Learning Efficient Communication in Cooperative Multi-Agent Environment

Reinforcement learning in cooperate multi-agent scenarios is important for real-world applications. While several attempts before tried to resolve it without explicit communication, we present a communication-filtering actor-critic algorithm that trains decentralized policies which could exchange filtered information in multiagent settings, using centrally computed critics. Communication could potentially be an effective way for multi-agent cooperation. We supposed that, when in execution phase without central critics, high-quality communication between agents could help agents have better performance in cooperative situations. However, information sharing among all agents or in predefined communication architectures that existing methods adopt can be problematic. Therefore, we use a neural network to filter information between agents. Empirically, we show the strength of our model in two general cooperative settings and vehicle lane changing scenarios. Our approach outperforms several state-of-the-art models solving multi-agent problems.