A genetic programming-based classifier system

In this paper we present initial results from the use of Genetic Programming in Holland's learning Classifier System architecture. Rules are of the form condition(s)/action in which the conditions are represented as binary strings and the actions are represented by S-expressions. The basic Classifier System architecture used here is more analogous to Wilson's ZCS system than Holland's original formalism since no internal message list exists; stimulus-response rules are developed here. Using two well-known classification tasks it is shown that our approach can develop useful feature extractors for the K-nearest-neighbour algorithm. We also show that the use of niche-based evolutionary search can improve performance.