Abstract: In order to avoid local optimum of Mamdani model parameter optimization, a novel algorithm for Mamdani neural network was proposed. The initial parameters of Mamdani Fuzzy Neural Network(FNN) were generated by Fuzzy Cmeans clustering, based on PSO, and then optimized by using PSO. Finally, Gradient descent method was adopted for further optimizing the parameters so that the fuzzy rules could be automatically adjusted, modified and improved. Numerical experiments show that the presented algorithm improves the approximation ability of Mamdani FNN.

Key words: particle swarm optimization(PSO); fuzzy cmeans clustering(FCM);fuzzy rules; Mamdani neural networks; optimization; gradient descent method