Structure Design of Fuzzy Neural Networks Based on Recursive Clustering and Similarity

Facing the structure design problem of fuzzy neural networks (FNNs), this paper proposed a structure design approach based on the recursive clustering and similarity methods. First, a recursive clustering method to identify FNN structure was proposed. Guided by the strength of output variations and using the recursive sub-clustering as the means, the proposed method determined the initial network structure through recursive iterations. Second, maintaining a high accuracy, the method calculated the similarity degree between each pair of fuzzy rules and then merged highly similar rules to simplify the initialized structure of the FNN. Finally, numerical experiments in function approximation and nonlinear system identification were used to verify the feasibility and effectiveness of the proposed approach.