A recursive SVD-based self-constructing rule generation for neuro-fuzzy system modeling

We propose a recursive SVD-based self-constructing rule generation (RSVD-SCRG) approach for structure identification in neuro-fuzzy system modeling. Fuzzy clusters are generated incrementally, with none at the beginning, by presenting the training data one by one. For each presented data, we evaluate its input similarities and output similarities to existing clusters. If the data is not similar enough to any of existing clusters, a new fuzzy cluster is created and its corresponding Gaussian input membership functions and TSK-type linear output function are initialized. Otherwise, we combine the data into the most similar existing cluster and update the corresponding membership functions and output function with statistical calculations and a recursive SVD-based least squares estimator. Therefore, our approach solves the parameter estimation problem encountered in processes of cluster generation and updating. Besides, experimental results have shown that our approach generates more precise initial fuzzy rules and produces lower approximation errors than other approaches.