Abstract:

Much of the current research interest in neurofuzzy hybrid systems is focused on how to generate an optimal number of fuzzy rules in a neurofuzzy system and investigate the automated methods of adding and pruning fuzzy rules. To deal with this problem, an adaptive network based fuzzy inference system (ANFIS) based on UKF is presented. Firstly, fuzzy rules and their parameters of ANFISRR are obtained by subtractive clustering. Secondly, the parameters are learned by linear least square and the back propagation algorithm. Thirdly, the nonlinear dynamical system expression of fuzzy networks is analyzed, and LLS and UKF are used to learn linear and nonlinear parameters respectively. Then, a method of error descending rate is used as the fuzzy rule pruning strategy, so that the rule which plays an unimportant role in the system is deleted. Finally, by typical experiments of function approximation and system identification indicate that fuzzy networks obtained by the proposed algorithm has a more tightened structure and better generalization than other algorithms.

Key words: unscented Kalman filter;adaptive network based fuzzy inference system (ANFIS);rule reduction;system identification;function approximation