Classifier Systems with Hamming Weights

Abstract

Classifier systems are learning systems that use a genetic algorithm to carry out their learning. With the exception of work by Lashon Booker, researchers in the field have employed exact matching procedures when using classifier systems. In this paper we describe three learning problems that create difficulties for classifier systems using exact matching. The first problem requires covering in order for exact matching classifier systems to succeed. The second requires extremely large populations of classifiers for exact matching systems. The third creates difficulties for matching techniques based on Hamming distance - the technique Booker studied - and for exact matching techniques using measures of specificity. We propose a new closeness matching scheme - “weighted Hamming matching” - and show that classifiers matching with Hamming weights are able to encode optimal solutions to the three types of learning problems described without the difficulties encountered by other matching methods.