Information theoretic fitness measures for learning classifier systems

During the course of evolution in a genetic algorithm, and in particular a Learning Classifier System, we conjecture that there is benefit in using multiple fitness functions including information theoretic fitness measures. We discuss multiple ways to evaluate rules and groups of rules, some utilizing the data, and others without. For example, to evaluate the fitness of any rule against a set of data, we propose the use of information theoretic fitness measures Sufficiency and Quality, which are derived from the classical concepts of entropy and mutual information, in order to assign fitness values to classifiers. We present the experimental results for one of the methods, viz. evaluating individual rules against the data showing that their use reduces the number of generations needed to achieve peak performance.