An Efficient Attention-Based Deep Reinforcement Learning Model for Traffic Signal Control

Traffic Signal Control (TSC) is a significant challenge within intelligent transportation systems. As Vehicle-to-Everything (V2X) technology advances, TSC systems are increasingly able to utilize extensive vehicle driving data to enhance decision-making. This paper proposes a reinforcement learning-based model for TSC at urban intersections. We design a vehicle driving information matrix that integrates vehicle position and speed information. The state space includes vehicle driving information and the current traffic signal phase. The action space includes various traffic signal phases. To efficiently extract traffic features from the large state space, the model incorporates an attention mechanism within the neural network. The simulation results on the Simulation of Urban Mobility (SUMO) demonstrate the convergence and generalization capabilities of our model, showing significant advantages in trip duration compared to several benchmark methods. Ablation studies further validate the effectiveness of our proposed position-speed fusion matrix and attention mechanism in extracting traffic feature representations.