Ranked Communication Channel Confidence for Multi-Agent Reinforcement Learning

Dongzi Wang, Mingyang Geng, Bo Ding, Dawei Feng

Abstract: Multi-agent reinforcement learning systems have received increasing attention in these years. Existing works have proved that communication between agents can accelerate the training process and improve the final performance. However, in real-world tasks, the communication bandwidth is always restricted and unstable either in the training or test process, which is usually neglected in previous works. In this paper, we propose a method called Ranked Communication Channel Confidence Multi-agent Reinforcement Learning (RC3MARL), which exploits a non-uniform information encoding mechanism to discriminate the message in finer granularity. RC3MARL is more flexible and can achieve better performance in various environments with different communication bandwidths than the methods with fixed communication mechanism. In concrete, when the bandwidth is limited, RC3MARL can guarantee an acceptable lower-bound performance, and when the bandwidth is adequate, RC3MARL will make full use of the bandwidth and provide better performance. The experimental result demonstrates that our method can achieve better performance with same training samples.