Classifier system learning of the Boolean multiplexer function

Classifier systems are dynamically adapting production rule systems that modify their rules through experience. This thesis is an empirical study of classifier system learning applied to a particular Boolean reinforcement learning problem, the multiplexer function. The multiplexer function is a special case of a concept expressable in a disjunctive normal form, the k + 2^k multiplexer is expressable as a k + 1 disjunctive normal formula. A variety of modifications to the classifier system are investigated to induce rapid learning. Some of the methods tried axe dynamic population sizes, controlled initial population generality, bit flipping, relative complement, dynamic operator control, and evaluation on demand. Experimental results suggest the following: the initial population generality influences learning performance; relative complement is a more effective operator than bit flipping; and dynamic operator control, and evaluation on demand axe necessary options for learning in the larger k size problems.