Multi-Agent Deep Deterministic Policy Gradient for Traffic Signal Control on Urban Road Network

Aiming at how to effectively use information in urban traffic signal control to optimize traffic conditions and ensure the adaptability and robustness of the control algorithm, this paper applies multi-agent reinforcement learning algorithms for cooperative traffic signal control. A multi-agent deep deterministic policy gradient (MADDPG) based method is proposed to reduce the average waiting time of vehicles, though adjusting the phases and lasting time of traffic lights. The environment in each intersection is abstracted by the method of matrix representation, which effectively represents the main information on the traffic network and reduces redundant information. The control method is evaluated via simulation platform SUMO, under different levels of road congestion. The results of the contrast experiment imply the efficiency and stability of MADDPG in multi-intersection signal control.