Learning Multi-Agent Communication with Policy Fingerprints for Adaptive Traffic Signal Control

Adaptive traffic signal control is widely recognized as an effective solution to improve urban mobility and reduce congestion in metropolises. Recently, reinforcement learning has been adopted for this transportation problem. While centralized reinforcement learning inevitably faces action space explosion, decentralized reinforcement learning allows agents to develop policies based on local observations but suffers from unstable training. In this paper, we present CommNetPF, a multi-agent decentralized reinforcement learning model incorporating communication and neighbourhood policy fingerprints for adaptive traffic signal control. With policy fingerprints in communication, agents learn to produce cooperative policies and the model converges faster. Experiments in scenarios of adaptive traffic signal control show that CommNetPF outperforms several strong baselines in terms of control performance and convergence speed.