A fuzzy theory refinement algorithm 1

A fuzzy theory refinement algorithm composed of a heuristic process of generalization, specification, addition and elimination of rules is proposed. This refinement algorithm can be applied to knowledge bases obtained from several sources (learning algorithms, experts), but its development is strongly associated with the SLAVE learning system. SLAVE was developed for working with noise-affected systems where the application of some conditions of classical learning theory do not produce good descriptidns. This learning system allows us to obtain the structure of the rule, i.e. it can determine among all the variables proposed those that are relevant for describing the system (feature selection). SLAVE uses an iterative approach for learning with genetic algorithms. This method is an alternative approach to the classical Pittsburgh and Michigan approach and it consists in obtaining a useful rule to describe the system in each iteration. So in this approach, the final solution is obtained from partial solutions. The refinement module appended to SLAVE (SLAVE + R) is proposed as a method for verifying that the union of the partial solutions is a good global solution. Furthermore, this module allows us to minimize the number of necessary rules, maintaining or improving the accuracy and understanding of these rules. © 1998 Elsevier Science Inc. All rights reserved.