Learning When to Communicate Among Actors with the Centralized Critic for the Multi-agent System

Centralized training and decentralized execution have become a basic setting for multi-agent reinforcement learning. As the number of agents increases, the performance of the actors that only use their own local observations with centralized critics is prone to bottlenecks in complex scenarios. Recent research has shown that agents learn when to communicate to share information efficiently, that agents communicate with each other in a right time during the execution phase to complete the cooperation task. Therefore, in this paper, we proposed a model that learn when to communicate under the centralized critic supporting, so that the agent is able to adaptive control communication under the centralized critic learned by global environmental information. Experiments in a cooperation scenario demonstrate the advantages of model. With our proposed cooperation model, agents are able to block communication at an appropriate time under the centralized critic setting and cooperation with each other at the task.