A hybrid evolutionary search concept for data-based generation of relevant fuzzy rules in high dimensional spaces

Abstract

We propose a hybrid fuzzy-evolutionary system for fuzzy modelling in high dimensional search spaces. The system architecture is based on a Michigan-style approach (one individual represents one fuzzy rule). The design of the evolutionary algorithm makes use of a distance measure in the search space that in turn reflects some heuristic assumptions about the fitness landscape. Additionally, strategy parameters are dynamically adapted by means of a fuzzy controller. The approach is successfully applied to a complex benchmark problem as well as to several real-world modelling tasks such as the cancellation behaviour of insurance clients and the classification of automatic gearboxes.