Research And Simulation On Urban Traffic System Signal Control Strategy Based On Deep Reinforcement Learning

Urban intersections are important traffic nodes that affect the traffic capacity of urban transportation systems.For a long time,researchers at home and abroad are committed to making reasonable improvements to the intersection signal control algorithms.However,due to the complexity and randomness of the traffic system,the improved signal control algorithms often lack real-time and accuracy,and can not be verified and popularized.Therefore,in order to improve the utilization of urban traffice system transportation resources,to reduce the impact of traffic bottlenneck on road traffic capacity,and to improve the flexibility and accuracy of traffic signal control,this paper proposes a traffic flow control strategy based on deep reinforcement learning,and builds a simulation plarform to verify the effect of the algorithm.The main work of this paper is as follows:1.The theoretical basis and key technical difficulties of reinforcement learning,deep learning and deep reinforcement learning applied to urban traffic system are elaborated.Combining the denoising autoencoder model in deep learning with the Q learning model in reinforcement learning,we solve the problem of state space explosion caused by reinforcement learning applied in complex traffic system.2.Based on the traffic signal control strategy via deep reinforcement learning and considering the influence of phase sequence on a single intersecton,the research on coordinated control system of phase sequence and signal timing strategy is carried out.For the urban traffic area traffic control system,a regional traffic signal control strategy based on distributed multi-agent system is proposed to ensure the real-time and reliability of regional traffic classification control.3.A Python-VISSIM online traffic simulation platform is designed.This platform can use Python to write custom traffic signal control algorithms,and can access and change the traffic model objects in VISSIM through VISSIM’s own COM interface.The collected data are then organized and visualized.The simulation results show that when the traffic scene is simple and the traffic volume is stable,the average vehicle delay at the intersection where the improved algorithm is applied is lower than the average delay of the intersection using the traditional deep reinforcement learning algorithm.The real-time performance and accuracy of the improved algorithm are optimized,but the control effect is not better than the fixed signal timing algorithm;when the traffic scene is complex,such as the number of phases increases or the traffic folw varies greatly,the influence of traffic anomaly data on the signal control effect is amplified,and function of the denoising autoencoder is demonstated.The effect of the control strategy proposed in this paper outperform the traditional deep reinforcement learning control stratrgy and fixed traffic control strategies.