Learning Team-Level Information Integration in Multi-Agent Communication

TL;DR — Double Channel Communication Model

Abstract

In human cooperation, both individual knowledge and group consensus play important roles in accomplishing tasks. However, existing multi-agent reinforcement learning (MARL) communication methods commonly focus on individual-level communication, which lacks the necessary global information for well-grounded decision-making. Meanwhile, individual-level communication is often infeasible when the communication bandwidth is limited. To tackle these problems, we propose a group-level information integration model called Double Channel Communication Network (DC2Net). DC2Net highlights the significance of independent group feature learning by separating individual and group feature learning into two independent channels. In this model, agents no longer communicate with each other in a peer-to-peer paradigm; instead, all interactions are carried out in the group channel. By combining individual and global features, decisions are made collaboratively. We conduct experiments on several multi-agent cooperative environments and the results show that the DC2Net not only outperforms state-of-the-art MARL communication models but also reduces the communication costs. Furthermore, the two independent channels enable adaptive balancing of individual and group feature learning based on task requirements.