Abstract: | A description of the problems, successes and failures encountered whilst attempting to encourage a Classifier System to learn to play a simple board game well. Classifier Systems are a kind of free-for-all Production Rule System where the pattern-matching rules compete on the basis of their (modifiable) strength values, and the population of rules is altered by a Genetic Algorithm. They have shown promise in problems where there is very little specific, (i.e. useful) information available from the environment, and the internal adjustments proceed without explicit direction from the environment (or the programmer). In this thesis an attenpt is made to 'coerce' a variant of Goldberg's Simple Classifier System to 'learn' how to play a simple board game (called Dodgems). Various options were tried, among them were: different internal representations, adding more powerful move operators, forcing every move to be valid, and others... The results, whilst not startling, do indicate increased performance with the use of the enhanced move operators over the initial representations. Larger population sizes appear to be beneficial. Also, there is a discussion of the problems involved in choosing the relevant data to study the internal workings of the Classifier System.
|