Alternatives for Classifier System Credit Assignment

Abstract

Classifier systems are production rule systems that automatically generate populations of rules cooperating to accomplish desired tasks. The genetic algorithm is the systems' discovery mechanism, and its effectiveness is dependent in part on the accurate estimation of the relative merit of each of the rules (classifiers) in the current population. Merit is estimated conventionally by use of the bucket brigade for credit assignment. This paper addresses the adequacy of the bucket brigade and provides a preliminary exploration of two variants in conjunction with enumerated rules and with discovery. In limited experiments, a variant that combines the bucket brigade, classifier chunking, and backwards averaging has yielded improved performance on simple maze problems. Tentative similarities between this hybrid and Sutton's Adaptive Heuristic Critic (AHC) are suggested.