Applications of Learning Classifier Systems

308 Pages·2004·10.718 MB·English

by  Larry Bull (auth.), Larry Bull (eds.)

1. Bull (Ed.)
Applications of Learning Classifier Systems
Springer
Berlin
Heidelberg
New York
Hong Kong
London
Milano
Paris
Tokyo
Studies in Fuzziness and Soft Computing, Volume 150
Editor-in-chief
Prof. Janusz Kacprzyk
Systems Research Institute
Polish Academy of Sciences
ul. Newelska 6
01-447 Warsaw
Poland
E-mail: [email protected]
Further volumes of this series Vol. 140. A. Abraham, L.c. Jain. B.J. van der
Zwaag (Eds.)
can be found on our homepage:
Innovations in Intelligent Systems, 2004
springeronline.com ISBN 3-540-20265-X
Vol. 141. G.C. Onwubolu, B.Y. Babu
New Optimzation Techniques in Engineering,
Vol 130. P.S. Nair 2004
Uncertainty in Multi-Source Databases, 2003 ISBN 3-540-20167-X
ISBN 3-540-03242-8
Vol. 142. M. Nikravesh, L.A. Zadeh, V. Korotkikh
Vol 131. J.N. Mordeson, D.S. Malik, N. Kuroki (Eds.)
Fuzzy Semigroups, 2003 Fuzzy Partial Differential Equations and
ISBN 3-540-03243-6 Relational Equations, 2004
ISBN 3·540-20322-2
Vol 132. Y. Xu, D. Ruan, K. Qin, J. Liu
Lattice-Valued Logic, 2003 Vol. 143. L. Rutkowski
ISBN 3-540-40175-X New Soft Computing Techniques for System
Modelling. Pattern Classification and Image
Vol. 133. Z.-Q. Liu, J. Cai, R. Buse Processing. 2004
Handwriting Recognition, 2003 ISBN 3-540-20584-5
ISBN 3-540-40177-6
Vol. 144. Z. Sun. G.R. Finnie
Vol 134. V.A. Niskanen Intelligent Techniques in E-Commerce, 2004
Soft Computing Methods in Human Sciences, 2004 ISBN 3-540-20518-7
ISBN 3-540-00466-1
Vol. 145. J. Gil-Aluja
Vol. 135. J.J. Buckley Fuzzy Sets in the Management of Uncertainty.
Fuzzy Probabilities and Fuzzy Sets for Web 2004
Planning, 2004 ISBN 3-540-20341-9
ISBN 3-540-00473-4
Vol. 146. J.A. Gamez. S. Moral. A. Salmeron (Eds.)
Vol. 136. L. Wang (Ed.) Advances in Bayesian Networks. 2004
Soft Computing in Communications, 2004 ISBN 3-540-20876-3
ISBN 3-540-40575-5
Vol. 147. K. Watanabe. M.M.A. Hashem
Vol. 137. V. Loia, M. Nikravesh, L.A. Zadeh (Eds.) New Algorithms and their Applications to
Fuzzy Logic and the Internet, 2004 Evolutionary Robots. 2004
ISBN 3-540-20180-7 ISBN 3-540-20901-8
Vol. 138. S. Sirmakessis (Ed.) Vol. 148. C. Martin-Vide. V. Mitrana. G. PilUn
Text Mining and its Applications, 2004 (Eds.)
ISBN 3-540-20238-2 Formal Languages and Applications. 2004
ISBN 3-540-20907-7
Vol. 139. M. Nikravesh, B. Azvine, I. Yager, L.A.
Zadeh (Eds.) Vol. 149. ).). Buckley
Enhancing the Power of the Internet, 2004 Fuzzy Statistics. 2004
ISBN 3-540-20237-4 ISBN 3-540-21084-9
Larry Bull (Ed.)
Applications
of Learning
Classifier Systems
, Springer
Larry Bull
University of the West of England
Faculty of Computing,
Engineering & Mathematical Sciences
BS16 lQY Bristol
United Kingdom
E-mail: [email protected]
ISSN 1434-9922
ISBN 978-3-642-53559-8 ISBN 978-3-540-39925-4 (eBook)
DOI 10.1007/978-3-540-39925-4
Library of Congress Cataloging-in -Publication-Data
A catalog record for this book is available from the Library of Congress.
Bibliographic information published by Die Deutsche Bibliothek.
Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliographie;
detailed bibliographic data is available in the Internet at http://dnb.ddb.de
This work is subject to copyright. All rights are reserved, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitations, broadcasting, reproduction on microfilm or in any other way, and storage in data
banks. Duplication of this publication or parts thereof is permitted only under the provisions
of the German copyright Law of September 9, 1965, in its current version, and permission for
use must always be obtained from Springer-Verlag. Violations are liable for prosecution under
the German Copyright Law.
Springer-Verlag is a part of Springer Science+Business Media
springeronline.com
© Springer-Verlag Berlin Heidelberg 2004
Softcover reprint of the hardcover 1st edition 2004
The use of general descriptive names, registered names trademarks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
Typesetting: Camera-ready by editor
Cover design: E. Kirchner, Springer-Verlag, Heidelberg
Printed on acid free paper 62/3020/M -543 2 1 0
Foreword
The field called Learning Classifier Systems is populated with romantics. Why
shouldn't it be possible for computer programs to adapt, learn, and develop while
interacting with their environments? In particular, why not systems that, like
organic populations, contain competing, perhaps cooperating, entities evolving
together? John Holland was one of the earliest scientists with this vision, at a
time when so-called artificial intelligence was in its infancy and mainly concerned
with preprogrammed systems that didn't learn.
Instead, Holland envisaged systems that, like organisms, had sensors, took
actions, and had rich self-generated internal structure and processing. In so doing
he foresaw and his work prefigured such present day domains as reinforcement
learning and embedded agents that are now displacing the older "standard Af' .
One focus was what Holland called "classifier systems": sets of competing rule
like "classifiers", each a hypothesis as to how best to react to some aspect of
the environment--or to another rule. The system embracing such a rule "popu
lation" would explore its available actions and responses, rewarding and rating
the active rules accordingly. Then "good" classifiers would be selected and re
a
produced, mutated and even crossed, la Darwin and genetics, steadily and
reliably increasing the system's ability to cope.
This breathtaking vision-certainly as romantic as any in science, since it
dares to relinquish control and leave the machine to its own devices, inspirations,
and fate-was noticed by some, and it inspired them to try to work the vision
out and test it. The way was not easy, because Holland's vision was-and still
is-vast. Always alert to progress in biology, economics, and everything to do
with what he calls "complex adaptive systems", Holland saw phenomena and
mechanisms he knew must be included in a realistic adaptive system. Still, the
others bit, and over the past twenty years, they gradually brought concreteness
via distillation, some revamping, and much experiment-to Holland's vision.
The ongoing results are now called Learning Classifier Systems (just a name
change-the learning was always there). There are several LCS versions, there
is solid theory, and there is an increasingly challenging range of environments
that can be adapted to. In addition, applications of LCS have begun to appear,
where the systems' ability to respond rapidly to ongoing environment (problem,
process) changes, and, through the rules, to capture environmental structure and
"show the knowledge", are leading advantages.
Larry Bull, one of the field's best-known and most innovative researchers, has
contributed significantly to our current understanding of LCS. In this book Dr.
Bull puts together a fascinating selection of applications of LCS theory and know
how in such domains as data mining, modeling and optimization, and control.
Now, as has occurred in other fields of science, Learning Classifier Systems-long
in gestation-are beginning to benefit from the challenge of increasing contact
with the real world.
Stewart W. Wilson
v
Contents
Learning Classifier Systems: A Brief Introduction ................................................... 3
Bull
Section 1 - Data Mining
Data Mining using Learning Classifier Systems ...................................................... 15
Barry, Holmes 6-Llora
NXCS Experts for Financial Time Series Forecasting ............................................ 68
Armano
Encouraging Compact Rulesets from XCS for Enhanced Data Mining. .•............. 92
Dixon, Corne 6-Oates
Section 2 - Modelling and Optimization
The Fighter Aircraft LCS:
A Real-World, Machine Innovation Application .................................................. 113
Smith, EI-Fallah, Ravichandran, Mehra 6-Dike
Traffic Balance using Learning Classifier Systems
in an Agent-based Simulation ................................................................................ 143
Hercog
A Multi-Agent Model of the UK Market in Electricity Generation .•................... 167
Bagnall
Exploring Organizational-Learning Oriented Classifier Systems
in Real-World Problems .......................................................................................... 182
Takadama
VII
Section 3 - Control
Distributed Routing in Communication Networks using the Temporal
Fuzzy Classifier System - a Study on Evolutionary Multi-Agent Control ...... 203
Carse, Fogarty & Munro
The Development of an Industrial Learning Classifier System
for Data-Mining in a Steel Hop Strip Mill ............................................................. 223
Browne
Application of Learning Classifier Systems to the On-Line Reconfiguration
of Electric Power Distribution Networks ............................................................... 260
Vargas, Filho & Von Zuben
Towards Distributed Adaptive Control for Road Traffic Junction Signals
using Learning Classifier Systems .......................................................................... 276
Bull, Sha'Aban, Tomlinson, Addison & Heydecker
Bibliography of Real-W orId Classifier Systems Applications ............................. 300
Kovacs
VIII
Data Mining
Learning Classifier Systems: A Brief Introduction
Larry Bull
Faculty of Computing, Engineering & Mathematical Sciences
University of the West of England
Bristol BS16 lQY, U.K.
[email protected]
[Learning] Classifier systems are a kind of rule-based system with general
mechanisms for processing rules in parallel. for adaptive generation of new
rules. and for testing the effectiveness of existing rules. These mechanisms make
possible performance and learning without the "brittleness" characteristic of
most expert systems in AI.
Holland et aI., Induction, 1986
1. Introduction
Machine learning is synonymous with advanced computing and a growing body of
work exists on the use of such techniques to solve real-world problems [e.g.,
Tsoukalas & Uhrig, 1997]. The complex and/or ill-understood nature of many
problem domains, such as data mining or process control, has led to the need for
technologies which can adapt to the task they face. Learning Classifier Systems (LCS)
[Holland, 1976] are a machine learning technique which combines reinforcement
learning, evolutionary computing and other heuristics to produce adaptive systems.
The subject of this book is the use of LCS for real-world applications.
Evolutionary computing techniques are search algorithms based on the
mechanisms of natural selection and genetics. That is, they apply Darwin's principle
of the survival of the fittest among computational structures with the stochastic
processes of gene mutation, recombination, etc. Central to all evolutionary computing
techniques is the idea of searching a problem space by evolving an initially random
population of solutions such that better - or fitter - solutions are generated over time;
the population of candidate solutions is seen to adapt to the problem. These
techniques have been applied to a wide variety of domains such as optimization,
design, classification, control and many others. A review of evolutionary computation
is beyond the scope of this chapter, but a recent introduction can be found in [Eiben &
Smith, 2003]. In LCS, the evolutionary computing technique usually works in
conjunction with a reinforcement learning technique.
Reinforcement learning is learning through trial and error via the reception of a
numerical reward. The learner attempts to map state and action combinations to their
utility, with the aim of being able to maximize future reward. Reward is usually
received after a number of actions have been taken by the learner; reward is typically
L. Bull (ed.), Applications of Learning Classifier Systems
© Springer-Verlag Berlin Heidelberg 2004
delayed. The approach is loosely analogous to what are known as secondary
reinforcers in animal learning theory. These are stimuli which have become
associated with something such as food or pain. Reinforcement learning has been
applied to a wide variety of domains such as game playing, control, scheduling and
many others. Again, a review of reinforcement learning is beyond the scope of this
chapter and the reader is referred to [Sutton & Barto, 1998].
Learning Classifier Systems are rule-based systems, where the rules are usually in
the traditional production system form of "IF state THEN action". Evolutionary
computing techniques and heuristics are used to search the space of possible rules,
whilst reinforcement learning techniques are used to assign utility to existing rules,
thereby guiding the search for better rules. The LCS formalism was introduced by
John Holland [1976] and based around his more well-known invention - the Genetic
Algorithm (GA)[Holland, 1975]. A few years later, in collaboration with Judith
Reitman, he presented the first implementation of an LCS [Holland & Reitman,
1978]. Holland then revised the framework to define what would become the standard
system [Holland, 1980; 1986]. However, Holland's full system was somewhat
complex and practical experience found it difficult to realize the envisaged
behaviour/performance [e.g., Wilson & Goldberg, 1989]. As a consequence, Wilson
presented the "zeroth-level" classifier system, ZCS [Wilson, 1994] which "keeps
much of Holland's original framework but simplifies it to increase understandability
and performance" [ibid.]. Wilson then introduced a form of LCS which altered the
way in which rule fitness is calculated - XCS [Wilson, 1995]. In the following
sections, each of these LCS is described in more detail as they form the basis of the
contributions to this volume. A brief overview of the rest of the volume then follows.
·2. Holland's LCS
Holland's Learning Classifier System receives a binary encoded input from its
environment, placed on an internal working memory space - the blackboard-like
message list (Figure 1). The system determines an appropriate response based on this
input and performs the indicated action, usually altering the state of the environment.
Desired behaviour is rewarded by providing a scalar reinforcement. Internally the
system cycles through a sequence of performance, reinforcement and discovery on
each discrete time-step.
The rule-base consists of a population of N condition-action rules or "classifiers".
The rule condition and action are strings of characters from the ternary alphabet
{0,1,#}. The # acts as a wildcard allowing generalisation such that the rule condition
1#1 matches both the input 111 and the input 101. The symbol also allows feature
pass-through in the action such that, in responding to the input 101, the rule IF 1#1
THEN 0#0 would produce the action 000. Both components are initialised randomly.
Also associated with each classifier is a fitness scalar to indicate the "usefulness" of a
rule in receiving external reward. This differs from Holland's original implementation
[Holland & Reitman, 1978], where rule fitness was essentially based on the accuracy
of its ability to predict external reward (after [Samuel, 1959]).
2