Traffic Signal Control Based on Composite State Representation

As the number of motor vehicles continues to increase, the problem of road congestion becomes increasingly significant. In recent years, research has focused on establishing intelligent transportation systems and optimizing the traffic signal timing based on real-time traffic information collected. Many recent studies use reinforcement learning for traffic signal control. It is important to model the traffic environments at intersections reasonably. However, there are few studies that focus on how to consider information at different scales simultaneously. In this paper, we incorporate microscopic information such as the specific positions of vehicles, as well as macroscopic information such as the length of each lane queue, and propose a single-intersection traffic signal control algorithm based on composite state representation. We use an improved DQN algorithm based on phase-switching action space to process the composite state representation for further decision-making. Simulation results on a real data set show that the proposed model can effectively reduce average travel time and queue length.