일반화된 CMAC 신경망의 민감도 특성

Abstract

Algorithm for calculation of derivative of trained GCMAC(Generalized CMAC) is proposed. If GCMAC is directly differentiated, the accuracy of derivative is not satisfactory. This is because of the quantization of input space and the shape of basis function used. A closer look into the output of GCMAC enables to obtain more accurate derivative of trained GCMAC. Because the output of GCMAC has periodic property, finite difference algorithm with finite difference set to the period of GCMAC gives excellent results. SISO and MISO examples show that the proposed algorithm gives almost no errors in the derivative calculations of trained GCMAC.