Learning Classifier Systems

1
Learning Classifier Systems

• Navigating the fitness landscape?
• Why use evolutionary computation?
• Whats the concept of LCS?
• Early pioneers
• Competitive vs Grouped Classifiers
• Beware the Swampy bits!
• Niching
• Selection for mating and effecting
• Balance exploration with exploitation.
• Balance the pressures
• Zeroth level classifier system
• The X-factor
• Alphabet soup
• New Slants - piecewise linear approximators
• Why don't LCS rule the world?
• Simplification schemes
• Cognitive Classifiers
• Neuroscience Inspirations
• Application Domains

2
International Workshop on Learning Classifier
Systems

• The workshop series was initiated in 1992, held
  at the NASA Johnson Space Center in Houston,
  Texas. Since 1999 the workshop has been held
  yearly in conjunction with PPSN in 2000 and 2002
  and with GECCO in 1999, 2001 and from 2003 to
  2009. Topics of interests include, but are not
  limited to
• Paradigms of LCS (Michigan, Pittsburgh, ...)
• Theoretical developments (behaviour, scalability
  and learning bounds, ...)
• Representations (binary, real-valued, oblique,
  non-linear, fuzzy, ...)
• Types of target problems (single-step,
  multiple-step, regression/function approximation,
  ...)
• System enhancements (competent operators, problem
  structure identification and linkage learning,
  ...)
• Cognitive control (architectures, emergent
  behaviours, ...)
• Applications (data mining, medical domains,
  bioinformatics, ...)

3
Advances at the frontier of LCS

• January 8th, 2007
• Advances at the frontier of Learning Classifier
  Systems has been shipped to Springer for the
  final stages of editing and printing. The volume
  is going to be printed as Springers LNCS 4399
  volume. When we started editing this volume, we
  faced the choice of organizing the contents in a
  purely chronological fashion or as a sequence of
  related topics that help walk the reader across
  the different areas. In the end we decided to
  organize the contents by area, breaking a little
  the time-line. This was not a simple endeavor as
  we could organize the material using multiple
  criteria. The taxonomy below is our humble effort
  to provide a coherent grouping. Needless to say,
  some works may fall in more than one category.
  Below, you may find the tentative table of
  contents of the volume. It may change a little
  bit, but we will keep you posted as soon as we
  learn from Springer.

4
Part I. Knowledge representation

• 1. Analyzing Parameter Sensitivity and Classifier
  Representations for Real-valued XCSby Atsushi
  Wada, Keiki Takadama, Katsunori Shimohara, and
  Osamu Katai 4399 - 001
• 2. Use of Learning Classifier System for
  Inferring Natural Language Grammarby Olgierd
  Unold and Grzegorz Dabrowski4399 - 018
• 3. Backpropagation in Accuracy-based Neural
  Learning Classifier Systemsby Toby OHara and
  Larry Bull4399 - 026
• 4. Binary Rule Encoding Schemes A Study Using
  The Compact Classifier Systemby Xavier Llorà,
  Kumara Sastry , and David E. Goldberg4399 - 041

5
Part II. Mechanisms

• 5. Bloat control and generalization pressure
  using the minimum description length principle
  for a Pittsburgh approach Learning Classifier
  Systemby Jaume Bacardit and Josep Maria
  Garrell4399 - 061
• 6. Post-processing Clustering to Decrease
  Variability in XCS Induced Rulesetsby Flavio
  Baronti, Alessandro Passaro, and Antonina
  Starita4399 - 081
• 7. LCSE Learning Classifier System Ensemble for
  Incremental Medical Instances by Yang Gao,
  Joshua Zhexue Huang, Hongqiang Rong, and Da-qian
  Gu4399 - 094
• 8. Effect of Pure Error-Based Fitness in XCSby
  Martin V. Butz , David E. Goldberg, and Pier Luca
  Lanzi4399 - 105

6
Part II. Mechanisms

• 9. A Fuzzy System to Control Exploration Rate in
  XCSby Ali Hamzeh and Adel Rahmani4399 - 116
• 10. Counter Example for Q-bucket-brigade under
  Prediction Problemaby Atsushi Wada, Keiki
  Takadama, and Katsunori Shimohara4399 - 130
• 11. An Experimental Comparison between ATNoSFERES
  and ACSby Samuel Landau, Olivier Sigaud,
  Sébastien Picault, and Pierre Gérard4399 - 146
• 12. The Class Imbalance Problem in UCS Classifier
  System A Preliminary Studyby Albert
  Orriols-Puig and Ester Bernadó-Mansilla4399 -
  164
• 13. Three Methods for Covering Missing Input Data
  in XCSby John H. Holmes, Jennifer A. Sager, and
  Warren B. Bilker4399 - 184

7
Part III. New Directions

• 14. A Hyper-Heuristic Framework with XCS
  Learning to Create Novel Problem-Solving
  Algorithms Constructed from Simpler Algorithmic
  Ingredientsby Javier G. Marín-Blázquez and Sonia
  Schulenburg4399 - 197
• 15. Adaptive value function approximations in
  classifier systemsby Lashon B. Booker4399 - 224
  
• 16. Three Architectures for Continuous Actionby
  Stewart W. Wilson4399 - 244
• 17. A Formal Relationship Between Ant Colony
  Optimizers and Classifier Systemsby Lawrence
  Davis4399 - 263
• 18. Detection of Sentinel Predictor-Class
  Associations with XCS A Sensitivity Analysisby
  John H. Holmes4399 - 276

8
Part IV. Application-oriented research and tools

• 19. Data Mining in Learning Classifier Systems
  Comparing XCS with GAssist by Jaume Bacardit and
  Martin V. Butz4399 - 290
• 20. Improving the Performance of a Pittsburgh
  Learning Classifier System Using a Default
  Ruleby Jaume Bacardit, David E. Goldberg, and
  Martin V. Butz4399 - 299
• 21. Using XCS to Describe Continuous-Valued
  Problem Spacesby David Wyatt, Larry Bull, and
  Ian Parmee4399 - 318
• 22. The EpiXCS Workbench A Tool for
  Experimentation and Visualizationby John H.
  Holmes and Jennifer A. Sager4399 - 343

9
Utilisation of LCSs

• Perpetually novel events accompanied by large
  amounts of noisy or irrelevant data.
• Continual, often real - time, requirements for
  actions.
• Implicitly or inexactly defined goals.
• Sparse payoff or reinforcement obtainable only
  through long action sequences.
  
• Booker 89
• Main problems found in domains
• - Multimodal
• - Lack of Separation
• - High Dimensionality

10
Known Data
11
Learning Classifier Systems
PARAMETERS-
COILING TEMPERATURE
WRAPPER ROLL SPEED
DOWNCOILER SET-UP
KNOWN CONDITIONS MONITORED MILL PARAMETERS
KNOWN ACTIONS MEASURED MILL OUTPUT
HEAD TELESCOPE
OUTPUT-
COIL QUALITY
COIL PRESENTATION
STRIP SURFACE QUALITY
PINCHING
12
Domains

• Toy problems
• MUXs
• Woods Mazes
• Monks
• Binary Royal Road Max ones
• Reinforcement learning
• Cart-pole
• Mountain-car
• Data mining
• Machine learning repository (MLR)
• Industrial
• Steel mill
• Bioinformatics
• Traffic lights

13
Domains

• Cognitive robotics
• Phototaxis
• Exploration
• Internet
• Music recommendation systems
• Finance
• Stock market prediction