Accuracy, Parsimony, and Generality in Evolutionary Learning Systems via Multiobjective Selection

Evolutionary learning systems (also known as Pittsburgh learningclassifier systems) need to balance accuracy and parsimony for evolving high quality general hypotheses. The learning process used in evolutionary learning systems is based on a set of training instances that sample the target concept to be learned. Thus, the learning process may overfit the learned hypothesis to the given set of training instances. In order to address some of these issues, this paper introduces a multiobjective approach to evolutionary learning systems. Thus, we translate the selection of promising hypotheses into a two-objective problem that looks for: (1) accurate (low error), and (2) compact (low complexity) solutions. Using the proposed multiobjective approach a set of compromise hypotheses are spread along the Pareto front. We also introduce a theory of the impact of noise when sampling the target concept to be learned, as well as the appearance of overfitted hypotheses as the result of perturbations on high quality generalization hypotheses in the Pareto front.