Towards a Theory of Strong Overgeneral Classifiers

Abstract

Learning Classifier Systems (LCS) typically use a genetic algorithm (GA) to evolve sets of if then rules called classifiers to determine their behavior in a problem environment. This chapter analyzes the concept of strong overgeneral rules, the Achilles' heel of traditional Michigan-style learning classifier systems, using both the traditional strength-based and newer accuracy-based approaches to rule fitness. It argues that different definitions of overgenerality are needed to match the goals of the two approaches, presents minimal conditions and environments which support strong overgeneral rules, demonstrates their dependence on the reward function, and provides some indication of what kind of reward functions will avoid them. Finally, it distinguishes fit overgeneral rules, describes how strength and accuracy-based fitness differ in their response to fit overgenerals and concludes by considering possible extensions to this work.