The local structure of CMAC (cerebella model articulation controller) neural networks results in faster learning of nonlinear functions. However, the learning accuracy of CMAC is too low to meet the requirements of application in many fields. Hence, an associative interpolation algorithm is proposed in this paper for improving the learning accuracy of CMAC. Meanwhile, a simulation experiment is described. Its result shows that the learning accuracy of the improved CMAC is ten times higher than that of the original CMAC, and the learning convergence is also faster.