An Improved Reinforcement Q-Learning Method with BP Neural Networks in Robot Soccer

In traditional reinforcement Q-Learning method, there exists two problems: difficulty of dividing the state information, complexity of extreme large dimension input. To solve these two problems, this paper proposed an improved reinforcement Q-Learning method with BP neutral network. In this method, the large Q table is replaced by a BP neural network. Continuous environmental information is the input. The Q value is the output. The Q value and weight of the network are also adjusted by the action rewards. This paper presents an algorithm for single agent´s action selection. Simulation shows proposed method is more stable and applicable for the agent´s strategy selection.