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
YCS

• A Simple Accuracy-based Learning Classifier
  System
• Larry Bull
• Learning Classifier Systems Group Technical
  Report UWELCSG03-005
• University of the West of England,
• Bristol, BS16 1QY, U.K.
• larry.bull_at_uwe.ac.uk
• Learning Classifier Systems use evolutionary
  algorithms to facilitate rule-discovery, where
  rule fitness is traditionally payoff based and
  assigned under a sharing scheme. Most current
  research has shifted to the use of accuracy-based
  fitness, after the introduction of XCS, where
  rule fitness is based on a rules ability to
  predict the expected payoff from its use. Whilst
  XCS has been shown to be extremely effective in a
  number of domains, its complexity can make it
  difficult to establish clear reasons for its
  behaviour. This paper presents a simple
  accuracy-based learning classifier system with
  which to explore aspects of accuracy-based
  fitness in general. The system is described and
  modelled, before being implemented and tested on
  the multiplexer task.

3
YCS Parameters

• k2 multiplexer.
• problems60000
• moving.avg50
• multiplexer.runs10
• classifiers400
• beta0.2 p0.6
• initial.payoff10.0
• initial.error10.0
• initial.nichesize10.0
• g0.25
• theta25.0
• chi0.5
• mu0.01
• init.rulesettrue
• Can be lowered to 16 parameters
• http//www2.cmp.uea.ac.uk/it/ycs/ycs.html

4
YCS Concept

• Accuracy based LCS, similar to ZCS
• WH Delta rule for payoff, niche size estimate and
  error
• Panmictic GA and covering,
• A probability of GA invocation
• Simple roulette wheel, 2 parents on inverse of
  error f 1 / (e 1)
• Explore/exploit scheme used
• Replacement based on an estimate of a niche size
• Not maximally general mapping as not triggered a
  niche GA

5
Condensation schemes

• Once the final rule base has been trained, then
  a separate process removes unnecessary rules
• A Preliminary Investigation of Modified XCS as a
  Generic Data Mining Tool
• Phillip William Dixon, David Corne, Martin J.
  Oates
• July 2001 IWLCS '01 Revised Papers from the
  4th International Workshop on Advances in
  Learning Classifier Systems, Publisher
  Springer-Verlag
• Use of supervised learning to reduce population
• Compact rulesets from XCSI
• Wilson, S. W
• Presented at the Fourth International Workshop on
  Learning Classifier Systems (IWLCS-2001), San
  Francisco, CA, July 7, 2001.
• Compact Ruleset Algorithm achieved through
  iteration