LFMAC: Low-Frequency Multi-Agent Communication

Cong Cong, Chuxiong Sun, Ruiping Liu, Rui Wang, Xiaohui Hu

Abstract: An essential challenge of cooperative multi-agent reinforcement learning lies in boosting the efficiency of communication. However, the full communication mechanism adopted by existing methods would generate large communication costs. Furthermore, the redundant messages might even degrade the collaboration performance. We believe that it is not necessary to share the similar information from continuous observations. Hence we propose a Low-Frequency Multi-Agent Communication (LFMAC) method, which enables agents to learn when to communicate (for both sending and receiving messages) and how to act in a multi-task manner. Concretely, we learn a behavioral policy and a communicational policy for each agent through a multi-head actor network. Then we implement a gating mechanism to cut unnecessary messages, and a communication-skipping trick to reduce communication frequency. In addition, we evaluate LFMAC in a variety of scenarios from StarCraft II micromanagement tasks. The results demonstrate that LFMAC can achieve efficient communication under lower communication costs.