Traffic Signal Control Using End-to-End Off-Policy Deep Reinforcement Learning

An efficient transportation system can substantially benefit our society, but road intersections have always been among the major traffic bottlenecks leading to traffic congestion. Appropriate traffic signal timing adapted to real-time traffic may help mitigate such traffic congestion. However, most existing traffic signal control methods require a huge amount of road information, such as vehicle positions. In this paper, we focus on a particular road intersection and aim to minimize the average waiting time. We propose a traffic signal control (TSC) system based on an end-to-end off-policy deep reinforcement learning (deep RL) agent with background removal residual networks. The agent takes real-time images at the road intersection as input. Upon sufficient training, the agent can perform (near-) optimal traffic signaling based on real-time traffic conditions. We conduct experiments on different intersection scenarios and compare various TSC methods. The experimental results show that our end-to-end deep RL approach can adapt to the dynamic traffic based on the traffic images and outperforms other TSC methods.