Abstract
About how to coordinate the behaviors of multiple agents, several studies have been made to allow multiple agents to synthesize the coordinated decision policy needed to accomplish their common goals through reinforcement learning effectively. When there are multiple goals in the environment, however, the agents can not obtain the policy that can allow them to split into teams and achieve the goals simultaneously and automatically through reinforcement learning. On the other hand, they attempt to achieve the same goal together and then the next respectively. In this thesis, we use modular reinforcement learning to allow agents to learn not only action-selection but also target-selection in multi-goal cooperative MAS. Through target-selection, we can achieve team partition. We apply our method to hunter-prey problems and the results show that our method has better a performance in the average capturing steps and the convergence speed of learning. We also demonstrate the result of team partition by showing a scenario in the hunting process.