CMAC neural network has two advantages: the local generalization and no local maximum value. Currently, ICA-CMAC and FCMAC models are used extensively. However, the two models cannot reasonably characterize the direction and magnitude of network weight in the weight correction algorithm. To solve the problem, an improved CMAC learning algorithm is proposed. It takes iterative errors, iteration number and a window function as the performance. Based on information fusion strategy, it introduces global information into the calculation to optimize the network weight. Through a simulation test, it can be found that the model has significant improvement in terms of convergence speed and prediction control.