Multi-rule-set Decision-making Schemes for a Genetic Algorithm Learning Environment for Classification Tasks

Over the last three years, we developed an inductive learning environment called DELVAUX for classification tasks that learns PROSPECTOR-style, Bayesian rules from sets of examples, using a genetic algorithm to evolve a population consists of rule-sets. Several problems complicate the search for the best rule-set. First, the search space that is explored by DELVAUX is enormously large, which makes it difficult to predict if a particular solution is a good solution. The second problem is the problem of convergence with outliers that perform well in training but not in testing. This paper describes efforts to alleviate these two problems centering on multi-rule-set learning techniques that learn multiple rule-sets and proposes several decision-making schemes that are employed by the multi-rule-set learning environment to derive a decision. Empirical results are presented that compare the single rule-set learning environment of DELVAUX with several multi-rule-set learning environments that use different decision-making schemes. Moreover, a more sophisticated fitness function for the multi-rule-set learning approach is introduced, and a genetic algorithm approach that finds the "best" multi-rule-set for a given set of rule-sets is discussed.