A hybrid learning algorithm with a similarity-based pruning strategy for self-adaptive neuro-fuzzy systems

An algorithm for the generation of a TS-type neuro-fuzzy system is presented. There are two stages in the generation: in the first stage, an initial structure adapted from an empty neuron or fuzzy rule set, based on the geometric growth criterion and the ɛ-completeness of fuzzy rules; in the second stage, the obtained initial structure is refined by a hybrid learning algorithm based on backpropagation and a proposed recursive weight learning algorithm to minimize the system error. The similarity analysis applied throughout the entire learning process attempts both to alleviate overlap among membership functions and to reduce the complexity of the obtained system. Benchmark examples, comparing the proposed algorithm with previous approaches, show the proposed algorithm is more effective in terms of both model accuracy and compactness.