Learning Communication with Limited Range in Multi-agent Cooperative Tasks

In a multi-agent system, communication is an effective way for multi-agents to cooperate. However, when the number of agents in the environment is large, receiving large amounts of messages requires high bandwidth and results in long latency and high computational complexity. Predefined communication architectures, such as master-slave communication architectures, may help, but they limit communication between specific agents, inhibiting potential cooperation. Therefore, we propose a model, which enables the agent to effectively select communication agents and aggregate communication messages, so as to carry out effective communication. Finally, we demonstrate the advantages of our model in multi-agent cooperative navigation scenarios where agents are able to develop more coordinated and complex policies than existing Methods.