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, a self-organising fuzzy networks training algorithm based on unscented Kalman filter(UKF) is presented. Firstly, a non-linear dynamical system expression of fuzzy networks is analyzed, and RLS and UKF are used to learn linear and non-linear parameters respectively. Secondly, guidelines of how to generate a new rule and update parameters are presented. Then, the method of the error descending rate is used as fuzzy rule pruning strategy, so that the rule which plays an unimportant role in the system is deleted. Finally the typical experiments of function approximation and system identification indicate that the fuzzy network obtained by the proposed algorithm has a more tighten structure and better generalization than other algorithms.