Learning Efficient Multi-agent Communication: An Information Bottleneck Approach

• DOI: 10.48550/arxiv.1911.06992
• Publication Date: 2019-01-01
• Authors: Rundong Wang; Xu He; Runsheng Yu; Wei Qiu; Bo An; Zinovi Rabinovich
• Keywords: FOS: Computer and information sciences; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; 0202 electrical engineering, electronic engineering, information engineering; Computer Science - Multiagent Systems; 02 engineering and technology; Multiagent Systems (cs.MA)

We consider the problem of limited-bandwidth communication for multi-agent reinforcement learning, where agents cooperate with assistance a protocol and scheduler. The scheduler jointly determine which agent is communicating what message to whom. Under limited bandwidth constraint, required generate informative messages. Meanwhile, an unnecessary connection should not be established because it occupies resources in vain. In this paper, we develop Informative Multi-Agent Communication (IMAC) method learn efficient protocols as well scheduling. First, from perspective theory, prove that constraint requires low-entropy messages throughout transmission. Then inspired by information bottleneck principle, valuable compact weight-based To demonstrate efficiency our method, conduct extensive experiments various cooperative competitive tasks different numbers bandwidths. show IMAC converges faster leads among under compared many baseline methods.