Techniques in Neural-Network-Based Fuzzy System Identification and Their Application to Control of Complex Systems

Abstract

This chapter discusses the identification of fuzzy systems with artificial neural networks. By using updated version of the pi-sigma neural network, both premise and consequent parameters of the fuzzy system can be efficiently identified online or offline. Methodologies in artificial neural networks, fuzzy systems, and evolutionary computation are successfully combined and new techniques called soft computing or computational intelligence is developed. Learning algorithms for both Gaussian and triangular forms of membership functions are presented. The consequent part of the fuzzy rules is represented by a subnetwork, which enables the algorithm to be applicable to high-order Takagi-Sugeno fuzzy systems. Some measures are taken to preserve the interpretability of the fuzzy system in the course of learning. The proposed method is applied to the nonlinear decoupled control of robot manipulators and satisfactory simulation results are obtained. One common problem for most neuro-fuzzy algorithms is that the interpretability of fuzzy systems is deteriorated. After adaptation, either the distinguishability among the fuzzy subsets in a fuzzy partitioning is blurred, or the fuzzy partitioning of the input space is incomplete.