Neural-network approach to structural tuning of classification rules based on fuzzy relational equations

Abstract

An adaptive approach to structural tuning of classification rules based on solving fuzzy relational equations, allowing to avoid the fuzzy knowledge base redundancy is proposed. The approach lies in building and training the neural-fuzzy backward chaining network, the isomorphic system of fuzzy relational equations, which allows to adjust the structure of fuzzy rules as new experimental data appear. Because of the lack of effective selection methods, there is no single methodological standard for structural tuning of fuzzy rules. Modern neural-fuzzy systems use heuristic selection methods to reduce the number of rules without losing the distinctive ability of the network. It was found that using the neural-fuzzy backward chaining network allows to avoid the knowledge base redundancy while preserving the inference precision. The number of rules in the class is equal to the number of solutions, and the form of membership functions of fuzzy terms is defined by intervals of values of input variables in each solution. This approach is an alternative to the classical approach, based on selection of rules from a set of rules-candidates that prevents obtaining compact fuzzy knowledge bases.