Learning Cooperative Intrinsic Motivation in Multi-Agent Reinforcement Learning

The cooperative behavior is important skill in many real-world applications. Recently, many works have used the multi-agent platform to solve the real-world applications. However, it is difficult to learn the cooperative behaviors with equal rewards that the environment provides without considering the contributions. In this paper, we propose a method for learning cooperative behaviors in the centralized multi-agent environment. Firstly, we implement a reward model to predict the average rewards of all agents. And then, we use the reward model for calculating the contributions. The proposed method allows the model to distinguish which agent behaves better for team success. In order to evaluate the performance of the proposed method, we compute the average team rewards on the multi-agent battle environment. Experimental results show that the proposed method has better performance than the baseline using the cooperative behaviors.