Communication Information Fusion Based Multi-Agent Reinforcement Learning for Adaptive Traffic Light Control

Applying reinforcement learning for adaptive traffic light control is an important research direction of smart transportation. In the scenario of multiple intersections, it is essential for the algorithm to alleviate the impact of the non-stationary environment to improve the performance. In this paper, we propose the communication information fusion based multi-agent deep deterministic policy gradient algorithm (CIF-MADDPG) for adaptive traffic light control. CIF-MADDPG applies local communication to realize multi-intersection cooperative control by integrating the traffic information of neighbor intersections. The performance of CIF-MADDPG is verified on the public datasets, and simulation results demonstrate that CIF-MADDPG can effectively relieve traffic jams.