Evaluation and Enhancement of Bayesian Rule-Sets in a Genetic Algorithm Learning Environment for Classification Tasks

The paper describes an inductive learning environment called DELVAUX for classiication tasks that learns PROSPECTOR-style, Bayesian rules from sets of examples. A genetic algorithm approach is used for learning Bayesian rule-sets, in which a population consists of sets of rule-sets that generate oosprings through the exchange of rules, permitting tter rule-sets to produce oosprings with a higher probability. Reward and punishment mechanisms are introduced that evaluate the performance of a Bayesian rule within a rule-set. For this purpose, fuzzy techniques that evaluate the "goodness" of a rule within a rule-set are provided. A new mutation operator is introduced that uses this evaluation information, which replaces bad rules with higher probabilities than good rules, when a mutation occurs. Empirical results that evaluate the presented reward-punishment techniques are presented. Finally, we compare our learning environment that learns fuzzy rules that rely on decision making by evidence combination with classical classiier systems that rely on non-fuzzy, data-driven rules and the bucket brigade algorithm.