DualLight: Intelligent Traffic Signal Control Method Based on Deep Reinforcement Learning

In recent years, research combining max pressure (MP) with reinforcement learning for traffic signal control has become a hot topic in the field of intelligent transportation. However, existing MP methods ignore the importance of moving vehicles, which prevents the model from focusing on sufficient key information when extracting state features, thus making it suffer from unstable learning, slow convergence, and limited performance during the training process. To address this problem, this paper proposes DualLight, a signal control algorithm based on deep reinforcement learning. In DualLight, a new traffic state representation, is designed, which can help the model extract key information at intersections. To stabilize the learning process of the agent and accelerate the convergence speed, the DMP control method is designed as an expert method for agent imitation in the pretraining module, through which the parameters of the model are initialized. Comprehensive experiments on several real-world datasets show that DualLight improves its performance on the average travel time metric by an average of 17.05%, which significantly improves the efficiency of the transportation network.