Abstract:The main advantage of CMAC neural network is the partial correction of weight coefficient. As we can change very little weight coefficient to get faster learning ability, CMAC is very suitable for real time control. But in the actual process of using,CMAC often faces the problem occurred by the uneven distribution of units’ credibility, therefore we use CA-CMAC instead of CMAC.Q-learning is an important method of reinforcement learning, In this article ,we combine Q-learning and CA-CMAC neural network and use the algorithm in Robocup simulation for improving the agent’s ability of intercepting .We get good results through the simulation which shows that the algorithm is deasible and effective.