A fuzzy CMAC structure and learning method for function approximation

Abstract

A fuzzy CMAC (cerebellar model articulation controller) structure is proposed in this paper. The basis functions in the original CMAC are replaced by fuzzy membership functions for smoothing the network's output and increasing the approximation ability in function approximation. A two-overlay structure of the fuzzy CMAC with the membership functions of different receptive fields is employed. These receptive fields are determined by the distribution of the training data. The proposed structure can reduce the memory requirement a great deal in the original CMAC, especially in high-dimensional structures, and can maintain the same performance as the original CMAC. Furthermore, the issue of generalization parameter selection and the need for considerable training data for updating all of the weightings in the CMAC can be solved in the proposed structure. A sinusoidal function approximation example is illustrated in order to compare the new fuzzy CMAC structure with the original one.