A Study of the Reasoning Methods Impact on Genetic Learning and Optimization of Fuzzy Rules

Abstract

A Genetic Algorithm based learning procedure was proposed earlier by the authors and the results, advantages and usefulness of the proposal have been reported in the literature. The procedure is based on the Pittsburgh approach and is divided in two separate phases: learning of candidate rules and selection of relevant rules. The learning and the optimization processes involve an evaluation function that considers the performance of the candidate rule base, requiring the selection and use of a particular reasoning method. With the objective of investigating further the robustness and usefulness of the previous approach, the authors developed a comparative study of the GA learning algorithm focusing on the impact of the reasoning method used. Two different methods were used: the one based on a single winner rule, and the one based on the combination of all rules. Following the description of the rules format and reasoning methods used, the GA learning and optimization approach proposed before is also reviewed. The comparison of simulation results is presented based on the criteria of correct classification rates and number of rules in the rule base. The results demonstrate that the knowledge base performance is similar in both cases, suggesting that the GA learning procedure derives good rule bases despite the reasoning method used.