The Research About Learning Algorithm Of Robot Soccer Behavior Control

As one of the grand challenges of artificial intelligence research, robot soccer combines multiple hot researches. It is a standard platform to research artificial intelligence and multi-agent system. The research topics about artificial intelligence research of robot soccer include multi-agent cooperation, behavior strategy decisions of single robot, action optimization, etc. The article focused on action autonomic learning of soccer robot, and introduces reinforcement learning. Series simulations of reinforcement learning about soccer robot actions were designed to demonstrate the reinforcement learning’s feasibility which is used in soccer robot actions.Firstly, the article generally introduced the robot soccer system, and explained multiple intelligence research directions of robot system. The traditional methods of soccer robot actions were introduced. And the shortcomings of these methods were illustrated. To solve these problems, the idea of using reinforcement learning to control soccer robot action was put forward.The paper elaborated the model of reinforcement learning in detail. It introduced the Markov decision process, raised Q learning algorithm aiming at the discrete state space. Then the successive approximation method for the continuous state space was introduced in the application of reinforcement learning. The paper showed the implementation process of reinforcement learning based on multilayer feed forward neural network.To achieve the intercepting action of soccer robot, the reinforcement learning based on Cerebellar Model Articulation Controller (CMAC) was introduced. The CMAC has the characteristics of simple structure, fast learning. The intercepting action simulation of soccer robot was designed to verify the effectiveness of the proposed algorithm. According to the shortage of the CMAC network, the CMAC network had made the improvement to achieve the continuous approximation of the neural network output. The reinforcement learning based on the improved CMAC network was used to the intercepting action simulation. And to achieve the avoiding dynamic obstacles of soccer robot, a reinforcement learning algorithm based on Parallel continuous CMAC was designed. It avoid dimension disaster caused by the input state space of high dimension.Finally, to achieve PID control of reaching target in any given direction, the Actor-Critic learning algorithm was used in PID control. Finally the adaptive PID control simulation that soccer robot reach target point in given direction is realized.