Multimodal Traffic Light Control with Connected Vehicles: A Deep Reinforcement Learning Approach

Multimodal traffic light control is a cost-effective way to deal with urban congestion. The development of V2X (Vehicle to Everything) technologies offers unprecedented data and hence new opportunities for situation awareness, but the conventional control algorithms fall short of fully exploiting the real-time vehicle information at the intersections. In this work, a Double Deep Q-learning (DDQL) approach is proposed for multimodal traffic light control with different priority requests in a connected vehicle environment. The proposed DDQL approach is integrated with the existing actuated controller and is readily implementable. The integrated system can terminate the DDQL controller and switch to the actuated controller for safety when an urgent issue occurs such as an electric power outage. The simulation results demonstrate the advantage of the proposed approach compared with actuated control and indicate the reduction of delays for both public transportation and personal vehicle by the proposed approach.