(The) architecture of decentralized multi-agent reinforcement learning with communication

Abstract

We consider the problem of cooperative multi-agent reinforcement learning in partially observable environment. In particular, it is essential to coordinate between agents in partially observable environment. In this paper, we introduce the compressed feature vectors for communicate between agents and how to design the decentralized network using them. We also introduce and apply the group dropout layer to train the ensemble of sub-network efficiently, and evaluate the proposed network on pursuit, which is a standard task in multi-agent systems.