Learning Strategies for Coordination of Multi Robot Systems: a Robot Soccer Application

This paper presents a hybrid method for learning a dynamic strategy for a robot soccer team. In this method, an imitation learning scheme based on observed robot soccer games is used as a seed for an experience-guided learning scheme based on reinforcement learning. A lack in the application of classic reinforcement learning to the robot soccer problem is the high number of states to be analyzed. In order to reduce this number, the proposed method was implemented using state abstraction and it was applied only in states and actions obtained by imitation. Thus, the algorithm was not applied over all states and actions as the traditional form of reinforcement learning. This approach was shown to be a good option for implementing a robot soccer team that improves its own performance while it acquires more experience. Keywords— Robot Soccer, Learning by Experience, Imitation Learning, Reinforcement Learning.