A Novel Parameter Learning Algorithm for a Self-constructing Fuzzy Neural Network Design

This paper proposes a novel parameter learning algorithm for a self-constructing fuzzy neural network (SCFFN) design. It concludes dynamic prior adjustment (DPA) which is employed to adjust parameters according to the distribution of the input samples and group-based symbiotic evolution (GSE) which is applied to train all the free parameters for the desired outputs. DPA considers the relevance between input samples space and the IF-part parameters, which intends to accomplish coarse adjustment. Then, GSE is adopted to search the global optimum solution. Unlike traditional GA with each gene representing a whole fuzzy system, GSE divides the population into several groups that each one only represents a fuzzy rule. The full solutions can be generated by all possible combinations of the groups. The simulations results have verified that the proposed algorithm achieves superior performance in learning accuracy.