RESEARCH ON A HIGH-ORDER CMAC NEURAL NETWORK

In this paper, a high order CMAC(HCMAC) neural network is proposed, in which the high order activation functions are utilized as the receptive field functions. The method of address mapping used by CMAC is adopted in the new network. Because of enhancement of the input pattern, the physical address in HCMAC is reduced highly, and by using HCMAC the approximation of complex functions can be obtained which is more continuous than using CMAC and has analytic derivatives. As a result of these characters, the computing amount and learning time are reduced more than RBF neural networks. This paper also introduces originally the Kalman filter algorithm to the CMAC-like networks, so learning effectiveness is improved further. By simulating, it is proved that HCMAC is feasible in many fields.