Automatic generation of fuzzy set membership functions and rules using a neural network

Abstract

In generating fuzzy systems, the primary task is to extract and adjust membership functions and fuzzy rules. However, using traditional hand coding methods, the amount of this work expands startlingly with the increasing number of variables. The paper presents a neural network based algorithm to automatically extract membership functions and rules for fuzzy systems. The complicated input-output relationship is firstly decomposed into the accumulation of simple input-output relationships. For each variable, a set of membership functions that are appropriate for all simple input-output relationships are generated, and multiple sets of fuzzy rules that reflect its efficacy on every simple input-output relationship are also extracted. The fuzzy rules for the whole system are then generated based on these sets of fuzzy rules. In that way, we obtain the membership functions and fuzzy rules of the whole system. Because a complicated problem is decomposed into the accumulation of simple ones, the complexity of its solution will not expand startlingly with the increasing number of variables and the algorithm can be put into practice.