In this thesis, a multi-agent cooperation system for robot soccer is designed and developed based on modular Q learning method. In the multi-agent system, action selection is important for the cooperation and coordination among agents. There are many techniques in choosing a proper action of the agent. As the environment is dynamic, reinforcement learning is more suitable than supervised learning. Reinforcement learning is based on the trial and error method through the experience. The experience of the agent is used for the reinforcement learning. Because of this, reinforcement learning is applied to many practical problems. However, straightforward application of reinforcement learning algorithm may not successfully scale up to more complex multi-agent learning problems. To solve this problem, the modular Q learning is employed for the multi-agent learning among the reinforcement learning methods. The mediator module of the modular Q learning selects the final action of the agent considering the Q value obtained by each learning module. To obtain better performance, the action selection of mediator module is modified by considering the state information. The effectiveness and applicability of the modified method are verified by real experiments.