Abstract: Aiming at the problem of physical mapping address collision in the learning algorithm of CMAC(Cerebellar model articulation controller) neural network caused by Hash mapping in the algorithm, a method based on setting a weight overflow area is proposed in this paper. Compared with traditional learning algorithms which solve the collision problem through increasing the size of real space, this method have the advantages of saving physical memory space and improving the precision of network-learning under the conditions of the same size of real space. Finally, the experiment results show that it works well in the applications of nonlinear system identification and color matching.