A Study of Generalization Techniques in Evolutionary Rule Learning

ABSTRACT A STUDY OF GENERALIZATION TECHNIQUES IN EVOLUTIONARYRULE LEARNINGJeffrey K. Bassett, M.S.George Mason University, 2002Thesis Director: Dr. Kenneth A. De JongGeneralization is an important aspect of all machine learning algorithms. Without it,learning can only occur in the simplest of problem domains. The most interesting domainstend to be more complex though. As we scale-up our algorithmsto work in these com-plex domains, an understanding of the different generalization techniques available and thetrade-offs between them becomes increasingly important.This thesis is primarily concerned with rule learning using evolutionary algorithms.We examine two commonly used generalization techniques called wildcards and partialmatching, as well as a third which is a hybrid of these two. We demonstrate that thewildcards are more effective at generalizing in some domains, and partial matching is moreeffective in others. It is our hypothesis that the hybrid will be the more robust of the threetechniques. In other words, the hybrid will generalize as well, or almost as well, as thebetter of the other two in a variety of domains.Two very different domains were chosen as testbeds for our experiments. The first isa concept learning domain, which tends to favor wildcards. The second is a multi-agentrobotics task, where partial matching is more effective. When the hybrid was tested inthese environments, the results show that it was either equivalent or superior to the other