Thoughts on Selection Methods in Michigan-style Classifier Systems

1
Thoughts on Selection Methods in Michigan-style
Classifier Systems

• When solving optimization problems it is
  desirable to replace good solutions whereas in
  LCS replacing a very good classifier is quite
  dangerous, because the systems real-time
  performance might drop significantly after a few
  good rules have been replaced.
• Usually, the amount of change in LCS is
  significantly less compared to to EC systems that
  solve optimization problems, especially after the
  system has evolved over a certain amount of time,
  and the environment is somewhat stable.
• Niche methods are much more attractive for LCS
  style system, because designing a classifier
  systems can be viewed as finding a good set of
  collaborating rules rules can be viewed as
  experts on what to do if a particular situation
  occurs. You are looking for a team not super-fit
  individuals.
• Coverage is another important aspect of
  classifier systems that does not exist in
  tradition EC systems a rule set has to be
  evolved that covers many different situations.
  Having a few good rules with narrow expertise
  might result in a system that guesses for most
  inputs.

2
Other Problems in Designing Classifier Systems

1. What strength/fitness value should be given to
   new rules? Giving them average values or slightly
   below average values seems to be a good strategy.
   Another problem with new rules is that they might
   get deleted prior to being used often enough to
   evaluate their usefulness.
2. What about duplicate rules? No clear answer to
   this question. For example in systems in which
   decisions are made using voting redundancy is
   desirable. Moreover, redundancy makes it less
   likely that good genetic material is lost. On
   the other hand, copy of the same rule can produce
   a significant overhead and might reduce
   situation coverage.
3. How often should we learn, and how much change
   should we apply to the current rule-set if we
   learn? I am afraid that this is a question that
   can only answered empirically.
4. What should I do if there is no rule that covers
   a given input? Obviously, you do not know what
   action to take. Moreover, generating a new rule
   that covers this situation might not be a bad
   idea in an XCS-style system.

3
Other Problems in Designing Classifier Systems
(cont.)

• How should a generate my initial rule-set? What
  mutation operator should I use? Generate the
  dont cares dependent on the rule-set size, the
  number of possible inputs, and how much coverage
  for a given input you might like to have.
• How should actions be selected, if rules with
  different actions match the input? Many
  approaches are possible here including random
  selection (might not be a bad strategy at the
  beginning of the evolution process), strength
  based (classifiers with higher past payoffs are
  preferred), accuracy based (classifiers whose
  performance is predictable are preferred), simple
  voting (each classifier has one vote),
  strength-weighted, accuracy-weighted, voting,
• How should I evolve strength, fitness, and other
  values based on system feedback? In my opinion,
  the strategies employed in XCS seem to be a good
  choice. Initially, set your learning rate between
  0.1 and 0.3 and adjust it based on experimental
  results. In general, if the experiment is short
  or the environment is changing a lot higher
  learning rates are useful.

4
Other Problems in Designing Classifier Systems
(cont.)

• The learning period seems to be short for my
  system to learn something what should I do?
  Either adjust your system parameters, or if you
  demonstrate that the performance of your system
  improves over longer learning (e.g. you ran it
  for 4x1200 examples) periods this is also good
  enough for the course project.
• How complex should the system be a should I
  design? Dont make your system to complicated
  unless you are a very fast programmer, because
  this will give you enough time for fine tuning
  based on experimental results. Moreover, if you
  understand what your system is doing, it might be
  easier to enhance it.
• Do I have to write a large report? No just a
  short description of the strategies your system
  employs is sufficient.