The unfavourable effects of hash coding on CMAC convergence and compensatory measure

The unfavourable effects of hash coding on the convergence of CMAC (Cerebellar Model Articulation Controller) learning are investigated in detail, based on the fact that CMAC learning is equivalent to the Gauss-Seidel iteration for solving a linear system of equations. A set of theoretical results are obtained concerning the convergence of CMAC learning. It is pointed out that hash coding may give rise to divergence, or at least deteriorate the convergence behavior, and the causes of such phenomena are revealed in a matrix-theoretic approach. We propose a compensatory measure which is shown to be effective in minimizing the unfavourable effects of hash coding by simulation