A Complete State Transition-Based Traffic Signal Control Using Deep Reinforcement Learning

Abstract

Traffic signal control is a fundamental but challenging real-world problem that manages traffic at roadway intersections by adjusting signal timing and phases sequences. With the advances in emerging transportation technologies (e.g., Connected and Automated Vehicles, roadside sensing, drones), richer information (e.g., vehicle position, speed, type) has now become available in real-time, which can be utilized to reduce congestion. This paper introduces a deep reinforcement learning (DRL)-based traffic signal control (TSC) method for isolated intersections, where vehicles’ positions and speeds (available via V2I, roadside sensing, or drone-based surveillance) are processed by a convolution neural network (CNN) and fed into the RL system as inputs. In addition, a complete state transition process using a dual-ring mechanism is introduced to enable flexible traffic signal control. Using a traffic simulator (SUMO: Simulation of Urban Mobility), the proposed algorithm is compared with both fixed-time and actuated TSC under different traffic conditions. Results show that the DRL-based strategy can reduce average delay and pollutant emissions by 34.7% and 18.5%, respectively.