Czech Rep.) ........................................................................................ 329 ANALYSIS METHODS FOR IMAGE PRE-PROCESSING AND 2-DIMENSIONAL PATTERN RECOGNITION Roman Weisser (Czech Rep.) ............................................................................................. 333 ON THE DISREGARDED FORMS OF COMPUTER CRIME Pichl Karel, Pichlová Markéta (Czech Rep./ France) ........................................................ 343 8 A message from the General Chairman of the Conference Dear guests and participants at this conference I would like to welcome you very kindly on behalf of myself and my colleges at the second Scientific conference ICSC held in the European polytechnialc institute. I would like to thank to all who actively acted in preparation in of the conference. Delivered papers of the top quality guarantee the high quality of the conference program. The personal attendance of the authors gives a possibility of both high professional level and the social level of the event. The specialization of the conference as in ICSC 2003 focuses on three study branches at our college: Marketing and Globalization, Soft Computing in the Financial Systems, Soft Computing in Computer Environments. Thanks to you there will be brought forward the most up-to-date contemporary reports in all the sections. I mean mainly the problems of using a neural networks, fuzzy logic and genetic algorithms presented in particular articles, and questions concerning the different optimization problems to the modelling of management systems, marketing optimization using genetic algorithms, as well as the role of the financial systems in the global economy. The first international conference ICSC 2003 held by our school in January last year fulfilled our imagination to getting the basic information about the actual trends in developing the branches, the state of knowledge of various problems and not the last to give the possibility to make a low of personal contacts with local institutions as well as those abroad. The papers of the conference ICSC 2004 follow up the themes of the last one but the large development of principles and possibilities of using neural networks and genetic algorithms in practice. In the future we expect that the attention will be paid to application of using the conference results in „business intelligence“. The conference ICSC 2004 has been organized by European Polytechnical Institute in cooperation with The Plehanov Russian Academy of Economics in Moscow and with mutual activities of Association of Small and Medium-sized Enterprises and Crafts in Czech Republic. Let me close with me wishing you all, by myself and in the name of all the college management, a pleasant stay and the establishing of useful professional and personal contacts. I believe that our conference will fulfil your expectations and that you will gain a lot of valuable information and ideas here. Kunovice, January 29, 2004 Dipl. Ing. Oldřich Kratochvíl rector 9 10 THE SOLUTION OF LOGISTICS AND TRANSPORT PROBLEMS BY MEANS OF GENETIC ALGORITHMS Pavel Dostál1), Karel Rais2) 1) Strakatého 15, 636 00 Brno, Czech Republic E-mail: [email protected], http://www.iqnet.cz/dostal/ Phone: ++420 5 44211639, Fax: ++420 5 44234750 2) Technical University of Brno, Faculty of Business and Management, Technická 2, 616 69 Brno, Czech Republic E-mail: [email protected], Tel: ++420-54114 2233, Fax: ++420-54114 2527 Abstract: The article presents the possible use of genetic algorithms for the solution of logistics and transport problems. The logistics problem is demonstrated on the case of pick-up places and collection centers. The transport problem is demonstrated on the case of traffic route. The process of optimalization is contribution to minimize the costs and maximize the profit. Keywords: Logistics problems, transport problems, optimalization, genetic algorithms 1. INTRODUCTION They are problems that are necessary to solve in practice and transport and logistics problems belong among them. The correct optimization of such problems enables us to minimize the cost and time. The genetic algorithms can help us with such problems. 2. LOGISTICS OPTIMALIZATION The solution of logistics problem can be presented in the following case. We have defined the coordinates of pick-up places (for example pick-up of mail, garbage, etc.) and we determine the number of collection centers where the fabricator will be placed. The problem is to determine the coordinates of such collection centers and respective pick-up places. The method of clustering can be used and the genetic algorithms optimize the solution. During the process of calculation the pick-up places are divided into clusters and then the coordinates of center of each cluster are calculated. The number of pick-up places and collection centers are not restricted. Part of Tab.1. shows the number and coordinates of pick-up places and their relative collection center. The Tab.2. shows calculated coordinates of collection centers A,B,C. Number ….. 22 23 24 25 26 27 28 29 ….. Pick-up places X Y ….. ….. 25.00 4.19 16.00 2.62 111.00 16.79 66.80 10.10 20.40 4.10 120.00 17.01 56.80 5.69 59.90 10.80 ….. ….. Collection center …... A A C B A Tab.2. Collection C B B …... Coordinates X Y 18.4 3.1 59.8 9.0 106.8 15.9 Collection center A B C centers Tab.1. Pick-up places INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 11 The process of optimalization determines the placement of collection centers and relative pick-up places in such a way to minimize the distances among them. The Fig.1. shows the pick-up places and collection center which solve the mentioned problem and thus to minimize the costs and time. 20 15 10 5 0 0 20 40 60 80 100 120 Fig.1. The pick-up places and collection center A,B,C 3. TRAVEL OPTIMIZATION The solution of travel problem can be presented in the following case. When we know the coordinates of pick-up places and collection centers we can determine the traffic distances among them for each group. The problem is to determine the shortest traffic route of places which we have to visit. The genetic algorithms can help us to solve such problem especially when the number of visited places is high. During the process of calculation is searched the shortest traffic route. Part of the Tab.3. shows the traffic distances among collection center A (CCA) together with pick-up places in group A (PPAx) in km. The Tab.4. shows the order of visited places before and after optimization. The traffic route measures after (before ) optimization 327 (441) km. The Fig.2. shows these two traffic routes. The shortest traffic route enables us to minimize the costs and time. CCA ….. PPA2 PPA3 PPA4 PPA5 ….. Order ….. ….. ….. ….. ….. ….. ….. ….. 1 CCA CCA PPA2 30 ….. 0 14 30 30 ….. 2 PPA1 PPA13 PPA3 24 ….. 14 0 18 17 ….. 3 PPA2 PPA12 PPA4 36 ….. 30 18 0 15 ….. 4 PPA3 PPA10 CCA 0 ….. 30 24 36 24 ….. 5 PPA4 PPA9 PPA5 24 ….. 30 17 15 0 ….. 6 PPA5 PPA8 ….. ….. ….. ….. ….. ….. ….. ….. 7 PPA6 PPA7 8 9 PPA7 PPA11 PPA8 PPA1 10 PPA9 PPA4 PPA10 PPA3 12 PPA11 PPA5 4.5 13 PPA12 PPA6 4.0 14 PPA13 PPA2 15 PPA14 PPA14 Tab.3. Travel distances Before optimalization After optimalization Tab.4. Traffic route 11 4 .5 4 .0 3 .5 3.5 3 .0 3.0 2 .5 2.5 2 .0 2.0 1 CCA CCA Total: 441 km 327 km INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED 1.5 IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 1 .5 EPI Kunovice, Czech Republic, January 29-30, 2004 12 1.0 1 .0 0 5 10 15 20 25 30 0 5 10 15 20 25 30 Fig.2a. Traffic route (441 km) Fig.2b. Traffic route (327 km) 4. CONCLUSION The genetic algorithms enable us to solve complicated logistics and traffic problems. The correct optimalization and application of results in practice enables us to minimize the costs, increase the profit and save our environment. LITERATURE: [1] DOSTÁL P., RAIS, K.: Methods of Large Investment Unit Modelling, Trento 2001, Italy, Transformation of CEEC Economies to EU Standards, Conference, s.53-57, ISBN 80-86510-27-1. [2] DOSTÁL P., RAIS, K.: Genetické algoritmy a jejich využití v modelování investic. Brno 2002, Manažerské mosty 2002, Konference, p..41-44, ISBN 80-7314-004-7. [3] DOSTÁL, P.: Moderní metody ekonomických analýz (Modern Methods of Economic Analyses). UTB – FAME - Zlín 2002, Study materials, 110 p., ISBN 80-7318-075-8 [4] DOSTÁL P., RAIS, K.: Production Schedule by Means of Genetic Algorithms. Kunovice 2003, International Conference on Soft Computing, p.32-34, ISBN 80-7314-017-9 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 13 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 14 EVOLUTIONARY ALGORITHMS FOR MARKETING AND MANAGEMENT Pavel Ošmera1), Oldřich Kratochvíl2) 1) Institute of Automation and Computer Science Brno University of Technology, Technická 2, 616 69 Brno, Czech Republic osmera @uai.fme.vutbr.cz 2) European Polytechnic Institute, Osvobození 699, 686 04 Kunovice, Czech Republic Tel. 572 548 035, Fax: 572 549 018, E-mail: [email protected] Abstract: There has been a recent resurgence of interest in using simulated evolution as a method for discovering solutions to economic optimization problems. These include prediction, modeling, control, decision etc. The adaptive significance of parallel GAs and the comparison with standard GAs are presented. Evolutionary algorithms can be used to solve the management and marketing problems. Few evolutionary algorithms have been applied to the Dynamic Economic Problem (DEP). Hybrid and parallel genetic algorithms (GAs) for solving the dynamic economic problems can be implemented. The adaptive significance of parallel GAs is presented. Keywords: parallel evolutionary algorithms, adaptive systems, optimization economic problems 1 INTRODUCTION Business decisions almost depend upon some forecast the course of events. Forecasting involves making the best possible judgment about some future event. In today’s rapidly changing business world such judgments can mean the difference between success and failure. Demand forecasts are also valuable as an input to sales and departmental budgets. Sales forecast, e.g. in dollars, are provided to the finance department as an input to the financial budgeting process. Managers use time-series analysis smoothing, regression, and judgmental methods in developing forecasts of sales. As forecaster, at one time or another, we have to ask ourselves why we should try to forecast in the first time. There is a three-part response to this question: First, the power of force such as economics, competition, markets, social concerns and the ecological environment to affect the individual firm is severe and continues growing. Secondly, forecast assessment is a major input in management’s evaluations of different strategies at business decision-making levels. Thirdly, the inference of no forecasting is that the future either contains “no significant change” or there is ample time to react “after the fact”[18]. The fundamental approach to decision-making is to formulate a single metric which summarizes the performance or value of a decision and iteratively improve this performance by selecting from among the available alternatives. Unfortunately, many problems do not lend themselves to optimization by classic methods (e.g., Newton-Gauss, steepest descent, and so forth). The mapping from each proposed solution to its corresponding performance index may not be differentiable, or even continuous, and may possess multiple local optima. In many modeling problems, the assumptions that are required to implement an estimation technique are violated. Unknown systems may be nonstationary or even intelligently interactive. Classic optimization methods often lead to unacceptably poor performance when applied to real-world circumstances. A more robust optimization technique is required. Applying the logical aspects of the evolutionary process to optimization offers several distinct advantages. There exists a large body of knowledge about the process of natural evolution which can be used to guide simulations. This process is well-suited for solving problems with unusual constrains where heuristic solutions are not available or generally lead to unsatisfactory results. A traditional science tended to emphasize stability, order, uniformity, and equilibrium. It concerned itself mostly with closed systems and linear relationships in which small inputs uniformly yield small results. Generally they center on the basic conviction that - at some level - the world is simple and is governed by time-reversible fundamental laws. What makes the new paradigm especially interesting is that it shifts attention to those aspects of reality that characterize today’s accelerated social change: ways, scientists have expressed that the puzzle fits disorder, instability, diversity, disequilibrium, nonlinear relationships (in which small inputs can trigger massive consequences), and temporality – a heightened sensitivity to the flow of time. Most phenomena of interest to us are, in fact, open systems, exchanging energy or matter (and, one might add, information) with their INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 15 environment. All systems contain subsystems, which are continually “fluctuating”. At times, single fluctuations or a combination of them may become so powerful, as a result of positive feedback, that they shatter the preexisting organization. Authors [6], [8] call it ”a singular moment” or “a bifurcation point”. It is inherently impossible to determine in advance which direction a change will take: whether the system will disintegrate into “chaos” or leap to a new, more differentiated, higher level of “order” or organization which is called “dissipative structure”. Today we are going from deterministic, reversible processes to stochastic and irreversible ones. Certain events go only one way - not because they cannot go the other way, but because it is extremely unlikely that they go backward. Initial conditions corresponding to a single point in unstable systems correspond to infinite information and are therefore impossible to find or observe. No equilibrium brings “order out of chaos”. However, as it was already mentioned, the concept of order (or disorder) is more complex than it was thought to be. Open systems evolve to higher and higher forms of complexity. Molecular biology shows that not everything in a cell is alive in the same way. Some processes reach equilibrium; others are dominated by regulatory enzymes far from equilibrium. Speculation that artificial intelligence, or more properly “machine learning”, could be created through a simulation of evolution goes back at least to Cannon (1932), who pictured natural evolution as a process that is similar in consequence to that which an individual undergoes, proceeding by random trial-and- error. The first attempt at the formal application of evolutionary theory to practical engineering problems appeared in two somewhat disparate areas: statistical process control and artificial intelligence. 2 ADAPTATION OF COMPLEX ECONOMIC SYSTEMS In dynamical systems, transition can be found: order, complexity, and chaos. Analogously, water can exist in solid, transitional, and fluid phases. Kauffman was busy on the autocatalytic set simulation for the origin of life [5], [6]. In nonlinear systems, a chaos theory tells us that the slightest uncertainty in our knowledge of initial conditions will often grow inexorably, and our predictions are nonsense. Complex adaptive systems share certain crucial properties (non - linearity, complex mixture of positive and negative feedback, nonlinear dynamics, emergence, collective behavior, spontaneous organization, etc.). In the natural world, such systems include brains, immune systems, ecology, cells, developing embryos, and ant colonies. In the human world, they include cultural and social systems. Each of these systems is a network of a number of “agents” acting in parallel: in a brain, the agents are nerve cells; in ecology, the agents are species; in a cell, the agents are organelles such as the nucleus and the mitochondria; in an embryo, the agents are cells, and so on. Each agent finds itself in the environment produced by its interactions with the other agents in the system. It is constantly acting and reacting to what the other agents are doing. There are emergent properties, the interaction of a lot of parts, the kinds of things that the group of agents can do collectively; something that the individual cannot. There is no master agent - for example - a master neuron in the brain. Complex adaptive systems have a lot of levels of organization (hierarchical structures), with agents at any level serving as building blocks for agents at a higher level. It is no wonder that complex adaptive systems (with multiple agents, building blocks, internal models, and perpetual novelty) are so hard to analyze with standard mathematics. We need mathematics and computer simulation techniques (a whole new mathematical art: programming) that emphasize internal models, emergence of new building blocks, and a rich web of interactions between multiple agents. We now have a good understanding of chaos and fractals showing how simple systems with simple parts can generate very complex behaviors. The edge of chaos is a special region onto itself, the place where you can find systems with lifelike, complex behavior. Living systems are actually very close to this edge-of-chaos phase transition, where things are much looser and more fluid. A natural selection is not an antagonist of selforganization. It is a force that constantly pushes emergent, self-organizing systems towards the edge of chaos. Evolution always seemed to lead to the edge of chaos. The complex evolutionary structure described in [10], [13] can be transformed to the structure of the computational intelligence (see Fig. 1). A random genetic crossover or mutation may give a species the ability to run much faster than before. The agent starts changing, then it induces changes in one of its neighbors, and finally you get an avalanche of changes until everything again stops changing. Systems get to the edge of chaos through adaptation: each individual (agent) tries to adapt to all the others. Co - evolution can also get them there; the whole system coevolves to the edge of chaos. In ecosystems or ecosystem models [10], three regimes can be found: ordered regime, chaotic regime, and edge-of chaos like a phase transition. When the system is at the phase transition, then - of course - order and chaos are in balance. There is an evolutionary metadynamics, a process that would tune the internal organization of each agent so that they all reside at the edge of chaos. The maximum fitness INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 16 occurs right at the phase transition. Life is clearly a chemical phenomenon, and only molecules can spontaneously undergo a complex chemical reaction with one another. The first source of chemistry’s power is a simple variety. The second source of power is reactivity: the structure A can manipulate the structure B to form something new – a structure C. A growth of complexity really does have something to do with far-fromequilibrium systems building themselves up, cascading to higher and higher levels of organization. Once they have accumulated a sufficient diversity of objects at the higher level, they go through a kind of autocatalytic phase transition and get an enormous proliferation of things at that level. Life is a natural expression of complex matter. It is a very deep property of chemistry and catalysis and being far from equilibrium. Evolutionary computation is generally considered as a consortium of genetic algorithms (GA), genetic programming, evolutionary programming, evolutionary strategies. While it has been studied for 40 years by computer scientists, its application to economics and finance has a much shorter history. After a decade-long development, we believe that it is high time to have a special interest on this subject. In 1988, when John Holland and Brian Arthur established an economics program at the Santa Fe Institute, artificial stock markets were chosen as the initial research project. Now many papers are directly related to agent-based artificial stock markets ACE (Agent-Based-Computational Economic). There are other softcomputing methods suitable for solving economic problems: Particle Swarm Optimization, Multi-Agent EA, Agent-based Multiobjective Optimalization Ant Colony Optimization, Team Optimization,Culture Algorithms, Evolutionary Computation: Evolution Strategies Genetic Programming Genetic Algorithms, Parallel GAs Fuzzy logic, Neural Networks, Fuzzy-rough Sets, Fuzzy-neural Modeling, Hybrid Learning, Intelligent Control, Cooperative co-evolutionary Algorithms, Parasitic Optimization, Bacterial EA, Artificial Life Systems, Differential EA, Parallel Hierarchical EA, Meta-Heuristics, Evolutionary Multi-objective Optimization, Evolvable control, Embryonic Hardware, Human-Computer Interaction, Molecular-Quantum Computing, Data Mining, Chaotic Systems, Scheduling 3 MARKETING OPTIMIZATION USING PARALLEL GENETIC ALGORITHMS For artificial foreign exchange markets, GA learning can be implemented in two different ways, namely, learning how to optimize and learning how to forecast. Many GAs are implemented on a population consisting of haploid individuals. However, in nature, many living organisms have more than one chromosome, and there are mechanisms used to determine dominant genes [9], [11]. A sexual recombination generates an endless variation of genotype combinations that increase the evolutionary potential of the population [11]. As it increases the variation among the offspring produced by an individual, it improves the chance that some of them will be successful in varying, and they will often encounter unpredictable changes of the environment. Mutation, genetic drift, migration, nonrandom mating, and natural selection cause changes in variation within population. It can be demonstrated that standard GAs with haploid chromosomes are unable to correctly locate optimal solutions for time-dependent objective functions. The adaptation of GAs depends on the speed of landscape changes through time. An applications of GAs concerns the effectiveness of different policies on pollution control. Pollution control is very hot topic in microeconomics because it provides a perfect illustration on how the optimal level of pollution can potentially be achieved by using the market mechanism. Cooperative computation with market mechanism provides another demonstration how to use Softcomputing. 3.1 Algorithm description The following GAs can be used to solve the DEP: Standard GA – One population with the size = 40 individuals. Individuals are sorted by their fitness. The higher is the fitness, the higher is the probability of selecting the individual to be a parent (a roulette wheel). The second parent is selected in the same manner. Then, the crossover, mutation, and correction are applied. The reproduction is repeated until the worse half of population is replaced. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 17 GA with two sub-populations (the sexual approach where the male and the female are distinguished) –Every sub-population with the size = 20 individuals. Individuals are sorted by their fitness. The first parent is selected from the male sub-population while the second parent is selected from the female sub-population. The selection probability of the first parent is performed with a uniform distribution function, while the selection probability of the second parent is different, using a modified roulette wheel approach that more often prefers better individuals. Crossover, mutation, and correction operators are then applied. The reproduction is repeated until the worse half of every sub-population is replaced. The GA with two sub-populations is a special case of the parallel GA with two sub-populations and sexual selection. The sexual recombination generates an endless variety of the genotype combinations that increase the evolutionary potential of the population [9-13]. As it increases the variation among the offspring produced by an individual, it improves the chance that some of them will be successful in the varying environment. Mutation, genetic drift, migration, a nonrandom mating, and natural selection cause changes in variation within the population. It can be demonstrated that standard GAs with haploid chromosomes that are unable to correctly locate optimal solutions for time-dependent objective functions. The adaptation of GAs depends on the speed of landscape changes through time. Parallel GA has a two - level structure. The first level is created by several populations with different GAs [10], [14]. The best or random individuals from the first level are sent to the second level. At this level, the standard GA with the elitism runs. This two-level structure allows us to find a better solution than that found by GAs in the first level; the best solution from the first level can never be lost but only overtaken in the second level. 4 CONCLUSIONS Evolutionary optimization can quicly achive near optimum solution even the space of possible solution is immense. Further, the nature of the algorithms indicates its potential utility for parallel processing machines. Each parent can be mutated independently and each offspring can be evaluated independently (unless fitness is an explicit function of the other members of the population and even then the process can be accelerated through parallelism). Competition and selection can alsou be parallelized. Thus the required execution time of evolutionary algorithms can be decreased dramatically if a number of small processors are devoted to these independent calculations. Mete-approaches can be applied to the continuum of evolutionary algorithms. This may lead to an increased rate of optimization. Natural evolution is the most robust yet efficient problem solving technique. Evolutionary algorithm can be made as robust. The same procedure can be applied to diverse problems with relatively litle reprogramming. While such efforts will undoubtedly successfully address difficult real-world problems, the focus of further development is in simulation evolution shloud remain “top-down”. The ultimate advancement of the field will, as always, rely on careful observation and abstraction of natural process of evolution (for example artifficial immune systems to solve complex economic problems). There is an optimal lifetime where almost all individuals in the population have high fitness. This strategy is used in nature by viruses and bacteria. A host-parasite arms race is one of the most basic and unavoidable consequences of evolution. It begins to look as if parasites are inevitable in any systems of life. Computer viruses have since become a worldwide problem, too. Sex is about disease. It is used to combat the threat from parasites. Organisms need sex to keep their genes one step ahead of their parasites. Females add sperm to their eggs because if they did not, the resulting offspring would be identically vulnerable to the first parasite that picked their genetic locks [7]. Sex stores genes that can be currently bad but have promise for reuse. The computer simulation helps identify the conditions under which the evolution of living world is running forever. The sexual reproduction is typical for complicated creatures. GAs with sexual reproduction can increase the efficiency and robustness, and thus they can track better optimal parameters in a changing environment [4], [7]. Diploid chromosomes increase the efficiency and robustness of GAs, and they can better track optimal parameters in a changing environment. Sex is distinguished by two-bit value stored in the chromosome [4]. It can be demonstrated that standard GAs with haploid chromosomes are unable to correctly locate optimal solutions for time-dependent objective functions. An interesting fact was found, namely that sexual reproduction (SR) and the immune system (IS) use two types selections working in parallel (IS: clonal and negative selection [11], SR: female and male selection [2]). Every selection mechanism solves one task (for example: convergence or adaptation). The system as a whole has then both desired features. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 18 ACKNOWLEDGMENTS. This work has been supported by MŠMT grant No: MSM 260000013 REFERENCES [1] PUVES,W. K., ORIANS, G. H., HELLER, H. C.: Life, the science of biology, Sinauer Associates, Inc. 1992. [2] BLACKMORE, S.: The Meme Machines, Oxford University Press 1999. [3] DAWKINS, R.: The Selfish Gene, Oxford University Press, Oxford 1976. [4] RYAN, C.: Shades: Polygenic Inheritance Scheme, MENDEL’97, Brno, Czech Republic (1997) 140-147. [5] KAUFFMAN, S.A.: The Origins of Order, Oxford University Press, New York 1993 [6] KAUFFMAN, S.A.: Investigations, Oxford University Press, New York 2000 [7] RIDLEY, M.: The Red Queen – sex and the evolution of nature, Penguin Books Ltd. 1993 [8] PRIGOGINE, I., STENGERS, I.: Order out of Chaos, Flamingo 1985 [9] OŠMERA, P., MASOPUST, P.: Schedule Optimization using Genetic algorithms, Proceedings of MENDEL’2002, Brno, Czech Republic (2002) 132 – 138. [10 OŠMERA, P.: Complex Evolutionary Structures, Proceedings of MENDEL’2002, Brno, Czech Republic (2002) 109 – 116. [11] OŠMERA, P., ROUPEC, J.: Limited Lifetime Genetic Algorithms in Comparison with Sexual Reproduction Based GAs, Proceedings of MENDEL’2000, Brno, Czech Republic (2000) 118 – 126. [12] ROUPEC, J.,OŠMERA, P., MATOUŠEK, R.: The Behavior of Genetic Algorithms in Dynamic Environment, Proceedings of MENDEL’2001, Brno, Czech Republic (2001) 84 - 90. [13] OŠMERA, P.: Complex Adaptive Systems, Proceedings of MENDEL’2001, Brno, Czech Republic (2001) 137 – 143. [14] OŠMERA, P.: Genetic Algorithms and their Aplications, the work to be habilitated, in Czech language (2002) 3 – 54. [15] DAVIS, L.: Job Shop Scheduling with Genetic Algorithms, International conference ICGA’85 (1985) 132138 [16] MORTON, T.E., PENTICO, D.W.: Heuristic Scheduling Systems, John Wiley and Sons 1993 [17] WALDROP, M.M: Complexity – The Emerging Science at Edge of Order and Chaos, Viking 1993 [18] FOGEL D.B.: A Brief History of Simulated Evolution, First Annual Conference on Evolutionary Programming, La Jolla, USA (1992) 1-16 [19] OŠMERA, P.: Complex Adaptive Systems, Proceedings of MENDEL’2001, Brno, Czech Republic (2001) 137 – 143 [20] OŠMERA, P.: Complex Evolutionary Structures, Proceedings of MENDEL’2002, Brno, Czech Republic (2002) 109 – 116. [21] OŠMERA, P.: Genetic Algorithms and their Aplications, the habilit work, in Czech language (2002) 3 – 114. [22] WALDROP, M.M: Complexity – The Emerging Science at Edge of Order and Chaos, Viking 1993 [23] COELLO C.A., CORTÉS N.C.: A Parallel Implementation of an Artificial Immune System to Handle Constrain in Genetic Algorithms, WCCI 2002, 819-824, Hawai [24] SANCHEZ-VELAZCO, J., BULLINARIA, J.A.: Sexual Selection with Competitive/Co-operative Operators for Genetic Algorithms, Proceedings of NCI’2003, Cancun, Mexico (2003) 308 – 316. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 19 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 20 DIAGNOSTIC APPROACH TO THE HUMAN CAPITAL DEVELOPMENT Alžbeta Kucharčíková, Jozef Vodák Department of Macro and Microeconomics Department of Managerial Theories Faculty of Management Science and Informatics University of Žilina [email protected], tel. 00421-041-565 10 12 [email protected], tel. 00421-041- 565 10 17 Abstract: This article describes the role of some diagnostics methods first of all identification and analysis of education and other needs of firms and balanced scorecard method for the improvement the enterprise processes and the development of the human capital. Keywords: enterprise diagnostics, STEEP analysis of external factors, SWOT analysis, benchmarking, cost-volume-profit analysis, human capital, identification and analysis of education and others needs of firms, balanced scorecard method. 1. INTRODUCTION The firms work in a new global economic age. Very fast installing of the new technologies, opening of the world market and increasing competition are some characteristics of present situation in which the companies must live. Firms must search for possibilities how to survive in this high turbulent environment. That means that they must find how to increase the quality of their production, how to better serve to their customers, how to improve their internal processes, how to improve potential of their human resources and how to measure all of this things. Good diagnostics seem very important for efficient way in increasing the ability of the companies to compete successfully on global markets. 2. ENTERPRISE DIAGNOSTICS Enterprise diagnostics is science approach which is using methods and techniques for the identification of problems in firms and for the problem solving. Advantage of this approach is its complex view on the company and its business processes. “Basically the term diagnosis originates from the Greek basis “dia” = though and “gnosis” = knowledge denoting the thorough knowledge which might metaphorically mean the distinguishing of the phenomena and processes, respectively the determination and defining of the state of the phenomena and processes using.” [1] There are several methods which may be applied for the diagnostic process, for example: General methods: · mathematical methods · comparative methods · time investigation methods · factor investigation methods Methods for evaluation of specific business process - traditional methods: · analysis of production · cost - benefit analysis · analysis of competition · marketing research · financial analysis INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 21 - modern methods · benchmarking · cost-volume-profit analysis · STEEP analysis of external factors · SWOT analysis · identification and analysis of education and other needs of firm · balanced scorecard method Benchmarking is a “continuous and systematic comparing of own performance in productivity, quality and production process with the enterprises and organisations presenting top- achievements. Benchmarking looks for the forms how to achieve competitive advantage. Enterprise production, services providing and procedures have to be continuously compared with the achievements of best competition and reputable marketleaders in all sectors.” [2] Cost-volume-profit analysis is a “systematic method of examining the relationship between changes in volume and changes in total sales revenue, expenses and net profit. As a model of these relationships, costvolume-profit analysis simplifies the real world conditions which a firm will face. Like most models which are abstractions from reality, cost-volume-profit analysis is subject to a number of underlying assumptions and limitations and it is s powerful tool for decision making. The objective of cost-volume-profit analysis is to establish what will happen to the financial results if a specified level of activity or volume fluctuates”. [3] STEEP analysis – external factors analysis This analysis consist of analysing the forces in social, technical, economical, environmental and political areas which influence the company. By characterisation each of starting points is suitable to start by external environment analysis. Performance of the company depend significantly on external forces and on interactions between the company and external environment. Good management is able to achieve excellent results in not very positive entrepreneurial environment. That means, good management is able to eliminate not very positive external forces for the company prosperity or is able to eliminate their impacts. SWOT analysis This tool consist of two views into company and two views outside of the company. Look inside consist of strengths and weaknesses of the company. The most important strengths are usually skilled employees, management, organisation culture and the ability of efficient production of the company products in expectation quality. Opportunities and threats are in the outside of a company. It is often changing entrepreneurial environment, changing competition conditions, changing needs and expectations of customers and economic situation in the territory. Good SWOT analysis is very suitable diagnostic tool for management of the company. Application of facility management enables to use some diagnostics methods, too. “Facility management for example in building trade is useful method for the future owners of buildings of various age, technical condition and facilities. It should be of benefit by optimising all supportive processes, specifying costs and their reduction at more effective area utilisation. Investors as well as developer´s organisations can benefit from facility management as well. The application of facility management helps them construct buildings of high practical value at low operation costs, providing projects of high quality and all services for tenants. “ [4] 3. THE HUMAN CAPITAL DEVELOPMENT At the present globalisation age are realised extensive industry changes. There are successful only firms which invest in new technologies and people working with this technologies. People are the major business assets. The firms need high skilled workers and so the human capital is one of the main factors of ability to compete growth of firms. Human capital is the sum of inborn or obtained knowledge, competencies, skills and experiences of the individuals. Investment in human capital is tended to development of knowledge and skills of workers. Firms invest to the human capital when they expect future benefits. Investments in human capital are realised as investments in education and training. “Correct organised education and systematic development of employees INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 22 competences is continuous process. This process of the employees`s education and development connects with the firm mission, objectives, philosophy, firm culture, strategy and policy of management and development of the human potential.” [5] There are some benefits of investment in education, for example growth of the production, services, quality, labour productivity, decrease in costs, other innovations, high quality relationship with customers, growth of the competitive ability on market. Application of the diagnostic approach for human capital development can be useful for growth of the quality of business processes and for effective business recourses decisions. Using the modern diagnostics methods – identification and analysis of education and other needs of firm is very effective in this case. 4. IDENTIFICATION AND ANALYSIS OF EDUCATION AND OTHERS NEEDS OF FIRM This method should identify skills gaps and also the other factors affecting performance problems in firm. Process of the identification starts by answering questions. For example: · Is performance of that skill really essential? · Is the employee rewarded for using that skill? · Does management discourage that behaviour? · What other barriers of the performance exist? The aim of this method is to identify a performance problems of employee by symptoms and to identify the reason of this problems. The reason can be in the changes of the motivation and behaviour of the people or in firm environment. This method`s result is a list of people`s education and others needs, design of the education programme and design of the others problems solving. Evaluation of obtained date gives us possibility to identify the other organisation problems, too. Data collection informations standards Data analysis · Identification of the achievement problems · Identification of problems causes Priorities of problem sphere Identification of education needs ******** Identification of others needs and measures Design of education programme Fig.1: Diagnostics model of identification and analysis of education and other needs process INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 23 5. USING BALANCED SCORECARD IN DEVELOPMENT HUMAN POTENTIAL Companies concern by questions about their strategy (creation and implementation), which will help them create pre-requisite for their successful life in high competition markets. Necessary part of the strategy is development of human resources, too. Therefore companies invest a great deal of their sources in this field. It is necessary to invest this resource cost-effective. This is the reason why companies try to find suitable learning methods for developing their employee potential and to find suitable methods for diagnostics and evaluation these learning processes. We want to point to one possible view of this problem through developing human potential in management of modern company by Balanced Scorecard (BSC) approach. This model is concern to achieve high effectiveness by investing into developing human resources for achieving company vision and strategy. Balanced Scorecard seems as very useful tool for creation and implementation of the company strategy in every day life. There are other reasons why is useful to use Balance Scorecard, too. One of them is need to look forward and create learning organisation. The learning organisation is characterise by developing, obtaining and carrying knowledge and by modification their behaviour towards new knowledge and opinion. In other words, learning organisation develop its ability to react, adapt and profit from changes in external and internal environment. The word “learning“ stress focus on knowledge, competencies and is tightly connected with intellectual capital. Learning of individuals is important as a base for collective learning of organisation. Balanced Scorecard can be used as management system for managing organisation or very effective measurement and diagnostic tool. Balanced Scorecard helps significantly match strategy and its implementation to company culture and help to recognise, understand and motivate employees. Next picture shows basic cause and effect principles through single perspectives in Balanced Scorecard. This picture re-cover clear understanding about this method and causality towards high level goals and company vision. Strategy focus of organisation begins with a translation of the strategy into operational terms. Balanced Scorecard is a very good method that helps translate strategy into operational objectives that drive employees both to the behaviour and performance. VISION and STRATEGY Financial perspective “To safety our shareholders, what, financial objectives must we accomplish?” Customer perspective “To achieve our financial objectives, what customer needs must we serve?” Internal processes perspective Learning and development perspective “To satisfy our customers, and shareholders, in which internal processes “To achieve our goals, how must our organisation learn and innovate ?” Fig.2: Perspectives in BSC and basic causality relations in Balanced Scorecard methodology Suitable system of evaluation together with suitable rewards system can very significantly motivate employees of the company. Measurement which is connected with strategy and evaluation of performance with feedback on strategy, is system, which realise organisation learning on strategic level (double loop learning). Realisation of implementation strategy with utilisation of potential Balanced Scorecard method brings: · · · · Objectives conformity on all levels of organisation with its strategy. Higher effectiveness measurement and control of the organisation performance. Realisation of strategic feedback and creation of clear communication platform . Maximization of investment effectiveness into human resources and other investments, too. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 24 Learning and development perspective of Balanced Scorecard - is basic presumption for successful achieving strategic company goals. In this perspective company develop goals and measures which support learning and development of the individuals and the whole organisation. Setting goals in finance, customer and perspective of internal processes determine where organisation must achieve excellent results for achieving strong increasing of its performance. The goals in learning and development perspective create assumption for successful achieving goals in other perspectives and move towards fulfilling its vision. Organisation which want to achieve success and fulfil stakeholders expectation must invest into infrastructure which consist of people, systems and procedures. To notify this ideas and fulfilling it helps successful implementation of the Balanced Scorecard methodology. Willingness to invest effort and money into this perspective is presumption for clear interconnection with goals at all corporate levels and understanding causal relation among perspectives and goals. Very important is question about evaluation of investing into employees. Basic areas of this perspective are: · · · Capability of systems (including information system ). Leadership, motivation, empowering, involvement. Capability and skills of employees. Some examples of measurement which can be useful for learning and development perspective with focus on human resources: · Leadership index (Number /No/) · Motivation index (No) · Number of employees (No) · Employee turnover (%) · Average employee years of service with company (No) · Average age of employees (No) · Time in training (days/year) (No) · Temporary employees / permanent employees (%) · Share of employees with university degrees (%) · Average absenteeism (No) · Number of woman managers (No) · Number of applicants for employment at the company (No) · Empowerment index (No), number of managers (No) · Share of employees less than 40 years old (%) · Per capita annual cost of training ($) · Full time or permanent employees who spend less than 50% of work hours at a corporate facility (No) In learning and development perspective companies often use some of mentioned measurements. It is important to know, that every measurement must be able to match concrete conditions and needs of the company. Probably only the ability of the company to learn more quickly as its competitors can be one of maintainable competition advantage for the future. For employees to see their own growth means motivation and one kind of satisfaction. Training learning activities are important for developing employee potential. Training means learning of training programmes for developing knowledge, skills and attitudes. It is important to notify that people first of all obtain skills which are important for their everyday work and for their work performance in training. Basic ideas about managing developing employee are based on assumption that human resources strategy contribute towards corporate strategy. This idea depends on that deal how organisation permit that employees create added value and they are therefore strategic source of the organisation. Through clear direction clear goals and effort which is concerned in one direction is possible by less energy achieve more, it is effect of synergy. In connection with developing human resources it means that by focusing into clear needs, which are derived from setting goals and metrics of the organisation (by BSC methodology) can company achieve its strategic goals more effectively than traditionally way. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 25 Next important dimension for human resources development is evaluation of training projects. In this model of development potential in modern company is therefore used four level model for evaluation investment into training projects. 6. CONCLUSION By ratification of the organisation budget is investment into developing human resources often on the last position but when the restrictions are needed they are at front line. Management know that investment into human resources are important, but the connection between development training programmes and achieving organisation goals is not clear enough. We believe that appropriate application of above mentioned diagnostics approaches can help managers in our organisations change this situation and in that way help their organisations. This work was supported by grant VEGA 1/0499/03. REFERENCES: [1] TOKARČÍKOVÁ,E.-KUCHARČÍKOVÁ,A.: Marketing diagnostics of 3G mobile communication possibilities. TRANSCOM 2003, University of Žilina, 23.-25.6.2003,s. 369-373 [2] STRIŠŠ,J.-REŠKOVÁ,M.: Benchmarking in service enterprises. ICSC 2003, European polytechnical institute Kunovice ČR, 30.-31.1.2003, s.45-50 [3] CHROMJAKOVÁ,F.- KNEŽNÍKOVÁ,M.: Cost-volume-profit analysis as an important input for financial analysis in the industrial enterprise. TRANSCOM 2001, University of Žilina 25-27.6.2001,s.231-235 [4] CHODASOVÁ,Z.: Application of facility management in Slovakia. Ekonomické a riadiace procesy v stavebníctve a v investičných projektoch, Slovak University of Technology in Bratislava – Faculty of Civil Engineering, 16.-17.9.2003, s.133-136 [5] BLAŠKOVÁ,M.: Riadenie a rozvoj ľudského potenciálu. EDIS Žilina, 2003 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 26 MARKETING ACTIVITIES OF SLOVAK BUSINESSES AFTER THE ENTRY TO EU Jozef Strišš1), Monika Rešková2) 1) European polytechnical institute, Osvobození 699, 686 04 Kunovice Phone: +420 572 549018, e-mail: [email protected] 2) University of Žilina, Faculty of management science and informatics, Department of management theories Moyzesova 20, 010 26 Žilina, Slovak republic, tel., fax: +421 41 5652 775 email: [email protected] Abstract: With entering the EU, new possibilities for operating on foreign markets will create for Slovak businesses; but at the same time, they will have to confront the foreign competition. The paper is dealing with marketing activities that Slovak businesses will have to accept if they want to be successful on foreign markets. Key marketing activities are: adaptation to global market conditions, choice of target markets and orientation on concrete segments, product innovation, services development, using information technologies within sale and communication, new forms of communication. Keywords: marketing, marketing communication, global markets, information technologies, ebusiness In May 2004, Slovak Republic and other countries will be accepted to EU. Joining economically forward European counties will mean outstanding change for Slovak business as well. New opportunities related to new, free markets will arise; but also competition conditions will change. Simpler approach to markets of new coming states for all the countries of EU, international companies with rich know-how and experience will be in advantage regarding long-term operating in market economics conditions. Business from new coming countrieswhere Slovakia belongs to as well- will have to be more qualitative and effective in the field of production quality, work productivity, cost reduction, more aggressive operating on the market, enhancing the communication with external environment. Successful entering of Slovak businesses on European Union markets is determined by marketing approach and higher marketing activities. Just marketing can be decisive factor within ensuring effectivity and prosperity. Marketing activities have to be evolved right now ; after the entry to EU their absence will have bad impact on market success. Most expressive marketing activities are: Adaptation to global market conditions. Global marketing is not only entering new markets and selling single product, but there is also ability of enterprise to understand single markets, differentiate them according to social, economical, cultural and national specifications. Using global marketing strategy, enterprise – using own strengths - tries to use all opportunities available on world markets with respecting the difference of single regions. Enterprise should learn to sell globally, but above all, think and consider in global merit. Selection of target markets and orientation on concrete segments. Basic marketing activity is selecting target markets – target customers. Every enterprise can offer to own target market something special with that, differentiate from the competition. Selecting target market should be activity made with special attention. When selecting decisive market segments, it is not enough to find target customer group, but it is necessary to feel customer thinking, to focus strategy on customer acquisition, to create business atmosphere (ability to sell products), to orientate the advertising on target customer group and on area of sale, to value (explore) the customer, to raise the customer to using new products. Activities of all enterprises’ bodies, every process has to be orientated on customer. It does not cover only marketing and commercial bodies, but complexly all the enterprise from management to executive employees. The enterprise has to evaluate own activities with the eyes of the customer and enterprise results should be focused on satisfying the customers. The enterprise objective should be permanent relationship with customers, attachment to the enterprise. It is INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 27 important to estimate customer value for the enterprise; it means profit that will reach will long-term (5, 10 years) contact. According to that, enterprise will determine expediency of contact and recourses that will invest to the customer (developing new products, delivery channels). Existence basis of every business is customer need. There is necessary to understand that customer does not look for product itself, but is searching for problem solution and benefit. Product innovation. Global marketing does not cover only single entrance on new markets, communication and selling relatively single product, but also its modification in global merit, it means product modification – its design and conversion in some ways jointly on all the markets with respecting national and cultural specifications and norms. Product should reach required standards on foreign and national markets. Global marketing includes also the enterprise ability to understand the markets and to differentiate them in term of cultural and national differences. Definite condition for success on the market is: · Permanent innovation and adaptation of products regarding the dynamics of demand to ensure apart from big innovations also smaller innovation, where experiences from competition products can be used ; it is possible to improve those products and services. When innovating, it is necessary not to forget customer wishes and to directly develop and produce products according to individual customer needs ; · High product quality – complete functional perfection based on newest science knowledge, adapted to customer needs. Enterprise can reach outstanding results if the products are far behind customer expectations and have such characteristics that will surprise him. Services development. In past few years, dynamics of services market is growing rapidly. Not only share of services as individual products is growing, but also services that are the part of main product arise. Basic principles that marketing has to respect are innovation and quality. Innovation can be based on better services organisation, or providing complement services that competition does not offer. It is suitable to use enterprise traditions when offering new services. Service marketing has to insist on quality of services provided. High standards, quality control and strong motivation of employees oriented on qualitative objectives have to be matter of course. Using new information technologies when acquiring commercial partners. Quick reaction to impulses in enterprise surrounding is basic factor of business success. Fundamental element of success is time. Enterprise has to react quickly on all the changes related to demand change, revenues, collision situations on markets and other problems. Building quality marketing information system has to be centred on information in enterprise surrounding. Lot of business absent marketing information systems, providing data about market changes, technical development, competition programmes etc. New information technologies enable the enterprise to create information and logistic networks, creation of intra-enterprise relations that lead to optimalisation of production, distribution, and communication and market activities and to cost reduction. New forms of communication with the market. Slovak enterprises will have to realise expressive changes in the field of marketing communication. In the past years, signals from market are clear – customer needs more that classic forms of marketing communication, aimed at one-way information flow. Market is becoming demanding and requires dialogue. It does not need to be personal meeting, but it can use modern forms of dialogue through interactive Internet abilities, e-mail, adds, questionnaires etc. These modern forms enable very fast product ordering by on-line system, optimal keeping of supplies and permanent customer informing. System of dialogue marketing is based on four fundamental phases: · privity, means offering the customer basic information with the possibility of feedback and convenience feeling; · interest, means by address dialogue or by personal meeting to convince the customer about the uniqueness of offer and product; · preferences, means dialogue has to be clear and straight for the purpose of acquiring the customer for purchase; · sale, means to achieve that customer will create friendly and long-term relationship with the enterprise, even after purchase. Using new information technologies in past years is also seen in business through the internet – in ebusiness. E-business includes using technologies, processes and management experience that improve competitiveness of the enterprise through strategic use of electronic information. E-business is new way of making business. Business that is based on global openness of economics, caused by information technologies INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 28 development, integrates of all the parts of the world into one „global village“. Term of e-commerce means exchange value of money, goods, services and information in electronic form. In connection with e-commerce, there is also the term of e-economics. E-economics means environment that e-commerce is running in and easily question some of basic macro economical thesis. When defining e-business, we have to see e-commerce as a factor that provides possibility of development. It is built on advantages and structure of tradition commerce, but it uses flexibility that electronic media and networks are offering. E-commerce of enterprise can connect critical business processes directly with its key parts. Key parts can be customers, employees, salesmen, suppliers, business partners and however having influence on business. 21. Century technology smoothes business and commerce. For being successful, enterprises have to understand basic parts of e-commerce. E-commerce is not ultimate objective itself, but developing process that enables the enterprise to work better. Enterprises transforming business styles know that, that is why they explore own strategies, techniques and tools in the light of new technologies. Companies have different looks and sizes; there is not one set of technologies of e-commerce suitable for all them. That is the keystone of ecommerce: Using technologies to build better relationships with customers, suppliers and employees. Corporate Identity is philosophy of enterprise operating and surviving in given market space, creating corporate image of enterprise – common factor for all the enterprise activities. Corporate identity is the mix of corporate culture (enterprise behaviour to customers, shareholders, public, employees), corporate design (external enterprise image – logo, symbols, colours, shapes) and corporate communication (advertising, PR, information). In competitive environment, there is often the corporate identity – form of enterprise communication, trademarks, and behaviour to public – that can persuade and acquire the customer. Creating the relationships with the customers. Lately, enterprise success is based on creating the relationships with the customers. Customers are becoming more a partner with who the enterprise creates not only formal, but very often informal, friendly relationships. Typical example is post-sale marketing. Characteristics of post-sale marketing can be in slogan: „Purchase of product or service to the customer does not mean the end of the relationship to the customer, it only begins.“ Process of post-sale marketing includes wide complex of activities that can be summed in areas: · processing of information files about customers, · elaborate method and form of raising and keeping the dialogue with the customers, · elaborate arrangements for keeping and improving customer satisfaction, · elaborate arrangements for acquisition of lost customers. Benchmarking – learning from the best. Benchmarking is permanent and systematic process of comparison and measuring products, processes and methods of own organisation with those, that were found as suitable for this measurement in order to defining the objectives of improving own activities. Practical meaning of benchmarking:: · helps to better understanding of customer requirements, · enables managers to gain information that would by only random process outcome · is the way of finding unbiased indicators of measuring own performance and productivity · one of the most effective processes of gaining impulses to own improvement. As the conclusion, it can be said that entering the EU does not have to be nightmare for our enterprises. When applying the principles of marketing, understanding customer meaning, its needs and satisfaction, Slovak enterprises can also be of use on foreign markets and be successful in competitive environment with foreign enterprises. REFERENCES: [1] HORÁKOVÁ, I. a kol.: Strategie firemní komunikace. Management Press. Praha. 2000. [2] JEDLIČKA, M.: Propagačná komunikácia podniku. MAGMA. Trnava. 2002. [3] KOTLER, P., Armstrong, G.: Marketing. GRADA. Praha. 2004. [4] MATEIDES, A., ĎAĎO, J.: Služby. EPOS. Ružomberok. 2002. [5] SMITH, P.: Moderní marketing. Computer Press. Praha. 2000. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 29 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 30 THE DECISION-MAKING PROCESS MODELLING Štefan Hittmár University of Žilina, Moyzesova 20 010 26 Žilina, Slovakia, tel.: ++421 41 5652 775 e-mail: [email protected], www.fria.utc.sk Abstract: Decision is a fundamental part of management because it requires choosing among variant courses of action. Decision making is a process of eight steps that include identifying a problem, selecting an variant, and evaluating the decision´s effectiveness. In the following you can find a processed model of decision making in management. Keywords: Management, Analyzing, Decision-making, Process, Information, Problem, Criteria, Variant, Implementing Evaluating 1. INTRODUCTION Management styles vary. Some managers follow their instincts and deal with situations on a case-bycase basis. Other managers use a more systematic and structured approach to making decisions. Decision making is part of every aspect of the manager´s duties, which include planning, organizing, staffing, leading, and controlling. For example, managers can formulate planning objectives only after making decisions about the organization´s basic mission. To accomplish the objectives within some time period, decisions must be made on what resources are required. When pursuing company objectives, management must make decisions on the division of labor and reporting relationships. Prospective staff members are identified and selected according to what the established positions require. The organization pursues its objectives through countless daily decisions required to make or sell the product or service. And decisions about whether to take corrective action when performance does not measure up to standards affect operations. Decision making is typically described as „choosing among variants“. But this view is too simple. Why? Because decision making is a comprehensive process, not just a simple act of choosing among variants. 2. MODEL OF DECISION-MAKING PROCESS The decision-making process is illustrated on Figure 1. This process is a set of eight steps that begins with identifying a problem and decision criteria, and allocating weights to those criteria; moves to developing, analyzing, and selecting a variant that can resolve the problem; implements the variant; and concludes with evaluating the decision´s effectiveness. . s t e p 1. Identification of a Problem s t e p 2. Identification of Decision Criteria s t e p 3. Allocation of Weights to Criteria s t e p 4. Development of Variants INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Ku novice, Czech Repu blic, Ja nu a ry 2 9 -3 0 , 2 0 0 4 31 . s t e p 5. Analysis of Variants s t e p 6. Selection of a Variant s t e p 7. Implementation of the Variant s t e p 8. Evaluation of Decision Effectiveness Figure 1. The Model of Decision-Making Process Step 1: Identifying a Problem The decision-making process begins the existence of a problem or, more specifically, a discerepancy between an existing and a desired state of affairs. Before something can be characterized as a problem, managers have to be aware of the discrepancy, they have to be under pressure to take action, and they must have the resources necessary to take action. (See Figure 2.) How do managers become aware that they have a discrepancy? They obviously have to make a comparison between their current state of affairs and some standard. What is that standard? It can be past performance, previously set goals, or the performance of some other unit within the organization or in other organizations. Avareness of discrepancy Pressure to act Problem Sufficient resources to do something Problem: A discrepancy between an existing and a desired state of affairs. Figure 2. The Characteristics of a Problem Step 2: Identifying Decision Criteria Once a manager has identified a problem that needs attention, the decision criteria important to resolving the problem must be identified. That is, managers must determine what is relevant in making a decision. Whether explicitly stated or not, every decision maker has criteria that quide his or her decision. Note that in this step in the decision-making process, what is not identified is as important as what is. Step 3: Allocating Weights to the Criteria The criteria listed in the previous step are not all equally important so the items must be weighted in order to give them the correct priority in the decision. How does the decision maker weight criteria? A simple approach is merely to give the most important INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Ku novice, Czech Repu blic, Ja nu a ry 2 9 -3 0 , 2 0 0 4 32 criterion a weight of 10 and then assign weights to the rest against this standard. Thus, in contrast to a criterion that you gave a 5, the highest factor would be twice as important. Of course, you could use 100 or 1,000 or any number you select as the highest weight. The idea is to use your personal preferences to assign a priority to the relevant criteria in your decision as well as to indicate their degree of importance by assigning a weight to each. Step 4: Developing Variants The fourth step requires the decision maker to list the viable variants that could resolve the problem No attempt is made in this step to evaluate these variants, only to list them. Step 5: Analyzing Variants Once the variants have been identified, the decision maker must critically analyze each one. The strengths and weaknesses of each variant become evident as they are compared with the criteria and weights established in steps 2 and 3. Each variant is evaluated by appraising it against the criteria. Step 6: Selecting an Variant The sixth step is the crucial act of choosing the best variant from among those listed and assessed. Since we have determined all the pertinent factors in the decision, weighted them appropriately, and identified the viable variants, we merely have to choose the variant that generated the highest score in step 5. Step 7: Implementing the Variant While the choice process is completed in the previous step, the decision may still fail if it isn´t implemented properly. Therefore, step 7 is concerned with putting the decision into action. Implementation includes conveying the decision to those affected and getting their commitment to it. Groups or teams can help a manager achieve commitment. If the people who must carry out a decision participate in the process, they are more likely to enthusiastically support the outcome. Step 8: Evaluating Decision Effectiveness The last step in the decision-making process appraises the result of the decision to see whether the problem has been resolved. The variant chosen in step 6 and implemented in step 7 accomplish the desired result. What happens if, as a result of this evaluation, the problem is found to still exist? The manager then needs to dissect carefully what went wrong. Was the problem incorrectly defined? Were errors made in evaluating the various variants? Was the right variant selected but improperly implemented? Answers to such questions might send the manager back to one of the earlier steps. It might even require starting the whole decision process over. CONCLUSION Decision-making process is very important part of managerial functions. It has been realised as systematic and structured approach to every management activity. The “eight steps model“ is a tool for improvement of managerial works. REFERENCES: [1] ROBINS,S.P.,COULTER,M. Management. Prentice Hall, Inc. 1999 [2] HITTMÁR,Š. The decision-making process. MOSMIC, VŠD Žilina, 1997 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Ku novice, Czech Repu blic, Ja nu a ry 2 9 -3 0 , 2 0 0 4 33 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Ku novice, Czech Repu blic, Ja nu a ry 2 9 -3 0 , 2 0 0 4 34 MARKETING AND ITS GLOBAL ORIENTATION ON TURN OF THE CENTURIES Milan Jedlička MtF STU Trnava, Paulínska 16, 917 26 Trnava, SLOVAKIA phone: +42133551032, e-mail: [email protected] Abstract:The exploitation of marketing in global market has to have the systemic fundamentals. The expression of systemic fundamentals particularly depends on these decisive premises: · mutual synergy of analytic and synthetic actions, · continual and consecutive execution of marketing actions, · mutual synergy among several managing elements as a part of marketing management, · creating various decisive variants which will dynamically respond to changing market situation, · providing feedback among plans and final evaluations of marketing activities to possible exploitation of experience and knowledge at further strategic decision making. · marketing communication and propagation communication and its essence and signification in business practice, · supply to demand and demand to supply philosophical orientation of propagation communication, · personal selling, sales promotion, public relations and advertising as the main techniques of propagation communication activating, · hierarchy business communication, Keywords: market, marketing manager, customer, global marketing, marketing communication, propagation communication 1. INTRODUCTION Marketing as a branch or particular management action reaches a rapid development at present. Such trend which has begun in post-war period (especially in sixties) became step by step more dynamic. It bounds up with several factors, from which these can be signified as decisive: · increasing the technological level and thereby the product manufacturing in single commodities, · branching out the supply not only by new material products but especially by services, · gradual territorial development and branching out the world-wide market, · forming the new types of managers, which have more complex training and approaches to customer, · using the new theoretical and practical knowledge and skills, · huge development of informatics and following increasing of communication quantity and quality, · growth of purchasing power almost all around the world etc. Of course, this trend was supported in last years by transformation of many states from "directive forms of management" to market economy. The most advanced states of socialist type had very well developed industrial production, high level of education and built up infrastructure on solid level. These were the positive starting-points for economy transformation. But to reach the effective change, there was, there is and there will be a lot of things to do. This is also not possible without appreciation of some most important trends in development and using of marketing in real management practice. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 35 Aim Therefore this paper aims to call the attention to some trends in marketing which are related with globalisation. But there are general formulating philosophical starting-points, which every enterprise subject must retransform to its own conditions. Problem Analysis Although most experts very quickly understand the fundamental idea of transformation process - if we want to create market and market relations, it is not possible without marketing - concrete practice already shows something different. Let me generalize some of my observations, which I have gained during the longer time activities in economic practice, but also during the activities at the University, especially during the special training courses for managers. I took part at about eighty courses of low, middle and top management. Besides I acted as the consultant in about ten companies. Decisive observation are following: · the necessity of special training of managers is understood almost by everyone, but it is necessary to train another ones and not yourself, · new organizational and decision approaches are accepted by everyone, but almost nobody uses them · in many companies the board "commercial or sale department" was repainted on marketing department and that was the end, · intro-preparation of training course the most frequents requirement was following: "teach us do marketing in one day", so as soon and cheap as possible. · minimal will to change the well-running procedures, to think strategic and long-time, to awake the need of all company departments co-operation etc. There is a lot of problems in markets where the transformation of economy is running. We can not solve all these problems and therefore I will discuss the problem of marketing communication in this paper. Marketing communication and especially its specific field, propagation communication, is today one of the determining activities of every company which makes business on markets where is big competition. Marketing communication has great signification in business activity and also in macrolevel, as well as in microlevel, in external and internal environment of company. It provides the necessary communication relations in existing market reality, in supply to demand processes. Well not only in this reduced understanding, in market environment, where in prevalent measure it came into being and branched out after several centuries, if we take to reflections the documented propagation, advertising or adman artefacts. In wide interpretation we can say even about many millenniums, during which it influenced not only the economic, but also the social life. In history there were rather singular cases in more adult civilisations or economics. Today we can observe, that its signification dramatically increased and its exploitation has massive character. It is the massive communication on all levels of company with activation of every business subject, which wants to be successful and it has effort to address and persuade almost every potential customer. Marketing communication mainly influences and enlarges not only social material consumption, but also immaterial one, which is represented by many social or cultural displays. It becomes generally accepted cumulation of several activities and professions, connection of social, economic, political and artistic information. Its aim is not only to sell but also to penetrate into awareness, to influence thought level or amuse. It creates synergical unity of the rational, but also the irrational, the pragmatical or emotional elements, which in final consequence influence most effectively on final decision of asking subjects. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 36 As it was mentioned the term marketing communication is called promotion or propagation communication. This fact is the result of this reasons: · in European area is typically used the term “propagation” e.g. against American, where it is used especially the term “promotion“, · the contents of this word in its simplificated form very exactly represents its signification for this field of communication, · some authors use as the covering term “advertising“ (e.g. J.Prachar), but this term is by most authors considered only like a subsystem of propagation (promotion), which even follows too from its contents. The propagation can be considered as the set of specific communication techniques and their active, systematic, synergical and effective activity in benefit of company aims. There are used the sources in the company but also out of the company (ideas, philosophies, techniques, equipment similarly) on the influence in market environment (external also internal company environment) with aim to spread the desirable information and multiply total number of the customers preferring and using company supply. Decisively it is a very arduous process, which needs to be performed by very knowledgeable and creative managerial collective. On the designation of suitable aims, providing the adequate procedures and gaining expected results propagation strategy must be work-out (long-time plan of activities) and also its particular executable plans at shortly time period. The philosophical starting-points for application of marketing communication Making clear the substance of philosophical orientation of marketing communication has great signification. Company must soon, as it starts, consider about its application, make sure this main philosophical starting points: · present, respectively expected state of market environment, · present, res. expected state of company environment, · specification of the primary aim of propagation and orientation of its influence. · Without needful analyse of main starting points is not possible to create and use propagating communication. Therefore it is necessary to know the state at market and in company, to define the determining characteristics, fortes and weaknesses. As a starting point for finding and decision making can help the designation of the determining factors which influence the social, respectively market communication (see figure No.1). Social level Need of communication Motive of communication Communicatio Supply assortment n environment Ability of communication Thought level Figure 1: Determining factors of the communication environment INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 37 · · · · · · · On one side it is necessary to consider factors connected with customer and its individual state: need of communication (the basis of consumption behaviour, substance of individual qualification and quantification, hierarchy of satiation of wants and their intensity etc.), motive of communication (the natural necessity, emotional reaction, motivation and involvement etc.), ability of communication (individual supposition, qualities, preparation, skills, know-how etc.). On the other side in general social and market level there are these determining factors: social level (conditions which creates concrete society in field of law, economic tools, services etc.), thought level (the state in which situation the population is as the demand communicator which except the individual factors has also social characteristics as the looking at political climate, living level, social welfare etc.), supply assortment (quality and quantity of supply, communication techniques, tools, equipment and forms, competitive level of communication etc.). The result of such a primary analysis should be the designation of the centrobaric type of customers, their nearby surroundings, ability of influence and motivation other ones, observing of social traditions, designation the tendencies and trends, climate etc. From present starting points we can designate the main aims of propagation impact and primary orientation which can have two primary philosophical tendencies: · supply to demand, · demand to supply. The position of marketing communication in business practice Marketing communication or specifically propagation communication in field of marketing management became the most important tool of marketing mix because if it is applied in effective way it mainly influences still not entirely decided or hands-off customers. And that was the reason why then new management method “customer relationship management” came into existence. While product, price, distribution, place or another used tools of marketing mix have in content of its activity more rationality and less widely interpreted matter, where propagation is essentially less merge, more creative, more striking, res. more reactive and dynamic. There is more taste of man in it, propagator or seller. Just personal approach, individual care, strength figure, something more than only selling goods for some price and obtained money, these are very important factors for customer. Today propagation did not became by mistake the first tool of marketing mix. In present market practice it is the main retaining starting point by which the company introduces its own supply and wants influence the market environment by it. In still increasing competition, in increasing amount of offers at market, there information chaos occurs, even some disorientation of customer. Without the right solution of customer and “its redirection to your own side“ the long-term market success can be very difficult expected. If the company wants to be heard and accepted, it has eminently to change and creatively develop communication approaches, it has to have clearly and clever estimated but also executed communication strategy. And exactly propagation and then propagation strategy are the bearing columns of communication activities of the company in market economic. Analogous like the task of marketing mix in its strategic expression is to properly use the single tools, so is that in case of propagation strategy. In prevalent number of technical literature there are mentioned as primary propagation tools these four (2): · personal sales, · sales promotion, · public relations, · advertising. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 38 Their synergical combination or multichannel affect on customer is sometimes also called “communication mix“ (4), res. for correct differentiation it is better to use term “propagation mix“ because as it was already mentioned, the number of communication tools of company is essentially wider as four specific tools of propagation. Besides other tools of marketing mix, that means product, price, distribution or place, it has also an immutable communication signification. In wide view, propagation communication is closely connected with and goes out from the general company communication, respectively in low hierarchical level from marketing communication (see figure No.2). Company Company information communication system Marketing Marketing information communication system Propagation Propagation communication information system Advertising information Advertising communication Figure 2: Hierarchy of company communication Application of management method CRM There exist several conceptions in application of CRM method. If the conception is oriented on specific market segments, company has to search systemic approach to marketing communication management with these segmetns, respectively with their main representatives (leader of opinion, etc.). Following this conception (see fig.3) the market is divide into four basic segments: · general public (any member of public), · potential customers (which preferred the company), · decided customers (evident order), · consumers (at least one purchase). P R active reactive conception General public Potential customers Decided customers A S P P S R M C Consu- CK mers R M Figure 3: The conception oriented on market segments · · · The main tasks in application of this conception approach are as follows: to better determine the content of the needs, motives and wishes of these segments, to determine the borders among the segments and also specify their quantity and quality of requirements, the are used as the tools of promotion or propagation communication (personal selling, promotion sale, public relation, advertising), also CRM, but there is another specification following the KCRM method (Key Costumers Relationship Management) which is especially concentrated on costumers). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 39 Designed Solutions The new trends, which would have come on after overcoming the present failures, can be summarised into five following suppositions: · mutual synergy of analytic and synthetic actions, · mutual synergy among several managing elements as a part of marketing management, · continuity, flexibility and dynamics of the marketing actions execution , · creating various decisive variants which will dynamically react on changing market situation, · providing feedback among plans and final evaluations of marketing activities to possible exploitation of experience and knowledge at further strategic decision making. 1. Supposition, which is associated with the synergy of analytic and synthetic activities, aims the consistent systemic connection of market research, its segmentation on decisive parts, and creating the marketing strategies following the specific requirements of its participants. 2. In advanced companies marketing isn't already only an individual activity, but it is collective concept of marketing management. The harmony among several managing elements, that means methods, instruments, facilities and forms, which are used for reaching the market effects, is inevitable. This supposition is straight linked up to the previous one and it elaborates and deepens this one moreover. 3. Continuity, flexibility and dynamics of the marketing actions execution also belongs to systemic suppositions. Market, its elements and relations among them are still complicated and they have mainly subjective and stochastic character. Therefore, if we want to reach large a telling ability of the information about market, we must recognise the market longer, watch its development and we must consequently recognise the mutual relations among various market subjects and objects. To eternally changing conditions it is necessary to respond flexible following the supply-demand motion and dynamically use various marketing instruments. 4. There must be the ambition to enforce strategic decision-making, but it is necessary to reflect, that in this case it is typical to decide with risk and in vagueness. Such reality requires the creating of several decision-making variants for different situations of the market surround. 5. The feedback as a systemic condition results in comparison of plans and results and it gives manager the real vision about the true state of the enterprise activities. It is mutually necessary to connect and synchronise rather regulation arrangements at the operative level and rather inspection mechanisms at the strategic level. CONCLUSION Marketing and its orientation and application on turn of the centuries must become the immanent part of company management and must accept the latest trends of market that becomes more globalized. It is not sufficient to discuss this problem, accept it on a formal way and reduce it only for example to the problem of advertising or creating of image. At pragmatic economic level such marketing orientation creates however the outlays first of all, while marketing should use in a complex view to create the market effects connected with the adequate profit. Without understanding the basic contents and substance of marketing communication, acquiring the necessary information, creating the philosophy and correct assigning the aim oriented on customers and without strategic oriented management, the serious result of entrepreneurial activities can not be expected. REFERENCES: [1] DOHNAL, J..: Řízení vztahů se zákazníky, Grada Publishing Praha, 2002 [2] JEDLIČKA, M.: Propagačná komunikácia podniku, Magna Trnava, 2000 [3] JEDLIČKA, M.: Marketingový strategický manažment, Magna Trnava, 2003 [4] LABSKÁ, H.: Marketingová komunikácia, JUP Nové Zámky, 1993. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 40 METHODS USED IN MARKETING SEGMENTATION PROCESS Monika Rešková University of Žilina, Faculty of management science and informatics, Department of management theories Moyzesova 20, 010 26 Žilina, Slovak Republic tel: +421415134458, e-mail: [email protected] Abstract: To compete successfully in today’s volatile and competitive business markets, mass marketing is no longer a viable option for most companies. Marketers must attack niche markets that exhibit unique needs & wants. Market segmentation is the process of partitioning markets into groups of potential customers with similar needs or characteristics who are likely to exhibit similar purchase behaviour. Market segmentation is the foundation on which all other marketing actions can be based. It requires a major commitment by management to customer-oriented planning, research, implementation & control. The paper deals with the term of market segmentation, describes market segmentation process and focuses on most frequent and useful techniques used in market segmentation process. Keywords: Market segmentation, marketing strategy, positioning, targeting, variables 1. WHAT IS MARKET SEGMENTATION? Market segmentation describes the division of a market into homogeneous groups which will respond differently to promotions, communications, advertising and other marketing mix variables. Each group, or “segment”, can be targeted by a different marketing mix because the segments are created to minimize inherent differences between respondents within each segment and maximize differences between each segment. Market segmentation was first described in the 1950’s, when product differentiation was the primary marketing strategy used. In the 1970’s and 1980’s, market segmentation began to take off as a means of expanding sales and obtaining competitive advantages. In the 1990’s, target or direct marketers use many sophisticated techniques, including market segmentation, to reach potential buyers with the most customized offering possible. 2. USING MARKET SEGMENTATION There are many good reasons for dividing a market into smaller segments. The primary reasons are: · Easier marketing It is easier to address the needs of smaller groups of customers, particularly if they have many characteristics in common (e.g. seek the same benefits, same age, gender, etc. · Efficient. More efficient use of marketing resources by focusing on the best segments for your offering – product, price, promotion, and place (distribution). Segmentation can help you avoid sending the wrong message or sending your message to the wrong people. · Find niches. Identify under-served or un-served markets. Using “niche marketing”, segmentation can allow a new company or new product to target less contested buyers and help a mature product seek new buyers. 3. CONDITIONS FOR USING MARKET SEGMENTATION Any time you suspect there are significant, measurable differences in your market, you should consider market segmentation. Identified segments must be: · Big enough. Market must be large enough to warrant segmenting. Don’t try to split a market that is already very small. · Different. Differences must exist between members of the market and these differences must be measurable INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 41 · · · · through traditional data collection approaches (i.e., surveys). Responsive. Once the market is segmented, you must be able to design marketing communications that address the needs of the desired segments. If you can’t develop promotions and advertising that speak to each segment, there is little value in knowing that those segments exist. Reachable. Each segment must be reachable through one or more media. You must be able to get your message in front of the right market segments for it to be effective. If one-eyed, green aliens are your best marketing opportunity, make certain there is a magazine, cable program or some other medium that targets these people (or be prepared to create one). Interested in different benefits. Segments must not only differ on demographic and psychographic characteristics, they must also differ on the benefits sought from the product. If everyone ultimately wants the same things from your product, there is no reason to segment buyers. However, this is seldom the case. Even commodities like sugar and paper plates can benefit from segmentation. Profitable. The expected profits from expanding your markets and more effectively reaching buyer segments must exceed the costs of developing multiple marketing programs, re-designing existing products and/or creating new products to reach those segments. 4. MARKET SEGMENTATION PROCESS There are two basic ways to segment a market: · A priori. A priori segmentation involves dividing a market into segments without the benefit of primary market research. Manager intuition, analysis of secondary data sources, analysis of internal customer databases or other methods are used to group people into various segments. Previous “post hoc” segmentation studies are considered to be “a priori” when applied to the same markets at some point in the future. · Post hoc. Primary market research is used to collect classification and descriptor variables for members of the target market. Segments are not defined until after collection and analysis of all relevant information. Multivariate analytical techniques are used to define each segment and develop a scoring algorithm for placing all members of the target market into segments. There are two types of information used in market segmentation: Classification variables. Classification variables are used to classify survey respondents into market segments. Almost any demographic, geographic, psychographic or behavioural variable can be used to classify people into segments. · Demographic variables – Age, gender, income, ethnicity, marital status, education, occupation, household size, length of residence, type of residence, etc · Geographic variables – City, state, zip code, census tract, county, region, metropolitan or rural location, population density, climate, etc. · Psychographic variables – Attitudes, lifestyle, hobbies, risk aversion, personality traits, leadership traits, magazines read, television programs watched, PRIZM clusters, etc. · Behavioural variables – Brand loyalty, usage level, benefits sought, distribution channels used, reaction to marketing factors, etc. Descriptor variables. Descriptors are used to describe each segment and distinguish one group from the others. Descriptor variables must be easily obtainable measures or linkable to easily obtainable measures that exist in or can be appended to customer files. Many of the classification variables can be considered descriptor variables. However, only a small portion of those classification/descriptor variables are readily available from secondary sources. The trick is to identify descriptor variables that effectively segment the market in the primary research effort which are also available or can be appended to individual customer records in customer databases. The process of market segmentation, targeting and positioning is outlined in Figure 1. Phases 1 and 2 of the process involve gathering and analysing market data to identify possible market segments. From this data a range of possible market segments is identified which then requires further detailed analysing and planning. These market segments are evaluated to determine which are worth targeting in terms of commercial viability and potential growth. The selected segments will represent the target market that the organisation wishes to reach, and the marketer moves on to phase 3. The product’s position in the market is determined and appropriate marketing strategies developed and implemented. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 42 Phase 1 – Segmentation 1. 2. 3. Market definition Choosing the basis for segmentation Dividing the market and profiling the segments Phase 2 – Targeting 4. Valuating the potential and commercial attractiveness of each segment 5. Choosing one or more segments Phase 3 – Product positioning 6. Defining product positioning 7. Develop appropriate marketing strategies 8. Implementation marketing strategies Figure 1: 8-step algorithm of market segmentation 5. METHODS OF MARKET SEGMENTATION Most multivariate analytical techniques can be used and probably have been used in some way to create post hoc market segments. There is no ideal methodology that works with every segmentation study. Each methodology has advantages and disadvantages. Segmentation studies generally require the use of two or more methodologies to produce the best results. In nearly every case, multiple techniques should be tested before selecting the "best" solution. There are 3 categories of analytical techniques applied to market segmentation: data preparation, data analysis, and classification. Data Preparation Numerous techniques can be used to aide the segmentation process. Factor analysis. Factor analysis can reduce the number of variables to a more manageable size while also removing correlations between each variable. Correspondence analysis. The coordinates produced by correspondence analysis, when calculated at the individual or group level, can be clustered to produce market segments. Correspondence analysis can also be used to convert nominal data (like yes/no answers) to metric scales. Conjoint analysis. Utilities from conjoint analyses can be used in segmentation because they represent the relative value individuals' place on all key attributes that define a product or service. In fact, conjoint utilities represent the most effective basis variables because they are derived from respondent preferences between product options or from actual choices of preferred products. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 43 Data Analysis Cluster analysis. Cluster analysis is the most frequently used method of segmenting a market. The underlying definitions of cluster analysis procedures mimic the goals of market segmentation: to identify groups of respondents in a manner that minimizes differences between members of each group while maximizing differences between members of a group and those in all other groups. However, there is one key difference between clustering and segmenting respondents -- clusters produce groups of respondents who have similar responses on key variables while segmentation finds groups of respondents who have similar behaviours when purchasing and seeking products in the market. Both hierarchical and iterative cluster analysis procedures can be used, but hierarchical procedures are difficult to evaluate once you exceed 100 or 200 survey respondents. Among the various iterative cluster analysis procedures, the K-Means method is most often used. K-Means cluster analysis can be found in all of the most popular statistical programs (SAS, SPSS, BMDP, Statistica, SYSTAT). Chi-square Automatic Interaction Detection (CHAID) or Classification and Regression Trees (CART). CHAID and CART are known as "Classification Tree Methods." These methods divide respondents into groups and then further divide each group into subgroups based on relationships between segmenting basis variables and some dependent variable. The dependent variable is usually a key indicator such as usage level, purchase intent, etc. These procedures create tree diagrams, starting at the top with all respondents combined and then branching into 2 or more groups at each new level of the tree. Subdivisions are determined by finding the survey variable that produces the greatest difference in the dependent variable among individual response categories or groups of response categories on that survey variable. CHAID is the most commonly used classification tree method, but it can not handle continuous dependent variables so a combination of CHAID and CART is sometimes used. Both CHAID and CART have the ability to process non-metric and non-ordinal data. Unlike cluster analysis, classification tree methods create true segments when they divide respondents. However, these segments are only based on one dependent variable. Other methods, including cluster analysis, divide respondents based on 10's or even 100's of data elements. Artificial neural networks. Artificial Neural Networks or ANNs offer another means to segment respondents. The Kohonnen architecture is one self-organizing ANN that can be used for segmentation. It is called self-organizing because, like cluster analysis, there is no dependent variable specified in the model. The ANN attempts to group respondents based on their similarities. It differs from cluster analysis in its ability to ignore noisy data. Atypical individuals have less impact on the segmenting calculations and each successive iteration makes ever smaller adjustments to the network weights so the calculations quickly stabilize, ignoring infrequent respondent characteristics. The greater the variation or uncertainty in respondents' answers, the better ANNs perform compared to cluster analysis. Latent class structure models. Latent class analysis is often described as "factor analysis for categorical variables." It is used to find underlying constructs within sets of variables. However, latent class analysis can also be used to cluster categorical variables into segments based on responses across a broad range of categorical variables. Latent classes attempt to find the underlying constructs which motivate people to buy a particular product or to desire certain features in that product. Classification There are a number of classification algorithms or analytical methods which can be applied to market segmentation. Discriminant analysis. Discriminant analysis can be used to classify respondents into predefined segments based on descriptor variables like census data. The segmentation scheme determines which respondents belong in each market segment. The classification or scoring program then creates the means of identifying potential members of each segment based on limited information (usually data which can be obtained from secondary sources). When a limited set of information can be used to accurately predict which market segment each individual belongs, you have a successful classification algorithm. Multiple regression. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 44 Multiple regression and multinomial logit can be used in the same manner to create classification schemes for your market segments. 6. MARKET SEGMENTATION OUTCOME Unfortunately, there is no definitive answer to the question how many segments should be the outcome of market segmentation. Experience, intuition, statistical results and common sense all must be applied to decide on the number of segments to retain. If you have several very small segments, you may need to change the criteria for segmentation or remove some of these respondents as outliers. Too many segments can lead to developing many different marketing programs for small, very similar, markets. But there are a few rules of thumb for segmentation: · Large enough. Majority of segments must be large enough to be economically feasible to target marketing and product design efforts. · Relevant. The segments must be relevant to your company's products/services. · Reachable. Segments must be reachable through one or more marketing mix variables (price, promotion, features or distribution). · Different. There must be clearly defined differences between market segments to make some segments more desirable than others. If many of the segments want essentially the same features and intend to buy at the same frequency or volume level, then these segments do not exhibit meaningful differences. 7. CONCLUSION The paper has outlined the market segmentation process. This process lies at the heart of overall philosophy of marketing. Success stories abound of companies who have successfully adopted and implemented market segmentation into their planning process. Equally, stories of failures are all to frequent as companies poorly define their markets, treat all customers in that market the same, do not evaluate market segments rigorously or finally, fail to position the product appropriately or communicate this position effectively. Market segmentation techniques are becoming increasingly sophisticated, but the important principles that underline this process will always remain essential for the development of effective market strategies. REFERENCES [1] KOTLER, P.: Marketing management. Grada Publishing, spol .s.r.o. 1998 [2] REŠKOVÁ, M.: Rozhodovacie procesy v marketingu so zameraním na výber cieľových trhov v osobnej železničnej doprave. (písomná práca k dizertačnej skúške), 2002 [3] McCARTHY, J., E., PERREAULT, W., D.: Základy marketingu. Victoria Publishing a.s., 1995 [4] SHARMA, A., LAMBERT, D., M.: Segmentation of Markets Based on Customer Service. International Journal of Physical Distribution and Logistic Management, vol. 24. MCB University Press, p. 50-58, 1994. [5] HRUSCHKA, H., NATTER, M.: Comparing Performance of Feedforward Neural Nets and K-means for Cluster-based Market Segmentation. European Journal of Operational Research, p. 346-353, 1999. [6] LIN, Ch., F.: Segmenting Customer Brand Preference: Demographic or Psychographic. Journal of Product and Brand Management, vol. 11. MCB UP Limited, p. 249-268, 2002. [7] SNYDER, R.: Market Segmentation: Successfully Targeting the Mature Population. The Journal of Active Aging, 2002. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 45 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 46 ADVERTISING AND GLOBALIZATION INFLUENCES Achalova Larisa Vladislavovna Plehanov Russian Academy of Economics, the faculty of International Economic Relations, Moscow, Stremiannii alley., 36, tel. (095) 236-40-94, (095) 237-95-52, [email protected] Abstract: The article will present the topic concerning advertising and the influence of globalization. Globalization is today’s reality. Advertising is an important part of the international marketing program of firms competing in the global marketplace. There is a chance for companies to use advantages of globalization in creating new global advertising campaigns. Keywords: advertising; advertising objective, budget, media, message; direct-response advertising, advertising to the consumer market, to the business and professional markets; international advertising Advertising is an integral part of our social and economic systems. In the complex society and economic system in which we live, advertising has evolved into a vital communication system for both consumers and business. The ability of advertising and other promotional methods to deliver carefully prepared messages to targeted audience has given them a major role in the marketing programs of most organizations. Companies ranging from large multinationals corporations to small retailers increasingly rely on advertising and promotion to help them market products and services. In market-based economies, consumers have learned to rely on advertising and other forms of promotion to provide them with information they can use in making purchase decisions. Hence, today companies must do more than make goods – they must inform consumers about product benefits and carefully position products in consumers’ minds. To do this, they must skillfully use the masspromotion tools of advertising, sales promotion, and public relation. We define advertising as any paid form of nonpersonal presentation and promotion of ideas, goods, or services by an identified sponsor. Several aspects of this definition should be noted. First, the paid aspect of this definition reflects the fact that the space or time for an advertising message generally must be bought. The nonpersonal component indicates advertising involves mass media (e.g., television, radio, magazines, newspapers) whereby a message can be transmitted to large groups of individuals, often at the same time. The nonpersonal nature of advertising means there is generally no opportunity for immediate feedback from the message is sent, the advertiser must attempt to understand how the audience will interpret and respond to the message. Also, advertising can be defined as a method of delivering a message from a sponsor, through an impersonal medium, to many people (the word ad comes from the Latin ad vertere, meaning “to turn the mind toward”). According to the Longman Business English Dictionary, advertising is a form of telling people publicity about a product or service in order to persuade them to buy it. The roots of advertising can be traced back to early history. Although advertising is used mostly by private enterprise, it also is used by a wide range of other organizations and agencies, postal service and branches of the armed services. Advertising is a good way to inform and persuade, whether the purpose is to sell Coca-Cola worldwide or to get consumers in a developing nation to drink milk or use birth control. Different organizations handle advertising in different ways. In small companies, advertising might be handled by someone in the sales department. Large companies set up advertising department whose job it is to set up the advertising budget, work with the ad agency, any handle direct mail advertising, dealer displays, and other advertising not done by the agency. Most large companies use outside advertising agencies because they offer several advantages. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 47 ADVANTAGES OF ADVERTISING. There are several advantages to the use of advertising in the firm’s promotional mix. Since the company pays for the advertising space, it can control what it wants to say when it wants to say it, and to some extent, to whom the message is sent. Advertising can also represent a costeffective method for communicating with large audiences, and cost per contract through advertising is often quite low. Advertising also can be used to create images and symbolic appeals for products and services, a capability that is very important to companies selling products and services that are difficult to differentiate. Another advantage of advertising is its value in creating and maintaining brand equity. Brand equity can be thought of as a type of intangible asset of added value or goodwill that results from the favorable image, impressions of differentiation, and/or the strength of consumer attachment to a company name, brand name, or trademark. The equity that results from a strong company or brand name is important because it allows a brand to earn greater sales volume and/or higher margins than it could without the name and also provides the company or brand with a competitive advantage. Yet another advantage of advertising is its ability to strike a responsive chord with consumers when other elements of the marketing program have not been successful. DISADVANTAGES OF ADVERTISING. Advertising has some disadvantages. The cost of producing and placing advertising can be very high. Other problems with advertising include its credibility and the ease with which it can be ignored. Advertising is often treated with skepticism by consumers, many of whom perceive it to be vary biased and are concerned by its intent to persuade. Not only are consumers skeptical about many of the advertising message they see and hear, but also it is relatively easy for them to process selectively only those ads of interest to them. Actually, with so many messages competing for our attention every day, it is out of necessity that we must ignore the vast majority of them. The high level of “clutter” is a major problem in advertising. The numerous commercials we see on television or hear on the radio, as well as the many ads that appear in most magazines and newspapers, make it very difficult for advertisers to get their messages noticed and attended to by consumers. MAJOR DECISIONS IN ADVERTISING. Company must make five important decisions when developing an advertising program. 1. SETTING OBGECTIVES. The first step in developing an advertising program is to set advertising objectives. These objectives should be based on past decisions about the target market, positioning, and marketing mix. The marketing positioning and mix strategy define the job that advertising must do in the total marketing program. An advertising objective is a specific communication task to be accomplished with a specific target audience during a specific period of time. Advertising objectives can be classified by purpose - whether their aim is to inform, persuade, or remind. Message decisions: Objectives setting: Budget decisions: Communication objectives Affordable approach Percent of sales Competitive parity Objective and task Sales objectives Message selection Message generation Message evaluation Message execution Campaign evaluation: Communication impact Sales impact Media decision: Reach, frequency, impact Major media types Specific media vehicles Media timing Figure 1 Major advertising decision Informative advertising is used heavily when introducing a new product category. In this case, the objective is to build primary demand. Thus, producers of compact disc players first informed consumers of the sound and convenience benefits of CDs. Persuasive advertising becomes more important as competition increases. Here, the company's INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 48 objective is to build selective demand. For example, when compact disc players became established and accepted, Sony began trying to persuade consumers that its brand offers the best quality for their money. Some persuasive advertising has become comparison advertising, in which a company directly or indirectly compares its brand with one or more other brands. Reminder advertising is important for mature products - it keeps consumers thinking about the product. Expensive Coca-Cola ads on television are designed to remind people about Coca-Cola, not to inform or persuade them. 2. SETTING THE ADVERTISING BUDGET. After determining its advertising objectives, the company next sets its advertising budget for each product. The role of advertising is to affect demand for a product. The company wants to spend the amount needed to achieve the sales goal. 3. CREATING THE ADVERTISING MESSAGE. A large advertising budget does not guarantee a successful advertising campaign. Two advertisers can spend the same amount on advertising, yet have very different results. Studies show that creative advertising messages can be more important to advertising success than the number of money spent. No matter how big the budget, advertising can succeed only if commercials gain attention and communicative well. Therefore, the budget must be invested in effective advertising message. Good advertising message are especially important in today’s costly and cluttered advertising environment. 4. SELECTING ADVERTISING MEDIA. The advertiser next chooses advertising media to carry the message. The major step in media selection are deciding on reach, frequency, and impact; choosing among major media types; selecting specific media vehicles; and deciding on media types. To select media, the advertiser must decide what reach and frequency are needed to achieve advertising objectives. Reach is a measure of the percentage of people in the target market who are exposed to the ad campaign during a given period time. Frequency is a measure of how many times the average person in the target market is exposed to the message. The advertiser also must decide o the desired media impact – the qualitative value of a message exposure through a given medium. The major media types are newspapers, television, direct mail, radio, magazines, and outdoor. Each medium has advantages and limitations. The best media vehicles – specific media within each general media type. Media planners also compute the cost per thousand persons reached by a vehicle; must consider the costs of producing ads for different media; must balance media cost measures against several media impact factors. The advertiser also must decide how to schedule the advertising over the cource of a year. Finally, the advertiser has to choose the pattern of the ads. Continuity means scheduling ads evenly within a given period. Pulsing means scheduling ads unevenly over a given time period. 5. ADVERTISING EVALUATION. The advertising program should evaluate both the communication effects and the sales effects of advertising regularly. Measuring the communication effect of an ad – copy testing – tells whether the ad is communicating well. Measuring the sales effect – compare past sales with past advertising expenditures. 6. CLASSIFICATIONS OF ADVERTISING. The nature and purpose of advertising differ from one industry to another and/or across situations. The target of an organization’s advertising efforts often varies, as does its role and function in the marketing program. To better understand the nature and purpose of advertising to the final buyer, it is useful to examine some classifications of the various types of advertising. So, we can subdivide advertising into two main groups: 1. Advertising to the consumer market; 2. Advertising to the business and professional markets. Advertising to the consumer market. Advertising done by a company on a nationwide basis or in most regions of the country and targeted to the ultimate consumer market is known as national advertising. The companies that sponsor these ads are generally referred to as national advertisers. Most of the ads for well-known brands that we see on prime-time television or in other major national and regional media are examples of national advertising. This form of advertising is usually very general, as it rarely includes specific prices, directions for buying the product, or special services associated with the purchase. This type of advertising makes known or reminds consumers of the brand and its features, benefits, advantages, and uses or reminds its image so consumers will be predisposed to purchasing it, wherever and whenever it is needed and convenient to do so. National advertising is the best-known and most widely discussed form of promotion, probably because INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 49 of its pervasiveness. Another prevalent type of advertising directed at the consumer market is classified as retail (local) advertising. This type of advertising is done by retailers or local merchants to encourage consumers to shop at a specific store or to use a local services such as bank, fitness club, or restaurant. Retail advertising tends to emphasize specific customer benefits such as store house, credit policies, service, store atmosphere, merchandise assortments, or other distinguishing attributes. Direct-response advertising is a method of direct marketing whereby a product is promoted through an advertisement that offers the customer the opportunity to purchase directly from the manufacturer. Another way of viewing advertising to the ultimate customer is in terms of whether the message is designed to stimulate either primary or selective demand. Primary demand advertising is designed to stimulate demand for the general product class or entire industry, whereas selective demands advertising focuses on creating demand for a particular manufacturer’s brands. Primary demand advertising is often used as a part of a promotional strategy for a new product to help it gain acceptance among customers. Selective demands advertising is getting consumers to consider using the product. Advertising to the business and professional markets. For many companies, the ultimate customer is not the mass consumer market but rather another business, industry, or profession. Business-to-business advertising is used by one business to advertiser its products or services to another. The target for business advertising is individuals who either use a product or service or influence a firm’s decision to purchase another company’s product to service. Three basic categories of business-to-business advertising are industrial, professional, and trade advertising. Advertising targeted at individuals who buy or influence the purchase of industrial goods or other services is known as industrial advertising. Industrial goods are those products that either become a physical part of another product (raw materials, component parts), are used in the manufacture of other goods (machinery, equipment), or are used to help the manufacturer conduct business (office supplies, computers, copy machines, etc.). Business services, such as insurance, financial services, and health care, are also included in this category. Industrial advertising is usually found in general business publications (such as Fortune, Business Week) or in publications targeted to the particular industry. Industrial advertising is often not designed to sell a product or service directly, as the purchase of industrial goods is often a complex process involving a number of individuals. An industrial ad helps make the company and its product or service better known by the industrial customer, assists in developing an image for the firm, and perhaps most important, opens doors for the company’s sales representatives when they call on these customers. Advertising that is targeted to professional groups – such as doctors, lawyers, dentists, engineers, or professors – to encourage them to use or specify the advertiser’s product for others’ use is known as professional advertising. Professional group are important because they constitute a market for products and services they use in their business. Also, their advice, recommendation, or specification of a product or service often influences many consumer purchase decisions. These classifications of the various types of advertising demonstrate that this promotional element is used in a variety of ways and by a number of different organizations. Advertising is a very flexible promotional tool whose role in a marketing program will vary depending on the situation facing the organization and what information needs to be communicated. Companies are also focusing attention on international markets because of the opportunities they offer for growth and profits. Advertising and promotion are important parts of the international marketing program of firms competing in the global marketplace. However, in addition to its importance, many companies are realizing the challenge and difficulties they face in developing and implementing advertising and promotion programs for international markets. Companies planning on marketing and advertising their products or services abroad face an unfamiliar marketing environment and customers with a different set of values, customs, consumption patterns and habits, as well as differing purchase motives and abilities. Not only may the language vary from country to country, but also several may be spoken within a country, such as in India or Switzerland. Just as with domestic marketing, companies engaging in international marketing must carefully analyze and consider the major environmental factors of each market in which they compete. The major environmental factors affecting international marketing include economic, demographic, cultural, and political/legal variables. Figure 2 shows some of the factors marketers must consider in each category when analyzing the environment of INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 50 each country or market. Consideration of these factors is important not only in evaluating the viability and/or potential of each country as a market but also in designing and implementing a marketing and promotional program. Economic environment: Cultural environment: Stage of economic development Economic infrastructure Standard of living Per capital income Distribution of wealth Currency stability Exchange rates Language Lifestyles Values Norms and customs Ethics and moral standards Taboos International MARKETING Demographic environment: Political environment: Size of population Number of households Household size Age distribution Occupation distribution Education levels Employment rate Income levels Government policies Laws and regulations Political stability Nationalism Attitude toward multinational companies Figure 2 Forces in the international marketing environment. The discussion of differences in the marketing environment of various countries suggests that each market is different and requires a distinct marketing and advertising program. However, in recent years, a great deal of attention has been focused on the concept of global marketing whereby a company utilizes a common marketing plan for all countries in which it operates, thus selling the product in essentially the same way everywhere in the world. Global advertising would fall under the umbrella of global marketing as a means of implementing this strategy by using the same basic advertising approach in all markets. The idea of a global marketing strategy and advertising program offers certain advantages to a company, including: · Economies of scale in production and distribution. · Lower marketing and advertising costs as a result of reductions in planning and control. · Lower advertising production costs. · Abilities to exploit good ideas on a worldwide basis and introduce products quick)y into various world markets. · A consistent international brand and/or company image. · Simplification of coordination and control of marketing and promotional programs. Advocates of global marketing and advertising contend that standardized products are possible in all countries if marketers emphasize quality, reliability, and low prices. They argue that people everywhere want to buy the same products and live the same way. The results of product standardization are lower design and production costs as well as greater marketing efficiency, which translates into lower prices for consumers. Additionally, product standardization and global marketing enable companies to roll out products faster into world markets, which is becoming increasingly important as product life cycles become shorter and competition increases. Opponents of the standardized, global approach argue that differences in culture, marketing, and economic development; consumer needs and usage patterns; media availabilities; and legal restrictions make it extremely difficult to develop an effective universal approach to marketing and advertising. International advertisers face many complexities not encountered by domestic advertisers. The basic issue concerns the degree to which global advertising should be adapted to the unique characteristics of various country markets. Some large advertisers have attempted to support their global brands with highly standardized worldwide advertising. Standardization produces many benefits - lower advertising costs, greater coordination of global advertising efforts, and a more consistent worldwide company or product image. However, standardization also has drawbacks. Most importantly, it ignores the fact that country markets differ greatly in their cultures, demographics, and economic conditions. Thus, most international advertisers think globally but act locally. They INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 51 develop global advertising strategies that bring efficiency and consistency to their worldwide advertising efforts. Then they adapt their advertising programs to make them more responsive to consumer needs and expectations within local markets. Global advertisers face several additional problems. For instance, advertising media costs and availability differ considerably from country to country. Some countries have too few media to handle all of the advertising offered to them. Other countries are peppered with so many media that an advertiser cannot gain national coverage at a reasonable cost. Media prices often are negotiated and may vary greatly. Countries also differ in the extent to which they regulate advertising practices. Many countries have extensive systems of laws restricting how much a company can spend on advertising, the media used, the nature of advertising claims, and other aspects of the advertising program. Such restrictions often require that advertisers adapt their campaigns from country to country. Thus, although advertisers may develop global strategies to guide their overall advertising efforts, specific advertising programs usually must be adapted to meet local cultures and customs, media characteristics, and advertising regulations. While globalization of advertising is viewed by many in the advertising industry as a difficult task, some progress has been made in learning what products and services are best suited to worldwide appears. Products that can take advantage of global marketing and advertising opportunities include: 1. Brands that can be adapted for a visual appeal that avoids the problem of trying to translate words into dozens of languages; 2. Brands that are promoted with image campaigns that lend themselves to themes that play up to to universal appeals such as sex or wealth; 3. High-tech products coming to the world for the first time; new technology products coming on the world at once and not steeped in the cultural heritage of the country; 4. Products with nationalistic flavor if the country has a reputation in the field; 5. Many companies and brands rely heavily on visual appeals that are easily adapted for use in global advertising campaigns. Summing up, it is important to note, that the process of globalization the sphere of marketing, especially advertising, is a reality, which represents the objective and completely inevitable phenomenon of nowadays. It can be braked by objective reasons or by means of economic policy, but it is impossible to stop or to cancel. REFERENCES: 1. BEARDEN W.O., INGRAM T.N., LAFORGE R.W.: Marketing: principals, perspectives. – Chicago: Irwin, 1995. – 631 p. 2. BELCH G.E., BELCH M.A.: Introduction to advertising and promotion. An integrated communications perspective. – Homewood: Irwin, 1993 – 836 p. 3. Cateora, PHILIP R.: International marketing. – 9th ed. – Chicago: IRWIN, 1996. – 770 p. 4. KAATZ RON: Advertising & marketing checklists: 107 PROVEN CHECKLISTS TO SAVE TIME & boost advertising & marketing effectiveness. – 2nd ed. –Lincolnwood: NTC Business Books/NTC a division of NTC Publishing Group, 1995. – 240 p. 5. KOTLER Ph.: Marketing Management. – Millenium ed., International ed. – Englewood Cliffs: Prentice-Hall International, Inc., 2000. – 718 p. 6. KOTLER Ph.: Marketing management: Analysis, planning, implementation, and control. – 8th ed. – Englewood Cliffs: Prentice-Hall International, Inc., 1994. – 801 p. 7. KOTLER Ph.: Principles of marketing. – 2nd ed. – Englewood Cliffs: Prentice-Hall, 1994. – 692 p. 8. PATTI C.H., HARTLRY S.W., KENNEDY S.L.: Business-to-bussiness advertising: a marketing management approach. – Chicago: Business/Professional Advertising Association (B/PAA), 1991. – 286 p. 9. SCHULTZ D., TANNENBAUM S., ALLISON A.: Essentials of advertising strategy. – 3rd ed. – Lincolnwood: NTC Business Books, a division of NTC Publishing Group, 1996.-155p. 10. SHIMP T.A.: Advertising, promotion, and supplemental aspects of integrated marketing communication. – 4th ed. – Fort Worth: The Dryden Press – Harcourt Brace College Publishers, 1997.- 589 p. 11. RUSSELL T., LANE R.: Kleppner’s Advertising procedure. – 2nd ed., - Englewood Cliffs: Prentice hall, New Jersey, 1990. - 718 p. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 52 CUSTOMER BEHAVIOUR Radoslav Jankal University of Žilina, Faculty of Management Science and Informatics, Department of Management Theories Moyzesova 20, 010 26 Žilina, Slovak republic, tel.:+421 41 513 4458, fax: +421 41 5652 775 email: [email protected] Abstract: The sense of marketing is to meet needs and wants of customers. To understand the customer behaviour is not easy. The consumers often say something else as they finally do. Their thinking and ideas can change very quickly by the new stimulus. When the firms want to be successful and profitable, they need to know very well, what and why their customers buy. Keywords: behaviour, customer, trends, technology 1 WHAT IS CUSTOMER BEHAVIOUR Mental and physical activities undertaken by household and business customers that result in decisions and actions to pay for, purchase and use products and services. This is one of the views on the customer behaviour. Customer behaviour reflects customers´ decisions with respect to: · the acquisition, consumption, and disposition · of goods and services, time, and ideas · by (human) decision making [over time]. For the understanding of customer behaviour, is useful to know the answers on the questions in the figure 1. [4] Why ar e o ur pr o duc ts bo ug ht Whe n ar e c o mpe titiv e c o nsume d H o w o fte n pr o duc ts H o w muc h ar e substitute s By who m Whe r e e v aluate d How Figure 1 Questions for better understanding of customer behaviour The start point of customer behaviour study is the model, which represent, how the customer reacts on different stimuli (see figure 2). Marketing stimuli and environs effect enter in the customer subconscious. Typical customer characteristics and his decision process lead in specific buyer response. The role of marketers is to recognize, what is going on in customer mind from the moment, when the specific stimuli from external environment get in his subconscious, to the moment, when he makes a purchase decision. [3] M ar ke ting stimuli Product Price Place Promotion O the r stimuli Economic Technological Political Cultural Custo me r c har ac te r istic s Cultural Social Personal Psychological Custo me r de c isio n pr o c e ss Buye r ’s r e spo nse s Problem identification Collection of information Evaluation of information Decision Behaviour after buying Product choice Brand choice Dealer choice Purchase timing Purchase amount Figure 2 Model of customer behaviour INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Ku novice, Czech Repu blic, Ja nu a ry 2 9 -3 0 , 2 0 0 4 53 2 TRENDS IN CUSTOMER BEHAVIOUR Predicting future customer behaviour requires predicting the factors that influence customer behaviour. Almost any of these diverse factors is capable of changing and, in turn, changing customer behaviour. Here is on the three factors that are expected to cause the most significant change in future customer behaviour: changes in demographics, advances in technology, and changes in public policy (see figure 3). Changes in other customer factors are likely as well, but they will be less discernible and perhaps less dramatic. For this reason, we explore these, projecting the trajectory of developments already visible. De mo g r aphic s · a ging popu la tion · women in work force · single-person hou seholds · declining middle cla ss · ethnic diversity · geogra phic redistribu tion Te c hno lo g y · control over informa tion · sma rt produ cts · a ccess to produ cts · ma ss cu stomisa tion Public Po lic y · pra gma tism over ideology · rights of pa ssive consu mers · regiona l economic integra tion Figure 3 Changes in the determinants of customer behaviour Anticipating trends in customer behaviour can give companies a key strategic advantage. Companies that will survive and thrive in the marketplace of tomorrow are those in which managers spend less time worrying about how to position their firm among current competitors and more time trying to envisage a new competitive space – i.e. a space defined by tomorrow’s customer needs and wants. new ma rk ets from cha nnelling a la tent need first to ma rk et positive pu blic opinion marketing executive Figure 4 Strategic advantages of anticipating customer trends Foreseeing the coming trends in customers’ needs and wants offers companies several advantages (see figure 4). First, if you see a market need first, you can be the first to start working to meet that need. Consequently, it reduces the fulfilment time, making the economic payback a lot quicker. For example, Sony anticipated the Walkman phenomenon, was the first in the marketplace with its product, and succeeded. Second, by sighting a trend, the industry can create a market by channelling a latent need. One example is the mobile phone industry. Being able to communicate without being grounded in one location (i.e. wireless communication) is certainly on many people’s wish list, but not everyone can afford the high initial equipment costs. The industry responded by offering a free phone with a service contract, and the market for mobile phones skyrocketed. Many people consider mobile phones an essential piece of equipment and the expenditure on mobile phones is considered in the same category as expenditures made on fast food, poker machines and electrical goods. As a nation, Australia has taken to mobile phones faster than anyone else. Third, anticipating trends in customer needs and wants and responding to them creates positive public opinion for the company and the industry, portraying them as responsive. For example, Australian airlines and INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Ku novice, Czech Repu blic, Ja nu a ry 2 9 -3 0 , 2 0 0 4 54 terminals are smoke free; many restaurants have moved to a no-smoking policy even before such a policy has become law; several local and foreign products are now produced and marketed as being environmentally friendly products, such as organically grown foods (hydroponic tomatoes at the supermarkets), recyclable paper products (toilet paper, writing paper) and cosmetics developed without being tested on animals. [4] 2.1 Demographic trends Changing demographics of the population serve as a good indicator of the future marketplace. As we have illustrated in preceding chapters, such demographic characteristics as age, income, race and geographic location of a customer intimately influence his or her customer behaviour. Consequently, when the demographic make-up of a population changes, the marketplace changes in terms of its needs and wants. Studying projections or trends in the demographic make-up of a population can therefore help marketers anticipate the needs and wants of their customers. Six demographic trends have already begun to transform the marketplace: · the ageing of the population · the rise in the number of working women · an increase in single-person households · the decline of the middle class · the increase in ethnic diversity · geographic redistribution. 2.2 Technological trends A second force shaping future customer values is technology. Advances in technology have already given customers increased access to information, newer generations of products, automation of transaction processes to provide customers with greater flexibility and control, and access to some customised products. Future developments will further change products and the very nature of customer behaviour. In particular, technological advances are expected to include products that give customers more control over information and information access-smart products, automation of processes that liberate customers from operations-driven processes and products and services that enable customised lifestyles. Customer responses to new technology Technological developments will in turn stimulate changes in customer behaviour. Customers will increasingly take on the role of co-producers. In addition, they will engage in disintermediation, outsourcing and automation of consumption (figure 5). [4] 1 . Custo me r s as c o -pr o duc e r s 2 . Disinte r me diatio n 3 . Custo me r o utso ur c ing 4 . Auto matio n o f c o nsumptio n Figure 5 Trends in customer behaviour as a result of improved technology 2.3 Trends in public policy Several projected trends in public policy are relevant to a study of customer behaviour. First, economic pragmatism is prevailing over ideology. In addition, many governments are concerned with protection of passive consumers. Finally, governments on an international level are pursuing regional economic integration. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Ku novice, Czech Repu blic, Ja nu a ry 2 9 -3 0 , 2 0 0 4 55 3. CONCLUSION Markets have to be understood before marketing plans can be developed. The consumer market buys goods and services for personal consumption. The customer’s behaviour is influenced by four major factors: cultural (culture, subculture, and social class); social (reference groups, family, and roles and statuses); personal (age and life cycle stage, occupation, economic circumstances, life-style, and personality and self-concept); and psychological (motivation, perception, learning, and beliefs and attitudes). All of these provide clues as to how to reach and serve the buyer more effectively. [2] Before planning its marketing, a company needs to identify its target consumers and the type of decision process they go through. When buying something, a buyer goes through a decision process consisting of problem recognition, information search, evaluation of alternatives, purchase decision, and postpurchase behavior. The marketer's job is to understand the various participants in the buying process and the major influences on buying behavior. This understanding allows the marketer to develop a significant and effective marketing program for the target market. Here is very necessary and useful good and effective communication. The communication is one of the most important points of the management of each company. It is the irreplaceable tool of the cooperation and relation of employees and management in their common effort for the achieving of the target vision, mission and specified aims. [1] With regard to new products, consumers respond at different rates, depending on the consumer's characteristics and the product's characteristics. Manufacturers try to bring their new products to the attention of potential early adopters, particularly those who are opinion leaders. REFERENCES: [1] BLAŠKOVÁ, M.: Riadenie a rozvoj ľudského potenciálu: uplatňovanie motivačného akcentu v procesoch práce s ľuďmi. - 1. vyd. - Žilina : Žilinská univerzita, 2003. ISBN 80-8070-034-6 [2] KOTLER, P.: Marketing essentials. Prentice Hall. 1984. ISBN 0-13-557232-0 [3] KOTLER, P.: Marketing management. Grada. 2001. ISBN 80-247-0016-6 [4] WIDING, R. et al.: Customer Behaviour: Consumer Behaviour and Beyond. Thomson Learning. 2002. ISBN: 0-17-010786-8 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Ku novice, Czech Repu blic, Ja nu a ry 2 9 -3 0 , 2 0 0 4 56 CONSUMER BEHAVIOUR IN GLOBAL MARKET Elvira Seifoullaeva The Russian Plekhanov Academy of Economics – the Faculty of International Economic Relations: 113054, Stremanniy per. 36, Moscow, Russia, (095) 237-95-52; [email protected], [email protected] Abstract: in modern world success of a company directly depends on the level of its orientation on the consumers and knowledge of major aspects of consumer behaviour.Consumer behaviour is influenced by many factors, such as culture and mentality, personal attitudes and social backgroud. Thus, market segmentation is nessesary for a company. Analysing consumer behaviour helps companies create more effective marketing strategies and stay leaders on the market, constantly widening the supply area. Keywords:consumer behaviour, marketing strategy, methods of research and evaluation, influencing factors, consumer preferences, marketing decisions, global market. It is well known that nowadays the success of a company mainly depends on how efficient this company is in satisfying consumers’ needs and desires. And the only way to be closer to the consumers is to learn as much as possible about their preferences and their behaviour. In modern world only a quality-focused and customer-driven strategy creates market leaders and inevitably results in profitability and success. Understanding consumer behaviour is the key to increasing the efficiency of a company, thus, creating the image of a consumer oriented producer and increasing volume of sales accordingly. Consumer behavior is often defined as “the study of individuals, groups, or organizations and the processes they use to select, secure, use, and dispose of products, services, experiences or ideas to satisfy needs, and the impacts that these processes have on the consumer and society”. Thus, the definition brings up some very important points: 1. Behavior occurs either for the individual, or in the context of a group (i.e., friends often influence personal decisions concerning buying sertain goods or not buying them) or an organization (management decisions on products the company should use); 2. Consumer behavior involves the use and disposal of products as well as the study of how they are purchased. Product use is often of great interest to the marketer, because it may influence how a product is best positioned or how the increase of consumption can be encouraged; 3. Consumer behavior involves services and ideas as well as tangible products; 4. The impact of consumer behavior on society is also relevant. In global marketing there are four major applications of consumer behavior: 1. For marketing strategy—i.e., for creating and adopting more effective marketing campaigns; 2. For public policy; 3. For Social marketing, which involves getting ideas across to consumers rather than selling something; 4. For helping managers evaluate the supply through the consumers’ eyes. In international marketing as well as in marketing in general there are several crucial units that should be analyzed. Consumer behaviour is one of them. It can be analysed through two main research methods: secondary and primary research. Secondary research presupposes the use of data, already collected in reports, statistics or some other sources. However, in some cases, the required information is too specific and then a nessesity to use original – primary research arises. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 57 A company has a variety of possibilities in using different instruments of primary research, such as: · Surveys – one of the most useful ways for getting much specific information. Surveys may contain either open-ended questions, where a respondent is not limited to the options listed and therefore is not influenced by them, or closed-ended questions, where the respondent is asked to select answers from a brief list.; · Experiments are used when only one explanation for a particular observation should be ruled out. However, this method has a serious drawback - the consumer is removed from his or her natural surroundings; · Focus groups are an important marketing category, which involve getting a group of 6-12 consumers together to discuss product usage. Focus groups are especially useful if there are no specific questions to ask yet, because they help find out what consumers’ concerns might be; · Still, there are also drawbacks such as high costs and the fact that generalization toward the entire population is difficult for such small sample sizes. Moreover, the fact that focus groups involve social interaction means that, participants might say not what they really think, but what in their opinion will make themselves look better, which can be defined as the social desirability bias; · Projective techniques can be used when a consumer feel embarrassed to admit to certain opinions, feelings, or preferences. For example, many older executives may not feel comfortable admitting to being intimidated by computers. It’s obvious, that in such cases people will tend to respond more openly not for themselves but for "someone else" ; · Observation of consumers is often a powerful tool. Looking at how consumers select products may yield insights into how they make decisions and what they are actually looking for. In consumer behaviour analysis segmentation also plays an important role. Understanding the consumer allows a company segment markets more meaningfully. Segmentation basically involves dividing consumers into groups in such a way, that members of a group are on the one hand as similar as possible to each other, but on the other hand they differ as much as possible from members of other segments. This enables companies to deal with each segment variously by: 1. Providing different products according to customers’ preferences; 2. Offering different prices; 3. Distributing the products where they are likely to be bought by the targeted segment. To icrease the effectiveness of a segment structure: · Each segment must have an identity—i.e., it must contain members that can be described in some way (like price sensitive) and that behave differently from another segment; · Each segment must engage in systematic behaviours (i.e., a price sensitive segment should consistently prefer the low price item rather than randomly switching between high and low priced brands); · Each segment must offer marketing mix efficiency potential—i.e., it must be profitable to serve. For example, a large segment may be profitable even though the competition it attracts tends to keep prices down. A smaller segment may be profitable even if it is price insensitive or can be targeted efficiently. It’s important to mention that consumer behaviour can be strongly influenced by culture as an external factor of the impact. Culture is said to represent influences that are imposed on the consumer by other individuals. It can be defined as "unit of knowledge, belief, art, morals, custom, and any other capabilities and habits acquired by person as a member of society." Culture has several important characteristics: 1. It is comprehensive. This means that all parts must fit together in some logical fashion; 2. Culture is learned rather than being something we are born with; 3. It is manifested within boundaries of acceptable behavior; 4. Conscious awareness of cultural standards is limited; 5. Cultures fall somewhere on a continuum between static and dynamic depending on how quickly changes are accepted. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 58 Moreover, it is worth mentioning, that some cultures tend to adopt new products more quickly than others, which is said to be based on following reasons: · Modernity - the extent to which the culture is receptive to new things. F.e. in some countries, such as Britain and Saudi Arabia, tradition is valued the most - thus, new products often don’t fare too well, whereas other regions long for progress; · Homophily – meaning that the more similar to each other members of one culture are, the more likely an innovation is to spread withing this culture and therefore, some region. The cause lies in the point of people being more likely to imitate similar than different models of consumer behaviour. · Nowadays the two most rapidly adopting countries in the World are the USA and Japan. · Physical distance – presupposes that the greater the distance between people is, the less likely an innovation is to spread. · Opinion leadership- an issue showing that the more opinion leaders are valued and respected, the more likely an innovation is to spread. The style of opinion leaders moderates this influence. However, in less innovative countries opinion leaders tend to be more conservative, i.e., to reflect the local norms of resistance. What is also vital when talking about factors, influencing consumer behaviour, is family decision making issue. Individual members of families often serve different roles in decisions that ultimately draw on shared family resources. Some individuals are information gatherers/holders, who seek out information about products of relevance. These individuals often have a great deal of power because they may selectively pass on information that favors their chosen alternatives. Influencers do not ultimately have the power to decide between alternatives, but they may make their wishes known by asking for specific products or causing embarrassing situations if their demands are not met. On the contrary, the decision makers have the power to determine issues like: which product or brand to buy; where to buy it; when to buy and, finally, whether to buy it all or not. However, it’s important to differentiate the roles of the decision maker and that of the purchaser. From the point of view of the marketer, this introduces some problems since the purchaser can be targeted by point-ofpurchase (POP) marketing efforts that cannot be aimed at the decision maker. Another point worth mentioning is group influence. Humans are inherently social beings, therefore individuals influence each other greatly. A useful framework of analysis of group influence on the individual is so called reference group, which obtain various degrees of influence. Primary reference groups come with much greater impact on single consumer than secondary reference groups, which tend to have less influence, usully said to be limited to consumption during some certain time period. Among factors that have strong impact on consumer behaviour personality is the most controversial one. Traditional research in marketing has not been particularly successful in finding a link between personality and consumer behavior. Still, personal preferences of a consumer inevitably play major role when the decision about the purchase is made. Unfortunately, it is very difficult to determine which factors actually formed that very specific attitude. What about emotional impact, on the one hand it is used by producers to attract attention to a brand or a company itself, and on the other hand is vital for the information processing. Appealing to emotions is almost always successful, unless only the expected reaction of consumers varies from the one in reality. Nevertheless, orientation on consumers doesn’t mean a company should be completely dependable on their preferences. Sometimes it is possible to change consumer behavior. It is the question of efficient marketing strategy and its implementaion, to make consumers switch from well known brands to new ones. One way to reach such result is temporary price discounts and coupons. However, such approach makes consumers buy a product on deal, which brings up a threat they may justify the purchase based on that deal (i.e., the low price) and later switch to other brands on the same basis. A better way is to make the product itself more convinient to obtain and prove its value and quality to a larger number of consumers. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 59 Still, the value of a product for a consumer also depends on many factors. The most important issue the consumer takes into consideration when making a decision over a particular purchase is satisfaction of his or her major needs. The needs of consumer were best structured in a hierarchy by Abraham Maslow. He suggested the intuitively appealing notion that humans must satisfy the most basic objectives before they can move onto "higher level" ones. Thus, an individual must satisfy physiological needs (such as food and liquid) before he or she will be able to expend energy on less fundamental objectives such as safety. The Maslow hierarchy of needs is as follows: It is a useful model for understanding different needs of consumers both in national and in global markets. Self – However, they shouldn’t be taken too literally, since people actualization may occasionally "swing" between these needs. Thus, the Personal needs: art of marketing presupposes flexibility and adequate status, respect, prestige reaction to changes in consumer behaviour. Social needs: Therefore, a company should be perfectly friendship, belongigng, love informed of its success or failure in adopting the marketing Safety needs: strategy to a changing consumer demands. In this concept financial, freedom from harm consumer satisfaction is an impportant issue. Calculating Physiological needs: financial value of a satisfied consumer is a serious problem food, water, oxygen, sex for any company. The most common approach to evaluating this is by focusing on consumer retention. On the half of consumer, behind the very decision over a purchase lies an important process of making a choice. In this purchase decision process a consumer passes through several stages, which are: Problem recognition: Perceiving a need Information search: Seeking value Evaluation of alternatives: Assesing value Purchase decision: Buying value Postpurchase behaviour: Value in consumption or use Postpurchase behaviour refers to the feelings consumer has after comparing the obtained products with his or her expectations. In case of dissatisfaction marketers will have to analyse whether the product was deficient or consumer’s expectations were too high. Deficiency of the product may require change of design or even quality of the product, and in case of high consumer expectations the company might need to change it’s advertising strategy or reorient to another target group. Thus, consumer behaviour is an integral part of marketing process and together with financial matters and long-term obligations form the basis for strategic marketing decisions. Besides marketing mix influences also pshychological, sociocultural and sometimes situational factors have impact on consumer behaviour, either motivationg or discoraging consumers to obtain some particular products. The knowledge of consumer behaviour and adequate response to its changes are two major factors for a company’s success in global market. REFERENCES: [1] BERKOWITZ, KERIN, HARTLEY, RUDELIUS: Marketing. Fifth edition: Irwin McGraw-Hill, 1997, p.p. 152-157; [2] www.ACNielsen/customer.com [3] www.consumerpsychologist.com [4] www.sciencecenter.com/services/customer.html INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 60 SALES PROMOTION Helen Berchik Plehanov Russian Academy of Economics, the faculty of International Economic Relations, Moscow, Stremiannii alley., 36, tel. (095) 236-40-94, (095) 237-95-52, [email protected] Abstract: Sales promotion includes a wide variety of promotion tools designed to stimulate earlier or stronger market response. Sales promotion can be broken into two major categories: consumeroriented promotions and trade-oriented promotions. Sales promotion plays an important role in the total promotion mix. To use it well, the marketer must define the sales-promotion objectives, select the best tools, and evaluate the results. Keywords: sales promotion, promotion tools, consumer-oriented promotions and trade-oriented promotions. Sales promotion has been defined as “a direct inducement that offers an extra value or incentive for the product to the sales force, distributors or the ultimate consumer with the primary objective of creating an immediate sale”. It consists of short-term incentives to encourage purchase or sales of a product or service. Whereas advertising offers reasons to buy a product or service, sales promotion offers reasons to buy now. Examples are found everywhere: · A family buys a camcorder and gets a free traveling case or buys a car and gets a check for a $500 rebate; · An appliance retailer is given a 10 percent manufacturer discount on January’s orders if the retailer advertises the product in the local newspaper. Sales promotion includes a wide variety of promotion tools designed to stimulate earlier or stronger market response. Sales promotion can be broken into two major categories: consumer-oriented promotions and trade-oriented promotions. Consumer-oriented sales promotion includes sampling, couponing, rebates, pricesoffs, premiums, contests, bonus packs, sweepstakes and event sponsorship. These promotions are directed at the consumers who purchase goods and services and are designed to provide them with an inducement to purchase the marketer’s brand. Trade-oriented sales promotion includes activities such as promotional allowances, free goods, merchandise allowances, cooperative advertising, push money, dealer sales contests; and salesforce promotion – bonuses, contests, sales rallies designed to motivate distributors and retailers to carry a product and make an extra effort to promote or “push” it to their customers. RAPID GROWTH OF SALES PROMOTION Sales promotion tools are used by most organizations, including manufacturers, distributors, retailers, trade associations, and nonprofit institutions. Estimates of annual sales-promotion spending run as high as $125 billion, and this spending has increased rapidly in recent years. A few decades ago, the ratio of advertising to sales promotion spending was about 60/40. Today, in many consumer packaged – goods companies, the picture is reversed, with sales promotion accounting for 75 percent or more of all marketing expenditures. Salespromotion expenditures have been increasing 12 percent annually, compared to advertising’s increase of only 7.6 percent. Several factors have contributed to the rapid growth of sales promotion, particularly in consumer markets. First, inside the company, promotion now is accepted more by top management as an effective sales tool and more product managers are qualified to use sales promotion tools. Furthermore, product managers face greater pressures to increase their current sales. Second, externally, the company faces more competition, and competing brands are less differentiated. Competitors are using more and more promotions, and consumers have become more deal oriented. Third, advertising efficiency has declined because of rising costs media clutter, and legal restraints. Finally, retailers are demanding more deals from manufacturers. The growing use of sales promotion has resulted in promotion clutter, similar to advertising clutter. Consumers are increasingly tuning out promotions, weakening their ability to trigger immediate purchase. In INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 61 fact, the extend to which U.S. consumers have come to take promotions for granted was illustrated dramatically by the reactions of Eastern European consumers when Procter & Gamble recently gave out free samples of a newly introduced shampoo. To P&G, the sampling campaign was just business as usual. To consumers in Poland, however, it was little short of a miracle: With nothing expected in return, Warsaw shoppers were being handed free samples of Vidal Sasoon Wash & Go shampoo. Just for the privilege of trying the new product no standing in the line for a product that may not even be on the shelf. Some were so taken aback that they were moved to tears. Although no sales promotion is likely to create such excitement among promotion-prone consumers in the United States and other Western countries, manufacturers now are searching for ways to rise above the clutter, such as offering large coupon values or creating more dramatic point-of-purchase displays. PURPOSE OF SALES PROMOTION Sales-promotion tools vary in their specific objectives. For example, a free sample stimulates consumer trial; a free management advisory service cements a long-term relationship with a retailer. Sellers use sales promotions to attract new triers to reward loyal customers, and to increase the repurchase rates of occasional users. There are three types of new triers – nonusers of the product category, loyal users of another brand, and users who frequently switch brands. Sales promotions often attract the last group – brand switchers –because nonusers and users of other brands do not always notice or act on a promotion. Brand switchers most are looking for low price or good value. Sales promotions are unlikely to turn them into loyal brand users. Thus, sales promotions used in markets where brands are very similar usually produce high short-run sales response but little permanent market-share gain. In markets where brands differ greatly, however, sales promotions can alter market shares more permanently. Many sellers think of sales promotion as a tool for breaking down brand loyalty and advertising as a tool for building up brand loyalty. Thus, an important issue for marketing managers is how to divide the budget between sales promotion and advertising. Ten years ago, marketing managers typically would fist decide how much they needed to spend on advertising and then put the rest into sales promotion. Today, more and more marketing managers first decide how much they need to spend on trade promotion, then decide what they will spend on consumer promotion, and then budget whatever is left over for advertising. There is a danger in letting advertising take a back seat to sales promotion, however. Reduced advertising spending can result in lost consumer brand loyalty. One recent study of loyalty toward 45 major packaged-goods brands showed that when share of advertising drops, so does brand loyalty. Since 1975, brand loyalty for brands with increased advertising spending fell 5 percent. However, for brands with decreased ad spending, brand loyalty dropped 18 percent. When a company price-promotes a brand too much of the time, consumers begin to think of it as a cheap brand. Soon, many consumers will buy the brand only when it is on special. No one knows when this will happen, but the risk increases greatly if a company puts a well-known, leading brand on promotion more than 30 percent of the time. Marketers rarely use sales promotion for dominant brands because the promotions would do little more than subsidize current users. Most analysts believe that sales-promotion activities do not build long-term consumer preference and loyalty, as does advertising. Instead, promotion usually produces only short-term sales that cannot be maintained. Small-share competitors find it advantageous to use sales promotion because they cannot afford to match the large advertising budgets of the market leaders. Nor can they obtain shelf space without offering trade allowances or stimulate consumer trial without offering consumer incentives. Thus, price competition is often used for small brands seeking to enlarge their shares, but it is usually less effective for a market leader whose growth lies in expanding the entire product category. The upshot is that many consumer packaged-goods companies feel that they are forced to use more sales promotion than they would like. Recently, Kellogg, Kraft, Procter & Gamble, and several other market leaders have announced that they will put growing emphasis on pull promotion and increase their advertising budgets. They blame the heavy use of sales promotion for decreased brand loyalty, increased consumer price sensitivity, a focus on short-run marketing planning, and erosion of brand-quality image. Some marketers dispute this criticism, however. They argue that the heavy use of sales promotion is a symptom of these problems, not a cause. These marketers assert that sales promotion provides many important benefits to manufacturers as well as to consumers. Sales promotions encourage consumers to try new products INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 62 instead of always staying with their current ones. They lead to move varied retail formats, such as the every-daylow-price store or the promotional-pricing store, which gives consumers more choice. Finally, sales promotions lead to greater consumer awareness of prices, and consumers themselves enjoy the satisfaction of felling like smart shoppers when they take advantage of price specials. Sales promotions usually are used together with advertising or personal selling. Consumer promotions usually must be advertised and can add excitement and pulling power to ads. Trade and salesforce promotions support the firm’s personal selling process. In using sales promotion, a company must set objectives, select the right tools, develop the best program, pretest and implement it, and evaluate the results. SETTING SALES-PROMOTION OBJECTIVES Sales-promotion objectives vary widely. Sellers may use consumer promotions to increase short-term sales or to help build long-term market share. The objective may be to entice consumers to try a new product, lure consumers away from competitors’ products, get consumers to “load up” on a mature product, or hold and reward loyal customers. Objectives for trade promotions include getting retailers to carry new items and more inventory, getting them to advertise the product and give it more shelf space, and getting them to buy ahead. For the salesforce, objectives include getting more salesforce support for current or new products or getting salespeople to sign up new accounts. In general, sales promotions should be consumer franchise building – they should promote the product’s positioning and include a selling message along with deal. Ideally, the objective is to build long-run consumer demand rather than to prompt temporary brand switching. If properly designed, every sales-promotion tool has consumer franchise building potential. SELECTING SALES-PROMOTION TOOLS Many tools can be used to accomplish sales-promotion objectives. The promotion planner should consider the type of market, the sales-promotion objectives, the competition, and the costs and effectiveness of each tool. Consumer-Promotion Tools The main consumer-promotion tools include samples, coupons, cash refunds, price packs, premiums, advertising specialties, patronage rewards, point-of-purchase displays and demonstrations, and contests, sweepstakes and games. Trade-Promotion Tools More sales-promotion dollars are directed to retailers and wholesalers than to consumers. Trade promotion can persuade retailers or wholesalers to carry a brand, give it shelf space, promote it in advertising, and push it to consumers. Shelf space is so scarce these days that manufacturers often have to offer price-offs, allowances, buy-back guarantees, or free goods to retailers and wholesalers to get on the shelf and, once there, to stay on it. Manufacturers use several trade-promotion tools. Many of the tools used for consumer promotions – contests, premiums, displays-also can be used as trade promotions. Or the manufacturer may offer a straight discount off the list price on each case purchased during a stated period of time (also called a price-off, offinvoice, or off-list). The offer encourages dealers to buy in quantity or to carry a new item. Dealers can use the discount for immediate profit, for advertising, or for price reductions to their customers. Manufacturers also may offer an allowance (usually so much off per case) in return for the retailer’s agreement to feature the manufacturer’s products in some way. An advertising allowance compensates retailers for advertising the product. A display allowance compensates them for using special displays. Manufacturers may offer free goods, which are extra cases of merchandise, to middlemen who buy a certain quantity or who feature a certain flavor or size. They may offer push money – cash or gifts to dealers or their salesforce to “push” the manufacturer’s goods. Manufacturers may also give retailers free specialty advertising items that carry the company’s name, such as pens, pencils, calendars, paperweights, matchbooks, memo pads, ashtrays, and yardsticks. Business-Promotion Tools Companies spend billions of dollars each year on promotion to industrial customers. These business promotions are used to generate business leads, stimulate purchases, reward customers, and motivate salespeople. Business promotion includes many of the same tools used for consumer or trade promotions. The most important are conventions and trade shows, and sales contests. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 63 DEVELOPING THE SALES-PROMOTION PROGRAM The marketer must make some other decisions in order to define the full sales-promotion program. First, the marketer must decide on the size of the incentive. A certain minimum incentive is necessary if the promotion is to succeed; a larger incentive will produce more sales response. Some of the large firms which sell consumer packaged goods have a sales-promotion manager who studies past promotions and recommends incentive levels to brand managers. The marketer also must set conditions for participation. Incentives might be offered to everyone or only to select groups. For example, a premium might be offered only to those who turn in box tops. Sweepstakes might not be offered in certain states, to families of company personnel, or to people under a certain age. The marketer then must decide ho to promote and distribute the promotion itself. The length of the promotion is also important. If the sales-promotion period is too short, many prospects will miss it. If the promotion runs too long, the deal will lose some of its “act now” force. Brand managers need to set calendar dates for the promotions. The dates will be used by production, sales, and distribution. Some unplanned promotions also may be needed, requiring cooperation on short notice. Finally, the marketer must determine the sales-promotion budget, which can be developed in one of two ways. The marketer may choose the promotions and estimate their total cost. However, the more common way is to use a percentage of the total budget for sales promotion. One study found three major problems in the way companies budget for sales promotion. First, they do not consider cost effectiveness. Second, instead of spending to achieve objectives, they simply extend the previous year’s spending, take a percentage of expected sales, or use the “affordable approach”. Finally, advertising and sales-promotion budgets are too often prepared separately. PRETESTING AND IMPLEMENTING Whenever possible, sales-promotion tools should be pretested to find out if they are appropriate and of the right incentive size. Companies should prepare implementation plans for each promotion, covering lead time and sell-off time. Lead time is the time necessary to prepare the program before launching it. Sell-off time begins with the launch and ends when the promotion ends. EVALUATING THE RESULTS Evaluating is also very important. Yet many countries fail to evaluate their sales-promotion programs, and others evaluate them only superficially. Manufacturers can use one of many evaluation methods. The most common method is to compare sales before, and after promotion. Consumer research also would show the kinds of people who responded to the promotion and what they did after it ended. Surveys can provide information on how many consumers recall the promotion, what they thought of it, how many took advantage of it, and haw it affected their buying. Sales promotions also can be evaluated through experiments that vary factors such as incentive value, length, and distribution method. Clearly, sales promotion plays an important role in the total promotion mix. To use it well, the marketer must define the sales-promotion objectives, select the best tools, and evaluate the results. REFERENCES: [1] BEARDEN W.O., INGRAM T.N., LAFORGE R.W.: Marketing: principals, perspectives. – Chicago: Irwin, 1995. – 631 p. [2] KOTLER Ph.: Marketing Management. – Millenium ed., International ed. – Englewood Cliffs: Prentice-Hall International, Inc., 2000. – 718 p. [3] KOTLER Ph.: Marketing management: Analysis, planning, implementation, and control. – 8th ed. – Englewood Cliffs: Prentice-Hall International, Inc., 1994. – 801 p. [4] KOTLER Ph.: Principles of marketing. – 2nd ed. – Englewood Cliffs: Prentice-Hall, 1994. – 692 p. [5] SHIMP T.A.: Advertising, promotion, and supplemental aspects of integrated marketing communication. – 4th ed. – Fort Worth: The Dryden Press – Harcourt Brace College Publishers, 1997.- 589 p. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 64 STRATEGIES OF AN INTERNATIONAL MARKET ENTRY Ponomareva Maria Alexandrovna Plehanov Russian Academy of Economics: Stremiannii alley, 36, 8 (095) 2379552, [email protected] Abstract:When a company plans to enter a new market, it should first of all decide which country to enter. There are several major criteria for the choice of country to enter which a company should think over. They are the following: market attractiveness (GNP per head, forecast demand, etc.); competitive advantage (prior experience in similar markets, language and cultural understanding); risk (political stability, possibility of government intervention and similar external influences, business risk, currency risk etc.). Once a company decides to enter a market of a particular country, it has to determine the best model of entry. It can choose from indirect exporting, direct exporting, licensing, joint ventures, and direct investment. Each succeeding strategy involves more commitment, risk, control, and profit potential. And the amounts of these factors increase from indirect exporting to direct investments. Keywords: Indirect Export, Domestic-Based Export Merchant, Domestic-Based Export Agent, Cooperative Organization, Export-Management Company, Direct Export, Domestic-Based Export Department or Division, Overseas Sales Branch or Subsidiary, Traveling Export sales Representative, Foreign-Based Distributor or Agent, Licensing, Management contract, Contract manufacturing, Joint Venture, Direct Investment. When a company plans to enter a new market, it should first of all decide which country to enter. There are several major criteria for the choice of country to enter which a company should think over. They are the following: market attractiveness (GNP per head, forecast demand, etc.); competitive advantage (prior experience in similar markets, language and cultural understanding); risk (political stability, possibility of government intervention and similar external influences, business risk, currency risk etc.). Furthermore, some activities should be carried out. I mean: 1) Strategic review/SWOT analysis; 2) Market and competitor research (marketing environment differences, market size, segmentation, cost factors, government support, etc.), 3) Establish a network of contacts; 4) Evaluation of entry strategies cost, investment, control, risk and return; 5) Evaluation of SLEPT factors: (social; legal; economical (Level and trend in per capita income, Balance of payments, Inflation situation, Exchange rate situation, Convertibles of currency),political and technological. 6) Rethink the marketing mix and conformity to standards in context of country chosen. 7) Choose of media. Press may not be appropriate in countries where levels of literacy are low. TV ownership may not be widespread. Outdoor tends to rely on visuals and it is therefore a good international medium. Cinema is experienced in different ways. The quality of films varies considerably. Radio is mainly a support medium across the world. Commercial stations may not be available. Once a company decides to enter a market of a particular country, it has to determine the best model of entry. It can choose from indirect exporting, direct exporting, licensing, joint ventures, and direct investment. It is up to a company to decide which strategy to choose to enter a market. Each succeeding strategy involves more commitment, risk, control, and profit potential. And the amounts of these factors increase from indirect exporting to direct investments. The normal way to get involved in a foreign market is through export. There are two types of exporting: occasional exporting and active exporting. When we talk about occasional exporting we mean a passive level of involvement where a company exports from time to time on its own initiative or in response to unsolicited orders from abroad. On the contrary, active exporting takes place when the company makes a commitment to expand exports to a particular market. In either case, the company produces all of its goods in the home country and it might or might not adapt them to the foreign market. Exporting involves the least change in the company’s INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 65 product lines, organization, investments, or mission. Companies typically start with indirect exporting, in other words they work through independent middlemen. Four types of middlemen are available to the company: 1) Domestic-Based Export Merchant – this middleman buys the manufacturer’s products and sells them abroad on its own account. 2) Domestic-Based Export Agent – this agent seeks and negotiates foreign purchases and is paid a commission. Such agents can be included in the group and in that case they’re called trading companies. 3) Cooperative organization – organization that carries on exporting activities on behalf of several producers and is partly under their administrative control. This type is often used by producers of primary products such as fruits, nuts, and so on. 4) Export-Management Company – this middleman agrees to manage a company’s export activities for a fee. Indirect exporting has two main advantages. Firstly, it involves less investment. The firm does not have to develop an export department, an overseas sales force, or a set of foreign contracts, because it uses intermediaries such as export houses, specialist export management firms, complementary exporting (i.e. using other companies’ products to pull your own into an overseas market) etc. Secondly, it involves less risk. International-marketing middlemen bring know-how and services to the relationship, and the seller will normally make fewer mistakes. However when he company wants to get more potential return in spite of greater investment and risk, it may decide to handle it’s own export. In this case it can carry on direct exporting in the following types: 1) Domestic-based Export Department or Division – an export sales manager carries on the actual selling and draws on market assistance as needed. It might evolve into a self-contained export department performing all the activities involved in export and operating as a profit center. 2) Overseas Sales Branch or Subsidiary – an overseas sales brand allows the manufacturer to achieve greater presence and program control in the foreign market. The sales branch handles sales and distribution and might handle warehousing and promotion as well. It often serves as a display center and customer-service center. 3) Traveling Export Sales Representatives – the company sends home-based sales representatives abroad to find business. 4) Foreign-Based Distributors or Agents – the company hires foreign-based distributors or agents to sell the goods on behalf of the company. They might be given exclusive rights to represent the manufacturer in that country or only general rights. To sum up, direct exporting is exporting to overseas customers, who might be wholesalers, retailers or users, without the use of export houses etc. Licensing is an alternative to foreign direct investment by which overseas producers are given rights to use the licensor's production process in return for a fee or royalty payments. Licensing agreement is a commercial contract whereby the licensor gives something of value to the licensee for exchange for certain performances and payments. Licensing represents a simple way for a manufacturer to become involved in international marketing. The licensor may provide any of following: 1) Rights to produce a patented product or use a patented production process, 2) Manufacturing know-how (unpatented), 3) Technical advice and assistance including the supply of essential materials, components, plants, 4) Marketing advice and assistance, 5) Rights to use a trademark or brand. The licensor gains entry into the foreign market at little risk; the licensee gains production expertise or a well-known product or name without having to start from scratch. Licensing is growing in extent and importance through the world. It is used by all sizes of firms, as it has many advantages for a licensor: 1) It requires no investment in other words save the continuing costs of monitoring the agreement, 2) It enables entry into markets that would otherwise be closed, 3) As a mode of entry, it is relatively simple and quick, 4) The licensor gains access to knowledge of local conditions, 5) New products can be introduced to many countries quickly because of low investment requirements, 6) It provides all the usual benefits of overseas production, 7) It can be a source of competitive advantage, if it spreads the firm's proprietary technology, giving it wider exposure than that of rival. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 66 Licensing also suffers from drawbacks, however: 1) Revenues from licensing are very low, usually less than 10% of turnover, 2) A licensee may eventually become the licensor's competitor, 3) Although the contract may specify a minimum sales volume, there is some danger that licensee will not fully exploit the market, 4) Product quality might deteriorate if the licensee has more lax attitude to quality control than licensor, 5) Governments may impose restrictions or conditions on payment of royalties to the licensor or on the supply of components, 6) It is often difficult to control the licensee effectively. The licensee's objectives often conflict with those of licensor and disagreements are common. Companies can enter foreign markets on other bases. A company can sell a management contract to manage a foreign hotel, airport, hospital, or other organization for a fee. In this case a firm is exporting a service instead of a product. Management contracting is a low-risk method of getting into a foreign market, and furthermore it yields income from the beginning. The arrangement is especially attractive if the contracting firm is given an option to purchase some share in the managed company within a stated period. On the other hand, the arrangement has no sense if the company can put its scarce management talent to better uses or if there are greater profits to be made by undertaking the whole venture. Management contracting prevents the company from going into competition with its clients. Another method of entry is contract manufacturing, where the firm engages local manufactures to produce the product. In other words a firm (the contractor) makes a contract with another firm (the contractee) abroad whereby the contractee manufactures or assembles a product on behalf of the contractor. The contractor retains full control over marketing and distribution, while the manufacturing is done by a local firm. The drawback of contract manufacturing is less control over the manufacturing process and the loss of potential profits on manufacturing. On the other hand, it offers the company a chance to start faster, with less risk, and with the opportunity to form a partnership or buy out the local manufacturer later. So Advantages of contract manufacturing include the following: 1) There is no need to invest in plant overseas, 2) Risk associated with currency fluctuation, is largely avoided, 3) The risk asset expropriation is minimized, 4) Control of marketing is retained by the contractor, 5) A product manufactured in overseas market may be easier to sell, especially to government customers, 6) Lower transport costs and, sometimes, lower production cost can be obtained. Disadvantages of contract manufacturing include the following: 1) Overseas contractee producers who are reliable and capable cannot always be easily identified, 2) Sometimes the contractee producer's personnel must relieve intensive and substantial technical training, 3) The contractee producer may eventually become a competitor, 4) Quality control problems in manufacturing may arise. Consequently, contract manufacturing is perhaps best suited : 1) to countries where small size of the market discourages investment in plant, 2) to firm whose main strengths are in marketing rather in production. The next strategy of entering a foreign market is a joint venture. Joint venture is an arrangement of two or more firms (foreign investors and local investors) to create a manufacture and to join forces for manufacturing, financial and marketing purposes in which each has a share in both the equity and management of the business. Forming a joint venture might be necessary or desirable for economic or political reasons. The foreign firm might lack the financial, physical, or managerial resources to undertake the venture alone. Or the foreign government might require joint ownership as a condition for entry. Advantages of joint venture include the following: 1) Some government discourage or even prohibit foreign firms setting up independent operation; 2) Joint ventures are especially attractive to smaller or riskaverse firms, or where very expensive new technologies are behind researched and developed; 3) When funds are limited, joint ventures permit coverage of larger number of countries since each one requires less investment by each participator; 4) A joint ventures can reduce the risk of government intervention as a local firm is involved; 5) The participating enterprises benefit from all sources of profit; 6) Joint venture can provide close control over marketing and other operation, 7) A joint venture with an indigenous firm provides local knowledge, quickly. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 67 Disadvantages of joint ventures are that there can be major conflicts of interests between the different parties. Disagreements may arise over: 1) profit shares, 2) amount invested, 3) the management of the joint venture; 4) the marketing strategy, or other policies. One partner might want to reinvest earnings for growth, and the other partner might want to withdraw these earnings. Furthermore, joint ownership can hamper a multinational company from carrying out specific manufacturing and marketing polices on a worldwide basis. The ultimate form of foreign involvement is direct ownership of foreign-based assembly or manufacturing facilities. The foreign company can buy part or full interest in a local company or build its own facilities. As a company gains experience in export, and if the foreign market appears large enough, foreign production facilities offer distinct advantages:1) The firm does not have to share its profits with partners of any kind; 2) The firm does not have to share or delegate decision making and so there are no loses in efficiency arising from inter-firm conflict; 3) There are none of the communication problems that arise in joint ventures, license agreements etc.; 4) The firm is able to operate a completely integrated and systematic international system; 5) the firm will gain a better image in the host country because it creates jobs; 6) the firm develops a deeper relationship with government, customers, local suppliers, and distributors, enabling it to adapt its products better to the local marketing environment; 7) the firm assures itself access to the market in case the host country starts insisting that purchased goods have domestic content. Disadvantages of wholly owned overseas manufacture include the following: 1) The substantial investment funding required prevents some firms from establishing operations overseas; 2) Suitable managers, whether required in the overseas market or posted abroad from home, may be difficult to obtain; 3) Some overseas government discourage or even prohibit 100% ownership of an enterprise by a foreign company; 4) This mode of entry forgoes of overseas partner's market knowledge, distribution system and other local expertise. However, the firm has no choice but to accept these risks if it wants to operate on its own behalf in the host country. Many companies show a distinct preference for a particular strategy of entry. One company might prefer exporting because it minimizes its risk. Another company might prefer licensing because it is an easy way to make money without investing much capital. Another company might favor direct investment because it wants full control. Yet insisting on one mode of entry is too limiting. Some countries will not permit imports of certain goods nor allow direct investment but will only accept a joint-owned venture with a foreign national. Consequently, companies need to master all of these entry methods. Even though a company might have preferences, it needs to adapt to each situation. For instance, most sophisticated multinationals manage several entry modes simultaneously. REFERENCES: [1] BEARDEN W.O., INGRAM T.N., LAFORGE R.W.: Marketing: principles, perspectives. – Chicago: IRWIN, 1996, 631 p. [2] CATCORA, PHILIP R.: International marketing. – 9th ed. – Chicago: IRWIN, 1996, 770 p. [3] LAMBIN J.: Strategic Marketing management. – Berkshire: The McGraw – Hill Companies, 1997, 692 p. [4] PHILIP KOTLER, GARY ARMSTRONG: Principles of marketing, sixth edition, Prentice-Hall, International, Inc., 1994, 692 p. [5] PHILIP KOTLER: marketing management the millennium edition, Prentice-Hall, International (UK) Limited, Sydney, 2000, 131 p. [6] PHILIP KOTLER: Marketing management analysis, planning, implementation, and control, eighth edition, Northwestern University, 1994, 801 p. [7] RANGAN V.K., SHAPIRO B.P., MORIARTY R.: Business Marketing Strategy: Cases, concepts, and applications. – Chicago: IRWIN, 1995, 328 p. [8] SUDHARSHAN D. : Marketing strategy: relationships, offerings, timing and resource allocation. – Englewood Cliffs: Prentice-Hall, Inc., 1995, 446 p. [9] WALKER O.C., BOYD H.W., LARRECHE J-Cl.: Marketing strategy: Planning and implementation. – Chicago: IRWIN, 1995, 329 p. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 68 SCOPES OF MARKETING STRATEGY FOR 3G MOBILE SERVICES Emese Tokarčíková Department of Macro and Microeconomics, Faculty of Management Science and Informatics, University of Žilina, Moyzesova 20, 010 26 Žilina, Slovak Republic Tel: 00421 – 041 – 513 4421, fax 0041 –041 –565 2044 e–mail : [email protected] Abstract: Third generation (3G) of mobile services are the first multimedia services and more or less general available communication systems. Mobile operators, using this technology are able to offer new services for their customers. The purpose of marketing strategy is directing mobile operators to offer suitable services to selected category of customers – to be custom oriented. Keywords : mobile services, 3 generation, marketing strategy, customer requirements 1. INTRODUCTION Digital communications in the world have changed greatly. Mobile phones have been one of the technology success stories of the last few years. As well more and more households have home Internet access. Mobile telephones allows us to talk on the move. The Internet turns raw data into helpful services that people find easy to use. Now these two technologies create 3G, which is a statement of International Telecommunication Union (ITU) for third generation of mobile communication services. In the mid-1980’s, the ITU created the single standard of a family of technologies entitled ‘IMT-2000’, “International Mobile Telecommunications”. Following these trends European UMTS (which stands for Universal Mobile Telecommunications System), falls within the ITU's IMT-2000 vision of a global family of 3G mobile communication systems. 2. 3G MOBILE SERVICES The simplest definition of services says that services are actions that satisfy people’s needs. Mobile services are not counting. 3G mobile services offer new ways to communicate, access information, conduct business, learn and be entertained – liberated from slow, cumbersome equipment and immovable points of access. 3G mobile services have different classes for four types of traffic: · Conversational class (voice, video telephony, video gaming) · Streaming class (multimedia, video on demand, web cast) · Interactive class (web browsing, network gaming, database access) · Background class (email, SMS, downloading) Technically, generations which allows to provide these class are defined as: UMTS TECHNOLOGY FEATURES 1G AMPS Advanced Mobile Phone Service 2G CDMA Code Division Multiple Access TDMA GSM PDC Time Division Multiple Access Global System for Mobile Communication Personal digital cellular CDMA-2000 TD-SCDMA Based on the Interim Standard-95 CDMA standard Time division synchronous code-division multiple access 3G Fig. 1 History of mobile technologies W-CDMA Wide-band Code Division Multiple Access - Analogue voice service No data service Digital voice service 9.6 to 14.4K bit/sec. CDMA, TDMA and PDCoffer one-way data trans. only No always-on data connection - Superior voice quality Up to 2M bit/sec. always on data Broadband services like video and multimedia Enhanced roaming INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 69 Forum's Market Aspects Group has identified some common lifestyle attributes for mobile multimedia applications. Here is a list of possible type of services that will be available in 3G networks: · · · · · · · · · · · Fun: WWW, video, post card, snapshots, text, picture and multimedia messaging, datacast, personalisation applications (ring tone, screen saver, desk top), jukebox, virtual companion / pet... Work: Rich call with image and data stream, IP telephony, B2B ordering and logistics, information exchange, personal information manager, dairy, scheduler, note pad, 2-way video conferencing, directory services, travel assistance, work group, telepresence, FTP, instant voicemail, colour fax... Media: Push newspaper and magazines, advertising, classified... Shopping: E-commerce, e-cash, e-wallet, credit card, telebanking, automatic transaction, auction, micro billing shopping... Entertainment: News, stock market, sports, games, lottery, gambling, music, video, concerts, adult content... Education: Online libraries, search engines, remote attendance, field research... Peace of Mind: Remote surveillance, location tracking, emergency use... Health: Telemedicine, remote diagnose and heath monitoring... Automation: Home automation, traffic telematics, machine-machine communication (telemetry)... Travel: location sensitive information and guidance, e-tour, location awareness, timetables, eticketing... Add-on: TV, radio, PC, access to remote computer, MP3 player, camera, video camera, watch, pager, GPS, remote control unit... 3G will become an essential part of our everyday lives and catalyst for a whole new array of high-speed mobile services, providing personal mobility, interactivity and access to advanced broadband and positioning services anywhere, anytime. 3G is not about technology, it is about services. Basically, 3G opens the door to anything costumers can imagine. They will be able to do a multitude of things while going through their daily schedule, whether at work or at leisure. 3. SCOPES OF MARKETING STRATEGY FOR 3G MOBILE SERVICES Mobile operators need marketing strategy, which is adapted to the very different market of mobile data. “Winning strategy needs excellent customer and marketing conception and winning marketing conception needs suitable framework in strategic management system in a company. The most important part of strategy is marketing conception with its focus to suitable market segments and their customers.” [2] Therefore strategic market planning is a very first step to create excellent marketing strategy and conception and includes the planning process that yields decisions in how a business unit can best compete in the markets it elects to serve. The strategic plan is based upon the totality of the marketing process. Supports for mobile operators in creating strategic plan for 3G mobile services are experiences and results from 2G mobile services. It is necessary to analyze the success of GSM mobile services to be able to prognose costumer perspectives for next generation of mobile services. “Basic measurement in costumer perspective usually consist of measurement as: · market share · customer loyalty · recruitment of new customers · customer satisfaction · customer profitability “ [2] Because of 3G mobile services make possible to individualize customer requirements, therefore mobile operators need better customer specification, close customer orientation and dialog communications. The basic question in aspects of 3G mobile services, which is need to be answered clearly in marking strategy, is: INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 70 How can mobile operators, software designers, and handset producers inspire vendors to use their mobile phones for longer lengths of time and for more applications? To answer this question, mobile operators at first have to define: · what kind of costumers want to address (e.g. following by age, profession, demand, need, habitation) · which kind of product or service they want to address costumers ( which 3G mobile service or pack of services to offer ) · in what mobile operator is meaningful for customer. After that mobile operators can assign frame of their marketing strategy: · to define firm’s identity according to external requirement (mobile market) · to control efficiency and expedience of impact to potential partners on the market · to define and adhere firms´ principles and firms ´ culture and clime ( operator can be more competitive) · to define long term targets, because in case of 3G the returns of investments are long term matters · to create special unit for close dialog with customers and public General requirements to 3G mobile services characteristic for all customers are identified as following: One number One bill One phone Global roaming What does the customer want from the 3G services? Internet access Useful services Ease of use Value Fig. 2 Customer requirements Beyond these general request, customers have more individual needs. Therefore qualitative processes oriented on customers’ requirements and focused on satisfying their demand are determined terms in marketing strategy. Answering further questions and processes they can solve the basic problems in this field: · Should 3G services be targeted at all users? · What are the main needs that manage to demand 3G mobile services? · What are the key market segments that should be addressed with 3G services? · How big will the market for 3G services be in the future? · Is there a killer application that will revolutionise the industry? · What are the revenue opportunities for 3G mobile services providers? · What should the focus of service creation be to ensure early time to profit in 3G? · Who will be the key players in the 3G-value chain and how will they share the revenue? · Make strategic analysis: competitive, regulatory, supplier and customer · Analysis of risks and opportunities, based on a scenario approach · Identify the service portfolio and the associated marketing mix · Assign factors of influence customer’ s preferences · Appoint life-cycle of service · Define tariff policy and communication policy INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 71 4. CONCLUSION Mobile operators have to decide not only whether to build out or upgrade 3G infrastructures, but mainly which mobile services they will offer to customers and how they will profit from them? There are no set answers to these questions. For all they have to create outright marketing strategy which will answer not "why" but "how” to improve odds bided by new mobile technology. This technology is a challenge for mobile operators and the right strategy can made them more competitive and successful. It is also a chance to realize vision of “ building a mobile future where their customers can enjoy richer communications - anytime and anywhere, to be a friendly partner, deliver easy-to-use and inspiring solutions that let us GET MORE out of life and to believe in openness, respect for people and commitment to excellence.”[7] REFERENCES [1] TOKARČÍKOVÁ, E: Nové služby pre tretiu generáciu (UMTS) mobilných sietí, Žilina, Písomná práca k dizertačnej skúške, Žilinská Univerzita v Žiline, 2003 [2] VODÁK, J.: Meaning of marketing conception for strategic management system on global markets, Text book of International scientific conference: Marketing of the companies in V4 countries one step before the entry to European union, Matej Bel University in Banská Bystrica, 2003 [3] KOTLER,P: Marketing podle Kotlera, Management Press, Praha 2000 [4] http://www. umtsworld.com [5] http://www.tomiahonen.com [6] http://www.3g.co.uk [7] http://www.eurotel.sk INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 72 CULTURE IN INTERNATIONAL MARKETING AND BUYER BEHAVIOUR Kalugin Evgeniy The Plekhanov Russian Academy of Economics,Stremanniy pereulok 36, 113054, Moscow, Russian Federation. tel: +7 095 2379552; e-mail:evgeshk@ mail.ru. Abstract: To develop effective marketing programs, the marketing manager must have knowledge of the needs and wants of potential buyers, how they arise, and how and where they are likely to be satisfied. Keywords: Characteristics of culture International Marketing and buyer behavior, examples of Cultural Blunders Made by International Marketers, the culture sensitivity of markets, the development of global culture, cultural analysis of global markets, cross- cultural analysis. 1. INTRODUCTION Culture is the learned ways of group living and the group’s responses to various stimuli. It is also the total way of life and thinking patterns that are passed from generation to generation. It encompasses norms, values, customs, art, and beliefs. Culture is the patterns of behavior and thinking that people living in social groups learn, create, and share. Culture distinguishes one human group from others. A people's culture includes their beliefs, rules of behavior, language, rituals, art, technology, styles of dress, ways of producing and cooking food, religion, and political and economic systems. Anthropologists commonly use the term culture to refer to a society or group in which many or all people live and think in the same ways. Likewise, any group of people who share a common culture—and in particular, common rules of behavior and a basic form of social organization—constitutes a society. Thus, the terms culture and society are somewhat interchangeable. Characteristics of culture: Culture is prescriptive. It prescribes that kinds of behavior considered acceptable in the society. The prescriptive characteristic of culture simplifies a consumer’s decision-making process by limiting product choices to those, which are socially acceptable. These same characteristics create problems for those products not in tune with the consumer’s cultural beliefs. Culture is socially shared. Culture cannot exist by itself. Members of a society must share it. Thus acting to reinforce culture’s perspective nature. Culture is learned. Culture is not inherited genetically; it must be learned and acquired. Socialization or enculturation occurs when a person absorbs or learns the culture in which he or she is raised. Culture facilitates communication. One useful function provided by culture is to facilitate communication. Culture usually imposes common habits of though and feeling among people. Thus, within a given group culture makes it easier for people to communicate with one another. But culture may also impede communication across groups because of a lack of shared common culture values. This one reason why a standardized advertisement may have difficulty communicating with consumers in foreign countries. How marketing efforts interact with a culture determines the success or failure of a product. Advertising and promotion require special attention because the play a key role in communicating product concepts and benefits to the target segment. Culture is subjective people in different cultures often have different ideas about the same object. What is acceptable in one culture may not necessarily be so in another. In this regard, culture is both unique and arbitrary. Culture is enduring, because culture is shared and passed along from generation to generation, it is relatively stable and somewhat permanent. Old habits are hard to break, and people and people tend to maintain its own heritage in spite of continuously changing world. Culture is cumulative. Culture is based on hundreds or even thousands of years of accumulated circumstances. Each generation adds something of its own of culture before passing the heritage on to the next generation. Therefore culture tends to be broader based over time, because new ideas are incorporated and become a part of the culture. Culture is dynamic. Culture is passed along from generation to generation, but one should not assume that culture is static and immune to change. Culture is constantly changing it adapts itself to new situations and new sources of knowledge. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 73 2. INTERNATIONAL MARKETING AND BUYER BEHAVIOR. An understanding of buyer behavior is central to successful marketing. To develop effective marketing programs, the marketing manager must have knowledge of the needs and wants of potential buyers, how they arise, and how and where they are likely to be satisfied. Buyer behavior is affected by many factors. Class, education, age, and psychosocial traits are just four of the many factors useful in distinguishing different buyer groups. Researching the relationships that exist between the marketing-mix variables and buyer needs and response. From this effort have evolved many buyer behavior models, concepts, and techniques. International Marketing’s Four Buyer Behavior Tasks Apparent similarities such as language can hide subtle but important differences between markets. International marketers have often shown a higher propensity to misinterpret a marketing situation when the cultural and economic environments of the foreign market are apparently the same as their own. For example, Philip Morris lost a considerable amount of money when tried to introduce a U.S. cigarette to the Canadian market. Management was under the erroneous impression that Canadians and Americans had similar smoking habits because the spoke the same language, had similar cultural heritages, dresses more or less the same, and watched many of the same television programs. Campbell Soups lost $30 million in Europe before it accepted the idea that British and U.S. soup consumers were different in three important ways. First British soups consumers have different taste preferences. Campbell soups made no attempt to modify the taste of their soups for the British palate. Second, British soup consumers had not been educated to the condensed soup product concept. Because of the smaller can size. Third, British soup consumers did not respond the same way to U.S. advertisement as U.S. consumer did. Examples of Cultural Blunders Made by International Marketers: Language: A U.S. toothpaste manufacturer promised its customers that they would be more “interesting” if they used the firm’s toothpaste. What the advertising coordinators did not realize, however, was that in Latin American Countries “interesting” is another euphemism for “pregnant”. Food: Chase and Sanborn met resistance when it tried to introduce its instant coffee in France. In the home, the consumption of coffee plays more of a ceremonial role than in the English home. The preparation of “real” coffee is a touchstone in the life of the French housewife, so she will generally reject instant coffee because its causal characteristics do not “fit” into the French eating habits. Values: In 1963, Dow Breweries introduced a new beer in Quebec, Canada; called “kebec” the promotion incorporated the Canadian flag and attempted to evoke nationalistic pride. The strategy backfired when major local groups protested the “profane” use of “sacred” symbols. Religion: England’s East India Company once caused a revolt when it did not modify a product. In 1857, bullets were often encased in pig wax, and the tops had to be bitten off before the bullets could be fired. The Indian soldiers revolted since it was against their religion to eat pork. Hundreds of people were killed before order was restored. Social Norms and time: A telephone company tried to incorporate a Latin flavor in its commercials by employing Puerto Rican actors. In the ad, the wife said to her husband, “ run and phone Mary. Tell her we will be a little late.” This commercial has two major cultural errors. Latin wives seldom dare order their husband around, and almost no Latin would feel it necessary to phone to warm of tardiness since it is expected. The Sociocultural Dimension of Buyer Behavior Culture does influence Consumption to a great extent. Consumption patterns, living styles, and the priority of needs are all dictated by culture. Culture prescribes the manner in which people satisfy their desires. Not surprisingly, consumption habits very greatly. The consumption of beef provides a good illustration. Some Chinese do not consume beef at all, believing that it is improper to eat cattle that work on farms, thus helping to provide foods such as rice and vegetables. 3. THE CULTURE SENSITIVITY OF MARKETS. Markets can be divided into consumer markets and industrial markets. Consumer markets can be further subdivided into durable goods markets and nondurable goods markets. A further profitable distinction in the international market place is to divide durable goods into technological products and nontechnological products. 1. Industrial markets: the main distinction between industrial markets and consumer markets is that industrial buyers are interested in solving problems. Generally, these problems are to reduce costs, to increase production administrative efficiency, to produce a particular type of product, or to effect a combination of these goals. Consequently, industrial buyers tend to be rational and emphasize economic goals. Cultural and social INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 74 considerations play a relatively less important role in the purchase decision. To some extent, industrial markets can be viewed as global markets. There are, however, differences that need to be considered when selecting the products to be marketed and the marketing programs to be used. Government regulations, the size and sophistication of the potential buyer’s operations, and the context within which the product or service is to be used all have an impact on the marketing effort. 2. Consumer Markets: it consists of buyers interested in satisfying a personal need or want. They are generally more susceptible to cultural and social forces than are industrial buyers. The influence of sociocultural factors is most apparent in the purchase of nondurable products such as clothing, foods, etc. there are expectations, however, that make the international marketer’s task more interesting and challenging. 4. THE DEVELOPMENT OF GLOBAL CULTURE. Rapid changes in technology in the last several decades have changed the nature of culture and cultural exchange. People around the world can make economic transactions and transmit information to each other almost instantaneously through the use of computers and satellite communications. Governments and corporations have gained vast amounts of political power through military might and economic influence. Corporations have also created a form of global culture based on worldwide commercial markets. Local culture and social structure are now shaped by large and powerful commercial interests in ways that earlier anthropologists could not have imagined. Early anthropologists thought of societies and their cultures as fully independent systems. But today, many nations are multicultural societies, composed of numerous smaller subcultures. Cultures also cross national boundaries. For instance, people around the world now know a variety of English words and have contact with American cultural exports such as brand-name clothing and technological products, films and music, and mass-produced foods. Many anthropologists have become interested in how dominant societies can shape the culture of less powerful societies, a process some researchers call cultural hegemony. Today, many anthropologists openly oppose efforts by dominant world powers, such as the U.S. government and large corporations, to make unique smaller societies adopt Western commercial culture. 5. CULTURAL ANALYSIS OF GLOBAL MARKETS. Whether a firm is pursuing a national-market or global-market strategy, it is interested in increasing the effectiveness and efficiency of its marketing programs within and across foreign markets. It must therefore know to what degree it can use the same product, pricing, promotion, and distribution strategies in more than one market. Unfortunately, the dual goals of program effectiveness and efficiency are in conflict. Market effectiveness is achieved by adapting marketing programs to marketing characteristics and conditions within markets. While doing so incurs additional marketing and production costs, the firm strengthens its market competitiveness by being more responsive to the needs of the marketplace. Efficiency, on the other hand, is achieved by minimizing marketing program changes across markets. Thus the firm minimizes marketing and production costs and strengthens its competitiveness vis-а-vis its competitors. The economic and competitive implications of both goals need to be taken into account when making program adaptation decisions. Both goals depend on understanding the cultural context of each market and the degree to which they are culturally similar. Thus, global companies need to develop a capability to conduct cross-cultural analysis of buyer behavior. Such a capability can help these companies optimally balance the competitive benefits to be derived from effectiveness and efficiency. 6.CROSS- CULTURAL ANALYSIS. “Cross-cultural analysis is the systematic comparison of similarities and differences in the material and behavioral aspects of cultures.” In the marketing, cross-cultural analysis is used to gain an understanding of market segments within and across national boundaries. The purpose of this analysis is to determine whether the marketing program, or elements of the program, can be used in more than one foreign market or must be modified to meet local conditions. The approaches used to gain this understanding draw on the methods developed by such social sciences as anthropology, linguistics, and sociology. Standard marketing research INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 75 techniques, such as multi attribute and psychographic techniques, can be used. For example, Berger, Stern and Johansson used to multi attribute method to study Japanese and American car buyers, and Boote used a psychographics approach to study the segmentation of the Europe community. In marketing, cross-cultural analysis most often involves identifying the effects culture may have on family purchasing roles, product function. Product design, sales and promotion activities, channel systems, and pricing. One approach suggested by Engel, Blackwell, and Miniard to the study of the effects of culture on buyer behavior, and thus in the marketing- mix elements. This involves answering a comprehensive list of questions, although these are neither exhaustive nor specific. For example, a manufacturer of processed foods would be interested in knowing the impact that culture has on such things as taste, purchasing habits, and eating habits. A manufacturer of household appliances, on the other hand, would be particularly interested in how potential buyers view a product’s reliability, durability, and reparability. 7. CONCLUSION There is no doubt that the international marketing process do faces a large set of variables as it take place over different countries and it does act in different environments. One of the most determinant environments to the success of the international marketing process is Culture, which hold the reason for many human acts and behavior. Reaching to that point international marketer should study deeply culture treaties of a country the company is planning to act in. so that special amendments in the organization overall plans and actions is made to act in accordance with the new market variables. REFERENCES: [1] [2] [3] [4] [5] International Marketing, Sixth Edition. Vern & Ravi. Dryden Press. International Marketing, Ninth Edition. Philip Cateora. IRWIN. International Marketing, Sixth Edition. Michael & Ilkka. Harcourt. International Marketing, Tenth Edition. Phillip & john graham. Consumer Behavior, Eighth Edition. James F. Engel & Roger Blackwell, Bowel Miniard. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 76 EFFECTIVE COMMUNICATION AND PROMOTION STRATEGY Oreshkin Alexey Gennadievich Plehanov Russian Academy of Economics: Stremiannii alley, 36, 8 (095) 2379552, [email protected] Abstract: In preparing marketing communications, the communicator has to understand the nine elements of any communication process: sender, receiver, encoding, decoding, message, media, response, feedback, and noise. The communicator's first task is to identify the target audience and its characteristics. Next, the communicator has to define the response sought, whether it be awareness, knowledge, liking, preference, conviction, or purchase. Then a message should be constructed with an effective content, structure, and format. Media must be selected, both for personal communication and nonpersonal communication. The message must be delivered by a credible source—someone who is an expert and is trustworthy and likable. Finally, the communicator must collect feedback by watching how much of the market becomes aware, tries the product, and is satisfied in the process. Keywords: Advertising, Sales promotion, Public relations and Publicity, Personal selling, Direct marketing, Sender, Receiver, Message, Media, Encoding, Decoding, Response, Feedback, Noise, Personal Communication Channels, Nonpersonal Communication Channels. Modern marketing calls for more than developing a good product, pricing it attractively, and making it accessible to target customers. Companies must also communicate with present and potential stakeholders, and with the general public. The marketing communication: advertising, sales promotion, public relations and publicity, personal selling, and direct marketing. Advertising. Because of the many forms and uses of advertising, it is hard to generalize about its unique qualities as a part of the promotion mix. Yet several qualities can be noted. Advertising's public nature suggests that the advertised product is standard and legitimate. Because many people see ads for the product, buyers know that purchasing the product will be understood and accepted publicly. Advertising also lets the seller repeat a message many times, and it lets the buyer receive and compare the messages of various competitors. Large-scale advertising by a seller says something positive about the seller's size, popularity, and success. Advertising is also very expressive, allowing the company to dramatize its products through the artful use of print, sound, and color. On the one hand, advertising can be used to build up a long-term image for a product (such as Coca-Cola ads). On the other hand, advertising can trigger quick sales (as when Sears advertises a weekend sale). Advertising can reach masses of geographically spread-out buyers at a low cost per exposure. Advertising also has some shortcomings. Although it reaches many people quickly, advertising is impersonal and cannot be as persuasive as a company salesperson. Advertising is able to carry on only a oneway communication with the audience, and the audience does not feel that it has to pay attention or respond. In addition, advertising can be very costly. Although some advertising forms, such as newspaper and radio advertising, can be done on small budgets, other forms, such as network TV advertising, require very large budgets. Personal Selling. Personal selling is the most effective tool at certain stages of the buying process, particularly in building up buyers' preferences, convictions, and actions. Compared to advertising, personal selling has several unique qualities. It involves personal interaction between two or more people, so each person can observe the other's needs and characteristics and make quick adjustments. Persona! selling also allows all kinds of relationships to spring up, ranging from a matter-of-fact selling relationship to a deep personal friendship. The effective salesperson keeps the customer's interests at heart in order to build a long-term INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 77 relationship. Finally, with personal selling the buyer usually feels a greater need to listen and respond, even if the response is a polite "no thank you." These unique qualities come at a cost, however. A sales force requires a longer-term commitment than does advertising—advertising can be turned on and off, but salesforce size is harder to change. Personal selling is also the company's most expensive promotion tool, costing industrial companies an average of almost $200 per sales call.10 U.S. firms spend up to three times as much on personal selling as they do on advertising. Sales Promotion. Sales promotion includes a wide assortment of tools— coupons, contests, cents-off deals, premiums, and others—all of which have many unique qualities. They attract consumer attention and provide information that may lead to a purchase. They offer strong incentives to purchase by providing inducements or contributions that give additional value to consumers. And sales promotions invite and reward quick response- Whereas advertising says "buy our product," sales promotion says "buy it now." Companies use sales-promotion tools to create a stronger and quicker response. Sales promotion can be used to dramatize product offers and to boost sagging sales. Sales-promotion effects are usually short-lived, however, and are not effective in building long-run brand preference. Public Relations. Public relations offers several unique qualities. It is very believable—news stories, features, and events seem more real and believable to readers than do ads. Public relations also can reach many prospects who avoid salespeople and advertisements—the message gets to the buyers as "news" rather than as a sales-directed communication. And, like advertising, public relations can dramatize a company or product. Marketers tend to underuse public relations or to use it as an afterthought. Yet a well-thought-out public relations campaign used with other promotion mix elements can be very effective and economical. The communication process consists of nine elements: sender, receiver, message, media, encoding, decoding, response, feedback, and noise. To get their messages through, marketers must encode their messages in a way that takes into account how the target audience usually decodes messages. They must also transmit the message through efficient media that reach the target audience and develop feedback channels to monitor response to the message. To communicate effectively, marketers need to understand the fundamental elements underlying effective communication. A communication model includes nine elements. Two represent the major parties in a communication sender and receiver. Two represent the major communication tools – message and media. Four represent the major communication functions – encoding, decoding, response, and feedback. The last element in the system is noise (random and competing messages that may interfere with the intended communication). The model underscores the key factors in effective communication. Senders must know what audiences they want to reach and what responses they want to get. They must encode their messages in a way that understands how the target audience usually decodes messages. They must transmit the message through efficient media that reach the target audience and develop feedback channels to monitor the responses. For a message to be effective, the sender's encoding process must mesh with the receiver's decoding process. The more the sender's field of experience overlaps with that of the receiver, the more effective the message is likely to be. This puts a burden on communicators from one social stratum (such as advertising people) who want to communicate effectively with another stratum (such as factory workers). The sender's task is to get his or her message through to the receiver. The target audience may not receive the intended message for any of three reasons: 1. Selective attention: People are bombarded by 1,600 commercial messages a day, of which 80 are consciously noticed and about 12 provoke some reaction. Selective attention explains why ads with bold headlines promising something, such as "How to Make a Million," have a high likelihood of grabbing attention 2. Selective insertion: Receivers will hear what fits into their belief system. As a result, receivers often add things to the message that are not there (amplification) and do not notice other things that are there (leveling). The communicator's task is to strive for simplicity, clarity, interest, and repetition to get the main points across. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 78 3. Selective retention: People will retain in long-term memory only a small fraction of the messages that reach them. If the receiver's initial attitude towards the object is positive and he or she rehearses support arguments, the message is likely to be accepted and have high recall. If the initial attitude is negative and the person rehearses counterarguments, the message is likely to be rejected but to stay in long-term memory. Because much of persuasion requires the receiver's rehearsal of his or her own thoughts, much of what is called persuasion is actually self-persuasion. The communicator considers audience traits that correlate with persuasibilily and uses them to guide message and media development. People of high education or intelligence are thought to be less persuasible, but the evidence is inconclusive, Those who accept external standards to guide their behavior and who have a weak self-concept appear to be more persuasible, as do persons who have low self-confidence. Fiske and Hartley have outlined some general factors that influence the effectiveness of communication: • The greater the monopoly of the communication source over the recipient, the greater the recipient's change or effect in favor of the source. • Communication effects are greatest where the message is in line with the receiver's existing opinions, beliefs, and dispositions, • Communication can produce the most effective shifts on unfamiliar, lightly felt, peripheral issues, which do not he at the center of the recipient's value system. • Communication is more likely to be effective where the source is believed to have expertise, high status, objectivity, or likability, but particularly where the source has power and can be identified with. • The social context, group, or reference group will mediate the communication and influence whether or not the communication is accepted. Developing effective communications involves eight steps: 1. Identify the target audience, 2. determine the communications objectives, 3. design the message, 4. select the communication channels, 5. establish the total communications budget, 6. decide on the communications mix, 7. measure the communications` results, and 8. manage the integrated marketing communication process. In identifying the target audience, the marketer needs to perform familiarity and favorability analyses, then seek to close any gap that exists between current public perception and the image sought. Communications objectives may be cognitive, affective, or behavioral – that is, the company might want to put something into the customer`s mind, change the consumer’s attitude, or get the consumer to act. In designing the message, marketers must carefully consider message content, message structure, message format, and message format, and message source. Communication channels may be personal (advocate, expert, and social channels) or nonpersonal (media, atmospheres, and events). PERSONAL COMMUNICATION CHANNELS Personal communication channels involve two or more persons communicating directly with each other face to face, person to audience, over the telephone, or through e-mail. Personal communication channels derive their effectiveness through the opportunities for individualizing the presentation and feedback. A further distinction can be drawn among advocate, expert, and social communication channels. Advocate channels consist of company salespeople contacting buyers in the target market. Expert channels consist of independent experts making statements to target buyers. Social channels consist of neighbors, friends, family members, and associates talking to target buyers. In a study of 7,000 consumers in seven European countries, 60 percent said they were influenced to use a new brand by family and friends. Many companies are becoming acutely aware of the power of "word of mouth,' They are seeking ways to stimulate social channels to recommend products and services. Regis McKenna advises a software company launching a new product to \w-mote it initially to the trade press, opinion luminaries, and financial analysts, who can supply favorable word of mouth; then to dealers; and finally to customers MH attracted customers with its INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 79 Friends and Family program, which encourages MCI users to ask friends and family members to use MCI so that both parties will benefit from | lower telephone rates. See the Marketing Memo "How to Develop Word-ofMouth Referral Sources to Build Business." Personal influence carries especially great weight in two situations. One is with products that are expensive, risky, or purchased infrequently. Here buyers are likely to be strong information seekers. The other situation is where the product suggests something about the user's status or taste. Here buyers will consult others to avoid embarrassment. Companies can take several steps to stimulate personal influence channels to work on their behalf: • Identify influential individuals and companies and devote extra effort to them; In industrial selling, the entire industry might follow the market leader in adopting innovations. • Create opinion leaders by supplying certain people with the product on attractive terms; A new tennis racket might be offered initially to members of high school tennis teams at a special low price. Or Toyota could offer its more satisfied customers a small gift if they are willing to advise prospective buyers. • Work through community influential such as local disk jockeys, class presidents, and presidents of women's organizations: When Ford introduced the Thunderbird, it sent invitations to executives offering them a free car to drive for the day. Of the 15,000 who took advantage of the offer, 10 percent indicated that they would become buyers, whereas 84 percent said they would recommend it to a friend. • Use influential or believable people in testimonial advertising: Quaker Oats pays basketball star Michael Jordan several million dollars to make Gatorade commercials. Jordan is viewed as the world's premiere athlete, so his association with a sports drink is a credible connection, as is his extraordinary ability to connect with consumers, particularly children. • Develop advertising that has high "conversation value": Ads with high conversation value often have a slogan that becomes part of the national vernacular. In the mid-1980s, Wendy's "Where's the Beef?" campaign (showing an elderly lady named Clara questioning where the hamburger was hidden in all that bread) created high conversation value. Nike's "Just do it" ads have created a popular command for those unable to make up their minds or take some action. • Develop word-of-mouth referral channels to build business: Professionals will often encourage clients to recommend their services. Dentists can ask satisfied patients to recommend friends and acquaintances and subsequently thank them for their recommendations. • Establish an electronic forum: Toyota owners who use an online service line such as America Online can hold online discussions to share experiences. NONPERSONAL COMMUNICATION CHANNELS Nonpersonal channels include media, atmospheres, and events. Media consist of print media (newspapers, magazines, direct mail), broadcast media (radio, television), electronic media (audiotape, videotape, videodisk, CD-ROM, Web page), and display media (billboards, signs, posters). Most nonpersonal messages come through paid media. Atmospheres are "packaged environments" that create or reinforce the buyer's leanings toward product purchase. Law offices are decorated with Oriental rugs and oak furniture to communicate "stability" and "success." A luxury hotel will use elegant chandeliers, marble columns, and other tangible signs of luxury. Events are occurrences designed to communicate particular messages to target audiences. Publicrelations departments arrange news conferences, grand openings, and sports sponsurships to achieve specific communication effects with a target audience. Although personal communication is often more effective than mass communication, mass media might be the major means of stimulating persona! communication. Mass communications affect personal attitudes and behavior through a two-step flow-of-communication process. Ideas often flow from radio, television, and print to opinion leades, and from these to the less media-involved population groups. This two-step flow has several implications. First, the influence of mass media on public opinion is not as direct, powerful, and automatic as INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 80 supposed. It is mediated by opinion leaders, people whose opinions are sought or who carry their opinions to others- Second, the two-step flow challenges the notion that consumption styles are primarily influenced by a "trickle-down" or "trickle-up" effect from mass media. People interact primarily within their own social group and acquire ideas from opinion leaders in their group. Third, two-step communication suggests that mass communicators should direct messages specifically to opinion leaders and let them carry the message to others. Pharmaceutical firms should promote new drugs to the most influential physicians first. Communication researchers are moving toward a social-structure view of interpersonal communication. They see society as consisting of cliques, small groups whose members interact frequently. Clique members are similar, and their closeness facilitates effective communication but also insulates the clique from new ideas. The challenge is to create more system openness so that cliques exchange information with others in the society. This openness is helped by people who function as liaisons and bridges, A Unison is a person who connects two or more cliques without belonging to either. A bridge is a person who belongs to one clique and is linked to a person in another clique. Although there are many methods used to set the promotion budget, the objective-and-task method, which calls upon marketers to develop their budgets by defining their specific objectives, is the most desirable. In deciding on the marketing communications mix, marketers must examine the distinct advantages and costs of each promotional tool. They must also consider the type of product market in which they are selling, whether to use a push or a pull strategy, how ready consumers are to make a purchase, the product's stage in the product life cycle, and the company's market rank. Measuring the marketing communications mix's effectiveness involves asking members of the target audience whether they recognize or recall the message, how many times they saw it, what points they recall, how they felt about the message, and their previous and current attitudes toward the product and company. Managing and coordinating the entire communications process calls for integrated marketing communications (IMC). REFERENCES: [1] BEARDEN W.O., INGRAM T.N., LAFORGE R.W.: Marketing: principles, perspectives. – Chicago: IRWIN, 1996, 631 p. [2] CATCORA, PHILIP R.: International marketing. – 9th ed. – Chicago: IRWIN, 1996, 770 p. [3] LAMBIN J. - .J: Strategic Marketing management. – Berkshire: The McGraw – Hill Companies, 1997, 692 p. [4] PHILIP KOTLER, GARY ARMSTRONG: Principles of marketing, sixth edition, Prentice-Hall, International, Inc., 1994, 692 p. [5] PHILIP KOTLER: marketing management the millennium edition, Prentice-Hall, International (UK) Limited, Sydney, 2000, 131 p. [6] PHILIP KOTLER: Marketing management analysis, planning, implementation, and control, eighth edition, Northwestern University, 1994, 801 p. [7] RANGAN V.K., SHAPIRO B.P., MORIARTY R.: Business Marketing Strategy: Cases, concepts, and applications. – Chicago: IRWIN, 1995, 328 p. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 81 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 82 MARKETING PREPAREDNESS OF SMALL AND MIDDLE-SIZE ENTERPRISES FOR THE ENTRY INTO THE EU Jaroslav Ďaďo Faculty of Economics, Matej Bel University, Banská Bystrica, SLOVAKIA Tel. 00421-48-4152 786, fax. 00421-48-411 6859, : [email protected] Abstract: Globalisation trends are also affecting the countries of Eastern Europe, which, until recently, were separate from the world of marketing. The signing of the accession protocols for the entry into the EU has an enormous influence on small and middle-size enterprises. This step opens up more room for the development of marketing activities of enterprises on a qualitative level. The development and marketing activities of small and middle-size enterprises, which are preparing for the entry into the international market or for those who, in their domestic market encounter international competition, have special importance. Keywords: small and middle-size enterprises, European Union, SWOT analysis; The process of globalisation and integration of countries reflects the change in the sale of production ratio between foreign exchange and domestic consumption. In the 1920s, the ratio was 93-95% to 7-5%. Now, this ratio is reversed. Present day integration trends, specifically in Europe, are quickly and strongly influencing Slovakia as well. On the Slovak market, these trends are more visible in the offers of foreign companies. Only a few Slovak companies can, in this field, dictate trends on the European scale. Small and middle-size enterprises are even less capable of this. Despite this fact, trends of globalisation are strongly showing their presence on the Slovak market as well. Therefore, even companies that do not directly enter the European Economic Community (EEC), have to accept global and European trends in marketing. Many companies exist that will not enter the EEC for a long time. Nevertheless, these will have to (and already do) face heavy foreign competition on the Slovak domestic market. The foreign competition does not only enter the market with a competing product, but also with thought through marketing formation and with a marketing strategy. When concerning what products are offered on the Slovak market, globalisation trends are more and more visible in consumption. The consumer meets on a daily basis products by foreign companies, and s/he compares this to domestic products. This whole process directly influences big, small and middle-size enterprises. The premise for the entry of small and middle-size Slovak enterprises into the EEC and their integration into the trends of international marketing are defined by a lot of factors. Core areas that need to be addressed for larger involvement of Slovak companies in European trends are mainly the following: · country’s macroeconomic indicators and their pro-export state policy · competitive abilities of companies · other factors 1. PREMISES FOR MARKETING Slovakia already started with an association agreement to the European Union as a part of the Czechoslovak Federation. Each government, since the separation of the federation, declared their attempt to become part of international economic aggregations such as the OECD, CEFTA or the European Union. Slovakia is a member of CEFTA, a group of countries V4, also a member of OECD, and it is also on the right track for the entry into the European Union. The integration efforts can be also supported by a number of bilateral agreements signed by the Slovak Government. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 83 Even if the role of the state in the market economy is declining, the state can eventually support the involvement of small and middle-size enterprises in international economic cooperation. The state can also create better or worse conditions for the entry of foreign capital into Slovak economy. The development of macroeconomic indicators (as well as the balance of the foreign trade balance, amount of export, amount of foreign investment, etc.) creates the development of globalisation and integration trends. These indicators, sometimes, strike with similarity in post-communist countries. The difference of individual countries’ positions depends on a thorough examination of separate indicators. The development of these factors directly influences both small and middle-size enterprises in Slovakia. Liberalisation of the market, which is a symbol of globalization, meant catastrophe for some countries and several sectors. In the conflict of global tendencies, disagreements occur for example between Slovakia and Hungary regarding the animal market. The Polish are facing pressure from Slovakia and the Czech Republic in the sugar market. Slovaks refuse to import Czech beer, Hungarians force their corn into Poland, and the Baltic countries are faced with the import of cheaper meat. Individual countries are competing with each other for foreign capital. By doing so, they mainly use various tax policies, tools to support active work policy, prices of land, and “gifts” of infrastructure, etc. The support of pro-export policy as constructed in several countries is based on the amount of international trade. This indicator indirectly reflects the level of active involvement of companies in the activities of international marketing. An important factor, which reflects the globalisation process, is the entry of foreign investment. This creates a premise for the development of cooperation of small and middle-size enterprises with large companies. The cooperation of small companies with middle-size enterprises, with a foreign investment, creates a premise for the development of companies, which will allow them to become familiar with the rules of their existence and determine their successful entry into the international market. The field related to foreign direct investment (FDI) is necessary, not only for the economy of a specific country, but also for SME*. This then creates an option for involving SME in cooperation with foreign capital and indirectly an entry onto the international market as well. In the year 2000, the field structure of the FDI in Slovakia, was dominated by production (54%), followed by transportation and communication (15%), money and insurance (13%), and trade (12%). Construction reached only 1% and services 2%. The importance of marketing in the activities of SME can also be derived from macroeconomic indicators in the specific field of a SME. The amount of production in Slovakia, in 2001, was % from GDP. SME share in the total export of Slovakia was % (from 11.9 billion USD). 2. COMPANIES’ ABILITY TO COMPETE The ability to compete, of small and middle-size enterprises in Slovakia, (as well as in the former Board of Common Economy Cooperation) are dominated by historical influences, which are the result of field structure, concentration of capacities, using of technologies, and the system of taking over a market valid until 1990. The structure of small and middle-size enterprises in Slovakia practically did not exist from 1950 to 1990. The capacities of numerous companies were, in the time, limited for the needs of an approximately 350 million market. The field industry was unbalanced. Historical factors are also strongly visible even in the human resources area of small and middle-size enterprises, starting with the top level management to the worker. This human resource phenomenon in the sphere needs to be presently seen as relationships between the owners, managers, and workers. This has also determined the formulation of the SME ability to compete. The current analysis of strengths and weaknesses of Slovak SME reflects its characteristic in the table number 1. The analysis is dominated by a marketing view. The indicated characteristics determine the premise for the integration into the process of global and European marketing. * Small and middle-size enterprises (translator’s note) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 84 The analysis represents a summary of information accumulated during workshops organized in the cooperation with PHARE. The aim of the workshops was to prepare SME management, with a potential for export, to make their own marketing plan for export. In the scope of the project, eight workshops were realized, which over 200 SME managers from around Slovakia attended. Table No.1 the analysis of strengths and weaknesses of SME Strengths Weaknesses - high level of management creativity - high analytical abilities of management - creativity of management - qualified work force - internal reserves of companies - ability to handle development of a product - ability to produce in conditions with limited resources - socially moderate employees - vacant production capacities - the quality of production bases on subsidies for special production - local perception of the market - missing real constructive production - low ability to acquire information - low level of marketing creativity - low ability to find options on the market - inappropriate usage of ownership rights of the owners - lack of financial resources - unclear as a whole, especially the marketing strategy of the company - orientation toward defensive strategies - low level of willingness of managers to cooperate with owners - low level of customer care The core research question in these workshops was, “Why is the level of involvement of Slovak SME low in international marketing?” Besides the characteristics shown in table 1, the research has also shown following shortcomings in approaches to marketing by SME in Slovakia: · the entry into the international market is prepared intuitively, without exact market research, without calculations of the potential of the targeted market, based on insufficient information about the market; SME understand the market as undifferentiated, they cannot closely identify the customer, the target segment, its amount, and its purchasing behaviour. On the target market, they do not know the real and potential competition as well as the substitution offer. The competition is analysed only superficially, they are not aware of their real ability to compete. · their goals are focused more on the means and less on the outputs; The goals are not clearly defined, without any substance or a time frame, they are not concrete. · the basic strategy of the entry is direct export; They do not have a clearly defined strategy. Companies, to a high extent, form long term ties with one customer, who they usually get by the interest of the customer to receive a shipment of their goods. Companies are satisfied by having one customer, which increase the risk of their position on the international market. · Slovak suppliers do not have sufficient information about his/her foreign partner; On the other hand, the foreign partner often “studies” the Slovak supplier more thoroughly · the basic factor of the offer is the price of the product (usually lower, based on the expenses of the company); SME do not spend time to create a special market position for their offer, their do not take into consideration the differences in products, they are unable to identify their USP, and they base their offer on a higher added value. They fall back on high quality of their production, which is neither declared by criteria, nor does it have a certificate, and it is without an analysis of a competing product. They cannot identify why, in what, and in comparison with what does their product have a higher quality. · lack of confidence in their abilities and competencies is spread among SME; Their identification of strengths and weaknesses is neither based on a real analysis of criteria nor do they compare themselves with competition. They are unable to systematically identify threats and possibilities on the market. They cannot guess the trends of the market development in relation to their offer and to a served segment. · planning of business and marketing activities does not have a formal expression, the planning process is underestimated; They do not take advantage of economic and financial modelling, they do not work with alternatives, and are unable to identify business risks connected with the entry on the international market. · SME also lag in organisation and managing of their marketing activities. This situation is also caused by insufficient preparedness of the human potential and by execution of marketing activities. Only on a small scale, do they try to predict the needs of customers, the marketing lacks creativity and a conceptual approach to the market. From the researched sample, 10-15% of managers, or people working in the marketing sphere, meet the difficult criteria used. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 85 One of the deciding factors, which could help to cope with the above mentioned shortcomings, which can help use their strengths, is education of people working in management positions in companies, people who work professionally in marketing, and people who specifically work in the sphere of international business. This education must entail theoretical preparation and acquisition of practical experiences by working on various world markets. The research also revealed a considerable pro-export potential of Slovak managers in SME. Among the deciding criteria of their success are: · managers show great effort in applying new information in practice, they show their interest in marketing training, they want to change the present state of marketing in their company; To a positive stimuli from around, they react proactively with the system task=challenge, they show high personal involvement. Personal attendance of owners or managers of SME in the workshops gives guarantees that these represent guarantors and the moving powers of changing process. · Attendees of the workshops showed high interest in information and information resources dealing with international market. More than 30% of participants had real-life experiences from working in the sphere of international marketing. These experiences guide them to the effort to establish their own information marketing system about the international market. · Export and other forms of entry into the international market are understood by them as an important factor for stable profits means of their company. In this field, there exist real cash flows for deliveries of goods and services. 3. OTHER FACTORS The process of integration of Slovakia (and not only of Slovakia) and Slovak companies influences a wide range of social and cultural factors. One can positively see, in this trend, the growth in the number of cosmopolitan consumers, who are dominated by younger consumers and representatives of middle and upperclass. On the other hand, the situation in the country is that the state policy, partially, stimulates the growth of so called, “nationally oriented consumers”- who prefer consumption of products made in Slovakia. To the professionals in marketing, it is known that globalisation trends in consumption strongly influence habits of consumption as well. Habits of consumption are in Eastern Europe (in comparison with Western Europe) concerning some needs and products a lot stronger. The consumer trend lays the premise for companies operating with import from foreign markets. Another important factor determining involvement of Slovak SME in international marketing is the actual self-perception. This reflects a perception of managers as well as that of consumers. With managers, this perception formulates their abilities and steps to be taken on the market where with consumers this, to a high degree, formulates their consumption of a large group of products. The position of Slovakia, within the scope of Eastern Europe based on self-perception, is shown in research conducted by the Fesel and GFK agency, which is indicated in table number 2 (the countries are identified according to international marking of motor vehicles). The respondents of the represented countries evaluated criteria chosen by the agency. These also influence the speed and the level of acceptance of European and global trends of consumption and management style by SME managers. A management style forms one of the alternative decision-makings about the entry of SME into the international market. These criteria have a strong influence, not only on consumption, but they also influence the level of how a company establishes itself on the international market and the domestic market in comparison with entering worldwide companies. Numbers of these above-mentioned characteristics directly reflect managing style. Table No. 2 Ratio of self perception (rating form 0 to 100) Country Self perception as: H CZ SK PL - moderní (modern) 36 36 32 41 - bystrí (clever) 47 34 31 34 - búrliví (loud) 39 44 56 52 - silní (masculine) 35 24 13 27 - seriózni (serious) 48 28 28 37 - mierumilovní (peace) 70 18 38 14 BG 32 51 47 39 36 61 SU 29 30 62 52 27 49 RO 35 58 25 29 72 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 86 - vzdelaní (educated) - cieľavedomí (goal-or.) - dôslední (pedantic) - úspešní (successful) - elegantní (elegant) - pesimistickí (pessimistic) - pomalí (slow) - spoločenskí (sociable) - tolerantní (tolerant) - konzervatívni(conservative) - súcitní (appealing) 55 50 28 28 19 41 48 66 50 22 68 59 40 10 19 24 24 42 35 36 28 63 59 44 20 23 16 23 41 47 36 25 64 37 52 11 41 12 25 31 63 36 20 53 61 33 27 21 18 25 46 59 41 28 44 46 16 14 7 8 17 40 62 68 32 46 53 36 14 22 22 13 36 73 65 26 50 Note: Counties are identified according international car codes Resource: According “Mirror“, Central European Economic Review, The Wall Street Journal Europe´s, December 1997 – January 1998, Vol. V. No. 10, p. 7. According to the information in table number 2, Slovaks characterise themselves as educated, energetic, and emphatic, but more intolerant and weak. The opposite are Russians, who feel rough, social and tolerant, but don’t feel modern, successful and think that they have little elegance. Romanians consider themselves pacifistic and with temperament. Poles feel modern and successful, but inconsistent, with little elegance and a bit little pugnacious. Currently, globalisation affects small and middle-size enterprises in Eastern, Western Europe, and around the world. Globalisation forms a premise for a new approach to product development, communication, and distribution. Globalisation penetrates the Slovak market via global products, it has a positive influence on Slovak companies when standardizing products, packaging, in services, in support as well as in distribution. New distribution channels (the Internet) of some products will allow the aversion of protectionist measures in a given country. Companies must formulate their strategy in consideration to their working on the market, where big and strong international companies are established. Both defensive and offensive measures should have a wider range of application in Slovak companies. Based on their ability, it is apparent that they should develop a strategy of taking closely defined segments. Globalisation and European integration, which lead to Euromarketing, suppress the importance of state borders. The above-mentioned trends lead to a higher product quality and its standardization. Within the trends of globalisation, Slovak companies must consider trends that are specific for the European market. Among the core trends, which SME must count with, in their marketing strategies and with other marketing activities, are following: · development of a common European market, taking down market barriers, a free flow of goods, of services, investment, labour force, and application of a common currency · the European Union will remain, for years, economically, legally, and culturally heterogeneous, in individual regions different levels will fulfil the needs · a change in the competition environment in Europe and outside will have a strong impact on Slovak companies as well; The competition will be stronger and a system of a so-called exclusive representation will be limited. · a state will restrain its measures against competition, with the help of which it tried to regulate or protect the domestic market · development of the market will bring establishment of new segments and present segments will change their behaviour This all will have a very strong impact on the marketing of Slovak small and middle-size enterprises. The changes will be visible more on these, than on large companies and multinational corporations. It will have a direct impact on the offer formation by SME quality production, price, distribution methods, marketing communication, protection of the consumer, and other fields of activities. The bottom line is that it will not only determine the economic outcomes of SME, but it will also determine their ability to survive on the market in the upcoming years. Essentially, this effort to increase the quality of marketing activities must be visible in the strengthening of the marketing orientation of Slovak SME and in the strengthening of their ability to compete on the market. A crucial factor of success will be a change in their work style and in education of top managers and marketing professionals. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 87 LITERATURE: [1] „Mirror“, Central European Economic Review, The Wall Street Journal Europe´s, December 1997 – January 1998, Vol. V. No. 10, p. 7. [2] JAWORSKI, B. J. – KOHLI, A. K. 1993. Market Orientation: Antecedents and Consequences, Journal of Marketing, Vol. 57, July, pp. 53-71. [3] KRIŽANOVA, A. 1995. Význam marketingového riadenia v podnikoch cestnej dopravy, In: Zborník prác F PEDaS č. 3, Žilina: ES VŠDS, str. 49-55. [4] SLATER, S. F. – NARVER, J. C. 1994. Does Competitive Environment Moderate the Market OrientationPerformance Relationship?, Journal of Marketing, Vol. 58, January, pp. 46-56. [5] URAMOVA, M. 2002. Slovak Enterprises and their Competitivnes in the V4 Conutries Area. In: Geopolitical Importance of Central Europe (V4) and its Prospects. Banská Bystrica: Univerzita Mateja Bela, Ekonomická fakulta. 2002. s. 267-273. ISBN 80-8055-732-2. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 88 INTERCULTURAL DIFFERENCES BETWEEN SLOVAK AND EU COUNTRIES Vladimír Laššák, Jaroslav Ďaďo Faculty of Economics, University of Matej Bel, Banská Bystrica, Slovakia Tel. 00421-48-4152786, e-mail: [email protected]; [email protected] Abstract:Intercultural management as one of the youngest disciplines of management starts its development after falling down trade barriers in international trade in the last century. The globally operating businesses are based on free goods, people, and the know-how and capital movements where the operations of companies are developed without respecting countries or continents borders. Studies of different culture interactions and relationship of people in international management became highly important. This article describes cultural differences between Slovak and EU entrepreneurs in selected areas of business life as we recognise it in our empirical research. Keywords: intercultural differences, European Union, INTRODUCTION It is well known that ability to adapt in foreign country is in direct relations to degree of cultural knowledge of foreigners. People feel cultural differences strongly when having a very short visit in foreign country, e.g. on holiday or business trip, but everybody is able overcome its because both visitor and also host people tolerate many partner’s habits and behaviour. To know smallest and the most specific differences in other country culture is sort of spice belonging to passion of travellers. In situation of long-term working relations between employees from different cultures ability to tolerate many habits and behaviour should not be so evident. Everyday interaction of people, which has been breeding from its pupil age in different cultures and different value systems, can change the workplace on fireplace with day-to-day latent conflicts. Doing job in different cultural management is much more demanding and challenging because relations and procedures running in “my home country” without any problems are going here by totally different way. Simply say, my action (behaviour or requirements) is perceives and evaluates differently by my foreign partners and I face usually unexpected response and reaction than that I had usually got in my home country business environment. Working in multicultural environment requires integration of many elements and requirements from foreign culture into my culture. Doing this does not mean adaptation to every rule and requirements of other culture – because one can lose his or her identity and spontaneous in his or her behaviour. A very frequent mistake the cultural unskilled managers have done is judging and qualifying behaviour of domestic staff or employee according criteria which are typical for manager’s culture environment. This means that everything what employee has done was evaluated and judged in relation to cultural criteria of manager’s country. For example, behaviour which most people form EU see as “no problem” (to come a little bit later, just twenty minutes, on meeting) is for them common and “o.k.”, but for my own (e.g. Slovak) culture is hardly acceptable. 1. EU MAN COMES TO SLOVAKIA If somebody from EU come to visit Slovakia he or her probably hear that Slovakia is a country in Central Europe and Slovak people are in principle same as EU people without any serious differences. “Well, then I shouldn’t be afraid of being there and my adaptation will be without problem”, thing EU people. This truth is only relative, until we compare mentality and behaviour of Slovak people to the other West Europe nations. So, this good advice can be also wrongful and having it results that foreigner will not be sensitive enough to situation in new environment. He or her probably forgets that “don’t be different” cannot mean to have similar attitudes and to be sovereign in every situation. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 89 UE entrepreneurs having analysed current daily situations from Slovak environment according France cultural criteria is not able interpreted it correctly because naturally using of different decoding system of this event. Being able to better understand of these cultural decoding systems we have discussed some cultural difference in attitudes and behaviour of EU and Slovak people. We focus primarily on two types of cultural issues: 1. differences in everyday life, 2. differences in working environment, refer to: 3. differences in business trade or sales negotiation, 4. differences in management situation. 2. DIFFERENCES IN EVERYDAY LIFE There are many cultural differences concerning everyday life situation of average EU man and Slovak. We concern in our research only to some of them. Fields describe below are basic artefacts through we can identify specific culture. Table 1 Cultural differences in everyday life EU and Slovak in EU in Slovakia Greetings Greeting in EU “Hello, how are you?” has a function to get first contact with others. Most EU people don’t expect real answer. The obvious answer to this greeting is “Hello, I´ m well thank you.” Perception of time and meeting time To be late for fifteen or twenty minutes is very common. Everybody tolerate this. When being inviting to family lunch or dinner it is even expected to be a little bit late. Meetings are often set to late evening hours e.g. 19:00 but not in the lunch time 12:00- 14:00 Meal The meal consists from starters, main meal, cake or cheese. The bottle with water is always on the table. EU man are happy in situation spending with others sitting around the table. Conversation topic concern usually gastronomy (food, vine). Spirits Spirits (like cognac, vodka etc.) are not offered as aperitif when starting, but usually at the end of meeting. Working hours Employee works for that he wants to live. Average EU man 35 hours-working time. In other countries it is 40 or 48 hours. Acceptance of foreigners in family It differ according time of arrive. The meal is offered only when coming in time of lunch or dinner. Otherwise are offered some soft drinks or coffee. Body contact When meeting friends they often symbolically touch each other on face. Greetings Greeting in Slovakia has different meaning. First contact should be an opportunity to say something about myself and share real personal information. The answer is very often like “My children are ill again” or “My car battery is go wrong this morning because tough winter, etc.” Perception of time and meeting time The agreed meeting time is expected to keep and this is also expected in private and business situations. Meetings are arranged usually before or after lunchtime, e.g. 13:00 or early in morning (7:30) exceptional in the late evening. Meetings in the evening could be seen as intervention into private live. Meal The lunch starts with soup in every season of year. Drinks are offered after the meal. Time spend by eating is much shorter. The food is very often much important then conversation. Discussion became much more intensive after having lunch or dinner. Spirits Spirits (like cognac, vodka or brandy) are offered like aperitif and also in any time of day if there is any appropriate opportunity (in official business meeting or in family). Working hours For middle or lower social classes – to have a job is necessary for survival. To have two or more jobs is quite common (in spite of 18% of average unemployment rate). Acceptance of foreigners in family It does not matter on the time when visitor comes. Always are offered meals and drinks (including spirits). Welcoming a visitor starts with meal offer. The offered meal is a symbol of hospitality. Body contact Slovaks touch others when shaking hands, touching arm or shoulder of visitor as a symbol of friendship but can be seen as too tight. 3. WORKING ENVIRONMENT DIFFERENCES To be able understand different business culture situations it is necessary to study misunderstandings and conflicts because through them we can easily understand to motives which lead to this accidents. The conflict, open or close, extrinsic or intrinsic is a part of everyday working live. People do not feel good in conflict situations and for that they try to keep out of them. By doing this they try to create friendly working environment. But essential part when managing personal conflict is respect to other side or partner. Generally INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 90 say, we cannot be successful in global business without respect differences in multicultural environment. This means to respect different style of life, thinking and different habits, behaviour, procedures and decisions made in business situations. In trade the better understanding of partner and colleagues enable create positive and effective working environment. For being successful in negotiations it isn’t just enough to present professional competence and sense for negotiations. The co-operation, listening, understanding a knowing habits of others is essential. This part of the article has two sub-parts. In the first one we have focused on negotiation affairs, in second one we will discuss differences relating to employees in the EU and Slovak working situation. Table 2 Cultural differences in trade situations and business life of the EU and Slovak in EU in Slovakia Business negotiations Agreement After signing agreement obligatory fulfilment is expected. Each side is obliged to fulfil what was promised. Terms Terms are firm and obligatory set and are very important. Planning Planning is done systematic step by step in each level of management. Organisation of meetings Seeming meeting a little bit jamming maybe chaotic doesn’t affect opinion about partner. In spite of the fact that program of meeting is structured it is let free for everybody to keep it. Legislation and relations To fulfil goals and done results have higher importance than good relations between partners. Laws win and are above personal interest of people. Negotiation process Partners may be seen as jumping from one topic to another and talking around using nice words. Negotiation is opportunity to exchange opinions and there is no problem to change topic of conversation, to interrupt of partner - “France negotiations”. Decision making The decision making is made after analysis and checking all information and knowing all possible circumstances. Way of thinking Strong sense for synthesis. Short and long term anticipation make no problem. Creativity Creativity is developed from pupil age at basic schools and has its own tradition. Schools are oriented to delivering know-how (knowledge, skills and attitude). Respecting rules French people are rebels. They struggle for changing laws (strikes, manifestations, etc.) Customers of services are very often hostages. Relations to boss Employees use to say very open what they think about their boss. Agreement After signing agreement obligatory fulfilment is expected but not always entirely. Agreement is perceive as guideline but not exact setting. Terms Terms are symbolic part of any plan. Planning Planning is also done, but people do not take real care when plan. Planning is very often just symbolic part of management. Organisation of meetings The way the meeting is organised can reflect how the company done its job in organisation of things. It is expected to strictly respect agreed program of meeting. Legislation and relations Very good relationships between partners are much more important than premeditate strategy of negotiation. Personal relations are much more important that legislation or rules. Negotiation process Negotiation process should be different. Partners can jump from one topic to another without sense. Skilled businessman is usually very structured and goal oriented in the negotiation. Decision making The decision making is intuitive in many cases. It is quite common don’t have analysis and precise information especially about markets. Way of thinking People think more analytically than synthetically. Difficulties are in general imagination (concept view) and short or long term anticipations. Creativity People are very creative what can be seen in many private situations, not in jobs, or in parallel economy. Schools are oriented to delivering of accumulated knowledge. Expected is exact reproduction of teacher words or book text. Respecting rules Slovak people don’t try change bad laws, they try go out of them. In evading the law and agreed rules creativity of Slovak people is without borders. Relations to boss Before velvet revolution 1989 employees was afraid to say own opinion. They obey bosses. After 1989 there is still distress. Managers are not able accept employees opinions and they are not able utilise employees skills and competencies. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 91 Table 3 Cultural differences concerning employee EU countries Business lunch or dinner Business lunch or dinner is usually realised before having official meeting. Time spend when having meal is use for assessment of partners. Managers Managers are accepted and have respect if they are competent and have appropriate results. Also freshly graduated can be set as managers. Diploma is seen as a symbol of being competent. Goals of communication Goals are declared very directly and clear. Confidence People are confident if fulfil what promise. Degree Academic degrees (like Dr., Dipl.ing.) are not important in private and business relations. Agreeing If you agree always with your boss it isn’t positively considering. Slovakia Business lunch or dinner Business lunch or dinner is usually realised after having official meeting. Time spend when having meal is use for short quick discussion. Managers Managers are usually older. Principles of seniority are very strong. But in the last years there are some positive changes especially in foreign subsidiaries. Goals of communication Goals are declared indirectly very often. Confidence People are confident if they are loyal to authority. Degree Academic degrees (like Dr., Dipl.ing.) are very important is society but in companies using it had declining tendency. Agreeing To agree with boss means that you are under his/her umbrella save and secure. CONCLUSION The goal of the topic was description of selected cultural differences between Slovak and EU partners in common business and working situation. We have studied main historical, political, social, cultural, economical, etc. reasons to understand factors and forces, which formed certain typical business culture. We provided comparison of behaviour, habits and attitudes of managers in both countries. Our recognition was structured into three tables concerning everyday live situation, negotiations and typical working situation. We believe that studying and thinking about this problems can lead to better knowing and understanding of cultural differences and improving relations in live and business of both countries. Recommendation from this study can be describe as follows: · For being successful in multicultural business it is necessary to be open, perceive and tolerant to the other country culture. · When need to acquire and involve foreign partners to successful co-operation it is necessary to understand which differences are significant in each culture and how can be interpreted by both. · Features significantly different from cultural the point of view (strategic or managerial issues) can be on the other side surprisingly reasons for co-operation in business situations (business issues) because looking for new stimulus. LITERATURE: [1] URAMOVÁ, M. 2001. Podnik, štát a konkurenčné prostredie. In: Ekonomika firiem 2001. Zborník z medzinárodnej vedeckej konferencie. Košice: Ekonomická univerzita v Bratislave, Podnikovohospodárska fakulta, 2001. S. 759 -764, ISBN 80-225-1446-2. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 92 FUZZY LOGIC AND FINANCIAL TIME SERIES Pavel Dostál1), Ladislav Žák2) 1) 2) Strakatého 15, 636 00 Brno, Czech Republic, E-mail: [email protected], http://www.iqnet.cz/dostal/ Phone: ++420 5 44211639, Fax: ++420 5 44234750 Department of Stochastic and Non-Standard Methods, Institute of Mathematics, Faculty of Mechanical Engineering, Technical University of Brno,Technická 2, 616 69 Brno, Czech Republic, Email: [email protected], Tel: ++420-54114 2550, Fax: ++420-54114 2527 Abstract: The article presents the possible prediction of time series by means of fuzzy logic. The prediction errors MAPE and the estimation of trends serve as the evaluation of used method. The obtained results show possible applicability to prediction of financial time series and their information efficiency during the process of decision making on the stock market. Keywords: Time series, finance, prediction, fuzzy logic 1. INTRODUCTION There are manifold methods used for the predictions of time series. In addition to the classical method such as Box-Jenkins methodology, Kalman filter, etc., we can use artificial neural networks, genetic algorithms or fuzzy logic. See articles 1,2,3,4,5,6,7. This article describes the search for the use of fuzzy logic for the purposes of prediction of financial time series. We used the Fuzzy Inference System (FIS) of Sugeno type for the prediction of time series. The FIS Sugeno is designed in such a way to give one output with n inputs. The inputs of FIS are represented by the values of time series preceding the value that we want to predict. The output value is prediction value. The longterm testing shows that the quality of prediction is given by the set up of the number of members of time series which creates the input of FIS. The FIS does not describe the course of time series when the number of members is either low or high (the low or high sensitivity of regulator). The FIS has been tuned to the initial part of time series to set up the number of input values of time series. The cluster method has been used for the purposes of assigning members of time series to clusters. The number of clusters define the number of input linguistics value of input linguistics variable. We choose the output of FIS Sugemo in the form of linear dependence. We found the constants for set up of the FIS suitable for the searched time series by the optimization of Sugeno methods above preceding members of time series. The design and tuning of FIS Sugeno has been made in the program Matlab – FuzzyTolbox. The quality of prediction of heat consumption has been calculated according to match of prediction of tendency of prediction of development of time series with reality and according to average error MAPE defined æ MAPE = 1 ç L Lèå i =1 ö (abs( Pi - R i ) / R i ) ÷ , ø where R1, R2, …, RL are real values of time series and P1, P2, …, PL are the predicted members of time series where L is the number of predicted values. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 93 2. EXAMPLE I. TIME SERIES OF NASDAQ INDEX The first case which has been tested was the time series of development of Nasdaq index negotiable on the New York stock exchange. The values were used in the interval from 09:30 13.06.2003 to 16:00 01.07.2003 within the following two days (10 minutes sampling, 520+40 values). The graph of time series is on the Fig.1. NASDAQ-History 1690 1680 1670 1660 1650 1640 1630 1620 1610 1600 1 26 51 76 101 126 151 176 201 226 251 276 301 326 351 376 401 426 451 476 501 Fig.1. Nasdaq index – history The suitable FIS Sugeno is represented by 3 input variables and 9 input values for the purposes of prediction of Nasdaq index time series. See fig. 2, 3. FIS surface Structure 1680 in1 (9) 1660 (sugeno) in2 (9) f (u) out1 NASDAQ 9 rules 1640 1620 1600 out1 (9) 1680 1660 1640 1620 1600 in3 in3 (9) 1640 1620 1600 in2 1660 1680 System NASDAQ: 3 inputs, 1 outputs, 9 rules Fig. 2. FIS Sugeno of Nasdaq index Fig. 3. The part of FIS surface The evaluation of match of tendency of prediction with reality gives the correct prediction in 62.5 %. The fig.4. shows the values +1 (-1), which present increasing (decreasing) real tendency and predicted tendency. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 94 Nasdaq-Prediction 1 0 -1 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 Fig. 4. Nasdaq index - prediction The prediction by means of fuzzy logic made the value of MAPE = 0.0013. The Fig.5. shows the real values (solid line) and values of prediction (dashed line) of time series. NASDAQ-Prediction 1680 1675 1670 1665 1660 1655 1650 1645 1640 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 Fig.5. Nasdaq index – prediction 3. EXAMPLE II. - TIME SERIES OF QQQ INDEX The second case which has been tested was the time series of development of QQQ index negotiable on the New York stock exchange. The values were used in the interval from 09:30 13.06.2003 to 16:00 01.07.2003 within the following two days (10 minutes sampling, 520+40 values). The graph of time series is on the Fig.6. QQQ-History 31.5 31.0 30.5 30.0 29.5 29.0 1 26 51 76 101 126 151 176 201 226 251 276 301 326 351 376 401 426 451 476 501 Obr.6. QQQ index - history The suitable FIS Sugeno is represented by 2 input variables and 8 input values for the purposes of prediction of QQQ index time series. See fig. 7, 8. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 95 FIS surface Structure 31 (sugeno) out1 QQQ in1 (8) f (u) 8 rules 30.5 30 29.5 out1 (8) 31 30.5 in2 (8) 30 in2 29.5 29.5 30 in1 30.5 31 System QQQ: 2 inputs, 1 outputs, 8 rules Fig. 7. FIS Sugeno of QQQ index Fig. 8. The part of FIS surface The evaluation of match of tendency of prediction with reality gives the correct prediction in 65.0 %. The fig.9. shows the values +1 (-1), which present increasing (decreasing) real tendency and predicted tendency. QQQ-Prediction 1 0 -1 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 Fig. 9. QQQ index - prediction The prediction by means of fuzzy logic made the value of MAPE = 0.0016. The Fig.10. shows the real values (solid line) and values of prediction (dashed line) of time series. QQQ-Predikce 31.0 30.9 30.8 30.7 30.6 30.5 30.4 30.3 30.2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 Fig. 10. Index QQQ – prediction INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 96 4. CONCLUSION The paper describes the design of the calculation method of prediction of financial time series by means of fuzzy logic. Other methods such as Box-Jenkins methodology, Kalman filter, Elliot’s waves, artificial neural networks, genetics algorithms, etc., (see 1,2,3,4,5,6,7) rank the use of fuzzy logic for prediction of financial time series to the most successful ones. The prediction error MAPE and the prediction of tendency of development of time series give evidence of usable prediction method for the process of decision making on the capital markets. LITERATURE: [1] DOSTÁL, P.: Neural Networks Decision Making and Stock Market Lisabon 2000, ISF2000, Konference, 5 s. [2] DOSTÁL, P.: Neural Networks Decision Making and Stock Market Lisabon 2000, ISF2000, Konference, 5 s. [3] DOSTÁL, P.: Neural Networks and Shares, Zlín 2000, Nostradamus 00, Konference, 22-27 s, ISBN 80-2141668-8. [4] DOSTÁL, P.: Stock Market and Financial Cybernetics, Zlín 2001, Nostradamus 01, Konference, 7 s, ISBN 80-7318-030-8. [5] DOSTÁL, P., Zmeškal O.: Chaos and Stock Market, Zlín 2001, Nostradamus 01, Konference, 8 s, ISBN 807318-030-8. [6] DOSTÁL, P.: Genetics and Shares Brno 2002, Small and Medium Firm Management with Computer Support, Konference, s.11, ISBN 80-86510-56-5 [7] DOSTÁL P.: Soft Computing and Stock Market Brno 2003, Mendel 03, Konference, s.258-262, ISBN 80214-2411-7 [8] ŽÁK, L.: Odhad vlivu počasí na odběr elektrické energie pomocí fuzzy regulátoru. . CR – Praha: Automatizace č.5-6. 2002. 5p. ISSN 0005-125X. [9] ŽÁK, L., Riaz M. K., Předpověď spotřeby elektrické energie s využitím fuzzy regulátoru. Acta Mechanica Slovaca 3/2001, pp 513-522, ISSN 1335-2393. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 97 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 98 DETERMINATION OF TAX SHIELD AND ITS INFLUENCE IN FINANCIAL DECISIONS # Jarmila Radova1), Petr Marek2), Tibor Hlačina3) 1) Department of Banking and Insurance, Faculty of Finance and Accounting, University of Economics Prague, nam. W. Churchilla 4, 130 67 Prague, Czech Republic Tel.: +420-2-24095161, Fax: +420-2-24095135, E-mail: [email protected] 2) Department of Corporate Finance and Valuation, Faculty of Finance and Accounting, University of Economics Prague, nam. W. Churchilla 4, 130 67 Prague, Czech Republic Tel.: +420-2-24095147, Fax: +420-2-24095135, E-mail: [email protected] 3) European polytechnical institute, Kunovice, Osvobození 699, Czech Republic Tel. 572 548 035, Fax: 572 549 018, E-mail: [email protected] Abstract:We aim at seeking of the real moment of tax shield realization and investigate the influence the time span between the moment of tax shield claim origination and the real moment of tax shield realization has on the present value of tax shield. The real moment of tax shield realization was recognized at the moment the tax shield is reflected in the taxpayer’s cash flow, in the form of income tax reconciliation (differences), and paid tax deposits. Subsequently, we have investigated the dependence of tax shield present value on the discount rate, and on the size of a time span between the moment of tax-shield claim origination and tax-period’s end. The largest difference between elementary value of the tax shield and its present value was founded at the maximum discount rate considered in our model, and at the maximum size of the time span between the moment of tax-shield claim origination and tax-period’s end. The dependence was, however, not always found as linear. We conclude suggesting modifications of financial models that employ tax shields by coefficients reflecting the real moment of tax shield realization. Key Words: Tax Shield, Financial Model, Time Value of Money JEL Classification: G3, H2 Taxes are not merely a financial managers’ „nightmare“, but principally they represent an important factor influencing most of financial decisions, primarily through tax wedges and shields. Whereas tax wedge diminishes investor’s return, tax shields, on the other hand, mitigate impact of certain decisions on financial performance. Tax shield can be regarded as a sum of money the tax liability is reduced by, due to occurrence of a transaction that decreases company’s tax base, or entitles to tax relief. The transaction can be understood as posting of tax deductible costs and expenses in taxpayer’s accounts, or occurrence of a different matter resulting in tax liability alteration, e.g. acquisition of fixed assets. Mainstream finance has recently turned away from highly abstract theories in favour of detailed models of actual functioning of real world. And in this world, tax shields play an irreplaceable part – they alter capital budgeting decisions, influence capital structure, etc. All classic models of finance1 assume that the real moment of tax shield realization is identical with incurrence of the particular tax-deductible expense, or other taxliability-affecting transaction. However, this does not correspond to the real world. In our contribution, we aim to elaborate the tax shield issue in the suggested direction, under conditions of the Czech Republic tax law, and # 1 This article has been elaborated as one of the outcomes of research project registered with the Grant Agency of the Czech Republic under reg. no. 402/03/1307 and 402/04/1313. For review of financial models development see e.g.. Copeland, T. E. - Weston, J. F.: Financial Theory and Corporate Finance. Addison Wesley, Reading 1988. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 99 subsequently apply learned findings through adjustments of financial models. In our discussion, we shall seek the answers to the following questions: 1. What is the real moment of tax shield realization? 2. What influence shall the real tax shield realization moment have on tax shield present value? Question no.1: What is the real moment of tax shield realization? Cash flow based financial models must logically reflect the real tax shield realization in company’s cash flows at the time of real execution of a relevant tax payment into state budget. For illustration, a tax-deductible payment2 incurred anytime in 2002 shall not be reflected in tax payments into state budget until the 2002 income tax is settled, i.e. by 31st March of the next consecutive year (2003 in our example), or by 30th June 2003 in case of taxpayers liable to annual financial statements audit, or their tax assessment is elaborated and presented by a tax advisors.3 Decreased tax base shall subsequently affect the amount of income tax deposits paid in the next consecutive tax deposit period. Consequently, the difference between tax deposits paid and real income-tax liability for the relevant period shall equal the amount the tax deposits payments were reduced by. Illustrative example:4 1. tax-deductible payment at amount of 1,000 (payment A) 9.2002 30. tax balance reconciliation date – difference paid for 2002 is decreased by the tax 6.2003 shield, i.e. 1,000 × 0.31 = 310 (payment B) 15. income tax deposit payment decreased by ¼ of tax shield originated in 2002, i.e. 9.2003 310 / 4 = 77.5 (payment C) 15. income tax deposit payment decreased by ¼ of tax shield originated in 2002, i.e. 12.2003 310 / 4 = 77.5 (payment D) 15. income tax deposit payment decreased by ¼ of tax shield originated in 2002, i.e. 3.2004 310 / 4 = 77.5 (payment E) 15. income tax deposit payment decreased by ¼ of tax shield originated in 2002, i.e. 6.2004 310 / 4 = 77.5 (payment F) 30. tax balance reconciliation date – difference paid for 2003 is increased by the equal 6.2004 amount the tax deposits paid in 2003 were reduced by, i.e. 77.5 × 2 = 155 (payment G) 30. tax balance reconciliation date – difference paid for 2004 is increased by the equal 6.2005 amount the tax deposits paid in 2004 were reduced by, i.e. 77.5 × 2 = 155 (payment H) Question no.2: What influence shall the real tax shield realization moment have on present value of tax shield? We shall primarily concentrate on the present value of tax shield calculation methodology. Calculation procedure can be viewed in the following illustration depicted in Table no. 1.5 2 3 4 5 For the sake of simplicity we assume the tax shield claim originates at the same time as realization of this payment, or execution of this payment shall sooner or later lead to origination of the claim, for example, based on incurrence of tax-deductible costs and expenses. Sec. 40, par. 3 of Tax administration Act no.337/1992 Coll. We have left out of considerations financial years not corresponding to a calendar year. Illustrative example assumes that a) investor uses the services of a tax advisor and exploit the opportunity of tax assessment submitting by end of June of the next consecutive year, b) tax period is calendar year, c) income tax rate is equal 31%, d) last known tax liability of the taxpayer exceeded 150 000 Kč. Here we shall add two additional assumptions: interest compounded on monthly basis and fixed monthly discount rate of 1% (i.e. annual discount rate of 12%). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 100 1. 9.2002 30. 6.2003 15. 9.2003 15.12.2003 15. 3.2004 15. 6.2004 30. 6.2004 30. 6.2005 Total A B C D E F G H X 0 10 12 15 18 21 22 34 X Present value of tax shield Tax Shield Original Payment Number of months Date Payment Table no.1: Present Value of Tax Shield – Illustrative Example -1000.0 -1000.0 +310/(1+0.01)10=259.17 +77.5/(1+0.01/2)/(1+0.01)12=62.89 +77.5/(1+0.01/2)/(1+0.01)15=61.04 +77.5/(1+0.01/2)/(1+0.01)18=59.24 +77.5/(1+0.01/2)/(1+0.01)21=57.50 -155/(1+0.01)22=-115.00 -155/(1+0.01)34=-102.05 +282.77 +310.00 +77.50 +77.50 +77.50 +77.50 -155.00 -155.00 +310.00 As the basic value of tax shield, there is understood the total (but not discounted) cash amount, we can decrease tax liability by. In case of tax-deductible payment, or other transactions entitling the company to claim income tax base reduction, it is calculated as a product of income tax-base reducing item value and income tax rate: TS = TBRI ´ t , where TS TBRI t = = = basic value of tax shield, income tax-base reducing item, income tax rate. In case the tax liability decrease is claimed on the basis of entitled income-tax relief, the basic value of tax shield equals the relief value: TS = relief , wher relief = value of income-tax relief. e Now, we shall approach the calculation of the present value of tax shield. As independent variables, we shall regard monthly discount rate [i], and number of months from the moment of tax shield claim origination till the end of tax period [n]. Other parameters reflect the model’s assumptions suggested in the illustrative example. ì 1 1 æ öü + ç ÷ï n +8 n +11 ï (1 + im ) (1 + im ) 1 1 ç ÷ï ï + ´ ÷ï 1 1 ï (1 + im )n +6 4 ´ (1 + im ) ç + + ç ï 2 è (1 + im ) n +14 (1 + im )n +17 ÷ø ïï ï PV (TS ) = TS ´ í ý 1 æ ö ï ï ç ÷ ï 1 ç (1 + im )n +18 ÷ ï ï- ´ ç ï ÷ 1 ï 2 ç+ ï n + 30 ÷ ïî ïþ ( 1 i ) + m è ø INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 101 wher PV(TS) = present value of tax shield, e n = number of months from the moment of tax shield claim origination till the end of tax period, im = monthly discount rate, = i / 12, i = annual discount rate. The issue of discount rate determination is not addressed in our article. In fact, the precise determination it is not even possible as far as theoretical approaches are concerned, because the determination of discount rate reflects the purpose of financial model, and vary with each individual investor reflecting their expectations. However, one rule shall be maintained, so that cash flow after tax shall be discounted by a taxadjusted discount rate. Financial models reflecting tax shields do employ cash flow after tax. Consequently, we shall draw our attention to investigation of dependence of [PV(TS)] on discount rate. Graph no.1 depicts dependence of [PV(TS)] on discount rate. Individual plotted curves represent the development of present value of tax shield for selected number of months from the moment of tax shield claim origination till the end of tax period, i.e. in our case till the end of calendar year. Graph no.1: Dependence of present value of tax shield on discount rate 0,00 330,00 0,04 0,08 0,12 0,16 Annual Discount Rate 0,20 0,24 0,28 0 1 320,00 2 3 310,00 4 Tax Shield 300,00 5 6 290,00 7 8 280,00 9 10 270,00 11 12 260,00 250,00 Curve [12] demonstrates the dependence of [PV(TS)] on discount rate in situation when the tax shield claim originates on 1st January of respected year, i.e. 12 months prior to the tax-period’s end. Similarly, e.g. curve [0] represents the researched dependence in case when the tax shield claim originates at the year (thus, tax period) end. All curves plotted can be characterized by a quadratic equation in common form: Cx 2 + 2 Dy + 2 Ex + F = 0 . One must be aware that coefficient [C] is the lower the longer the time span is between the moment of tax shield claim origination and tax period’s end. Therefore, we can state that in case the tax shield claim has originated at the time of the tax period's beginning, the curve depicting the tax shield present value [PV(TS)] dependence on discount rate can be approximated by a linear function. Such situation is demonstrated by curve [12]. In addition, we also state that a curve extreme (maximum) can be found for each curve representing a INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 102 different time span remaining to tax period's end and characterized by parabolic function, with the maximum located at individual parabolas’ vertexes. Such maximum represents the largest [PV(TS)], and each of the discussed functions are maximized at a different level of discount rate. The closer the tax-shield claim origination is to the tax-period’s end, the higher the parabola-maximizing discount rate is. The dependence of [PV(TS)] on the time span between the moment of tax-shield claim origination and tax-period’s end represents another researched relationship. This relationship is depicted in Graph no. 2 for various discount rates. Graph no.2: Dependence of Tax Shield Present Value on the Moment of Tax Shield Claim Origination 0 1 2 3 4 5 6 7 8 9 10 11 12 330,00 Number of Months from the Origination of Tax Shield Claim till Tax Period's End 320,00 0,01 310,00 0,04 0,08 Tax Shield 300,00 0,12 290,00 0,16 0,2 280,00 0,24 0,28 270,00 260,00 250,00 Investigated relationship is linear for any selected discount rate. The largest [PV(TS)] logically occurs with the minimal time span between the moment of tax-shield claim origination and tax-period’s end, thus when the claim is realized on the last day of tax period, i.e. in our case on 31st December. In this point, the [PV(TS)] is paradoxically even larger than the basic value. This paradox is, however, merely ostensible. This fact has been caused by the existence of income tax deposits, by and discounting of individual tax payments. With increasing span, the [PV(TS)] declines, with its minimum at the point of 12 months prior to tax-period’s end, i.e. on 1st January in our case. The intersection of lines representing [PV(TS)] for various discount rate we consider to be rather interesting graphical finding. In case the tax-shield claim origination moment falls on the last day of tax period, the maximum [PV(TS)] is reached for discount rate of 20 %. Conversely, should the tax shield claim originates 12 months prior to tax-period’s end, the maximum [PV(TS)] shall be reached for the lowest discount rate (in our illustration for 1 % annual discount rate). With respect to above research, there is no doubt the present and basic value of tax shield can differ. However, the question remains whether the difference is significant enough so that it becomes necessary to adjust financial models to employ the present rather then basic value. For the purposes of further research we shall introduce a function [α(i;n)] producing for a given discount rate [i], and a given time span between tax shield claim origination and tax-period’s end [n] an absolute deviation of the difference between [TS] and [PV(TS)] as a percentage of the basic tax shield value [TS]. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 103 TS - PV (TS ) ´ 100 , TS wher α(i;n) = absolute value of a deviation between the present and basic value of tax e shield. a (i ; n ) = Graph no. 3 shows a map of function’s values [α(i;n)] for various discount rates [i] and various time spans (number of months) between the time of tax shield claim origination and tax-period’s end [n]. For the sake of better clarity, we have used discount rates up to 24 % level even they can be regarded as unrealistic, though. As follows form the graph and a consecutive Table no. 2, the deviation between the present and basic value of the tax shield does not represent in most cases more than 2.5 % of the basic value. 12 11 10 9 8 7 6 5 4 3 2 1 0 · · · ○ ○ ○ · ■■■ · · · ○ ○ ○ · ■■ ■ · · · · · · ■ ■ ■■ · · · · · · ■■ ■■ · · · · · ■ ■■ ■■ ○ ○ ○ · · · · · ■ ■ ■ ■ ■ ○ ○ ○ ○ ○ ○ · · · · · · ■ · · · ○ ○ ○ ○ · · · · ○ ○ ○ ○ · · · · · ○ ○ · · · · · · ○ · · · ■ · · · ○ ○ · · ○ ○ ○ · · · · · · · · ○ ○ ○ ○ · · · · · · · · ○ ○ ○ · · · · · · · · · · · · · · · · · · · · · · · · · ○ ○ ○ ○ ○ ○ ○ · · · · · · · · · · · · · · · · · · ○ ○ ○ ○ ○ ○ · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ○ ○ ○ ○ ○ ○ ○ ○ ○ ○ ○ ○ ○ ○ ○ ○ ○ · · · 0,01 0,02 0,03 0,04 0,05 0,06 0,07 0,08 0,09 0,10 0,11 0,12 0,13 0,14 0,15 0,16 0,17 0,18 0,19 0,20 0,21 0,22 0,23 0,24 Number of months prior to tax-period’s end Graph no. 3: Function α(i;n) Value Map Annual discount rate 10,0 % < α ≤ 12,5 % · α ≤ 2,5 % ○ 2,5 % < α ≤ 5,0 % · 12,5 % < α ≤ 15,0 % · 5,0 % < α ≤ 7,5 % ○ 15,0 % < α ≤ 17,5 % ■ 7,5 % < α ≤ 10,0 % · 17,5 % < α ≤ 20,0 % With annual discount rate between <0.01; 0.24>, almost 50 % (49.68 %) of all cases, and with the rates falling into the interval of <0.01; 0.12> that most corresponds to the real world, even 71.15 % of all cases do not exceed the 2.5 % level. At the same time, in the interval of <0.01; 0.12> deviations exceeding 10 % do not occur at all, in <0.01; 0.24> only in 10.9 % of all cases. The largest deviations occur in cases, when tax shield claims originate at the tax-period’s beginning, whereas in cases when the claims are originated at the tax-period’s end, the deviation values are rather negligible. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 104 > 17.5 % ≤ 20.0 % ∑ α > 15.0 % ≤ 17.5 % α > 12.5 % ≤ 15.0 % α > 10.0 % ≤ 12.5 % α > 7.5 % ≤ 10.0 % α > 5.0 % ≤ 7.5 % α > 2.5 % ≤ 5.0 % α ≤ 2.5 % α Frequency Discount rate Table no.2: Function value frequency [α(i;n)] up to absolute 0.24 relative 155 62 35 26 15 12 6 1 312 49.7 19.9 11.2 8.3 4.8 3.9 1.9 0.3 100.0 up to absolute 0.12 relative 111 30 13 2 - - - - 156 71.2 19.2 8.3 1.3 - - - - 100.0 When shall the deviations be considered significant enough and when not? The answer is rather straightforward and intuitive. The deviation is significant enough only when substitution of basic tax shield value with its present value leads to alteration in investment choice, or modification of other financial or capital decision. Thus, the deviation shall be with each case more or less significant in accordance with the scope of differences in results of applied decision-making criteria for various investment alternatives. This is, however, not always possible to find out prior to calculations, but very often impossible even after the calculations are done. As a certain guideline on whether it is necessary to use the present tax shield value instead of the basic one, we might use the map of function values [α(i;n)] depicted in Graph no. 5. Therefore, we recommend using the present value of tax shield instead of the basic one in all cases, where one can rationally expect that the value switch shall affect the relevant decision. For simplicity we can use the coefficients of tax shield present value [kTS(i;n)]: kTS (i; n ) = PV (TS ) TS , or PV (TS ) = TS ´ kTS (i; n) , wher kTS(i;n) = coefficient of present value of tax shield. e CONCLUSION Paper conclusions can be summarized in the following points: 1. Real moment of tax shield realization is effected at the moment, when the tax shield is reflected in taxpayer's cash flows in form of tax payment. 2. In relation to time span between the tax shield claim origination and tax period's end, holding other variables equal, the maximal present value of tax shield is reached when the claim originates at the end of tax period. Extending the time span the tax shield present value decreases. 3. In relation to discount rate, ceteris paribus, the maximal present value of tax shield is located at the parabola vertex, i.e. the tax shield present value grows until a certain level of discount rate is reached; on reaching the level the present value decreases. 4. In most of the cases, the absolute value of deviation between the present and basic value of tax shield does not exceed 2.5 % of basic value within economic real values of discount rate. The maximal deviation is reached with the longest possible time span between the tax shield claim origination and tax period's end, i.e. 12 months. 5. We recommend using the present value of tax shield instead of the basic value in all cases where we can expect the switch of values would influence financial decision. For simplicity, coefficients of present value of tax shield [kTS(i;n)] can be used. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 105 LITERATURE: [ 1] BREALEY, R. A., MYERS, C. M.: Principles of Corporate Finance. New York : Irwin / McGraw Hill, 2002. [ 2] COPELAND, T. E., WESTON, J. F.: Financial Theory and Corporate Finance. Reading, Addison Wesley, 1988. [3] COPELAND, T. E., KOLLER, T., MURRIN, J.: Valuation. Measuring and Managing the Value of Companies. New York, John Wiley, 2000. [4] DVOŘÁK, P., RADOVÁ, J.: Finanční matematika pro každého (Financial mathematics for Everybody). Prague : Grada, 2003. [ 5] HOLEČKOVÁ, J.: Analýza vlivu zdanění na finanční rozhodování na základě metody výpočtu efektivní daňové sazby - daňových klínů) (Analysis of the Impact of Taxation on Financial Decisions based on the Method of Calculations of Effective Tax Rates - Tax Wedges). Acta Oeconomica Pragensia, 2002, vol. 10, no. 1, p. 90-103 [ 6] MAREK, P., RADOVÁ, J.: Skutečný okamžik realizace daňové úspory (Actual Moment of Tax Shield Realization). Acta Oeconomica Pragensia, 2002, vol. 10, no. 1. p. 104-116 [ 7] MAŘÍK, M.: Určování hodnoty firem (Corporate Valuation). Prague : Ekopress, 1998. [ 8] VALACH, J., OTHERS: Finanční řízení podniku (Corporate Financial Management). Prague : Ekopress, 1999. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 106 MODELLING OF INFLATION FUZZY TIME SERIES USING THE COMPETITIVE NEURAL NETWORK Dušan Marček University of Žilina, Faculty of Management Science and Informatics, e-mail: [email protected] Abstract: Based on the work by [5], in this paper a dynamic process with linguistic values of inflation rates is defined and studied. A complete fuzzy modelling approach is proposed which includes: fuzzifying of observed data, developing a fuzzy time series model and computing the outputs. The fuzzy rules are by competitive neural network derived. Keywords: Fuzzy time series, Clustering technique, Neural network, Supervised Competitive Learning (SCL), 1. INTRODUCTION In this paper we describe one application of the fuzzy set theory for inflation time series. Fuzzy logic started over thirty years ago as a theory to deal with uncertainty and vague linguistic notions. The first applications of the fuzzy sets have been realised in industrial machinery. There are lists available with hundreds of application in all areas of the electrical and electronic market [1] and primarily for control and predictive diagnostics of complicated systems and processes. Fuzzy time series is proposed in order to deal with such situations in which the traditional time series model is no longer applicable, i.e., if the historical data are not real numbers but they are linguistic variable. In the case if the historical data are real numbers, the data should be fuzzified first. In this paper, we assume there is a relationship between observations at time t and those at previous time. Thus, the modelling process is to develop fuzzy relations among the observations at different times. The concept of fuzzy time series is applied to represent simple relationship between observation at time t and at the previous times, i.e., the observation at time t is the accumulated results of the observations at the previous times (see also [5] ). In next section a possible new method is proposed as a fuzzy system product with heuristic rules describing the behaviour of inflation with linear membership functions for the fuzzification of inflation time series. 2. MODELLING INFLATION WITH FUZZY TIME SERIES As we mentioned above, in practice it is difficult to forecast the expected stock price since economic and financial decision is highly erratically. The objective value of the expected inflation should be vague and uncertain. Therefore, we can apply a fuzzy analysis to the inflation time series. Since the first two methods are frequently used in technical or technological systems, we will apply the third method. As we mentioned earlier, for its applying it is supposed that a database describing previous inputoutput behaviour of a system and the adequate model of the observed process are available. A powerful tool for generating fuzzy rules purely from a database in a fuzzy system are neural networks. The neural networks can adaptively generate the fuzzy rules in a fuzzy system by supervised productspace clustering technique [2]. Next, in a numerical example, we will illustrate and show how to obtain fuzzy rules using fuzzy set theory and neural networks. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 107 Let us consider a simple example. The data set used in this example (the 514 of monthly inflation observations from February 1956 to November 1998) was published at http://neatideas.com/data/inflatdata.htm). As do many economic time series, the original data exhibit considerable inequality of variance over time, and the log transformation stabilises this behaviour. Fig. 1 illustrates the time plot of this time series. This time series shows no apparent trend or periodic structure. We would like to develop a time series model for this process so that a predictor for the process output can be developed. To build a forecast model the sample period for analysis y1 , ..., y344 was defined, i.e. the period over which the forecasting model can be developed and estimated, and the ex post forecast period (validation data set), y345 , ..., y514 as the time period from the first observation after the end of the sample period to the most recent observation. By using only the actual and forecast values within the ex post forecasting period only, the accuracy of the model can be calculated. inflation 2.5 2 1.5 1 0.5 56 -0.5 61 66 71 76 81 86 91 96 t -1 Fig. 1. Natural logarithm of monthly inflation from February 1956 to November 1998 After some experimentation, we formulated a model relating the value yt of the series at time t that depends only on its previous value yt-1 and on random disturbance, i.e. y t = x + f1 yt -1 + e t (1) where the variable yt (in our case the first difference of inflation rate) is explaned by only on its previous values, and et is a white noise disturbance term. Next, analogously to the conventional time series models (1), it is assumed that the observation at the time t accumulates the information of the observation at the previous times, i.e. there exists a fuzzy relation Rij(t, t - 1): and equivalently Rij(t, t-1): yti-1 ® ytj (2) Yt ® Yt -1 (3) or ytj = yti-1 ° Rij(t, t - 1) (4) which can be interpreted as fuzzy implication Rij(t, t-1) º if yt -1 then yt i j t Î t j (5) i t -1 where y Y , y Î Yt -1 , i Î I, j Î J, I and J are indices sets for Yt and Yt -1 respectively, “ ° “ is the maxmin composition. The Eq. (4) is called first-order model of the fuzzy time series with lag p = 1. The describing a basic structure of the fuzzy rule based system for fuzzy time series modelling was made in [4]. At this stage, we will only give some outlines to model fuzzy time series in a fuzzy environment. The fuzzy time series modelling procedure consists of a implementation of several steps, usually as follows: 1. Define input-output variables and the universes of discourse. 2. Define (collect) linguistic values, i.e. fuzzy sets on universes discourse. 3. Define (find) fuzzy relations (fuzzy rules). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 108 4. 5. Apply the input to the model and calculate the output. Defuzzify the output of the model. The proposed fuzzy time series modelling procedure is divided into five steps. In Step 1 to Step 2, the input data are fuzzified, in Step 3, analogously to the conventional model (1), the fuzzy time series model, i.e. the fuzzy relation is created. The Steps 4, 5 are considered as application of the model (i.e., analysis of economic structures and forecasting). In the literature this modelling approach is known as fuzzy rule based system (see Fig. 2). Below, we will discuss these steps and apply them to the inflation time series at a more detailed level. In Fig. 2 the fuzzy rule based system has three blocks: (a) block for fuzzification of input variables, (b) knowledge base block, and (c) defuzzification block. Fig. 2. Structure of the fuzzy system for forecast of the inflation In the fuzzification block firstly, we specified input and output variables. The input variable indicated as { xt -1 } is the laged first difference of inflation values { yt } and calculated as xt -1 = yt -1 - yt -2 , t = 3, 4, ... . The output variable indicated as { xt } is the first difference of inflation rates { yt } and calculated as xt = y t - yt -1 , t = 2, 3, ... . The variable ranges are as follows: - 0,75 £ xt , xt -1 ³ 0,75 These ranges define the universe of discourse within which the data of xt -1 and xt are and on which the fuzzy sets have to be specified. The universes of discourse were partioned into the seven intervals. Next we specified the fuzzy set values on the input and output universe. The fuzzy sets numerically represented linguistic terms. Each fuzzy variable assumed seven fuzzy set values as follows: NL: Negative Large, NM: Negative Medium, NS: Negative Small, Z: Zero, PS: Positive Small, PM: Positive Medium, PL: Positive Large. Fuzzy sets contain elements with degrees of membership. Fuzzy membership function can have different shapes. The triangular membership functions were chosen. Fig. 3 shows membership function graphs of the fuzzy subsets above. Fig. 3. The membership functions for each linguistic fuzzy set value The input and output spaces were partioned into the seven disjoint fuzzy sets. From membership INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 109 function graphs mt -1 , mt in Fig. 3 is shown that the seven intervals [-0,75; -0,375], [-0,375; -0,225], [-0,225 0,075], [-0,075; 0,075], [0,075; 0,225], [0,225; 0,375], [0,375; 0,75] correspond respectively to NL, NM, NS, Z, PS, PM, PL. Next we specified the fuzzy rule base or bank of fuzzy relations. The competitive neural network was applied which uses supervised competitive learning to derive fuzzy rules from the database. The bank contains the following 5 fuzzy rules i j IF xt -1 = NL THEN xt = NS i j IF xt -1 = PS THEN xt = NM i i j i j IF xt -1 = PS THEN xt = PS IF xt -1 = PS THEN xt = PL j IF xt -1 = PM THEN xt = PS (6) Finally, we determined the output action given the input conditions. We used the Mamdani´s implication [3]. Each fuzzy rule produces the output fuzzy set clipped at the degree of membership determined by the input condition and the fuzzy rule. When the input, say xt -1 = x344 , is applied to the model (4), the i j i j j output xt = x345 can be calculated. It is possible to compute the output fuzzy value xt by following simple procedure consisting of three steps: · Compute the membership function values mNL( xt -1 ), mNM( xt -1 ), ..., mPL( xt -1 ) for the input xt -1 using the the membership functions pictured in Fig. 3. · Substitute the computed membership function values in fuzzy relations (5), (6). · Apply the max-min composition to obtain the resulting value xt of fuzzy relations. j j Following the above principles, we have obtained the predicted fuzzy value for the inflation xt = x345 = 0,74933. j The inflation values in the output xt , t = 345, 346, ... are not very appropriate for a decision support because they are fuzzy sets. To obtain a simple numerical value in the output universe of discourse, a conversion of the fuzzy output is needed. This step is called defuzzification. The simplest defuzzification scheme seeks for the value x$ t that has middle membership in the output fuzzy set. Hence, this defuzzification method is called middle of maxima, abbreviated MOM. Following this method, we have obtained the predicted value for the x$345 = - 0,15 . The remaining forecasts for ex post forecast period t = 346, 347, ... may be generated similarly. As a final point, let us examine what has been gained by use of a fuzzy time series model over an ordinary AR(1) model for the output x345 . For this purpose we computed prediction limits on the one-stepahead forecast from the AR(1) model, and fuzzy time series model. The 95 percent interval around the actual inflation value based on statistical theory is xˆ345 m u1-a / 2sˆ e (1 + f12 ) 2 = 0,00312 m 1,96 0,15476(1 + (-0,1248) 2 ) 2 = (-0,0442 ; 0,05043) where x$ 345 represents the forecast for period t = 345 made at origin t = 344, u1-a is a 100 (1 - a 2 ) percentile 1 1 2 of the standard normal distribution, and s$e an estimate of the standard deviation of the noise. An intuitive method for constructing confidence intervals for fuzzy time series model is simply the defuzzification method firs of maxima and first of minima to obtain prediction limits on the one-step-ahead forecast. In our example, the „confidence“ interval for fuzzy time series value x$ 345 = 0,00312 is (-0,30256 to 0,3088). The actual value for the AR(1) model fell not within the forecast interval, and moreover, its sign is apposite to the forecast value sign. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 110 3. CONCLUSION In this paper we have discussed the fuzzy model for inflation forecast. Using the technical analogy, we have designed the fuzzy system structure. From the data we have derived the short time behavior in the inflation movements. The interval linear membership functions are employed in order to avoid non-linearity. We have evaluated inflation forecasts by fuzzy model. To determine the fuzzy relation of the first-order fuzzy time series model, the neural network with SCL clustering technique was used to derive the rules directly from the database. The method may be of real usefulness in practical applications, where usually the expert can not explain linguistically, what control actions the process takes or there is no knowledge of the process. ACKNOWLEDGEMENT This work was supported by Slovak grant foundation under the grant No. 1/9183/02. REFERENCES [1] HELLONDOOM, H.: Industrial Application of Fuzzy Control. Fuzzy Structures, Current Trends. Tatra Mountains Mathematical Publications, Volume 13, 1997, pp. (69-91) [2] KOSKO, B.: Neural networks and fuzzy systems - a dynamical systems approach to machine intelligence. Prentice-Hall International, Inc. 1992 [3] MAMDANI, E., H.: Application of a fuzzy logic to approximate reasoning using linguistic synthesis. IEEE Trans. Comput. 26 (1977), pp. 1182-1191 [4] MARČEK, D.: Forecasting Stock Prices With Fuzzy Time Series. Studies of University in Zilina. Mathematical-Physical Series. Vol. 12., December 1999, ISSN 80-7100-690-4, pp.(31-35) [5] SONG, Q. and CHISSON, B.,S.: Fuzzy time series and its models. Fuzzy Sets and Systems 54 (1993) 269277, North-Holland INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 111 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 112 USE OF NEW METHOD FOR LEARNING OF NEURAL NETWORKS Jindřich Petrucha, Vladimír Mikula European polytechnical institute, Kunovice, Osvobození 699, Czech Republic Tel. 572 548 035, Fax: 572 549 018, E-mail: [email protected], [email protected] Abstract: Neural networks are used for classification and prediction of data at future. For training of neural networks we can use the backpropagation procedure, although there are the limitations associated with this technique and these problems are well documented. Global search techniques such as genetic algorithm and tabu search have been used for training of neural networks, too. These algorithms have been focused more on accuracy than on the training speed because the training speed is not important in cases where the training procedure can be done off- line. In this paper a training procedure is described, that can achieve the needed accuracy at limited training time. We then use the data from simulation of time series and we compare the proposed methods with the backpropagation method in our simulator. Keywords: artificial neural networks, artificial neurons, scatter search, improvement method, synaptic weights, layers of neurons, hidden layers, input patterns, learning (training) of neural nets, error function, time series, prediction of future values, single step prediction. 1. INTRODUCTION Training of neural networks is a complicated problem. There are the situations, where the training procedure spends hour of the computer time and we need to find the results of the problem in some minutes. The speed of the training arises in context of optimising a simulation. Sometimes the data are not inside a large database but they are generated on-line, this situations needs the on-line training of the neural network. There is e.g. a simulation model of the factory – for the simulation of the flow of products through the factory over a period. The product managers make virtual changes of the production, to predict the impacts without ever changing of real equipment or manufacturing. The models are applied in financial planning and in marketing strategy. The limitation of simulation is an inability to evaluate then fraction of the large of options. Practical problems need to consider numbers of interconnected alternatives. In fact these complexities and uncertainties are primary reasons that simulation is chosen as a basic means for handling of such problems. The area that incorporates artificial intelligence evolutionary processes led to the creation of new approaches that successfully integrated simulation and optimization. Specifically the neural network can be used to filter out the solutions that are likely to perform poorly when the simulation is executed. 2. NEURAL NETWORKS AND PROCESS OF THE OPTIMIZATION We try to formalize the use of neural network as a prediction model for the simulation and we describe on-line training procedure that is producing fairly accurate prediction within a reasonable time. Let us introduce the notation that helps us to formalize the process of optimization and training process. x f(x) p(x,w) x* l u -a solution problem -output of the simulation -the output as predicted by neural network with weights w when the solution x is used as inputs -the best solutions of the problem -set of lower for decision x -set of upper for decision x INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 113 During the optimization process we try to find : Min f(x), where it is valid l ≤x≤u · · · The training problem consists of finding the set of weights w that minimize an aggregate error, we use MSE (mean squared errors). During the search for the optimal values of x, the procedure generates set of ALL solutions x. Let the TRAIN be a random sample of solutions in ALL, such that TRAIN < ALL. We define the problem: Min g ( w) = 1 å | TRAIN | x ÎTRAIN ( f ( x ) - p ( x, w ) 2 w - is the set of optimization weights of neural network 2.1 Training algorithm Scatter search is an evolutionary method that is founded on strategies that only piecemeal come to propose as augmentations to GA. The approach uses principles and strategies that are still not emulated by other evolutionary methods. We can find good results for solving variety of complex optimization problems, including neural network training. Outline of training procedure 0. Normalise input and output data 1. Start with P = 0 Use diversification method to construct a solution w between wlow and whigh If w Ï P add w to P, otherwise discard w. Repeat this step until P| = |PSize| Apply the improvement method to the best b/2 solutions in P to generate w(1)..w(b/2) Generate b/2 more solutions, where w(b/2+i)=w(i) (1+[U-0.3,0.3] for i=1..b/2 RefSet {w(1), ... w(b)} 2. Order RefSet according to their objective function value w(1)is best w(b) is worst while ( NumEval < TotalEval ) 3. Generate NewPairs, which consist of all Paris of solutions in RefSet that include at lest one new solution. Make NewSolutions = 0 for ( all NewPairs ) do 4. select the next pair (w(i), w(j)) in NewPairs 5. obtain new solutions w as linear combinations of (w(i), w(j))and add them to NewSolutions end for 6. Select NewSolutions and apply the improvement method. for ( each proved w) do if ( w is not in RefSet and g(w) < g(w(b)) ) then make w(b) =w and reorder RefSet end for while ( IntCount < Intimit) 7. Make IntCount=IntCount + 1 and w = w(1) 8. Make w(b/2+i)=w(i) (1+[U-0.05,0.05]) and apply the improvement method. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 114 if ( g(w) < g(w(1)) then make w(1) = w and IntCount=0. end while 9. Apply the improvement method to w(i) for i=1,…, b/2 in RefSet 10. Make w(b/2+i))= w(i)(1+ U[ -0.01, 0.01] ) i=1,…, b/2 in RefSet end while Let us describe each steps of this outline : · The main goal of the diversification is to generate solutions that are diverse with respect to those solutions that have been generated in the past. · In step 2 we order RefSet according to quality, we use MSE on training set of data for measure error function. The best solution is first one in the list. · In step 3 we construct the NewPairs by using linear combinations when the NewPairs are selected one at time in lexicograpical order. · In step 6 New Solutions are subject to the improvement method. Each improvement solution is tested for admission into RefSet.If new solution improves upon the worst solution in RefSet, the new solution replaces the worst and RefSet is ordered. · In step 7 the procedure intensifies the search around the best-known solutions. · In step 8 the solutions is perturbed and improvement method is used · In step 9 and 10 procedure uses improvement method to the best b/2 solutions in RefSet and replaces the worst b/2 solutions with perturbation of the best b/2. 2.2 Architecture of Neural Network Our aim was to develop the program simulator that can be able to predict data for future period on the base of known data in past. We used well-known feed forward multilayer perceptron neural network architecture consisting of at least three layers. The input layer is for distribution of input pattern. x1 Output of predicted value x2 X xn0 Output layer of neurons Input (fun out ) layer of neural network The first hidden layer of neurons Fig. 1. MLP architecture of neural network for data prediction There are several ways to compute the forecast: Let us assume we have a simple time series D1, … ,D100 values of data, we want to forecast future data values Di , i > 100, and we desire to use five lagged values as inputs Single step – open loop forecasting: Train a network with target Di and inputs Di-1 , Di-2 ,….. Di-N , (N – number of train patterns). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 115 Then forecast D 101 = Net (D 100 , D 99 , D 98 , D 97 , D 96 ) D 102 = Net (D 101 , D 100 , D 99 , D 98 , D 97 ) D 103 = Net (D 102 , D 101 , D 100 , D 99 , D 98 ) Multi-step – closed loop forecasting: Forecast D 101 as P 101 = Net (D 100 , D 99 , D 98 , D 97 , D 96 ) Forecast D 102 as P 102 = Net (P 101 , D 100 , D 99 , D 98 , D 97 ) Forecast D 103 as P 103 = Net (P 102 , P 101 , D 100 , D 99 , D 98 ) N– step ahead forecasting: Assume N=3 Compute P 101 = Net (D 100 , D 99 , D 98 , D 97 , D 96 ) Compute P 102 = Net (P 101 , D 100 , D 99 , D 98 , D 97 ) Forecast D 103 as P 103 = Net (P 102 , P 101 , D 100 , D 99 , D 98 ) Forecast D 104 as P 104 = Net (P 103 , P 102 , D 101 , D 100 , D 99 ) Forecast D 105 as P 105 = Net (P 104 , P 103 , D 102 , D 101 , D 100 ) Which method we choose for computing of forecasting depends in part on the requirements of our application. For our simulator we choose single step – open loop method and we try to compare the experimental results with the simulator that uses the backpropagation method. 3. SIMULATOR – PROGRAM CHARACTERISTICS Our aim is to make a program for using times series, that can be able to predicted data for future period. For making this program we use neural network architecture that uses three layers. In our simulator we focus on problematics of learning the neural networks from past data and on finding the weight parameters of the net for the best time prediction. Program is designed on the principle of three layer architecture with the possibilities to determine the number of neurons for input layer and for hidden layer . Number of neurons for output layer is one. This information we can load from the text file that obtain information about architecture and times data. The program is written in PASCAL DELPHI, using graphics component for visualising the data. Following figures show in graphical form the result obtained by the proposed program. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 116 Fig. 2. Fundamental screen of program Graph of MSE during the learning Graph of testing and prediction of experimental data 4. TESTING During our tests we tried to find the best dimension of input vector and the number of neurons in hidden layer for architecture of neural nets. The number of neurons is written in the header of text files before the data of time series. Architecture of neural network 5 - 6 -1 5 - 7 -1 5 - 10 -1 10 - 15 -1 Table 1. Experimental results MSE 0.1606 0.21 0.203 0.17 The best results have been reached by using the architecture 5 – 6 – 1. The another parameters for algorithm are constants b = 10 , PSize = 100 , TotalEval = 1000 cycles and IntLimit=20. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 117 5. CONCLUSION The neural network was used for the prediction of data at future and real result that we have got we have compared with actual data. There are some differences between the prediction and actual data. The scatter search algorithm has good ability to find the results in the first phase of searching when the procedure intensifies around the best known solution in the step 7. The experiments show that the Scatter Search algorithm can be used for practical using when we need a reasonable prediction accuracy during on-line learning of the neural network. REFERENCES: [ 1] NOVÁK, M..: Umělé neuronové sítě - teorie a aplikace. Nakladatelství C.H.Beck, Praha 1998 [ 2] KOSKO, B.: Neural Networks and Fuzzy Systems. Prentice – Hall, 1992 [ 3] KUTZA, K.: Neural Networks at your Fingertips. http://www.geocities.com/CapeCanaveral/1624/ [ 4] LAGUNA, M. MARTÍ, R.: Neural Network Prediction in System for Optimizing Simulation. University of Colorado, University of Valencia, 2001 [ 5] GLOVER, F. : A Template for Scatter Search And Path Relinking. University of Colorado, 1998 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 118 NEURAL NETWORK AND JOINT TRANSFORM CORRELATOR APPLIED FOR THERMAL IMAGERY-BASED RECOGNITION Krzysztof Kościuszkiewicz1), Joanna Kobel2), Jacek Mazurkiewicz3) 1,3) Institute of Engineering Cybernetics, 2) Institute of Physics Wroclaw University of Technology ul. Wybrzeże Wyspiańskiego 27, 50-370 Wrocław, POLAND Phone: +48-71-3202681, +48-71-3202825 Fax: +48-71-3212677 E-mail: [email protected], [email protected], [email protected] Abstract: This paper focuses on two different implementations of infrared-based biometric system. Artificial neural network and computer simulated joint transform correlator realize the classification and identification of human face’s thermograms. Accuracy of both systems is evaluated using standard ratios for biometric systems: FRR, FAR and EER. Keywords: Multilayer Perceptron, thermogram, identification 1. INTRODUCTION Authentication methods based on recognition and comparison of infrared face images are considered to guarantee a third best security level among other biometric methods. Infrared imagery cross-over error rate places it just behind two eye recognition methods: retina and iris scanning. Infrared face recognition is based on the fact that thermal image of each human’s face is individual and unrepeatable. Even identical twins have different infrared images. The IR approach posses some advantages over visible wavelength images. First, thermal image is not vulnerable to disguises, can’t be altered or camouflaged. The systems based on facial thermogram technology are external lighting independent and passive. This means that authentication can be done in low light or even in the total darkness, without physical contact and human cooperation [6, 8]. This technology also has some drawbacks. The main disadvantage is a low stability of thermal pattern. Thermogram can be disturbed by many external and internal factors, like environment temperature, illnesses, emotional state etc. Fig. 1. Visible wavelength and thermal photographs of the same face 2. RESEARCH MATERIAL The experiment was performed on 10 adult healthy volunteers. The group consisted of five males and five females, aged 23 to 40. The faces were recorded with AGEMA 900 LW system, at the ambient temperature 21,5°C and the 28% humidity of the environment. Spectral region from 8 to 13 mm was chosen as a sensing range. For every subject 10 images in different conditions were taken. From original photographs recorded in format of 135x270 pixel bitmaps, the square sample of central face part was cropped. This way database of 100 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 119 images of size 85x85 pixels was created. 3. ARTIFICIAL NEURAL NETWORK BASED SYSTEM Neural implementation is based on a classic model of multi-layered perceptron. This topic is well described in numerous publications, just to mention [2, 3, 5]. Good introduction to general neural processing can be found in [3]. The system itself consists of three main blocks (Fig. 2). The first stage performs data acquisition and preprocessing. Artificial neural network is the second block - main processing is done in this stage. The third (output) block transforms output vector of the NN into response of the whole system. training vectors image preprocessing (block 1) input vector ANN (block 2) threshold output vector output vector interpretation (block 3) response Fig. 2. Neural processing system model Input to the system is presented as 85x85 pixel 8-bit grayscale bitmap. In the first stage size of the bitmap is rescaled to 25x25. As value of each pixel falls between 0 and 255 another rescaling has to be done in order to effectively utilize linear part of activation function. Also the system response should be invariant with respect to changes in environment temperature. Both problems can be solved with rescaling all pixel values in the input image to fall into range (0, 1). This is achieved by applying pi = where: xi - min{ x} , max{ x} - min{ x} (1) pi is i-th element of the input ANN vector p, xi is i-th element of the input image vector x, to each pixel. As it was stated before, the second layer is a classic implementation of MLP. The ANN has 625 neurons in the input layer, one hidden layer consisting of 12 units (number of neurons in the hidden layer has been chosen experimentally). Response of the ANN is coded in the "1 of N" manner. Unipolar sigmoid function f (ui ) = 1 , 1 + exp(- bui ) (2) where: ui is weighted sum of i-th perceptron input values, β is parameter defining “steepness” of activation function f, was chosen for the nonlinear activation element, so all neuron output values fall in range (0, 1). Training set consisted of 3 photos of five first subjects. The net was trained with use of backpropagation algorithm with momentum and variable learning rate. Mean square error measure (MSE) MSE (o, t ) = 1 N N å (oi - ti )2 , (3) i =1 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 120 where: oi is i-th element of network output vector o, ti is i-th element of target (desired) output vector t, N is the number of output neurons, was used as the base of network energy function. The third block of the system as an input takes an output vector of the ANN. It performs MSE calculation of output vector o against the set of target output vectors T. Subject is considered authorized iff MSE of at least one of target output vectors ti ÎT with o is lower than threshold d. Index i of neuron with the highest response, max{oi}, indicates the class to which the corresponding input vector (and the subject itself) belongs. 4. JOINT TRANSFORM CORRELATOR APPROACH The joint transform correlator (JTC) is one of the main optical architectures (except for the Van der Lugt correlator) that are being used for purposes of pattern recognition. There exists a comprehensive bibliography describing the applications of JTC, as e.g. [1, 4] In this study we have used computer simulation of the JTC. As the basis of an optical correlator the fundamental relation for optical signal correlation was applied: s * f = FT -1 ( S × F * ) , where: (4) s and f are functions (images) being correlated, S and F are Fourier transforms of s and f, respectively, FT-1 is the inverse Fourier transform. For evaluation of the correlation quality two criterions were used: Discriminant Capability (DC) and Peak-to-Correlation Energy (PCE). DC characterizes the ability to recognize the target image against non-target (the recognized thermogram) and is defined as DC = where: CC , A (5) CC is non-target correlation signal (cross correlation), A is target correlation signal (autocorrelation). Peak-to-Correlation Energy is defined as c(0,0 ) PCE = , Ec 2 where: (6) c(0,0) is the highest value of the correlation peak, Ec is the correlation plane energy: ¥ ¥ -¥ -¥ Ec = ò dx ò dy c( x, y ) . 2 (7) The PCE calculated for a high and sharp correlation peak has a larger value in comparison to the case of low and broad correlation peak [7]. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 121 5. ACCURACY MEASURES FOR BIOMETRIC SYSTEMS Performance of biometric system is measured with two main parameters: false acceptance rate (FAR) and false rejection rate (FRR). FAR is the probability that a biometric system will incorrectly identify an individual or will fail to reject an impostor. FFR is the probability that a system will fail to identify a subject, or verify the legitimate claimed identity of an individual. Usually both ratios depend on one parameter. In ANN-based system it is the threshold of MSE (3) calculated for output vector o against desired output vector t. In the JTC case it’s the threshold computed for described above criterions values, which are correlation measures. It is obvious that change in any of two measures implies change of another one in opposite direction. When threshold decreases the system becomes more rigorous, which results in decrease of FAR but an increase of FRR. Analogous situation occurs when threshold increases. Equal error rate (EER), sometimes called cross-over error rate, is evaluated at the intersection point of FAR and FRR plotted against threshold. Corresponding threshold value, called equal error point (EEP) is often used in a final system implementation. Another way to evaluate optimal threshold is finding minimum of weighted FAR+FRR function, which minimizes total number of authorization errors (regardless of their type). 6. ACCURACY ANALYSIS OF ANN-BASED SYSTEM Error ratios have been averaged for 20 training sessions. For each rate sessions with different MSE training goals have been computed. To preserve clarity FAR and FRR have been presented on two different plots (Fig. 3., Fig. 4.). One can notice that averaged FAR increases linearly with threshold value. This fact can be simply explained: during training sessions images of subjects to be rejected weren’t presented to the network. It wouldn’t be advisable, as during training ANN in a real implementation one cannot present images of all unauthorized subjects to the network. It’s obvious, that network response for images from classes, that haven’t been used during training sessions, has a statistical character. Fig.3. Average FRR of neural system Fig. 4. Average FAR of neural system In order to maximize system performance ANN should be trained to lowest possible MSE goal. Table 1 presents EER values for the same ANN configuration and for different training goals. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 122 Table 1. Average neural system EER 1 2 3 Training MSE goal 0.01 0.005 0.001 Average EER 0.1393 0.0858 0.0558 7. ACCURACY ANALYSIS OF JTC BASED SYSTEM JTC based thermogram recognition was used in identification mode, where the biometric system identifies a person from entire enrolled population by searching a database for a match. For accuracy evaluation DC (5) and PCE (6) criterions were used. The results are presented in Fig. 5. and Fig. 6. Fig. 5. JTC error rates based on DC criterion Fig. 6. JTC error rates based on PCE criterion In the next step, for each person the mean images based on 5 different thermograms were created and accuracy evaluation was repeated (Fig. 7. and Fig. 8.). Generally in both cases using the DC criterion for correlation accuracy evaluation yield better results. Moreover the applying of thermograms averaging improved results of recognition from EER=11,4% to ERR=8,3% for DC and form EER=30,4% to EER=28,3% for PCE. Fig. 7. JTC error rates based on DC criterion for averaged thermograms Fig. 8. JTC error rates based on PCE criterion for averaged thermograms 8. CONCLUSIONS AND FURTHER RESEARCH In this paper we have shown two complete neural biometric systems. Accuracy has been measured using standard ratios for this kind of applications. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 123 Further research in this area will include testing the systems with larger databases of subjects and development of better preprocessing methods in order to improve accuracy. Visible wavelength and infrared-based systems shall be compared. We will also make attempts to implement both systems in hardware. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] COLIN J., LANDRU N., LAUDE V. et al., 1999: High-speed photorefractive joint transform correlator using nonlinear filters, Applied Optics, 1, pp.: 283-285 HECHT-NIELSEN R., 1990: Neurocomputing, Addison-Wesley, USA HERTZ, J., KORGH, A., PALMER R., 1991: Introduction to the Theory of Neural Computation, Addison-Wesley, USA KODATE K., INABA R., WATANABE E., KAMIYA T., 2002: Facial recognition by compact parallel optical correlator. Measurement Science and Technology, vol. 13, pp.: 1756-1766 OSOWSKI S., 1996: Sieci neuronowe w ujęciu algorytmicznym, Wydawnictwa Naukowo-Techniczne, Warszawa SOCOLINSKY D. A., WOLFF L. B., NEUHEISEL J. D., EVELAND C. K., 2001: Illumination Invariant Face Recognition Using Thermal Infrared Imagery, IEEE Proceedings on Computer Vision and Pattern Recognition VIJAYA KUMAR B. V. K., HASSEBROOK L., 1990: Performance for correlation filters, Applied Optics, 129, pp.: 2997-3006 WILDER J., PHILLIPS P., JIANG C., WIENER S., 1996: Comparison of visible and infrared imagery for face recognition, Proceedings of the 2nd International Conference on Automatic Face & Gesture Recognition, Killington, pp.: 182-187 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 124 NEURAL NETWORK IN ADAPTIVE CONTROL Václav Veleba, Petr Pivoňka Dept. of Control and Instrumentation, FEEC, BUT, Božetěchova 2, 612 66 Brno, Czech Republic Tel.: +420 5 4114-1193, Fax: +420 5 4114-1123, E-mail: [email protected] Abstract: This paper describes some approaches to using neural networks as adaptive controllers. We used three different heterogeneous structures with neural network working as controller, operator and estimator: · Direct neural controllers · PSD controller tuned by neural networks · Adaptive controllers tuned by Ziegler-Nichols method with process identification using neural networks The controllers were implemented in MATLAB-Simulink and programmable logic controller. They were tested on different dynamic models and real process. Their principle, advantages and disadvantages were compared. 1. INTRODUCTION Used neural networks in adaptive control have several employments: they were applied as controller, for model inverse dynamics or for parameter identification of controlled system. The tested algorithms were implemented in ANSI C, which is easy portable. Algorithms performances were validated by simulation in Matlab and then algorithms were implemented in the programmable logic controller B&R 2005. During implementation was strictly keeping principle of three-step design of control algorithm [4]. The first step is simple simulation and algorithm is tested in MATLAB on a number of mathematic models. In the second step, we keep the same simulating scheme, but numeric model is replaced by an interface connected in PLC with real process. Heterogeneous algorithm is now tested on real process and comfort of simulation environment is retained. Direct implementation in PLC is the last step. The previous steps make it very simple with minimum changes in code. 2. NEURAL ADAPTIVE CONTROLLERS 2.1. Direct inverse controller A controller shown in figure 1 is one of the simplest neural controllers. Classical neural network is trained using known inputs and targets, but network used as this controller is trained to mineralize control error e(k). Network is trained on-line in any control step. When error is nearly zero, we could say, then, that neural network is able to approximate an inverse dynamic of controlled process. w(k) Neural network u(k) Process y(k) e(k) Fig. 1: Control with direct inverse controller INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 125 Signals, which give information of process dynamic, are connected to inputs of neural networks, for improving properties of control. These inputs usually contain delay signals of set point w(k) and output y(k) according to figure 2, but exist variants with simple state reconstruction [5]. During control of different process was established following properties of direct inverse controller: · Simple and efficient control mechanism · Low memory cost – last process data are not saved · Zero control error despite non integral character of process · Relative fast and damped oscillation response of the process to a step change · Slow adaptation · Impossibility of batch learning and selecting significant training instance · Steady control error during linear changing of set point · Casual oscillations due to over train Regulating proces of direct inverse controller is shown in figure 3. w(k) w(k) z z z e(k) w(k-1) -1 -1 w(k-2) NN y(k-1) -1 z u(k) F(s) y(k) y(k-2) -1 z -1 y(k-3) Fig. 2: Direct inverse controller Set hodnota point and process output Zadana a regulovana velicina 6 4 2 0 -2 -4 -6 0 100 200 300 t [s] 400 500 600 400 500 600 Control Akcni zasah value 1.5 1 0.5 0 -0.5 -1 -1.5 0 100 200 300 t [s] Fig. 3: Control process of direct inverse controller. Transfer function of process is F2(s). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 126 2.2. Adaptive PSD controller tuned by neural networks The PID controller and discrete PSD controller are the most usual controllers in world. Robustness, simplicity and stability are controller rated features in many industry applications. Adaptive PSD controller tuned by neural networks is based on idea of neural network working as human operator settings parameters of PSD. Although he provides less efficient transient response, it is simple to implement and standard algorithm PSD ensure stability between change of plant dynamic and response of adaptive mechanism. Direct access to tuned controller parameters is decided advantage. A structure shown in figure 4 uses multilayer neural network. These networks inputs are connected to delayed control value and control error signals. Neural network is trained to mineralize control error, analogically to direct inverse controller. Feature of adaptive PSD controller tuned by neural networks: · Easy to implement · Fast computed control value · Access to tuned controller parameters · Robustness due to conventional PSD algorithm · Passable noise immunity · Damped oscillating step response · Casual steady control error during zero order signal Regulating proces of adaptive PSD controller tuned by neural networks is shown in figure 5. u(k) w(k) e(k) K TI F(s) z TD z z -1 y(k-3) u(k-1) -1 z NN TI u(k-2) -1 z Fig. 4: -1 y(k-2) -1 z z -1 K y(k-1) -1 z y(k) -1 u(k-3) TD Adaptive PSD controller tuned by neural networks 2.3. Adaptive Ziegler-Nichols rule based PSD controllers with neural network identification In this chapter, we discuss a controller that uses an algorithm [2] for computing the ultimate gain KULT and the ultimate period TULT. We compute the ultimate gain and the ultimate period from parametric model of controlled process. This model is based on neural network. We used a popular Ziegler-Nichols algorithm to set the parameters of PSD controller. This sequence (figure 7) ensures adaptive response for process dynamic change. Neural network dynamic is represented by step delays as above. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 127 Set point and process output Zadana hodnota a regulovana velicina 10 5 0 -5 0 100 200 300 t [s] Akcni Controlzasah value 400 500 600 0 100 200 300 t [s] 400 500 600 10 5 0 -5 -10 Fig. 5: Control process of adaptive PSD controller tuned by neural networks. Transfer function of process is F2(s). The algorithm for minimization can be the back propagation (BP) or Levenberg-Marquardt (LM). An advantage of (LM) is foremost a speed of convergence, but a disadvantage is a complicated implementation due to complex matrixes computations. LM is sensitive to disturbances and offset. From this point of view, LM is similar to most frequent method for identification – least square method. For this reason, we prefer BP that provides very fast sampling even high noise level. Nonlinear structure together with powerful learning algorithm has ability to overcome some nonlinearity in control loop as quantization effect. We used two standard controllers in heterogeneous control structure: · Takahashi controller · PSD controller with derivate part filtration using Tustin approximation The different between two controllers is evident from transient step shown in figure 6. The start of step was set after end of adaptation (stability of estimated parameters). Set point and process output Set point Tustin approximation Takahashi controller Fig. 6: Comparison between neural identification based controller. Takahashi controller and PSD controller with Tustin approximation control process F2(s). Adaptation was over after 32s. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 128 KULT Critical parameters finding TULT a1 y (k-1) a2 y (k-2) NN b1 u (k-2) y (k) w (k) Controller u (k) z -1 z -1 -1 u (k-1) b2 Z-N algorithm z z -1 e (k) y (k) F(s) Fig. 7: The structure of controller tuned by Z-N method with neural network identification The Takahashi controller provides faster transient response then PSD with Tustin approximation, but it has notable overshoot. The PSD with Tustin approximation has derivate character otherwise. Both adaptive controllers have similar features: · Simply implementation of controller and adaptation · Fast transient response · Stability even noise at sensor · Control of complex root and integrating systems · High ability to adapt to rapid change process dynamic and process gain · Damped control value · Access to controller parameters · Non-filtered derivation cause oscillation of controller value if exist noise in control loop · The convergence speed of adaptation is decreased with increasing buffer length Set point and process output Control value Fig. 8: Control process of adapting PSD controller with derivate part filtration using Tustin approximation with neural networks identification. Transfer function of process is F2(s). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 129 3. TRANSFER FUNCTIONS F1 (s ) = 1 (5s + 1)(4s + 1) F2 (s ) = 5 (3s + 1)(4s + 1) 4. CONCLUSION The paper describes several adaptive structures based on neural networks. The standard adaptive control approach could fail because some aspects are omitted in theory. These aspects are too short sampling period, nonlinearities in control loop, highly decreased numerical precision due to used A/D and D/A converter. In that case, we could use neural network as effective substitution for standard control element. All algorithms described in this paper were implemented in ANSI C, they were validated by simulation in MATLAB and then they were implemented in PLC B&R 2005. The results of simulations are shown in figures 3, 5, 6 and 8. The output of real process controlled by algorithm implemented in PLC is shown in figure 9. Set Žádaná point and hodnota process a regulovaná output 4 veličina 2 0 2 40 1 0 2 0 3 0 4 0 5 t0 [s] 6 0 7 0 8 0 9 0 10 0 6 0 7 0 8 0 9 0 10 0 Control Akčnívalue 1 0 zásah 5 0 5 10 0 1 0 2 0 3 0 4 0 5 t0 [s] Fig. 9: Programmable logic controller B&R 2005: Control process of adapting PSD controller with derivate part filtration using Tustin approximation with neural networks identification. Transfer function of process is F1(s). ACKNOWLEDGEMENTS This research was particularly supported by GACR 102/01/1485 – Environment for Developing, Modeling and Application of Heterogeneous Systems, MSMT CEZ MSM: 260000013 – Automation of Technological and Manufacturing Processes. REFERENCES [1] OMATU, S., KHALID, M. , YUSOF, R.: NEURO-Control and its Applications, Springer, Berlin, 1996. [2] BOBÁL, V., BŐHM, J., PROKOP, R., FESSL, J.: Practical aspects of self tuning controllers: algorithms and implementation (in Czech). Vutium, Brno, 1999. [3] PIVOŇKA, P.: Digital control (in Czech). Electronic textbook, ÚAMT FEKT VUT, Brno, 2000. [4] VELEBA, V.: Adaptive neural controllers (in Czech). Diploma work, ÚAMT FEKT VUT 2003, Brno. [5] KRUPANSKÝ, P., PIVOŇKA, P.: Adaptive neural controllers and their implementation. ÚAMT FEI VUT, Brno, 2000. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 130 FUZZY INFERENCE SYSTEM AND PREDICTION Libor Žák Department of Stochastic and Non-Standard Methods, Institute of Mathematics Faculty of Mechanical Engineering, Technical University of Brno, Technická 2, 616 69 Brno, Czech Republic, Email: [email protected] Tel: +420-541142550, Fax: +420-541142527 Abstract: This paper describes the implementation of fuzzy set theory and Fuzzy Inference System (FIS) for prediction of electric load. The proposed technique utilizes fuzzy rules to incorporate historical weather and load data. The use of fuzzy logic effectively handles the load variations due to special events. The fuzzy logic has been extensively tested on actual data obtained from the Czech Electric Power Company (ČEZ) for 24-hour ahead prediction. Test results indicate that the fuzzy rule base can produce results better in accuracy than artificial neural networks (ANNs) method. Keywords: fuzzy sets, fuzzy logic, fuzzy inference system, prediction. 1. INTRODUCTION Short-term load forecasting (STLF) of power demand plays a very important role in the economic and secure operation of power systems [1]. Improvement in accuracy of load forecasts results in substantial savings in operating cost and also increase the reliability of power supply. In order to predict the future load accurately, numerous forecasting techniques have been used during the past 40 years. Fuzzy logic model has been selected as an alternative method for the load forecasting problem in this paper. It is a suitable technique in case when the historical data are not real numbers, but linguistic values [4]. This paper presents the results of a preliminary investigation of the feasibility of use of a fuzzy logic model for short-term load forecasting. In this research, historical load and weather data are converted into fuzzy set theory to produce fuzzy forecasts and defuzzification is performed to generate a point estimate for system load. 2. FUZZY INFERENCE SYSTEM Fuzzy set is one, which has vague boundaries. Fuzzy sets can successfully represent a human's ambiguous estimations. They are represented by a membership function, which takes values between 0.0 and 1.0. Fuzzy reasoning uses this membership function to describe the human's reasoning process successfully. 2.1 Description of Fuzzy Inference System Fuzzy set A is defined in terms (U, mA), where U is relevant universal set and mA: U ® á0,1ñ is a membership function, which assigns each elements from U to fuzzy set A. The membership of the element xÎU of a fuzzy set A is indicated mA(x). We call F(U) the set of all fuzzy set. Then „classical“ set A is fuzzy set where: mA: U ®{0, 1}. Thus xÎA Û mA(x) = 0 and xÏA ÛmA(x) = 1. Let Ui , i = 1, 2, ..., n, be universals. Then fuzzy relation R on U = U1 ´ U2 ´...´ Un is a fuzzy set R on universal U. Fuzzy Inference System: One of the possible applications is a fuzzy inference system (FIS) (fuzzy regulator). There are a few types of regulators. We use the regulator of type P: u = R(e) for our purposes, where the action values depend only on a regulation deviation: Input variables: Ei = (Ei, T (Ei ), Ei, G, M), i = 1,..., n. Output variables: U = (U, T (U), U, G, M). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 131 We consider the fuzzy regulator as the statement of the type:  = Â1 else Â2 else ,..., else Âp, where  k is in the form:  k = if e1 is X E1 ,k and e2 is X E2 , k and ..... and en is X En ,k then u is YU ,k , where e1ÎE1,..., enÎEn, uÎU, X Ei ,k Î T (Ei), YU ,k Î T (U) for "i= 1,...,n, for "k=1,.., p. We mark meaning of the mathematical statement  by M(Â)=R. M(Â) and the fuzzy relation on p E1´E2´...´En´U is defined by R=M(Â)= U M(Âk), where else it is considered as an union and M(Âk) is k =1 defined in the form: M(Âk) = AE1 ,k ´ AE2 ,k ´...´ AEn , k ´ AU , k . This is the fuzzy relation on E1´ E2´...´En´U, where AEi , k = M( X Ei ,k ) is the fuzzy set on universal Ei for "i= 1,…, n and AU , k = M( YU ,k ) is the fuzzy set on universal U for "k=1,…, p. Let a Ei i=1,…,n be a regulation deviance. Let a Ei be any fuzzy set on Ei. Then the dimension of action values aU is defined by term aU = ( a E1 ´ a E 2 ´...´ a E n )°R. This is a composition of fuzzy relation ( a E1 ´ a E 2 ´...´ a E n ) on universal E1´E2´...´En and relation R defined on the universal E1´...´En´U. The result of this composition is the fuzzy set on the universal U. We do not require the output of fuzzy regulator to be a set in many cases, but we require the concrete value z0ÎZ, e.i. we want to make a defuzzification. The centroid method is the most frequently used method of defuzzification. The FIS defined in such a way is called Mamdami. Fuzzy rule base Nonfuzzy input Fuzzification interface Defuzzification interface Nonfuzzy output Fuzzy interface machine Fig. 1: Block diagram of a fuzzy inference system. 2.2 Description FIS for prediction of load The four principal components of a fuzzy system is shown in Figure 1. The fuzzification interface performs a scale mapping that changes the range of values of input variables into corresponding universe of discourse. It also performs fuzzification that converts nonfuzzy (crisp) input data into suitable linguistic values, which may be viewed as labels of fuzzy sets. Fuzzy rule base, which consists of a set of linguistic control rules written in the form “If a set of conditions are satisfied, Then a set of consequences are inferred”. Fuzzy inference machine, which is a decision-making logic that employs rules from the fuzzy rule base to infer fuzzy control actions in response to fuzzified inputs. Defuzzification interface performs a scale mapping that converts the range of values of output variables into corresponding universe of discourse. It also performs defuzzification that yields a nonfuzzy (crisp) control action from an inferred fuzzy control action [2]. A commonly used defuzzification rule known as centroid method is used here, according to which the defuzzification interface produces a crisp output defined as the center of gravity of the distribution of possible actions. This centroid approach produces a numerical forecast sensitive to all the rules. Fuzzy logic is a tool for representing imprecise, ambiguous, and vague information. Fuzzy set operations are grounded on a solid theoretical foundation, although they deal with imprecise quantities, they do so in a most precise, and well-defined way. Fuzzy operations that act on the membership functions lead to INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 132 ExL 1 VL L M H VH ExH Degree of membership Degree of membership consistent and logical conclusion [3]. If we use appropriate membership function definitions and rules, we can achieve useful results. One of the most useful properties of the fuzzy set approach is that contradictions in the data need not cause problems. Fuzzy systems are stable, easily tuned and can be conventionally validated. Designing of fuzzy sets is very easy and simple. Abstract reasoning and human-like responses in cases involving uncertainty and contradictory data are the main properties of fuzzy systems [4]. 0.8 0.6 0.4 0.2 VC 1 ExC C N W H VH ExH 0.8 0.6 0.4 0.2 0 0 4000 6000 8000 Input Load [MW ] 1000 0 -10 0 10 20 Temperature [°C] 30 Fig. 2: Triangular type membership functions for input load, input temperature. ExL VL L M H VH ExH ExC 1 VC C N W H VH ExH 0.8 0.8 Degreeof membership Degree of membership 1 0.6 0.4 0.2 0.6 0.4 0.2 0 0 4000 5000 6000 7000 8000 Input Load [MW] 9000 -15 10000 -10 -5 0 5 10 15 20 Temperature[°C] 25 30 Fig. 3: Gaussian curve membership function for input load and input temperature. ExL VL 1 L M H VH ExH 1 0.8 VC C N W H VH ExH Degree of membership 0.8 0.6 Degreeofmembership ExC 0.6 0.4 0.4 0.2 0.2 0 0 4000 5000 6000 7000 8000 InputLoad[MW] 9000 10000 -15 -10 -5 0 5 10 15 20 25 30 Temperature [°C] Fig. 4: Trapezoidal membership function for input load and input temperature. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 133 2.3. Fuzzy rules for prediction of load Heuristic and expert knowledge are often expressed linguistically in the form of If-Then rules. These rules can be extracted from common senses, intuitive knowledge, general principles and laws, and other means that reflect actual situations. For example “If the temperature is extremely cold, Then load demand will be very high” is an example of logic statement. The input temperature is sorted into eight categories and labeled as Extremely Cold (ExC), Very Cold (VC), Cold (C), Normal (N), Warm (W), Hot (H), Very Hot (VH), and Extremely Hot (ExH). The input and output load is sorted into seven categories and labeled as Extremely Low (ExL), Very Low (VL), Low (L), Normal (N), High (H), Very High (VH), and Extremely High (ExH) as shown in Figure 2. A tentative list of input and output variables using statistical analysis, and engineering judgements was compiled. The input and output variables within the [0, 1] region was normalized. The next step was the selection of the shape of the fuzzy membership for each variable. This is purely arbitrary, but one usually starts with a particular shape of membership function and changes it if the forecasting accuracy is not good. The triangular shaped mappings shown in Figure 2 are very common, however, alternative shapes, such as Gaussian and trapezoidal curve with different amounts of function overlapping, were also used as shown in Figure 3 and 4, respectively. VL L M H VH ExH Output Load [MW] Degree of membership 1 ExL 0.8 0.6 0.4 0.2 6000 8000 8000 6000 4000 30 20 10 0 Temperature [°C] -10 0 4000 10000 10000 8000 4000 10000 6000 Input Load [MW] Output Load [MW] Fig. 5: Triangular type membership functions for output load and FIS surface for triangular input. 3. TEST RESULTS Two years of historical load and temperature data were used, one year (1998) for the fuzzy rule base design and the following year (1999) for testing the model performance. The inputs used to forecast hourly load for day (t) are; hourly load of the previous day (t-1) and minimum, maximum and average temperature for the day (t). In case, the accuracy is un-satisfactory, then the number of fuzzy membership functions and/or shape of the fuzzy membership functions can be changed and a new fuzzy rule base is obtained. The iterative process of designing the rule base, choosing a defuzzification algorithm, and testing the system performance may be repeated several times with a different number of fuzzy membership functions and/or different shapes of fuzzy memberships. The fuzzy rule base that provides the minimum error measure for the test set is selected for realtime forecasting. The quality of prediction of heat consumption has been calculated according to average error MAPE. Let (R1, R2, …, Rk) be the real values of time series and ( P1, P2, …, Pk ) are the predicted members of time series where k is the number of searched members. Thus æ k ö MAPE = 1 k çç å (abs( Pi - R i ) / R i ) ÷÷ . è i =1 ø INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 134 MAP = max (abs( Pi - R i ) / R i ) . i=1,..., k Using different membership functions, the mean absolute percentage error (MAPE) and maximum absolute percentage error (MAP) for working days of the week, weekend and special days are computed as given in Tables 1, 2 and 3, respectively. Similarly, MAPE for each day of the week during 3rd week of January 1999 using fuzzy logic and back-propagation neural network is also shown in Figure 6. Working Days Monday Tuesday Wednesday Thursday Friday Membership Functions Gaussian Curve Triangular Trapezoidal MAPE (%) MAP (%) MAPE (%) MAP (%) MAPE (%) MAP (%) 2.8 1.3 0.99 0.88 0.89 6.1 4.7 2.7 2.3 2.4 2.3 1.1 1.4 1.8 1.6 8.8 3.9 4.9 4.0 4.3 2.2 1.4 1.2 1.1 1.2 4.8 5.4 3.5 2.5 3.2 Table 1: MAPE and MAP for working days in 3rd Week of January 1999 using various membership functions. Weekend Days Saturday Sunday Membership Functions Gaussian Curve Triangular Trapezoidal MAPE (%) MAP (%) MAPE (%) MAP (%) MAPE (%) MAP (%) 3.6 3.1 8.5 7.6 5.4 4.9 7.8 8.6 6.0 6.2 8.6 9.8 Table 2: MAPE and MAP for 3rd Weekend in January 1999 using various membership functions. Special Days New year day Easter Labor day Victory day Yanhose day Independence day Students’ day Christmas eve Christmas 1st day Christmas 2nd day Triangular Membership Functions Gaussian Curve Trapezoidal MAPE (%) MAP (%) MAPE (%) MAP (%) MAPE (%) MAP (%) 2.6 3.0 2.6 3.2 2.0 1.6 2.6 3.1 3.4 3.2 8.0 7.1 7.8 7.9 4.1 4.4 5.1 5.6 8.7 9.8 3.4 3.0 4.0 1.3 4.8 2.2 1.8 3.6 4.6 4.2 7.6 8.1 9.2 4.1 4.4 5.0 6.8 5.1 8.9 7.9 3.8 2.7 3.3 4.0 2.2 4.8 2.7 1.7 4.9 5.1 7.4 6.8 9.2 9.9 5.6 8.5 5.4 4.4 7.3 10.1 Table 3: MAPE and MAP for special days/holidays during the year 1999 using various membership functions. To compare the performance of the fuzzy logic model, a back-propagation network (BPN) was also developed. A separate BPN with only one hidden layer and 24 hidden neurons was used for each day using the same inputs and single output as the fuzzy logic model. The training and testing of BPN were performed with the same training and testing data sets as mentioned earlier. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 135 F u zzy BPN 4 3 .5 3 MAPE [%] 2 .5 2 1 .5 1 0 .5 0 M onday Tuesday W ednesday T h u rs d a y F rid a y S a tu r d a y S unday D a y s o f th e W e e k Fig. 6: MAPE for each day of the Week during 3rd Week of January 1999 using fuzzy logic and back propagation neural network. 4. CONCLUSIONS A forecasting technique based on fuzzy logic approach has been presented in this paper. This approach can be used to forecast loads with better accuracy than ANN technique. The flexibility of the proposed approach, which offers a logical set of rules, readily adaptable and understandable by an operator, may be a very suitable solution to the implementation and usage problem that has consistently limited the adoption of STLF models. This technique is simple to implement on a personal computer and allows for operator intervention. Various membership functions have been discussed, and for the particular application data sets, their effects on model performance have been demonstrated. The proposed model has been able to generate forecasts with a MAPE frequently below 2.8 % for working days, 3.6 % for weekends and 3.4 % for special days. The simulation results demonstrate the effectiveness of the fuzzy model for 24-hour ahead prediction. REFERENCES: [1] [2] [3] [4] [5] [6] YANG HONG-TZER, HUANG CHAO-MING: “A new short-term load forecasting approach using self-organizing fuzzy ARMAX models”, IEEE Transactions on Power Systems, Vol. 13, No. 1, (1998), 217-225. HAYKIN, S.: “Neural networks _ A comprehensive foundation“, Macmillan College Publishing Company, New York, 1994. MASTERS, T.: “Practical neural network recipes in C++, Academic Press Inc., 1993. KHAN, M.R., ŽÁK, L., ONDRŮŠEK, Č.: "Forecasting Weekly electric load using a hybrid fuzzyneural network approach", submitted for publishing in Engineering Mechanics _ International Journal of theoretical and applied mechanics, Technical University of Brno, Czech Republic, December 2000. KHAN, M.R., ŽÁK, L., ONDRŮŠEK, Č.: Fuzzy Logic Based Short_Term Electric Load Forecasting, ELEKTRO 2001, 4th International Scientific Conference, Žilina, 2001, pp 19-25, ISBN 80-7100-836-2. ŽÁK, L. Fuzzy Objects and Fuzzy Clustering. Proc. 8th Zittau Fuzzy Colloquium, Zittau, 2000, pp 293 – 302, ISBN 3-00-006723-X. ACKNOWLEDGEMENT The paper was supported by research design CEZ: J22/98:261100009 “Non-traditional methods for investigating complex and vague systems” and the Grant Agency of Czech Republic under grant No: 101/01/0345. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 136 EVOLVING CONTROLLERS Pavel Ošmera 1), Uday K. Chakraborthy2), Imrich Rukovanský 3) 1) 2) 3) Brno University of Technology, Technická 2, 616 69 Brno, Czech Republic E-mail: osmera @uai.fme.vutbr.cz Dept. of Math & Computer Sc., Univ. of Missouri, St. Louis, MO 63121, USA E-mail: [email protected] European Polytechnical Insstitute Kunovice, Osvobozeni 699, 686 04 Kunovice, Czech Republic Tel. 572 548 035, Fax: 572 549 018, E-mail: [email protected] Abstract: The adaptive significance of genetic algorithms (GAs) with diploid chromosomes and an artificial immune system has been studied. An artificial immune system was designed to support a parallel evolutionary algorithm. We implemented hybrid and parallel genetic algorithms for design of evolvable controllers. A flexible structure with PID controllers can be designed by parallel evolutionary algorithms. The adaptive significance of parallel GAs and the comparison with standard GAs are presented. Keywords: artificial immune systém (AIS), parallel genetic algorithms, neural networks, evolving controllers, adaptive systems, immune network model 1 INTRODUCTION Darwin suggested that slight variation among individuals significantly affects the gradual evolution of the population. He called this differential reproductive process of varying individuals natural selection. Gregor Mendel (1822-1884) accurately observed patterns of inheritance and proposed a mechanism to account for some of the patterns. Genes determine individual traits. Various kinds of offspring appear in proportions that can be predicted from Mendel’s laws. We often use the term Mendelian genetics (see memory I in Figure 1) to refer to the most basic patterns of inheritance in sexually reproducing organisms with more than one chromosome. Mendel presented his classic paper Experiments in Plant Hybrids for Natural Science Society in Brno in 1865. Mendel observed that the spherical seed trait was dominant, being expressed over the dented seed trait, which he called recessive. In diploid organisms, chromosomes come in pairs (memory Ia and memory Ib in Fig. 1). Dawkins [3] described organisms as vehicles for the genes, built to carry them around and protect them. Pictures, books, tools and buildings are meme vehicles (memory III). In our own bodies, thousands of genes cooperate to produce organs and to result in a machine that effectively carries all genes around inside it. However, when we examine a biological cell or an organism, the situation is quite different: not only are these systems open, but they exist only because they are open. The feed of the flux of matter and energy is coming to them from the outside world. The free energy E can create P different organisms (species), every species with Ni copies that are created by DNA information Ii , where Qin_i is the metabolic heat inside of an organism released by their activities, Qout is the metabolic heat that is lost. Ti is the temperature, mi is a mass, Vi is a volume, W is a scrap (waste) and c1, c2 are constants [7, 8]: E = E stru cture + E heat = P å (c N V o i =1 1 i i i + Q in _ i ) + W + Q o ut ,where Qin _ i = c2 N i miTi (1) oi, is the measure of energetic order (defined by the energetic density = Estruxture_i/Vi). On the Earth there appeared a complex biosphere with the food chains that must satisfy (1). An increase of energy DE is covered by sun or earth activity. In a far-from-equilibrium condition, various types of self-organization processes may occur. The INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 137 goal of evolution (fitness Fe) is to maximize the accumulated energy in living systems that have a complex evolution structure (see Fig.1): ì P ü Fe = max through _ evolution íc1.å NiVi oi ý î i =1 þ (2) A living system appears very complex from the thermodynamic point of view. Certain reactions are close to equilibrium, while others are not. Open systems evolve to higher and higher forms of complexity. Molecular biology shows that not everything in a cell is alive in the same way. Some processes reach equilibrium and others are dominated by regulatory enzymes far from equilibrium [4]. 2 SELF-ORGANIZATION AND ADAPTATION OF COMPLEX SYSTEMS In dynamical systems, transitions can be found: order, complexity, and chaos. Analogously, water can exist in solid, transitional, and fluid phases [4, 8]. In nonlinear systems, a chaos theory tells us that the slightest uncertainty in our knowledge of initial conditions will often grow inexorably, and our predictions are nonsense. Complex adaptive systems share certain crucial properties (non - linearity, complex mixture of positive and negative feedback, nonlinear dynamics, emergence, collective behavior, spontaneous organization, etc.). In the natural world, such systems include brains, immune systems, ecology, cells, developing embryos, and ant colonies. In the human world, they include cultural and social systems. Each of these systems is a network of a number of “agents” acting in parallel. In a brain, the agents are nerve cells; in ecology, the agents are species; in a cell, the agents are organelles such as the nucleus and the mitochondria; in an embryo, the agents are cells, and so on. Each agent finds itself in the environment produced by its interactions with the other agents in the system. It is constantly acting and reacting to what the other agents are doing. There are emergent properties, the interaction of a lot of parts, the kinds of things that the group of agents can do collectively; something that the individual cannot. There is no master agent - for example - a master neuron in the brain. Complex adaptive systems have a lot of levels of organization (hierarchical structures), with agents at a given level serving as building blocks for agents at a higher level. The immune system is governed by local interaction between cells and antibodies, there is no central controller in distributed control. Similar behavior can found in the development of the Internet. We can use biological laws to describe the development of the Internet. It is no wonder that complex adaptive systems (with multiple agents, building blocks, internal models, and perpetual novelty) are so hard to analyze using standard mathematics. We need mathematics and computer simulation techniques that emphasize internal models, emergence of new building blocks, and a rich web of interactions among multiple agents. We now have a good understanding of chaos and fractals showing how simple systems with simple parts can generate very complex behaviors. The edge of chaos is a special region onto itself, the place where you can find systems with lifelike, complex behavior. Living systems are actually very close to this edge-of-chaos phase transition, where things are much looser and more fluid. A natural selection is not an antagonist of selforganization. It is a force that constantly pushes emergent, self-organizing systems towards the edge of chaos from a chaos area. Mutation and crossover are opposite forces pushing the systems from an order to a chaos areas. Evolution always seemed to lead to the edge of chaos. The complex evolutionary structure described in [7, 8] can be transformed to the structure of the computational intelligence (see Fig. 2). A random genetic crossover or mutation may give a species the ability to run much faster than before. The agent starts changing, then it induces changes in one of its neighbors, and finally you get an avalanche of changes until everything again stops changing. Systems get to the edge of chaos through adaptation: each individual (agent) tries to adapt to all the others. Co-evolution can also get them there; the whole system coevolves to the edge of chaos. In ecosystems or ecosystem models, three regimes can be found: ordered regime, chaotic regime, and edge-of chaos like a phase transition. When the system is at the phase transition, then - of course - order and chaos are in balance. There is an evolutionary metadynamics, a process that would tune the internal organization of each agent so that they all reside at the edge of chaos. The maximum fitness occurs right at the phase transition. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 138 3 AN ARTIFICIAL IMMUNE SYSTEM The immune system is a complex, self-organizing and highly distributed system that has no centralized control and that uses learning and memory when solving particular tasks. An artificial immune system (AIS) fully exploits self-organizing properties of the vertebrate immune system. The biological immune system is an efficient natural protection system whose primary function is to generate multiple antibodies from the antibody gene libraries and to keep them alive even if the unknown foreign pathogen infects the body. The AIS has several desired features for optimization purposes, such as robustness, flexibility, learning ability and memory. The AIS is self-organizing, since it determines survival of newly created clones, and it determines its own size. This is referred to as meta-dynamics of the system. These characteristics are often useful in scheduling problems. Two distinct algorithms have emerged as successful implementations of artificial immune systems: the immune network model described by Jerne and the negative selection algorithm developed by Forrest. The immune system implements two types of selection [11] – negative selection and clonal selection. These two processes have to be dealt with separately by researchers. Negative selection, which operates on lymphocytes maturing in thymus (called T-cells), ensures that these lymphocytes do not respond to self-proteins. This is achieved by killing any T-cell that binds to a self protein while maturing. The second selection process, called clonal selection, operates on lymphocytes that have matured in the bone marrow (called B-cells). Any B-cell that binds to a pathogen is stimulated to copy itself. The copying process is subject to high probability of errors (“hypermutation”). The combination of copying with mutation and selection amounts to an evolutionary algorithm that gives rise to B-cells that are increasingly specific to the invading pathogen. If the same or similar pathogens invade in the future the immune system will respond much more quickly because it maintains a memory of successful responses from previous infections. We used a constraint-handling approach based on emulation of an immune system (particularly, using the negative selection model) that was incorporated into a parallel GA [11]. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 139 ecosystems Kauffman’s enzyme nets Dawkins’ theory of memes memory III (memes of mankind ) culture social environment books & school & television & newspaper & Internet & religions & .. laws & morals & society & friends & enemy & terrorism & .. & family mother & father & wife/husband & sexual partner & children & .... Lamarckian imitation of memes memes memů body carrier of genes imunitní systém hormones Baldwin effect individual (descendant) brain memory II carrier of memes chemical messenger immune system learning behavior (rules) (instincts) instinctive behavior conscious behavior central nervous system mitochondrial genes (epigenetic information), order of cells (epigenetic structure) memory Ia genes for structure of the body memory Ib genes for structure of the brain and instincts diploid chromosomes & sexual reproduction integrated fitness Mendelian genetics memory IVa Darwinian selection process parasites memory IVb prey-predator interaction (living part of nature) flow of memes environment (energy of sun and earth....) influence of genes direction of influence Fig. 1 Complex evolutionary structure INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 140 Computational Intelligence Game Theory and Collective Behavior Intelligence memory III (genome & memes) social environment culture Particle Swarm Optimization, Multi-Agent EA Agent-based Multiobjective Optimalization Ant Colony Optimization, Dynamical Systems Modeling Niches, Team Optimization Culture Algorithms family Evolutionary Computation: Evolution Strategies Genetic Programming Genetic Algorithms, Parallel GAs body individual (descendant) brain memory II Hormone Systems Clonal Selection Enzyme Behavior Artificial Immune Algorithms Fuzzy Systems Fuzzy-rough Sets Hybrid Learning Neural Networks Fuzzy-neural Modeling Intelligent Control Mitochondrial Systems memory Ia genes for structure of the body memory Ib genes for structure of the brain and instincts Diploid GA with Sexual Reproduction DNA Computing , Messy GA memory IVb memory IVa Cooperative co-evolutionary Algorithms Parasitic Optimization, Bacterial EA Artificial Life Systems, Differential EA Parallel Hierarchical EA, Meta-Heuristics Evolutionary Multi-objective Optimization Evolutionary Design , Robotics Evolvable Hardware, Embryonic Hardware Human-Computer Interaction Molecular-Quantum Computing Data Mining, Chaotic Systems, Scheduling Fig. 2 Soft Computing 4 DESIGN OF EVOLVABLE CONTROLLERS In recent years, there has been growing interest in using intelligent approaches such as fuzzy, neural network, evolutionary methods, and their combined technologies for the PID controller [13], [14]. The Proportional-Integral-Derivative (PID) controller has been widely used owing to its simplicity and robustness in chemical processes, power plants, and electrical systems. Its popularity is also due to easy implementation in hardware and software. However, with only P,I,D parameters, it is very difficult to control a plant with complex dynamics, such as large dead time, inverse response and highly nonlinear characteristics. That is, since the PID controller is usually poorly tuned, a higher of degree of experience and technology is required for tuning in the INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 141 actual plant. Also, a number of approaches have been proposed to implement mixed control structures that combine a PID controller with fuzzy logic. GAs can be used for the optimal design of a fuzzy controller. The GA chooses the number and position of membership functions and modifies the rules. The fuzzy method does not require professional expertise or mathematical analysis of the plant model. Unfortunately fuzzy controllers are still difficult to implement on commercial low-cost microcontroller, and may not offer significant improvements in terms of robustness and performance. The use of the classical PID controllers is preferred whenever non-linear techniques are not strictly required. A GA is used to optimize both the structure and the associated parameters of controllers. We used several versions of GAs: Standard GA – One population with population size (= number of individuals) = 40. Individuals are sorted by fitness. The higher the fitness, the higher is the probability that the individual will be selected as a parent (a roulette wheel mechanism). The second parent is selected in the same manner. After that, crossover, mutation, and correction are applied. The reproduction process is repeated until the worse half of population is replaced. GA with two sub-populations (the sexual approach where the male and the female are distinguished) -Each sub-population has 20 individuals. Individuals are sorted by fitness. The first parent is selected from the male sub-population while the second parent is selected from the female sub-population. The selection probability of the first parent is determined from a uniform distribution, while the selection probability of the second parent is obtained using a modified roulette wheel approach that more often prefers better individuals. Crossover, mutation, and correction operators are then applied [12]. The reproduction is repeated until the worse half of every sub-population is replaced. The GA with two sub-populations is a special case of the parallel GA with two sub-populations (male and female) and sexual selection. Sexual recombination generates a variety of genotypes that increase the evolutionary potential of the population [2,7]. As it increases the variation among the offspring produced by an individual, it improves the chance that some of them will be successful in the varying environment. Mutation, genetic drift, migration, nonrandom mating, and natural selection cause changes in variation within the population. The adaptation of GAs depends on the speed of landscape changes through time. A parallel GA can have a two - level structure. The first level is created by several populations with different GAs [7], [8]. The best or random individuals from the first level are sent to the second level. At this level, the standard GA with elitism runs. This two-level structure allows us to find a better solution than that found by GAs in the first level; the best solution from the first level can never be lost but only overtaken in the second level. We tested the following modifications of GAs: Limited lifetime – an individual is not removed (replaced) until the minimum lifetime is reached [2]. The minimum lifetime is randomly generated when the individual is born. Individuals can survive for several generations even if they are not good. They have an opportunity to improve their fitness during the lifetime. This approach can slow down the evolution process, but also prevents the loss of potentially good individuals. HC operator – a hill climbing approach is used. In randomly selected individuals and in randomly selected genes, several small modifications are carried out; the best modification with the best fitness is retained for further use. Adaptive version of the GA with artificial immune system (as in [11]) -- We have used a constrainthandling approach based on emulation of immune system (particularly, using the negative selection model) that was incorporated into a parallel GA. Comments on figures 3 and 4: G_ standard genetic algorithm S_ GA with diploid chromosomes (with sexual reproduction, with two sub-populations) H_ hill climbing approach is applied L limited lifetime is applied A adaptive version of GA with artificial immune system is applied P parallel structure of GA is applied INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 142 250 500 G_H 450 G_LH 200 400 G_HA G_LHA 350 G 300 S_H 150 G_L S_H G_H 250 S_LH G_LH 200 S_LH 100 G_HA S_HA G_LHA 150 S_HA S_LHA 50 100 S_LHA P1 50 P2 0 0 1 10 19 28 37 46 55 64 73 82 1 91 100 Fig. 3: The comparison of standard methods 5 9 13 17 21 25 29 33 37 41 45 49 53 57 Fig. 4: GAs with parallel architecture 5. DISCUSSION The artificial immune system and sexual reproduction are typical of complicated creatures. Parallel GAs with the sexual reproduction and an artificial immune system can increase the efficiency and robustness of systems, and thus they can track better optimal parameters in a changing environment. From the experimental session it can be concluded that modified standard GAs with two sub-populations can design controllers much better than classical versions of Gas do. The HC modification of the standard GA (see Fig.3) can improve the speed of the convergence, but the lifetime modification can support the higher variety of population. In the beginning, the HC modifications have better results in most cases, but the lifetime modifications have slower convergence; in the end, they can find better solution than the HC modifications. It is not easy to say which individual modifications are the best (see Fig. 4). If we join them together by the parallel GA, then - in the higher level - it is not important which method will contribute more to the final solution. Diploid chromosomes can increase the efficiency and robustness of GAs [1, 7], and they can better track optimal parameters in a changing environment. It can be demonstrated that standard GAs with haploid chromosomes are unable in many cases to correctly locate optimal solutions for time-dependent objective functions. Acknowledgments: This work has been supported by MŠMT grant No: CZ J22/98 260000013 REFERENCES [1] OŠMERA, P.: An Application of Genetic Algorithms with Diploid Chromosomes, Proceedings of MENDEL’98, p. 86 – 89, Brno, Czech Republic, 1998. [2] OŠMERA, P. - Roupec, J.: Limited Lifetime Genetic Algorithms in Comparison with Sexual Reproduction Based GAs, Proceedings of MENDEL’2000, Brno, Czech Republic, 2000. [3] DAWKINS R.: The Selfish Gene, Oxford Univrsity Press, Oxford 1976 [4] KAUFFMAN, S.A.: Investigations, Oxford University Press, New York 2000 [5] PRIGOGINE, I., STENGERS, I.: Order out of Chaos, Flamingo 1985 [6] OŠMERA, P.: Complex Adaptive Systems, Proceedings of MENDEL’2001, Brno, Czech Republic (2001) 137 – 143 [7] OŠMERA, P.: Complex Evolutionary Structures, Proceedings of MENDEL’2002, Brno, Czech Republic (2002) 109 – 116. [8] OŠMERA, P.: Genetic Algorithms and their Aplications, the habilit work, in Czech language (2002) 3 – 114. [9] WALDROP, M.M: Complexity – The Emerging Science at Edge of Order and Chaos, Viking 1993 [10] DAVIS, L.: Job Shop Scheduling with Genetic Algorithms, International conference ICGA’85 (1985) 132138 [11] COELLO, C.A., CORTÉS, N.C.: A Parallel Implementation of an Artificial Immune System to Handle Constrain in Genetic Algorithms, WCCI 2002, 819-824, Hawai INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 143 [12] OŠMERA, P., MASOPUST, P.: Schedule Optimization Using Genetic Algorithms, Proceedings of MENDEL’2002, Brno, Czech Republic (2002) 109 – 116. [13] CUPERTINO, F., NASO, D., SALVATORE, L., TURCHIANO, B.: Design of Cascaded Controllers for DC Drivers using Evolutionary Algorithms, Proceedings of CEC’2002, Honolulu, USA (2002) 309 – 316. [14] KIM, H. D., HONG, P.W., PARK, J.I.: Auto-tuning of Reference Model Based PID Controller Using Immune Algorithm, Proceedings of CEC’2002, Honolulu, USA (2002) 509 – 516. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 144 AN IMPROVED ALGORITHM BASED ON STOCHASTIC HEURISTIC METHODS FOR SOLVING STEINER TREE PROBLEM IN GRAPHS Daniel Smutek VUT v Brně, Fakulta strojního inženýrství, Ústav automatizace a informatiky, Technická 2, 61669 Brno e-mail: [email protected] Abstract: The Steiner Problem in a Graph (SPG) is one of the classic problems of combinatorial optimization. Given a graph and designated subset of the vertices, the task is to find a minimum cost subgraph spanning the designated vertices. In this paper we present an improved algorithm based on stochastic heuristic methods (tabu search, simulated annealing and genetic algorithm) which approximate SPG instances up to 500 vertices in a feasible time. Computational results are given for randomly generated problems from OR-Library having 500 vertices. The SPG arises in a large variety of diverse optimization problems such as network design, multiprocessor scheduling and integrated circuit design. Keywords: heuristic, Steiner tree, minimum spanning tree, graph 1 INTRODUCTION Steiner Tree Problem in a Graph (SPG) is a variant of the minimal spanning tree problem. Given a connected, undirected graph G = (V, E) with a vertex set V, an edge set E, a positive weight function defined on its edges and a non-empty set C Í V of terminals. The problem is to find a shortest tree interconnecting C. The tree satisfying these properties is a Steiner minimum tree (SMT) or Steiner tree (see Fig. 1). Vertices in C are called terminal vertices or customer vertices. All terminal vertices must be in the Steiner tree. Other vertices in V that aren’t terminal vertices, are called Steiner points or optional vertices and they are in set S. Steiner points can be in the Steiner tree, but they don’t have to. The SPG has been shown to be NP–hard [11] for general graphs. As a result, the existing exact methods can only solve moderately sized instances and heuristic approaches are required to tackle larger instances likely to be encountered in real–life applications of the problem. Since the SPG can be designed as a combinatorial problem, we can use stochastic heuristic methods such as genetic algorithms (GA), simulated annealing (SA) or tabu search (TS). The purpose of this paper was to present an improved algorithm for the SPG that is based on three steps. First step is a graph reduction, the second step is using of the stochastic heuristic and the final step is a simple deterministic heuristic. A computer program was created to test efficiency of the designated algorithm. Then the graphs with 500 vertices presented in OR-Library were tested by the program. We will compare also our algorithm with two well-known approximation algorithms (Shortest paths approximation (SPA) created by Takahashi and Matsuyama [20] and the distance network approximation (KMB) created by Kou, Markowsky, and Berman [13]). This work is the continuation of the previous work which solved the SPG with the GA, SA [19] and TS [18] for graph instances up to 100 vertices only. 1 1 3 ...... Terminal vertices 1 10 8 2 13 8 3 7 2 6 14 4 ..... Steiner points 4 7 8 9 Fig. 1: Steiner minimum tree (bold lines) in a graph INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 145 2 GRAPH REDUCTIONS Before the GA, SA or TS itself is executed, an attempt to reduce the size of the given problem is performed using standard graph reduction techniques. Routine GraphReductions of Figure 2 performs three kinds of rather simple reductions all of which are described in [21, 22]. More elaborate reductions as well as proofs of the correctness of the reductions used here can be found in [2]. Let evw denote the edge between vertices v and w, and let sp(v,w) Í E denote the shortest path between vertices v and w. The three reductions used are following: a) Any Steiner vertex of degree 1 can be removed along with the edges incident with it. b) If v Î V – C, deg(v) = 2 and euv, evw Î E, then v, euv and evw can be deleted from G and replaced by a new edge between u and w of equivalent cost. More specifically, if euw Ï E, then E = E È {euw} and c(euw) = c(euv) + c(evw). If there is an edge from u to w already, i.e., euw Î E, then c(euw) = min { c(euw), c(euv) + c(evw)}. c) If evw Î E and c(evw) > c(sp(v, w)), then no SMT can include evw, which, therefore, can be deleted. To obtain the largest possible overall reduction of G, the above reductions are performed repeatedly [3] as described bellow. Knowledge of the cost of a shortest path is required whenever a reduction of type c is performed. We can find a shortest path with e.g. Dijkstra’s algorithm [1]. The chosen scheme for performing reductions in the routine GraphReductions is shown in Figure 2. The routine GraphReductions terminates when no reduction of any type succeeded for a complete iteration, i.e., when no reduction can reduce G further. repeat repeat Reduction(a); until no improvement in one iteration; Reduction(b); Reduction(c); until no improvement in one iteration; Fig. 2: Outline of routine GraphReductions 3 STOCHASTIC HEURISTIC METHODS 3.1 Tabu search The general principles of tabu search algorithms were first described by Glover [5, 6, 7, 8] and, independently, by Hansen and Jaumard [9]. Given a search or solution space defining possible solutions, the basic idea of the approach is to explore this search space by moving at each step from the current solution to the best one in its "neighborhood". A neighborhood is defined as a set of possible solutions that are found by applying an appropriate transformation of the current solution. A key feature of tabu search is that it allows moves resulting in a degradation of the objective function, thus avoiding the trap of local optimality. One of the fundamental objectives of the method is to prevent the search from cycling. This is accomplished by forbidding the choice of moves to recently encountered solutions or moves that "reverse" the effect of recent moves. Such moves are said to be "tabu". The search can be intensified on some specific types of solutions or it can be diversified to previously unexplored portions of the solution space. 3.2 Simulated Annealing Simulated Annealing is an optimization method applicable to searching for the global minimum of the cost function. Annealing denotes a physical process, where the solid is first heated to a high temperature and then cooled slowly down to the original temperature. Annealing gives the system the opportunity to jump out of local minima with a reasonable probability while the temperature is still relatively high. The success of this method depends especially on choosing starting temperature and rapidity of cooling. This method is described e.g. in [17]. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 146 3.3 Genetic algorithms Genetic algorithms [17] are based on natural evolution. The algorithm maintains a population of individuals, each of which corresponds to a specific solution to the optimization problem at hand. Starting with a set of random individuals, a process of evolution is simulated. The main components of this process are crossover, which mimics propagation, and mutation, which mimics the random changes occurring in nature. After a number of generations, highly fit individuals will emerge corresponding to good solutions to the given optimization problem. A phenotype is the physical appearance of an individual, while a genotype is the corresponding representation or genetic encoding of the individual. Crossover and mutation are performed in terms of genotypes, while fitness is defined in terms of phenotypes. For a given genotype, the corresponding phenotype is computed by a decoder. One of the most important steps in designing GA is to specify the representation of Steiner tree by a chromosome (see Fig. 3). 4 IMPLEMENTATION OF HEURISTIC METHODS 4.1 The Solution Space For all of presented heuristic methods (GA, SA, TS) the solution is determined by a set of Steiner points S thus this solution can be represented by a bit–string, where 1 corresponds to whether the Steiner vertex is included in the solution tree. The length of the bit–string is equal to the number of Steiner points (see Fig. 3). 0 1 1 0 1 0 1 1 1 0 1 0 0 0 1 0 The number of Steiner points Fig. 3: Representation of the solution space by a bit–string 4.2 The Bit–string decoder If the representation of the solution space is made, we have to build a decoder which computes minimal Steiner tree. For that purpose we use the procedure which decodes the bit–string, returns the Steiner tree and the cost of this tree. The decoding procedure works as follows (see Fig. 4): 1) Assign Steiner points from an input graph G to corresponding bits in the bit–string. If the value of the bit is 0, then set an attribute of the corresponding Steiner point to FALSE. 2) Remove all Steiner points with attribute FALSE and remove all edges that are incident with them too. We obtain the subgraph F, which is made from terminal vertices and from the rest of Steiner points. 3) Search through the subgraph F with the breadth-first search algorithm [1] and determine if the spanned tree is connected over all terminal vertices. · If the subgraph F is disconnected between terminal vertices (e.g. terminal vertex is standalone), then value of the corresponded bit–string is evaluated with the penalty function. In other case we can continue with the next step. · Make the minimal spanning tree K(F) from the graph F using minimal spanning tree algorithm. In our case we used Kruskal’s algorithm (e.g. in [1]). The value of the corresponded bit–string is set to the weigh of K(F). In this case K(F) is the Steiner tree. If K(F) is minimal from all possible combinations of spanning trees, then K(F) is the minimal Steiner tree. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 147 Input graph G … Terminal vertices … Steiner points The bit–string: 9 4 1. Assign values from the bit–string to Steiner points. 2 0 1 8 14 2 10 1 5 3 5 20 3 6 7 Steiner points 3. Spanning tree K(F) (bold lines) is the minimal Steiner tree. 2 1 1 8 8 5 4 5 Fig. 4: Decoding procedure (part one) 2 2 7 5 3 10 1 6 2 1 7 2. Removal of Steiner points with attribute FALSE and all incident edges. Subraph F is build. 10 1 6 2 1 1 2 20 5 3 2 7 5 20 6 3 3 Fig. 4: Decoding procedure (part two) However this scheme does not ensure that every bit string corresponds to a valid solution. This is a disadvantage of the decoding procedure. Algorithms based on this procedure are practicable used only on the graphs where the number of vertices is between 50 and 100 [19]. In the next part of this paper we will show, how we can correct this disadvantage. 4.3 Correction techniques In our algorithm we used four correction techniques: If the terminal vertex has degree = 1 (i.e. the terminal vertex is incident with only one edge) and his neighbor is a Steiner vertex, then the Steiner vertex must be in minimal Steiner tree. Thus we add number 1 into the corresponding position of the bit–string. This step is performed always, when a new bit–string is created. 2) If the initial solution (i.e. bit–string) is penalized, then we find the Steiner vertex with the maximal number of the incident edges. Then we add number 1 into the corresponding position (corresponding to the Steiner point) of the bit string and we recalculate the value of the bit string with decoding procedure. This step is performed until the initial solution (or initial chromosome) is valid. This step is performed only at generating of the initial solution. 3) If the initial solution (i.e. bit–string) is penalized, then we find all of the Steiner vertices, which are on the shortest path between randomly chosen terminal vertices. Then we change all of the corresponding bits (corresponding to the Steiner points) in the bit–string to the number 1 and we recalculate the value of this bit–string. This step is performed until the initial solution (or initial chromosome) is valid. This step is performed only at generating of the initial solution. 4) If the initial solution (i.e. bit–string) is penalized, then we select the low-cost edge and its incident Steiner points. Then we change all of the corresponding bits (corresponding to the Steiner points) in the bit–string to the number 1 and we recalculate the value of this bit–string. This step is performed until the initial solution (or initial chromosome) is valid. This step is performed only at generating of the initial solution. Now we must choose the correction method on the concrete graph. The choosing was performed after testing the correction techniques on the graph instances. Result of the test is shown on the table Tab. 1. 1) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 148 Graph c01 c02 c03 c04 c05 c06 c07 c08 c09 c10 c11 c12 c13 c14 c15 GA Correction 4 Correction 2 Correction 2 Correction 3 Correction 2 Correction 3 Without method Correction 3 Without method Without method Without method Without method Without method Without method Without method TS Correction 4 Correction 2 Correction 2 Correction 3 Correction 2 Correction 3 Correction 3 Correction 3 Correction 3 Correction 3 Without method Without method Without method Without method Without method SA Correction 4 Correction 2 Correction 2 Correction 3 Correction 2 Correction 3 Correction 3 Correction 3 Correction 3 Correction 3 Without method Without method Without method Without method Without method Tab. 1: Correction techniques for graphs 4.4 The Initial Solution The initial solution is a randomly generated bit–string. However, the randomly generated initial solution does not ensure to be a valid solution. This disadvantage is corrected with the special designed correction techniques (see chapter 3.3). 4.5 The Neighborhood Structure Given a particular solution xÎ X , the neighborhood of such a solution is denoted by N(x). N(x) consists of the set of minimum spanning trees that can be obtained by the removal or the addition of a single Steiner node j Î S from/to the current solution x. Thus, the neighborhood from the initial solution in TS and SA is generated by inverting one position in the bit–string. The neighborhood structure in GA is given by the crossover operator (e.g. uniform crossover). 4.6 Tabu Search t:=1; generate randomly initial solution Po; correction of the Po; Pmin := Po(t); fmin := f(Po(t)); TL := (); {Tabu List} while t <= max_iterations do begin Ploc := Po(t); floc := f(Po(t)); change := false; for i:=1 to count_of_Steiner_points do begin Q := Transform(Po(t),i); {inverts a position i in the bit–string} if (f(Q) < floc) and ((not InTabuList(i)) or (f(Q) < fmin)) then begin Ploc := Q; floc := f(Q); Vloc := i; change := true; end;{if} end; {for} if change = false then {if all of transformations are unsuccessful then ivert one bit randomly} begin i := (Random(count_of_Steiner_points))+1; {generate randomly a number in an interval [1,SP]} Q := Transform(Po,i); {invert one position in the bit–string randomly} Ploc := Q; floc := f(Q); end; if floc < fmin then begin fmin := floc; Pmin := Ploc; INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 149 end; if change = true then if |TL| < k then TL := TL + Vloc else TL := TL – V1 + Vloc; t := t+1; Po(t) := Ploc; end; {while} Pmin is the approximation of the optimal solution. Fig. 5: The Tabu search algorithm The initial solution Po is generated randomly and it is corrected with the correction technique. TL is a Tabu list (FIFO) which contains the last progressive transformations. The tabu list is set to the length 4. The number of iterations t was chose to be t=200. The aspiration criterion is implemented in the command: if (f(Q) < floc) and ((not InTabuList(i)) or (f(Q) < fmin)), < fmin). where aspiration criterion is represented by the expression (f(Q) 4.7 Simulated annealing randomly generate starting bit string P0; correction of the Po; Pmin := P0; set starting temperature T0 > 0; set terminal temperature Tf; T := T0; nT := number of Steiner points; repeat for i := 1 to nT do begin randomly generate neighborhood state P from a set of neighborhoods N(P0); Df := f(P) – f(P0); if Df < 0 then begin P0 := P; { better solution is always accepted} if f(P) < f(Pmin) then Pmin := P; { updating of the best solution} end else begin randomly select r from the steady-state distribution (0,1); if r < exp(–Df / T) then P0 := P; { moving to the worse solution} end; end; T := decr * T; until T < Tf; { crystallization in the annealing } { Pmin is the approximation of the optimal solution } Fig. 6: The Simulated annealing algorithm The initial solution Po is generated randomly and it is corrected with the correction technique. The starting temperature is chosen to be 10000 and the terminal temperature is chosen to be 1. The number of repeating of the Metropolis algorithm is chosen to be equal 100. The cooling schedule is given: T = T * decr, where decrement decr = 0.99. 4.8 Genetic algorithm randomly generate starting population P(0) := {P1, ..., PN}; correction of the P(0); compute fitness of all chromosomes: {f(P1),..., f(PN)}; seek for best chromosome Pmin from starting population: f(Pmin) £ f(P), "P Î P(0); t := 1; while t < iteration_count do begin repeat select 2 parents R1, R2 from population P(t) ; generate offspring D crossovering parents R1 and R2; INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 150 mutate offspring D; until not (D in P(t)); compute fitness of offspring D; seek for worst chromosome Pmax in population P(t): f(Pmax) ³ f(P), "P Î P(t); replace chromosome Pmax with offspring D; if f(D) < f(Pmin) then Pmin := D; {update best solution} t := t + 1; end; {Pmin is approximation of optimal solution} Fig. 7: The Genetic algorithm The starting population is randomly generated and it contains 200 chromosomes. Parents of new offspring are selected with the binary tournament selection. Mutation operator randomly inverts one bit in offspring’s chromosome with probability 1/n, where n is length of the chromosome. The least-fit member replacement is implemented in the program. The least-fit chromosome is replaced by a new offspring in each of GA iteration. The algorithm is stopped after achieving of 50000 iterations. 5. FINAL ALGORITHM The final algorithm is designed as follows (see Fig. 8): GraphReductions(InputGraph); SteinerTree := Stochastic heuristic(TS or SA or GA) with Correction algorithm; ApproximationOfSteinerMinimumTree := FinalOptimization(SteinerTree); Fig. 8: The Final algorithm Routine FinalOptimization performs simple local hill–climbing by executing a sequence of mutations on bit–string, each of which improves the solution. An exhaustive strategy is used so that when the routine has been executed, no single mutation exist, which can improve solution further. 6. COMPUTATIONAL RESULTS The FinalAlgorithm was tested on the SPG instances from the OR-Library [12] (http://elib.zib.de/steinlib/testset.php), denoted c01 – c15. The problem sizes after performing the initial graph reductions are listed in the Tab. 2. Computational results of the final algorithm are described in the Tab. 3. Each graph (c01 – c15) was tested ten times to obtain the closest solution to the optimum. In the Tab. 3 we present a minimum from ten achieved solutions (column “Min”), average cost obtained from ten tests (column “Æ”) and the computational time (column “time”; h=hour, m=minute, s=second, ms=millisecond). Empty lines in the result table presents that the algorithm SPA was not found a solution within 5 hours. The testing program was built in Borland Delphi 6.0 and the algorithm was tested on the computer P4 Celeron, 2018 MHz, 512 MB RAM and Windows XP Professional. Graph c01 c02 c03 c04 c05 c06 c07 c08 c09 c10 c11 c12 c13 c14 c15 |V| 500 500 500 500 500 500 500 500 500 500 500 500 500 500 500 |E| reduced |V| reduced |E| 625 149 267 625 133 242 625 208 329 625 250 374 625 336 457 1000 369 847 1000 383 870 1000 395 882 1000 424 916 1000 443 934 2500 499 2184 2500 499 2239 2500 498 2210 2500 500 2195 2500 500 2174 |C| Optimum 5 85 10 144 83 754 125 1079 250 1579 5 55 10 102 83 509 125 707 250 1093 5 32 10 46 83 258 125 323 250 556 Tab. 2: Original and reduced graphs „C“ INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 151 INPUT GRAPH Graph Optimum c01 c02 c03 c04 c05 c06 c07 c08 c09 c10 c11 c12 c13 c14 c15 GA Min Æ SA time Min Æ TS time Min Æ KMB time Min time SPA Min time 85 85 85,8 4m 85 87,2 4m 85 92,2 2m 88 31ms 87 94ms 144 148 153 3m 145 148,1 3m 148 156,3 1m 144 78ms 144 391ms 754 754 757,3 7m 754 755,1 8m 758 762,6 2m 783 8s 772 6m 1079 1079 1079 9m 1079 1080,4 11m 1079 1084,9 3m 1118 25s 1098 20m 1579 1579 1579 12m 1579 1579,2 20m 1579 1579,1 4m 1602 3m 1595 3h 11m 55 55 63 15m 55 61,9 23m 61 73,3 18m 60 187ms 55 594ms 102 107 113 15m 107 114,7 25m 114 126,1 20m 114 827ms 102 4s 509 516 520,8 16m 514 522 32m 524 532,6 19m 533 56s 523 33m 707 713 720,9 19m 713 716,6 38m 720 734 21m 727 4m 722 1h 55m 1093 1093 1095 26m 1093 1097,5 47m 1093 1095,2 17m 1122 18m 32 36 42,8 46m 35 39,2 1h 20m 36 41,2 1h 24m 35 4s 34 7s 46 53 60,5 51m 58 61,4 1h 24m 58 69,2 1h 30m 49 14s 48 1m 258 264 268,4 53m 272 275,7 1h 27m 281 283,6 1h 17m 276 14m 323 324 325,9 47m 335 339,6 1h 50m 342 343,6 1h 24m 342 37m 556 556 557 56m 564 565,9 1h 45m 560 561,8 50m 574 2h 9m - Tab. 3: Computational results In the result table Tab. 3 we can see, that the GA found the global optimum 7 times, SA found the global optimum 6 times and TS found the global optimum 4 times. If the set of terminal vertices is too small (5 or 10), then best results we obtained using of the KMB or SPA heuristic. For the much more terminal vertices (80 – 250) were approximation algorithms KMB and SPA practically unusable due to the infectivity and runtime consumption. 7. CONCLUSION In this paper the improved algorithm based on stochastic heuristic methods for Steiner Problem in a Graph was presented. The main idea behind the algorithm is the application of the spanning tree algorithm on the subgraph. Subgraph is created by removal of some Steiner points from the input graph. The state space is represented by a bit string. This scheme does not ensure that every bit string corresponds to a valid solution. This disadvantage we corrected with the special correction techniques. At the beginning the input graph is reduced before starting of the heuristic algorithm, thus we applied heuristics on lesser graph instances. Algorithms presented in this paper are practicable used on the graphs where the number of vertices is 500 and the number of edges is 2500. Testing of the algorithms is verified on graphs from OR-Library using a program. In the future work we can try to improve the time complexity of base algorithms (Kruskal’s and Dijkstra’s) with heap data structures to obtain results in a shortest time. Then, we can try to combine also stochastic heuristic methods with the approximation algorithm to obtain better solutions. REFERENCES [1] ChARTRAND, G. and OELLERMANN, O.R.: Applied Algorithmic Graph Theory. McGraw Hill, New York, 1993. [2] DUIN, C. W. and VOLGENANT, A.: Reduction tests for the Steiner problem in graphs. Networks 1, 1971, pp. 549–567. [3] ESBENSEN, H.: Computer Near-Optimal Solution to the Steiner Problem in Graph Using a Genetic Algorithm. Networks, 1995, Vol. 26, pp. 173-185. [4] GENDREAU, M., LAROCHELLE, J–F. and SANSO, B.: A Tabu Search Heuristic for the Steiner Tree Problem, Networks, 1999, Vol. 34, pp. 162–172. [5] GLOVER, F., TAILLARD, E. and de WERRA, D.: A user's guide to tabu search, Ann Oper Res 41 (1993), pp. 3–28. [6] GLOVER, F.: Future paths for integer programming and links to artificial intelligence. Comput Oper Res 13 (1986), pp. 533 – 549. [7] GLOVER, F.: Tabu search–Part I, ORSA J Comput 1 (1989), pp. 190–206. [8] GLOVER, F.: Tabu search–Part II, ORSA J Comput 2 (1990), pp. 4–32. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 152 [9] HANSEN, P. and JAUMARD, B.: Algorithms for the maximum satisfiability problem, RUTCOR Research Report, Rutgers, New Brunswick, NJ, 1986, pp. 43–87. [10] HERTZ, A. and de WERRA, D.: Tabu search techniques – A tutorial and application to neural networks. Oper Res Spekt 11 (1989), pp. 131–141. [11] KARP, R. M.: Reducibility among combinatorial problems. In R.E. Miller and J.W. Thatcher, editors, Complexity of Computer Computations, pages 85-103. Plenum Press, 1972. [12] KOCH, T., MARTIN, A. and VOß S., SteinLib Testdata Library, http://elib.zib.de/steinlib/steinlib.php [13] KOU, L., MARKOWSKY, G. and BERMAN, L.: A fast algorithm for Steiner trees. Acta Info. 15, 1981, pp. 141–145. [14] PLESNÍK, J.: Grafové algoritmy. ALFA, Bratislava, 1983. [15] ŠEDA, M.: Aplikace Steinerových stromů v síťové optimalizaci. In Proceedings of the 4th Scientific – Technical Conference PROCESS CONTROL 2000 (ŘÍP 2000), Kouty nad Desnou, 2000, 10 str. [16] ŠEDA, M.: Solving Steiner Tree Problem Using Local Search Metheds. In Proceedings of the 22nd International Conference Telecommunications and Signal Processing – TSP’99, Brno, 1999, pp. 102-105. [17] ŠEDA, M.: Využití moderních heuristických metod v rozvrhování. [Dissertation thesis], Brno 1998. – VUT Brno. Faculty of Mechanical Engineering. Department of Automation and Computer Science. [18] SMUTEK, D.: A Tabu Search Heuristic for the Steiner Tree Problem in Graphs. In Proceedings of the First International Conference on Soft Computing Applied in Computer and Economic Environments ICSC 2003. Evropský polytechnický institut, Kunovice, 2003, pp. 124–128. [19] SMUTEK, D.: Steinerovy stromy v grafech. [Diploma thesis] Brno 2000. - VUT FSI Brno. Faculty of Mechanical Engineering. Department of Automation and Computer Science. [20] TAKAHASHI, H. and MATSUYAMA, A.: An approximate solution for the Steiner problem in graphs, Math. Jap. 24, 1980, pp. 573–577. [21] WINTER, P. and MACGREGOR SMITH, J.: Path–distance heuristics for the Steiner problem in undirected networks. Algorithmica 7, 1992, pp. 309 – 327. [22] WINTER, P.: Steiner problem in networks: A survey, Networks 17, 1987, pp. 129–167. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 153 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 154 DETECTING PARETO OPTIMAL SOLUTIONS WITH PSO Ulrike Baumgartner, Ch. Magele, W. Renhart Kopernikusgasse 24/3, A-8010 Graz, Austria, email: [email protected], phone: +43 316 873 7257 Abstract - Real world optimization problems often require to minimize/maximize more than one objective which, in general, are conflicting each other. These problems (multiobjective optimization problems, vector optimization problems) are usually treated by using weighted sums or other decision making schemes. An alternative way is to look for the Pareto Optimal Front. In this paper the Particle Swarm algorithm is modified to detect the Pareto Optimal Front. Keywords - Stochastic Optimization, Particle Swarm Optimization, Pareto Optimality, Multimodal Problems I. INTRODUCTION Stochastic optimization methods have successfully been applied to scalar and vector optimization problems. In the latter case, the different objectives, which are in general conflicting each other have to be transferred into a single scalar objective function. This is usually done by normalizing and processing the contributions to the objective function (weighted sums, fuzzy membership functions, ...) [1]. Another approach is to find the Pareto optimal front which summaries all Pareto optimal solutions [2]. A Pareto optimal solution is, by definition, the best that could be achieved for one objective without disadvantaging at least one other objective. Particle Swarm Optimization (PSO) [3] is a stochastic optimization technique which imitates the social behaviour of a bird flock flying around and sitting down on a pylon. The idea of the particular PSO implementation, discussed in this paper, is that the flock should now sit down on the transmission line (the pareto front) and the pylons are controlling points (sampling points) for the flock. The reliability and the performance of the proposed method will be tested using analytical examples. 2. PARETO OPTIMALITY The solution to a multiobjective problem is, as a rule, not a particular value, but a set of values of optimization variables such that, for each element in this set, none of the objective functions can be further increased without a decrease of some of the remaining object functions (every such value of a optimization variable is referred to as pareto-optimal). The simplest way illustrating this behaviour is shown in Fig. 1. The point xt+1 is in both cases pareto optimal because (1) or (2) is true (OF ... objective function value). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 155 Figure 1. Pareto optimal points - single dimension Figures 2 and 3 are illustrating pareto optimality in a two dimensional parameter space on simplified Schaffer functions (3) [4]. In point A both objective function values can be improved by moving into the region enclosed by the (negative) gradient vectors, while in point B the two objectives are in conflict; an improvement of one objective causes a deterioration of the other objective, which corresponds to the definition of a pareto optimal point. The gradient vectors can be formulated as a positive linear combination. It is obvious that all objective function values can be improved for point C, while Point D is again pareto optimal; none of the objective functions can be increased without a decrease of some of the remaining objective functions. In point D it is also possible to state g1k1 + g2k2 + g3k3 = 0 where k1, k2 and k3 are positive numbers. Figure 2. Two conflicting objectives - 2D Figure 3. Three conflicting objectives - 2D INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 156 Figure 4. Two conflicting objectives - 3D Figure 5. Three conflicting objectives - 3D The three dimensional cases are shown in Fig. 4 and Fig. 5, the behaviour of pareto optimal solutions and the corresponding gradient vectors are the same as in two dimensions. The mathematical interpretation of this considerations needs the use of linear algebra and the treatment of a system of equations using the gradient direction vectors. A point is identified as pareto optimal if (4) is fulfilled, that means that the rank of the gradient matrix G is smaller than m. 3. PARTICLE SWARM OPTIMIZATION In the population-based search procedure σ particles fly around in a multidimensional search space. During flight each particle adjusts its position xt and velocity vt according to its own experience and the position of the best of all particles. The discrete ordinary differential equation controlling the motion of a single swarm member is given in (6). The PSO algorithm for single objectives works as in the following described: The swarm is uniformly spread in the INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 157 parameter space, bounded by [xmin,xmax]. The fitness of all swarm members is evaluated and the first swarm leader is going to be specified. The position and velocity of each swarm member are updated according to (6). After evaluating the performance of all particles a new swarm leader (the member with the best performance) is found. If the swarm collapses (i. e. the swarm radius is below a certain positive small constant ε) or the maximum number of iterations has been exceeded, the algorithm terminates. 4. PARTICLE SWARM OPTIMIZATION AND PARETO OPTIMALITY To implement the idea of a bird flock landing on a transmission line using pylons as orientation points the whole particle swarm is portioned into sub swarms. The ”pylons” are additional objective functions, which are weighted sums of the main objectives (no accessorily function calls - only multiplications and additions). The algorithm starts again with a randomly distributed starting configuration, the initialization swarm, which will be generated in the range [xmin,xmax]. The objective function values OF1, OF2, ... ,OFm of each swarm member are determined and (7) is calculated for a set of weights. For each weighted sum, the particle yielding the best objective function value becomes a swarm leader x t,k best. The swarm is equally partitioned and each part of the swarm evolves into the direction of its own swarm leader x t,k best according to (8) where x t,k best = min(OFwk) is the position of the current swarm leader of the swarm part k. The size of a swarm part k is set to s/n. The major modification compared to the standard algorithm is the detection of pareto optimal solutions. To speed up the algorithm this process is two fold. The first step (preliminary pareto decision) is narrowing down the choice by testing all swarm members against the simple Pareto formulations (1) and (2). The fitness OFj(xt+1) and the previous fitness OFj(xt) are already calculated and therefore no additional solutions of the forward problem are needed in this step . If (1) or (2) is fulfilled it is possible that the point is pareto optimal and the procedure is followed by a substantially stricter selection process (main pareto decision), taking the gradient information (9) into account. To do so, additional objective function values OF1, OF2, ...,OFm are evaluated at positions slightly displaced from the current position xt+1 of a particle under investigation (OFj,Δx). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 158 If a particle is recognized to be pareto optimal, it has reached its final position and is removed from the swarm. If the number of remaining swarm members has reached a prescribed threshold or the maximum number of iterations maxit has been exceeded, the algorithm stops. 5. ANALYTICAL TEST FUNCTIONS The simple test functions (3) were scaled up to the five dimensional space. To obtain the pareto optimal front, the proposed pareto PSO version was applied and compared to a classical PSO algorithm, both in reliability of the results and in computational effort. The pareto PSO version was started with a population size s= 100 and the algorithm was stopped when, at least, 25 pareto optimal points were detected. Only one solution for one objective can be found by a classical PSO run, therefore the standard algorithm was applied for each weighted sum of objectives (7) sequently. The same weighted sums were used for the pareto PSO version as orientation points. The standard PSO procedure was applied twice, firstly with a constant swarm size sS for all dimensions and secondly with a swarm size sS depending on dimensions and the number of main objectives. Two test cases were investigated, one with two conflicting objectives and the other one with three conflicting objectives. The complete series has included ten optimization runs for two, three, four and five optimization parameters (dimensions), respectively. A. Analytical Test Functions - Two Conflicting Objectives For the two objectives test series the weights were chosen as l = [0 0.25 0.5 0.75 1] leading to five objective function values OFwk = lkOF1 + (1 - lk)OF2. Table 1 summarizes and compares the results. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 159 The proposed pareto PSO was able to locate at most 36, 34, 32 and 31 pareto optimal points (pop) during a single run in the 2D, 3D, 4D and 5D case, respectively, while the mean value of found pops was between 25.1 and 28.6. As can be seen, the number of function calls increased from 2198 in the 2D case to 10391 in the 5D case. This is not the case when running the classical PSO with a constant swarm size sS = 20. But this approach fails completly if higher dimensional problems are tackled. In the 5D case, the best run was able to determine two (out of five) optimal points, the mean value turned out to be even worse, namly 0.6 (out of five) orientation points. Scaling the swarm size with the dimension of the problem (sS =20, 40, 60 and 80 in the 2D, 3D, 4D and 5D case, respectively) brings the classical PSO back into the game, at the cost of additional function calls, which are comparable to the ones of the pareto PSO. The mean value of obtained optimal points corresponds in all cases with the number of orientation points, which means that all possible solutions are found. Nevertheless one should prefer the solutions of the pareto PSO, which represent the pareto front much better than the discrete orientation points. Figure 6. Two conflicting objectives - 2D Figure 7. Pareto optimal front - OF1 and OF2 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 160 Figure 8. Two conflicting objectives - 3D Figure 9. Two conflicting objectives - 3D - all pops Figures 6 and 8 show the pareto optimal solutions (denoted by dots) obtained during a single run in the 2D and in the 3D case respectively. They are distributed nicely along the Pareto optimal front, spanned by the orientation points (denoted by stars). In Fig.9 all pareto optimal solutions in the 3D case obtained by the 10 successive runs are plotted. Figure 7 displays an OF2(OF1) diagramm, which is the same for all tested cases with two conflicting objectives. B. Analytical Test Functions - Three Conflicting Objectives When three conflicting objectives have to be taken into account, 15 orientation points are set up as given in (7). From Fig. 10 it can be seen, that in the 2D case these orientation points (denoted by starts) are distributed along and inside a triangle with the three individual optimal points (max(OF1), max(OF2) and max(OF3)) as verteces. Figure 11 also shows the distribution of pareto optimal points (denoted by dots) obtained by a single pareto PSO run, while in Fig. 8 the pops of the 3D case are given. Figure 10. Three conflicting objectives - 2D Figure 11. Three conflicting objectives - 3D Having a closer look at Tab. 2 reveals results similar to the previous ones, but much more in favour of the pareto PSO. The classical PSO with a constant swarm size sS fails completely, once the 2D space is left. Increasing the swarm size corresponding to the dimension of the problem brings back the reliability of the method but at the cost of an enormous number of function calls (5D about 50000). The pareto PSO, though, is INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 161 able to approximate the pareto front very accurately with a remarkable low computational effort. 6. CONCLUSION The particle swarm optimization algorithm was applied to multiobjective optimization problems with two and three objectives. The algorithm revealed an excellent performance concerning the number of solutions of the forward problem and completed with a very reliable representation of the pareto optimal front. The pareto optimal front gives a survey of all reasonable solutions of the multiobjective problem. Now the user is free to choose the preferred realiziation from the whole set of solutions. REFERENCES: [1] ALOTTO, P.G., BRANDSTÄTTER, B., CELA, E., FÜRNTRATT, G., MAGELE, Ch., MOLINARI, G., NERVI, M., PREIS, K., REPETTO, M. and RICHTER, K.R.: “Stochastic Algorithms in Electromagnetic Optimization” IEEE Trans. Magn., vol. 34, No. 5, pp 3674-3684, 1998 [2] BARTA, P. Di, FARINA, M., SAVINI, A.: “An improved technique for enhancing diversity in Pareto evolutionary optimization of electromagnetic devices” A COMPEL: Int J for Computation and Maths. in Electrical and Electronic Eng.; Volume 20 No. 2; 2001 [3] BRANDSTÄTTER, B., BAUMGARTNER, U.: “Particle swarm optimization - mass-spring system analogon” Magnetics, IEEE Transactions on , Volume: 38 Issue: 2 Part: 1 , March 2002 [4] SCHAFFER, J.D.: “Multiple Objective Optimization with Vector Evaluated Genetic Algorithms”. In Genetic Algorithms and their Applications: Proc. first Int. Conf. on Genetic Algorithms, pages 93–100, 1985. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 162 THE RISK ANALYSIS OF SOFT COMPUTING PROJECTS Branislav Lacko Institute of Automation and Computer Science, Faculty of Engineering Brno University of Technology, Technicka 2, 616 69 Brno, Czech Republic Lacko @ uai.fme.vutbr.cz Abstract: The paper provides basic information on the RIPRAN method which has been developed at VUT Brno to support project teams in risk analysis of soft computing projects. The method is suitable especially for the risk analysis of company information system projects. Key words: Project Control, Risk Analysis Method, Risk Identification, Risk Quantification 1. THE IMPORTANCE AND POSITION OF PROJECT RISK MANAGEMENT 1.1. The goal of project management Project Risk Management includes processes dealing with identification, analysis and response to risks in projects, with the aim to minimize their impact on the project.[3] The goal of project management is represented by a successful project. A project can be deemed successful if complying with the following: - The goals planned have been achieved and - The project has been finished as scheduled and - The budget planned has been observed and - Sources available have been used and - Optimum efforts have been made to carry out the project and - Design and documentation have been done in required quality and - The project has no negative impact on environment, project participants or other projects The above-mentioned factors characterizing a successful project show us that the requirements for a successful project are very demanding. At the same time we must reject such an approach stipulating deliberately small goals and low requirements to make their fulfilment easy – a so-called soft project. A project which enables its easy performance as a successful project is not a quality project! Projects of soft computing are very complicated: · Goal is very sophisticated · Goal is changed through project · Goal is difficulty described The main goal of project management is to design and carry out successful projects. In practice, we meet unsuccessful projects very often. A project is considered unsuccessful in the following cases: · Any of the goals planned has failed to be achieved · The project has not been finished in a term scheduled INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 163 · · · · · · Costs planned have been exceeded Available sources have not been used or necessary sources have been lacking Inadequate efforts have been made to achieve the goals, meet the deadline and costs The project has been done in confusion and a whole number of critical situations occurred, the reasons of which could have been assumed beforehand The project has negative impact on environment or other projects A whole number of projects even show combination of several unsuccessful characteristics mentioned above. There are many general reasons of a project's failure: · The project's definition (goals, deadlines, costs, sources) does not correspond to real needs · Changing requirements for the project due to changes in the project's surroundings, which have not been accepted · Underestimated planned costs · Lack of necessary works · Wrong choice of project's individual parts contractors · Wrong identification of the initial state · Wrong project team management · Bad team work · Underestimation of various unfavourable factors or even their negligence · Insufficiently precise project design · Insufficient qualification of team members · Insufficient involvement of the project results end user in the project · Failure in project management methods mastering · Bad environment and bad standard of work management in the company carrying out the project · Wrong link to other projects · Unclear definition of goals · Wrong estimates of time and costs with specific works · Failure of project manager · Fluctuating team staff · Bad team line-up · etc. Moreover, with software projects the danger of project impairment can emerge due to programme errors occurrence. As we can see, the list of frequently repeated, general failures is not and cannot be complete. Furthermore, each of us is sure to remember specific reasons of some concrete software project failure! Project management, however, wants to assure the greatest hope possible of achieving a successful project by means of a project risk analysis. That is the main purpose of the project risk analysis done. [1] Therefore, we must ask the following questions as early as in the stage of project design: · To what extent can we expect our project to be successful? · To what extent can we expect our project to fail to be successful? · What can endanger our project's success? · What can support our project's success? · What can we do to increase the success expected of the project? · What can we do to decrease the failure expected of the project? Answering the above mentioned questions and a broader analysis of the reasons helps us in preparing measures, moderating the possibility of the project's failure and increasing the probability of the project to be successful. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 164 1.2. The importance of the environment's impact to the project In many cases, the source of the project's failure lies in the mistakes committed by the individual project team members. We can understand them as internal reasons of the project's failure. Those can be eliminated relatively easily by thorough preparation of the team staff and improving their professional knowledge and experience. With software projects those mistakes include mistakes in programme products in consequence of programmers´ bad work. The influences coming from the project's environment have a far more substantial, unfavourable and sometimes hardly predictable impact. Those can be called the external causes of the project's failure. These influences may often come as a response to those exerted by the project to its environment. Project environment PROJECT A project and its environment influence mutually! What changes can the project environment cause? · Changing the goal · Changing the road to the goal · Changing the sources available · Changing the conditions of the project implementation A good project manager thus does not focus exclusively on the project – he observes the project's environment as well. The relation of project management and risk engineering can be shown as the following scheme: PM RE Project Risk Management INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 165 2. BASIC RISK ENGINEERING TERMS The RIPRAN method (see further) is based on the risk engineering principle [1], namely that to analyse the risk we must determine the following quartet first and prepare a relevant list thereof: Thread - Scenario - Probability - Loss As the number of accidental events can never be predicted precisely, the list cannot be complete. The incomplete list is also caused by the knowledge or respectively lack of knowledge of the project team members, therefore we talk about a representational list, i.e. a list presenting all the significant risks we were able to determine and which we take as a base for specific risk analysis. Let us present briefly the meaning of those expressions, as understood by the RIPRAN method. Thread Hazard which threatens and causes disastrous consequences and troubles within the project. ( E.g. violent windstorm, insufficient loan, glaze, devaluation of currency, strike, project manager resignation, bad subdelivery for the project, ....) Scenario Course which we assume in the project as a consequence of the thread occurrence. ( E.g. We are not granted a loan – the project will not be covered financially, Johnny falls ill – we will lose a single staff member who is able to do the work for our project...) Probability The probability of implementing the Scenario expressed as PÎ< 0,1> We relate the probability to the project duration *1) and/or so-called reference time*2) when we feel to be endangered by the thread. We must note that it is probability, that there is a scenario allotted with certain probability to the thread with certain probability. Usually we expect both phenomena to be mutually independent. If the windstorm probability is 0.03 and the probability of windstorm coming and makes the crane fall down is 0.7, then the resulting probability which we will taken into account in the respective case is 0.7 x 0.03 = 0.021 1) E.g. for a strong windstorm of 11 degrees of Beaufort scale in our geographic conditions, the probability of occurrence within one year is 0.01, but for 100 years it makes 0.63. 2) E.g. Cables laying is to be done from 1 March to 25 March – then we are interested in the occurrence of ground freezing in this period. Loss A loss for a project occurred through implementation of the scenario. We usually express it by financial units (we can do it in some else way of course, such as time delay, losses of the staff lives, etc. ). To which n-group we can adjoin the risk value Risk value = probability * loss 3. THE RIPRAN METHOD 3.1. Characteristics of the method. The RIPRAN method (RIsk PRoject ANalysis) represents a simple empiric method for risk project analysis, especially for medium-size enterprises. It is based strictly on process approach of the risk analysis. It understands the risk analysis as a process (inputs into process – outputs of the process – activities transforming the inputs to the outputs with certain target). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 166 The method accepts the Philosophy of quality management (TQM) therefore it comprise activities assuring the risk project analysis quality, as required by ISO 10 006 [2] The method is designed to respect the principles of Risk Project Management, described in PMI and IPMA literature. It is focused on processing the project risk analysis which must be carried out before the implementation itself. It does not mean that we should not work with threads in the other stages. On the contrary – in each stage of the project's life cycle we must carry out activities which gather basic data for the very project risk analysis for the project implementation stage and evaluate the possible risks of failure in the stage being currently in progress. The risks recorded will then be used for the general project risk analysis. The entire project risk analysis process [3] consists of three activities: · Risk identification · Risk quantification · Risk reduction These activities are conceived as processes related mutually. 3.2. Risk identification Goal: Finding threads and scenarios Inputs: · · · · · · Project description Historical data on previous projects ( Post Implementation Analysis, Trouble List) Prognoses of possible external influences Prognoses of possible internal influences Experience Activies: Application WHAT-IF questions Application WHAT-IS questions Output: A list of pairs thread – scenario № Thread Scenario Quality supporting activities: · Input data validity and completeness test · Test of the team competency and completeness · Test of prognostic input data up-to-date condition · Test of output pair list completeness 3.3. Risk quantification Goal: Evaluate the probability of scenarios, scope of damages and assess the risk rate Inputs: · A list of pairs thread - scenario · Statistical data on previous projects and various other statistical data · Experience Outputs: Complete n-groups (thread, scenario, probability, loss)– interim result № Thread Scenario Probability Loss Risk Value INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 167 · · · · List I. for the project design completing List II. for information to operative actions possible List III. for the follow-up Risk reduction process Preliminary standard of the general project risk Quality supporting activities: · Input data list validity and completeness test · Test of the team competency and completeness · Test of statistical data up-to-date condition · Test of output pair list completeness 3.4. Plans for the risk reduction Goal: Based on information and awareness of the danger, measures are to be prepared to reduce the risks impact. Input: List of n-items (thread, scenario, probability, loss), which must be taken into account (List III.) Activities: Creative application of 9 prototyping risk reductions Output: · Plans of risk reduction · (A plan of measures to reduce the risk) · Evaluation of associated project risk based on the measures Quality supporting activities: · Test of the input list validity and completeness · Test of the team competency and completeness · Checking the plans for project risk reduction 4. CONCLUSION Soft computing projects usually miss the chapter on risks analysis. It is certainly one of the reasons why a whole number of soft computing projects fail. When designing software, our software firms should start using modern project management to much greater extent, considering also the problems of risks in project design and implementation. The RIPRAN method is a suitable means, how to carry out the risk analysis and elaborate a design to decrease the risk in top quality. The Institute of automation and computer science wants to focus on concrete company projects of control and information systems, soft computing project, etc. At present the 1st version is available and the 2nd version will be prepared, which will include computer aid of the RIPRAN method plus a form of documents complying with the ISO 9000:2000 standards requirements. REFERENCES: [1] BOEHM, W.B.: Risk Management. IEEE Computer Society Press. Los Alamos 1975 ISO 10006 [2] IPMA Competence Baseline. Report of International Project Management Association Project Management Body of Knowledge. Project Management Institute. Upper Darby 1996, (Chap. 11 Risk Project Management) [3] LACKO,B.: RIPRAN Method. Research Report 02/2000. TU of Brno – Faculty of Mechanical Engineerig – Institute of Automation and Computer Science, Brno 2000, 57s. The paper was elaborated within the frame of the research plan of the Ministry of Education No. CZ J22/98 260 000013 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 168 NETWORK VIRTUAL LABORATORY – INDUSTRIAL ROBOT Wlodzimierz M. Baranski, Tomasz Walkowiak Institute of Engineering Cybernetics, Wroclaw University of Technology ul. Janiszewskiego 11/17, 50-372 Wroclaw, POLAND, Phone: +48-71-3202969, +48-71-3203996 Fax: +48-71-3212677, E-mail: {mwbar, twalkow}@ict.pwr.wroc.pl Abstract: The paper describes an approach to distance learning in computer engineering. A virtual simulated laboratory available for students through Internet is presented. The aim of laboratory is to teach students how to control external device: an industrial robot type b. Students has a possibility to test robot’s functionality by setting individual bits of control ports. Furthermore, distant access to programming environment is implemented which gives the user possibility to write programs controlling a virtual robot. The whole system has been implemented in Java and requires on user end the web browser with Java plugin. Keywords: distance learning, virtual laboratory, industrial robot, Internet 1. INTRODUCTION Development of Internet has a big influence on global information society. As it is stressed in [7], education policy is of one of the most importance for the success of the transformation towards the global information society. We think that one of solutions for the global education is the internet based distance learning. In science and engineering the laboratory is a lifeblood of teaching. We are focusing here on computer engineering laboratory exercises where students write and verify their own programs on real peripheral devices. In this testing phase equipment is very often destroyed what causes that during the next laboratory test students are not able to work. Service and price of such devices are usually high and reparation requires a qualified service man. The idea of replacement of real devices with simulator with an Internet access allows having laboratory exercises in distance learning. It is important to remember that simulator must behave in the same way like a real device. The distance learning laboratory presented in this paper is functionally compatible with real device [1]. In the case of distance learning there is no possibility to access a real laboratory. Therefore, practical classes are real challenge. Generally, there are two approaches to this problem and each of them has its own advantages and disadvantages: · real laboratory with distance access; · virtual laboratory. In work reported here the concept of virtual laboratory was used. The idea of distance access to real devices was rejected because of following disadvantages: limited number of simultaneously working students (depending on the number of real devices) and the total cost of the system - devices, webcams (to view the actual state of the process) and converters of distance user commands. The virtual laboratory doesn't exist in reality. Real devices, on which students practise, are replaced by programs simulating behaviour of those devices in real time. Access to such laboratory gained through the Internet is limited only by the availability of free resources on the server and the network bandwidth and not by the number of real devices or time of the day. There is no need to use additional devices such as webcams or INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 169 special interfaces; therefore no special service is needed in case of a device malfunction. The simulation software can run on the client side or on the server (when simulator needs more resources). There is also an option of running one part of the program on the user’s machine and the second one on the server. The choice depends on many factors such as quality of the network connection, simulation algorithm complexity, difficulty of installation on the user’s machine, system maintenance (new versions and patches distribution) and widely interpreted computer system security. 2. NETWORK VIRTUAL LABORATORY Network Virtual Laboratory (NetVL) [2] is a computer aided learning system developed at Wroclaw University of Technology. The aim of the NetVL is to support learning of programming such types of external devices as a milling table [4] or industrial robot. There is also a possibility of programming microprocessor devices such as PCs. The main purpose of the system is a distance access to those devices through the internet. The user is expected to have only a Java enabled internet browser and of course, the internet access. Beside the ability to view the simulated device and its registers, the user can operate in the programming and running environment. He can write programs controlling devices compile and run them without having any compiler installed on his machine. Access to the compiler and the computer, where the user's program actually is being run, is possible through the internet. Like in the case of inspecting the simulated device, only a Java enabled browser is required. A Java applet which provides communication with the compiler runs in the browser's window. NetVL consists of three main modules (see Fig 1): · Management subsystem (MS) – allows user identification and acts as an communication point for all other modules ; · Virtual external device (VED) – simulates and visualises behaviour of a given external device; · Virtual microprocessor (VM) – makes possible to create and run on the device simulator user programs which interacts with VED. visualisation applet TCP/IP simulator instance for user 1 VED server manipulation applet TCP/IP web browser Internet MS server simulator instance for user 2 Java program Intranet Java program Fig. 1. NetVL modules communications 3. VIRTUAL EXTERNAL DEVICE (VED) Internal structure of virtual device (Fig .1) consists of: Virtual Ports. Module that represents (communication) ports in a real control computer (but that are accessed through the net). Acts as an interface between the user (VM and Manipulator) and the Simulator. This module runs on the server side. · Internal Ports. The same functionality as the Virtual Ports, but having the additional information of the internal state of the device. It sends (on-line) info to the Visualization module that visualizes the device’s functionality. This module runs on the server side. · Simulator. Module that simulates the device’s behavior. This runs on the server side. The simulator works on-line. The outside world could change the internal state of the simulator at any time by · INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 170 · · accessing the virtual ports (by playing with the manipulator or through any outside program). The work of the simulator is visualized on-line in the visualization module, where the user can see the movement of the simulated device and the results of playing with the virtual ports. Manipulator. Gives the ability to control the device (Simulator), by simply playing with the input ports in Virtual Ports (via the Manipulator). This module runs on the user’s computer. Visualization. Module that visualizes the behavior of the device based on the data written to the Internal Ports. This module runs on the user’s computer. 4. ROBOT EXTERNAL DEVICE One of realised Virtual External Devices is an industrial robot [7]. The External Device simulates the behaviour of industrial robot type b (IRb) [5]. The industrial robot type b, shown in Fig 2, consists of: arm motor, arm, forearm, shoulder motor, hand motor, handswivel motor, gripper, shoulder, base motor and base. The robot is made up from several rigid bodies, called links, connected sequentially by revolute or prismatic joints, to achieve the required rotational and/or translation motion. Mechanically, a robot manipulator is composed of an arm and a wrist attached to a tool fixture (a gripper). The arm typically has three degrees of freedom, which accomplish the major positioning of the robot arm and place the wrist unit at the work piece. The wrist consists of up to three rotary motions that help to obtain an appropriate orientation of the tool toward the object. The aim of this lesson is to give familiarity with an industrial robot (Fig. 4), and to teach how to control and operate it. It teaches the student how to write programs for controlling such devices. The user is able to control a virtual robot Simulator of robot allows the user to play with its functionality (Fig 3). Practical work (writing own programs) allows the user to practise the methods on real problems. Fig. 2. Industrial robot type b Fig. 3. Robot applet 1.1. The ports – general description The robot is connected to controlling unit (say PC) and controlled through one 16-bit input virtual port (Control Port, we assigned addresses to 300h) and ten 16-bit output virtual ports (at addresses 300h-312h). The robot has ten output ports: · the first at address 300h and known as a Status Port, · the second at address 302h and known as a Sensor Port, · the next five known as Motor Ports: · the Base Motor Port at address 304h, INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 171 · the Shoulder Motor Port at address 306h, · the Arm Motor Port at address 308h, · the Hand Motor Port at address 30Ah, · the Handswivel Motor Port at address 30Ch, · and next three known as Position Ports: · the X-Coordinate Port at address 30Eh, · the Y-Coordinate Port at address 310h. · the Z-Coordinate Port at address 312h. 1.2. The ports – detailed description The Control/Status Port consists of an input port (Control Word) and an output port (Status Word) located at the same address (300h). The Control Word controls the state of the robot, making it possible to do the following: · turn the power on or off for each motor in the joints of the robot, · turn the power on or off for the gripper, · set the direction of motion of each motor (up/down or left/right), · set the type of the gripper operation (close or open: when the gripper is on and the gripper operation is set to close the robot can write in space). The Status Word includes the same information, but is read-only. Its function is to show the dynamic state of the robot: · which motors are on or off, · if the gripper is on or off, · what direction of motion is set for each motor (up/down or left/right), · what type of gripper operation is being performed (close or open). The Sensor Port (302h) lets you check the state of the sensors that show when the extreme positions of each motor (right/left or highest/lowest) are reached and check the state of the gripper (closed or open). The Motor Ports (304h-30Ch) give the actual state of each motor (deflection in degrees relative to their standard position at reset). Each port has its own range of motion: · Arm Motor from -30 to 30 degrees, · Shoulder Motor from -40 to 20 degrees, · Hand Motor from -20 to 200 degrees, · Handswivel Motor from -90 to 90 degrees, · Base Motor from -170 to 170 degrees. The Coordinate Ports (30Eh-312h) give the current position of the gripper (edge). The position of the gripper is described by three coordinates: X, Y and Z (at addresses 30Eh, 310h, 312h, respectively). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 172 a) b) c) d) Fig. 4. Robot simulator in action 1.3. Simulator architecture The robot simulator was written in Java and it’s components agrees with VED concept presented in part. It allows communication with NetVL system components as presented in Fig. 5. It consists of two separte programs: robot server (Java program runs as a demon) and robot client (with front end presented in Fig. 3) which is a Java applet. RobotClient Manipulator PortRegister Wizualizator ViewController new Client (A pplet) RobotServer LO G IN RobotServer Thread send W irefr (0) Wireframe Matrix[] LO G IN PCClient Internal Ports RobotApple PortServer RobotPrgm tThread Thread Control new Client (P rogram) Status Sensor Auth. module UserTable IBM PS/2 terminal emulation Control ApplMotion Thread NetVLServer Virtual Ports RobotCom Internal Ports (local) RobotPortClient check LO G IN PrgmMotion Thread NetV L.h Status Sensor DOF DOF Fig. 5. Robot Simulator architecture 5. VIRTUAL MICROPROCESSOR (VM) The Virtual Microprocessor (VM) subsystem of the NetVL is focused on teaching how to write programs. Internal structure of virtual microprocessor consists of following elements: · Programming environment. Allows the user to prepare (write, compile), run and test controlling programs on a remote computer. It will run some commercial compilers and processor simulators INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 173 · · · · · · · and send terminal-like outputs to the user. The executable programs will stay on the server side. It will also allow execution of these programs. The program running on the server side will communicate with the outside world: with a user through the Terminal and with the VED via the Virtual ports. This module will be run on the server side. Control. Module that allows the user to edit any programs and execute compilation or programs running on the remote computer. Also any compiler-linker outputs will be seen on the user’s screen. This module will be run on the client side. Terminal. Module that allows terminal-like communication of an executing program with the user. Real device. There is a possibility that the program will be run on a real device fitted in the computer on the server side [2]. The user of VM can: create a new file or directory; edit and save an existing file (Fig 6) ; change a file or directory name (Fig 7); delete a file or directory. All users have their own disk space that can only be accessed by them (and of course by the administrator). When an authorized user accesses a given VM for the first time, the VM server automatically creates a user disk space. The user can create executable programs. The user applet in the VM subsystem allows the user to run compilers or linkers on remote machines and to see the program outputs on a terminal-like window. Programs can be run on a remote machine and can communicate through a terminal-like interface with the user. One of realization of VM [2] (so called PC VM) is is based on the Borland C++ 5.5 freeware compiler. It allows the user to write any C or C++ language program (though only terminal user interface is supported) and run it. Using a terminal window, the user can communicate with a program running on a remote computer. Fig. 6. VM applet – contents of directory Fig. 7. VM applet – contents of directory The PC VM also has the ability to write programs that control Virtual External Devices (VED), such as a milling table (see Fig. 9). We have developed a special C library giving a user two functions (inb() and outb()) that allow reading or writing to a VED port. Writing to or reading from a VED port is performed by a TCP/IP connection with the VED server. The visual effects of controlling the VED can be seen in the client applet. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 174 TCP/IP VM C lient stdin, stdout Sofware calls inb(), User outb() NetVL progr am Librrary TCP/IP TCP/IP MSServer VM Server TCP/IP VED server VED Client Fig. 8. NetVL modules communications 6. SUMMARY We have presented a virtual laboratory system available for students equipped with a computer and internet access. This system distance learning laboratory system is very useful for teaching students how to control an external device: industrial robot. All materials are delivered by the internet. Tests proved that ordinary phone-line modem access is sufficient for effective usage, however the main drawbacks are form long delays. So the main requirement for All the software has been written efficiently and computing resources are not heavily challenged. It’s possible [3] to include part or whole the system in the standard Learning Management Systems (like for example WebCT). 7. ACKNOWLEDGMENT The work reported here was partly sponsored by European Union in the framework of „Telematics Application Program” INCO-Copernicus project „Multimedia Education: An Experiment in Delivering CBL Material” [6]. REFERENCES [1] BARANSKI W., MAJEWSKI J., DOBROWOLSKI A., Functional simulation of microprocessor external devices, Proceedings of 4-th International Conference Computer Aided Engineering Education, CAEE '97. Krakow, Poland, September 11-13, 1997, vol II, pp.280-287. [2] BARANSKI W., WALKOWIAK T., MM-EDU: network virtual laboratory, Technical Reports Institute of Engineering Cybernetics, SPR 2/2001, Wroclaw University of Technology, 2001. [3] BARANSKI W., WALKOWIAK T., Multimedia approach to distance learning, First International Conference on Soft Computing Applied in Computer and Economic Environments. ICSC 2003., January 30- 31, 2003 Kunovice, Czech Repubic, pp. 150-158. [4] BARANSKI W., WALKOWIAK T., Network Virtual Laboratory - Milling Table. ICTINS’2003, International Conference on Information Technology and Natural Sciences, October 19-21, 2003, ALZaytoonah University, Amman, Jordan, pp.. 122-126. [5] CRAIG, J.J. Introduction to Robotics: Mechanics and Control. Addison-Wesley, 1986. [6] MM-EDU project http://mm-edu.ict.pwr.wroc.pl. [7] Poland and Global Information Society: Logging on. Human Development Report, UNDP, Warsaw 2002. [8] WALKOWIAK T., Storyboard of integrated MM-EDU module full version, Technical Reports Institute of Engineering Cybernetics, SPR 43/2000, Wroclaw University of Technology, 2000. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 175 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 176 SIMD APPROACH TO RETRIEVING ALGORITHM OF MULTILAYER PERCEPTRON Jacek Mazurkiewicz Institute of Engineering Cybernetics, Wrocław University of Technology ul. Janiszewskiego 11/17, 50-372 Wrocław, POLAND, Phone: +48-71-3202681 Fax: +48-71-3212677, E-mail: [email protected] Abstract: The paper is a proposal related to partial parallel realisation of retrieving phase of Multilayer Perceptron algorithm. The method proposed is based on pipelined systolic arrays – an example of SIMD architecture. The discussion is realised based on operations which create the following steps of the algorithm. The data which are transferred among the calculation units are the second criterion of the problem. The efficiency of proposed approach is discussed based on implementation quality criteria for systolic arrays. The results of discussion show that it is possible to create the architecture which provides massive parallelism and reprogrammability. Keywords: Multilayer Perceptron, systolic array, SIMD architecture 1. INTRODUCTION The paper is related to partial parallel realisation of retrieving phase of Multilayer Perceptron algorithm. The method proposed is based on pipelined systolic arrays. The described methodology can be used as the theoretical basis for hardware of software simulators of Multilayer Perceptron. The discussion is realised based on the following assumptions: · the outcome of algorithms realised true to proposed methodology is exactly the same like the outcome of classical Multilayer Perceptron algorithm, · three-layer Multilayer Perceptron is taking into account, but the presented approach can be easily adopted to more sophisticated Multilayer Perceptron networks, · the systolic structure is realised using only digital elements, input and output data are represented in proper binary code, · the number of neurons which create the layers of Multilayer Perceptron is unrestricted, but the maximum number of elementary processors can be limited. 2. MULTILAYER PERCEPTRON ALGORITHM – RETRIEVING PHASE The Multilayer Perceptron network is composed by sets of neurons which create the layers. The classical net includes three layers: · input layer, · hidden layer, · output layer. Each layer consists of the proper number of neurons. The size of input layer equals to the size of input vector, the size of output layer is related to the code used for the output of the network description and finally the number of hidden layer neurons is estimated by different ways. For our discussion lets assume that we have: · N neurons in input layer, · K neurons in hidden layer, · L neurons in output layer. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 177 It means that we operate with N-elements size input vectors and L-elements size output vectors. It is possible – of course – to discuss more than single hidden layer, but it does not change the general idea of presented approach. here are no connections among the neurons from the same layer, but output signal from single neuron is transmitted as input signal to all neurons from the next layer. So the outputs from the input layer neurons are inputs from the hidden layer neurons and the outputs from the hidden layer neurons are the inputs for the output layer neurons. The only task related to the neurons from input layer is input vector components transfer to neurons from hidden layer. This way there is no need to discuss their implementation – we have only to guarantee the input vector components transfer to all neurons from hidden layer. Neurons from hidden and output layers realise exactly the same operation, but using different data. For hidden layer neurons we can describe the following equation: N ui = f (å xl wli ) ( 1) (1) l =1 where: ui - output value calculated by single neuron from hidden layer – single component of K-elements size vector generated by neurons from hidden layer, xl - component of N-elements size input vector, wli (1) - weight associated with connection from component of input vector xl and neuron from hidden layer indexed by i, f( ) - non-linear, usually sigmoid, neuron activation function. For output layer neurons we can describe the following equation: K yi = f (å u l wli ) (2) (2) l =1 where: yi - output value calculated by single neuron from output layer – single component of L-elements size vector generated by neurons from output layer, ul - component of K-elements size vector generated by neurons from hidden layer, wli (2) - weight associated with connection from component of vector generated by neurons from hidden layer ul and neuron from output layer indexed by i, f( ) - non-linear, usually sigmoid, neuron activation function. As we can notice each neuron calculates the weighted sum which is an argument of neuron activation function. Calculations related to neurons from the same layer can be done in parallel mode, but the sequence of operations ought to be preserved in order to succeeding layers. We assume – of course – that the values of weights are fixed and ready to use – as a product of any proper for Multilayer Perceptron learning algorithm. 3. DATA DEPENDENCE GRAPHS FOR MULTILAYER PERCEPTRON DURING RETRIEVING PHASE A Data Dependence Graph is a directed graph that specifies the data dependencies of an algorithm. In a Data Dependence Graph nodes represent computations and arcs specify the data dependencies between computations. For regular and recursive algorithms, the Data Dependencies Graphs are also regular and can be represented by a grid model. Design of a locally linked Data Dependence Graph is a critical step in the design of systolic array [8]. The Data Dependence Graphs for Multilayer Perceptron retrieving algorithm ought to be discuss individually for each layer. Let’s start from the hidden layer. The input layer – as we noticed before is responsible only for proper distribution of input vector components. So this layer is absent during retrieving algorithm realisation.. The hidden layer includes K neurons (according to the assumption in part 2 of this document) and each neuron from this layer collects signals from N neurons related to input layer. For such topology there are (N ´ L) weights – obtained during learning phase [8]. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 178 The hidden layer ought to be described by rectangular Data Dependence Graph (Fig. 1.). Each node in this graph – excluding the last horizontal line of nodes - is responsible for elementary multiplication product calculation. This means that each node realises the operations described by rule (1), but without summing using proper value of component of N-elements size input vector and the proper weight and without the value of activation function calculation. The local memory of each node should put the value of the single weight obtained earlier after learning algorithm. The size of the graph equals to the size of the weight matrix plus single extra horizontal line with nodes responsible for activation function calculation. Each node of the graph should be loaded by two signals. The first one is the component of the input vector. The second one is the current value of weighted sum calculated by single neuron of Multilayer Perceptron. So each node ought to add the calculated product of multiplication to the loaded previous value of weighted sum signal (1) an pass updated value to the next node in the same column. Such proposed solution requires capacity of the local memory of each node large enough to store the single weight, but this way reduces to minimum the number of data which ought to be transmitted during retrieving algorithm realisation by presented Data Dependence Graph. The operations realised by single neuron from Multilayer Perceptron are described by the single column of the graph. The components of input vector are loaded to nodes by horizontal arcs of the Data Dependence Graphs. These values are passed to the next neighbour on the right hand. The current value of weighted sum generated by single neuron of Multilayer Perceptron is loaded by vertical arcs (Fig. 1.). The update this value is passed to the next bottom neighbour. This way we can observe only point-to-point communication among the nodes which means that presented Data Dependence Graph is local graph. This property guarantees construction of well defined systolic array to realise the retrieving algorithm of Multilayer Perceptron [4]. The nodes from the last – extra horizontal line calculates the value of activation function when the previously calculated weighted sum is the argument. We propose to realise this operation using lookup table instead of analytical calculation. Such solution guarantees quite short time of operation preserving satisfactory level of precision. The only problem is that the local memory ought to store the lookup table. Fig. 1. Data Dependence Graph for hidden layer of Multilayer Perceptron during retrieving algorithm Now let’s try to discuss the Data Dependence Graph for output layer of Multilayer Perceptron. In general steps which ought to be realised are the same, but the number of neurons is different and neurons are loaded by the vector generated by neurons from hidden layer. The output layer includes L neurons (according to the assumption in part 2 of this document) and each neuron from this layer collects signals from K neurons related to hidden layer. For such topology there are (L ´ K) weights – obtained during learning phase [8]. The output layer ought to be described by rectangular Data Dependence Graph (Fig. 2.). Each node in this graph – excluding the last horizontal line of nodes - is responsible for elementary multiplication product calculation. This means that each node realises the operations described by rule (2), but without summing using proper value of component of K-elements size vector generated by hidden layer neurons and the proper weight, and without the value of activation function calculation. The local memory of each node should put the value of the single weight obtained earlier after learning algorithm. The size of the graph equals to the size of the weight matrix plus single extra horizontal line with nodes responsible for activation function calculation. Each node of the graph should be loaded by two signals. The first one is the component of the vector generated by hidden layer neurons. The second one is the current value of weighted sum calculated by single neuron of Multilayer INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 179 Perceptron. So each node ought to add the calculated product of multiplication to the loaded previous value of weighted sum signal (2) an pass updated value to the next node in the same column. Such proposed solution requires capacity of the local memory of each node large enough to store the single weight, but this way reduces to minimum the number of data which ought to be transmitted during retrieving algorithm realisation by presented Data Dependence Graph. The operations realised by single neuron from Multilayer Perceptron are described by the single column of the graph. The components of the vector generated by hidden layer neurons are loaded to nodes by horizontal arcs of the Data Dependence Graphs. These values are passed to the next neighbour on the right hand. The current value of weighted sum generated by single neuron of Multilayer Perceptron is loaded by vertical arcs (Fig. 2.). The update this value is passed to the next bottom neighbour. This way we can observe only point-to-point communication among the nodes which means that presented Data Dependence Graph is local graph. This property guarantees construction of well defined systolic array to realise the retrieving algorithm of Multilayer Perceptron [4]. The nodes from the last – extra horizontal line calculates the value of activation function when the previously calculated weighted sum is the argument. We propose to realise this operation using lookup table instead of analytical calculation. Such solution guarantees quite short time of operation preserving satisfactory level of precision. The only problem is that the local memory ought to store the lookup table. In general Data Dependence Graphs presented above are quite similar and this is a chance to create the single set of Elementary Processors (PE) with switched functions to operate first with hidden layer and next with output layer. The same idea can by adopted for more than single hidden layer if the Multilayer Perceptron has more sophisticated topology. Fig. 2. Data Dependence Graph for output layer of Multilayer Perceptron during retrieving algorithm 4. MAPPING DATA DEPENDENCE GRAPHS ONTO SYSTOLIC ARRAY STRUCTURE 4.1. Processor assignment via linear projection r Mathematically, a linear projection is often represented by a projection vector d . Because the Data Dependence Graph of a locally recursive algorithm is very regular, the linear projection maps an n-dimensional Data Dependence Graph onto an (n-1) dimensional lattice of points, known as processor space [8]. It is common to use a linear projection for processor assignment, in which nodes of Data Dependence Graph along a straight line are projected to an Elementary Processor in the processor array (Fig. 1.). 4.2. Schedule assignment via linear scheduling A scheduling scheme specifies the sequence of the operations in all Elementary Processors. More precisely, a schedule function represents a mapping from the n-dimensional index space of the Data Dependence Graph onto a 1-D schedule (time) space. Linear scheduling is very common for schedule assignment (Fig. 1.). A linear schedule is based on a set of parallel and uniformly spaced hyperplanes in the Data Dependence Graph. These hyperplanes are called equitemporal hyperplanes - all the nodes on the same hyperplane r are scheduled to be proceed at the same time. A linear schedule can also be represented by a schedule vector s , which points in the direction normal to the hyperplanes. For anyr computation node indexed by a vector n in the Data Dependence Graph, its scheduled processing time is sn . INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 180 4.3. Mapping policies r r Given a Data Dependence Graph and the projection direction d not all schedule vectors s are valid for the Data Dependence Graph. Some rmay violate the precedence relations specified by the dependence arcs. For systolic design, the schedule vector s in the projection procedure must satisfy the following two conditions [1]: r r - causality condition: sTe > 0 (3) r e - represents any of the dependence arcs in the Data Dependence Graph r r - positive pipeline period: sTd ¹ 0 (4) This way the rectangular Data Dependence Graphs are converted into linear pipelined systolic arrays. This situation we can observe both for hidden and output layers for Multilayer Perceptron (Fig.3., Fig.4.). The number of elementary processors which are used for array construction equals to the number of neurons in simulated layers of Multilayer Perceptron neural network. Each elementary processor combines all functions described by nodes of Data Dependence Graph placed at the same column of single slab of Data Dependence Graph. If we want to reduce the number of elementary processors we can change the classical systolic structure into ring structure – where each elementary processor is responsible for modelling of greater number of neurons. Of course the reduction of number of elementary processors ought to be done in the way which preserve the same number of neurons for single processor. Fig. 3. Systolic array for hidden layer of Multilayer Perceptron during retrieving algorithm Fig. 4. Systolic array for output layer of Multilayer Perceptron during retrieving algorithm 5. EFFICIENCY OF PROPOSED APPROACH 5.1. Computation time This is time interval between starting the first computation and finishing the last computation of r problem. Given a coprime schedule vector s , the computation time of a systolic array can be computed as [1]: r r r T = max { s ( p - q )} + 1 T r r p ,q ÎL (5) where L is the index set of the nodes in the Data Dependence Graph. In the presented architecture the schedule r vector is defined as: s = 1 , 1 . The total computation time is a sum of computation time related to hidden layer and computation time related to output layer. These two values ought to be calculated independently because operations of each layer are realised in sequence and we have two independent Data Dependence Graph. So total computation time equals: [ ] Tsystol = Tsyshid + Tsysout (6) where: INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 181 Tsyshid - computation time for hidden layer neurons, neurons Tsysout - computation time for output layer For hidden layer we have N+1 elements within vertical axis and K within horizontal axis of space for Multilayer Perceptron (Fig.1.). Based on this remark the computation time we can estimate as: Tsyshid = (N + K )t (7) where t - processing time for elementary processor For output layer we have K+1 elements within vertical axis and L within horizontal axis of space for Multilayer Perceptron (Fig.2.). Based on this remark the computation time we can estimate as: Tsysout = (K + L )t (8) Tsystol = (N + 2 K + L )t (9) Based on (6) finally we can say that: 5.2. Pipelining period rThis isr the time interval between two successive computations in a processor. As previously discussed, if both d and s are irreducible, then the pipelining period equals [1]: r r a = sT d (10) r In the presented approach the schedule vector is defined as: s = [1 , 1] , the projection direction vector r equals: d = [ 0 , 1] . The pipelining period is constant: a = 1. It means the time interval between two successive computations in an elementary processor is as short as possible. The pipelining period is exactly the same for both Data Dependence Graphs and there is no sense to combine them into single value. 5.3. Processor utilization rate Lets define the speed-up factor as the ratio between the sequential computation time and the array computation time, then the utilization rate is the ration between the speed-up factor and the number of processors [1]. sequential computation time (11) speed - up = array computation time speed - up (12) utilization rate = number of processors Sequential computation time always is proportional to the number of nodes, which are responsible for calculations. On the other hand time of computations related to each layer ought to be analysed independently because of operation sequence in Multilayer Perceptron. Based on these remarks the total sequential computation time equals: Tseq = Tseqhid + Tseqout (13) For hidden layer sequential computation time we can estimate as: Tseqhid = (N + 1)Kt (14) For output layer sequential computation time we can estimate as: Tseqout = (K + 1)Lt (15) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 182 Based on (13) finally we can say that: Tseq = ((N + L + 1)K + L ) t (16) where: t - processing time for elementary processor For systolic implementation we need K elementary processors to realise calculations for hidden layer and L elementary processors to realise calculations for output layer respectively. If we assume that all elementary processors need the same time-period to proceed their calculations the speed-up and utilization rate factors for three layer Multilayer Perceptron can be estimated as: speed - up = (N + L + 1)K + L utilizationrate = N + 2K + L (N + L + 1)K + L (N + 2 K + L )(K + L) (17) 6. CONCLUSIONS Summarising, the paper proposed a new methodology for retrieving algorithm of Multilayer Perceptron neural network simulation based on systolic array structure. The discussion is focused on operations which are realised during the following steps of algorithm and the data which are transferred among the calculation units. It is clear which operations can be done in parallel way and when the sequence is necessary. We can notice as very promising the results of estimation of total computation time, speed-up and utilization rate parameters. The proposed methodology is presented for three layer network – because such Multilayer Perceptron is very often used in different tasks. On the other hand there are no barriers to adopt the solution for more sophisticated Multilayer Perceptrons. The main idea is to create next Data Dependence Graphs for next layers and transfer them into linear systolic structures. The time parameters in efficiency calculation will increase in additive way. The proposed approach generates no barriers to tune the Multilayer Perceptron to completely new tasks. The proposed methodology can be used as a basis for VLSI structures which implement Multilayer Perceptron or as a basis for set of general purpose processors – as transputers or DSP processors - which can be used for Multilayer Perceptron neural network implementation. At the end proposed methodology can be also useful for parallel programme realisation of Multilayer Perceptron and can be easily adopted for other feedforward neural nets. The main advantage of the proposed architecture is that it can be easily extended to larger networks and – this way – to different practical task where Multilayer Perceptron can be used as a classification tool for example. REFERENCES [1] KUNG S. Y., 1993: Digital neural networks, PTR Prentice Hall [2] MAZURKIEWICZ J., 1998: A processor pipeline architecture for Hopfield neural network, II SCANN’98 Slovak Conference on Artificial Neural Networks, Smolenice, Trnava, Slovakia, pp.: 158 – 163 [3] MAZURKIEWICZ J., 2001: Efficiency of systolic simulator based on elementary processors with switched functions for Hopfield neural network, VII International MENDEL 2001 Conference on Soft Computing, Brno, 6 – 8 June, 2001, Czech Republic, pp.: 349 - 354 [4] MAZURKIEWICZ J., 2001: SIMD-type simulator for recurrent neural nets, 35th Spring International MOSIS’01 Conference Modelling and Simulation of Systems, Ostrava, 9 – 11 May 2001, Czech Republic, vol. 1, pp.: 173 – 180 [5] MAZURKIEWICZ J., 2003: Systolic realisation of self-organising neural networks, 1st International ICSC 2003 Conference on Soft Computing Applied in Computer and Economic Environments, Kunovice, 30 – 31 January 2003, Czech Republic, pp.: 116 – 123 [6] PETKOV N. 1993: Systolic parallel processing, North-Holland [7] SHIVA S. G., 1996: Pipelined and parallel computer architectures, Harper Collins Publishers [8] ZHANG D., 1999: Parallel VLSI neural system design, Springer-Verlag INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 183 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 184 ACOUSTIC LOCALIZATION OF MULTIPLE VEHICLES Wojciech Zamojski, Tomasz Walkowiak Institute of Engineering Cybernetics, Wroclaw University of Technology ul. Janiszewskiego 11/17, 50-372 Wroclaw, POLAND, Phone: +48-71-3203996 Fax: +48-71-3212677, E-mail: {zamojski,twalkow}@ict.pwr.wroc.pl Abstract: In this paper we describe an approach to localize of multiple vehicles based on sound emitted by them. The problem is known as the acoustic direction of arrival (DOA) in case of multiple sources. The standard techniques for solving this problem were found inadequate in the presence of normal disturbances (such as produced by wind) and wide-band non-stationary signals. The paper presents a method which is based on the MUSIC beam forming technique. Also, a usage of a multilayer perceptron neural network to determine the number of sound sources on a scene monitored by an array of linearly spaced microphones is presented as well as some heuristics algorithms for false sources cancelling. The method is experimentally shown to deal with this problem. Field experiments included scenes with zero, one or two moving vehicles. Keywords: direction of arrival estimation, localisation o f vehicles, neural networks 1. INTRODUCTION The paper discusses a problem of localization of military ground vehicles based on sound emitted by them [5][7][8]. The sources of those sounds are mainly the engines, tracks and wheels of those vehicles. Their characteristics depend on many factors such as speed of a vehicle, environmental conditions and surface type. However, a priori knowledge of those characteristics is not needed. The localization here is understood as a direction angle, not a position on XY space. This problem is known in signal processing as a direction of arrival estimation [2]. There are standard algorithms for dealing with the problem, even in case of multiple sources [4]. These assume that the sound sources are narrow banded with presence of white noise and that the knowledge of the number of sources to be localised and tracked are known. This is not a case of our work since the sound emitted by vehicles is a wide band not-stationary signal. Also the noise could not be assumed as white, it is mainly produced by wind and trees. Moreover the number of vehicles presented in the analysed filed is unknown a priori and could not be determined by non-acoustic observation and measurement. Therefore, the number of presented sound sources (vehicles) has to be estimated on the basis of the received sound signals. In our experiments there were 4 microphones placed in an ULA (uniform linear array) configuration with d=0.6 m span. Therefore, we could detect direction angles form -90° to +90° not the full 360° space. The number of microphones (4) in theory allows to localise and track 3 sound sources, however due to limitation in our field experiments we are focusing on a case when we have 1, 2 or non vehicle present. 2. BACKGROUND The input data are the sounds emitted by military wheeled vehicles such as BRDM, Star and caterpillar vehicles such as BWP and Gozdzik assault gun in a military filed [7]. Besides that, inspected vehicles have particular routes and speeds in certain range (Fig 1). We have marked several points on vehicles routes which allow later analysis of results. The experiments were repeated at two different locations, at dissimilar weather conditions. The vehicles moved around with speeds ranging from 0 to 30 km/h (in most of the case it’s around 15 km/h). We assumed that the area can be treated as flat and without any obstacles obscuring sound propagation and INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 185 producing unwanted reverberations. The mentioned factors allow us to assume that the vehicle are far field point sources which implies that the sound wave coming to the measure point is flat. The sounds were gathered using 4 Brüel & Kjær 4188 microphones. The microphones were connected to 4 Brüel & Kjær 2671 preamplifiers. Next, the signals were gathered in a NEXUS 2693A0S4 amplifier. Finally, the data was converted into digital form using a PC and a dSpace DSP set. Used format was 16 bit and 5 kHz sampling. We have the additional signal which recorded information when which control point was passed by a vehicle. Y X Fig. 1. Experiment field 3. BEAMFORMING The microphones placed in the uniform linear array allow creating a beamforming system. Such system magnifies signals from a given direction, and lowers from all other directions. It acts as a kind of directional microphone. The beamforming is performed by summing M delayed signals (yi (t)) according to equation: M (1) z( t ) = å y i ( t - D i ) i =1 Setting proper values of delays Di one could position a beamforming system for a given direction angle a. The power P(a) of output signal z(t) reaches the maximum when the beam is directed on source of the signal [2]. Therefore, our problem is equivalent to finding K -largest maximum in function P(a), where K is number of vehicles presented in analysed filed. More convenient for digital signal processing is an operation of beamforming system in frequency domain. It could be obtained by Fourier transformation of equation 1 [8]. However in such case we obtain a set of signal powers for each analysed frequency. So the global power is calculated as a sum of individual powers for given frequencies. Since the most of the signal energy is concentrated in the 50-200Hz band (lower band limit is set to ignore the sound generated by wind, i.e. most of the power of signal up to 50 Hz was generated by wind) we limit our frequency analysis to a band 50-200Hz. So we got a new equation for the output power of the beamforming system: P( a) = 200 å e' R f e (2) f =50 where e are steering vectors (denoting the shifts of phase from individual microphones for given frequency): é exp(- jvD1T ) ù ê exp(- jvD T ) ú 2 ú, e=ê ê ú ... ê ú ëexp(- jvDM T )û (3) and Rf is so called spatial correlation matrix, given as a product of the Discrete Fourier Transformation (Yf) of a windowed (multiplied by Hamming window) snapshot (snapshot) of input signals from each microphone Let denote it as: INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 186 R f º Y fY f '. (4) It is worth to mention that the Rf matrix (with size 4x4, since we have 4 microphones) is really a function of a frequency (f). In our experiments we have used an adaptive method of estimating covariance spatial matrix Rf so called MUSIC algorithm. 4. MUSIC ALGORITHM The MUSIC (from MUltiple SIgnal Classification) algorithm has been first presented in [4]. The data model comprises the one assumed in this system. Having an ULA microphone array with M omnidirectional sensors which is pinged by a set of (K<=M) narrowband zero-mean, far-field signals (M – number of signal sources), according to [2] eigenvectors corresponding to K largest eigenvalues li,f of the spatial covariance matrix Rf (given by equation 4) span the signal subspace. This space is orthogonal to the noise subspace which is spanned by the rest of the eigenvectors. Therefore, the adaptive version of spatial covariance matrix Rf could be calculated as [4]: R MUSIC f -1 M = å v i, f v i, f ' , (5) i = K +1 where eigenvectors (vi,f ) are sorted according to corresponding eigenvalues (li,,f): l1, f ³ l2, f ³ K ³ lM -1, f ³ lM , f > 0 . (6) Therefore, the inversion of MUSIC covariance matrix is a base for calculation of the beamforming adaptive power: 200 [ -1 P(a ) = å e' R MUSIC e f f =50 ] -1 . (7) 5. NUMBER OF VEHICLES RECOGNITION BY NEURAL NETWORK In our case we don’t know the number of the sources K, so we are unable to form the definite signal subspace matrix. On the other hand decomposition of matrix R gives information on the eigenvalues (li,f) which can be interpreted as the powers of consecutive potential sound sources. Since N is unknown the assumption is that there are M pseudo-sources some of which correspond to the real sources and the remaining ones to noise. One has also to admit that preceding considerations are about narrow-band sources; however in this case the moving vehicles are rather wide-band emitters. 6 8 l1 l1 6 4 4 2 2 0 0 50 100 150 200 250 0.8 f [Hz] 0 50 100 150 200 250 f 50 100 150 200 250 [Hz] 0.4 l2 0.6 0.2 0.2 0.1 0 l2 0.3 0.4 0 0 50 100 150 200 250 f[Hz] 0 0 f [Hz] INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 187 0.06 0.5 l3 0.4 l3 0.04 0.3 0.2 0.02 0.1 0 0 50 100 150 200 250 0 f 0.4 0.04 l4 0.3 0.02 0.1 0.01 0 50 50 100 150 200 250 50 100 150 200 250 f l4 0.03 0.2 0 0 100 150 200 250 f 0 0 f Fig. 2. Pseudspectro for 1 (left side) and two vehicles (left side) The proposed solution is introduction of pseudo-spectra of the pseudo-sources. A pseudo-spectrum is a set of k-th eigenvalues in all processed bands (e.g. 1st pseudo-spectrum comprises of largest eigenvalues of the estimate of the covariance matrix R in each band). In case of this system, there are 4 pseudo-spectra (four eigenvalues, since we have had four microphones). As one can see on the figure 2, noise pseudo-spectra have common structure (i.e. follow the 1/f pattern) and contain much less energy than the pseudo-spectra associated with real sources. Decision on what the real number of sources is taken by the neural classifier [1]. The multilayer perceptron neural network, with it’s very good ability to estimated the unknown function, uses the date from the pseudo-spectra to determine the number of sources. Using all data contained in the pseudo-spectra would imply too much complexity so we decided on deriving some measures describing the pseudo-spectra. The firs class of such measures used is connected with the power of the sources whereas the second one deals with the structure of each pseudo-source. The power of a pseudo-source is expressed in the following equation: Pi = å l i , f . (8) f The measure describing the structure is the relation between the power below 50Hz and the whole Pi: ål i, f Ki = f <50 Hz Pi . (9) The final feature vector consisted of 8 components: { log(P1),…,log(P4),log(K1),…,log(K4) }. The logarithmic vector space was dictated by the fact that most relations between powers are either divisions or multiplications. Therefore, the neural network has 8 input neurons corresponding to feature vector, 22 hidden layer neurons - selected experimentally and 2 output neurons. The output coding was 0,0 for no vehicles; 1,0 for one vehicle and 1,1 for two sources. The network was trained using the Levenberg-Marquardt [1] algorithm which gave the fastest convergence. [6] 6. MAXIMUM DETECTION As it was stated in chapter 4, the localisation is equivalent to finding local maximums in function P(a) given by equation 7. In a case when the neural network gives an answer that there is only one vehicle the problem is trivial - just a simple maximum detection. However, in a case of two vehicles, it is more sophisticated. Several disturbances are presented in a power signal due to background noise. Therefore, a heuristic algorithm for maximum detection was developed. In the first step the power function is convoluted INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 188 with 7-point Hanning window. Next, local minimums MINq and maximums MAXq in smoothed power function is detected (i.e. local maximum is presented for a given angle when power values for previous and next arguments are lower). Maximums, below given threshold value Tp are rejected. The threshold value is calculated based on all detected maximums: T p = 0.1 1 Q å MAX q , Q q =1 (10) where Q is a number of maximums. In the next step, some disturbance resulting in false maximums are rejected. Two kinds of disturbance could be distinguished, presented in Fig. 3. The detection of false maximums is done by analysis of relative values of maximums compared with nearest minimums: (MAXq- MINq-1) / MAXq i (MAXq- MINq)/ MAXq.. When these both values are below a given threshold (experimentally set to 0.2) then the maximum is rejected (assumed as type A disturbance). And similarly in a case of disturbance type B, when one of theses values is lower below other threshold (experimental set to 0.05). Finally, two largest maximums are selected and its argument values give the estimated values of position (angles) of two vehicles. Type A P o w e r Type B P o w e r MAX q MAXq MINq MINq MINq-1 angle angle Fig. 3. Kinds of disturbance in power signal a) a α’’ D2 d) a 22 next g) 2 2 next α’ D3 D4 a 12 next D4 D2 D1 α’’ a 12 next α’’ D3 α’’ D1 c) α’ a 22 next a 12 next D4 D3 D2 f) D1 α’ D3 D2 a 12 next a 22 next D2 D4 a2 D2 D3 2 next α’ D1 D1 a 12Dnext3 D1 D4 a 12 next D2 α’’ D1 D3 D4 a 22 next D2 a 22 next D3 D1 a 12next D3 α’ α’ α’’ D2 a 22 next a 22 next l) α’’ a 12 next α’ D2 D4 k) D3 α’ α’ D4 i) a 12 next a 22 nextα’’ α’’ D1 D2 α’’ α’ D4 j) α’’ D3 a 22 next h) α’ a 12 next a 22 next D4 D2 e) D1 α’ D3 a 12 next D1 α’’ D4 b) D1 D4 a 12next Fig. 4. Possible estimated position of vehicles in analyzed field INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 189 7. TRACING ALGORITHM The results of previous localizations step requires further preprocessing. The main aim of this part of the system is to assigned two angles to a given object and averaging of results. At first two angles: α’ i α’’ from previous step are assigned to a calculated by MUSIC method directions: a 12 and a 22 . This is done based on the nearest neighborhood rule. At first four distances: D1, D2, D3 and D4 are calculated (see Fig. 4) according to following equations: D1 = || α’ - a 12 ||, D2 = || α’’ - a 12 ||, D3 = || α’ - a 22 ||, D4 = || α’’ - a 22 ||. (11) Next, directions are assigned according to simple rules, deduced from figure 4: If D1+D2 ≤ D3+D4, then a 12 = α’ and a 22 = α’’. (12) If D1+D2 > D3+D4, then a = α’’ and a = α’. 2 1 2 2 Having assigned the directions to traced objects the directions are averaged. The averaging is based on an idea that the vehicles are driving in linear directions, i.e. changes of directions are slow. This assumption is justified for military vehicles. Therefore, linear regression model is used. However in some cases the directions achieved from MUSIC algorithm are far away A single frame of N from previous direction assumptions. It is caused samples by signal disturbances, for example by bad weather conditions. Therefore, the heuristic y1(t) algorithm was developed. It is performed for y2(t) each of two analyzed direction as follows: y3(t) · Linear prediction of current y 4(t) direction based on ten previous direction, resulting in a new . · Calculation of current position FFT based on MUSIC algorithm, i.e a. · Linear prediction of next direction MUSIC power MUSIC power Control vectors based on nine previous directions determination for determination for generation for hypothesis that hypothesis that and current one form MUSIC angles in range n is 1 n is 2 a Î - p, p block, resulting in a next . · Finally, current position is given by: 2 greatest local maxima of Power maximum for all a Neural power for all a a ' = Aa new + Ba next - if a differs · network more then a given threshold for number of vehicles from a new or a next ,(A+B=1). Tracing Tracing determination · a ' = Ca new + Da next + Ea otherwise, (C+D+E=1). a 12 , a 22 – DOA's for two vehicles a1 – DOA of one vehicle (13) Final result Fig. 5. System overview 8. SYSTEM OVERVIEW AND PERFORMANCE The presented above system (see Fig 5 for system overview) was tested on real data gathered in a military filed (see chapter 2). The exemplar results for BWP and Gozdzik vehicles are presented on Fig 6. Some problems with estimation of the direction angle of the vehicle localized far away from the microphones (for large values of angles) could be noticed. Such effect is due to the fact that the vehicle sound was “blown INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 190 Direction [°] away” by the very strong wind. 80 60 40 20 0 -20 -40 -60 -80 10 20 30 40 50 60 Time [s] 70 80 90 100 Fig. 6. Exemplar results of two vehicles localizations Performed experiments showed quite good effectiveness of the presented method. The usage of the MUSIC algorithm, tracking and neural networks for number of vehicle estimation resulted in a good overall estimation of the direction angle of the vehicles. In the area from -45° to +45° the direction angle could be estimated with an error below 3°. For distances up to 100 m the error of localization of the vehicle position is within the length of the observed vehicle (± 5.24 m). In the area beyond -45° to +45° part of the half-plane results are less satisfactory. It is probably due the fact of a larger distance from the vehicle to the position of array of microphones (more then 100 m) and therefore low signal to noise ratio. But also in this area, the system is able to estimate the position quite well. However, some real problems could be encountered in a bad weather conditions. On the other hand, this is a limitation of all methods based on acoustic signals. The overall result for neural network estimation of number of vehicles is more then 90% [6]. The Neural network estimator has some problems when two vehicles are meeting together resulting in one sound source (two vehicles on the same position), therefore it it’s physically impossible to reach 100% efficiency. 9. CONCLUSION AND FURTHER WORK The researches related to the mobile object localization based on acoustic information can be used as the basis for the new systems to support the contemporary military equipment. The system observes the stage in a passive way, so it is very hard to find and destroy the military devices equipped by such intelligent attachment. The proposed algorithm of localisation can be implemented in a very efficient way using DSP technology (some primary work has been done on it) and microcontrollers as an intelligent robot. We think that future investigation it to extend the system the other kinds of signals achieved from moving vehicles, like visual one (from video camera) or seismic. It could improve results especially in bad weather conditions. Seismic signals could be analysed in similar way than acoustic signals presented here however video signal requires more sophisticated localisation algorithms, see for example [3] for some primary experiments. REFERENCES: [1] BISHOP, Ch. M.: (1995) Neural Networks for Pattern Recognition, Clarendon Press, Oxford. [2] JOHNSON, D.H., Dudgeon D.E. (1993) Array Signal Processing: Concepts and Techniques. Prentice Hall, Englewood Clifs, New York. [3] MAMICA, J., WALKOWIAK, T. (2003) Genetic Algorithm for Motion Detection, MENDEL’2003, 9th International Conference on Soft Computing. June 4-6, 2003, Brno, Czech Republic, pp. 29-34. [4] SCHMIDT, R.O. (1979) Multiple emitter location and signal parameter estimation. Proc. RADC Spectrum Estimation Workshop, Rome, 243-258 [5] WALKOWIAK, T. (2003), Acoustic Localization of Vehicles, ICTINS’2003 International Conference on Information Technology and Natural Sciences, October 19-21, 2003, AL-Zaytoonah University, Amman, Jordan, pp. 303-306. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 191 [6] WALKOWIAK, T., ZOGAL, P (2003) Neural Network Approach to Acoustic Detection of Number of Vehicles, INNSC 2002 – 6th International Conference Neural Networks and Soft Computing, June 11-15, 2002, Zakopane. Heidelberg; New York: Physica-Verlag, cop. 2003, series Advances in Soft Computing, pp. 909-914. [7] ZAMOJSKI, W. et al. (1999) A System for Recognition and Tracking of Moving Objects on the Basis of Acoustic Information (in Polish). Reports of the Institute of Engineering Cybernetics, Wroclaw University of Technology, PRE 17/99, Wroclaw, Poland. [8] ZAMOJSKI, W., MAZURKIEWICZ, J., WALKOWIAK, T. (1997), Mobile Object Localisation Based on Acoustic Information, Proceedings of the IEEE International Symposium on Industrial Electronics, Guimaraes, Portugal, July 7-11, 1997, vol. 3, pp. 813-818. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 192 THE FUZZY RATIONAL DATABASE SYSTEM FSEARCH 2.0 Josef Bednář Department of Stochastic and Optimisation Methods, Institute of Mathematics Faculty of Mechanical Engineering, Brno University of Technology, Technická 2, 616 69 Brno, Czech Republic E-mail: [email protected], Phone: +420-541142550, Fax: +420-54114 2800 Abstract: The main purpose of this paper is the theoretical development of methods of fuzzy logic and fuzzy searching in fuzzy sets especially in fuzzy sets with a polygonal membership function. On the basis of these methods a new approach has been developed to the theoretical solution of fuzzy relational database systems. The results will be applied to the creation of the fuzzy database system Fsearch 2.0 as an upgrade of the one-dimensional fuzzy database system Fsearch 1.0. Keywords: fuzzy metric, fuzzy distance, fuzzy metric space, fuzzy number, fuzzy point 1. INTRODUCTION In a classical database, the information is represented by exact data. Such database systems can only deal with unambiguous and well-defined data, but a database is a computer model of the real world. There exist uncertain or ambiguous data and information which cannot be defined in a certain and well-defined form by any means in the real world. Since in everyday life, we often make decisions based on such fuzzy data. An approach exists in which the fuzzy set theory is used to designing the fuzzy relational database. The emphasis of this paper is put on the fuzzy searching in fuzzy sets because information, which is represented as a fuzzy set in a fuzzy database, is available only when we can find it in great quantity other information. If we search item in classical databases, then we probably set an interval and we look for all items, whose values lie within the interval and we can order these items by different rules. This way is a special case of a fuzzy question, because the interval is a special case of a fuzzy number. We generalise a question so that it is represented by a fuzzy number. The values of quantity are numeric data of a vague character, therefore they are represented by fuzzy numbers. These fuzzy numbers are called information. We look for all items, whose intersection of information and a question is not empty, this intersection is called the answer. We can order these items by different rules. 2. FUNDAMENTAL DEFINITIONS OF FUZZY SETS THEORY We will use the following basic notion of the theory of fuzzy sets [ 1, 2 ]. Definition: Suppose that X ¹ Æ is a universe of discourse (i. e. the set of all possible elements to be considered with respect to a vague property. Then a fuzzy set A in X is defined as a set of all ordered pairs {(x, μ A (x)); x Î X; μ A (x) Î 0,1 }, where mA: Xàá0,1ñ is the membership function of a fuzzy set A. The number mA(x) is the grade of membership of x in A, from 1 for full belongingness to 0 for full nonbelongingness, through all intermediate values. Definition: The support of a fuzzy set A = (X, mA) is defined as the ordinary set supp A = {x Î X; mA(x) > 0}. The kernel of a fuzzy set A = (X, mA) is defined as the ordinary set ker A = {x Î X; mA(x) = 1}. The complement of a fuzzy set A = (X, mA) is defined as the fuzzy set A =(X, m A ), such that m A (x) = 1 - mA(x) for " x ÎX. The height of a fuzzy set A=(X, mA) is defined as a number INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 193 h(A) = sup m A ( x ) , X if h(A) = 1, then a fuzzy set A is called a normal fuzzy set. Definition: A fuzzy set A = (X, mA) is included in (is a subset of) a fuzzy set B = (X, mB), A Í B, if and only if mA(x) £ mB(x), "xÎX. Two fuzzy sets A = (X, mA), B = (X, mB) are equal, A = B, if and only if mA(x) = mB(x), "xÎX. A fuzzy set A = (X, mA) is empty, A = Æ, if and only if mA(x) = 0, "xÎX. Definition: The intersection of two fuzzy sets A = (X, mA), B = (X, mB), is defined as a fuzzy set C = (X,mC) = = A Ç B, such that mC(x) = min{mA(x), mB(x)} for " x ÎX. The union of two fuzzy sets A = (X, mA), B = (X, mB), is defined as a fuzzy set C = (X, mC) = A È B, such that mC(x) = max{mA(x), mB(x)} for " x ÎX. Definition: Suppose aÎá0, 1ñ, then an a-cut of a fuzzy set A = (X,mA) is defined as the ordinary set Aa Í X such that, Aa = {xÎX; mA(x) ³ a}, for "aÎá0,1ñ. Definition: A fuzzy number A = (X, mA) is defined as a fuzzy set, where X Í R, which is normal and convex (i. e. whose all a-cuts are convex) and mA is continuous in parts. 3. THE CHARACTERISTICS OF FUZZY SETS The answers (fuzzy sets) C on a question B can be ordered by different means. We order answers from a most satisfactory answer to a least satisfactory answer. The fuzzy database will be used by different users. Therefore, we select several right characteristics of answer. The answers will be ordered downward by height of answers C and following characteristics. These characteristics and their properties are described in Mendel’99 [3]. Definition: The measure of a fuzzy set A = (X, mA) is defined as a number m( A ) = ò mA (x)dx = ò mA (x)dx . X sup pA Definition: The relative measure of a fuzzy set C = (X, mC) with respect to fuzzy set B = (X, mB) is m(C) m(C/B) = , if m(B) ¹ 0. defined as a number m(B) Note: To comprehend following n-dimensional characteristics we must note that the question (information, answer), in more dimensions which are designated as a1, a2, …, an, is defined as the cartesian product of partial questions ( information, answers) in single dimensions. The measure and the relative measure of an n-dimensional fuzzy set can be defined, but this approach is unsuitable with respect to the numerically demanding calculation. Therefore the characteristics of an n-dimensional fuzzy set are defined by the medium of characteristics of partial questions (information, answers) in single dimensions. Designation: the question B(ai) the question B in dimension ai, the information A(ai) the information A in dimension ai, the answer C(ai) the question C in dimension ai, the answer C(a1, a2, …, an) the n-sizeable fuzzy set C(a1, a2,…, an), where C(a1, a2,…, an) = C(a1) ´ C(a2) ´… ´ C(an). Definition: The height of a n-dimensional answer C(a1, a2, …, an) is defined as the height of the fuzzy set C(a1, a2,…, an) = C(a1) ´ C(a2) ´…´ C(an). Height of an n-dimensional answer C(a1, a2,…, an) is denoted by h(C(a1, a2,…, an)). Therefore h(C(a1, a2,…, an)) = min( h(C(a1)), h(C(a2)),…, h(C(an))). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 194 Definition: The average height of an n-dimensional answer C(a1, a2,…, an) is defined as the number n hg(C(a1 , a 2 ,..., a n )) = n Õ h(C(a )) . i i =1 The minimal measure of an n-dimensional answer C(a1, a2,…, an) is defined as the number m(C(a1, a2,…, an)/B(a1, a2,…, an)) = min( (C(a1)/B(a1)), h(C(a2)/B(a2)),…, h(C(an)/B(an))). The average relative measure of an n-dimensional answer C(a1, a2,…, an) is defined as the number The measure of an n-dimensional answer is not defined, it is not standardised. n mg(C(a1 , a 2 ,..., a n ) / B( a1 , a 2 ,..., a n )) = n Õ m(C(a ) / B(a )) . i i i =1 4. THE MODIFICATION OF INTERSECTION Question Ç Information Answers Evaluation characteristics of answers Optimal item selection Chart 1. Chart 1 of fuzzy searching in fuzzy information, which is stored in a fuzzy database, has been developed on the basis of foregoing considerations. Different users will use the fuzzy database and they can require a less uncertainty of the answers. Therefore we define two operations, which are modifications of the intersection, but they are more rigorous. Definition: The confident intersection of two fuzzy sets A = (X, mA), B = (X, mB), is defined as a fuzzy set C = (X,mC) = A Ç B, such that mC(x) = max{0, mA(x) + mB(x) - 1}. Definition: The algebraic product of two fuzzy sets A = (X, mA), B = (X, mB), is defined as a fuzzy set C = (X,mC) = A · B, such that mC(x) = mA(x) . mB(x). Theorem 1. If two fuzzy sets A = (X, mA), B = (X, mB) exist, then A Ç B Í A · B Í A Ç B. Theorem 2. Suppose that classical sets A, B with characteristic functions cA,cB: U ® {0, 1} are defined in a universe X. If we understand characteristic functions cA and cB as membership functions mA and mB, then A Ç B and A · B and A Ç B coincide with the intersection of classical sets. Theorem 3. If a fuzzy set A = (X, mA), then A Ç A = Æ , A Ç A ¹ Æ and A · A ¹ Æ. Definition: A question B1 is a contraction of a question B, if the fuzzy number B1 = (X, mB1) is included in the fuzzy set B = (X, mB). Theorem 4: Suppose that fuzzy sets and fuzzy numbers are defined in a universe X Í Rn, information A(a1, a2, …, an) is fixed, a sequence of questions {Bm(a1, a2,…, an)} is defined so that question Bm+1(a1, a2,…, an) is a contraction of a question Bm(a1, a2, …, an), a sequence of answers {Cm(a1, a2, …, an)} is defined either so that answer Cm = A Ç Bm or so that answer Cm = A · Bm or so that answer Cm = A Ç Bm , mÎN. Then 1. The sequence of the heights of answers Cm(a1, a2, …, an), {h(C1),h(C2),…,h(Cm),…} is non-increasing and the domain of this sequence is in interval H = á0,1ñ. 2. The sequence of the average heights of answers Cm(a1, a2,…, an), {hg(C1), hg(C2),…, hg(Cm),…} is nonincreasing and the domain of this sequence is in interval H = á0,1ñ. 3. The sequence of the minimal relative measures of answers Cm(a1, a2,…, an), {mg(C1/B1), mg(C2/B2),…,mg(Cn/Bn),…} is not generally monotonous and the domain of this sequence is in interval H = á0,1ñ. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 195 4. The sequence of the average relative measures of answers Cm(a1, a2,…, an), {mg(C1/B1), mg(C2/B2),…,mg(Cn/Bn),…} is not generally monotonous and the domain of this sequence is in interval H = á0,1ñ. 5. The sequence of products of the average height of a answers Cm(a1, a2,…, an) and the relative measure of answers Cm(a1, a2,…, an), { hg(C1).mg(C1/B1), hg(C2).mg(C2/B2),…, hg(Cm).mg(Cm/Bm),…} is not generally monotonous and the domain of this sequence is in interval H = á0,1ñ. 5. THE FUZZY DATABASE SYSTEM FSEARCH 2.0 The foregoing results were applied to the creation of the fuzzy database system Fsearch 2.0 for metal materials. In this database, fuzzy sets are represented as trapezoid fuzzy numbers [3, 4] whose membership æ æ x -a x -d ö ö , ,1÷, 0 ÷÷ . The trapezoid fuzzy number is defined by four function is given by μ A (x) = maxçç minç è b-a c-d ø ø è numbers and that is evident from its membership function, therefore information are stored in the fuzzy database as four-dimensional vectors and a question is represented as four-dimensional vector too. Since the search algorithm has been described by now, next we demonstrate a run of the fuzzy database in a particular example. Example: The identification marks of metal materials and values of material characteristics are stored in a material.dat file. The editing window, which is exemplified in Fig. 1, is intended for file management. For illustration, dimensions, which are designated as a1, a2, a3, a4, are in turns tensile strength, hardness, normalisation anneal, annealing point. If fuzzy database contains some data, then we can insert a question, which is exemplified in Fig. 2. · Select materials characteristics for searching. · Complete Table 1 on editing panel. · Press a button Solve. The characteristics, which are defined in part 2, are displayed on Table 2. We can order downward every material in accordance with these characteristics in panel, which is above the Table 2. · If we press a button Draw, then graphic window, which is exemplified in Fig. 3, is open and particular answers for materials, which are specified in Table 2, are displayed in this window. Standard number of material Material characteristics Figure 1. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 196 The question Table 1 Solve Downward order by Draw Figure 2. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 197 Figure 3. Acknowledgement: This paper was supported by research design CEZ: J22/98: 261100009 “Nontraditional methods for investigating complex and vague systems”. The author would like to acknowledge Associate Professor Zdeněk Karpíšek for his advice and support. Thanks go to Pavel Štarha for his assistance in programming fuzzy database FSearch 2.0 . REFERENCES [1] NOVÁK, V.: Fuzzy množiny a jejich aplikace. SNTL, Praha 1990. [2] DUBOIS, D. and PRADE, M.: Fuzzy Sets and Fuzzy Logic.Prentice Hall, New York 1980. [3] KLIR, G and YUAN B.: Fuzzy Set and Fuzzy Logic: Theory and Applications. Prentice Hall PRT, New Jersey 1995. [4] BEDNÁŘ, J.: The Theoretical Properties of Fuzzy Searching. In: Mendel´99. 5th International Conference on Soft Computing. Brno 1999, pp. 199-204. [5] ROGER JANG, J. S. and GULLEY, N.: Fuzzy Logic Toolbox User´s Guide. Matlab Works, Inc., Natick 1995. [6] KACPRZYK, J.: Fuzzinessin Database Management Systems. Physica-Verlag Heidelberg 1995. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 198 A SPECIAL DATABASE FOR THE MEASURING IN FIELD Dalibor Bartoněk Institute of Geodesy, Faculty of Civil Engineering, Brno University of Technology. Veveří 95, 662 37 Brno Czech Republic. Tel. 54114 7204, Fax: 54114 7218, E-mail: [email protected] Abstract: The measuring in field is the most commonly used activity in various industrial branches. To be the results useful the data must be stored in a suitable database. There is a problem: to use either one of most of commercial database systems or to create the special purpose designed database? This question is discussed and an own approach of the database design is presented in this paper. The goal is the designed database with minimum of memory capacity occupied and the most rapidly information provided. It is supposed that the basic database design rules e.g. to avoid redundant data (normalization), ensure that the relationships among attributes are represented or facilitate the checking of updates for violation of database integrity constraints are respected. A special optimization method for common attributes reduction, for storing strings of variable length and for database reorganization is used. The database model was implemented in the information system for anticorrosive protection of pipeline used in gas enterprises in the Czech Republic. Keywords: Data model, entities, attributes, object, period of measuring, measured values.Database, object, measured values, quantities, relational model, E-R diagram, entities, common and individual attributes, equivalent classes, overlay table, optimization. I. INTRODUCTION The measuring of quantities on the object in field belongs to common activities in many spheres of individual usage. The gained values serve either to ensuring the run of certain devices or they inform the users e.g. in the frames of prophylactic check up. In order to take effective advantage of these values it is necessary to store it in optimal designed database. There are two possibilities: 1. to use a commercial database software product (dBASE, MS Access Oracle etc.), 2. to create own special purpose database in any programming language (Borland Pascal, Delphi, C etc.). The commercial database systems, however, were designed for general purpose with the wide offer of functions which is connected with 2 main problems: · a lot of functions lead to complications of the system and inconvenient usage (some functions would hardly be made use of, others the missing ones would have to be programmed by the experts themselves), · by complications we mean tedious training of the operators who lack in experience with computers. Having discussed all “pros and cons“ the authors together with a group of potential users decided to make their own solution in Borland Pascal programming language (MS DOS application), later in Borland Delphi (application for Windows). The results were used in the system for anticorrosive protection of gaspipeline. 2. ANALYSIS OF THE MEASURING PROCESS IN FIELD Lets ponder about measuring on the objects of various types. On each of them we measure different quantities in different time. The moment of measuring can be: · regularly (periodical e. g. monthly, quarterly, yearly etc.) · irregularly that means the measuring are done at random according to our needs. The value of the measured quantity is under these hypotheses a function of two variables: v = f(t, p), (1) where INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 199 v is measured value, t is the type of the object, p is the period of the measuring. The value of the measured quantity depends upon the type of the object and the period of the measuring. It means that of every type of the object various quantities depending upon the time of the measuring can be measured. In every period (monthly, yearly etc.) a different kind of quantities on the given type of the object of all types bat at the same time the required measuring plan which is the description of the individual period of the measuring further function relationship (1). There are several basic types of the object. Every object can be either of basic type or the combination of several basic types. This fact should be taken into consideration as well. 3. CONCEPTION OF THE DATA MODEL For the storage of all data measured in field we have decided to choose the most commonly used the relational database. Relational data model will contain these entities: · Object and sub-objects with the ISA hierarchy, · Period of the measuring, · Measuring (with measured values), · List of quantities representing the equation (1). Object and period are the strong entity sets, whereas measuring and measured values are the weak entity sets, because they are dependent on the previous entities according to the equation (1). The list of quantities represents the relationship between period and measuring and realizes the function in the equation (1). The objects attributes from the point of view of belonging to the objects can be divided into 3 classes: a) key attributes - Ak, b) common attributes for each type of the object - Ac, c) individual attributes, which distinguish the individual types of the objects - Ai. These attributes can be from the point of view of data stored length divided into two groups: · fixed length, · variable length. E-R diagram of the data model of the measuring in field shows figure No 1. *Akq *Ako Aco object list of quantities measuring ISA *Akp Aom *Ako period +Adm subobject1 subobjectn Aio1 Aion measured values +Adv Aop Aoq +Adm *Ako Aov Explanations: Ako, Akp and Akq – primary key attributes of the object, period and list of quantities, Aco – common attributes of the object, Aio1 … Aion – individual attributes of the sub-object No 1, … ,n, Adm – discriminator of the measuring entity, Adv – discriminator of the measured values entity, Aom – other attributes of the measuring entity, Aov – other attributes of the measured values entity, INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 200 Aop – other attributes of the period entity, Aoq – other attributes of the list of quantities entity, …… cardinality 1, cardinality N …… non obligatory relationship double rectangles – weak entity set, double ellipses – grouped attributes Fig. 1. E-R diagram of the data model of the measuring in field 4. IMPLEMENTATION OF THE DATA MODEL The data model on Fig. 1 was transformed into tables of relational database according to the rules described in (Korth, H. F., Silberschatz, A., 1996), (pokorný, J., 1998). The tables are of two types: · Static, that means they have a fixed structure (tables of object, sub-objects, period, measuring and list of quantities), · Dynamic – the structure of which is variable according to the equation (1) – table of measured values. As for the static tables, the most complicated problem was with the creation of tables of object and subobjects (ISA hierarchy). To avoid data redundancy the method of attributes overlay was used. The principle consists in dividing the sub-objects into basic types so that any sub-object can be composed by combination of one or some objects of these basic types. Thus the number of sub-objects is very (approximately by one order) reduced. Then is the overlay table created the structure of his is shown in Fig. 2. Rows in overlay table consists attributes Aio1 – A ion, in columns are basic types of sub-objects. In the individual cell we write either symbol “*” (string date type) or “#” (other date type). Another reduction of the database structure is by creation of overlay table achieved. The process consists in 2 phases: 1) 2) Deleting all of the columns in the table 1, which have the null values in all own rows. Objects of these types are sufficiency covered by attributes in the basic table of object (common attributes). We obtain a new table. In the new table we step by step join columns according these rules: a) We join these columns that have in his rows the some symbols. b) Further we join columns that is different in n symbols only according these criterion: n m i =1 i =1 nå ci tOj < må d i tOk where n m ci di tOj tOk is number of attributes, in which are lower-level entity sets Oj and Ok (sub-objects) different, is number of attributes, in which are lower-level entity sets Oj and Ok (sub-objects) the same, is length of i. attribute (common for lower-level entity sets Oj and Ok) in Byte, is length of i. attribute (in which are lower-level entity sets Oj and Ok different) in Byte, is estimation of frequency of occurrence lower-level entity sets Oj in database, is estimation of frequency of occurrence lower-level entity sets Ok in database, Table 1: Attributes overlay table Attributes/SubO1 objects A1 * A2 Am (2) O2 On * # # # After the reduction process every row in the table 1 must contain at least one character * or # (completely overlay). Columns in the table 1 represent equivalent classes of sub-objects. We can define 2n tables in the database, where n is number of columns in the table 1. N tables are for attributes of string date type (there are a special compression method for string saving used on the internal database level) – see Fig. 3, n tables for INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 201 attributes of non-string attributes) – see Fig. 4. Internal structure of the database for saving records of objects is in Fig. 2 and. We use two files; first for the basic data e.g. common attributes Aco1, that must have a value (not null), second file contains two groups of attributes: Aco2 (common attributes, that may be null value), Aio, (individual attributes of object type). The field “Info object” uses trigger for record resize. The measuring data are saved also in two files. First file for all attributes of measuring Am = Ako + Adm + Aom, second file for measured values. This file has 3 parts: 1. “info-measuring” serves for trigger to record resize, 2. “bitmap” is sequence of bits (flags) corresponding to measured values. If the flag bit is zero, the correspondent value is null, if it is “1”, the correspondent value is valid. D1 is length of bitmap, D2 is length of array of measured values. Dynamic table for measured values is created with the help of the table of quantities list part of which is shown in the table 2. This table realizes the dependency in equation (1). The quantity Qi is loaded into the dynamic table of measured quantities if is measured on one of the sub-objects types <T1, … Tm> and in one of the periods <P1, … Pn> written in the row - see table 4. Basic data of the object Extended data of the object File 2 – extended Aco1 File 1 head Aco2 Aio Info_object pointer size Fig. 2. Internal structure of object records saved in the file Table 2. Part of text attributes of sub-object type Attribute i Attribute i+1 … … … … string i string i+1 … … … … Attribute m … string m … file ni string i ni+1 string i+1 nm ni+2 string m number of string characters Fig. 3. Compression method of saving the text attributes into the file File 1 – head data Bitmap File 2 - values Am Info_meas. 2B D1 pointer Table 3. Measured quantities Q1 Q2 … Qk … … … … 1.45 12.61 … 4587 … 2B … 4B …4B … 4B 1261 4587 BM D2 145 D1 D2 Fig. 4. Method of saving non-text attributes and measured values into the file INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 202 Table 4: Part of the quantity list Quantities Qi Sub-object types T1 ..... Periods of measuring Tm P1 ….. Pn 5. DATABASE UPDATE ACCORDING TO THE USE IN WINDOWS At present the database model was updated to run uder the Windows operating system. The MS Access was used as the base relational database system. The source code of the application is written in Borland Delphi and it is connected with the MS Access by ODBC interface through the data modul. The conception is shown in Fig. 5. MS Access database file (tables) Data module (database, data sources) + BDE components (Borland Delphi Engine) – (tables, queries) ODBC link Application Fig. 5. Conception of the current database solution The advantage of this conception consists in simple data manipulation (queries, filters etc.), whereas the insert database operation has to be in code source ensured against the primary key collision. If the primary key collision occurs, the MS Access causes exception error in the application. 6. CONCLUSION The database model was used in the information system for anticorrosive gas-pipeline protection in gasworks in the Czech Republic. This system with the name GAS-ACOR has already been running for more then 10 years and it had been upgraded according to the users needs. First version was made in Borland Pascal for MS-DOS. The last upgrade consisted in introducing module for universal file print in the network environment. This module was created in Borland Delphi. The authors provide the users with help in the form of hot line or personally. Last year it has been worked on the converting of the GAS-ACOR system from the MS-DOS system into Windows. Source modules are made in Borland Delphi. The testing run of the new system is operated in South-Moravia gasworks in Brno. The introduction of GAS-ACOR into practice will mean a considerable time saving and the increase of the efficiency of the anticorrosive protection. REFERENCES [1] KORTH, H.F, SILBERSCHATZ, A., (1996), Database System Concepts. Third Edition. McGRAW-HILL. [2] POKORNÝ, J., (1998), Databázová abeceda. SCIENCE, Veletiny, [in Czech]. [3] SEDLÁČEK, S., BARTONĚK, D., (1996), GAS – ACOR. User manual. SHINE, Brno, [in Czech]. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 203 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 204 INTERNET - A MODERN E-LEARNING MEDIUM Marek Woda, Tomasz Walkowiak Institute of Engineering Cybernetics, Wroclaw University of Technology ul. Janiszewskiego 11/17, 50-372 Wroclaw, POLAND, Phone: +48-71-3202969, +48-71-3203996 Fax: +48-71-3212677, E-mail:{ mwoda,twalkow}@ict.pwr.wroc.pl Abstract: The Word Wide Web, popular called Internet, as contemporary medium for distant learning will be presented.Some techniques and technical aspects of substitution traditional methods will be also discussed. Conventional techniques of teaching are in experts’ opinion rather outdated and should be substituted by new ones. The author reveals a routine to improve traditional way of e-learning by a usage of Learning Agents. 1. INTRODUCTION Distant learning is a process of teaching, which takes place between teacher and students usually separated in time and space. This kind of teaching is become even more popular, so wiling to vary this way of education educators started using all available tools possible to apply [6] specified by the modern e-technologies. Though idea of distant learning came into being about three century ago along with appearance of first mail courses, in informational and computerization age is being even more attractive form of teaching. Growing need of gaining larger and larger knowledge and reduction of time required for stationary methods, especially for the beginners, caused development and improvement in distant learning routines. This kind learning is often only one alternative form for people who are handicapped who for health reasons cannot afford for studies far away from home. It is also alternate proposition in case of the unemployed training. On the other hand development of this form of education gives universities opportunity to reach potential wider audience. Essentially exist only two categories of distant learning: synchronous and asynchronous [2]. Synchronous one requires simultaneous participation both students and tutors (educators). Main advantage of such kind of education is mutual interaction that takes place in real time. Exemplar forms of synchronous teaching are interactive television, teleconference, different kind of internet chat-rooms etc. Asynchronous teaching does not require students and instructor's simultaneous participation. It isn't required from students, to come together in specific place and time. The students have plenty of possibilities: choice of place, range of time, what they intend to learn and what materials are going to use according to the tasks that they ware entrusted to. Asynchronous education is more elastic than synchronous one. The asynchronous forms of teaching appear as electronic letters, audio-video broadcast transmissions, all sort of correspondence or Web pages. The biggest disadvantage of this form appears in possibility of accumulating significantly big amount of unfeasible tasks by the student [4]. Nowadays the utilization of Internet in Poland for distant teaching is on initial stage of development. The Internet, first of all, serves for asynchronous access to educational materials. This is the kind of access to information became substitute for traditional textbook. Dynamic development of new techniques in sending and processing data thru the net let us improve and develop more effective forms of teaching. Researches on this subject are carried out on whole world. One of the academic centers in Poland, which carries out this kind of experiments, is Wroclaw University of Technology that is one of the leaders in putting into practice multimedia and Internet in e-learning. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 205 2. TECHNOLOGICAL ASPECTS OF DISTANT LEARNING Research group of Institute of Engineering Cybernetics in Wroclaw University of Technology developed experimental course MultiMedia Education: “An Experiment in Delivering CBL Material” (nr. PL1046) [1]. That course exploits newest internet technologies. All discussed below (in point 2) techniques concern this course but could be applicable in general to the most e-learning solutions. Material of this course concerns microprocessor systems and it is accessible for users equipped only with computer along with internet browser and Internet access. All materials are delivered through net. Tests have proven that access by telephone line is sufficient to effective using all materials. During developing distant learning courses many conventional solutions ought to be replaced by the alternative ones and the main goal is to integrate various techniques of learning to make this process as much possible attractive to learners. Traditional lectures could be replaced by the previously recorded of video presentations. Every user then have possibility with assistance of internet browser to view and listen earlier recorded lectures (quick preview of content is great support) along with additional on screen notes that appear on screen. Traditional textbook should be replaced too. For this purpose multimedia presentation must be used. Taking advantage of animations and necessity of interaction from user side greatly increase effectiveness of this form of learning. Fig. 1. Main menu of Multimedia Education Course (based on: mm-edu.ict.pwr.wroc.pl) Another task is to create an internet form of laboratory activity. This is frequently the greatest challenge in all projects but if it is successful it results in the development of a virtual laboratory. Such kind of laboratory exists only virtually not physically. Devices accessible by students are entirely simulated by computers’ software. The purpose of such laboratory is to make available advanced learning contents in a network environment. Furthermore, distant access to programming environment is ought to be implemented, which gives the user the possibility to write programs for controlling any device. Certainly we cannot stop thinking about checking learning progress. Therefore we must include several sets of test questions to check student’s knowledge. Results of those tests should be immediately available to the teacher and the student. To test practical skills, additional sets of tasks for students that verify gained knowledge should be placed in the course content. Students are required to perform special tasks with virtual devices. To integrate all mentioned above subtends activities course management system must be used e.g. like commercial WebCT system. Everything needed is included in its environment, it provide decent set of administrative tools to assist the instructor in the process of management and continuous improvement of the course. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 30-31, 2004 206 2.1. Internet University First element that is crucial during construction a base for e-learning solutions is the virtual equivalent of university (understood as institution, facility) that have to meet three essential functions as: student’s management, dissemination of knowledge, and testing progress of its absorption. Such assignments are fulfilled by distant education supporting systems called Course Management Software. With assistance of CMS systems integrating many techniques is feasible. Such systems can make possible course material sharing, checking progress of knowledge absorption by each student, provide users’ authorisation policy, present large statistics and many more. Virtual Learning Environments [9] is an important part of all distance learning courses. In the many elearning solutions the WebCT core system is commonly used for the designed course. WebCT is a widely used course management system in higher education, enabling the delivery of online education around the world. The WebCT solution integrates the richest and most flexible pedagogical tools with existing campus infrastructure. Secure "virtual classroom" environment can be deployed enterprise-wide to supplement the traditional classroom or for pure distance programs. Use of the WebCT course management system allows institutions to efficiently leverage campus resources to both extend their offerings and enhance the teaching and learning experience. For example, one of the many benefits of using WebCT includes the ability to offer an "always on" environment, providing more time for the student to interact with professors and classmates as well as with the course material in an efficient, engaging and effective manner. Course management is possible thru the browser window. Worth to add is complex students’ progress indicator module. It usually integrates wide range of tests, quizzes along with open questions and simple content analysis. It automatically checks the answers along with result analysis for entire students group that easily allow check students knowledge. WebCT provides rich capabilities for tracking student learning data within a course. Among the numerous pieces of student performance data that instructors can analyse are portions of the online course that students are focusing their time on, frequency that students participate in online learning activities, student performance assessments. Fig. 2. Assessment canter in WebCT Core INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 30-31, 2004 207 These data provide a rich source of information for improving the learning experience, for individual students as well as for continually improving the, course experience for all students, current and future. Other useful features of the WebCT system are the questionnaire and quiz facilities. WebCT is one of many available distant learning systems. There are a many different solutions available: LearningSpace (Lotus Notes), TopClass (WBT Systems) or BlackBoard (BlackBoard Inc.). 2.2. Internet lectures One of the most important ways of training is a lecture. In the case of remote learning usually TV or video broadcast is used. When producing an internet lecture it is necessary to transfer audio and video recordings to computer using special video equipment (cameras, video cards). After initial processing, video material is compressed to lower bandwidth demand when transmitting it to the audience. Files with recorded audio-video content are stored on internet server and later transmitted on demand to the students’ computer as a data stream. The quality of audio and video in student’s computer highly depends on data bandwidth with server. One of producers who make software enabling video transmission is Real Networks Inc. Not only audio-video recordings could be used but there also exists a possibility to transmit to students a virtual view of lecturer’s whiteboard. Putting all together we used the following equipment called MiMio Whiteboard. MiMio is exceptionally interesting equivalent of conventional blackboard and it should be used in any distant learning course to improve comprehensibility of lectures. It transforms a standard whiteboard from a simple writing medium into a powerful collaboration and communication tool. It is easy to make any whiteboard digital by simply attaching MiMio. Fig. 3. Application of MiMio Whiteboard technology After that, everything written or drawn is captured electronically into the computer in colour and realtime - ready to be saved and shared with anyone, anywhere, anytime. MiMio improves productivity and collaboration by allowing meeting participants, team members and students to concentrate on sharing ideas, participating in discussions and solving problems instead of taking notes. But the most important thing in our situation is that the whole process of creating a picture on a whiteboard can be recorded. In connection with audio and video recording this gives us a possibility to make a presentation containing a real lecture. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 30-31, 2004 208 Fig. 4. MiMio tool 2.3. Internet textbook One of most popular forms of learning is simply a book. In computer aided learning the ideal substitution for a book is a so called multimedia presentation that integrates many technologies in one dynamic presentation. One can say that this is even more than just a book. Possibility the intercalation with users has similar effect as practice classes. The most popular system for creating guided dynamic presentations is Macromedia Authorware. Authorware is an integration program. It is designed for putting multi-media together in one structured, interactive program. It is not only one, but basically the best tool for creating images, sounds or animations. The metaphor used is that of a procedural program's execution flow line. The developer adds graphics, animation or sounds to the flow line in the order he wants them to appear. Media components can be created in Photoshop, Director, SoundEdit, FreeHand, Authorware itself, etc, and can stay up on the screen for as long the developer wants by setting the components' properties. Media components are added to the flow line using Icons, hence Authorware's Icon-Based authoring scheme. There are different kinds of icons that represent different media components. You drag an icon from the Icon Palette to the main flow line and double-click on it to add desired media. The icons that distinguish Authorware from other authoring tools are: Framework, Interaction and Decision Icons. The Framework Icon allows you to navigate through parts of your program, the Interaction Icon allows you to interact with the user, and the Decision Icon allows you to take specific paths according to criteria you set. It is important to note that other programs will allow you to create the same functionality, but they will inevitably require heavy scripting, while Authorware Icons have their functionality self-contained. 2.4. Laboratory exercises One of the most indispensable forms of classes is a laboratory exercise – working with genuine objects. It allows the students to attain practical skills and to combine theoretic knowledge with practical proficiency. When we deal with distance learning on a daily basis there is no opportunity to contact with real devices. Therefore, practical classes are the real challenge. Generally, there are two approaches to this problem and each of them has its own advantages and disadvantages: real laboratory with distant access and virtual laboratory. The plan of remote access to real devices in a laboratory is mainly rejected because of following shortcomings limited number of simultaneously working students (depending on the number of devices available in classroom) and the entire cost of the system - devices, video cameras (to view the actual state of the process) and any other additional devices that make communication possible. Therefore more often the Virtual Laboratory is used. This kind of facility actually does not exist. Genuine devices, on which students could operate, are being replaced by the programs which can simulate activities of those devices usually in real time. This kind of laboratory can be accessed thru user computer using Internet connection and is limited only by amount of resources on the server and the network bandwidth, neither by the number of real devices nor the time of the day. There is no additional requirements needed to start exercising, neither additional devices such as video cameras nor any other special interfaces are required; therefore no supplementary human service is INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 30-31, 2004 209 needed in case of a device malfunction. The simulation software usually runs on the client side or rather seldom on the server only when simulator consumes extraordinary large number of resources. Sometimes there is an option of running one part of the simulator (responsible merely for visualisation) on the user’s machine and the second one on the server (where server simulates behaviour of device). The choice depends on many factors such as quality of the network connection, simulation algorithm complexity, difficulty of installation on the user’s machine, system maintenance (new versions and patches distribution) and widely interpreted computer system security. The aim of the each virtual laboratory is to support learning of as much types of devices as possible along with possibility of programming microprocessor based devices such as PCs. The main intention of the system is a remote access to those devices through the Internet. The user is expected to have internet browser with any runtime environment such as Java and at least dial-up connection. Naturally all virtual peripheral devices could be are programmed in previously fixed programming language e.g. C/C++ or Java. The students can see the controlled device thru the window in the Internet browser. Usually, in the same window there is a module presenting the status of the device components, which can be accessed from the program. There is also a possibility to manipulate any moving part of device by sending to it suitable sequence of commands. Beside the ability to view the computer-generated device, the user can operate in the programming and running environment. He can write programs that control devices, compile and run them. In order to prevent unauthorised access to the laboratory every admission is controlled. Each time student logon to the lab he must obey the rules and provide his login and password. Only after the successful login the access to all devices and tutorials is granted. To ensures portability most of virtual laboratories are written using the Java language to make possible operating on many hardware platforms. 3. DISADVANTAGES AND PROBLEMS OF DISTANT LEARNING Despite many of undeniable strong points distant learning with utilization Internet has its own weak points. It Students feel isolated without any real tutor, from whom they can seek advice, they often lose their eagerness to learn [2]. Distant learning process leads also to isolation of human being, by delimitation interpersonal contacts and finally to dehumanization of gaining knowledge. To the worst disadvantages we may include time-consuming and expensive process of preparing complicated teaching materials (e.g. course content, lesson, electronic lectures, quizzes, tests)[8], complexity of management or impossibility teaching some disciplines of science such as medicine. Not infrequently we can deal with phenomenon when well prepared distant learning course is characterized by worst efficiency than the regular one. The cause of this phenomenon is a human factor. It could be well noticed in case of students who are not proficient in operating computers [6]. Main reason is that students are incapable to select essential information from so-called “information noise” and lack of interpersonal contacts with other disciples. Apart from mentioned before problems elementary problem in distant learning lays in difficulty to match material that will suit student needs and will be proper for his background or his level of knowledge. Another huge difficulty constitute problem how to manage the students with minimal amount of workers. If number of students exceeds lets say hundred or more it is considerably more difficult to manage them or track their learning progress. With growing number of disciples the assessment, scheduling their classes or exams become highly inconvenient. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 30-31, 2004 210 During the work with students there is a need to draw attention to different times of knowledge absorption by individual students. It indirectly follows that each disciple has own knowledge perception and absorption mechanism. In classical form of teaching all presented procedures are executed by a “human factor” what could cause (and usually causes) crisis situations, such as lack of control and supervision over entire group during learning process (because of tutor’s tiredness), delivering inadequately prepared learning content (because of ignorance of students’ skill). All these actions have an effect on disciple’s eagerness factor. Remedy for formerly presented inconveniences connected with e-learning and the manner how improves efficiency of absorbing knowledge could be application of Learning Agents [11]. 4. AGENT AND DISTANT LEARNING 4.1. What is agent? Speaking in most general terms, agent is processes that operates in background and performs some activities due to its own agenda after any event occur. In Internet environment Agent is a system component that reside on client side - in user computer. It collects information about users’ state and sends data back to the server that should be perceived as repository of data and management system [5]. Agents’ management system could operate on main computer in the net, though in case of distributed management systems can be also located in many nodes of net, and some local harvesting data could be also possible that send periodically gathered data to the main node. Measured vector data Data processing Data Analysis Observatio n of process Image of activities user Fig 5. Illustration of agent activities In simplification „agent” is a computer program, which is capable to execute and supervise complex operations such like retrieving data, “customer maintenance”, etc. In academics spheres term agent is still being discussed. Scientists cannot present one coherent definition for the agent. That is why this term is widely used for many, sometimes, different products and technologies. Some called agents just “robots”, abbreviate by “bots” – because idea of the agents is to help people out in doing their common boring action. What is more, agents are capable to perform their actions, even when users are temporarily not connected to Internet – agents’ autonomy is one of their essential features. Multi-agent systems, as a rule, have their application almost everywhere – nonetheless net applications of agents are their fastest developing domain of science. The agents are commonly used for collecting user data and for realization very complex tasks thru activity analysis and adjusting to suit user needs. Such systems have a great influence on distant learning techniques that exploitation of Internet as a transport medium, where they are perceived as some kind distributed supervision over learners. Agents are capable significantly relieve human supervisor of their duties and help in a proper selection of the course content to be optimal for every individual. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 30-31, 2004 211 There is a need during work with students to draw attention on a various level of knowledge absorption by particular individuals and it involves with user intelligent profile calculation. One of the best known means, to determine person profile, its abilities, interests or weak points is application of multi-agent system, which is one of many techniques that support e-learning [10]. Minimal aim in this case is to create an agent that will navigate and select material within the one Web course page limits. This aim can be easily achieved after the ability recognition and classification of learners is attained. For example if system is able to differentiate at least two groups of students. First dilemma that we encounter is the problem how to estimate if given portion of material has been absorbed or not and next what course level will suitable for him as next step. These problems cannot be solved directly by the agent but with its help tutor can get very useful hints for each individual. Assigning a proper level, on a basis of estimation some previously established parameters, for each learner is a right task for an agent. Next step that should be undertaken is user interests analysis by the his prior activity analysis in order to presented learning material not to be bored for him but encourage to further learning. It can take place on the basis of complex synthesis of course movement path across course content on initial skill analysis. Each node in hierarchy could posses ascribed some values that correspond to agreed worth – units that describe user skill. Thorough analysis of navigation process along with units counting and their proper interpretation reveals such desirable user profile. In further stages constant analysis of user progress allow to automatic material selection and will relive supervisors duties in significant degree. Both learner’s behavior and his analysis made by agent could be used by the other agents – for example retrieving agents in order to find information that could be useful in the near future. Action of retrieving data is merely, unconsciously triggered by the learner and it happens without his knowledge in independent way. 4.2. Characterization of user profile Single agent or even groups on them could be scheduled to determine the intellectual profile of users. Agent that models user behavior, adapts interface and marks nodes in movement path in knowledge tree, it passes information about user’s material interest factor. All information are collected and passed on to classification agent. It’s the first step in ascribing one of possible advancement levels to a particular student. The analysis of collected data allows for general evaluation and assigning to one of three levels (basic, intermediate, advanced). Qualification to one of these levels requires earlier selection of material in a knowledge base. 4.3. Navigational hints Navigational agent, as the name indicates, is responsible for a proper “navigation” across course material placed on WWW page. Materials could have placed on many locations – so that the task for such agent is facilitate the access and collecting information about them depending on user profile obtained form agent that analyze user movement across course content. Thru material classification performed by mentioned above agent the most vital materials or information are presented using interface agent. Materials strongly connected with course topic are in queue and wait for user access. Every time, when student made any choice each movement is tracked and analyzed by the bunch of agents, which are keep trying to adapt course content to suit user needs. 4.4. Harvesting data Finding materials useful in a given course / lesson is not an easy matter. Not only for the sake for its appropriate selection, but their placement. It is extremely seldom situation when complete course is placed on www site along with thorough set of materials concerning the course subject. Usually additional materials are INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 30-31, 2004 212 placed on multiple web pages outside of course main page. Process of retrieving such data, for the average user, is usually too much time-consuming and not always data found fully match the course subject. Harvesting agent could turn out be very exceptionally useful in such situation – it helps retrieve data and additional information gathered by the previously mentioned agents (info about type of user and his preferences). Results of its work are strictly connected with performed (or not) tasks of other agents. And the efficiency of harvesting agent will be strictly dependent on them. 5. SUMMARY Distant education plays even more important role in process of education many social groups. Its importance is growing considerably year by year, among other things, thanks to spreading Internet accessibility. And its popularity, especially among students form large area countries (like Canada, USA, Australia), matches traditional one. Both in Europe and USA majority of well known universities started teaching using e-learning courses. For these reasons distant learning should not be consider as a margin phenomenon, and the interest of it should be aroused by academics. It should be emphasized that e-learning is more effective and more economical than conventional one, but only in one condition – adequate preparation of material along with putting great effort into its realization. Distant education posses many advantages but quite a lot of negatives too. Some of them could be removed effortlessly. But unfortunately to get rid the others disadvantages it ought to be introduced a brand new intelligent model of teaching employing so called smart agents. Multi-agent systems become more popular at present in e-learning solutions. In particular individual Agents are perceived as some kind supervisors of simultaneous distributed education process for many participants. They could be very helpful in preparing personal paths of learning for individual students and come in handy to estimate progress of learning for each student. Agent’s architecture used in e-learning solutions creates new possibilities of efficient and quick work in the distance learning area. Agent-specific techniques are mainly used for estimation of knowledge absorption, adjusting tasks to be suitable for an individual and optimization a whole process of gaining knowledge to be optimal for each student. Presented approach that incorporates Learning Agents into teaching process may considerably shorten time of gaining knowledge by increasing efficiency and allow for increasing number of students without employing more human supervisors or even additional control. So that we think that highly specialized distant learning systems with built-in agent’s techniques are more effective for delivering e-learning multimedia content and solving distance problems than classical one. Multi-agents systems, in near future, will most probably constitute the basis of the majority systems for distance learning education. REFERENCES [1] BARÁNSKI, M., WALKOWIAK, T., MM-EDU: network virtual laboratory. Raporty Inst. Cyb. Tech.PWr., Seria - SPR 2/2001, 2001. [2] BOCZUKOWA, B., „Edukacja na odległość” - Akademia Podlaska, Siedlce, 2000 [In Polish] [3] CRADDOCK, I., MENDDRELA, M., "Capture and Asynchronous Remote Delivery of Handwritten, Audio and Video Lecture Content"ALT2001 [4] Distance Learning Week: What is Distance Learning? Public Broadcasting Services, 1999; http://www.pbs.org/adultlearning/als/dlweek/whatis.htm INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 30-31, 2004 213 [5] [6] [7] [8] [9] [10] [11] [12] HECZKO, R., „Systemy Wieloagentowe” - http://ie.silesnet.cz/mas-pl.html [In Polish] KUBIAK, M., „Internet dla nauczycieli – nauczanie na odległość”, Mikom, Warszawa, 1997 [In Polish] KUBIAK, M., „Wirtualna edukacja po polsku”, Computerworld, 2000/26 [In Polish] Online Journal of Distance Learning Administration, 1998 – http://www.westga.edu Projekt MM-EDU http://mm-edu.ict.pwr.wroc.pl STEINER, V., "Technology in education", 1995 - http://www.wested.org/tie/dlrn/distance.html Ulises Cortes García,Ramon Sanguesa Solé, Javier Béjar,Tim Hall “Improving Learning Tools by Means of Cooperative Agents Technology” - Dept. Llenguatges i Sistemes Informŕtics, Universitat Politčcnica de Catalunya, INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 30-31, 2004 214 USING THE RUP FOR SMALL WEB INFORMATION SYSTEMS Petr Švec Brno University of Technology, Faculty of Mechanical Engineering, Institute of Automation and Computer Science Technická 2, 616 69 Brno, Czech Republic e–mail: [email protected] Abstract: This paper deals with describing and customizing the method for proceeding in the software development. The objective is to reduce quantity doing operations at the Rational Unified Process thereby to accelerate development of the small information systems with special demands for deploying on the Internet. Keywords: Rational Unified Process, Unified Modeling Language, Web Application Extension, customizing RUP. 1. INTRODUCTION It is necessary to apply the project management during development of large projects as well as of small ones. The Rational Unified Process (RUP) is a clearly defined and structured software engineering process framework. It is a software development approach that is iterative, architecture–centric, use–case–driven but mainly it is customizable for all needs of the users. The various configurations can be made to support small or large teams or more or less formal approaches to development. The paper is concerned with particular parts of the RUP that are presented in [4, 5]. The books [1, 6, 7] introduce use–case–driven approach and the Unified Modeling Language (UML) is explained in [2, 8, 9]. 2. THE STRUCTURE OF THE RUP The RUP consists of two dimensions: a dynamic aspect (horizontal) and a static aspect (vertical). The dynamic aspect expresses lifecycle, phases, iterations and milestones. The static aspect expresses activities, disciplines, artifacts, and roles. Figure 1: The structure of the RUP INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 215 The dynamic structure deals with a lifecycle or a time dimension of the project. The RUP uses an iterative approach that divides a project into four phases: Inception, Elaboration, Construction, and Transition. Each phase consists of one or more iterations. The iteration focuses on producing the technical matters necessary to achieve business objectives of that phase. At the end of each phase is a milestone that helps to specify whether we can continue with the next phase or not, thereby it decreases possible risks. The Inception phase establishes a good understanding of all customers’ needs, requirements and a scope of the system. Further it mitigates many of the business risks and produces the business analysis for building the system. The Elaboration phase takes care of many of the most technical tasks (design, implementation, tests, key components, etc.) and it tries to pinpoint major technical risks (performance, data security, etc.) by implementing and validating actual code. The Construction phase does most of the implementation over whole project. Finally, the transition phase builds the final version of the product and delivers it to the customer. The static structure defines how individual disciplines are ordered in each iteration of the concrete phase. The disciplines are logical containers for all process elements – roles, activities, artifacts and associated concepts, guidelines and templates. The disciplines can be divided into two main parts – fundamental and supporting. The fundamental part includes business modeling, requirements, analysis and design, implementation, test and deployment. The supporting part includes configuration and change management, project management and environment. The individual processes explain who does what, how and when. The role is an individual or a group that has a competence and a responsibility for certain part of the system. The artifact is a piece of information that can be produced, modified or used by another processes. The activity is a unit of work that is assigned to specific role. Workflows are used to describe meaningful sequences of activities that produce some valuable result and to show interactions between roles. The most common workflows are the disciplines. 3. PRINCIPLES OF THE RUP APPROACH The essential principles, so–called best practices, are iterative development, requirements management, architecture and use of components, modeling and UML, quality of process and product, and configuration and change management. A major goal with the iterative development is to reduce risk early on project. This is done by analyzing and reducing top risks in each iteration. As shown at Figure 2, a few requirements, analysis, design, implementation, and testing is done in each iteration. Each iteration finishes with an executable application. Figure 2: The structure of iteration Tools called “use cases” are used to capture customer’s requirements in the requirements management. They are used during following analysis, design, implementation, and test disciplines. The component–based development encapsulates data and the functionality operating upon that data into a component. Every component has an interface in order to allow the other applications or part of system to connect to. This feature of component–based development is called encapsulation, which makes components easy to reuse. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 216 In the RUP maintaining and developing models describe individual parts of the system. To construct these models the UML is used. The UML is a graphical language for visualizing, specifying, constructing, and documenting the concrete parts of the system. Besides the UML we have to use the Web Application Extension (WAE) for representing Web pages and other architecturally significant elements in the model. Figure 3: System modeling from different views A model is simplification of reality that completely describes a system from different views (Figure 3). The models are used for better understanding the system we are modeling. They are important because they help the development team and the customer to communicate. As mentioned above, the software is rebuilt in each iteration. As soon as the software is built, testing is executed. Regression testing ensures that new defects are not included in new iterations. Finally, the change management is a systematic approach to managing changes in the requirements. 4. CUSTOMIZING THE RUP FOR SMALL WEB INFORMATION SYSTEM In our example we introduce a small web information system for outsourcing organization. The outsourcing service must be founded upon regular communication (feed–back) between the customer and the provider. Our system includes secure access for users, administration of customer’s requirements, administration transactions, generating reports etc. We adapt all of the processes (disciplines) in the RUP to our needs. The RUP is very customizable process. We have two dimensions that are shown in Figure 4. The horizontal axis expresses the documentation quantity and the vertical axis expresses a development approach. In our example, creation of a small web information system, we follow slender documentation approach and the iterative development. Our first discipline, which is used in the initial of each iteration, is the business modeling. The business modeling is used to understand the structure and the dynamics of the organization, current problems in the target organization and to identify improvement potentials. The business modeling is a common language for understanding the customer and the provider each other. The business model consists of business use cases. The use cases describe the services provided with our business to our customers. The next advantage of the business modeling for our system is the possibility to map the business models with object models into the subsequent models. For our outsourcing information system, the business modeling is valuable for understanding how this new system affects the way we conduct our business. In the case of building the small project we focus on only the parts of the business that are the most unclear or crucial and the information is captured less formally. There are six business modeling scenarios – organization chart, domain modeling, one business – many systems, generic business system, new business, and revamp. The organization chart scenario for our purposes is suitable. In this scenario we build a simple chart of the organization and its processes so that we get a good understanding of the requirements of the information system we are building. The business modeling is a part of the software– engineering project, primarily performed during the inception phase. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 217 Figure 4: The possibilities customizing the RUP for projects The business object model mentioned above captures the responsibilities, organizational units, and items within organization, and how they are related to each other. The business object model gives insight into the organization. In our example project we have simplified business use case model capturing different requirements (orders) and confirmed protocols for made services. Further we have a couple of business class diagram and activity diagram that implement the object model. The first class diagram represents individual relations between entities and employees (their responsibilities) and the second represents associations between workers. The unique activity diagram describes a flow customer’s requirement until the filling and confirming of the protocol is done. In the requirements discipline arise the most important models of the system, the use case models. They are created from the vision document, supplementary specification and partly from the business model. The use case models are used to establish and maintain agreement between the customers and the providers. The agreement on what should be done and why in order to provide system developers better understanding of the system requirements, to define the boundaries of the system, to provide a basis for planning the technical contents of iterations, to provide a basis for estimating cost and time to develop the system and to define a user interface for the system. In our example, we collect the customer’s requests and develop a brief vision document containing a set of key customers’ needs and the high–level features of the system. Further we compile a risk list enumerating hazards that could arise. The most important requirements decided to include in the vision document are based on analysis of the cost of implementing desired features. Then we translate these features into the detailed software requirements for designing, building a system and for identifying test cases to repeated testing of the system behavior. These detailed and the most essential requirements are captured in the use–case model and other supplementary specifications. While we are detailing these requirements, we will certainly find flaws that must be repaired by getting clarifications from our customer. The same person can take the role of analyst and developer, which means that the person who describes a use case will also implement it, so the second eventuality capturing requirements is that we spend less time on documenting the detailed requirements and come back to them as we implement the use case. Then we design the user interface by using a prototyping tool. The main inputs to the design of the user interface are the vision document, the use–case model and supplementary specifications. Finally, we must build a glossary that defines a common terminology that is used across the project. Further there are analysis and design discipline whose purpose is to translate the requirements into a specification that describes how to implement the system – that is, to a set of classes and subsystems. This transformation is driven by the use cases so that it ignores many of the nonfunctional requirements and the constraints of the implementation environment. The design of the system is captured in a design model. The design model is the primary artifact of the analysis consisting of co–operations of classes providing the system INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 218 behavior. For a transparence in our project we can have classes aggregated into packages and subsystems, e.g. it consists of subsystems for administering users, administering requirements and administering protocols. We elaborate the design of the system to the point when it can be implemented through a direct and systematic transformation of the design into code. We omit an analysis model that is an abstraction, or generalization, of the design to represent only the most important details of how system works. If we used the analysis model, we would have to ensure that the analysis and design models remain consistent. From that reasons the analysis model is superfluous for our small project. The set of operations performed by classes, subsystems and components is called an interface. We use interface because it improves the flexibility of designs by reducing dependencies between parts of the system and therefore making them easier to change. Interfaces are used by components that are dependent only on the interfaces of other components. The purpose of the implementation discipline is to implement classes and objects in terms of components and to test the developed components. The tests are limited to test units of individual components. During the iterative software development are created operational versions of a system or parts of the system which represent a subset of the capabilities to be provided in the final product, so–called builds. They represent ongoing attempts to demonstrate functionality. After we develop some subsystem of the project, we will integrate it into a complete system. The RUP approach is the incrementally software integration. For example, while we are gradually developing some subsystems, e.g. administration customers or requirements, we are integrating it into the whole system. The benefit of this approach is the testing individual components properly in order to locate unsafe faults. To reduce uncertainty affecting the stability, performance, project commitment, funding, understanding of requirements, usability, and ultimately the look and feel of the product, we use prototypes. In the RUP we can optionally use two kinds of prototypes – a behavioral prototype and an evolutionary prototype. The behavioral prototype focuses on exploring specific behavior of the system whereas the evolutionary prototype evolves to become the final system. The evolutionary prototype is used in our small project. The key artifacts of an implementation are an implementation subsystem, component, and integration build plan. The implementation subsystem is a collection of components and other implementation subsystems which are used to structure the implementation model by dividing into the smaller parts. The component is a piece of software code (e.g. source or binary). The integration build document defines the sequence in which the components and subsystems should be implemented. In our small project we create a component diagram. The component diagram contains components which are connected to each other. The purpose of the test discipline is to assess product quality and to provide feedback. The test discipline is being done throughout the lifecycle. It is used primarily for verifying that all requirements have been implemented correctly without defects. The simplified test model contains test procedures, test cases and notes. The test cases are a set of execution conditions and expected test results developed for a specific test. The test cases can be derived from use cases. The test procedure is the set of the detailed instruction for the setup, execution, and evaluation of test result for test cases. Finally, the notes are textual information describing constraints or additional information. We can also use the prototype as a tool for testing. The deployment discipline guarantees that the finished software is delivered to the customer. It contains testing the software in its final operational environment above all, distributing software and installing. There are several kinds of deployment processes – deployment of software in custom–built systems, deployment shrink– wrapped software and deployment of software that is downloadable over the Internet. Since we are developing the information system we use deployment of software in custom–built systems. The key artifact of the deployment discipline is a deployment diagram. The deployment diagram depicts the implementation and environment of a system. 5. CONCLUSION In this paper was outlined a manner of the customization of the RUP for small web information systems. It describes the most significant steps which are used in this method omitting the less important parts, using less formal approach which is not burdened by enormous quantity of documentation. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 219 REFERENCES [1] BITTNER, K., SPENCE, I.: Use Case Modeling. 1st ed., Addison–Wesley, 2002. [2] BOGGS, W., BOGGS, M.: UML with Rational Rose 2002. 1st ed., SYBEX Inc., 2002. [3] CONALLEN, J.: Building Web Applications with UML. 1st ed., Addison–Wesley, 2002. [4] KROLL, P., KRUCHTEN, P.: The Rational Unified Process Made Easy: A Practitioner’s Guide to the RUP. 1st ed., Addison–Wesley, 2003. [5] KRUCHTEN, P.: The Rational Unified Process: An Introduction. 2nd ed., Addison–Wesley, 2000. [6] KULAK, D., GUINEY, E.: Use Cases: Requirements in Context. 2nd ed., Addison–Wesley, 2003. [7] LEFFINGWELL, D., WIDRIG, D.: Managing Software Requirements: A Use Case Approach. 2nd ed., Addison–Wesley, 2003. [8] MELLOR, S. J., BALCER, M. J.: Executable UML: A Foundation for Model–Driven Architecture. 1st ed., Addison–Wesley, 2002. [9] SINAN, S. A.: Learning UML. 1st ed., O’Reilly & Associates, Inc., 2003. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 220 SHAPE OPTIMISATION OF THE FERROMAGNETIC CORE OF FLUXSET MAGNETOMETER Gábor Vértesy1, J. Pávó2 1 Research Institute for Technical Physics and Materials Science Hungarian Academy of Sciences H-1121 Budapest, Konkoly Thege út 29-33, Hungary phone: +361-3922677, fax: +361-3922226, e-mail: [email protected] 2 Department of Broadband Infocommunications and Electromagnetic Theory Budapest University of Technology and Economics H-1521 Budapest, Egry József u. 18., Hungary phone: +361-4632913, fax +361-4633189, e-mail: [email protected] Abstract: Optimisation of the sensing core’s shape of a high sensitivity magnetic field sensor (Fluxset sensor) is presented. An optimisation method is discussed that can be used for the enhancement of the sensitivity of the Fluxset sensor by varying the shape of the ferromagnetic core in the sensor. The task is solved by electromagnetic modelling of the ferromagnetic conductor materials in a non-conducting medium. The presence of the ferromagnetic object is represented by volumetric and surface magnetic currents flowing in the homogeneous medium. The actual distribution of these magnetic currents are determined by the solution of the integral equation derived in the paper. Numerical examples have been presented for the demonstration of the results obtained for the modelling of ferromagnetic films. Keywords: Magnetic sensors, Numerical electromagnetic field modelling INTRODUCTION Magnetic field sensors play an important and continuously increasing role in many fields of science and of modern technique. Fluxgate type sensors are solid-state devices for measuring the absolute strength of a surrounding magnetic field or the difference in the field strength between two different points within a magnetic field. Their measuring range and their resolution are just within the gap between inexpensive sensors such as the magnetoresistive or Hall type sensors and very expensive magnetometers based on quantum effects such as SQIUDs and others. A new type of magnetic field sensor (Fluxset), which belongs to the family of fluxgate sensors has been developed recently [1] for measuring DC and AC (up to 100 kHz frequency) low level magnetic fields (less, than 1 nT) with high accuracy. Its principle of operation is close to the so called pulse-position type fluxgate magnetometers [2]. The particular advantage of these magnetometers is an output signal that can be simply converted into binary signal. The measurement of a small magnetic field is reduced to a high accuracy time measurement through the displacement of the magnetization curve produced by the field. The Fluxset sensor has excellent sensor properties because of its principle of operation. These are the good linearity, good stability, calibration accuracy, high signal to noise ratio, high sensitivity of magnetic field measurement, ideal spatial resolution and wide operation temperature range. The probes are suitable for axial measurement of the magnetic field. The transverse sensitivity is negligible. The sensor has a small size, it is versatile, inexpensive and sufficiently robust to meet the demands of the industry and requires no microelectronic technology. It can be easily integrated due to the simple construction of both the probe and electronics. One of the most important factors, which determines the sensitivity of the device is the sensing element. The optimisation of the shape of this ferromagnetic core might help to improve the performance of the sensor. In order to find the optimal shape of the core, exact modelling of the ferromagnetic field, i.e. rigorous solution of INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 221 the corresponding direct problem is reguired. In this paper we present a method, which can be used for the shape optimisation of the ferromagnetic core of the Fluxset sensor. We use a recently developed method for the modelling of the electromagnetic field in the presence of a ferromagnetic conductor [3]. The method is based on the solution of an integral equation. Since the size of the ferromagnetic core in the Fluxset sensor is small, the developed method is extremely efficient because only the discretization of the volume occupied by the ferromagnetic object is required. The ferromagnetic material is assumed to be linear. The presence of the ferromagnetic object is represented by volumetric and surface magnetic currents flowing in the homogeneous medium. The actual distribution of these magnetic currents are determined by the solution of the integral equation derived in this paper. The main advantage of the presented model is that the numerical solution of the integral equation requires only the discretisation of the volume occupied by the ferromagnetic object. FLUXSET SENSOR The schematic drawing of a Fluxset sensor is shown in Fig. 1. The sensor is made of two solenoids wound on each other. The inner and outer solenoids are called driving and pick-up solenoids, respectively. The sensing element of the probe is an amorphous alloy ribbon with high initial permeability and low saturation. Manufacturing techniques and properties of amorphous magnets are well known for a long time and a number of monographs devoted to these aspects were published (see, e.g. [4-6]). The idealized magnetization curve of the core material is shown in Fig. 2. The actual BH curve and other parameters of the material strongly depend on the different mechanical, heat, chemical, etc. treatments of the raw material carried out to improve the magnetic properties of the core [7]. The approximate length and diameter of the different realizations of the sensor are between 5 and 15 mm and around 2 mm, respectively. The thickness of the ribbon core is about 25 mm. Pick-up coil A vs(t) Driving coil A id(t) Glass sensor support Ferromagnetic core Fig. 1: Schematic drawing of a Fluxset sensor The core material is periodically magnetized to saturation in both directions by the triangular current excitation, id(t), of the driving coil. Considering the idealized magnetization curve of the core material shown in Fig. 2, we can see that the induced voltage impulse in the pick-up solenoid, vs(t), is almost zero when the core is saturated, on the other hand this induced voltage has a relatively large value (proportional to the time derivative of the driving current) while the driving magnetic field changes its direction (i.e. the core material is magnetized in the linear range). If an external magnetic field (i.e. the field to be measured) is superimposed on the periodic driving magnetic field, the time spent in saturation in one direction (e.g. when B=Bs) is longer than the time spent in saturation in the other direction (e.g. when B=-Bs), consequently the impulse, vs(t), is shifted. The time shift of the impulse, vs(t), can be accurately measured, as a result the external field can be predicted. One can see that the measured magnetic field is actually the field of the sensor core. Since the volume of the sensor core is very small, we assume that it does not significantly disturb the magnetic field to be measured, consequently the sensor measures approximately the magnetic field that can be detected around its core if we neglect the disturbing effect of the presence of the sensor core. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 222 Fig. 2: Idealized BH curve of the core material The accuracy of the measurement is basically depends on the shape of the pulse, vs(t), detected on the sensor coil. Since the pulse shape is formed by the fast change of the magnetic flux density inside the core when the exciting field changes its sign, the time dependence of the pulse can be calculated by assuming a linear ferromagnetic material that permeability is obtained as the gradient of the BH curve. Of course, the results of such calculations are useful only if the core material is not saturated. It is easy to see that, the sharper the detected voltage signal, the easier to detect small time differences, consequently the magnetic field is measured more accurately. Based on this consideration we may conclude that the more uniform the magnetic field in the core (i.e. the different regions of the core are saturated in the same time instant), the higher accuracy can be achieved in the measurement. The requirement of the uniformity of the magnetic field in the core might be fulfilled by changing the shape of the sensor core. Since our goal is to find a configuration where the whole volume of the sensor core is saturated in the same time, the assumption of linear ferromagnetic material is acceptable. CALCULATION OF THE ELECTROMAGNETIC FERROMAGNETIC CONDUCTOR FIELD WITH THE PRESENCE OF Integral equation describing the ferromagnetic material Assume that a linear ferromagnetic conductor occupies the volume region V1 in the otherwise homogeneous non-conducting medium. The schematic drawing of the studied geometrical configuration is shown in Fig. 3. Fig. 3: Ferromagnetic conductor in a homogeneous non-conducting medium. The ferromagnetic object is surrounded by air and its permeability and conductivity are s1 and m1, respectively. Assuming that the electromagnetic field is varying in time as the real part of exp( jwt ) , the Maxwell equations are written in the following form: (1) Ñ ´ H (r ) = [s (r ) + jwe 0 ]E (r ) , Ñ ´ E (r ) = - jwm (r )H (r ) , (2) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 223 where, r ÎV1 , ìs s (r ) = í 1 î 0 otherwise (3) r Î V1 . ìm m (r ) = í 1 otherwise m î 0 (4) Taking the curl of (1) and substituting (2) into this equation we obtain, [ ] Ñ ´ Ñ ´ H (r ) - w 2 m 0e 0 H (r ) = w 2e 0 (m (r ) - m 0 ) - jwm (r )s (r ) H (r ) + Ñs (r ) ´ E (r ) . (5) Assuming (1), E(r) can be written as, E (r ) = 1 Ñ ´ H (r ) . s (r ) + jwe 0 (6) Using (6) E(r) can be eliminated in (5). As a result we obtain, [ ] Ñ ´ Ñ ´ H (r ) - w 2 m0e 0 H (r ) = w 2e 0 (m (r ) - m 0 ) - jwm (r )s (r ) H (r ) + + d (r - r0 ) s1 Ñ ´ H (r ) ´ nˆ s (r ) + jwe 0 , (7) where r0 is the co-ordinate vector pointing to the surface S1 of the volume V1, n is the outward normal unit vector of surface S1 (see Fig. 3) and d(r) is Dirac's delta function. Note that the first term on the right hand side of (7) is zero outside of V1 and the second term is not zero only on the surface S1. Based on (7), the difference between the electromagnetic field that can be calculated with and without the presence of the ferromagnetic object may be assumed as the field generated by a secondary source concentrated in the volume V1 inside a homogenous non-conducting medium (in the present case in air). This secondary source might be a volumetric magnetic current flowing in V1 and a magnetic surface current concentrated on the surface S1 (see Fig. 3). In order to find the actual values of these secondary sources, the Maxwell equations in homogeneous nonconducting medium with the assumption of the magnetic current, M(r), and magnetic surface current, Ms(r), are written as, Ñ ´ H (r ) = jwe 0 E (r ) , Ñ ´ E (r ) = - jwm 0 H (r ) + M (r ) + d (r - r0 )M s (r ) . (8) (9) Note that M(r)=0 outside of V1 and Ms(r)¹0 only on S1. Taking the curl of (8) and substituting (9) into this we get: Ñ ´ Ñ ´ H (r ) - w 2 m0e 0 H (r ) = jwe 0 M (r ) + d (r - r0 ) jwe 0 M s (r ) . (10) The comparison of (7) and (10) results that the effect of the presence of the ferromagnetic conductor can be modelled by placing the following magnetic current and magnetic surface current into the homogeneous non-conducting medium: ìw 2e 0 (m1 - m 0 ) - jws 1m1 , r ÎV1 , ï M (r ) = n (r )H (r );n (r ) = í jwe 0 ïî 0, otherwise s1 ì , r Î S1 . ï M s (r ) = J (r )Ñ ´ M (r ) ´ nˆ;J (r ) = í w 2e 0 (m1 - m 0 ) - jws 1m1 (s 1 + jwe 0 ) ïî 0, otherwise [ ] (11) (12) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 224 The magnetic field may be written as the sum of an external field, H0(r), and the field generated by the magnetic currents, Hf(r), (13) H (r ) = H 0 (r ) + H f (r ) . Here H0 represents the field induced by sources outside of the examined homogeneous medium (consequently it satisfies the homogeneous wave equation), while Hf represents the magnetic field due to the presence of the ferromagnetic conducting object, i.e. due to the secondary sources in (7) or (10). Hf can be expressed with the help of dyadic Green's functions as [8], (14) H f (r ) = jwe 0 ò G (r | r ') × M (r ')dv' + jwe 0 ò G (r | r ') × M s (r ')ds' , V1 S1 where G(r|r') is the dyadic Green's function that transforms the magnetic excitation into the magnetic field. By substituting (14) into (13) and multiplying the equation with n(r) (see (11)) the following integral equation is obtained: M (r ) = M 0 (r ) + jwe 0n (r )ò G (r | r ') × M (r ')dv' + jwe 0n (r ) ò G (r | r ') × [J (r ')Ñ ´ M (r ') ´ nˆ ]ds' V1 (15) S1 where (16) M 0 (r ) = n (r )H 0 (r ) , is assumed to be known (in most practical problems M0(r) represents the excitation). By solving (15), M(r) can be calculated, consequently the electromagnetic field of the considered configuration is obtained. Note that only the volume region V1 is to be discretised for the numerical calculation of the field. NUMERICAL SOLUTION OF THE INTEGRAL EQUATION In Descartes co-ordinate system the discrete approximation of (15) is found by subdividing the region of space occupied by the ferromagnetic object, V1, into a regular lattice of N cells, each cell being a rectangular parallelepiped. The distance between the nearby points in the x, y and z co-ordinate directions are Dx, Dy and Dz, respectively. The magnetic current is approximated as, N N n =1 n =1 ( ) M (r ) » å M n an (r ) = å M xn xˆ + M yn yˆ + M zn zˆ an (r );r ÎV1 , (17) where r=x x̂ +y ŷ +z ẑ ( x̂ , ŷ and ẑ are the unit vectors of the co-ordinate directions). The approximating function, an, is chosen to be linear, i.e., æ x - xn y - yn z - z n ö an ( x, y, z ) = f çç , , ÷;n = 1,2, K , N ; Dy Dz ÷ø è Dx . ì1- | x | - | y | - | z | + | xy | + | xz | + | yz | - | xyz |, - 1 < x, y, z < 1 f ( x, y, z ) = í 0, otherwise î After testing the discrete approximation of (15) by the testing function, (18) (19) t m ( x, y , z ) = am ( x, y , z );m = 1, 2, K, N , n a system of 3N linear equations is obtained. Solving this equation system, the unknown coefficients Mi (i=x,y,z; n=1,2,...,N) can be calculated. The most serious difficulties in the solution of the integral equation arise when the coefficient matrix is to be calculated. The coefficients are obtained as the linear combination of the following type of integrals, é ù m, n = 1,2,K, N ò êê ò G(r | r ')× a (r ')xˆdv'úú × t (r )zˆdv; xˆ,zˆ = xˆ, yˆ , zˆ . n m (20) ë û Using similar considerations that was published in [9], the spatial Fourier transform of the integral in the square brackets can be evaluated analytically. This analytical formula is particularly useful because the kernel of the integral (i.e. the dyadic Green's function) is singular. This singularity might cause serious problems in the numerical evaluation of the integral. Having the mentioned analytical formula, fast and numerically stabile solution of the integral equation is obtained. V1 V1 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 225 MODELLING OF FERROMAGNETIC FILMS; NUMERICAL EXAMPLES For modelling of the ferromagnetic core of Fluxset sensor, the above introduced integral equations should be applied for a high permeability ferromagnetic film in homogeneous magnetic field. In this section we deal with linear ferromagnetic object that thickness is very small compared to its other dimensions. The permeability of the material is assumed to be high, on the other hand the electric conductivity of the film is sought not to be considerably high. Based on these assumptions, the ferromagnetic film is modelled as an infinitesimally thin film, i.e. a mathematical surface where the tangential component of the magnetic field is zero. The prescribed boundary conditions are fulfilled by placing tangential magnetic surface current, Ms(r) on the surface representing the ferromagnetic film. The eddy currents flowing inside the film are neglected in this model. As an example, assume that a rectangular ferromagnetic film is placed in the xy plane. The sizes of the film in the x and y directions are 5 mm and 1 mm, respectively. The external magnetic field is x directed, its time variation is assumed to be sinusoidal with the frequency of 100 kHz, the magnitude of the magnetic field is 1 A/m. The x component of the magnetic surface current representing the ferromagnetic material in the case of the given external field is plotted in Fig. 4. The y component of the magnetic surface current (not shown here) is considerably smaller than its x component. Note that the magnetic current has p/2 phase shift with respect to the external magnetic field. In Fig. 5, the z component of the magnetic field due to the presence of the ferromagnetic film is plotted above the surface representing the film (note that this field is in phase with the external field). Knowing the x component of the magnetic surface current and using (9), the induced voltage can be easily calculated in the solenoid around the ferromagnetic core. For this calculation we assume that the voltage induced by the magnetic field around the core is negligible compared to the induction due to the magnetic currents. 0.035 0.03 0.025 0.02 0.015 0.01 0.005 0 1 0 1 2 x [mm] 0.5 3 4 5 y [mm] 0 Fig. 4: x component of the magnetic surface current, Msx [V/m]. 5 0 -5 -3 -2 -1 0 x [mm] 1 2 3 -1 0.5 0 y [mm] -0.5 Fig. 5: z component of the magnetic field above the ferromagnetic film, Hz [A/m]. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 226 OPTIMISATION OF THE SHAPE OF THE CORE Based the outlined method for the calculation of the magnetisation of the sensor core, the optimal core shape is found by simulated annealing optimisation procedure. The objective, expressing the requirement of smoothness of the magnetisation of the core material, is the minimisation of the expression, 1 N N åH n =1 t - Htn (21) where Ht denotes the average of the absolute value of the tangential magnetic field on the surface, S, and Htn is the absolute value of the tangential component of the magnetic field at the n-th point. Several constrains are posed on the optimisation. One of the constrains comes from the symmetry of the sensor arrangement, consequently the core is designed to be also symmetric. By the outer dimension of the sensor, the maxium size of the core is defined, it cannot be bigger than 5mmx0.8mm rectangle. To obtain a sensor signal with sufficient amplitude, the minimal size of the core is also fixed. Due to the symmetry and the special elongated shape of the sensor, the shape of the core is described with a small number of parameters, accordingly, the number of degrees of freedom of the optimisation is 4. The particular choice of the simulated annealing optimisation procedure [10] is supported by the facts that it is very easy to apply this algorithm and it works almost for sure if it is properly implemented. On the other hand, the efficiency of this method is not particularly good because of the high number of function calls required to find the optimum. However, the design of a Fluxset sensor is not a kind of problem that has to be solved very frequently, that is why the robustness of the optimisation is the primary concern and the efficiency is not crucially important. As a demonstration of the result of the optimisation, in Fig. 6 the tangential magnetic field component parallel to the axis of the sensor in the optimum shape sensor core is plotted. As the final result, ellipsoidal shape of the sensor core was found as optimal one. Note that the tangential magnetic field perpendicular to the axis of the core is negligible compared to the one that is plotted in Fig. 6. In the presented plot the function values at the points outside the core surface are set to zero, these values have no meaning. Fig. 6: x component of the tangential magnetic field for the optimum shape core (in phase with the external magnetic field) CONCLUSIONS An optimisation method targeting the enhancement of the sensitivity of the Fluxset sensor by varying the shape of the ferromagnetic core of the sensor has been discussed. As part of the optimisation method, the solution of the related direct problem, i.e. the calculation of the magnetisation of a ferromagnetic conductor thin film due to a given external field, has been outlined. An integral equation has been derived for the modelling of the electromagnetic field of a ferromagnetic conductor in a homogeneous non-conducting medium. The presence of the ferromagnetic object is represented by secondary sources, namely by volumetric and surface magnetic currents concentrated in the region occupied by the ferromagnetic object. Numerical solution of the derived integral equations have been outlined by using analytical expressions for the spatial Fourier-transform of the singular kernel of the integral equations. Numerical examples have been presented for the demonstration of the results obtained for the demonstration of the result of the optimisation. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 227 ACKNOWLEDGMENTS This work was supported by the Hungarian Scientific Research Fund (T035264). REFERENCES [1] VÉRTESY, G., GASPARICS, A., SZÖLLŐSY, J.: High sensitivity magnetic field sensor, Sensors and Actuators A, 85 (2000) 202-208 [2] RIPKA, P.: Review of fluxgate sensors, Sensors and Actuators A, 33 (1992) 129 [3] PÁVÓ, J., SEBESTYÉN, I., VÉRTESY, G., DARÓCZI, Cs.S.: Electromagnetic field modelling of the ferromagnetic core of Fluxset type sensors, in Studies in Applied Electromagnetics and Mechanics 11, Applied Electromagnetics and Computational Technologies, eds.: H. Tsuboi and I. Sebestyén, IOS Press, Amsterdam, pp.172-181, 1997 [4] LUBORSKY, F.E.: Amorphous Metallic Alloys, Butterworths, London 1983 [5] MOORJANI, K., COEY, J.M.D.: Magnetic Glasses, Elsevier, Amsterdam 1984 [6] LACHOWICZ, H.K.: Magnetic materials – progress and challenges, Journal of Technical Physics, 42 ( 2001) 127 [7] VÉRTESY, G., GASPARICS, A., VÉRTESY, Z.: Improving the sensitivity of Fluxset magnetometer by processing of the sensor core, J. Magn. Magn. Mat., 196-197 (1999) 333 [8] TAI, C.T.: Dyadic Green's Functions in Electromagnetic Theory, Intext, Scranton, 1971 [9] PÁVÓ, J. and MIYA, K.: Reconstruction of crack shape by optimization using eddy current field measurement, IEEE Trans. Magn., 30 (1994) 3407 [10] PRESS, W.H. et al.: Numerical Recipes, Cambridge University Press, Cambridge, 1986 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 228 EVALUATION OF DEFECT OF PIPELINE ISOLATION DETECTED BY PEARSON METHOD Dalibor Bartoněk1), Imrich Rukovanský2) 1) Institute of Geodesy, Faculty of Civil Engineering, Brno University of Technology. Veveří 95, 662 37 Brno Czech Republic. Tel. 54114 7204, Fax: 54114 7218, E-mail: [email protected] 2) Evropean popytechnical institute, Kunovice, Osvobození 699, Czech Republic Tel. 572 548 035, Fax: 572 549 018, E-mail: [email protected] Abstract: This article deals with Pearson system programme assessing defects of pipeline isolation evaluation that were found when measured by Pearson method. It describes a new method how to apply this, nowadays already a classical way of detecting isolation defects including a special software which automatically processes the data taking into consideration the individual defects of pipeline isolation on given routes. The conclusion sums up practical experience complemented with a graphical output from the software. Keywords: Defect, pipeline isolation, Pearson method, locator, organizer, measuring evaluation. 1. INTRODUCTION The extremly wide steel pipeline network of gas pipeline and productline in the Czech Republic is usually placed in the soil exposed to an agressive medium of the soil which causes material corrosion. This negative phenomenon can be avoided if the pipeline is protected: 1. actively by the system of anticorrosive devices spread over territories, working on an electrochemical principle [1], 2. passively by isolation of pipeline which will not enable the access of H2O + O2 to the steel mantle and thus it will be prevented from corrosion. Even the best insulation has a limited durability therefore its state must be checked up not only when the pipeline is installed, but also in certain time intervals. There are a lot of check up methods; some of them can be found in ČSN 03 8375 (Czech standard). Except for some visual methods during which it is necessary to uncover the pipeline, the two remaining ways of check up of the quality of insulation can be divided into two groups: 1. 2. methods, finding out the individual insulation defects, methods, finding out the average quality of insulation in a given section of pipeline. Pearson method, we will speak about, belongs to the first group. 2. PRINCIPLE OF PEARSON METHOD When used this method the alternate voltage from tone generator is conveyed into the pipeline in a certain point (check up output or connecting object). The signal in a defected places of pipeline insulation starts spreading into the environment where it causes voltage drop down. In order to detect defects on the pipeline locator and signal receiver are necessery. There must be two persons wearing special shoes with metal tips ensuring galvanic contact with the earth, who search for these insulation defects. They move either one after INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 239 another or next to each other and receiver detects difference of voltage between points, of their positions – see fig. No 1. If both of them are in the points with undeffected insulation no voltage difference is detected. If one of them comes across the deffected points the receiver detects the change of signal level which causes both the increasement of deflection of measuring device-receiver and the acustic signal. The defects are located towards the fixed points on the pipeline (devices on the route). These points have their own positioning which is the distance in [km] from the beginning of the given route. The measurement requires experience, because the device can record the change of the signal even if there is not any defect e. g. the change of environment. Fig. No 1. Principe of Pearson method with undeffected insulation no voltage difference is detected. If one of them comes across the deffected points the receiver detects the change of signal level which causes both the increasement of deflection of measuring device-receiver and the acustic signal. The defects are located towards the fixed points on the pipeline (devices on the route). These points have their own positioning which is the distance in [km] from the beginning of the given route. The measurement requires experience, because the device can record the change of the signal even if there is not any defect e. g. the change of environment. 3. INNOVATION OF THE METHOD In the original variant, the Pearson method was done by at least three persons. The two of thems worked as explained in previous chapter, and another pearson had to write the measuring results together with positioning of measured places into a special form. The positioning values were found out by measuring of distances with a tape from the fixed points of the route in order the defects could be located. The measured data were manually rewritten into the computer, to be later evaluated. The greatest disadvantages of this variant were: · complicated manipulation with the tape, especially in the broken field. · a great probability of making mistakes both in manual recording of measured data and in the course of rewrtiting into computer. Therefore in the half of the 90’s a more effective variant of pipeline defects detection started to be pondered about. The innovation of this method consisted in implementing of a pipeline locator with the special A-shaped detection sond and connected with serial cable with the organiser PSION type – see fig. 2. sond locator switch organiser PSION tip for contact with the earth Fig. 2. Block scheme of device configuration for measuring by Pearson method Locator RD 432 PDL type is the product of English Radiodetection company producing also special measuring sond to order. PSION holder with switch is produced by Elgas Pardubice company (CZ). How to measure in the field: INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 240 Generator of the AC signal with frequency of 8 or 16 kHz is connected with the check point of the pipeline according to the special variant as follows. In this case the measuring is done by only two persons; one of thems identifies the route by locator, another one equipped with devices – see fig. 2 follows him. In one hand he holds a sond connected with locator and he places the tips into the surface above the pipeline, with the other hand he handles the locator with PSION organiser. Both PSION and sond are connected by serial cable with the locator. The sond receives the signal from the pipeline and sends it into the locator, where processed in the form of text message. This is a standard message, which goes through the serial interface into the organiser where, with the help of a special software, written into a file. There are two methods how to write measured values: · · manually processing of a given button on the locator, automatically in regular time intervals which can be set on the control panel of PSION. The positioning of the measured point is, in this case, set automatically on the basis of a number of steps walked by a workman with measuring equipment between two writing of measured values. The accuracy of positioning calculation requires the handling person to keep the same number of steps (the same length) during measuring. This disadvantage taxes the person havily (it requires a great experience) is ballanced because this unpleasant manipulation with the tape can be avoided. Further the measuring can be interrupted by mechanical switch at any time when necessary (being tired etc.) – see fig. 2. The software in the organiser adds a number of steps to the initial positioning during of every interval between values writing. The initial positioning, length and a number of steps can be set up before measuring. When the worker comes to the fixed point on the route he writes his precise positioning into the file either manually or semiautomatically by the choice of the menu list. This list must be transformed from PC into organiser long before measuring in the terrain and serves as a basis for calculation of approximate positioning in steps into precise positioning in km. This file must obligatorily contain at least two fixed points, otherwise the precise positioning can not be calculated. Every measuring record can be interactivelly complemented with a note. This is important at the moment, when the state of terrain disables to continue measuring (natural obstacle, private allotment etc.). In this case the message in the measuring record (both at the beginning and end of defected section) writes conventional character “*” into the note. This given section is excluded from the statistic evaluation by the program in the computer. When measuring, one of these three variants of Pearson method can be chosen: · · · current, used for pipeline defect detection or distances ordered in hundreds of meters, voltage, detecting defects in the section in tens of meters, CD (current direction), used for measured distances in units of meters. Current or voltage method is used if we need to test some longer section of the route to find out pipeline defects. If the amplitude of the signal in a section of a given lengths (current method) drops down by more then 80% or (voltage method) by more then 70 dB it is very obvious that some pipeline defects in a given section are to occur. This section should be measured by CD method detecting pipeline defects with the accuracy of 0,5 m. 4. SOFTWARE FOR PEARSON METHOD Software for measuring and evaluation of results made by SHINE firm (S. Sedláček owner) has two parts: 1. 2. Procedures for PSION organiser, software for PC ad 1) Software for PSION was created by the author of this paper in OPL (Organiser Programming Language), which is in the organiser at disposal. It concerns procedures ensuring following functions: · scanning the file list with measured values, · procedure for creation of a new file of measuring, · data transmittion between PS and organiser, · deleting and editting records of measuring, · scanning the file with measuring or the file with fixed points of the route, INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 241 · · · module for help, subroutine for the system information, procedure for measuring belonging to the most complicated one because apart from others it analyses the message from the locator, out of which only the most important information is choosen (reduction of memory capacity). ad 2) Application PEARSON for PC processes and evaluates measured values from the terrain. It is composed of relatively separated modules each of them performes a certain group of functions. It concerns following modules: · communication module ensuring the data transmittion between computer and organiser, · database module with basic functions for database operations with the files (database of the routes, fixed points and measuring), · module for graphical presentation of measured results. The gained values can be shown on the screen, printed on the printer or drawn on the plotter – see fig. 3. On the graph we can see considerable clusters of defects on positioning 62, 62.4 and 62.8 km, · module of system functions. It concerns utilities for data backup, exit into operation system and running of the favourite text editor for possible file editing etc. · the main modul for the processing of measurings containing procedures for a) data transformation from the formate of organiser into the database of measuring b) calculation of approximate positioning of the measured place (in steps) which is found c) between two fixed points into accurate positioning according to interpolate equation: ( xm = x + xmk - x 1 p 1 pk ) x 2p - x1p x 2pk - x1pk (1) where x1p or. x2p, is accurate positioning 1. or. 2. fixed point on the route in [m], x1pk or. x2pk is approximate positioning 1. or. 2. of fixed point of fixed point of the route in steps, xm or. xmk is calculated in [m] or. approximate in steps positioning of measured place a) b) filtres serve for scanning of database of measured values according to various criteria creation of protocol with measuring that is composed of three parts: · header of measuring which contains common data e. g. the name of organisation, names of pearsons who measured, measuring conditions, the name and number of the route, used method and devices, which were used, · from the detailed message about the measuring, which contain records of the measured places of the route, · from the statistics of measuring. In this part the defects are divided according to the intensity into: small, medium and large. Criteria of division can be set in the configuration file. In the fig. 3 there are denoted values between 0.5 – 3.5 – 7.5. Within the statistic the following data are processed above all: · number of defects in the given section, · cluster analysis e. g. a number of clusters of defects and a number of defects in every cluster. The cluster means a great number of defects of the pipeline their mutual distance does not cross a certain limit (in the graph on fig. 3 has limit a value of 5 m) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 242 Datum: 29/05/1996 Route: 5320400000 ZNOJMO – KVĚTNOV; Method : CD Intensity (length) [m] Positioning [km] Small Medium Large <F3> - Raster; <F4>/<F5> - Zoom; <F6> - Previous zoom; <F7> - Original size; <ESC> Return to menu Fig 3. Graphical output of measuring of pipeline defects detected by Pearson method – variant CD 5. CONCLUSION The above described Peason method is used in these gas plants in the Czech Republic: · Jihomoravská plynárenská a. s. Brno, · Severomoravská plynárenská a. s. Ostrava, · Středočeská plynárenská a. s. Praha, · Východočeská plynárenská a. s. Hradec Králové, · Severočeská plynárenská a. s. Ústí nad Labem · Východoslovenské plynárny Michalovce (Slovak Republic). Further the method is used as a part of system of anticorrosive protection in enterprises: · MERO ČR, Kralupy nad Vltavou and · ČEPRO a. s. The practical experience proved that this method is reliable because in the areas with the occurance of great clusters defects the insulation of pipeline was always considerably damaged. REFERENCES [1] ČLUPEK, O., DAVIDOVÁ, H.: Anticorrosive protection. GAS s. r. o., Praha, 1998, 141 s., ISBN 80-902339-8-8, in Czech. [2] SEDLÁČEK, S., BARTONĚK, D.: System PEARSON. User manual, SHINE Brno, 1996, in Czech. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 243 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 244 SYSTEMS FOR EVALUATION OF ANTICORROSIVE PROTECTION OF PIPELINE Dalibor Bartoněk1), Jaroslav Nesvadba2) 1) 2) Institute of Geodesy, Faculty of Civil Engineering, Brno University of Technology. Veveří 95, 662 37 Brno Czech Republic. Tel. 54114 7204, Fax: 54114 7218, E-mail: [email protected] European Polytechnical Insstitute Kunovice, Osvobozeni 699, 686 04 Kunovice, Czech Republic Tel. 572 548 035, Fax: 572 549 018, E-mail: [email protected] Abstract: Program systems GAS-ACOR and GASSERV made for the processing of the periodical measurement within anticorro sive protection of pipelines are described in this paper. These systems have been used since 1990’s in Czech gas-enterprises and in some gas-enterprises in Slovakia as well. They have also been used in companies Transgas Inc., ČEPRO Inc. and MERO ČR Inc. Keywords: Gas-pipeline, corrosion, protective devices, database, quantities. 1. INTRODUCTION Czech gasworks make use of transport pipeline net in the lengths over 33 000 km. Out of this number 14 000 km of remote and transit gas conduits and 19 000 km local gas conduits net [2]. Lately, for the dislocation of gas, the pipeline made by polyethylene has been widely used. This material is not suitable for remote gas conduit due to the limited dimension and pressure. Commonly used insulated steel pipeline located in the earth is connected with very serious problem, which is corrosion of metal. Metal materials located in the earth can be attacked by the corrosion for various reasons. Physical and chemical influences connected with the character of soil cause the so called simple corrosion. In the earth there can flow through electric currents their effect upon the devices located in the earth causes corrosion called „corrosion caused by the wondering currents“. Under certain conditions colonies of microorganisms can attach the surface of metals. This is called microbial corrosion. The direct corrosive losses calculated on the share of rough national product are moving in all countries arround 4%; 10% - 15% of these direct losses concern pipeline located in the earth. Undirect losses e. g. losses brought about the primary corrosive problem such as break out in the production, ecological disasters etc. are even several times higher than direct losses. This given data confirm, that anticorrosive protection of pipeline has a great importance both from the point of view of economy (especially with gas conduits) and from the point of view of safety of operation. The application of anticorrosive protection is rather expensive but due to its rapid economic restitution and compared to damages which can appear when the care is neglected, we should not regret our investments. 2. THE PRINCIPE OF ANTICORROSIVE PROTECTION As early as 1819 a member of French Science academy L. J. Thénard published in one of his books, that the corrosion is an electrochemical phenomenon. Sir Humphrey Davy and his pupil Michael Faraday who in 1934 discovered quantitative dependence between corrosive falls and electrical current. Two reactions are proceeding during the corrosion of metals: · Anodic, in the course of which the metal dissolves, · Cathodic also called depolarisation reaction during which electrons rising from anodic reaction are consumed and therefore corrosive medium is reduced. The aim of anticorrosive protection is to minimalise losses brought about by corrosion (both economic and ecological). As far as the supply of outer power to the protected object is concerned we distinguish anticorrosive protection: INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 255 · · passive, active. Corrosive processes on the pipeline or other devices located it the earth can appear only when water or oxygen come into the contact with the metal. In the passive protection it is necessary to separate the surface of the metal from the environment with a suitable material insulation in such a way to prevent none of depolarisators from the access to the metal. Further there is a possibility of making use of many organisation measures leading to the reduction of corrosion (the choice of the route, the way of pipeline location, the choice of the accessories etc.). In the active protection we come out of the fact, that corrosive procedures are caused by electrochemical reactions of the surface of the metal, thus it is possible to influence them by other electrochemical reactions. According to polarisation we differenciate anodic protection (the protected device appears as anode) and cathodic (the protected device appears as cathode). For the protection of outer surfaces of the devices located in the earth only cathodic protection can be used. The principle of cathodic protection is drawn on fig. No 1. Graph in the fig. No 1a outlines the dependence of anodic Ja and cathodic Jc current density on the U potential. At the value of corrosive potential Ucor is Ja characterising dissolving of the metal equals Jc which determines exclusion of the metal back to the environment. By lowering the voltage under Ucor the value Ja drops down which results +J Ja anode Jo A + – U insulation Jc -J Umin Ucor pipeline a) potential graph b) scheme of configuration Fig. No. 1. Principe of the cathodical protection in dissolving of the metal. Our aim is to lower value U under Ucor so that the corrosion can drop under technically possible limit. The value of the potential under which this state has been achieved can be denoted as a minimum protection potential Umin. It is important to supply the power from the external source. The value of a protective current corresponding to the current density J must be always somewhat higher than a corrosive current abstracted from the graph in the fig. 1a for Umin.The scheme of the cathodic protection of the pipeline see fig. 1b. The device we want to protect must be plugged in the negative pole DC and a suitable anode in plus pole thus we can cathodically protect any steel device. The routes of gas conduits called station of cathodic protection (SCP) electric polarised drain (EPD), saturation (S) work on the principle of cathodic protection. 3. ANALYSIS OF ANTICORROSIVE PROTECTION SYSTEM IN GASWORKS An extensive net of gasworks in the Czech Republic is divided into the individual parts – routes according to the topological principe. On each route are placed devices of anticorrosive protection both active (SCP, EPD, S) and passive (joining object - JO, insulation fittings – IF, protective fittings – PF) – detailed description see [2]. On every device various quantities characterising the state of the protection against corrosion are measured. The measuring is done periodically in various intervals (14 days, one month, within three months, yearly); it depends on the type of the device and the quantities. According to the valid enterprise standards every measurement must be written into the protocol which is implemented with a graph of protective quantities for each route once a year. In order to achieve an unambiguous identification the routes are denoted with numerical codes, each device has its own INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 256 positioning which is a distance from the beginning of the route in km. Let us define the requirements for the database of device and measuring which are not the only ones within the systém of anticorrosive protection nevertheless they have a great meaning for its right function. The database of the objects on the routes must be designed in such a way so that it could be possible: · to insert or to delete the certain device on any place of the route, · to insert several devices into the place with the same positioning on the route, · to keep all information concerning the individual devices the range of information depends on the type of the device. There are about 100 various devices which can be divided from the point of view of data attributes into the following groups: 1. station of cathodic protection (SCP), controlled station of cathodic protection (CSCP), 2. electric polarised drain (EPD), 3. joining object (JO), 4. joining object with diode element (JOD), 5. joining object between gas conduit and another device (JOA), 6. insulation fitting (IF), 7. protection fitting (PF), 8. potential scanner (PS) Each group within the data model is represented by the same set of attributes. From the point of view of measured quantities we must implemented the above mentioned groups with some other groups: · critical point – any device where the cathodic protection is almost over (T), · all the devices except for SCP, CSCP or EPD – (X), Every device can be characterised by some of the described types or combinations of these types (e. g. joining object with insulation fitting which is at the some time a critical point – JOIFT). We must admit, that some combinations are impossible to be made e. g. CSP + T or EPD + SCP. 1. to file the measured values of quantities (Q) on all devices (D) in various period (P). Each quantity is a function of two variables: Q = f(P, D) 2. 3. 4. (1) Novadays about 200 – 300 of various quantities that are sorted according to the locality and number of devices placed on these routes have been used. to research measured values on the device which does not exist at the given moment (change of the positioning, change of the type of device and so on), simply to restructuralize record in the device and measuring database without the loss of data integrity. These are the most important requirements for the database of devices and measurings. Some of other important functions concerning the system of anticorrosive protection are mentioned in another paragraph. 4. SOFTWARE SYSTEMS GAS-ACOR AND GASSERV In late 80’s the team of research workers of former Czech gaswork headed by Svatopluk Sedláček in cooperation with the author of this paper gave birth to the idea to create the system for automatization of anticorrosive protection [3] - [6]. A great experience of these technicians of all the involved enterprises in the Czech Republic and Slovakia was made use of. It was pondered about taking advantage of all accessible softwars e. g. dBASE, later MS Office and so on that were already tested in practice. These systems, however, were designed for general purpose with the wide offer of functions which is connected with 2 main problems: · a lot of functions lead to complications of the system and uneffective usage (some functions would hardly be made use of, others the missing ones would have to be programmed by the experts themselves), · by complications we mean tedions training of the operators who lack in experience with computers. Having discussed all “pros and cons“ the authors together with a group of potential users decided to make their own solution in Turbo Pascal programming language consequently Borland Pascal. In 1990 – 1992 basic moduls of programes GAS-ACOR were made (used in Czech gasworks since 1997 in ČEPRO) and GASSERV (used in Transgas, since 1996 in MERO CR). Both systems come out of the some conception and they have following groups of function at their disposal: INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 257 1. Database. Except for these described databases of devices and measurings there are also databases of pipeline deffects at our disposal, further database of all routes (GAS-ACOR), measuring in areals (GASSERV) and the database of the soil resistance and measuring pipeline defects by method pigl-sigl (GASSERV for MERO). The database of the routes serves as the control dials so that it cannot be possible to assign non-existing route numbers. The same measure holds for the code dial of the device. 2. Measuring. In this part of the application we can read or edit the measured values of the quantities in the form of tables. Various types of measuring are at our disposal (monthly, within 3 months, yearly) but at the same type it is possible to scan as many as of six different periods on the same route. Also the output of measured quantities in the form of protective quantities for the screen, printer and plotter ranks among very important functions of the system. In the graph can be facultativelly shown: · description of the device in short form, · as many as 6 different curves of the same quantity measured in various periods, · curves for various quantities measured at the same period for a given route, · places with defected insulation measured by Pearson method, · auxiliary raster in both horizontal and vertical direction for easy substraction of the values on x and y axes, · points on the routs with pipeline defects being found by inspection of the “pigl-sigl” method (only in GASSERV for MERO). The protective quantities can be seen in the fig. No 2. It is possible to import or export the measured values into the organiser of PSION, CASIO or BVComp type, with the help of which measuring is done directly in the terrain. A very important function of the system is the transmission of measuring between ČEPRO and MERO organizations through special convert file (contains corresponding numbers of routs used in ČEPRO and MERO). In the tables of measuring there are built in filters of values of the individual quantities and names of the devices with the possibility to print. 1. Application. This menu item was made for the needs of the user because all the submenu is writeen into the text file GACMENU.TXT. For each item of the pull-down menu it is necessary to write into the GACMENU.TXT file 3 lines: · the first line with keyword “Menu Item” followed by the name of the item, · second line begins with keyword “Menu Help” with the text of help, · third line contains the batch statement which is run when this menu item is activated. It has its advantage because everybody can make his own submenu by inserting 3 lines into the text file in any text editor. The number of items is limited to 100. These menu items are preset: · yearly backup of measuring, · daily backup measuring, · communications with PSION, BVComp or CASIO organisers, · run of own special interpreter of SQL statements with the in-built list of the most frequent questions. 2. Outputs. This menu item contains the pull-down menu with following items: · converting into TXT or DBF formate, · compress of all database files, · joining database files of various enterprises. 3. Services. This menu item comprises: · configuration and initial settings of parametres for the output on the printer and plotter, · run of external favourite text editor, · exit to operation system. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 258 Year: 2000; Routes Number: 108 ... 115 Positioning [km] <F3> - Raster; <F4>/<F5> - Zoom; <F6> - Previous zoom; <F7> - Original size; <ESC> - Return to menu Fig. No. 2 Graph of the protective quantities 5. CONCLUSION The GAS-ACOR and GASSERV systems have already been in run for more then 10 years and they have been upgraded according to the users needs. The last upgrade consists in introducing module for universal file print in the network environment. This module vas created in Borland Delphi. The authors provide the users with help in the form of hot line or personally. This year it has been worked on the converting of the GAS-ACOR system from the MS-DOS system into Windows. Source modules are made in Borland Delphi and database is created in MS Access. Link between source modules and database is realized by ODBC (Open Database Connectivity) interface. The testing run of the new system will be operated in South-Moravia gaswork in Brno. The introduction of GASACOR into practice will mean a considerable time saving and the increase of the efficiency of the anticorrosive protection. REFERENCES [1] THÉNARD, L. J.: Annales de Chimie et de Physique 11, 1819, p. 40. [2] ČLUPEK, O., DAVIDOVÁ, H.: Anticorrosive protection. GAS s. r. o., Praha, 1998, 141 s., ISBN 80-902339-88, in Czech. [3] SEDLÁČEK, S., BARTONĚK, D.: GAS – ACOR. User manual. SHINE, Brno, 1996, in Czech [4] SEDLÁČEK, S., BARTONĚK, D.: GASSERV. User manual. SHINE, Brno, 1997, in Czech [5] SEDLÁČEK, S., BARTONĚK, D.: GASSERV pro MERO. User manual. SHINE, Brno, 1998, in Czech [6] SEDLÁČEK, S., BARTONĚK, D.: GAS-ACOR pro ČEPRO. User manual. SHINE, Brno, 1999, in Czech INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 259 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 260 COURSE TIMETABLING BY AN ARTIFICIAL IMMUNE SYSTEM Jaroslav Čechman Institute of Automation and Computer Science Brno University of Technology, Technická 2, 616 69 Brno, Czech Republic E-mail: [email protected] Abstract: This paper is aimed at design an artificial immune system that is used to solve a basic school course timetabling problem. The design of basic components and methods is progressively depicted. The experimental section includes parametres that have a bearing on behave of an artificial immune system. Keywords: timetabling, optimization, artificial immune system 1 INTRODUCTION Scheduling is difficult for two reasons. Firstly, it is computationally complex problem, described in computer science terms as NP-complete. This means that search techniques that deterministically and exhaustively search the space of possibilities will probably fail because of time requirements. In addition, search techniques that use heuristic to prune the search will not be guaranteed to find an optimal (or even good) solution. Secondly, scheduling problems are often complicated by details of a particular scheduling task. Timetabling is a common example of a scheduling problem and can manifest itself in several different forms. The particular form of timetable required is specific to the environment or institutions in which it is needed. The application of computers to timetabling problems has a long and varied history. Almost as soon as computers were first built there were attempts to use them to solve these problems. The first generation of computer timetabling programs in the early 1960’s was largely an attempt to reduce the associated administration work. Soon, programs were presented with the aim of fitting classes and teachers to periods. There has been a growing interest in the use of metaphors extracted from the immune system for the development of the so-called artificial immune systems. AIS have been applied to a wide variety of domain areas, such as pattern recognition and classification, optimization, data analysis, computer security and robotics. The paper is arranged as follows: the following section describes a background and a framework to general design of artificial immune systems. The design of fundamental components and methods of AIS is detailed in section 3. In section 4 the timetable requirements are defined and section 5 describes experiments and results. The final section discusses all findings and draws conclusion. 2 ARTIFICIAL IMMUNE SYSTEMS The establishment of the field of the artificial immune systems (AISs) has been difficult for a number of reasons. Firstly, the number of people active in the research area is still small but has been increasing in the past few years. Secondly, researchers found it difficult to identify the difference between an AIS and the work undertaken in theoretical immunology. Thirdly, the application domain of artificial immune systems is very wide range. Finally, only recently the first textbook proposing general framework to design AIS has been published. There were limited number of attempts to define the field of artificial immune systems. The present work adopts the concept in which artificial immune systems are defined as computational systems inspired by theoretical immunology and observed immune functions, principles and models, applied to solve problems [4]. This definition covers some of the aspects mentioned obove by drawing a fine line between AIS and theoretical immunology: the applicability. While works on theoretical immunology are usually aimed at modelling and INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 261 providing a better understanding of the immune functioning and laboratory experiments, works on AIS are applied to solve problems in computing, engineering and other research areas. This is more akin to a soft computing paradigm. It is beyond the scope of this paper to present and discuss theoretical aspects of artificial immune systems in more detail. However, in order for further comprehensibleness, we summarize a framework to design artificial immune system, at least, in the following basic elements: · a representation for the components of system; · a set of mechanisms to evaluate the interactions of individuals with the environment and each other; · procedures of adaptation that govern the system dynamics. A basis of framework to design artificial immune systems: a representation to create abstract models of immune organs, cells and molecules; a set of functions, termed affinity functions, to quantify the interactions of these „artificial elements“; and the set of general purpose algorithms to govern dynamics of the AIS. Figure 1 summarizes the elements involved in the framework to engineer an AIS: this can be thought of as a layered approach of the design procedure. Figure 1: The framework to engineer AIS and its layered structure. 3 AIS DESIGNING – TIMETABLE IMPLEMENTATION Educational timetabling is a major administrative activity for a wide variety of institutions. A timetabling problem can be defined to be the problem of assigning a number of events into a limited number of time periods. A. Wren defines timetabling [9] as follows: “Timetabling is the allocation, subjects to constraints, of given resources to objects being placed in space time, in such a way as to satisfy as nearly as possible a set of desirable objectives.“ In this paper, we will concentrate on basic school course timetabling problem. This problem is subject to many constraints that are usually divided into two categories: “hard“ and “soft“ ([2]). Hard constraints are rigidly enforced. Examples of such constraints are: · no resource (pupils or stuff) can be demanded to be in more than one place at any one time; · for each time period there should be sufficient resources (e.g. rooms, teacher etc.) available for all the events that have been scheduled for that time period. Soft constraints are those that are desirable but not absolutely essential. In real-world situations it is, of course, usually impossible to satisfy all soft constraints. Such problems can be: · time assignment: a school subject may need to be scheduled in a particular time period; · time constraints between events: one course may need to be scheduled before/after the other; · spreading events out in time: pupils should not have two same lessons on the same day; · coherence: teachers may prefer to have all their lessons in a number of days and to have a number of lecture-free days – these constraints conflict with the constraints on spreading events out in time; · resource assignment: teachers may prefer to teach in a particular room or it may be case that the particular school subject must be scheduled in a certain room. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 262 3.1 A Representation Scheme The fundamental components of artificial immune systems are immune cells (T-cells, B-cells). They present surface receptor molecules whose shapes are complementary to the shapes of antigens, allowing them to recognize disease-causing agents and then perform an effector function. The immune cells and molecules are therefore elements that have to be modelled and used to create AIS. Perelson and Oster [8] first proposed the concept of shape-space (Sh). Bearing in mind that the recognition of antigens is performed by the cell receptors, shape-space allows a quantitative descriptions of receptor molecules and antigens interactions. As in the biological immune system, in a shape-space, the degree of binding (degree of match or affinity) between an antigenic receptor (antibody Ab or T-cell receptor TCR) and an antigen (Ag), is measured via regions of complementarity. The set of features that describe the relevant properties of a molecule from a recognition perspective is termed its generalized shape. The generalized shape of an antibody is described by a set of L parameters. Thus, a point in a L-dimensional shape-space specifies generalized shape of antibody binding region with regard to its antigen binding properties. Mathematically, the generalized shape of a molecule (m), either an antibody (Ab) or an antigen (Ag), can be represented as an attribute string (set of coordinates) m=m1,m2,..,mL, mShL, or other more elaborate structures such as a neural networks or a Petri net. In a course timetabling case, in addition to basic attributes (teacher, subject and classroom), the string should contains the information about scheduled time (day, hour). If we imagine timetables as a sequence of each group timetable we should use matrix notation, where matrix coordinates code a scheduled time of an item on each group. The item consists information like teacher‘s name, number of classroom, subject‘s name etc. Figure 2 depicts this conception: there are three groups (1.A, 1.B, 1.C) and three subjects (History, Mathematics, Physics). Group 1.A has a history lesson on Monday morning (first time slot) and a physics lesson on Thursday afternoon (sixth time slot). Group 1.B has a mathematics lesson on Tuesday morning (eighth time slot). In other words, timetables form matrix S(g,d,h), where item si,j,k is a record that contains information about subject, teacher, classroom etc. In our case item s0,0,0 corresponds to a history lesson in 1.A group on Monday morning. Figure 2: Schedule Scheme 3.2 Affinity Measures The type of shape-space (representation), used to model an antibody and an antigen, will partially determine a measure to calculate their affinity. As the Ag–Ab affinity is related to their distance, it can be estimated via any distance measure between two strings or vectors, such as the Euclidean, the Manhattan, or the Hamming distance. Hence, if coordinates of an antibody are given by Ab=Ab1, Ab2, .., AbL and those of an antigen are given by Ag=Ag1, Ag2, .., AgL , then the distance D between them can be defined as: L D ‡”(Abi - Agi )2 (1) i =1 L D ‡”Abi - Agi (2) i =1 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 263 L D= ‡”δ1 ,where δi = { i=1 1 if Abi •‚Agi 0 otherwise (3) where Eq. 1 is the Euclidean distance, Eq. 2 Manhattan distance and Eq. 3 the Hamming distance. It is said that constraints of timetabling problem form two categories: hard and soft. To quantify these constraints we establish a certain functions (discrete or continuous) that depend on how important the constraint is. In case of basic school we could prefer to reschedule the history lesson to the afternoon and subjects as mathematics and physics are occurred in the morning. Figure 3 depicts two examples of priority graphs. The time flow at the Fig. 3(a) signifies our request for situating particular subject to fourth time slot (the best solution). The Fig. 3(b) refers to case when a subject is learned twice a week. Figure 3: Samples of Priority Graphs. (a) Daily subject characteristic. (b) Weekly subject charactristic. The cell affinity corresponds to quantitative evaluation of the solution. If the solution consists of many collisions (conflict in the schedule), then affinity of the cell that represents this solution is lower than cell affinity that represents solution without collisions. The criterial function is resulted from cell affinity: g fa = d h ‡” ‡” ‡”St , j ,k •¨ max . (4) t =1 f =1 k =1 where fa is cell affinity, g, d and h are boundaries of timetable (max. number of groups, days per week, hours per day), si,j,k is the item of matrix S(g,d,h). 3.3 Immune Algorithms Immune algorithms govern the behaviour of the system. The fundamental functions are cell generator, clonal selection and affinity maturation. The immune cells and molecules are generated in the bone marrow. The genes used to encode the receptor molecules are stored in separate and distinct libraries. The encoding of these molecules occurs through the concatenation of different gene segments that are randomly selected from each of the gene libraries. The bone marrow model is used to create the matrix that represents the immune receptors. The simplest bone marrow model is the one that generates matrix items by using random number generator. The more sophisticated model takes advantage of knowledge about timetable arrangement. This means that we can control the new cell growth and the system doesn‘t produce transparently useless cells. In [4] the authors focused on clonal selection principle and affinity maturation process of adaptive immune response in order to develop an algorithm suitable to perform task such as machine learning, pattern recognition and optimization. The main immune aspects taken into account to develop the algorithm were: selection and cloning the most stimulated cells proportionally to their affinity; death of non-stimulated cells; affinity maturation and selection of cells proportionally to their affinity; and generation and maintenance of diversity. On the basis of this algorithm two types of mutation were designed: hypermutation and somatic mutation. The difference between hypermutation and somatic mutation is that the hypermutation uses whole timetable range of group to reschedule given subject but somatic mutation shifts the subject through a day only. Moreover, somatic mutation takes subject priority time flow into account. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 264 It is said that number of mutations is proportional to cell affinity. This means that the highest affinity cell (with affinity Dmax) undergoes a mutation where the only one subject is rescheduled. The mutation rate () of all cells is determined by Eq. 5: ( ) α D* = exp(- p.D) (5) where is a parameter that controls the smoothness of the inverse exponential function, and D is the normalized affinity that can be determined by D=D/Dmax. 4 TIMETABLE SPECIFICATIONS The basic school course timetabling problem is solved. Table 1 summarizes all requirements that must be observed. Category of subjects represents the identical subject characteristic, i.e. first category covers all subjects that have same characteristic as priority graph that is depicted in Fig. 3(a). All subjects and teachers have only symbolic names. Table 1: Timetable specifications Category of Subject Subjects Subjects a Week 1. S01 2 S02 2 S03 5 S04 2 S05 2 2. S06 2 S07 2 3. S08 5 S09 1 4. S10 1 S11 2 S12 2 S13 2 5 Teacher T01 T02 T03 T04 T05 T06 T07 T08 T09 T10 T11 T12 T13 8.A 301 101 301 301 301 301 301 301 301 102 301 103 301 8.B 302 101 302 302 302 302 302 302 302 102 302 103 302 Classroom 8.C 303 101 303 303 303 303 303 303 303 102 303 103 303 8.D 304 101 304 304 304 304 304 304 304 102 304 103 304 EXPERIMENTS AND RESULTS The experiment incorporates two levels. On the first level we work with simplified timetabling problem, where no conflicts of the schedule are presented (take only the one group into account). On the second level we work with timetabling problem exactly as is denoted in table 1. Figure 4 depicts two kind of mutation: hypermutation (M1) and combination of hypermutation and somatic mutation (M2). Figure 5 depicts how is important to find a suitable ‘smoothness‘ parameter (cS) for optimal system working. The value of cS2 () is counted inverse to Eq. 5, where (D)=1. In Fig. 6 different sizes of population (PS) are explored. It is said that the clonal selection produces new clones of the cells proportionally to their affinity. But the size of the population is limited. Clonal ratio parameter (CR) determines the size of cloned cells that constitute whole population. Parameter CR is evaluated in order that clone population size is 10% (CR1), 50% (CR2) and 100% (CR3) of whole population. Figure 7 depicts how the clone population size is effective in exploration of the search space. Figure 8 is similar to Fig. 6, but all groups in timetable are presented. Figure 9 show the average value of the conflicts in the schedule. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 265 Figure 4: Type of mutation. M1 – hypermutation; . M2 – hypermutation and somatic mutation Figure 5: Mutation rate. cS1 < cS2 < cS3. Figure 6: Population size. PS1=100, PS2=500, PS3=1000. Figure 7: Clone ratio. CR1≈10%, CR2≈50%, CR3≈100% of population size Figure 8: Population size. PS1=500, PS2=1000 6 Figure 9: Conflict in the schedule. PS1=500, PS2=1000. CONCLUSIONS The paper shown the power of AIS to solve course timetabling problem. In previous section we can see that is very important to make the right evaluation of such parameters as population size, mutation rate and clone ratio. These parametres are essential to effective system functionality. Two types of mutation procedure show that we can use mutation control to gain in optimization efficiency. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 266 REFERENCES [1] BURKE, E.; ROSS, P. (Eds.) The Practice and Theory of Automated Timetabling. Springer Lecture Notes in Computer Science Series, 1996, sv 1153. [2] BURKE, E.; KINGSTON, J.; JACKSON, K.; WEARE, R. Automated University Timetabling: The State of the Art. The Computer Journal, 1997, str. 565–571. [3] BURKE, E.; PETROVIC, S. Recent Research Directions in Automated Timetabling. European Journal of Operational Research, 2002. [4] de CASTRO, L. N.; TIMMIS, J. Artificial Immune Systems: A New Computational Intelligence Approach. Springer, 2002. ISBN 1-85233-594-7. [5] de CASTRO, L. N.; TIMMIS, J. Artificial Immune Systems as a Novel Soft Computing Paradigm. To appear in the Soft Computing Journal, 2003, sv. 7. [6] ECHMAN, J. Optimalizace systému s využitím umlého imunitního systému. Diploma Thesis, Brno University of Technology, 2003. [7] OŠMERA, P.; MASOPUST, P. Schedule Optimization Using Genetic Algorithms. Proceedings of MENDEL 2002, Brno, Czech Republic, 2002, str. 132–138 [8] PERELSON, A.; OSTER, G. Theoretical Studies of Clonal Selection: Minimal Antibody Repertoire Size and Reliability of Self-Nonself Discrimination. Journal of Theoretical Biology,1979, sv.81, str. 645–670. [9] WREN, A. Scheduling, Timetabling and Rostering – A Special Relationship? In: (Burke1996a), str. 46–75. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 267 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 268 MANAGING A HIGH SPEED LAN USING DISTRIBUTED ARTIFICIAL INTELLIGENCE Ibrahiem M.M-El-Emary Faculty of Applied Science, Al-Balqa Applied University, Al Salt, Jordan. E-mail:[email protected] Abstract: This paper is concerned with a practical application of distributed artificial intelligence for managing the high data rate bus structured local area computer network that uses deterministic multiple access protocol. In the selected network that is managed using distributed artificial intelligence, the dynamic sharing of the available bandwidth among stations is achieved by forming “train to which each station may append a packet after issuing a reservation. Reservation and packet transmissions are governed by the reception of control packets (token) issued by the network end stations. Managing approach that was suggested depends on using intelligent autonomous agents, which are responsible for various tasks among it: election of the end stations, the recovery from failures, and the insertion of new stations in the network. All these tasks are based on the use of special tokens. Keywords: Distributed Artificial Intelligence, Autonomous Agents, neural Networks, EXPRES-NET, CSMA/CD 1. INTRODUCTION The main task of the network management system researchers is to develop new tool that work better than current by available tools for doing this laborious work. The work presented in this paper directed about how to delegate as much as possible to the machine, using network administrators as knowledge engineers that teach the machine how it should perform its work based on what is called Intelligent Autonomous Agents. There are a few predictable advantages of using Intelligent Autonomous Agents for any management system, network management being just one. First of all, there is the intelligent nature, which sounds promising. Having an intelligent and adaptable system is usually better than having dedicated applications for specific solution. Likewise, the word Autonomous gives us the idea of something that can work by itself or need almost no human interference. Finally, agent gives the impression of a helper or a wizard that some how works between the machine and human. So, the main objective proposed here is to project a system where the human system administration does not have to work as the truly intelligent-meaning cognitive- element in the management process, feeding the system with the rules of work or knowledge-it needs to operate, no more boring work, but intelligent work [5]. The network that is managed by the above approach categorized as a local area network labeled by TOKENET [4], its architecture based on a linear bidirectional bus, and on a deterministic channel access protocol, whose efficiency is similar to that of EXPRESS protocols [3]. Advantages of this network over the linear topology (L-EXPRESS-NET) are that stations need not be numbered, and that no silence counter is needed in order to schedule transmissions. The used LAN is named TOKENET since tokens (control packets) are used in the operations of the channel access scheme. The protocol algorithms were developed for a scenario in which Ethernet-like networks cannot be used, i.e. considering a large population of users that can be connected using several kilometers of cable, and whose traffic requirements must be satisfied with a very high data rate (>100 Mbps). The packet transmission time is assumed to be much shorter than the end-to-end propagation delay, so that the CSMA/CD yields extremely poor performance. In order to manage the above mentioned network using the Intelligent Autonomous Agents, we present this paper which is organized from the following: section two describes the architecture of the managed network. In section three, we explain the approach that is used in accessing such a network type. The procedures that indicate the system initialization was discussed in section four. The concept of how to recover a failure as well as INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 269 inserting a new station was devoted in section five. Section six was dedicated to clarify how to manage these TOKENET using the Intelligent Autonomous Agents. Finally, the paper was enclosed in section seven with conclusions and further works to be done by others. 2. ARCHITECTURE OF THE TOKNET NETWORK TOKENT is based on a linear topology. It is thus possible to identify the rightmost and leftmost active stations (where by active we mean that stations are executing the protocol algorithm). These two stations are named “end stations”. Stations are connected to the bus by means of passive taps that can simultaneously receive and transmit [1]. Packet collisions can be detected. Interfaces are capable of recognizing a limited set of short control packets, named tokens, whose functions will be described later on. The reception of tokens as well as the reception of packets may produce some response by a station. The time needed in order to start this response is td seconds; this includes the time for the recognition of the token (or the detection of the end of a packet reception) and the time to initiate the response. td should be of the order of a few bit times[4]. Data packets are transmitted in trains, composed of a token (the locomotive) followed by some packets (possibly none, in which case the train is said to be empty). The time separation between successive elements of the same train is td seconds. Packets may board the train only after a reservation has been issued. By listening to other stations reservations each station can determine the correct position of packet in the train. Transceivers can operate in four different modes depending on the state of the station: the CSMA-CD mode and the polite sensing mode are used during system initialization in order to elect the end stations; the reservation mode is used when a station has a packet ready for transmission in order to reserve a place on the next train; the scheduling mode is used after a reservation has been made in order to transmit the packet in the train in the reserved position. The CSMA-CD mode implies the transmission of packets following the CSMA-CD protocol [1, 2]. It is assumed that tc seconds are necessary to recognize a collision. The time tc is a few times longer than td. When a station is in the polite sensing mode it starts transmitting its packet after recognizing the end of a packet transmission from another station. If a collision is detected in the initial tc seconds the transmission is aborted and a new attempt will be made at the end of the colliding packet. When a station is in the reservation mode it transmits a short burst (any predefined bit pattern) when it receives the reservation token. Stations in the scheduling mode count other station reservations, determine which their received wagon in the train is and transmit their own packet in the reserved position. If a collision occurs in the initial tc seconds the transmission is immediately halted. Receivers are always enabled. Packets (and burst) must be separated by at least td seconds to be received. Packets can be of variable length. Before each packet a preamble is transmitted in order to synchronize the destination station receiver. The preamble duration is of the order of several tens of bits. In some cases a long preamble is necessary. It is not necessary, insteade, to transmit a preamble before the reservation bursts, since they must only be detected and need not be decoded. 3. ACCESS SCHEME USED IN TOKENET NETWORK Assume the two end stations have been found, and denote, as in Fig.1, by L the leftmost active station, and by R the rightmost active station. Upon recognizing the end of a train, R issues an R-token, denoted by Tr , and after td seconds transmit a reservation burst. The token and the burst propagate towards all stations to the left of R. Each station with packets ready for transmission, upon reception of Tr immediately transmits a burst of short duration that serves as a packet reservation. Due to the station latency time the burst is separated from Tr by a gap of duration td. Denote by tb the time duration of the burst. Provided that the distance between any two adjustment stations is larger than the distance over which the line signal propagates in (td+tb)/2 seconds, stations receive bursts transmitted from station to the left, separated by intervals of duration larger than td. It is thus possible for each station to count the number of bursts transmitted by stations to the left. Denote by Nj the count of station j. Bursts transmitted by the station to the right are instead seen as completely overlapping and are thus undistinguishable. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 270 Fig.1 TOKENET topology; L and R are the network end stations Finally Tr reaches L, that may transmit a reservation burst as any other station if it has a packet ready for transmit ion. After either transmitting the burst or just hearing the overlapped bursts of all stations with ready packets, L issues on L-token T1 that serves as locomotive of a train of packets. If L has just transmitted a reservation burst it also append the data packet to T1, leaving a silent gap of duration Td in between. The train propagates now in the L to R direction and those stations that just made a reservation may append packets. Before appending its own packet to the train station j must recognize T1, and count Nj packets in the train beyond T1. The train propagates all the way to R collecting new wagons. R may append a packet to the train too, and when R recognizes the end of the train it issues a new R token, thus starting a new cycle. A fault in the operation of the station may cause a truncation of the train: if a station has a wrong reservation count, it will either transmit too early (the transmission is stopped after Tc seconds because of the collision), or not transmit because it will not find enough packets in the train before its own. Both events prevent all stations to the right of the faulty one to transmit in the current train. The same thing happens if the station issues a reservation and then dose not transmit the packet. If instead, a station transmits a packet without transmitting the corresponding reservation, a collision may take place id some station to the right transmits a packet, but other stations can still board the train. Otherwise, if no collision occurs, then the train maybe larger than expected. All above faults conditions only influence next cycle. Other faults concerning either token losses or failures of the end station must be reserved through a system reinitializing. Quantatively, when we show the impact of the constraint on the minimum distance between adjustment stations, we see that for a 100 Mbps transmission speed, assuming that the sum (td+tb) is equivalent to ten bit times, and that the propagation delay is 10µs/km. The time interval corresponding to five bits is 50 ns, that is equivalent to a five-meter distance. Stations must hence be separated by at least five meters of cable. 4. INITIALIZATION PROCEDURE FOR TOKENET USING CSMA/CD PROTOCOL I f an active station doesn’t hear any packet, token, or reservation burst on the bus for more than To seconds (where To is of the order of several round trip delays), it goes onto initialization mode in order to elect the end station and restart the protocol operation. The initialization procedure consists of the following steps: *(A) Using the CSMA/CD protocol one station (say station A) acquires the channel and issues an initialization token. (Note: this may require several retransmission attempts. At each attempts only the station which collided in the previous attempt participates). To avoid ambiguities the initialization token length is such that transition time is longer than the end-to-end propagation delay. *(B) Upon hearing the initialization token, the remaining active stations compete again (as in step 1) until one station (say station B) acquires the channel and issues a response token. *(C) Station A issues a token which is called token A. Station B issues token B immediately upon hearing token A. All the other stations are listening and can determine their relative position with respect to station B as follow: for the station in region 1 (see Fig. 1) there is a gap (larger than Td seconds) between the time token A is received and the time token B is received. For the stations in region 2 token A is received immediately (more precisely td seconds) before token B. Station in region 1 are temporarily inhabited to respond to R station location tokens (see next step). *(D) Station A issues an R station location token that starts a train of tokens since all stations in region 2 (stations that are not inhabited) append a similar token to the train using the polite sensing mode. Each station monitors the channel after transmitting its token. If a station does not hear any other transition after its own it concludes that it is the right end station (R). The token has a special format, namely, the preamble length is such INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 271 that reamable transmission time is larger than the end-to-end propagation delay. This is necessary to guarantee that all region 2 stations correctly receive tokens. *(E) The newly elected R station issues an L station location token, and the left end station (L) are located using the same procedure above. Again the token preamble transmit ion time is longer than the end-to-end propagation delay. *(F) The L station issues an L token to notify the end of the initialization phase. *(G) The R station receives the L token and issues an R token, thus starting the data transfer phase. 5. OPERATION CONCEPT OF FAILURE RECOVERY AND INSERTING A NEW STATION The failure of an intermediate station (i.e., a station located between L and R) has no impact on the protocol; thus it does not require a specific recovery procedure. The failure of left or right end station, instead, requires the selection of replacement. Assume that the R(L) station fails. The L(R) station will detect the failure from the fact that the R-token (L token) does not come back after the line has been idle for a round trip time. After learning that the R(L) station is down, the L(R) station decides it is the new R station, issues an L station location token, and the initialization procedures is executed step 5 on. At the end the new end stations are elected. Note that if the L station fails, the old R station remains the new R station after the recovery, whereas, if the R station fails, the old L stations becomes the new R station. When a new station become active, it is necessary to allow it to join the other stations in the dynamic sharing of the communication channel. If the new station is in an intermediate position, i.e, between station L and station R, then it may immediately participate in the protocol procedures. If, instead, the station is either off to the right or off to the left, then it is necessary to reinitialize the whole network. The decision the new stations must take whether to start the initialization procedure itself by listening to the channel. If the new station receives the R token and reservation bursts immediately followed (after td seconds) by the L token, then it concludes it is off to the left. If the new station detects the end of the train immediately followed by an R token then it concludes that it is off to the right. If the new station, by listening to the channel, determines that it is not located between L and R, then it decides it is the new R station, and issues an L station location token, thus triggering the execution of the initialization procedure from step 5 on. At the end of the execution of the algorithms the new end stations are elected, and the new station has become station R. 6. TOKENET NETWORK MANAGEMENT USING AUTONOMOUS AGENT APPROACH The main objective in this session is to show how the agents fit in a standard network management system. The used approach depends on presenting a diagram from a regular system and then replaces its internal component by another diagram; using agents based components with similar functionalities. An ordinary network management could have a workflow like the one presented in Fig.2 [6] in which data is collected from a another managed device (1) using some standard management protocol as SNMP. After that, in step (2), the collected data is analyzed, and then condensed. In step (3) creating the real management information. Later on, managers make their analysis in step (4) and, if required, take reaction in order to correct weak points of the system. Based on Fig.2, we have replaced the management approach with another one shown in Fig.3 adapted to agents’ technology oriented. The types of used agents are: data collected (AgF) used to coordination of knowledge exchange between agents of the same community. In the following, we describe each part of the generic agent’s framework shown in Fig.3 as follows: (a) Data Collector agents (AgP) ; this type of agents retrieve data from devices and saves this information on a local knowledge database. This knowledge could be kept and processed locally and then shared to a whole community through the agents’ framework shown in Fig.4. The most part of data collector agent functionalitywhich is collecting data- is implemented by special knowledge rules loaded inside its knowledge database. These rules, in association with native knowledge implemented in the code of the agent, allow them to interface managed equipment through SNMP protocol. A signal data collector may collect information from several different devices. It could be setup with multiple different goals, with one or more goals for each device. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 272 Fig.2 Network management workflow. Fig.3 Network management with agents Symbols used in Fig.3 are given by the following: AgP: collector agent. AgM: manager agent. AgI: Interface agent. AgF: Facilitator agent. KQML: knowledge query management language. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 273 Fig.4 Generic agent structure (b) Management agent (AgM); are responsible for the organization of operation between agents of the same community .In addition, they can implement a higher level of data analysis based on shared knowledge (data) coming from several different collector agents. Finally, management agents coordinate multiple data collector agents on working on a special set of devices and later correlating that data. Besides, we have other types of management agents associated to agent community and not directly to network management environments. What we mean by this is that those agents will be presented on any agent community, not just on those dedicated to management self-creation, agents that neural network structure, wizard agents- those that hold and share massive knowledge data base-and many others being projected in order to create an agent virtual world [7]. The types of agents specially setup for network management purposes are: * Data analyzer; which analyze data retrieved from different knowledge bases, usually from distinct data collector agents. *Data consolidator; which combines or correlate data coming for a set of distinct data collector agents, creating new consolidated information. *Future value predictors; which are implemented by using neural network structures designed for time series prediction. With this class of agents our intention is to implement entities that theoretically could predict future values in a time series. This feature is quite helpful in designing proactive management system. *Value classifiers; which are implemented by using neural network as well, permit data classification in scales, like HIGH, NORMAL, or LOW instead of linear values. This classification is implemented by pattern machines and high adjustable being even possible to create neural network structures that adjust themselves based on past values of the time series. So far the neural agents have been studied as an improvement for the system functionality. Neural agents work as servers for the neural network structures, where batches of data are submitted for processing and result returned. In the future we are idealizing to have hybrid agents, with both rules- based and neural structures, inside INTERFACE modules. (c) Interface agents(AgI); are the bridges between the agent communities and the human managers. There are various possibilities for these interfaces, like e-mail interfaces, terminal like interfaces and web interfaces. Selecting which interface the human manager prefers the friendlier interface in some aspects are the terminal like interface, where humans can "chat" with agents using structured language similar to human's natural language. Interface agents, like any other, have their functionality structured inside the community and also the agent world. When a new sentence is input, the local parser, which is implemented by means of rules inside the knowledge database, will try to translate and understand it. If this is not possible locally, usually due to lack of local knowledge, the task will be escalated a more skilled-inside the community or not- that will try to get translation for it. If in the end the agents can't still understand the human entered sentence it will reply-as of current implementation an "I can’t understand you" sentence as ask to repeat. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 274 (d) Facilitator agents (AgF); are communication management agents needed in the agent community to work as brokers for the information exchange process. They also implement some useful tasks related to data exchanging like list of available service names, routing, meditation, and translation. 7. CONCLUSIONS AND FUTURE WORKS Managing the TOKENET local area network has been the aim of this paper. The reason behind selecting this network type comes from the ability of this network to be used for a high speed local area computers network using bidirectional bus, all network separations which must be executed through the agents are based on the transmission and reception of control packets (token). Tokens issued by the network and stations are used to schedule packets transmission in the from of a train on which a packet is admitted only after a reservation has been issued by the original station. The algorithm for the election of the end station, the recovery from failure, and the insertion of a new station are all based on special token transmission. The architecture of TOKENET is completely distributed: any station may be elected and station depending on which station are at some time connected to the network. When end station fail, replacements are elected among those stations that survived the failure. Our future work will be driven by the following: To integrate even closer the promising field of neural networks inside current interface module, in a transparent and extensible way. So, furthermore , we will have truly hybrid agents that will be able to take advantage of the best of both words. For network management, we use a direct utilization by creating neural structures to predict temporal series dynamically. This would be used in proactive network management through future value prediction. REFERENCES [1] TANENBAUM, A.S.: Computer Networks, Fourth Edition, Pearson Education, Inc.,Prentice Hall PTR, Upper saddle River, New Jersey, 2003. [2] STALLINGS, W.: Data and Computer Communications, 6TH Edition., Upper Saddle River, N.J., Prentice Hall, 2000. [3] FRATTA, L., BORGONOVO, F. and TOBAGI, F.A.: "The EXPRESS-NET: a Local Area Communication Network, Integration Voice and Data "Proceedings of the International Conference on Performance of Data Communication Systems and their Applications, Paris, France, September 1981. [4] MARSON, M.A., GERLA, M.: " TOKENET-A Token based Local Area Network", Proceedings of Mediterranean Electron Technical Conference, Atens,Greece, May 1983. [5] CHIKHOUHOU, M. M., CONTI, P., MARCUS, K., LABETOULLE, J.: A software Agent Architecture for Network Management: Case Studied and Experience Gained, INSM Special Issue on Intelligent Agents for Telecommunications Management, vol.8, No.3, Sep.2000. [6] MAES, P.: Agents that Reduce Work and Information Overload, Comm. ACM, vol.37, No.7, U.S.A, 1994. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 275 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 276 MARKOV CHAINS AND EXAMPLE OF THEIR USE Vítězslav Ševčík Brno University of Technology, Faculty of Mechanical Engineering Institute of Automation and Information Technology, Technická 2, 616 69 Brno Tel.: +420 541 143 349, Fax: +420 541 143 334 Email: [email protected] Abstract: This article clarifies and summarises the fundamentals of Markov analysis and presents some fields of usage of Markov chains as an example. Markov Analysis helps us to understand a certain sort of stochastic processes that can be described by the so-called Markov Chain. This article clarifies the fundamental methods and shows utilising of the Markov Chain, focusing on the descriptive methods. Keywords: Markov chain, Markov analysis Optimising, Theory, Finite state first order Markov Chain 1. INTRODUCTION The Markovian model is applicable in many practical situations. Therefore it can be a component of many algorithms, which deal with an appropriated sort of stochastic processes. This article was written to present a basic concept in Markov analysis and optimisation under uncertainty. 0 Stage 1 h Stage 2 2h Stage 3 3h 4h … Time Figure 1 Stages of the process First of all, we present some attributes for processes named in honour of a Russian mathematician Andrei Andreevich Markov. Markov process is a stochastic process with the so-called Markov property. In general there can be discrete-time and continuous-time processes assumed. In practice, the process time can be divided into intervals, which makes them discrete. This time intervals are called stages. Furthermore, let us consider that all stages are of the same duration h , as shown in Figure 1. The process can be found in one of any different states in each period of time. Let us consider that this set of process possible states is finite and also discrete. We say that we have a finite and discrete state space denoted by U = {1, 2, K , u} . We consider that set U includes all the possible states of the system, in other words, the system cannot appear in any other unexpected state, which is leading to some limitations in practice. In this case, the set of states is said to be collectively exhaustive. Furthermore, let us consider the system cannot be found in more than one state in each period. The states are said to be mutually exclusive. a1 Set of possible states of the process a1 1 a2 1 1 1 3 a3 2 3 2 2 a2 a3 Stage n + 1 Stage n Fiure 2 Systém with three states – current and following stage INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 277 The modelling of stochastic process behaviour Stochastic processes must be of the Markov Property, so that Markov Chain might describe these. We say that a stochastic process is of the Markov property if “the future is only depending on the present, not on the past”. In other words, following stage depends only on current stage, not on the “path” which is leading to current state. That means if we have a function, which describes transitions between proper states in the course of a stage and information about current state of the system, we are able to model the system behaviour throughout all the following stages. Thus, the question is: “If we are in some state now, in what state we will be in the following period?” Since system is stochastic, there is no clear answer, of course. We do not know the following state, but we know the probabilities that the system will be in a particular state after the following stage, under condition of the current state of the system in the present stage. This is called conditional state probability. If we denote the state of system in period n as X n and the state of system in the following period n + 1 as X n +1 , then for the probability of being in state j in the period n + 1 , provided that we are in state i in the period n , we write: Pr[X n +1 = j | X n = i ] (1) Thus, the answer to our above-mentioned question, about following stage is not a number of state, but the vector of the conditional state probabilities a = (a1 , a 2 , K, a s ) where s is the number of possible states. This probability vector can be interpreted as all chances of moving from a particular state to all s possible states in the following stage. Since the system cannot appear in any other states and must occupy one and only one of the possible states, the sum of all these probabilities must be equal to one: s åa j =1 (2) j =1 The conditional state probability represented by (1) is the basic information about the system together with the number of possible states. This probability is called the transition probability because it represents the possibility of the transition from one state to another across one stage. Let us present an illustration of a system with three possible states a1 , a 2 , a3 , then Figure 2 illustrates the 1 1 transition probabilities of this system. As written above if we know 2 the actual system stage, we can compute a probability vector that a2 represents the condition of the system in the following period. Let a1 us have the system studied in the first state in the current stage. Therefore, the probability vector, which represents the current 1 stage, is a ( n ) = (1, 0, 0) . (Here, number one represents certainty 2 1 because we know it.) As we can see, in the following stage the 3 a3 system transfers into the second state because probability of this transition is 1, and any other transition must have a zero probability according to Equation 2. The probability vector of following stage 2 3 is a ( n +1) = (0, 1, 0) . As noted in [9], the transition probability concept is the key to the Figure 3 Transition diagram Markov analysis. To clarify the situation, we can show another interpretation of the transition probabilities of To our example, by the transition diagram. In State at the following period Figure 3 we can see a mathematical graph which From a a2 L aj L as 1 represents the system possibilities of the moving p p L p L p 1s ö æ a 12 1j from state to state. It can be interpreted as a 1 ç 11 ÷ fraction of set of elements (for example the a 2 ç p 21 p 22 L p 2 j L p 2 s ÷ number of broken machines) or as a probability State in the M O M M ÷ M çç M ÷=P of one element reaching a particular state. The current period a ç p p i 2 L p ij L p 11 ÷ i1 i most frequently used representation of transition ç ÷ M M O M ÷ M ç M probabilities is the transition matrix. It is a a s çè p s1 p s 2 L p sj L p ss ÷ø natural way to represent a Figure 4 General structure of transition matrix INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 278 model of the Markovian processes in memory of universal computer. It also enables usage of matrix algebra for the Markov analysis. Let us denote this matrix P . The general structure of this matrix can be seen in Figure 4. Each particular element pij represents the transition probability from state i to state j . The rows of this matrix are probability vectors discussed above, and Equation 2 is applicable to each row of matrix P ; therefore we can write: s åp ij =1 (3) j =1 We can get the answer to our question about the probability vector in particular stages if we know probability vector of the initial stage and transition matrix of the Markov chain, by means of matrix algebra. Any probability vector of n -th stage a (n ) , multiplied by matrix P , gives us probability vector of following stage a ( n +1) : a ( n +1) = a ( n ) P (4) For our example the transition matrix can be constructed as follows: æ0 1 0ö ç ÷ P = ç 0 12 12 ÷ ç1 0 2÷ 3ø è3 (5) If we continue in our numerical example, the probability vector of the following stage can be computed: æ0 1 0ö ç ÷ (0 1 0)ç 0 12 12 ÷ = (0 ç1 0 2÷ 3ø è3 1 2 1 2 ) By means of the matrix algebra we obtain the resulting probability vector a ( n + 2 ) = (0, 12 , 12 ) , the correctness of this outcome can be proved by a glance at the graphical representation of the system transition probabilities in Figure 2 or Figure 3. Please note that the same result can be obtained by multiplying the initial state probabilities and the corresponding transition matrix power: a (n) = a (0) P n (6) where a (n ) is the vector of state probabilities in stage n and a ( 0 ) in the initial stage. The proof of Equation 6 is quite simple, and can be found in [8]. For our simple numerical example we get: 2 æ0 1 0ö æ0 ç ÷ ç 2 P = ç 0 12 12 ÷ = ç 16 ç1 0 2÷ ç2 3ø è3 è9 1 2 1 4 1 3 1 2 7 12 4 9 ö ÷ ÷ ÷ ø and then for the second stage: a (2) æ0 ç = (1 0 0)ç 16 ç2 è9 1 2 1 4 1 3 1 2 7 12 4 9 ö ÷ ÷ = (0 ÷ ø 1 2 1 2 ) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 279 Types of the Markov chain a1 a3 P reducible = a2 a4 a1 a2 a3 a4 a1 a 2 a 3 a 4 æ 12 12 0 0 ö ç ÷ ç 13 23 0 0 ÷ ç1 1 1 1÷ ç4 4 4 4÷ ç0 0 1 3 ÷ 4 4 ø è Figure 5 System with an ergodic set As mentioned in the introduction, we deal with the Markov Chains, which have the finite number of states, and we consider only the dependence of the following state on the current one. We call them finite-state first order Markov chains (as opposed to infinite-state and higher order). The dependence of successive stages is stochastic and we describe it by transition probability matrix P . The transition probabilities pij are assumed to be unchanging over time, we call them stationary. The system states can be classified according to their probability of incidence through time. When state i is accessible from state j , then we say they communicate. If i communicates with both j and k , then j communicates with k . Therefore, communication is a property of a certain class of states, and we call them ergodic and such a set of states is called an ergodic set (closed set). (We use term ergodic if every part of the observed sample demonstrates the same statistics parameters as the whole sample.) Whenever the system enters this closed set, this set cannot be left. a3 a1 Pabsorb a4 a1 a1 æ 1 ç a2 ç 0 = ç1 a3 ç 2 a 4 çè 0 a 2 a3 a 4 0 0 0ö ÷ 1 0 0÷ 0 15 103 ÷ ÷ 3 1 1 ÷ 10 2 5 ø a2 Figure 6 System with two absorbing states The example of such a system can be seen in Figure 5. States 1 and 2 communicate one another, but do not communicate with two other states; we call them ergodic states. In the corresponding transition diagram we can verify that states 1 and 2 cannot be escaped once entered. We say that the process has been absorbed in this ergodic set. No matter where the process starts out, it will end in the ergodic set after a2 the finite number of stages. In other words transition probabilities up to states 3 and 4 will approach to zero in high power of transient matrix (to be compared with Equation 6). Thereof states 3 and 4 are called transient (transient set). The processes that can be split out in this way are called reducible. The processes may have more a1 than one ergodic set. Another case is illustrated in Figure 6. This process will be absorbed either in state 1 or 2, which is evident from the transition diagram. States 1 and 2 are called absorbing a3 states, and in the corresponding transition matrix can be distinguished from other states by the ones in the leading diagonal. Another sort of finite Markov Chains is the Figure 7 Cyclic system case where systems always move from state to state. Such a simple system is INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 280 illustrated in Figure 7. These Markov systems are called cyclic. The chain that is not cyclic is called aperiodic. The system that was used as the first example in this article (see Figure 3) is called regular. This system has a set of states which all communicates one another. We say that the system is irreducible (cannot be split out to parts which have different properties). Moreover, this system is aperiodic. Such a Markov chain is called regular and its behaviour is well predictable especially for a long-term planning. Steady State Probabilities It can be observed that the probability vector of a current stage tends to remain constant after a number of transitions. In other words, when we deal with the transition matrix power of a particular stage, we say that matrix P k approaches a unique limiting matrix. The process is said to have reached a steady state. For the regular Markov chain this is always the case. 100% 100% 50% 50% 0% 0% 1 4 7 10 13 a) One half of the machinery is prepared; the second one is in operation at the beginning. 1 4 7 10 13 b) All machines are under maintenance at the beginning. Figure 8 Behaviour of regular system Since the system reaches steady state, the probability distribution of the following stage remains the same: a ( 0 ) P k +1 = a ( k +1) = a ( k ) (7) where a ( k ) = (a1( k ) , a 2( k ) , K , a s( k ) ) is the vector of steady state probabilities. Let us denote this vector as q = (q1 , q 2 , K , q s ) and limiting matrix as Q . In view of the fact that this probability distribution remains constant over all stages when this situation occurred, we can write: Q = QP (8) q = qP . (9) and From the theory of matrix calculus, Equation 9 can be expressed as: a3 a1 a1 a P = 2 a3 a4 a2 a1 a 2 a 3 a 4 7 3 0 æ0 10 10 ö ç ÷ 3 2 0 ç0 5 5 ÷ ç 1 4 0 0 ÷ 5 ç 5 ÷ 2 ç 3 0 0 ÷ø 5 è 5 a4 Figure 9 More complicated cyclic system INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 281 n qj = åp q ij for j = 1, 2, K , s . j (10) i =1 Therefore we can express Equation 9 as a system of s simultaneous linear equations: q1 = p11 q1 + p 21 q 2 + L + p s1 q s q 2 = p12 q1 + p 22 q2 + L + p s 2 q s . M M M M q s = p1s q1 + p 2 s q 2 + L + p ss q s (11) In the system of linear equations (11) one equation is always redundant (i.e. it presents the same information as that given by another equation in a different form) and hence it cannot be solved for unique solution. Therefore we must replace any of these equations by another one. We can expediently use Equation 2 in the following form: s åq j =1 . (12) j =1 If we want to obtain steady state probabilities of our system described by transition matrix (5), the following equations can be used: q1 = 0q1 + 0q 2 + 13 q3 q 2 = 1q1 + 12 q 2 + 0q3 . q3 = 0q1 + 12 q 2 + 23 q3 (13) We can directly substitute the first equation for q1 in the equations for q 2 and q3 . Then we get two equations which both express the same thing, thus one is redundant: q2 = 13 q3 + 12 q2 Þ 12 q2 = 13 q3 q3 = 12 q 2 + 23 q3 Þ 13 q3 = 12 q2 This is a common property of all three equations, and generally, for all systems of equations constructed this way, because of the transition matrix condition (see Equation 3). Therefore we replace one of three equations by the equation q1 + q 2 + q s = 1 and then we obtain the following solution: q1 = 16 , q 2 = 13 , q 3 = 1 2 This system will soon end up in this probability distribution for each individual state in whatever state it starts. Now we can join our example with a practical planning problem: Consider that we have a group of machines. 100% 100% 75% 75% 50% 50% 25% 25% 0% 0% 1 4 7 10 13 16 19 22 25 28 31 1 4 7 10 13 16 19 22 25 28 31 Figure 10 Steady state and cyclic behaviour of the system INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 282 Each machine can be in one of three states in the course of operation q1 = “prepared to operate”, q 2 = “in operation”, q3 = “out of order”. Then if there is a need of 150 machines in production and we know the steady state probabilities, we can easily compute how many machines we need to buy and how much space we need for the machinery maintenance. In our example 150 machines corresponds to one third of the whole set; therefore we need to buy 3 × 150 = 450 machines. From this number we must reserve a place for the expected 450 × 12 = 225 machines in the repair shop. In Figure 8 the expected system behaviour through 13 stages can be seen. In the first graph (8a) we can see a system, consisting of 50 % of machines in operation and 50 % machines prepared at the beginning. In second graph (8b) is the same system, but at the beginning there are all machines under maintenance. However, it can be seen that the system soon will end in stable form. 1.1. More complicated systems Consider the system that is characterised in Figure 9. This system creates one ergodic set and we can obtain the vector of steady state probabilities. (Computed by machine a=[.19642 .30357 .25892 .24107].) However, there can occur situations where the system cannot reach these steady state probabilities depending on the initial state. In Figure 10 are two graphs, showing distributions of the expected probabilities in particular stages through 33 periods of time. In the first graph stable behaviour of this system can be seen. It reaches the equilibrium probability very soon, when initial probability vector is a ( 0 ) = ( 14 , 14 , 14 , 14 ) . The second graph in Figure 10 shows the same system under another initial condition a ( 0 ) = ( 35 , 0, 25 , 0) . From this graph we can see that in this case the system behaviour is cyclic. Therefore this system is not regular. More generally, we can say that systems that are not regular need not reach steady state probabilities even if they exist. First Passage Time In our example of “machines in operation” a question might be of interest for us “ How long will it take to go from state ‘out of order’ to state ‘prepared to operation’? ”, which means from state 3 to 1. In the terminology of Markov analysis it is called first passage time and this means the expected number of transitions required before the state moves from i to j for the first time. Let mij be the number of transitions from state i to j . It is possible to calculate the expected number of transitions (sometimes we say average number of transitions) using the following equation: mij = 1 + åp ik mkj (14) k¹ j thus, in our example, we can compute m31 as follows: m31 = 1 + p32 m21 + p33 m31 m31 = 1 + 0m21 + 23 m31 m31 = 3 Therefore we can say that the machine remains in the repair shop through three stages of process on average. In this case it depends only on transition probability p33 , as we can easily discover from the transition diagram (Figure 3). Generally, there can be much more dependencies, and therefore more computing steps to find out the first passage time. More information can be found in [2] or [10]. CONCLUSION In this paper I tried to demonstrate the some properties of the Markovian analysis. Thanks to these properties it is suitable to use the Markov analysis as a descriptive tool for the stochastic process in different regions of industry. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 283 The paper is focused on the possibilities of utilisation in the management, production and process planning. The above analysis is also a suitable tool of the computer aid especially in the management of various enterprises. Therefore, I recommend a larger utilisation of the software applied in this way. Due to the applicability of the above methods, it will be good to motive Czech software firms to produce such a software that will include them. REFERENCES [1] GRINSTEAD, C. M., SNELL, J. L.: Introduction to probability. GNU version, distributed under GPL, 1997 [2] KEMENY, J. G., SNELL, J. L.: Finite Markov Chains. Springer-Verlag, New York 1976 [3] KEMENY, J. G., SNELL, J. L., GERALD, L. T.: Introduction to Finite Mathematics. New GNU version, distributed under GPL [4] KLAPKA, J., DVOŘÁK, J., POPELA, P.: Metody operačního výzkumu. VUTIUM, Brno 2001 [5] REIBMAN, A., SMITH, R., TRIVEDI, K. S.: Markov and Markov reward model transient analysis: An overview of numerical approaches. European Journal of Operational Research, 40, pp. 257-267, 1989 [6] STEVENSON, W. J.: Introduction to Management Science. Irwin, Boston 1998 [7] STEWART, W. J.: Introduction to the Numerical Solution of Markov Chains. Princeton University Press, Princeton 1995 [8] ŠEVČÍK, V: Využití stochastických procesů markovského typu k řešení ekonomických problémů. Diploma work, VUT Brno 2002 [9] TURBAN, E., MEREDITH, J.: Fundamentals of Management Science. Irwin, Boston 1991 [10] WALTER, J.: Stochastické modely v ekonomii. SNTL/Alfa 1970 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 284 APPLICATIONS OF THE TWO-DIMENSIONAL HELLINGER AND SHANNON QUASI-NORM Petr Jurák, Zdeněk Karpíšek Department of Statistical and Optimization Methods of the Institute of Mathematics at the BUT Faculty of Mechanical Engineering, Technická 2, 616 69 Brno E-mail: [email protected], Tel.: +420-541 142 532 Abstract. This paper is concerned with the solution to a statistical problem of finding the twodimensional simultaneous discrete probability distribution of a discrete random vector by minimizing the Hellinger distance and Shannon pseudo-distance with given constraints. The constraints are chosen in a general form using suitable functions linear in unknown probabilities. An economic example is used to illustrate the theoretical results. Keywords: Hellinger quasi-norm, Shannon quasi-norm, probability estimations, distribution fitting 1. INTRODUCTION Knowing the probability distribution is an important but often neglected part of finding a solution to the majority of statistical problems. Estimates are usually made of such distributions that use previous experience, asymptotic properties of statistics etc. However, problems with multi-modality and the dimension of the probabilities to be found are often encountered. Therefore, methods of searching for the probability distribution in question based on entropy [1,2] or on distribution distances seem to be a reasonable option [3,4]. The paper is concerned with making estimates of simultaneous discrete probability distributions of a random vector ( X , Y ) based on its observed values. Certain constraints are specified on the searched for functions j ijk ( x, y ) , i = 1,..., n1 , j = 1,..., n2 , k = 0,1,..., K . In addition, this probability probabilities using distribution p = ( pij ) , i = 1,..., n1 , j = 1,..., n2 must have a minimum distance from a given probability distribution q = ( qij ) , i = 1,..., n1 , j = 1,..., n2 . Since this is a discrete problem in 2D, we can use a Hellinger distance [3] given as n1 n2 D ( p, q ) = åå i =1 j =1 ( pij - qij ) 2 (1.1) and, next, use a Shannon pseudo-distance [4] defined below n1 n2 S ( p, q ) = åå ( pij ln pij - qij ln qij ) . (1.2) i =1 j =1 The distribution p can then be found either by a non-linear optimization [5] or, directly, deriving a system of non-linear equations. Assume that an observed random vector different pairs of values (x , y ) i j with ( X ,Y ) takes on a finite number of unknown probabilities pij = P ( X = xi , Y = y j ) , i = 1,..., n1 , j = 1,..., n2 , n1 > 1, n2 > 1 . Observing this random vector yields a twodimensional sample ( ( x , y ) ,..., ( x , y ) ) , which, when sorted, results in a sorted sample ( ( x , y ) , f 1 1 n i n f ij is the frequency of the observed pair ( xi , y j ) . The following theorems are proved in [8]. j ij ) / n where 2. HELLINGER QUASI-NORM INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 285 Table 4.2 – Probability estimates using the Hellinger quasi-norm y j \ xi 5 10 15 20 25 30 35 40 45 1 2 3 4 5 6 7 8 9 10 11 12 0.015334 0.014640 0.013992 0.013386 0.012818 0.012286 0.011787 0.011317 0.010875 0.010458 0.010064 0.009693 0.014296 0.013617 0.012986 0.012398 0.011848 0.011335 0.010854 0.010403 0.009980 0.009582 0.009207 0.008854 0.013360 0.012698 0.012085 0.011515 0.010985 0.010490 0.010028 0.009596 0.009191 0.008811 0.008455 0.008119 0.012513 0.011869 0.011275 0.010724 0.010212 0.009736 0.009293 0.008879 0.008492 0.008130 0.007791 0.007472 0.011743 0.011119 0.010543 0.010011 0.009518 0.009061 0.008635 0.008239 0.007870 0.007525 0.007202 0.006900 0.011043 0.010438 0.009881 0.009367 0.008893 0.008453 0.008045 0.007667 0.007314 0.006985 0.006678 0.006390 0.010404 0.009817 0.009279 0.008783 0.008327 0.007905 0.007514 0.007152 0.006815 0.006501 0.006209 0.005936 0.009818 0.009250 0.008730 0.008253 0.007813 0.007408 0.007034 0.006687 0.006365 0.006066 0.005787 0.005528 0.009281 0.008731 0.008229 0.007768 0.007346 0.006957 0.006598 0.006266 0.005958 0.005673 0.005408 0.005160 Table 4.3 – Probability estimates using the Shannon quasi-norm y j \ xi 5 10 15 20 25 30 35 40 45 1 2 3 4 5 6 7 8 9 10 11 12 0.015094 0.014470 0.013872 0.013298 0.012749 0.012222 0.011716 0.011232 0.010768 0.010323 0.009896 0.009487 0.014165 0.013559 0.012979 0.012424 0.011892 0.011384 0.010897 0.010431 0.009985 0.009558 0.009149 0.008758 0.013292 0.012705 0.012143 0.011607 0.011094 0.010603 0.010135 0.009687 0.009259 0.008849 0.008458 0.008084 0.012474 0.011905 0.011362 0.010843 0.010349 0.009876 0.009426 0.008996 0.008585 0.008194 0.007820 0.007463 0.011706 0.011155 0.010630 0.010130 0.009654 0.009199 0.008766 0.008354 0.007961 0.007586 0.007229 0.006889 0.010985 0.010453 0.009946 0.009464 0.009005 0.008569 0.008153 0.007758 0.007382 0.007024 0.006684 0.006360 0.010309 0.009794 0.009306 0.008841 0.008400 0.007981 0.007583 0.007205 0.006845 0.006504 0.006179 0.005871 0.009674 0.009178 0.008707 0.008260 0.007836 0.007434 0.007053 0.006691 0.006347 0.006022 0.005713 0.005419 0.009078 0.008600 0.008146 0.007717 0.007310 0.006924 0.006559 0.006213 0.005886 0.005575 0.005281 0.005003 Probability estimates using the Hellinger quasi-norm pij , fij /n 0,0161 0,0149 0,0136 0,0124 0,0112 0,0099 0,0087 0,0074 0,0059 0,0047 0,0035 0,0022 0,0009 5 10 15 20 25 y 30 35 40 45 12 11 10 9 8 7 6 5 4 3 x 2 1 fij /n pij Fig. 4. 1 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 288 Probability estimates using the Shannon quasi-norm pij , fij /n 0,0161 0,0149 0,0136 0,0124 0,0112 0,0099 0,0087 0,0074 0,0059 0,0047 0,0035 0,0022 0,0009 5 10 15 20 25 y 30 35 40 11 45 12 10 9 8 7 6 5 4 3 2 1 x Fig. 4. 2 Next we perform a chi-square test for both estimated distributions at a significance level of a = 0.05 . For the number of degrees of freedom df = 12(9) - 3 - 1 = 104 , the critical interval for not rejecting the hypothesis that the distributions fit well is á 0 ; 127.689288ñ . The value of the test statistic for the distribution obtained by minimizing the Hellinger quasi-norm is 2 æ f ij ö - pij ÷ n1 n2 ç n ø = 16.74351002 c H2 = n1n2 åå è pij i =1 j =1 and, for the Shannon quasi-norm, it is 2 æ f ij ö - pij ÷ n1 n2 ç n ø = 16.10787788 . c S2 = n1n2 åå è pij i =1 j =1 Thus we do not reject the good fit hypothesis at a significance level of a = 0.05 . However, by adding further constraints, we may calculate the probability values with more precision. 5. CONCLUSION This paper is an extension of the theoretical results of [2,3,4] and their applications. These results may be implemented on a PC either using a non-linear optimization or finding a solution to a system of non-linear equations for Lagrange multipliers. In Example 4.1 a non-linear system was implemented and solved in Maple. However, the numerical solution of the system of non-linear equations appeared to be strongly contingent on the starting values. Example 4.1 also shows that the results achieved very well correspond to the values observed but, to achieve a greater accuracy, further constraints have to be added (on variance for one). Due to the statistical properties and flexibility, the Shannon quasi-norm seems to be more suitable than the Hellinger one. However, both quasi-norms produce very similar estimates. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 289 REFERENCES [1] KARPÍŠEK, Z. Statistical Properties of Discrete Probability Distributions with Maximum Entropy. Folia Fac. Sci. Nat. Univ. Masarykianae Brunensis, Mathematica 9, Brno 2001, p. 21-32, ISBN 80-210-2544-1. [2] KARPÍŠEK, Z., JURÁK, P. Modelling of Probability Distribution with Maximum Entropy. In: MENDEL ´01.7th International Conference on Soft Computing. Brno 2001, p. 232-239, ISBN 80-214-1894X. [3] KARPÍŠEK, Z., JURÁK, P. Estimate of Discrete Probability Distribution by Means of Hellinger Distance. In: MENDEL ´02. 8th International Conference on Soft Computing. Brno, 2002, p. 301-306, ISBN 80-2142135-5. [4] KARPÍŠEK, Z., JURÁK, P. Odhady diskrétních rozdělení pravděpodobnosti pomocí kvazinorem, In: 1. Mezinárodní matematický workshop – Brno 2002, 7 pp, CD-ROM, ISBN 80-86433-16-1. [5] KLAPKA, J., DVOŘÁK, J., POPELA, P. Metody operačního výzkumu. Brno: PC-DIR, 1996. [6] PITMAN, E. J. G. Some Basic Theory for Statistical Inference. New York: John Wiley & Sons, 1978. [7] SHANNON, C., WEAVER, W. The Mathematical Theory of Communications. Urbana, Illinois, 1949. [8] JURÁK, P. Diskrétní rozdělení pravděpodobnosti s maximální entropií. PhD Thesis (to appear). [9] JURÁK P., KARPÍŠEK Z. Hellingerova a Shannonova kvazinorma ve dvourozměrném prostoru. Aplimat 2004, Bratislava (to appear). The paper is part of the work on the CEZ: J22/98: 261100009 research design entitled „Non-traditional methods for studying complex and vague systems“. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 290 THE NEW MODEL OF SYSTEM FUZZY RELIABILITY Pavel Jelínek, Zdeněk Karpíšek Department of Statistical and Optimization Methods of the Institute of Mathematics at the BUT Faculty of Mechanical Engineering, Technická 2, 616 69 Brno E-mail: [email protected], Tel.: +420-541 142 532 Abstract: Using Yager-type fuzzy probability, the paper defines fuzzy reliability generated by one probability measure. The algorithm computing fuzzy reliability of combined system for fuzzy time to failure is presented. The results can be used to determine the reliability of real objects or systems in cases where the observed random data are of a vague numerical type. Keywords: fuzzy possible time of failure, fuzzy reliability generated by probability measure, fuzzy reliability for fuzzy possible time of failure 1. INTRODUCTION The basic notion in the theory of probability is a random event (a subset of the basic space W), which may or may not occur depending on the implementation of a certain set of conditions (of a random experiment). We assume that, for a particular implementation, we can decide whether this event has or has not occurred. However, in practice, this requirement may not be complied with in a simple way. This is the case, for example, if a die is cast on a rugged surface or if we formulate a random event using vague linguistic terms such as: „a small number will be the result“, „the quantity will assume a value approximately equal to x0“ or „a small time to failure“. Such events may be suitably interpreted by fuzzy sets. On the other hand, the probability value itself may be of a vague nature expressed, for example, as „a very probable event“, „a more or less probable event“ or „the probability is approximately 0.9“. These inaccurate values may also be described by a fuzzy sets and fuzzy numbers [2,3]. 2. YAGER-TYPE FUZZY RELIABILITY ( ) Fuzzy number B = ! + , m B ( t ) with non-decreasing membership function m B ( t ) where m B ( 0 ) = 0 is % % % % called a fuzzy possible time of failure. The set of all fuzzy possible times to failure is marked B% . Let P be crisp probability measure defined on universal set ! + and P be fuzzy probability measure % generated by measure P [4,5,6]. Let, further, B = ! + , m B ( t ) be fuzzy possible time of failure and B% be set of % % all fuzzy possible times to failure. Let A * be the set of all generalized fuzzy numbers [4]. The fuzzy function R : B% ® A * where R ( B ) = P ( B ) is called a fuzzy reliability generated by probability measure P. R ( B ) is % % % % % % % called fuzzy reliability for fuzzy possible time of failure B . % Let P be fuzzy probability measure defined on universal set ! + and R be crisp reliability ( ) ( corresponding to measure P. Let, further, R be fuzzy reliability generated by measure P and B = ! + , m B ( t ) % % % be fuzzy possible time of failure. We have m R( B ) ( p ) = % % sup R ( Ba 1 ) = p ) {a a Î [0;1]} 3. ALGORITHM COMPUTING FUZZY RELIABILITY OF SYSTEMS A number of methods are used to calculate the reliability of a combined system of mutually independent elements using their reliabilities. The algorithm that follows is based on a combination of a list INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 291 method and a method of paths. The list method is based on establishing the set of all the possible logical events in the system that, subsequently, serves to calculate the actual reliability of the system using mutually independent random events. A path in the graph of our system is defined as a sequence of arcs connecting different vertices between the input and output nodes of the graph. The path can then be used to determine the resulting system state. To calculate all the paths and their states the following property [1] is used. We assume that a combined reliability system S of n elements has a non-empty simple acyclic digraph with an adjacency matrix A = ( akl )k ,l =1 . Then an element ckl of the matrix C = ( E - A ) - E where E is the n -1 unit matrix expresses all the paths from vertex k to vertex l in a given graph if the arithmetic operation of addition + is replaced by the logical addition Ú and the arithmetic product × by the logical product Ù in the strings ckl of elements auv , u, v = 1,..., n . Moreover, for k ¹ l , we have ckl = Dlk where Dlk is the algebraic complement of the element dlk of the matrix D = E - A . In matrix A, we put akl = 0 exactly if no arc goes from vertex k to vertex l, otherwise akl = Ai where Ai is an element of the system S. This is actually a one-to one mapping of the set of the elements Ai of the system S to the set of arcs akl . If the graph is not simple (with multiple arcs), it can be transformed into a homeomorphic simple graph by, for example, suitably halving arcs. This will make it possible to calculate all the paths from the input vertex k = 1 to the output vertex l = n of the graph for the system S using Dkl for any particular variation of the state of the list elements and, subsequently, to determine the state the system is in. If S denotes directly the state f the system and the state of the given element is substituted for Ai , then the state of the system is S = sgn ( c1n ) = sgn ( Dn1 ) where, in the algebraic complement Dn1 , we put auv = 1 or 0, if the state of the corresponding element is, according to the list, Ai = 1 or 0 respectively. This is because clk expresses, for given states Ai , the number of uninterrupted paths and clk = 0 exactly if no path from vertex k to vertex l exists for given states Ai . Next we assume that all the system elements are mutually independent and the fuzzy possible time of failure B has a continuous membership function m B ( t ) . Now, at an arbitrary (non-negative) time t, we can % % establish the resulting fuzzy reliability of the system by performing the steps of the so-called FJKS - algorithm: 1. Generate the list of all possible states of the elements of the given system in the form of a matrix with 2n rows that are formed by all the n-th class variations ( A1 , K, An ) of the two2. element set { 0,1} with repetition (these are in fact the binary numbers from 0 to 2n-1 ). Using the algebraic complement Dn1 , for each variation, calculate the state of the system S j = sgn ( c1n ) = = sgn ( Dn1 ) , j = 1,..., 2n where, for the elements akl of the adjacency matrix, 3. we substitute the logical value of the state of the corresponding system element in step 1. Denote 2n ì n ü ï é(1 - Ri (t ) )1- Ai ( Ri (t ) ) Ai ù ïý RS ( t ) = S í j ûï ï i =1 ë þ j =1 î å Õ the reliability of system S where Ri ( t ) is the reliability of Ai (we put 00 = 1 ). For the fuzzy possible time of failure B , the resulting fuzzy reliability RS ( B ) of the system has the membership function % % % ì1 for p < RS ( min Ker ( B ) ) , % ïï m RS ( B) ( p ) = ía such that p = RS ( Ba 1 ) for p Î éë RS ( min Ker ( B ) ) , RS ( minSupp ( B ) )ùû , % % % % ï for p > RS ( minSupp ( B ) ) . ïî0 % INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 292 4. EXAMPLE Figure 1 shows the combined system of three mutually independent elements. 2 1 3 Fig. 1: Combined system S The reliabilities of all elements are the same R1 ( t ) = R2 ( t ) = R3 ( t ) = exp ( -t / 60 ) . Fuzzy possible time of failure B (t ) depends on time-parameter t . Its membership function (Fig. 2, 3) is % 3t ì for t < , ï0 4 ï ï 4t - 3t é 3t ù for t Î ê ,t ú , m B(t ) ( t ,t ) = í t % ë4 û ï ï1 for t > t , ï î where the parameter t Î ! to failure. + means end of the time to failure and the value 3t / 4 means the possible end of time Fig. 2: Graph m B (t ) ( t ,t ) % Fig. 3: Graph m B ( 40 ) ( t , 40 ) % Using FJKS - algorithm we get the reliability of the system S RS ( t ) = ( exp ( -t / 60 ) ) (1 - exp ( -t / 60 ) ) + ( exp ( -t / 60 ) ) (1 - exp ( -t / 60 ) ) + ( exp ( -t / 60 ) ) = 2 2 3 = 2 ( exp ( -t / 30 ) ) - ( exp ( -t / 20 )) . Figures 4, 6 shows the graphs of the membership function of fuzzy reliability of an element for the fuzzy possible time of failure B (t ) . Figures 5, 7 shows the graphs of fuzzy reliability of the membership % INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 293 function of the system S for the same fuzzy possible time of failure B (t ) . % Fig. 4: Graph m R ( B (t ) ) ( p,t ) for an element 1 % % Fig. 5: Graph m R ( B (t ) ) ( p,t ) for the system S K % % Fig. 6: Graph m R ( B ( 40 ) ) ( p, 40) for an element 1 % % Fig. 7: Graph m R ( B ( 40) ) ( p, 40 ) for the system S % % 5. CONCLUSION The selection of the membership function for fuzzy possible time of failure is of a subjective nature and can be based either on an expert's opinion or statistical data of a vague nature can be used for expert evaluations. This restricts the applicability and interpretation of the conclusions arrived at. On the other hand, using the fuzzy set theory is of some advantage since it can be used to obtain more plausible even though approximate solutions to actual problems than those achieved using „classical“ methods. We have presented a fuzzy reliability model, which can deal with possible uncertainty precarious information about the observed probability distribution. This fuzzification contrary to Zadeh-type fuzzy reliability [2] assumes only one reliability measure and its concept is nearer the classical theory of probability. We expect implementation of the model on PC for system fuzzy reliability calculation. REFERENCES [1] KARPÍŠEK, Z., JELÍNEK, P. Algoritmus pro výpočet fuzzy spolehlivosti systému. In: Sborník konference 1. Mezinárodní matematický workshop – Brno 2002. FAST VUT, CD ROM, s. 6, 2002, ISBN 80-86433-161. [2] KARPÍŠEK, Z., JELÍNEK, P. Fuzzy stochastické metody modelování spolehlivosti. In: Sborník celostátního INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 294 semináře Analýza dat 2002/II. Lázně Bohdaneč, p. 90 – 103, 2002, ISBN 80-239-0204-0. [3] KARPÍŠEK, Z., SLAVÍŠEK, K. Fuzzy stochastické metody. In: 2nd International Conference APLIMAT 2003 (part I). Bratislava, p. 141 – 148, 2003, ISBN 80-227-1813-0. [4] KARPÍŠEK, Z., SLAVÍČEK, K. Two Fuzzy Probability Measures. In: 3rd Conference of the European Society for Fuzzy Logic and Technology EUSFLAT 2003. Zittau, p. 669 – 674, 2003, ISBN 3-9808089-4-7. [5] SLAVÍČEK, K. Fuzzy pravděpodobnostní míra (Fuzzy Probability Measure). PhD Thesis. FSI VUT, Brno 2002. [6] YAGER, R. R. A Note on Probabilities on Fuzzy Events. In: Information Science (18), Elsevier, p. 113 – 129, 1979. The paper is part of the work on the CEZ: J22/98: 261100009 research design entitled „Non-traditional methods for studying complex and vague systems“. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 295 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 296 ON THE APPLICATION OF INTELLIGENT AUTONOMOUS AGENTS FOR MANAGING THE REAL TIME COMMUNICATION NETWORK Ibrahiem M. M. El Emary Faculty of Applied Science, Al Balqa Applied University, Al Salt, Jordan. E-mail:[email protected] Abstract. This paper is concerned with a practical application of distributed artificial intelligence for managing the real time communication networks. This type of network was chosen because it is needed in interactive applications and all those requiring low data delivery time. A double bus broadcasting architecture is presented after a short description of the three principal network families, and its characteristics are outlined, showing how it is possible to introduce a priority scheme, to reduce data delay. To manage such network well, we present a new tool that work better than current depending on using Intelligent Autonomous Agents which is mainly a software that assists people and acts on their behalf, since the Autonomous Agents help in automating repetitive tasks, remembering rules that users frequently forget and summarizing complex data in understandable human friendly reports or alerts. Keywords : Distributed Artificial Intelligence, Autonomous Agents, Neural Networks, High Priority Minipackets (HPMP), LAN, CSMA/CD, SNMP, KQML, CORBA, SNMP, MLOG, KB, INFERENCE, MIB, NMS and CMIP. 1. INTRODUCTION When the computer networks started to grow in the early years, a management protocol is needed badly, to monitor the LAN faultiest, specialists first made in the first SNMP protocol using the ICMP (Internet Control Message Protocol) messages and the Ping messages were used to test the connection settings between the stations to make sure that there is no cut or any cable down in the network [1]. The main task of the network management system researchers is to develop new tool that work better than current by available tools for doing this laborious work. The work presented in this paper directed about how to delegate as much work as possible to the machine, using network administrators as knowledge engineers that teach the machine how it should perform its work based on what is called Intelligent Autonomous Agent. . There are a few predictable advantages of using Intelligent Autonomous Agents to any management system, network management being just one. First of all, there is the Intelligent Nature, which sounds promising. Having an Intelligent and adaptable system is usually better than having dedicated applications for specific solutions. Likewise, the word Autonomous gives us the idea of something that can work by itself or need almost no human interference. Finally, Agent gives the impression of a helper or a wizard that somehow works between the machine and human. So, the main objective proposed here is to project a system where the human system administration does not have to work as the main workforce available. Instead, we want them to work as the truly intelligent-meaning cognitive-element in the management process, feeding the system with the rules of work – or knowledge – it needs to operate, no more boring work, but intelligent work. . The network that has been suggested to be managed using the above tool is local computer network since these LANs have registered a great development in various fields as in office and industrial automation, so permitting the sharing of expensive resources, satisfying the need for local collection of information, and a speed data exchange for process control. A great proliferation has so arisen of different architectures characterized by simple topologies and simple interfaces with the network, communications are normally very fast. There is however some applications requiring real time communications, that is communications where data delivery time Td is short, compared to evolution time of processes needing to communicate. As such a requirement must be necessarily satisfied if needed by application, it is necessary to make Td as short as possible. . INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 297 Surely if transmission data rate increases, communication channel holding time decreases, so that at a parity of traffic, delay time in the queue to gain the access to the network decreases. This is not however an efficient way to have real time communication. Infact, increasing data transmission rate, requires to be performing in hardware some functions that normally can be software performed. Moreover it is necessary an expensive hardware to work with high bit rate. At last, also communication protocols must be accurately designed to be suitable for the high speed required. With the aim of examining real time communication network under using the Intelligent Autonomous Agents as a management tool to improve its performance, in section 2, three types of local networks are described: bidirectional broadcasting, token passing, and unidirectional broadcasting, examining how data requirements is influence their real time behavior. In section 3, a double bus network architecture is presented, pointing out benefits produced by using two different paths for reservation and data frames, and examining how even with a distributed management, it is possible to achieve a message priority. Section 4 presents the criteria that should be verified in well planed agents. Section 5 describes the architecture of the Intelligent Autonomous Agents. Section 6 presents the implementation of the Autonomous Agents. Section 7 terminates this paper with conclusions and future work. . 2. ARCHITECTURE OF THE WELL KNOWN LAN Three families of Local Area Network (LAN) are currently considered; bidirectional broadcasting networks, ring networks and unidirectional broadcasting networks. In the first, we find Ethernet, net- one, Z-net, etc. Such system uses a common coaxial transmission cable connecting all the communicating devices, by means of a passive interface, so achieving a simple and reliable broadcast communication. The problem of controlling the access to the communication channel has given rise to many different access techniques; the most used is "carrier sense multiple access with collision detection" (CSMA/CD) that is satisfactory when exchanging files, graphics, or other non interactive data, but can too long delay voice (more than the maximum valueTd max allowed by vecoder's rate) when there is a heavy load. A better behavior can be obtained with slotted Ethernet technique, where collision is minimized by means of "0" time slotting mechanism and message transmission is guaranteed within a fixed time [1]. In the second, we find Primenet, Cambridge ring, DCS ring etc. In a ring network message are passed from node to node along an unidirectional cable, by means of an active interface, a token is circulated around the ring and a station has the right to transmit data in the ring only when it has the token. Such access technique, by using an appropriate protocol ensures a maximum message delivery time, so that it can satisfy user requirements when there is a light load; when traffic increases, and voice and data are mixed, delay can however reach a bad value [2]. In the third family, we find Express Net. In a unidirectional broadcasting network, transmission signal are forced to propagate on the cable, in only one direction, by means of special tapes that attenuate the signal in the opposite direction. In this Express network type, there are two interconnected unidirectional channel, implemented by folding a single cable, so that each station can receive messages on a channel and can transmit data on the other one. This system provides advantages of a broadcasting communication and regularity of a token ring communication. This architecture has shown a good behavior with voice and data. All the above families' architectures show a satisfactory behavior in normal conditions, but in a hard environment, they cannot satisfy all the user requirements in terms of delivery time and reliability. It is possible to decrease delay time Td by increasing the transmission bit rate, but this is not an efficient way to use in the communication channel, because in the network there is burst traffic normally, so that, only for a short period of time, a data traffic jam can arise with long data queues. For this reason it is better to distinguish data on the basis of their delivery time requirements, so giving a high priority to those short data that need real time communications, as voice packets, and a lower priority to those data which can be delayed, or that reserve for a long time the communication channel. In this way it is possible to integrate interactive data on every network by means of an adequate protocol. If the network has to be used in distributed process control system- the matter of this paper- it is moreover necessary to take the environmental constraints and the application requirements into consideration. 3. STRUCTURE AND OPERATION CONCEPT OF THE DOUBLE BUS NETWORK A considerable improvement of the throughput, reliability and a low delivery time can be achieved by a double bus network. Reliability is often a very important parameter, many networks provide a redundant cable to INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 298 set going in case of failure of the main one so avoiding to stop the entire network. Such a use of the redundant communication path is however a waste; in fact with a little hardware more it is possible to use both the channels in the same time. In this way we can improve throughput, as there are two independent paths where data can travel, and in case of failure of a bus, communication can continue ( in a reduced way ) on the other one. Moreover, it is easy to implement a protocol that makes provision for reservation, and high priority packets so that urgent message can travel with low delay, at last, it is possible to provide a full duplex transmission ( all the previous mentioned tasks represent the main jobs that is required from the network management technique). . Double bus architecture improves surely the behavior of any type of network, in this section a bidirectional broadcasting local network is considered. The block diagram of simplified controller hardware is shown in Fig.1. From this figure, each station uses two independent buffers for received data, and one for transmitted data, so permitting a full duplex service. There are two receivers permanently connected to each cable and only one transmitter; which can be switched on one of the cables. In this way, both the channels are continuously monitored, where as it is possible to transmit on a channel each time. Bus 1 is used to transmit data a packet, after a successful reservation, Bus 2 is used to transmit reservation mini packets, voice packets, and also data packets. When a user has to send data, it must perform a reservation by sending on Bus 2, a reservation mini packet. If the reservation is successful, all users update their reservation table, by including the one just performed. As reservations are performed with CSMA/CD technique, a collision can occur; in this case, conflicting users must reschedule transmission after an opportune delay, depending on adopted contention resolution protocol. All mini packets must have a length at least equal to 2Td (where Td is the time necessary to connect the most remote users) so that a collision is surely detected if there is one. When a user has finished its reservation, Bus 2 is available for other reservations; with this mechanism, a priority chain on Bus 1 is produced, according to the following rule: the first reserved user is the first to transmit data. A maximum number N of reservations can be performed to avoid a too long reservation queue. Data transmission on Bus 1 is collisionless transmission, as it can be performed on the basis of the reservation order. When a user has become the first in the reservation table it, after detection of end of carrier on Bus 1 (EOC) can begin to transmit its data packets. Using Bus 1, only to perform reservation, it is obviously a restrictive way to employ the communication channel bandwidth; at this point it can be useful to introduce the high priority mini packets (HPMP) concept. Such a mini packet can be a speed data packet (for example a voice packet), or a reservation packet for a long data packet which need a speed delivery. To avoid the possibility of collision between HPMP and normal mini packets, a time slot mechanism can be used, so that, after each transmission on Bus 1, a time slot t=2.5Td is reserved for high priority mini packets; within this time no normal reservation can be performed. Moreover HPMP can be transmitted also when the maximum number N of normal reservations which can be performed, has been reached. Of course collisions can occur between HPMP, but in this case a procedure to solve contention must act. A HPMP reservation is identified by all users and causes a back shift to all normal reservations, whereas the user that has performed the prioritized reservation is moved forward in the reservation table. A further improvement in using the communication channel can be achieved by transmitting data packets on bus 1 when there are long data queues. If the reservation table is full, the first user in the list instead of waiting for bus 2 available sends its data packet on bus 1. No collision is possible with normal mini packets as reservation table is full and no user will try to perform a normal reservation, to avoid collision with HPMP, it will be sufficient to delay such a data transmission for 2.5Td. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 299 Bus 1 Bus 2 Reciever1 Collision Detect. Encoder Output high speed buffer Reciever2 E.O.C E.O.C Decoder Decoder Control E status Input high speed buffer 1 Input high speed buffer 2 Control bus Fig.1 Block diagram of the simplified controller hardware. Single lines represent serial data paths and double lines are bit parallel data paths 4. CRITERIA TO BE VERIFIED IN A WELL PLANNED AGENTS Our objective is to develop a decentralized computer network management platform borrowing existing concepts from distributed artificial intelligence. The guidelines for the system are: high degree of adaptability, module reusability, mobility, self generation and environment plasticity. These guidelines sound simple but actually they hide what we have deep in our mind: "To create a system that can really work by itself, generate a new system for a new environment alone, develop new components for new circumstances, and learn as much as possible by itself asking us only when there is no other way". From these guidelines, we see that there are seven commandments that drove us towards what we could call" measurable "goals. These commands are given by the following: INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 300 Agents must be truly autonomous; this means that any agent must be able to control itself and its own actions. Also, they should be able to work independently from what happened to other agents and have enough freedom to adapt themselves to any possible new environment dynamics. Without this item we will have just a usual system where whatever new happens it demands full system administration attention to properly react to new environmental conditions and setup new operating parameters. Agents should be goal-driven and rule-based; a set of rules constitutes the knowledge base for any administration system-here including human ones. And goals, which are rules by themselves, are the motivations agents have to work. We define static goals as parameters setup during agent creation, and dynamic goals those added to agent behavior derived from environmental changes or incoming messages. The agent environment must behave as a global knowledge base; where each data item or information stored by one agent is sharable throughout the whole system (by other agents). Agents' intercommunication and knowledge exchange is essential to ensure that all needed information would be available for decision making. Besides, in the global system- meaning worldwide over internet- information one agent could learn a new rule from another community and share that within its own community, thus increasing the overall community's knowledge. We intend to reduce human interaction when we have one automatic system learning new tricks from another automatic system more knowledgeable than the first one in some field. Agents should be capable of learning new skills; either from other agents, the community of agents, the global community or, as last resort, human beings. These new rules should be dynamically being stored on agent's local knowledge databases. The other possibility is inferring new rules based on environment data along with current knowledge. Agents should be self-generating; they should adapt themselves to new environments and generate subsets of their own characteristics creating new elements for new problems. This allows an agent to become a selfgeneration system, reducing human interaction as much as possible. Specifically for network management agents; they should interface with commercial solutions using standard protocols; this means that agents should know, through rules in their databases, how to interact with well-known protocols such as SNMP and CMIP (Communication Management Information Protocol) protocols, how to retrieve information from log files, and also how to interact with other existing management systems. Finally, Agents should have a human friendly interface; the friendly interface between human and machines is a natural language-like protocol. In mean time, we should use a more formal specification for the humanagent communication language and provide Application Program Interfaces (APIs) for developing other interfaces- like HTML, NNTP and TELNET. 5. ARCHITECTURE OF THE INTELLIGENT AUTONOMOUS AGENTS In the work presented here, we define Autonomous agents as any software that assists people and acts on their behalf. How they are internally implemented or externally appear is up to the environment where they are inserted and their developers. These discrepancies between implementations can be clearly observed taking on different works about this subject like [3, 4, and 5]. Agents must be truly autonomous; this means that any agent must be able to control itself and its own actions. Also, they should be able to work independently from what happened to other agents and have enough freedom to adapt themselves to any possible new environment dynamics. Without this item we will have just a usual system where whatever new happens it demands full system administration attention to properly react to new environmental conditions and setup new operating parameters. Therefore the main idea is that an autonomous agent is a software application that works by allowing human users to delegate tasks to these applications. Also, these applications should work independently, but on the other hand, should interact with human masters in order to acquire new knowledge and learn new tricks. For network management application, autonomous agents could help in automating repetitive tasks, remembering rules that users frequently forget and summarizing complex data in understandable human friendly reports or alerts. To illustrate the framework for a generic agent; we say that the objective of the generic agent concept is to define the basic structures, mechanisms and basic abilities that guarantee a minimum standard behavior, internal and external, to any agent that is created. In other words, a generic agent is a template that the agency system itself uses to create new agent. Over those templates new functionality can be added to make this new agent being created best fit into its destination environment [6]. The generic agent concept defines the following three modules as shown in Fig.2. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 301 AGENT INFERENCE INTERACTION COMMUNICATION Fig.2 Generic Agent Structure INFERENCE INTERACTION R R cv COMMUNICATION CC INFERENCE KB Fig.3 Detail for the INFERENCE Module *Inference module; which implements the rule deduction engine and also stores the knowledge database as shown in Fig 3. Symbols that denote the various components of the inference module are given by the following: RR: Rule resolution component CV: life cycle component CC: cooperation component KB: knowledge base *Interaction module; which performs the interface between the agent and the environment. *Communication module; which implement the message exchange procedures between agents. Communications are implemented by means of knowledge Query and manipulation language (KQML) [7] which was chosen due to its strong commitment to agent's applications. For extensibility purposes, the generic agent's architecture is being adapted to be compatible with common object request broker architecture (CORBA).We want agents to interface with CORBA services, either making use of distributed object or serving as CORBA repositories sharing internal functionalities. 6. IMPLEMENTATION OF THE AUTONOMOUS AGENTS The kernel part for every agent on the system is the very same defined in the generic agent structure described in the above section. Thus we have only one single source code for every agent. The distinction for each agent is made through the rules loaded on its knowledge base and the native rules loaded on its code. Native rules follow the java definition and are used strictly when the task is so specific that a knowledge rule could not be implemented instead. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 302 The rules language is called Mlog. Mlog is a subset of the prolog language optimized for size and resource consumption reasons. It is possible to submit knowledge to an agent in the IF-THEN-ELSE rule format and the agent will translate it to Mlog format-being helped by Interface agents that have IF-THENELSE to Mlog translation knowledge. A lot of prolog programs will run on Mlog-based agent. There should be Prolog converters to allow any Prolog application to be compatible. Programming in Mlog consists of: (a) Declaring some facts about objects and their relationships for network management, we can see objects as devices. (b) Defining some rules about objects and their relationships. (c) Asking question about object and their relationships. One major difference between Prolog and Mlog is that Mlog handles exception like unknown clauses or invalid values instead of throwing out an error message .If an unknown clause is reached inside the rules resolution rules are not analyzed when added to knowledge base (KB) but during execution-an unknown clause exception is thrown and the cooperation component activated. Thus, the agent will try to learn that clause from the community before deciding to interrupt that execution sequence. The same applies for invalid values and other exceptions. Once Mlog is learned, it becomes straightforward to understand how the inference module is implemented (see Fig 3). The knowledge base (KB) is implemented by a set of Mlog sentences, like in Prolog. The rules resolution component (RR) in the Mlog engine and is based on a full but reduced Prolog engine that process the (KB) rules based on an expression being executed. The cooperation component (CC) is a new feature, not usual to Prolog system. It implements learning capabilities and is used when either a new knowledge is required to accomplish a rule resolution, or a new knowledge is offered for the agent from the community. In other words, it handles how to add to or remove knowledge from internal knowledge bases and how to ask for new knowledge and later add it from the community. A flowchart for CC Vs. RR co-existence is presented in Fig 3. Execute rule(par1,par2,…) Variables initializing Knowledge base Load information about the rule being executed No Known rule? Learn from community Yes Analyze parameters Yes Is native ? Execute native rule(par1,par2,..) No Mlog rule resolution Fig.3 Basic Rule Resolution Flowchart INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 303 The life cycle component (CV) is an internal Mlog implemented component that complies the set of goals and has an agent to pursuer these in a certain amount of time. After that, it submits these goals, one by one, to execution in the rules resolution component. The interaction module is not a processing module per se but a set of rules and native methods declared inside an agent to interface with environmental variables. In network management application, SNMP interface clauses are implemented in this module. The goal of creating an independent module for those rules and native rules is to better understand the system functionality. From the compilation point of view, the independent modules are linked inside the code depending on the functions that will be exercised by agent being built. Finally, the communication module is much like the interaction module: a set of rules and native methods targeting communication procedures. Things like TCP/IP package exchanging KQML protocol formatting communication session control, and other related to information exchange are treated in this module. 7. CONCLUDING COMMENTS AND FURTHER DEVELOPMENT As we have seen in this paper, there are quite a few predictable advantages of using intelligent autonomous agents for managing a computer network. First of all, there is the intelligent nature which sounds promising. Having an intelligent and adaptable system is usually better than having dedicated applications for specific solution. It was also shown that applying a distributed artificial intelligence offer use a significant advantages over large monolithic system among it the following: System modularity; it is easier to build and maintain a collection of quasi independent modules than one huge one. Efficiency; not all knowledge is needed for all tasks. By modularizing it, we gain the ability to focus intelligent autonomous agent efforts in ways that are most likely to pay off. A communication structure that enables information to be passed back and forth among agents. As a future work to be done by others, we recommend with the following points: * To create a series of new agents modules and module plug-ins, in order to make them available to an increasing number of new working environments. * To create a global wizard structured that is a set of Internet available agents to whom local agents will refer in case they need locally unknown new information. In these structures we will store as much information as possible for all different fields where agents will be applied. REFERENCES [1] TANENBAUM, A.S.: Computer Networks, Forth Edition, Prentice Hall PTR, Upper Saddle River, NJ 07458, 2003 [2] KALYANAMAN, S.: Simple Network Management Protocol, Resselaer Polytechnic Institute, 2000 [3] KAUTZ, H.A.: Botton-up Design of Software Agents, Communications of The ACM, Vol.37, No.7, U.S.A, 1994 [4] MAES, P.: Agents that Reduce Work and Information Overload, Communications of the ACM, vol.37, NO.7, U.S.A, 1994 [5] FERBER, J.: Multi- Agent System an Introduction to Distributed Artificial Intelligent, pp.4-8, Ed. Addison – wesley, France .1999 [6] ICHOCKIL, A.: Architectures and Electronic Implementation of Neural Network Models – Neural Network for Optimization and Signal Processing, John Willey & sons, U.K, 1993 [7] KAHANI, M., BEADLE, P.H.V.: Decentralized Approaches for Network Management, Computer Communications Review, ACM-SIGCOMM, Vol.27, NO.3, pp.36-47, U.S.A, 1997 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 304 ON THE APPLICATION OF INTELLIGENT AUTONOMOUS AGENTS FOR MANAGING THE REAL TIME COMMUNICATION NETWORK Ibrahiem M. M. El Emary Faculty of Applied Science, Al Balqa Applied University, Al Salt, Jordan. E-mail:[email protected] Abstract. This paper is concerned with a practical application of distributed artificial intelligence for managing the real time communication networks. This type of network was chosen because it is needed in interactive applications and all those requiring low data delivery time. A double bus broadcasting architecture is presented after a short description of the three principal network families, and its characteristics are outlined, showing how it is possible to introduce a priority scheme, to reduce data delay. To manage such network well, we present a new tool that work better than current depending on using Intelligent Autonomous Agents which is mainly a software that assists people and acts on their behalf, since the Autonomous Agents help in automating repetitive tasks, remembering rules that users frequently forget and summarizing complex data in understandable human friendly reports or alerts. Keywords : Distributed Artificial Intelligence, Autonomous Agents, Neural Networks, High Priority Minipackets (HPMP), LAN, CSMA/CD, SNMP, KQML, CORBA, SNMP, MLOG, KB, INFERENCE, MIB, NMS and CMIP. 1. INTRODUCTION When the computer networks started to grow in the early years, a management protocol is needed badly, to monitor the LAN faultiest, specialists first made in the first SNMP protocol using the ICMP (Internet Control Message Protocol) messages and the Ping messages were used to test the connection settings between the stations to make sure that there is no cut or any cable down in the network [1]. The main task of the network management system researchers is to develop new tool that work better than current by available tools for doing this laborious work. The work presented in this paper directed about how to delegate as much work as possible to the machine, using network administrators as knowledge engineers that teach the machine how it should perform its work based on what is called Intelligent Autonomous Agent. . There are a few predictable advantages of using Intelligent Autonomous Agents to any management system, network management being just one. First of all, there is the Intelligent Nature, which sounds promising. Having an Intelligent and adaptable system is usually better than having dedicated applications for specific solutions. Likewise, the word Autonomous gives us the idea of something that can work by itself or need almost no human interference. Finally, Agent gives the impression of a helper or a wizard that somehow works between the machine and human. So, the main objective proposed here is to project a system where the human system administration does not have to work as the main workforce available. Instead, we want them to work as the truly intelligent-meaning cognitive-element in the management process, feeding the system with the rules of work – or knowledge – it needs to operate, no more boring work, but intelligent work. . The network that has been suggested to be managed using the above tool is local computer network since these LANs have registered a great development in various fields as in office and industrial automation, so permitting the sharing of expensive resources, satisfying the need for local collection of information, and a speed data exchange for process control. A great proliferation has so arisen of different architectures characterized by simple topologies and simple interfaces with the network, communications are normally very fast. There is however some applications requiring real time communications, that is communications where data delivery time Td is short, compared to evolution time of processes needing to communicate. As such a requirement must be necessarily satisfied if needed by application, it is necessary to make Td as short as possible. . INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 297 Surely if transmission data rate increases, communication channel holding time decreases, so that at a parity of traffic, delay time in the queue to gain the access to the network decreases. This is not however an efficient way to have real time communication. Infact, increasing data transmission rate, requires to be performing in hardware some functions that normally can be software performed. Moreover it is necessary an expensive hardware to work with high bit rate. At last, also communication protocols must be accurately designed to be suitable for the high speed required. With the aim of examining real time communication network under using the Intelligent Autonomous Agents as a management tool to improve its performance, in section 2, three types of local networks are described: bidirectional broadcasting, token passing, and unidirectional broadcasting, examining how data requirements is influence their real time behavior. In section 3, a double bus network architecture is presented, pointing out benefits produced by using two different paths for reservation and data frames, and examining how even with a distributed management, it is possible to achieve a message priority. Section 4 presents the criteria that should be verified in well planed agents. Section 5 describes the architecture of the Intelligent Autonomous Agents. Section 6 presents the implementation of the Autonomous Agents. Section 7 terminates this paper with conclusions and future work. . 2. ARCHITECTURE OF THE WELL KNOWN LAN Three families of Local Area Network (LAN) are currently considered; bidirectional broadcasting networks, ring networks and unidirectional broadcasting networks. In the first, we find Ethernet, net- one, Z-net, etc. Such system uses a common coaxial transmission cable connecting all the communicating devices, by means of a passive interface, so achieving a simple and reliable broadcast communication. The problem of controlling the access to the communication channel has given rise to many different access techniques; the most used is "carrier sense multiple access with collision detection" (CSMA/CD) that is satisfactory when exchanging files, graphics, or other non interactive data, but can too long delay voice (more than the maximum valueTd max allowed by vecoder's rate) when there is a heavy load. A better behavior can be obtained with slotted Ethernet technique, where collision is minimized by means of "0" time slotting mechanism and message transmission is guaranteed within a fixed time [1]. In the second, we find Primenet, Cambridge ring, DCS ring etc. In a ring network message are passed from node to node along an unidirectional cable, by means of an active interface, a token is circulated around the ring and a station has the right to transmit data in the ring only when it has the token. Such access technique, by using an appropriate protocol ensures a maximum message delivery time, so that it can satisfy user requirements when there is a light load; when traffic increases, and voice and data are mixed, delay can however reach a bad value [2]. In the third family, we find Express Net. In a unidirectional broadcasting network, transmission signal are forced to propagate on the cable, in only one direction, by means of special tapes that attenuate the signal in the opposite direction. In this Express network type, there are two interconnected unidirectional channel, implemented by folding a single cable, so that each station can receive messages on a channel and can transmit data on the other one. This system provides advantages of a broadcasting communication and regularity of a token ring communication. This architecture has shown a good behavior with voice and data. All the above families' architectures show a satisfactory behavior in normal conditions, but in a hard environment, they cannot satisfy all the user requirements in terms of delivery time and reliability. It is possible to decrease delay time Td by increasing the transmission bit rate, but this is not an efficient way to use in the communication channel, because in the network there is burst traffic normally, so that, only for a short period of time, a data traffic jam can arise with long data queues. For this reason it is better to distinguish data on the basis of their delivery time requirements, so giving a high priority to those short data that need real time communications, as voice packets, and a lower priority to those data which can be delayed, or that reserve for a long time the communication channel. In this way it is possible to integrate interactive data on every network by means of an adequate protocol. If the network has to be used in distributed process control system- the matter of this paper- it is moreover necessary to take the environmental constraints and the application requirements into consideration. 3. STRUCTURE AND OPERATION CONCEPT OF THE DOUBLE BUS NETWORK A considerable improvement of the throughput, reliability and a low delivery time can be achieved by a double bus network. Reliability is often a very important parameter, many networks provide a redundant cable to INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 298 set going in case of failure of the main one so avoiding to stop the entire network. Such a use of the redundant communication path is however a waste; in fact with a little hardware more it is possible to use both the channels in the same time. In this way we can improve throughput, as there are two independent paths where data can travel, and in case of failure of a bus, communication can continue ( in a reduced way ) on the other one. Moreover, it is easy to implement a protocol that makes provision for reservation, and high priority packets so that urgent message can travel with low delay, at last, it is possible to provide a full duplex transmission ( all the previous mentioned tasks represent the main jobs that is required from the network management technique). . Double bus architecture improves surely the behavior of any type of network, in this section a bidirectional broadcasting local network is considered. The block diagram of simplified controller hardware is shown in Fig.1. From this figure, each station uses two independent buffers for received data, and one for transmitted data, so permitting a full duplex service. There are two receivers permanently connected to each cable and only one transmitter; which can be switched on one of the cables. In this way, both the channels are continuously monitored, where as it is possible to transmit on a channel each time. Bus 1 is used to transmit data a packet, after a successful reservation, Bus 2 is used to transmit reservation mini packets, voice packets, and also data packets. When a user has to send data, it must perform a reservation by sending on Bus 2, a reservation mini packet. If the reservation is successful, all users update their reservation table, by including the one just performed. As reservations are performed with CSMA/CD technique, a collision can occur; in this case, conflicting users must reschedule transmission after an opportune delay, depending on adopted contention resolution protocol. All mini packets must have a length at least equal to 2Td (where Td is the time necessary to connect the most remote users) so that a collision is surely detected if there is one. When a user has finished its reservation, Bus 2 is available for other reservations; with this mechanism, a priority chain on Bus 1 is produced, according to the following rule: the first reserved user is the first to transmit data. A maximum number N of reservations can be performed to avoid a too long reservation queue. Data transmission on Bus 1 is collisionless transmission, as it can be performed on the basis of the reservation order. When a user has become the first in the reservation table it, after detection of end of carrier on Bus 1 (EOC) can begin to transmit its data packets. Using Bus 1, only to perform reservation, it is obviously a restrictive way to employ the communication channel bandwidth; at this point it can be useful to introduce the high priority mini packets (HPMP) concept. Such a mini packet can be a speed data packet (for example a voice packet), or a reservation packet for a long data packet which need a speed delivery. To avoid the possibility of collision between HPMP and normal mini packets, a time slot mechanism can be used, so that, after each transmission on Bus 1, a time slot t=2.5Td is reserved for high priority mini packets; within this time no normal reservation can be performed. Moreover HPMP can be transmitted also when the maximum number N of normal reservations which can be performed, has been reached. Of course collisions can occur between HPMP, but in this case a procedure to solve contention must act. A HPMP reservation is identified by all users and causes a back shift to all normal reservations, whereas the user that has performed the prioritized reservation is moved forward in the reservation table. A further improvement in using the communication channel can be achieved by transmitting data packets on bus 1 when there are long data queues. If the reservation table is full, the first user in the list instead of waiting for bus 2 available sends its data packet on bus 1. No collision is possible with normal mini packets as reservation table is full and no user will try to perform a normal reservation, to avoid collision with HPMP, it will be sufficient to delay such a data transmission for 2.5Td. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 299 Bus 1 Bus 2 Reciever1 Collision Detect. Encoder Output high speed buffer Reciever2 E.O.C E.O.C Decoder Decoder Control E status Input high speed buffer 1 Input high speed buffer 2 Control bus Fig.1 Block diagram of the simplified controller hardware. Single lines represent serial data paths and double lines are bit parallel data paths 4. CRITERIA TO BE VERIFIED IN A WELL PLANNED AGENTS Our objective is to develop a decentralized computer network management platform borrowing existing concepts from distributed artificial intelligence. The guidelines for the system are: high degree of adaptability, module reusability, mobility, self generation and environment plasticity. These guidelines sound simple but actually they hide what we have deep in our mind: "To create a system that can really work by itself, generate a new system for a new environment alone, develop new components for new circumstances, and learn as much as possible by itself asking us only when there is no other way". From these guidelines, we see that there are seven commandments that drove us towards what we could call" measurable "goals. These commands are given by the following: Agents must be truly autonomous; this means that any agent must be able to control itself and its own actions. Also, they should be able to work independently from what happened to other agents and have enough INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 300 freedom to adapt themselves to any possible new environment dynamics. Without this item we will have just a usual system where whatever new happens it demands full system administration attention to properly react to new environmental conditions and setup new operating parameters. Agents should be goal-driven and rule-based; a set of rules constitutes the knowledge base for any administration system-here including human ones. And goals, which are rules by themselves, are the motivations agents have to work. We define static goals as parameters setup during agent creation, and dynamic goals those added to agent behavior derived from environmental changes or incoming messages. The agent environment must behave as a global knowledge base; where each data item or information stored by one agent is sharable throughout the whole system (by other agents). Agents' intercommunication and knowledge exchange is essential to ensure that all needed information would be available for decision making. Besides, in the global system- meaning worldwide over internet- information one agent could learn a new rule from another community and share that within its own community, thus increasing the overall community's knowledge. We intend to reduce human interaction when we have one automatic system learning new tricks from another automatic system more knowledgeable than the first one in some field. Agents should be capable of learning new skills; either from other agents, the community of agents, the global community or, as last resort, human beings. These new rules should be dynamically being stored on agent's local knowledge databases. The other possibility is inferring new rules based on environment data along with current knowledge. Agents should be self-generating; they should adapt themselves to new environments and generate subsets of their own characteristics creating new elements for new problems. This allows an agent to become a selfgeneration system, reducing human interaction as much as possible. Specifically for network management agents; they should interface with commercial solutions using standard protocols; this means that agents should know, through rules in their databases, how to interact with well-known protocols such as SNMP and CMIP (Communication Management Information Protocol) protocols, how to retrieve information from log files, and also how to interact with other existing management systems. Finally, Agents should have a human friendly interface; the friendly interface between human and machines is a natural language-like protocol. In mean time, we should use a more formal specification for the humanagent communication language and provide Application Program Interfaces (APIs) for developing other interfaces- like HTML, NNTP and TELNET. 5. ARCHITECTURE OF THE INTELLIGENT AUTONOMOUS AGENTS In the work presented here, we define Autonomous agents as any software that assists people and acts on their behalf. How they are internally implemented or externally appear is up to the environment where they are inserted and their developers. These discrepancies between implementations can be clearly observed taking on different works about this subject like [3, 4, and 5]. Agents must be truly autonomous; this means that any agent must be able to control itself and its own actions. Also, they should be able to work independently from what happened to other agents and have enough freedom to adapt themselves to any possible new environment dynamics. Without this item we will have just a usual system where whatever new happens it demands full system administration attention to properly react to new environmental conditions and setup new operating parameters. Therefore the main idea is that an autonomous agent is a software application that works by allowing human users to delegate tasks to these applications. Also, these applications should work independently, but on the other hand, should interact with human masters in order to acquire new knowledge and learn new tricks. For network management application, autonomous agents could help in automating repetitive tasks, remembering rules that users frequently forget and summarizing complex data in understandable human friendly reports or alerts. To illustrate the framework for a generic agent; we say that the objective of the generic agent concept is to define the basic structures, mechanisms and basic abilities that guarantee a minimum standard behavior, internal and external, to any agent that is created. In other words, a generic agent is a template that the agency system itself uses to create new agent. Over those templates new functionality can be added to make this new agent being created best fit into its destination environment [6]. The generic agent concept defines the following three modules as shown in Fig.2. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 301 AGENT INFERENCE INTERACTION COMMUNICATION Fig.2 Generic Agent Structure INFERENCE INTERACTION R R cv COMMUNICATION CC INFERENCE KB Fig.3 Detail for the INFERENCE Module *Inference module; which implements the rule deduction engine and also stores the knowledge database as shown in Fig 3. Symbols that denote the various components of the inference module are given by the following: RR: Rule resolution component CV: life cycle component CC: cooperation component KB: knowledge base *Interaction module; which performs the interface between the agent and the environment. *Communication module; which implement the message exchange procedures between agents. Communications are implemented by means of knowledge Query and manipulation language (KQML) [7] which was chosen due to its strong commitment to agent's applications. For extensibility purposes, the generic agent's architecture is being adapted to be compatible with common object request broker architecture (CORBA).We want agents to interface with CORBA services, either making use of distributed object or serving as CORBA repositories sharing internal functionalities. 6. IMPLEMENTATION OF THE AUTONOMOUS AGENTS The kernel part for every agent on the system is the very same defined in the generic agent structure described in the above section. Thus we have only one single source code for every agent. The distinction for each agent is made through the rules loaded on its knowledge base and the native rules loaded on its code. Native rules follow the java definition and are used strictly when the task is so specific that a knowledge rule could not be implemented instead. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 302 The rules language is called Mlog. Mlog is a subset of the prolog language optimized for size and resource consumption reasons. It is possible to submit knowledge to an agent in the IF-THEN-ELSE rule format and the agent will translate it to Mlog format-being helped by Interface agents that have IF-THENELSE to Mlog translation knowledge. A lot of prolog programs will run on Mlog-based agent. There should be Prolog converters to allow any Prolog application to be compatible. Programming in Mlog consists of: (a) Declaring some facts about objects and their relationships for network management, we can see objects as devices. (b) Defining some rules about objects and their relationships. (c) Asking question about object and their relationships. One major difference between Prolog and Mlog is that Mlog handles exception like unknown clauses or invalid values instead of throwing out an error message .If an unknown clause is reached inside the rules resolution rules are not analyzed when added to knowledge base (KB) but during execution-an unknown clause exception is thrown and the cooperation component activated. Thus, the agent will try to learn that clause from the community before deciding to interrupt that execution sequence. The same applies for invalid values and other exceptions. Once Mlog is learned, it becomes straightforward to understand how the inference module is implemented (see Fig 3). The knowledge base (KB) is implemented by a set of Mlog sentences, like in Prolog. The rules resolution component (RR) in the Mlog engine and is based on a full but reduced Prolog engine that process the (KB) rules based on an expression being executed. The cooperation component (CC) is a new feature, not usual to Prolog system. It implements learning capabilities and is used when either a new knowledge is required to accomplish a rule resolution, or a new knowledge is offered for the agent from the community. In other words, it handles how to add to or remove knowledge from internal knowledge bases and how to ask for new knowledge and later add it from the community. A flowchart for CC Vs. RR co-existence is presented in Fig 3. Execute rule(par1,par2,…) Variables initializing Knowledge base Load information about the rule being executed No Known rule? Learn from community Yes Analyze parameters Yes Is native ? Execute native rule(par1,par2,..) No Mlog rule resolution Fig.3 Basic Rule Resolution Flowchart INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 303 The life cycle component (CV) is an internal Mlog implemented component that complies the set of goals and has an agent to pursuer these in a certain amount of time. After that, it submits these goals, one by one, to execution in the rules resolution component. The interaction module is not a processing module per se but a set of rules and native methods declared inside an agent to interface with environmental variables. In network management application, SNMP interface clauses are implemented in this module. The goal of creating an independent module for those rules and native rules is to better understand the system functionality. From the compilation point of view, the independent modules are linked inside the code depending on the functions that will be exercised by agent being built. Finally, the communication module is much like the interaction module: a set of rules and native methods targeting communication procedures. Things like TCP/IP package exchanging KQML protocol formatting communication session control, and other related to information exchange are treated in this module. 7. CONCLUDING COMMENTS AND FURTHER DEVELOPMENT As we have seen in this paper, there are quite a few predictable advantages of using intelligent autonomous agents for managing a computer network. First of all, there is the intelligent nature which sounds promising. Having an intelligent and adaptable system is usually better than having dedicated applications for specific solution. It was also shown that applying a distributed artificial intelligence offer use a significant advantages over large monolithic system among it the following: System modularity; it is easier to build and maintain a collection of quasi independent modules than one huge one. Efficiency; not all knowledge is needed for all tasks. By modularizing it, we gain the ability to focus intelligent autonomous agent efforts in ways that are most likely to pay off. A communication structure that enables information to be passed back and forth among agents. As a future work to be done by others, we recommend with the following points: * To create a series of new agents modules and module plug-ins, in order to make them available to an increasing number of new working environments. * To create a global wizard structured that is a set of Internet available agents to whom local agents will refer in case they need locally unknown new information. In these structures we will store as much information as possible for all different fields where agents will be applied. REFERENCES [1] TANENBAUM, A.S.: Computer Networks, Forth Edition, Prentice Hall PTR, Upper Saddle River, NJ 07458, 2003 [2] KALYANAMAN, S.: Simple Network Management Protocol, Resselaer Polytechnic Institute, 2000 [3] KAUTZ, H.A.: Botton-up Design of Software Agents, Communications of The ACM, Vol.37, No.7, U.S.A, 1994 [4] MAES, P.: Agents that Reduce Work and Information Overload, Communications of the ACM, vol.37, NO.7, U.S.A, 1994 [5] FERBER, J.: Multi- Agent System an Introduction to Distributed Artificial Intelligent, pp.4-8, Ed. Addison – wesley, France .1999 [6] ICHOCKIL, A.: Architectures and Electronic Implementation of Neural Network Models – Neural Network for Optimization and Signal Processing, John Willey & sons, U.K, 1993 [7] KAHANI, M., BEADLE, P.H.V.: Decentralized Approaches for Network Management, Computer Communications Review, ACM-SIGCOMM, Vol.27, NO.3, pp.36-47, U.S.A, 1997 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 304 HOLOGRAPHIC REDUCED REPRESENTATION IN ARTIFICIAL INTELLIGENCE AND COGNITIVE SCIENCE Vladimír Kvasnička, Jiří Pospíchal Slovak Technical University, 812 37 Bratislava, Slovakia I INTRODUCTION A modern view of the relation between brain and mind is based on the neuroscience paradigm [3], according to which the architecture of the brain is determined by connections between neurons, their inhibitory or excitatory character and also by the strength of the connections. Human brain displays a great plasticity, synapses are perpetually formed (but also deleted) during a learning process. It can be stated, that an ability of brain to perform not only cognitive activities, but also to serve as memory and control center for our motoric activities, is fully encoded by its architecture. The metaphor of a human brain as a computer should be therefore formulated in such a way, that a computer is a parallel distributed computer (containing many billions of neurons, elementary processors interconnected into a complex neural network). A program in such a parallel computer is directly encoded in the architecture of the neural network, i.e. human brain is a single-purpose parallel computer represented by a neural network, which can not be reprogrammed without a change of its architecture. It follows from the above general statement that the mind with the brain creates one integral unit, which is characterized by a complementary dualism. The mind is in this approach understood as a program carried out by the brain, while this program is specified by architecture of the distributed neural network representing the brain. The brain and the mind are two different aspects of the same object: · When talking about the brain, we have in mind a „hardware“ structure, biologically determined by neurons and their synaptic connections (formally represented by a neural network), on the other hand. · When talking about the mind, we have in mind cognitive and other similar activities of the brain, which are carried out on a symbolic level, where the transformation of symbolic information is processed on the basis of (simple) rules. A complementary dualism between brain and mind causes certain difficulties in the interpretation of cognitive activities of mind. A purely neural approach to the interpretation of cognitive activities of mind focuses on the search of neural correlates of neural activities and cognitive activities (connectionism). The application of the neural paradigm for the interpretation of symbolic cognitive activities has a „side effect“ in „dissolving“ of these activities in their microscopic description, symbols quasi „disappear“ in the detailed description of activities of neurons, strengths of synaptic connections etc. On the other side, the absolute acceptance of symbolic paradigm in interpretation of cognitive activities of mind (cognitivism), ignoring of the fact, that mind is thoroughly embedded in brain, leads to a conceptual sterility, to an effort to explain cognitive activities of human mind only in the phenomenological terms derived from the concept of symbol. It leads to symbolic constructs (methods, algorithms etc.), for which there usually does not exist any experimental support in neuroscience. The goal of this paper is to highlight an alternative approach, which may overcome the gap between the connectionist and cognitivistic approach in the description and interpretation of cognitive activities of the human brain [10-13]. We shall show, that the application of a distributed representation allows to integrate connectionism and cognitivism, where mental representations (symbols) are specified by distributed patterns of neural activities, while over these distributed patterns we can introduce formal algebraic operations, which not only allow to mathematically model cognitive operations, but also allow to simulate processes of storage and retrieval of information from memory. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 305 .... distributed representation neural network Figure 1. A visualization of the transition from neural network to distributed representation. The state of neural network in the time t is given by the activities of single neurons, which are determined by the activities in the previous time t-1 and by weight coefficients of single oriented connections. Using a certain abstraction, these activities can be ordered into one big one-dimensional array (vector) of real numbers (their size is determined by the level of gray of the corresponding component – neuron). In the distributed representation the architecture of the neural network is ignored, i.e. two distributed representations must be understood as totally independent without mutual relations, their incidental connections derived from neural network are completely ignored. New unary and binary operations are introduced in the distributed representation, which enable to create new distributed representations from the original ones. We shall turn our attention to a nontraditional style of performing calculation by using distributed patterns. This approach is substantially different from classical numeric and symbolic computations and it is a suitable model tool for understanding of global properties of neural networks. We shall show, that such a „neurocomputing“ is based on extensive randomly created patterns (represented by multidimensional vectors with random entries), see fig. 1. This approach, which basic principles were formulated already at the end of sixties [2,4,5,9,14], was crowned by a series of works by Tony Plate [7-9] about „holographic reduced representation“ (HRR). We shall show, which types of computation can be implemented in this approach and whether they help us to understand the processes in the brain during cognitive activities. Our addition to the development of HRR consists in its application to modeling of cognitive processes of reasoning by application of rules modus ponens a modus tollens. Kanerva [16-18] in the middle of nineties proposed a certain alternative to HRR, which is based on randomly generated binary vectors. 2. A MATHEMATICAL FORMULATION OF HOLOGRAPHIC REPRESENTATION The aim of this chapter is a presentation of basic properties of a holographic representation, which was developed by A. Plate [7-9]. Its basic notion is a conceptual vector, which is represented by an n-dimensional vector a Î R n Þ a = ( a0 ,a2 ,...,an-1 ) (1) where its components are random numbers with a standard normal distribution ai = N ( 0 ,1 n ) "Î {0 ,1,...,n - 1} (2) where N(0,1/n) is a random number with a mean equal to 0 and a standard deviation 1/n. Over conceptual vectors there is defined a binary operation „convolution“, which assigns to a couple of vectors a third vector, Ä : R n ´ R n ® R n , or c = aÄb The components of the resulting vector c = ( c0 ,c1 ,...,cn -1 ) are determined by a formula n -1 ci = å a j b[i - j ] ( i = 0 ,1,...,n - 1) (3) (4) j =0 where the index in the square brackets, [k], is defined using a modulo n operation as follows1 1 A standard definition of an arithmetic operation k modulo n is determined as a remainder after integer division by a number n. It is necessary to comment, that the used definition of the operation k modulo n is different from INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 306 k ¢ = k mod n ìk ¢ [ k ] = ïí ïîn + k ¢ ( if ( if (5a) k ¢ ³ 0) (5b) k ¢ < 0) Since it is one of the basic notions of the holographic representation, we shall give a table of values [k] for –4£k£4 and n=4. k -4 -3 -2 -1 0 1 2 3 4 k’ 0 -3 -2 -1 0 1 2 3 0 [k] 0 1 2 3 0 1 2 3 0 b a 0 1 2 0 00 01 02 1 10 11 12 2 20 21 22 c2 c0 c1 Figure 2. Convolution of two vectors a and b for n=3 (see (6)). A convolution of two vectors a and b for n=3 has the following form (see fig. 2) c0 = a0b0 + a1b2 + a2b1 c1 = a0b1 + a1b0 + a2b2 (6) c2 = a0b2 + a1b1 + a2b0 The convolution satisfies the following properties: (1) commutativity, a Ä b = b Ä a (2) associativity, ( a Ä b ) Ä c = a Ä ( b Ä c ) (3) distributiveness, a Ä ( ab + b c ) = a ( a Ä b ) + b ( a Ä c ) (4) an existence of a unit vector, 1 Ä a = a (1 = (1,0,...,0 ) ) The convolution can be also expressed by a circulant matrix [1] this standard definition for negative numbers k. While the standard definition provides a result with a negative value, if the result is negative in our definition, than it is transformed by adding n to it. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 307 c0 = a0b0 + a1b2 + a2b1 ü æ c0 ö æ a0 a2 a1 öæ b0 ö ï ÷ c1 = a0b1 + a1b0 + a2 b2 ý Þ çç c1 ÷÷ = çç a1 a0 a2 ÷ç ÷ç b1 ÷ Þ a Ä b = circ ( a ) b b ÷ c2 = a0b2 + a1b1 + a2b0 ïþ çè c2 ÷ø çè a2 a1 a0 ÷ç 144244 3øè 2 ø (7) circ ( a ) This specific example is generalized for an arbitrary dimension n as follows æ a0 ç a ç 1 c = a Ä b Û c = circ ( a ) b Û circ ( a ) = ç .. ç ç an -2 ça è n-1 an-1 a0 .. an -3 an - 2 .. .. .. .. .. a2 a3 .. a0 a1 a1 ö a2 ÷÷ .. ÷ ÷ an -1 ÷ a0 ÷ø (8) where the general circulant matrix has its elements ( circ ( a ) ) ij = a[i - j ] (9) The circulant matrix has the following properties circ ( a Ä b ) = circ ( a ) circ ( b ) (10) and since the convolution is a commutative operation, then circulant matrices are mutually commutative circ ( a ) circ ( b ) = circ ( b ) circ ( a ) (11) Let X is an inverse matrix to a circulant matrix circ(a) Xcirc ( a ) = circ ( a ) X = E Its alternative form is -1 Let a (12) X = circ -1 ( a ) be an inverse vector to the vector a, a -1 (13) Ä a = 1 = (1,0 ,...,0 ,0 ) , then assuming that the circulant matrix is regular, circ ( a ) ¹ 0 , it follows circ -1 ( a ) = circ ( a -1 ) (14) Let us define a unary operation involution (see fig. 3) ( )* by a formula : Rn ® Rn ( b = a* = a[0] ,a[-1] ,...,a[- n + 2] ,a[- n +1] (15) ) (16) (a 0,a 1,a 2,...,a n-2, an-1)* =(a 0,a n-1,an-2,...,a 2,a 1) Figure 3. Visualization of the unary operation of involution. The operation of involution satisfies the equations ( a + b ) = a * + b* ( a Ä b )* = a * Ä b* ( a Ä b* ) × c = a × ( b Ä c ) (17b) a ** = a circ ( a* ) = circT ( a ) (17d) (17e) * (17a) (17c) One of the ba8sic aspects of the holographic representation is the possibility of reconstruction of the original components, which were used for construction of convolution of two vectors. This possibility is very INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 308 important, since it allows us to decode the original information from the complex conceptual vectors. Reconstruction of x from cÄx is based on the formula x% = c * Ä ( c Ä x ) » x (18) according to which the convolution c* with the vector cÄx produces the vector x% , which is similar to the original vector x, x% » x . In the component formalism the vector x% has a form n -1 x%i = å c j ( c Ä x )[i + j ] (19) j =0 where ( c Ä x )[ k ] = n -1 åc x [ ] l =1 n -1 l éë k -l ùû x%i = xi å c 2j + j =0 , by using this formula in (19) one gets after simple algebraic manipulations n -1 åc cx[ j ,l = 0 ( j ¹l ) j l éë i + j ]- l ùû æ ö ç ÷ ç ÷ æ n -1 2 ö ç 1 n -1 = ç å c j ÷ xi + n -1 å c j cl xé[i + j ]-l ù ÷ ç ë û ÷ è j =0 ø ç c 2j (j j,l¹=l0) ÷ å j =0 ç 144424443 ÷÷ ç hi è ø (20a) This result can be reformulated in the form æ x0 + h0 ö ç x +h ÷ 1 %x = ç 1 1 ÷ = x + h ( c × c ) ç ............. ÷ çç ÷÷ è xn -1 + hn -1 ø (20b) where h is interpreted as a random noise with a normal distribution with a zero mean and a standard deviation much smaller than x. The overlap of the resulting vector x% with the original vector x is determined from a scalar product by -1 £ overlap ( x , x% ) = x × x% £1 x x% (21) where the inequalities result directly from the Schwartz’s inequality from linear algebra. The more this value is close to its maximum value, the more similar2 are the vectors x% and x. In the fig. 4 a histogram of overlaps is shown for the product c Ä x , containing a couple of randomly generated different conceptual vectors c and x of the dimension n=1000. It is evident from the figure, that the most common overlap between x% = c * Ä c Ä x and x is around 0.7, from which follows, that the vectors x% and x are similar, x% » x . frequency of occurrence 1,0 0,5 0,0 0,4 0,5 0,6 0,7 0,8 overlap(x,x) 0,9 1,0 2 In the case, that the overlap value approaches -1, then the vectors x% and x are also similar, even though they have opposite orientation (they are anticolinear). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 309 Figure 4. A histogram of overlaps between vectors x% and x (of dimensions n=1000) has highest frequency around 0.7, from which follows, that the vectors x% and x are similar. Let’s turn our attention to the second possibility of the verification of the formula (20b) with the application of the approach called the „superposition memory“. Let us have a set containing p+q randomly { } generated conceptual vectors, X = x1 , x2 ,..., x p , x p +1 ,..., x p + q , while p<q. Using the first p vectors from X allows us to define a memory vector t as their sum p t = å xi (22) i =1 0,6 overlap(xi,t) 0,4 0,2 0,0 -0,2 1 2 3 4 5 6 7 8 9 10 11 12 13 14 vectors included in vectors not included in the superposition memory the superposition memory Figure 5. Illustration of the superposition memory for the first 7 vectors of the set X, which contains 14 randomly generated conceptual vectors of the dimension n=1000. The threshold value J can be in this case set to 0.2. The vector t represents a superposition memory, which by a simple additive way contains vector from the set X. The decision, whether some vector x Î X is contained in t must be based on the value of the overlap (21) overlap ( x , t ) = x×t x t (23) If this value is greater than a predefined threshold value, overlap ( x , t ) ³ J , then the vector x is included in the superposition memory t, in the opposite case, if overlap ( x , t ) < J , then the vector x is not included in t (see fig. 5). In both of the previous examples (see figs. 3 and 4) there was used the same method of determination of conceptual vectors, which can „appear“ in some other different complex conceptual vector (which can be the result of complicated previous calculations – transformations). The used method is called „clean-up“ and it is specified as follows: Let us have a set of vectors X = { x1 , x 2 ,..., xn } and some vector t. We face the decision, whether the memory vector (trace) t contains a superposition component, which is similar (or which is not similar) to some vector from the set X. This problem can be solved by calculating so called overlap (23), formally ìï yes x »t =í ïîno ( overlap ( x , t ) ³ J ) ( overlap ( x , t ) < J ) (24) where J is a chosen threshold value of acceptance of the size of the overlap as the positive answer. The result of this cleaning-up process is a subset of vectors X ( t ) = { x Î X ; x » t} Í X (25) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 310 We can put the question also in a rather different form, which is, whether the memory vector t is similar to any of the vectors from the set X? The answer to this more general question shall be decided from the maximum value of the overlap overlap ( t , X ) = max overlap ( t , x ) (26) xÎX Then we can rewrite (24) in the form ( overlap ( x , X ) ³ J ) ( overlap ( x , X ) < J ) ìï yes x» X =í ïîno (27) 3. ASSOCIATIVE MEMORY The construction of the associative memory belongs to the main results of the holographic reduced representation, which can be further generalized by so called chunking. Let us have a set of conceptual vectors X = { x1 , x 2 ,..., xn } and a training set Atrain = {ci xi ; i = 1, 2,...,m} , which contains m<n associated couples of conceptual vectors ci xi , where ci is the input to the associative memory (cue) and xi is the output from the memory. Let’s create a memory vector t representing the associative memory created from the training set Atrain m t = c1 Ä x1 + ... + cm Ä xm = å ci Ä xi (28) i =1 Let us suppose, that we know in advance only the inputs ci to the associative memory, we do not know the possible outputs from the set X train = { x1 , x2 ,..., xm } . The response of the associative memory to the input - clue ci is determined by the process of „clearing-up“ represented by the formula (27). In the first step we shall * calculate the vector x% i = ci Ä t , then by a process based on the maximum value of the overlap we shall find whether x% i » xi Î X overlap ( x% i , X ) = max overlap ( x% i , x ) (29) xÎX train The associative memory will be illustrated by the following two examples. 1st example This example uses only the training set Atrain = {ci xi ; i = 1, 2,...,m} , which is randomly generated for m=8, while the dimension of conceptual vectors is n=1000. For each associated couple ci xi there are ( * calculated ti = ci Ä xi . The values of overlap ci Ä ti , x j 1 2 3 4 5 6 7 8 1 0.71703 -0.03998 -0.02757 0.00785 -0.00466 -0.01467 0.02966 -0.00344 2 -0.01820 0.73804 -0.01736 0.00374 0.00426 0.02522 0.00892 -0.01080 3 0.01452 0.01510 0.64667 -0.01899 -0.01831 -0.01403 -0.00301 0.00843 4 0.02776 0.01430 0.00474 0.68728 -0.00827 -0.01316 -0.00358 -0.01871 ) are presented in the table 5 -0.01488 0.00276 -0.11580 -0.15340 0.70767 -0.03000 0.01285 0.00324 6 -0.01922 0.02346 -0.00812 0.00005 0.04175 0.71444 0.00971 -0.02629 7 -0.02442 -0.00545 0.01476 -0.00561 -0.03384 0.00078 0.70790 0.00851 8 0.01358 -0.01626 0.00379 0.00136 -0.00668 -0.00526 0.01816 0.58957 It is evident from the table, that the overlaps are sufficiently great just for diagonal values, while the non-diagonal overlaps are smaller by an order of magnitude. We can therefore unambiguously decide from the * overlap, whether ci Ä ti » xi is associated with the cue ci. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 311 2nd example In this illustrative example we shall use the training set Atrain = {ci xi } , generated for m=10 associated couples – vectors of dimension n=1000. This memory is represented by a memory vector t = c1 Ä x1 + ... + cm Ä xm . The following table shows 20 experiments of „clean up“, where we used with a 50% probability as an associative entry a vector ci from the training set or a randomly generated conceptual vector. The table contains maximal values of overlaps (29), by which we can unambiguously determine, whether the used input has an associated counterpart in the training set. # max. overlap Input index 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 0.311 0.047 0.383 0.373 0.316 0.397 0.074 0.065 0.069 0.039 0.344 0.402 0.032 0.073 0.017 0.004 0.033 0.056 0.373 0.037 6 rand. gener. 5 10 3 4 rand. gener. rand. gener. rand. gener. rand. gener. 7 8 rand. gener. rand. gener. rand. gener. rand. gener. rand. gener. rand. gener. 10 rand. gener. index of output with max.overlap 6 nonexistent 5 10 3 4 nonexistent nonexistent nonexistent nonexistent 7 8 nonexistent nonexistent nonexistent nonexistent nonexistent nonexistent 10 nonexistent It follows from the table, that the associative memory with the clean up process is unambiguously identifying, the values of a maximum overlap for conceptual vectors well specify the existence (or nonexistence) of corresponding associative outputs. 1.0 overlap 0.8 a b d c 0.6 0.4 0.2 0.0 1 2 c d e f g h i j 3 4 5 6 7 8 9 10 11 12 superposition memory associative memory Figure 6. Illustration of analysis of a combination of superposition memory and associative memory. In the first step we carried out the process of clean up, by which we found, that the memory contains only two items – atomic vectors a and b, which in the next step serve as an input for further analysis of the associative memory, where we found as outputs the vectors c and d. During the clean up process we moreover verified with a negative INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 312 result, whether the superposition memory contains also further vectors c,d,...,i,j. With a great probability we can therefore decide, that the memory vector t is the combination of the superposition and associative memory t = a + b + a Ä c + b Ä d . The dimension of the used vectors is n=1000. Combination of superposition memory and associative memory We shall show, that also a combination of can be superposition and associative memory reliably analyzed, which will prove in our further applications as an advantageous feature of the holographic representation. Let us presume, that we have 10 conceptual vectors a,b,c,d,...,i,j, and from the first four we shall construct a combination of a superposition and associative memory as follows t = a+ b+ aÄc +bÄd (30) By the clean up procedure we can find out, that the vector t contains as its parts the vectors a and b, which we shall in the following step remove from the vector t t¢ = t - a - b = a Ä c + b Ä d (31) From the remaining superposition part we can find out by the application of the analysis (as from the associative memory) that it contains two couples a c and b d , see fig. 6. 4. SEQUENCE OF SYMBOLS The construction of the associative memory does not allow storing of structured data, the aim of this chapter is to show, that a holographic distributed representation is able to process a linear sequence of symbols, which are represented by a sequence of conceptual vectors. To concretize our thoughts, let us study a sequence of 6 conceptual vectors of the dimension n=1000 sequence = {a ® b ® c ® d ® e ® f } (32) For these vectors we shall construct a memory vector t0 = a + a Ä b + a Ä b Ä c + a Ä b Ä c Ä d + a Ä b Ä c Ä d Ä e + a Ä b Ä c Ä d Ä e Ä f (33) We know, that this vector contains the sequence of vectors coded by (33), but we do not know, which vectors these are and in what order. We shall show, that by the clean up procedure we can from the vector t0 reconstruct the original sequence (32) step by step using the following procedure (see fig. 7): 1. step: a = clean _ up ( t0 ) , t1 := t0 - a , t%1 := a* Ä t1 , 2. step: b = clean _ up ( t%1 ) , t2 := t1 - a Ä b , * t%2 := ( a Ä b ) Ä t2 , 3. step: c = clean _ up ( t%2 ) , t3 := t2 - a Ä b Ä c , * t%3 := ( y1 Ä y2 Ä y3 ) Ä t3 , 4. step: d = clean _ up ( t%3 ) , t4 := t3 - a Ä b Ä c Ä d , * t%4 := ( a Ä b Ä c Ä d ) Ä t4 , 5. step: e = clean _ up ( t%4 ) , t5 := t4 - a Ä b Ä c Ä d Ä e , * t%5 := ( a Ä b Ä c Ä d Ä e ) Ä t5 , 6. step: f = clean _ up ( t%5 ) . INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 313 (a a b a b c a b c d a b c d e a b c d e f )» a step 1 a *Ä a b a b c a b c d a b c d e a b c d e f )» b step 2 a b c d a b c d e a b c d e f )» c step 3 ) ( a b c d a b c d e a b c d e f )» d step 4 (a b c d *Ä a b c d e a b c d e f )» e step 5 b c d e *Ä a b c d e f )» f step 6 ( ( a b )* Ä( a (a b c b c *Ä ) ( (a ) ( Figure 7. Representation of the reconstruction of the sequence of symbols abcdef (single symbols are represented by randomly generated conceptual vectors). The clean up procedure is successively applied to the sequence of t vectors, by which we shall obtain the corresponding sequence vector. 1.0 overlap 0.8 0.6 0.4 0.2 0.0 a b c d e vectors of the sequence f Figure 8. The overlap for single vectors with a sequence from (30), which were obtained by the reconstruction from the vector t0. It is apparent from the figure, that a degradation of reconstruction relatively quickly appears, already the sixths vector f is reconstructed with a probability smaller than 0.20. The function clean_up(×) carries out the clean up process for the given vector t with respect to the set of vectors X = {a , b ,..., f , g , h,...} . The single steps of the reconstruction of the sequence of conceptual vectors – symbols are shown in the fig. 8, from which follows, that the process of the reconstruction of a sequence of symbols rather quickly degrades, already for the sixth vector the overlap is smaller than 0.2. The sequence of symbols can be coded also by an associative memory, where the vector of the entry ci specifies ith position of the given symbol. The above mentioned illustrative example is represented by a memory vector t = c1 Ä a + c2 Ä b + c3 Ä c + c4 Ä d + c5 Ä e + c6 Ä f (34) The recognition of this sequence consists in the search of the associate vector to the input vector ci , by application of the clean up process there can be constructed a „training set“ Atrain = {c1 a , c2 b , c3 c , c4 d , c5 e , c6 f } (35) which unambiguously specifies the sequence of vectors. The advantage of such a technique is in the accuracy of recognition, which does not degrade so fast as during the original procedure based on the memory vector (33). The associative approach to the implementation of the memory for a sequence of symbols (represented by atomic vectors a, b, ) can be easily modified into the form of so called „stack memory“. Associative entries ci are determined as follows ci = pi ( i = 1,2 ,...) (36) where pi is the ith (convolutive) power of randomly generated conceptual vector p, p = p memory vector (32) has a form i i-1 Ä p . Then the INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 314 t = p Ä x1 + p 2 Ä x2 + p3 Ä x3 + ... + p n Ä xn (37) where xi are single items from memory {a, b,...}. Such an interpreted associative memory for the sequence of symbols is called the „stack memory“, with the help of power entries pi we can easily change their contents, see fig. 9. Over this memory we can define three different operations, by which we can change its contents: t pop(t) top(t) push(t) Figure 9. Three possible operators for the stack memory represented by a vector. (1) push ( t , x ) = p Ä x + p Ä t , the new item x is placed to the top of the stack. ( ) (2) top ( t ) = clean _ up p Ä t , recognizes the top item in the stack. (3) pop ( t ) = p * Ä t - top ( t ) , removes the top item from the stack. -1 The most problematic is the third operation, which removes the top item from the stack. The correct implementation needs an application of the exact inverse vector p-1, the approximation of this inverse vector by involution, p -1 ! p* , leads to a fast degradation of the stack memory. 5. MEMORY CHUNKS Memory chunks help to overcome the problems with a degradation of memory for a sequence of symbols (see chap. 4). Let us have a set of conceptual vectors S = {a , b ,..., k , l ,...} , this set shall be divided into disjoint subsets - chunks (S È S S = S1 È S2 È S3 È S4 È ... i j = Æ , pre i ¹ j ) (38) Let’s study a set S = {a,b,c,d ,e, f , g,h} , its decomposition into chunks looks as follows (see fig. 10) S1 = {a,b,c} , S 2 = {d ,e} , S3 = { f } , and S 4 = { g,h} a s s1 (39) d f b e s3 s2 h c i s4 g Figure 10. Illustration of the decomposition of the set S into disjoint chunks, see (38). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 315 Chunks are represented by a vector, which represents a sequence of chunks {s1 ® s2 ® s3 ® s3 } t = s1 + s1 Ä s2 + s1 Ä s2 Ä s3 + s1 Ä s2 Ä s3 Ä s4 (40b) while single chunks are defined by corresponding sequences of vectors (see fig. 11) s1 = a + a Ä b + a Ä b Ä c s2 = d + d Ä e s3 = f s4 = g + g Ä h (40c) (40d) (40e) (40f) The processing of the memory chunks can be divided into two steps: 1. step – by a clean up process we shall identify chunks contained in t (we presume, that the clean up process has the set X = { x1 , x 2 ,..., xn } from the end of the 2nd chapter, where this process was specified, enlarged also by chunks s1 , s 2 , s3 , s4 , i.e. in our illustrative example X = {a , b , c , d , e , f , g , h, s1 , s2 , s3 , s4 } ). 2. step – the identified chunks are further processed by the clean up technique. t s1 s2 s3 a b c d e s4 f 1st stage 2nd stage g h Figure 11. The chunking of 8 conceptual vectors onto 4 chunks. In the 1st stage the clean up process identifies the chunks, which are then in the 2nd step further analyzed up to vectors describing elementary concepts. 1,0 a abc 0,8 de e b overlap f d g h 0,6 f gh c 0,4 0,2 0,0 1 2 3 set of chunks (abc,de,f,gh) 4 5 6 7 8 9 10 11 12 1st chunk 2nd chunk 4th chunk (cde) 3rd chunk (gh) (abc) (f) Figure 12. The representation of the two-step clean up process, where in the first step the chunks are identified, while in the second step there are identified the vectors corresponding to atomic concepts from the already identified chunks. The result of the two-step process of clean up is shown in the fig. 11. It is evident from this figure, that in the case of a long sequence of conceptual vectors a fast degradation during the clean up process can be partially overcome by the chunking of concepts onto chunks, which are at the highest level separately coded. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 316 t=s12345 s1 s2 a b c d e 1st step s345 s12 s3 f s4 s5 g h i j k 2nd step 3rd step Figure 13. Illustration of chunks of higher order, where some of the used chunks are composed from simpler chunks. 1.0 overlap 0.8 0.6 0.4 0.2 0.0 s12 s345 s1 s2 s3 s4 s5 a s12345 s12 s345 b c d e s1 s2 f g h s3 s4 i j k s5 Figure 14. Illustration of a 3-step clean up of the chunked trace t specified by the formula (41h). In the first step the trace t is analyzed, it is found, that it contains two chunks s12 and s345. In the second step the two chunks from the previous step are analyzed, and found to contain chunks s1, s2, ..., s5. In the last third step the 5 chunks identified in the previous step are successively analyzed. They contain the vectors a, b, ..., k. The chunking method can be generalized, so that the chunks of higher order are created, i.e. chunks composed of chunks, see fig. 12, where single chunks are defined as follows (the used vectors of concepts a, b, ..., c have the dimension n=1000) s1 = a + a Ä b + a Ä b Ä c s2 = d + d Ä e s3 = f s4 = g + g Ä h s5 = i + i Ä j + i Ä j Ä k s12 = s1 + s1 Ä s2 s345 = s3 + s3 Ä s4 + s3 Ä s4 Ä s5 (41a) (41b) (41c) (41d) (41e) (41f) (41g) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 317 t = s12345 = s12 + s12 Ä s345 (41h) The process of clean up of the chunked memory trace t specified by (41h) is shown in the fig. 13. In the 1st step we analyze the trace t, its analysis tells us, that the trace t contains two chunks s12 a s345. In the second step we analyze chunks from the previous first step, and we recognize their contents as chunks s1, s2,..., s5. In the last third step we analyze chunks from the previous step, which already contain atomic conceptual vectors a, b, ..., k. The overlaps of the resulting vectors in the clean up process are shown at fig. 14. The presented illustrative examples show, that the memory chunk approach represents an effective way to overcome fast degradation of the original version of a successive analysis of the vector (33). By combining of several conceptual vectors into a chunk, we shall gain a simple opportunity to extend our ability to analyze correctly greater sets of conceptual vectors. The chunking process can have several hierarchic levels, which removes the limits from our ability to store and recall conceptual vectors. 6. CODING OF RELATIONS Holographic reduced representation can serve also as a suitable means for encoding relations (predicates). Let us study a binary relation P(x,y), when the Pascal code is used, this relation is formally specified by the head function P ( x : type1 ; y : type2 ) :type3 (42) The single arguments of the relation are specified by the types type1 and type2, which specify the domain, over which are these variables defined; similarly also the relation P itself is understood as a function, which domain of values is specified by the type type3. In many cases the domain of variables and also the domain of the relation itself are equal to each other; therefore their specifications can be omitted, which substantially reduces the holographic representation of relations. The reduced form of relation (42) looks as follows function P ( x ; y ) (43) where we know in advance the type of variables x, y, and also the type of the relation P itself. The holographic representation of the relation (42) can have the following form t = P + variable1 + variable2 + (44) P Ä ( type3 + variable1 Ä ( x + type1 ) + variable2 Ä ( y + type2 ) ) Their decoding is carried out step-by-step. In the first step we use the clean up procedure to recognize the name (identifier) of the relation P and also the names (identifiers) of its variables x and y. In the second step we identify the type type3 of the relation P, in the last, third step we use previous results to identify variables x, y and also their types type1 and type2. In many cases the representation of the relation P(x,y) is satisfactory in the following simplified form (see (43)) t = P + variable1 Ä x + variable2 Ä y (45) The chosen method of the holographic representation of relation can be easily generalized also for more ( ) complex (higher order) relations, where the variables are predicates as well, e.g. P x,Q ( y,z ) , where the „inner“ predicate Q is characterized by function Q ( y : type3 ; z : type4 ) :type5 ( (46) ) In order to create a higher order relation P x,Q ( y,z ) , we must presume a type compatibility of the second variable of the relation P and of the type of relation Q, i.e. type2=type4. In the simplified approach, where all the types are the same, it is not necessary to distinguish the types of single variables and the relations themselves. A simplified holographic representation of relation (46) has the following form t ¢ = Q + variable1 Ä y + variable2 Ä z (47) By exchanging the representation (47) for the variable y in the representation (45) we get the following ( resulting representation of the higher order relation P x,Q ( y,z ) ) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 318 t = P + variable1 Ä x + variable2 Ä ( Q + variable3 Ä y + variable4 Ä z ) = P + variable1 Ä x + variable2 Ä Q + (48) variable2 Ä variable3 Ä y + variable 2 Ä variable 4 Ä z 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 29 30 31 32 Figure 15. A set of 48 similar figures, which contain two objects, placed either next to each other, or above each other and the objects are either small or big. Every column contains a couple of similar objects, which differ only by their placement or size. 1st illustrative example –a similarity between geometric figures In the figure 15 there are presented 48=6´8 geometric patterns, which contain either in horizontal or in vertical settings two objects, which moreover can be of two sizes, small and big. Let us mark holographic representations of corresponding atomic concepts as follows: Objects: tr (triangle), sq (square), ci(circle), st (star) Unary relations: sm (small), lg (large) Binary relations: hor (horizontal), ver (vertical) Variables: ver_var1 (1st variable for binary relation ver), ver_var1 (2nd variable for binary relation ver), hor_ver1 (1st variable for binary relation hor), hor_ver2 (2nd variable for binary relation hor) Single figures from z fig. 15 are characterized by relations given in the following table. row 1 2 3 4 5 6 specification ver(lg(x),lg(y)) hor(lg(x),lg(y)) hor(sm(x),lg(y)) and hor(lg(x),sm(y)) ver(sm(x),lg(y)) and ver(lg(x),sm(y)) ver(sm(x),sm(y)) hor(sm(x),sm(y)) Holographic representations of single cases from this table have the following form (compare with the equation (45)). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 319 t1,x , y = ver + ver _ var1 Ä lg Ä x + ver _ var2 Ä lg Ä y t2 , x , y = hor + hor _ var1 Ä lg Ä x + hor _ var2 Ä lg Ä y ìïver + ver _ var1 Ä lg Ä x + ver _ var2 Ä sm Ä y t3 , x , y = í ïîver + ver _ var1 Ä sm Ä x + ver _ var2 Ä lg Ä y ìïhor + hor _ var1 Ä lg Ä x + hor _ var2 Ä sm Ä y t4 , x , y = í ïîhor + hor _ var1 Ä sm Ä x + hor _ var2 Ä lg Ä y t5 ,x , y = ver + ver _ var1 Ä sm Ä x + ver _ var2 Ä sm Ä y (49) t6 ,x , y = hor + hor _ var1 Ä sm Ä x + hor _ var2 Ä sm Ä y where x and y are holographic representations of single objects (tr, sq, ci, st) and the bracket u indicates, that the vector u is normalized. The similarity between single figures is determined by the overlap of their holographic representations similarity ( X , X ¢ ) = overlap ( t , t ¢ ) 0.42 0.58 0.13 0.09 0.52 0.16 0.48 0.47 0.15 0.43 0.16 0.54 (50) 0.12 0.54 0.52 0.45 Figure 16. Illustrative presentation of similar figures for two chosen figures 1 and 48 (see fig. 16). Single arrows are marked by the overlap between the figures calculated by formula (50). The obtained results are shown in the fig. 16. The dominant feature controlling similarity value is the horizontal or vertical arrangement of objects. The overlap (i.e. also the similarity) between two figures, which have different arrangement is usually smaller than 0.1. In general, holographic reduced representation allows fairly simple determination of similarity of objects specified by a predicate structure (42) or by its generalization through further nested predicates (see (48)). This possibility opens new horizons on future developments in fundamental methods of search for similar objects or analogies, which are considered very difficult problems for artificial intelligence requiring special symbolic techniques [7]. 2nd illustrative example – similarity between binary numbers We shall study similarity between binary numbers of the length 3, which are represented by a sequence ( a1a 2 a3 ) Î {0 ,1}3 . This number can be understood as an ordered triple of binary symbols, which are in the distributed representation represented as follows (see (33)) t( a1a 2a3 ) = ta1 + ta1 Ä ta2 + ta1 Ä ta2 Ä ta3 (51a) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 320 ìï zero ta = í ïîone ( for a = 0 ) ( for a = 1) (51b) where zero and one are distributed representations of numbers ‘0’ or ‘1’. For example, a binary number (101) is represented as follows t(101) = one + one Ä zero + one Ä zero Ä one 7 0.1250 6 0.2500 5 0.3750 4 0.5000 3 0.6250 (52) 0.7500 2 1 0.8750 0 1.000 0 1 2 3 4 5 6 7 Figure 17. Graphic representation of similarity between distributed representations of binary numbers, the size of overlap between couples is proportional to the brightness of the corresponding square area (the areas on the skew diagonal are the brightest ones). The darkest areas are in the bottom lower corner and in the left upper corner. It corresponds to the fact that these areas are assigned to representations of maximally distant couples of numbers. The similarity between single representations is reflected also by a similarity between corresponding binary numbers; the representation of two close binary numbers is inversely proportional to their distance (e.g. to the absolute value of their difference). In the following table there are given the similarities between representations of binary numbers, which were calculated from their overlap (23). 000 001 010 011 100 101 110 111 000 1.00 0.74 0.50 0.46 0.21 0.16 0.14 0.11 001 0.74 1.00 0.71 0.42 0.45 0.15 0.12 0.15 010 0.50 0.71 1.00 0.75 0.70 0.43 0.09 0.08 011 0.46 0.42 0.75 1.00 0.46 0.70 0.37 0.04 100 0.21 0.45 0.70 0.46 1.00 0.74 0.43 0.39 101 0.16 0.15 0.43 0.70 0.74 1.00 0.70 0.35 110 0.14 0.12 0.09 0.37 0.43 0.70 1.00 0.69 111 0.11 0.15 0.08 0.04 0.39 0.35 0.69 1.00 The maximum similarity is between couples of representations assigned to two neighboring integers, minimum similarity of two representations occurs, if the corresponding numbers have a maximum distance, which equals 7. These values from the table are graphically represented in the fig. 17. This simple illustrative example shows, that in the framework of the holographic distributed representation one can use (at least potentially) associative representations of the type (28), where associative cues correspond to numbers. It means that in this distributed approach there exists a possibility of associative simulation of an arbitrary function, which substantially increases the potential of the method to be used universally. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 321 7. REASONING BY MODUS PONENS AND MODUS TOLLENS Simulation of reasoning processes (inference) belongs to the basic problems, which are repeatedly solved in artificial intelligence and cognitive science [15]. Fodor‘s critique of connectionism [19] was based precisely on the brash conclusion, that artificial neural networks are not able to simulate higher cognitive activities, which are purported to be an exclusive domain of the classical symbolic approach. This Fodor’s opinion was proved to be incorrect, further development of theory of neural networks showed, that connectionism is a universal computational tool, which does not have limits of applicability, it does not have domains of inapplicability, which would be forbidden for it. Of course, it can transpire, that in some domains its application is extremely cumbersome and exceedingly complicated, that there exist other approaches, which in the given domain provide substantially simpler and direct solution, than the one provided by neural networks. In this chapter we shall show a possibility of representation of two basic modes of deductive reasoning of modal logic, pÞq p and q pÞq q (53) p which are called modus ponens resp. modus tollens. These modes of reasoning are equivalent to the following tautologies of the predicate logic (( p Þ q ) Ù p ) Þ q (( p Þ q) Ù q ) Þ p (54a) (54b) Implication ••• can be understood as a binary relation, which can be in holographic distribution represented like this (see formula (47)) t p Þ q = op Ä impl + var1 Ä p + var2 Ä q (55) which contains a sum of three parts, the first part specifies the type of relation (implication), the second and third parts specify the first (antecedent) resp. the second (consequent) variable of the relation of implication. This conceptual vector representing relation of implication can be transformed as follows t%p Þ q = t p Þ q Ä T (56a) where T = var1* Ä p * Ä p * Ä q + var2* Ä q * Ä q * Ä p (56b) The transformed representation of implication is represented by a sum of two associated couples t%p Þq » p* Ä q + q * Ä p (57) which gives the holographic representation of the rules modus ponens and modus tollens p Ä t%pÞq » q q Ä t%pÞ q » p (58a) (58b) The first formula (58a) can be understood as a holographic representation of modus ponens (see (53) and (54a)), while the other formula is a holographic representation of modus tollens (see (53) and (54b)). A similar result can be obtained also by an alternative approach, which is based on the disjunctive form of implication ( p Þ q) º ( p Ú q) (59) The distributed representation of implication in this alternative form can be expressed by t p Ú q = op Ä disj + var1 Ä p + var2 Ä q (60) By a transformation of this representation we can get (see (55)) t%p Ú q = t p Ú q Ä T » p Ä q * + q Ä p* (61a) where T = var1* q * + var2* p* (61a) This transformation is much simpler than the one in the previous case (56b). The rules modus ponens and modus tollens are now realized by formulas similar to (58a-b). Moreover, we get also the following two „rules“ q* Ä t%pÚ q » p* (62a) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 322 p* Ä t%pÚ q » q * (62b) which remind us of the well known fallacies pÞq q pÞq p a p (63) q that are known as „affirming the consequent“ resp. „denying the antecedent“. This fault is caused by the fact, that the transformed representations of implications t%p Þ q and t%p Ú q are not identical, the representation t%p Ú q leads to unexpected results (63), which represent erroneous modes of reasoning (which are however often used by people without knowledge of principles of logic). 8. PREDICATE LOGIC We shall further deal with a simple form of predicate logic, which is based on unary predicates, P(x), where the distributed representation has a form (see chapter 6) t P( x ) = pred Ä P + pred_var Ä x (64) ( "x ) P ( x ) , can be represented in t( "x ) = uni _ quant Ä uni + uni _ quant_var Ä x (65a) The connection of this predicate with the universal quantifier, the following way t( "x) P( x ) = t( "x) + t( "x) Ä ( pred Ä P + pred_var Ä x ) (65b) Both conceptual vectors t P ( x ) a t( "x ) P ( x ) can be recognized and extracted by a clean up procedure. This process is unnecessarily complicated for our purposes of further study of reasoning processes in the framework of predicate logic and their distributed representation; the application of the conceptual vector t(x) for the representation of the symbol (x) basically only unnecessarily complicates the process of analysis of composed conceptual vectors containing as a constituent t(x). We shall therefore cease to use the symbol (x) explicitly, its meaning will be substituted by usage of a „universal variable“ x, i.e. predicate P(x) containing the universal variable x is interpreted as ( "x ) P ( x ) , we can therefore with a certain caution use a „formula“ ( "x ) P ( x ) º P ( x ) . In the predicate logic there exists a rule of universal instantiation, which concretizes a predicate with a universal quantifier onto a predicate with a concrete variable a, ( "x ) P ( x ) Þ P ( a ) , which is a result of a simple tautology of propositional logic (( p Ù q ) Ù p ) Þ q . With an application of the universal variable x we shall rewrite this concretization into a simpler form P ( x) Þ P (a) (66) We shall construct a distributed representation of this universal instantiation of a simple unary replicator by a transformation vector T, see equations (56-58). The distributed representation (66) looks as follows t P( a) » t P( x ) Ä T (67) t P( a) » t P( x ) Ä x * Ä a (68) * where T = x Ä a , then We can recapitulate our thoughts in saying that distributed representation of the universal instantiation is concretized by a transformation vector T , which helps to substitute a universal variable x by a „specified“ variable a. We shall use this simplified representation of quantified predicates to study so called generalized modus ponens and generalized modus tollens INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 323 ( "x ) ( P ( x ) Þ Q ( x ) ) P (a ) Q (a) ( "x ) ( P ( x ) Þ Q ( x ) ) Q (a ) P (a) a or in a simplified form using a universal variable x (69a) P ( x) Þ Q ( x) P ( x) Þ Q ( x) P (a ) a Q (a) Q (a) P (a) (69b) These generalized schemes of deductive reasoning follow directly from their standard sentential form (53) and concretization (63). The distributed representation of the main (top) premise of these rules has a form t P( x )ÞQ( x ) = op Ä impl + var1 Ä ( pred Ä P + pred_var Ä x ) + (70) var2 Ä ( pred Ä Q + pred_var Ä y ) The concretization of implication P ( x ) Þ Q ( x ) onto P ( a ) Þ Q ( b ) , can be formally expressed by an implication (see (66)) ( P ( x ) Þ Q ( x ) ) Þ ( P ( a ) Þ Q ( a )) (71) where the right hand side has the following distributed representation t P( a)ÞQ( a) = op Ä impl + var1 Ä ( P + a ) + var2 Ä ( Q + b ) (72) Similarly as in the introductory illustrative example (see (67)), this transfer expressed by an implication (71) can be in a distributed representation written by the following transformation * t%P( a)ÞQ( a ) = t P( a )ÞQ( a ) Ä T » ( P + a ) Ä ( Q + b ) + ( Q + b ) Ä ( P + a ) (73) where the new transformed distributed representation t%P a Þ Q a satisfies the formulas, which represent the rules * ( ) ( ) (69) modus ponens resp. modus tollens ( P + a ) Ä t%P( a)ÞQ( a) » ( Q + b ) ( Q + b ) Ä t% ( ) P a ÞQ( a ) (74a) » (P + a) (74b) Illustrative example – modeling of reflexive reasoning In this illustrative example we shall show, that the holographic distributed representation provides formal tools, which can be used to simulate the reasoning process based on generalized modus ponens (69). This process was widely studied by Shastri a Ajjanagadde [15] by the connectionist system called SHRUTI, which was able to simulate reflexive reasoning based on predicate logic. Similar results are achieved also by a formalism of holographic distributed representation. give(x,z,y)Þown(y,z) give(John,Mary,book) buy(x,y)Þown(x,y) buy(John,something) own(Mary,book) own(x,y)Þcan_sell(x,y) own(John,something) can_sell(Mary,book) can_sell(John,something) Figure 18. Illustration of an application of the generalized rule modus ponens (66) for deduction or knowledge discovery (marked by gray shading and also by incoming arrows) from implications (1-3) and from input facts (a-c), marked by outgoing arrows. Let us have a formal system containing three general rules (see fig.18): 1. give ( x, y,z ) Þ own ( y,z ) , type x : donor; type y : acceptor; type z : object, INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 324 2. buy ( y,z ) Þ own ( y,z ) , type y : buyer; type z : object, 3. own ( y, z ) Þ can _ sell ( y,z ) , type y : owner; type z : object, and three observations (facts) · give ( John,Mary,book ) , · own ( Mary,book ) · buy ( John,something ) . What are the deductive conclusions of this system? The results are shown in the figure 18, we shall now deduce them with an application of distributed representation based on conceptual vectors and operations over them. Let us analyze the first generalized modus ponens from the figure 18 give ( x, y,z ) Þ own ( y,z ) give ( John,Mary,book ) (75) own ( Mary,book ) With an application of the approach described by (70-74) we can realize this scheme of reasoning by a representation of conceptual vectors, its single items (going top down) are represented as follows 1 t1( ) = op Ä impl + var1 Ä tdať ( x ,y ,x ) + var2 Ä tvlastniť ( y ,x ) t2( ) = t give( John ,Mary ,book ) = give + give_var1 Ä John + (76a) 1 give_var2 Ä Mary + give_var3 Ä book t3(1) = town( Mary ,book ) = own + own_var1 Ä Mary + own_var2 Ä book (76b) (76c) where conceptual vectors t give( x ,y ,x) and town( y ,x ) are constructed analogically as in (45). In the first step we must carry out a concretization of the implication give ( x, y,z ) Þ own ( y,z ) , so that the general variables x, y, z are substituted by concrete variables John, Mary, book. This concretization is carried out by a transformation T, which is specified by formulas (67-68) ˆt1(1) » t1(1) Ä T where ˆt1(1) = op Ä impl + var1 Ä t + var2 Ä town( Mary ,book ) give ( John ,Mary ,book ) T = t1*(1) Ä ˆt1(1) (77) (78a) (78b) Thus concretized representation give ( John,Mary,book ) Þ own ( Mary,book ) is in the next step applicable for modus ponens carried out by formulas (73-74) 1 1 t%1( ) » ˆt1( ) Ä T ¢ (1) where the resulting conceptual vector t%1 (79a) already represents modus ponens, i.e. the next formula holds t give( John ,Mary ,book ) Ä t%1( ) » town( Mary ,book ) 1 (80) The other three generalized modus ponens from the figure 18 can be realized in a similar way. Illustrative example –generalization by induction Let us have a „training“ set composed of sequences of simple unary predicates, which can be interpreted as observations Atrain = {m ( ai ) ; i = 1, 2,...,q} (81) Our goal will be to generalize these particular predicates into the form with a universal quantificator (or in our simpler formalism, with a universal variable) INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 325 (82) ( "x ) éëm ( x )ùû º m ( x ) It means, that the single particular cases m ( ai ) from the training set (77) are generalized into a formula, which is not directly deducible from them (see fig. 19) m ( a1 ) Ù ... Ù m ( a p ) Þ m ( x ) ? m(a1) m(a2) ...... {m(a1), m(a2),...., m(ap)} (83) m(x) m(ap) stand-alone observations unification of observations into a single set inductive generalisation Figure 19. Diagrammatic scheme of inductive generalization, which consists from three stages. In the first stage we have isolated observations m ( ai ) , which are not related. In the second stage the isolation of single observations is replaced by their unification into a single set, which expresses the fact that observations are not isolated and independent, but they have something in common. In the final, third stage, the unified form of observations is inductively generalized using a universal variable x, which represents a class of objects with the same property m. Let each predicate m(ai ) from the training set Atrain is holographic represented in the following way tm( ai ) = rel Ä m + m_var Ä ai (84) These represented conceptual vectors can be mutually transformed by transformational vectors Ti,j tm( a ) » tm( ai ) Ä Ti , j j Ti , j = t m* ( ai ) Ä t m( a ) » ai* Ä a j j % Let us define a new transformation vector T (85a) (85b) i q T%i = å Ti , j = ai* Ä x (86a) x = a1 + a2 + ... + a p (86b) j =1 where the conceptual vector x is assigned to the new ”generalized” vector, which represents each argument ai from the training set Atrain (in our further consideration about induction it is understood as a stand-alone conceptual entity equal to the representations of the original objects a1, a2, ..., ap) tm( ai ) Ä T%i » tm( x ) (87) This formula can be interpreted as an inductive generalization, where particular objects ai from the training set in representations tm( a ) are substituted by a new object x, that can be interpreted as a new universal i object. 9. CONCLUSIONS Holographic reduced representation offers new unconventional solution to one of the basic problems of artificial intelligence and cognitive science, which is to find a suitable distributive coding of structured information (sequence of symbols, nested relational structures, etc.). The used distributed representation is based on two binary operations: unary operation „involution“ and binary operation „convolution“ over a domain of n-dimensional randomly generated conceptual vectors, which elements satisfy normal distribution N(0,1/n). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 326 Application of this distributed representation allows us to model various types of associative memory, which are represented by a conceptual vector and also to decode a memory vector, i.e. to determine the conceptual (atomic) vectors it is composed of. Such an analysis of the memory vector is carried out by a clean-up procedure that determines from the overlap of the vectors, which of the vectors is the most similar to the memory vector. We have also described the process of chunking of vectors, which allows us to overcome the undesirable fast degradation of success in retrieval of all the components of the memory vector. The countermeasure against the degradation consists in chunking of several vectors into one, which is then put into the memory vector. Holographic reduced representation allows to measure similarity between two structured concepts by a simple algebraic operation of scalar product of their distributed representations. This fact can be very useful, when we want to model processes, which search through memory to find its similar (analogical) single components. In the last part of the paper we have demonstrated, that the holographic reduced representation may be used also to model an inference process based on the rules modus ponens and modus tollens, and moreover, we have demonstrated the effectiveness of the approach for modeling of an inductive generalization process. ACKNOWLEDGMENTS This work was supported by the grants # 1/0062/03 and # 1/1047/04 of the Scientific Grant Agency of Slovak Republic. REFERENCES [1] DAVIS, P. J.: Circulant Matrices. Chelsea Publishing, New York, 1999. [2] GABOR, D.: Holographic model for temporal recall. Nature, 217 (1968), 1288–1289. [3] McCLELLAND, J. L. M., RUMELHART, D. E., and the PDP research group (eds.): Parallel distributed processing: Explorations in the microstructure of cognition, volumes 1 and 2. The MIT Press, Cambridge, MA (1986). [4] METCALFE Eich, J.: Levels of processing, encoding specificity, elaboration, and charm. Psychological Review, 92 (1985), 1–38. [5] MURDOCK, B. B.: A theory for the storage and retrieval of item and associative information. Psychological Review, 89 (1982), 316–338. [6] PLATE, T. A.: Distributed Representations and Nested Compositional Structure. Department of Computer Science, University of Toronto. Ph.D. Thesis, 1994. [7] PLATE, T. A.: Holographic Reduced Representation: Distributed Representation for Cognitive Structures, CSLI Publications, Stanford, CA, 2003. [8] PLATE, T. A.: Holographic Distributed Representations. IEEE Transaction on Neural Networks, 6 (1995), 623-641. [9] SLACK, J. N.: The role of distributed memory in natural language processing. In O’Shea, T. (ed.): Advances in Artificial Intelligence: Proceedings of the Sixth European Conference on Artificial Intelligence, ECAI84. Elsevier Science Publishers, New York, 1984. [10] SMOLENSKY, P., LEGENDRE, G., MIYATA, Y.: Principles for an Integrated Connectionist/Symbolic Theory of Higher Cognition. Technical Report CU-CS-600-92, Department of Computer Science and 92-8, Institute of Cognitive Science. University of Colorado at Boulder, 1992, 75 pages. [11] SMOLENSKY, P.: On the proper treatment of connectionism. The Behavioral and Brain Sciences, 11 (1988), 1-74. [12] SMOLENSKY, P.: Tensor product variable binding and the representation of symbolic structures in connectionist networks. Artificial Intelligence, 46 (1990)., 159-216. [13] SMOLENSK, P: Neural and conceptual interpretations of parallel distributed processing models. In J. L. McClelland, D. E. Rumelhart, & the PDP Research Group, Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Volume 2: Psychological and Biological Models. MIT Press - Bradford Books, Cambridge, MA, 1986, pp. 390-431. [14] WILLSHAW, D. J., BUNEMAN, O. P., LONGUET-HIGGINS, H. C.: Non-holographic associative memory. Nature, 222 (1969), 960–962. [15] SHASTRI, L., AJJANAGADDE, V.: From simple associations to systematic reasoning: connectionist representation of rules, variables, and dynamic bindings using temporal synchrony. Behavioral and Brain Sciences, 16 (1993), 417-494. [16] KANERVA, P.: The Spatter Code for Encoding Concepts at Many Levels. In Marinaro, M., Morasso, P. G. (eds.): ICANN ’94, Proceedings of the International Conference on Artificial Neural Networks. Springer Verlag, London, 1994, pp. 226–229. [17] KANERVA, P.: Binary Spatter-Coding of Ordered K-tuples. In von der Malsburg, C., von Seelen, W., INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 327 Vorbruggen, J. C., Sendhoff, B. (eds.): ICANN'96, Proceedings of the International Conference on Artificial Neural Networks. Springer Verlag, Berlin, 1996, pp. 869-873. [18] KANERVA, P.: Analogy as a Basis of Computation, In Uesaka, Y., Kanerva, P., Asoh, H.: Foundations of real-world intelligence. CSLI Publications, Stanford, CA, 2001. [19] FODOR, J.A., PYLYSHYN, Z.W.: Connectionism and cognitive architecture: A critical analysis., Cognition, 28 (1988) 3-71. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 328 A NOTE ABOUT NON-LINEAR FUNCTION APPROXIMATION WITH FUNCTION BOUNDARY Radomil Matoušek Institute of Automation and Computer Science, Brno University of Technology, Technická 2, 616 69 Brno, Czech Republic. Institute of Scientific Instruments, Academy of Sciences of the Czech Republic, Královopolská 147, 612 64 Brno, Czech Republic. [email protected] Abstract: The approximation of engineering data by means polynomial techniques is widely known as well as advantages and disadvantages these techniques. This paper discus different method of approximation based on special kind of function denoted as Radial Basis Function (RBF). Next, this paper proposes to embed a genetic algorithm (GA) in traditional learning algorithm of Radial Basis Function Networks (RBFN). Keywords: Genetic Algorithms (GA), Radial Bases Function (RBF) and approximation with function boundary. INTRODUCTION In engineering applications existing approximation tasks, which cannot be solving by means traditional mathematical techniques based on polynomial regression. Cause agent is a physical substance of problem. For example, absolute temperature has not go bellow zero, deformation resistance must by positive, and so on. In these cases the behaviour of polynomial can be unacceptable. This paper recommended special kind of RBF functions to use. Approximation method is take over from neural network denoted as RBFN. This paper proposes to embed a genetic algorithm in the traditional learning algorithm of radial basis function networks. Each function centre, function widths and output an RBF network is encoded as a binary string, and the concatenation of the strings forms chromosome. In each generation cycle, the GA determines the parameters of RBFN. RADIAL BASIS FUNCTION NETWORK Radial Basis Function Networks (RBFN) are new and extremely powerful type of feedforward artificial neural network. A RBFN are very often used to implement the mapping f :  ®  n M æ x -cj f k (x) » wk 0 + å wkj F ç ç sj j =1 è m such that ö ÷ = yk (x), 1 £ k £ m ÷ ø where x is an input vector in  , yk ( x) is the kth network output, wkj denotes the weight connecting hidden n node j and output node k, cj denotes the jth function centre , and s j denotes the jth function width. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 329 Common forms of basis function are: 2 F ( r ) =e - r F ( r ) =r 2 log r (Gaussian) (thinplate spline) -1/ 2 (inverse multiquadratic) F (r )= r 2 +c 2 (multiquadratic) 1/ 2 2 2 F (r )= r +c ( ( ) ) ü ï ï ý, r ï ïþ ³0 PHASE A PHASE B GA training The topology of RBFN and RBF unit can by describe as follows: INPUT xi A general RBFN and a RBF on the planar feature space. OPTIMISATION AND GENETIC ALGORITHMS The optimization task by GA is assumed to be: max {F (a) | a Î {0,1}l } , min { f (a) | a Î {0,1}l } , where F is objective function denoted as fitness function – satisfies F(a)>0, "a. For next discussion we will assume minimization optimization task with objective function f(a). The current simple GA consists of n binary strings a (individual), each of them of length l. This set of n individuals is denoted as population P. Each individual a is a feasible solution to problem. The transition between successive populations (Pi,Pi+1,…) is achieved by applying the genetic operators of selection S, crossover C and mutation M. This transition is denoted as a generation (in numerical terminology as an iteration). The simple GA with a generation model was used. The short formal description of the GA is: Pi +1 = g ( Pi , x ), x Î ( S , C , M ) , where Poº pseudo-random or heuristic setting. As we know, selection, crossover and mutation operators determine genetic algorithm’s behavior. Selection operator has influence on the convergence behavior of GA. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 330 TEST PROBLEM As test example of the RBF function approximation and GA learning, we used a set of data points from measuring of deformational resistance of steel. y 150 Polynomial 100 50 0 0 0.2 RBF 0.4 0.6 0.8 1 x 1.2 1.4 1.6 1.8 2 CONCLUSION This paper shows method for function approximation, which combines the RBFN and GA, to find the parameters of RBF networks. The result was support by means of simple statistic. D MSE 100 % 90 % Approximation The best polynomial RBF GA learning Training time < 1s 5s Different power of polynomial was confronted with RBF. MSE statistics don’t show demonstrably better quality of RBF. But power of RBF solution was presented in always-suitable solutions. Id est., the function values fulfilled the condition f ( x ) ³ 0 . 1 A different power of polynomial 0.8 0.6 y Polynomial function is unacceptable 0.4 0.2 0 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 x Center of RBF unit REFERENCES [1] OŠMERA, P.: An Application of Genetic Algorithms with Diploid Chromosomes, Proceedings of MENDEL’98, p. 86 – 89, Brno, Czech Republic, 1998. [2] GOLDBERG, D. E.: Genetic Algorithms in Search, Optimization and Machine Learning, Wesley, 1989. [3] LIGHT, W.: Advances in Numerical Analysis, Oxford Science Publications, Oxford, 1992. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 331 ACKNOWLEDGEMENTS The paper was supported by research designs CEZ: J22/98: 261100009 “Non-traditional methods for investigating complex and vague systems”, VZ MŠMT ME 526 andVZ MŠMT No: 260000013. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 332 ANALYSIS METHODS FOR IMAGE PRE-PROCESSING AND 2-DIMENSIONAL PATTERN RECOGNITION Roman Weisser Orlová, 73514, Czech Republic, [email protected] Abstract: Subject of my thesis is the analysis of algorithms for pre-processing of visual data and object recognition in 2-dimensional technological scenes. This work was developed in the research project CEZ: J22/98: 261100009 „Non-traditional methods for investigating complex and vague systems“. ITRODUCTION Image pre-processing: Methods for image pre-processing used for improvement image from look next processing. Subject image pre-processing is reduce noise created at digitalizing and picture transmision, remove distortion existent property sensor arrangement, or reduce or emphasize other charakteristics at standpoint next processin. Methods used for image pre-processing: transmission to greyscalle, histogram equalization, filtration, image sharpening, contrast and brightness correction Segmentation: Segmeting is procedure which disperze on part, having near context with objects. Method used for segmenting picture: threshold, edge detection, colorings Object description: Next levels pattern recognition is description found region. Method for object description: symptomatic description, syntactical description Pattern recognition: Pattern recognition consist of sorting into the clases. Class is subset objects, whose elements have common feature from classification stadtpoint. Object have physical character, which in computer vision understand most frekquently parts sectional picture. TESTING METHOD Testing single method was implementation on computer: Microprocessor: AMD Thunderbird 1132 @ 1267MHz, 266DDR RAM: 256MB / 266, CL 2 Methods for image pre-processing action Sharpening time (s) Roberts operator Laplace operator Coloring transformation Greyscalle Brightness, transformation Segmentation contrast Brightness, contrast Histogram Equalization Threshold 0,105 0,12 0,055 0,05 0,125 0,02 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 333 This methods was testing on scene: Width: 512 pixels, Height: 512 pixels Number of objects: 5 Coloring table: coloring objects type 1 2 number of 3 objects 4 5 bitmap size 512 * 512 1024 * 1024 randomly coincident coincident 0,06 0,05 0,19 0,07 0,06 0,25 0,1 0,07 0,3 0,131 0,07 0,32 0,161 0,171 0,391 Coloring table: From introduced table is resulting, that coloring method is depending not only on sizes picture, but on recontouring objects. Type of object named as coincident contain coincident objects, consistently turning and beyond they are set so as to with minimizeusing table colours equivalence. Objects named as randomly contain randomly turning objects and each other variety. Filtration table: Filtration Averaging mask mask 3 mask 5 mask 7 mask 9 time (s) 0,661 1,161 1,783 2,553 0,451 Median 0,1 0,5 1 1,5 2 time (s) 0,1 0,12 0,14 0,171 0,2 Rotaring mask 1 5 Number of 10 iteration 15 20 time (s) 0,14 0,681 1,362 1,993 2,724 Gauss filter sigma This table is summary of method for filtration scenes. Fastest method for filtration picture is Gaussian filter, which produced also very good results. Parameter sigma for Gaussian filter indicating peace smudge scenes. Filtration mask means size array for filtration. Whereby bigger is this array, thereby more neighbouring pixels is into filtration inclusion thereby happen also to bigger smudge (smooting). But as shows table averaging, so at using bigger filtration mask greatly increasing time for filtration. METHODS FOR EDGE DETECTION INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 334 In this part are comparsion single edge detectors. Introduced method are comparsion according to few charakteristics: Safe edge: Indicating cleannes edge. Contain residual noise, If edge width has everyvere identical count pixels, atd. Doubling edge: Critize edge reduplication or double display object with displacement. Noise: Critize cleannes scenes. Thickness edge: This item critize edge width throughout edge length. Recontouring preservation: Indicating formative modification object after edge detection. Erasure edge: Has been watched mistaken erasing all section edge. Last item inform about setting input values. From results is resulting, that fastest method for edge detection is zero crossing operator. But resultant scene contains considerable noise. This is posible precede, if entrance for this method will thresholding modified picture. Prewitt operator, Laplace operator, Gradient operator, Zero corossing table: Second fastest method is gradient method. Results this method is picture has less noise, than zerro crossing operator. If entry for this method is thresholded picture, then generate 2 pixel width edge (ideal for next processing). Next method are Izotropic operator, Sobel operator, Prewitt operator. This method have approximately coincident exit and differentiate only sometimes edge detecting. Fastest response has the Sobel operator, then Izotropic operator and on last places is Prewitt operator. Sobel, Izotropic operator table: Has been waylaid, that accelerated calculation at Sobel operator and Izotropic operator do not affect markedly output picture and this method accelerate about double. This method can cleanness separate real edge INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 335 of noise and thereby is resulting scene almost without noise. Next tested method is Laplace operator. This method os slowest method of all allude to method and on exit these method is too large noise. If major criteria is not speed detection, then for beforehand unpreparation scenes of these mentioned method is most suitable use Prewitt operator. Next method is deviation method. When in use mask 3 ´ 3, generate this method very first-rate edge. Deviation method has some handicap and though, that is slow in comparsion with competetive operators. When in use bigger masks happen to expansion lines and increase time requirement and this scene after edge detection must skirt thinning action (erosion method). Deviation method table: Next method for edge detection is Canny detector, which is using in medicine in machine computer termography (CT). It is very first-rate edge detector, generating edges width 2 pixels and beyond smooth. With reductive sigma soars speed edge detection, however in the same way quickly soars noise in resulting picture. With increasing sigma happen to modification shape single objects in picture. Sharp edges these objects passage in curve. That is one from reasons what for this detector suitable to detection objects, that are representation largely curved shapes. Modification shape is extra percetible at classification grammar, which is sensitive to any diversity. In the event of, that is finding optimal set this detector, then produces results executing all criteria with first-rate results. Only handicap of these detector is speed. In comparsion with different here tested methods be ranked among slowest algorithm. Canny detector table: Next edge detector is Rothwell edge detector, which detection edge in the same way first-rate as a Canny detector, but exit picture hasn`t been shape modification. But this method at increasing value sigma trimming margin scene and csequently near margin may be so much modification, that will badly classified. This edge detector leave sharp quoins edges and that is why is suitable for syntactical method image recognition. Edge detected this methom have 2 pixels width edges. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 336 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 337 Rothwell detector table: INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 338 Wavelet transformation table: Symbols in following tables: Single methods was classification reciprocal collating. METHOD FOR CLASSIFICATION OBJECT This methods constitute last and upper-most step in theory computer vision. This method was collation each other: Recognition with the aid of moment, Recognition with the aid of method Fourier`s descriptors, Recognition with the aid of gramatic, which describe to edge of object, INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 339 Recognition with the aid of neural network (back propagation). Fastest methods for pattern recognition are recognition with the aid of grammar and neural network. Next method is Fourier`s descriptors and lastly is method recognition with the aid of moment. Recognition with the aid of moment method give verz good results. This method classify faultlessly most object alredy at precept on some model etalon. Moment method it can be use for other edge detector, than for what was moment enumeration, since aren`t too sensitive on changes edge objects. e.g. Canny detector (at higher sigma rounded edge) and Sobel operator. This method is unfit for application expectant recognition minimum shape dissimilar, because this method insn`t sensitive on less shape changes. This method is fit for application that have require fast recognition dissimilar objects in different turning. Next method is Fourier`s descriptors. This method isn`t no-variant in the face of rotation, thereby at classification scenes no-recognition objects, which are otherwise turning than objects in etalon scenes. Problem with turning is possible sort out consecutive tutorage models, but increases hazard creation mistaken classification. This method is faster, then moment pattern recognition and is fit for application no-expectant recognition objects in dissimilar turning. Next method is recognition with the aid of grammar. If primitives marked in singles partitions edge, grammar is very sensitive to small mistakes in edge detection. Grammar is required set together „on peace“ for definite type edge detector. e.g. is fault set together grammar for objects which was detected e.g zero crossing operator and then recognition objects by means of Canny edge detector with sizeable sigma (edge modifing). Grammar fit in application which require recognition objects in dissimilar and when is emphasis on recognition small changes at segment edge. Creation grammar requires time and knowledge grammar description edges. Creation rules for grammar must work manual and unattended automatically as other method. Last tested method is algorithm Back Propagation. This method is the fastest of all collation method. For description objects this method was used 70 symptomatic vector orient of gravity to edge object. This method recognise considerably shape forced objects, but can she faulty identify shape similar objects. This method recogniting dissimilar turning objects. Fault these method growing if constrictive objects. CONCLUSION Subject of my thesis was analysis algorithms for pre-processing of visual data and object recognition in 2-dimensional technological scenes. Methods for pre-processing are difficult on the time, so is suitable use them only necessary eventuality. Every operation on scenes decreases number information contained in this scenes. Use method for preprocessing isn`t necessary always and consists in qualities scenes. If picture contains heavy expenses quantiy noise, that advise insert Gaussian filtration, which is very fast and has been exit good results for minimal s. After implementation filtration is also much simler find threshold in picture, than at non-thresholding picture. If being asked bigger edge quality (e.g. at syntactical recognition, when are primitive single edge partitions), is fit insert before edge detection threshold picture. For threshold segmentation picture unambigously advise gradient edge detection, or zero crossing operator, which in combination with threshold have very good time even qualitative results. Very good results have even Canny edge detector and Rothvel edge detector even without using threshold. Recommended / unredommended methods for pre-processing image: For recognition object isn`t necessary brightness transformation (change brightness and contrast, histogram equlization, etc. are methods suitable when human has qualitatively analyse scenes). For most cases isn`t necessity use higher order masks at averaging method. Recommended / unredommended methods for edge detection: Worst results serve Wavelet transformation, which isn`t in methods, where is requirement accurately determine edge object, on difference for instance of method concentrative on face rocognition. There is no need to use higher order masks at deviation method, because is dilatancy edge width and thereby would necessity for next procedure insert thinning method. Appearances at identical results in quality output picture at Prewitt operator, Sobel operator and Izotropic operator, is possible critize this methods according for the duration of calculation. Thereby I would INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 340 recommended only Sobel operator and Sobel accelerated calculation. Recommended application stocking for pattern recognition methods: Grammar recognition – is fit, when is requirement recognition turning objects and where is requirement detect high accurate single edge partitions, without possibilities creation meaningful mistakes and where is exaction high speed classification. Meaningful mistakes understand mistake, which it is imposible implicate on rules. Recognition with the aid of Fourier`s descriptors – is fit, where is demending high speed classification, without supposed turning objects. Rocognition with the aid of moments – is fit, where isn`t that much important accurate edge course, but be sufficient only „hard“ partition to the single classes. e.g. it does not matter, whether is between the two edge sharp passage, or short curve. Advantage compare to Fourier`s descriptors is invariant in turning objects, to be sure at the expense of slowly calculation. Recognition neuronal network – is fit, whete exaction high speed classification with randomly turning scenes and where we need tolerate differences between learned etalons and classified scenes. LITERATURE [1] WEISSER R., Software pro klasifikaci objektů [thesis], Brno, VUT FSI 2003 [2] KOZÁK M., Využití genetických algoritmů v systémech pro rozpoznávání objektů [thesis], Brno, VUT FSI 2002 [3] MÁLEK M., Programový systém pro rozpoznávání objektů [thesis], Brno, VUT FSI 2002 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 341 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 342 ON THE DISREGARDED FORMS OF COMPUTER CRIME Karel Pichl,1) Markéta Pichlová2) 1) Military Academy in Brno, K 303. Kounicova 65, P.B. 13, 612 00 BRNO, CZECH REPUBLIC Tel. +420-973442336, Fax: +420-973442987, E-mail: [email protected] 2) Gaz de France, Département Etudes Générales; 361, Avenue du president Wilson – B.P. 33 93211 SAINT-DENIS LA PLAINE CEDEX, FRANCE. Tel. +33-1-49225000, Fax +33-1-49225142, E:mail: [email protected] Abstract: This paper reports on the unjustifiably disregarded forms of computer crime, focussing on the present status in the Czech Republic. They include: ”An assault on a computer, program, data, or communication facility; unauthorised use of a computer or communication facility; unauthorised access to data; theft of a computer, programme, data, or communication facility; changing of programs, data and technical equipment, and finally, abuse of information technology to commit a crime of another form.” The paper mentions also some weak points of computer systems and the methods used in an unauthorised access to information systems. Keywords: Computer crime, criminal acts, dematerialised crime, intellectual theft, copyright violation, unauthorised used of a computer or communication facility, unauthorised access to data, weak points of computer systems, methods of unauthorised access to information systems, computer viruses, computer espionage, software piracy. I. INTRODUCTION While ”classical” crime has remained practically unchanged over the centuries, modern technology has brought about also new forms of crime. A murder remains a murder, no matter if the criminal uses a laser gun instead of a slaughter axe. Together with an increasing dematerialisation of the commonly used tools (in particular information, financial tools), a new phenomenon has emerged, which could be termed a dematerialised crime. It is rooted in dishonesty in the fields of business, enterprise, banking and insurance. It exploits computer systems with databases filled with information, the knowledge or the manipulation of which bestows significant advantages. It is committed by persons of a good and very high level of expertise. Dematerialised crime is a result of dishonesty of people with a high level of responsibility and powers, working at positions enjoying trust – the ”white collars”, as the clerk positions are sometimes called. In committing the crime, they use their brain and often highly sophisticated methods. This type of crime does not carry any signs of violence as it is commonly understood, but is almost always based on an illegal manipulation or fraud of the offender. It is reasonable to say that if these culprits would exert the same energy, creativity and smartness in doing a business, they would be highly successful. A prevailing motive of these crimes is a profit, a personal gain. Dematerialised crime is very lucrative – yielding hundred, thousand or more times the proceeds of the ”classical” crime. However, there may be also other motives or combinations thereof: · The feeling of superiority over the employer, police, or public, the feel of impunity and of being undiscoverable. · The belief that small losses cannot harm the company. · A wish to compensate dissatisfaction about work or personal life. · A feeling of being exploited or injured by the employer. · Desire for adventure and risk. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 343 The victims of this crime are mostly large banks and multinational companies (which often have no interest in making public the information that a fraud or other threat has taken place), or the public funds. In both cases, the injured parties are major operators of computers and in particular of computer networks and longdistance data transmissions. In such systems, there is a greater anonymity of access to the network due to a greater openness of the network and a variability of possible ways of access. In a majority of cases of dematerialised crime, modern technologies are involved that in fact enable the very existence of the intangible assets. Therefore, they fit very well as the objects of as well as tools for this type of crime. Regarding the computer, the dematerialised crime may be divided into two basic categories: - offences where the computer (program, data etc.) is a tool, - offences where the computer (program, data etc.) is the target. Examples of the first category include credit card frauds, bank account frauds, withholding business proceeds in a commerce, etc. The other group covers software piracy, hacking, unauthorised use of computers (theft of processor time), etc. However, a typical feature of dematerialised crime is that its various aspects and forms are combined, including an international character of the crime. Computer crime is in general very little discussed in the literature and the authors of the few articles that there are concentrate almost exclusively on software right or copyright or licence and their application on software. And that is too little. Completely neglected, other forms of computer crime remain disregarded, including: · An assault on a computer, data or communication facility. · Unauthorised use of a computer or communication facility. · Unauthorised access to data, retrieving confidential information (computer espionage) or other information on persons. · Theft of a computer, data or communication facility. · Changing of programs and data (and sometimes of the technical connection of a computer or communication facility). · Abuse of information technology to commit a crime of another form. · Frauds committed in relation to information technology. In case that any criminal activity related to information technology takes place, it is indispensable to seek an assistance of a computer expert who can prevent the deletion of ”mined” data, restore deleted information, prevent further spreading of a virus, etc. II. WEAK POINTS OF COMPUTER SYSTEMS In general, there are six major ”weak points” of computer systems, which may be abused to commit crime. They are: · system software, · user programs, · databases, · terminals, · telecommunication connections, · microcircuits. The perpetrators of computer crimes focus very often on computer software, targeting both the operation system or testing programs ensuring and controlling the operation of the computer, and software serving for solving of particular tasks. Such software may be stolen or degraded in some way. Databases concentrate a large amount of information to be further processed by a computer. The objective of the offenders is to capture the information and use it to their advantage or to degrade or destroy the information in order to inflict significant damages on the owner of the information. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 344 Terminals are essential components of mainframe computers and computer networks. They allow for communication of the user with a machine that may be a few meters or hundreds of kilometres far away. In a terminal network, the computers are mostly connected by telephone lines. The perpetrators seek to get access to terminals or they connect secretly to the respective connection lines. They are thus able to obtain input and output information from the computer processing and sometimes even intervene with the user programs. The telecommunication lines also give an opportunity for an illegal connection to the network. Another vulnerable part of computers are microcircuits, where undesired actions are difficult to identify. Making use of the LSI and VLSI technologies employed, ”secondary records” of the information processed may be created. III. METHODS OF UNAUTHORISED ACCESS TO COMPUTER SYSTEMS The literature reports on a number of methods of illegal access to computer systems. More than fifty methods have been identified. Each of the methods usually has various modifications and variations, some methods are used combined. In principal, the methods may be divided into: · Active unauthorised access methods (direct methods), where the offender deliberately exercises an activity seeking in particular to obtain information processed in the computers. · Passive unauthorised access methods (may be qualified as unintentional, however, this often involves disguising of the true intentions). Active unauthorised access methods Active unauthorised access methods include in particular: · destruction or theft of data, · the Trojan horse method, · method of alienating small amounts from various sources, · unauthorised use of application software, · misuse of asynchronies in computer processing, · misuse of identifiers and passwords, · misuse of transmission lines, · time bomb, · logical bomb, · information purging method, · wiretapping, · role sharing, · misuse of the failure and repair tasks, · simulation and modelling. Passive unauthorised access methods To degrade or misuse the processed information, these methods rely on: · damaging of memory media, · malfunction of reading (writing) equipment, · wrong marking or misnaming of a file, program or privilege, · modification or replacing of data or a program, · incorrect user instructions, · wrong user address, · switching of terminals (redirection of output to another terminal), · error in identification (of files, terminals, authorised users), · failure to observe applicable organisational guidelines, · input of wrong information into the computer memory, · external effects (e.g. a magnetic field). INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 345 Due to a limited scope of this paper, the individual methods will not be discussed in detail here (an attentive reader can anyway appreciate the operational applications of these methods). The methods will be shortly introduced in our conference presentation. For more details, the reader is referred to the literature in the field. And the following part of this paper is dedicated to methods that are often being unjustifiably disregarded. IV. UNJUSTIFIABLY DISREGARDED FORMS OF COMPUTER CRIME 1. Assault on a Computer, Program, Data or Communication Facility In considering this case, it is necessary to recognise that there are two areas of actions that are or rather may be combined. An assault on the computer hardware and accessories (theft, damage, etc.) and an assault on the information stored in the computer. An example of a combination of the two is a theft of a notebook with valuable software including text and data files storing e.g. a large work of literature, a technical specification, or the complete information system of a company. As regards an assault on the computer hardware as a material, the case is quite clear since the penal law does not make a difference between a computer and another thing. In this respect, a computer may be stolen, damaged or illegally used. Probably the simplest to describe are offences of the category of physical assaults on computer facilities, magnetic media, computer network lines, etc. However, the issue is how to determine the price (value) of the damaged data or any other damages occurring or threatening as a result of such a crime. While the price of a computer, its part or a magnetic medium as such is rather easy to determine, assessing the value of information (data stored) is a difficult task. There are two factors relevant for the price of information: the costs expended in obtaining and compiling the information, and the value that the information may have for its owner (or user). It must be admitted that the second factor – i.e. the value for the owner – belongs to an area requiring much of further work. Indeed, there exists no established methodology and any experiments will only lay the very foundations for any general approach. Logical assaults are much more sophisticated. The offenders directly use the properties (or the weaknesses) of the computer system. Examples include in particular the so called logical bombs (programs activated under certain conditions and performing certain actions) and viruses (harmful programs dragged in from the exterior by a deliberate action or failure to observe rules in handling of diskettes and in remote communication), which delete the memory content, block computer operation by performing a large number of useless operations, etc. In this area, there are two categories of offenders: those that are proud of their authorship, and those who remain cautiously clandestine. Also crimes committed by negligence of a programmer may occur, because not every programmer can appreciate all effects that may cause his program to behave as a logical bomb or a virus. Unfortunately, the law does not recognise a damage to other person’s thing due to negligence. 2. Unauthorised Used of a Computer or a Communication Facility The second category of computer crime includes mainly two types of actions: calculation on a computer, use of a program or communication facility belonging to another person without its knowledge or permission – i.e. an unauthorised use of computers. Probably the most typical offence of this type committed from ”inside” are private business activities of programmers. The statistics show that over a half of the perpetrators of computer crimes are data processing experts – system programmers, managers, technical staff. The legal awareness of this staff is often inversely INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 346 proportional to their commitment to IT work and their fascination by technical and software capabilities. A member of staff who has good knowledge, skills and an interesting work in which he has remarkable success may still not be or not feel being sufficiently remunerated. In such a situation, an opportunity, an offer or other trigger may occur that leads to a crime. And with people who deliver little work for little money, the limit is even easier to trespass. In all of the above cases, two typical features of computer crime are present: · Computing on a computer (and using the software) of the employer. · Selling of programs created within an employment contract to other users on the offender’s account and in his name. While the case b) carries clear signs of a crime, the case a) is usually judged very tolerantly by the environment (not to mention the offender). It is always necessary to distinguish between a failure to observe working obligations, a violation of owner rights, and a crime. It is practically impossible to track unauthorised use of a personal computer that is not connected to a computer network and is used by a single user. The operations are not documented in any way because personal computers usually do not use any registration of processor time. Larger computer systems (mainframes) or computer networks usually allow for determining that some tasks were carried out and in most cases also for identifying the user who launched them. However, the difficult issue is to find out whether the use was appropriate or unauthorised. 3. Unauthorised Access to Data In this area, covering unauthorised access to data, retrieving of confidential information (computer espionage) or information that are in other countries protected based on personality protection laws, a crime may in practical cases be proved only after such information is actually used. On the other hand, it may be difficult to penalise database hacking, in particular when done through a use of a modem. The essential problem is to prove an intent to cause a damage, as it is required by the law. Indeed, many hackers perform their activities only as an intellectual hobby, having no material or other ambitions, and not needing to share their enjoyment with anyone. They may well be satisfied by knowing that they are smarter that the designer of the security system of the computer. However, it may also happen that a hacker succeeds to make a particularly successful hack, and as he browses the database, he may take the decision to cross over beyond the line of a hobby and misuse the information obtained, usually against some kind of redress. However, at the time he gained access to the data carrier, he had no intention to misuse the information. And the illegal use of such information is yet another issue. We doubt that the ministry of the interior or defence or a major bank would be happy to find out that some post-pubescent spends the nights browsing through their databases. But unless they can establish and prove that he used or at least intended to use the information, they can do nothing (at least as far as the penal law is concerned) but quietly improve their data protection system. A mere copying of data (and programs) without using it is practically impossible to disclose. Investigating and proving such acts is then rather difficult, as there is usually a significant time lag between illegal obtaining of the information and its misuse. The only evidence is once again the above mentioned activity tracking and registration. More than ten years ago, on 29.4.1992, Act 256/1992 Coll. on protection of personal data in information systems was passed. This legislation provides for the protection of personal information; in particular, it stipulates the obligations and responsibilities related to protection of information in the operation of any information system that processes personal data of individuals. The law is constructed entirely in terms of a civil code, not stipulating any penal liability, but defining the claims (Art. 22) arising to the injured party in case of a breach of obligations on the part of the operator, intermediary or another liable person. 4. Theft of a Computer, Program, Data or Communication Facility INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 347 While a theft of a tangible property in this fourth category of crime remains a theft, the so called ”intellectual theft” consisting in a copying of a program or data is once again more difficult to qualify legally. In the former case, the perpetrator must physically take away the things, while on the latter case, the original objects remain at their place and the perpetrator takes away only there copies (or alternatively, he may remove the originals and leave copies). In this respect, it is I most case impossible to find out whether a digital record is an original or a copy. The only evidence available is secondary evidence, such as invoices, delivery certificates, witness testimony, printer outputs or files in magnetic memory and in particular the registered date of their creation. Such software piracy is in the Czech Republic as well as in most post-communist and developing countries regarded by many as a completely normal way of a acquiring the program needed. In this respect, the provisions of the copyright law should be recalled, stipulating that also a computer code may be subject to a copyright. This has important repercussions also in the penal law area. The cited copyright law precisely stipulates the right of the author, so that it is likely that not only the person spreading programs in a pirate way, but also anyone having such programs on his computer runs a risk of a penal prosecution. The value of a software may be determined based on the selling price or otherwise (based on expert opinion), so that everyone is able to calculate not only how much he has saved, but also how much he will be liable to pay if it is found that the software used has been obtained illegally. And to divulge the software illegality, it is enough that there is one unsatisfied employee or former employee. Because everyone knows very well that legal software comes with original manuals, original diskettes and licence certificates, while illegal programs are characterised by the fact that thee is either no manual or it is photocopied. The nature of programs, their structure and other evidence available usually provide no means to determine conclusively which is the original product and which is a copy. 5. Changes in Programs, Data and Technical Equipment This fifth category includes in particular: · Changing of programs and data by other programs (viruses) or by direct programmer intervention (which is common in particular in computer assisted fraud). · Less commonly, an alteration of the connection or another attribute of a technical equipment or communication facility. In case of changes of programs or data, it is possible (next to the above mentioned possibilities) to try to find older or deleted versions of files on magnetic media and make a comparison. However, it is advisable so charge a qualified professional with his task. Therefore, it is necessary to ensure that the whole computer or at least the magnetic memories are untouchable. This can be ensured also by removing the hard disk from the computer and depositing it in a safe deposit box. An alteration of the technical connection is a crime that is easier to disclose and document and where it should not be a problem to gather evidence on the case. The most appropriate evidence is the equipment itself. Detailed documentation of original connection and assembly diagrams is also indispensable, as well as signals and tests history. As has been noted above, investigating assaults on the information content of computers for penal law purposes will obviously pose a greater problem. Despite several unclear stipulations, the law penalises virtually all unauthorised interference with software. In the authors’ opinion, this should include also viruses, regardless of whether the virus destroys or just damages the data stored. Also in this respect, a lack of penal liability for acts in negligence is apparent, because a co-worker who introduces a highly destructive virus to a computer through e.g. a computer game may inflict much more damage than someone who deliberately makes pirate copies. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 348 6. Misuse of Information Technology to Commit Other Crime The easiest and therefore probably the most wide-spread way to illegal incomes using IT is a manipulation of data – goods on stock, revenues, health insurance and number of employees, fulfilment of the plan, account balance, etc. Unlike the manipulation of hard-copy documents, the manipulation of electronic data has several advantages for the offender: · Deletion or replacement of data on magnetic media is much easier and leaves practically no comebacks. · People (inspector, customer, etc.) psychologically a priori consider computer outputs reliable and trust them (though subconsciously). The latter aspect results in a high success of offences committed using information technology. It s precisely the high level of trust in computer outputs that renders IT frauds successful. And nothing is easier than to ”improve” the outputs in this sense. Indeed, who is going to re-calculate columns of numbers added up by a computer or by just a simple calculator with an output slip printer? There are dozens of thousands of people who were attracted to participate in a financial ”game” of a pyramid type by the slogan ”No need to look for followers – everything is controlled by a computer”. Data may be changed by manipulating the documents based on which they are recorded, by altering the data on the storage media, in the course of a computation performed by the computer, or in the printer output. Making a change in the already printed output is technically most tricky. However, if also the printer output is saved as a file, the task becomes rather easy. With less qualified staff, the most common form of a fraud is a manipulation of the input documents (i.e. inputting wrong data). A co-worker who processes his admin on a computer – whether it is a PC or a terminal of a mainframe machine – may practically arbitrarily change data prior to, during and after the processing. However, he or she must obviously take care in some way of the remaining information systems of the company, with which there are logical links. The main sources of evidence for criminal offences are records kept correctly and in full, in compliance with the stipulations of the Law, carrying out of all required bookkeeping and other operations, and comparing the results of manual and computer processing. If offences of this type are suspected, it is necessary to secure all hard copy documents as well as to record the current status of data and programs in the computer. If the computer cannot be put off service, it is necessary to archive its status to provide evidence for experts. 7. Frauds Committed in Relation to Information Technology As a result of the assaults mentioned above, yet a more involved criminal activity may take place. The common criterion to judge on a crime taking place is in general the presence of the factual content of the criminal act of fraud. For example a programmer who copies and then sells on his account and in his name a program developed within an employment contract commits several crimes: unauthorised use of other person’s thing (use of a computer to copy the program, safe for the issues listed in section Unauthorised use…), a theft of or a misappropriation (of the source code), violation of copyright (of the employer or the remaining members of the authors team), a fraud (in selling to other persons), and possibly also unauthorised business activity (in case he or she does not possess a trade licence for the activity). It is obvious that a number of these cases is often at the limit of legality and the impunity of the offenders stands and falls with the results of an expert analysis. A typical analysis includes e.g. an investigation of how much the product sold in the programmer’s name differs from the product developed in the employer’s organization. A frequent opinion on this case is that it is a theft of know-how, be it a program or a database. The authors are rather careful to take a conclusive standpoint, as there exists no clear definition of what all is included under the term know-how. In some cases, the legal aspects of the case are highly complicated and in many cases, it is difficult to establish or provide sufficient evidence for all facts. INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 349 Similarly, as a result of an unauthorized access to a computer network by a hacker, a manipulation of the data on account balance in a bank or the results of elections may take place (the criminal act of thwarting the course or the results of elections). V. CONCLUSIONS This paper has no ambitions, within its limited scope, to describe the issues discussed in an exhaustive manner. However, it should alert the managers about the unjustifiably disregarded issues of computer crime and activate them to combat it. The struggle against all forms of computer crime deserves much higher and coordinated attention (than it receives so far from e.g. the lawmakers or the computer users), because the damages suffered are significant an may be very extensive, and the methods and tools of the perpetrators are of a high technical and intellectual level. The investigation is usually very complicated and unless evidence is secured in time and team cooperation established between the prosecutors and the specialists from the Criminological Institute and other experts, little success can be expected in capturing and convicting of the offenders. And finally, a further pressing issue is a systematic approach to a possible amendment of the Penal Code with respect to the factual content of computer crime. REFERENCES [ 1 ] ANDREW S. TANENBAUM: Computer Networks. Third Edition. Prentice-Hall International Inc. .Internationale Edition, 1996 [ 2 ] PICHL,K. A KOL.: Výstavba, řízení a provoz informačních pracovišť AČR I.VA, Brno 2003 [ 3 ] BÍMOVÁ, A.: Počítačová kriminalita a naše doba. IDG Czechoslovakia, Praha 1990 [ 4 ] HONOMICHL, V.: Problematika ochrany osobních počítačů proti používání neoprávněnými osobami. MAA č. 8/1989 [ 5 ] SMEJKAL,V. – SOKOL, P.: Počítačová kriminalita a její trestněprávní aspekty. Softwarové noviny 6.4.1993 [ 6 ] DALER, T. – GULBRANDSEN, R. – MELGARD B. – SJOLSTAD, T.: Security of Information and Data. Horwood Ltd., USA, 1989 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 350 AUTHOR INDEX A L ACHALOVÁ, L.V. .................. 47 AZIZ, A.H.A. ......................... 245 LACKO, B. ............................163 LAŠŠÁK, V. .............................89 B M BARANSKI, W.M. ................ 169 BARTONĚK, D. ..... 199, 239, 255 BAUMGARTNER, U. ........... 155 BEDNÁŘ, J. .......................... 193 BERCHIK, H. .......................... 61 MAGELE, CH. .......................155 MARČEK, D. .........................107 MAREK, P. ..............................99 MATOUŠEK, R. ....................329 MAZURKIEWICZ, J. ..... 119, 177 MIKULA, V. ..........................113 C, Č N ČECHMAN, J. ....................... 261 NESVADBA, J. ......................255 CH CHAKRABORTHY, K.U. ..... 137 O D, Ď ORESHKIN, A.G. ....................77 OŠMERA, P. ....................15, 137 ĎAĎO, J. ........................... 83, 89 DOSTÁL, P. ...................... 11, 93 P PÁVÓ, J. ................................221 PETRUCHA, J. ......................113 PICHL, K. ..............................343 PICHLOVÁ, M. .....................343 PIVOŇKA, V. ........................125 PONOMAREVA, M.A. ............65 POSPÍCHAL, J. ......................305 E EMARY, I.M.M. .............269, 297 H HITMÁR, Š. ............................ 31 HLAČINA, T. .......................... 89 HUSSEIN, A.S. ..............229, 245 R JANKAL, R. ............................ 53 JEDLIČKA, M. ........................ 35 JELÍNEK, P. .......................... 291 JURÁK, P. ............................. 285 RADOVÁ, J. ............................99 RAIS, K. ...................................11 RENHART, W. ......................155 REŠKOVÁ, M. ...................27, 41 RUKOVANSKÝ, I. ........ 137, 239 K S,Š J SEIFOULLAEVA, E. ...............57 ŠEVČÍK, V. ............................277 SMUTEK, D. ..........................145 STRIŠŠ, J. ................................27 ŠVEC, P. ................................215 KALUGIN, E. .......................... 73 KARPÍŠEK, Z. ...............285, 291 KHATTAB, D.R. ................... 245 KOBEL, J. ............................. 119 KOSCIUSZKIEWICZ, K. ...... 119 KRATOCHVÍL, O. .................. 15 KUCHARČÍKOVÁ, A. ............ 21 KVASNIČKA, V. .................. 305 T TOKARČÍKOVÁ, E. ................69 TOLBA, M.F. .........................229 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 351 WEISSER, R. .........................333 WODA, M. .............................205 V VELEBA, V. .......................... 125 VÉRTESY, G. ........................ 221 VODÁK, J. .............................. 21 Z, Ž ŽÁK, L. ............................93, 131 ZAMOJSKI, W. ......................185 W WALKOWIAK, T. ..169, 185, 205 INTERNATIONAL CONFERENCE ON SOFTCOMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENTS ICSC 2004 EPI Kunovice, Czech Republic, January 29-30, 2004 352 10 Název: ICSC 2004 – Second international conference on soft computing applied in computer end economic environments Autor: Kolektiv autorů Vydavatel, nositel autorských práv, vyrobil: Evropský polytechnický institut Osvobození 699, 686 04 Kunovice Náklad: 120 ks Počet stran: 352 Rok vydání: leden 2004 ISBN 80-7314-025-x 11