Learning Group-Level Information Integration in Multi-Agent Communication

In multi-agent systems, it’s hard to make proper decisions for agents due to the partial observability of the environment. Among categories of multi-agent reinforcement learning (MARL) algorithms, communication learning is a common approach to solving this problem. However, existing work focus on individual-level communication which usually leads to significant communication costs. Meanwhile, the group feature couldn’t be well captured at the individual level. To tackle these problems, this paper proposes a group-level information integration model called Double Channel Communication Network (DC2Net). In DC2Net, individual and group features are learned in two independent channels. Agents no longer interact with each other at the individual level and all information interaction is carried out in the group channel. This model ensures effective learning of group features while reducing individual-level communication costs. Empirically, we conducted experiments on several environments and tasks. The experimental results show that the DC2Net not only has a better performance compared to other state-of-the-art MARL communication models but also reduces the costs of communication. Furthermore, it’s a natural communication topology with the ability in balancing individual and communication learning.

Authors

• Xiangrui Meng School of Intelligence Science and Technology, Peking University Key Laboratory of Machine Perceptron (MOE), Peking University
• Ying Tan School of Intelligence Science and Technology, Peking University Key Laboratory of Machine Perceptron (MOE), Peking University Institute for Artificial Intelligence, Peking University National Key Laboratory of General Artificial Intelligence

Keywords

• Multi-Agent Communication
• Multi-Agent Reinforcement Learning
• Deep Reinforcement Learning