Evolving granular fuzzy min-max modeling

Abstract

This paper develops an evolving fuzzy min-max algorithm for fuzzy rule-based systems modeling. Starting with an initially empty rule base, the algorithm may add, modify, or delete fuzzy rules of the rule base while processing input stream data. The data space is granulated using hyperboxes. Membership functions and affine functions are assigned to the hyperboxes, and each hyperbox defines a corresponding functional fuzzy rule. The model output is found combining the affine rule consequents weighted by the normalized activation degrees of the rules. The parameters of the consequent affine functions are updated with the recursive least squares with forgetting factor. The algorithm is intrinsically incremental, learns in one-pass, and allows gradual model changes in an online like manner. Computational experiments suggest that evolving granular fuzzy min-max modeling procedure is competitive with state of the art approaches.