Research On Intelligent Control Method Of Signalized Intersection Based On Deep Reinforcement Learning

Traffic congestion has become a stumbling block on the road to sustainable development in major cities.With current road resources and infrastructure,the search for more advanced and intelligent traffic signal control methods has become a popular trend recently.Thanks to the rapid development of artificial intelligence technology,traffic signal control methods have also made great progress,however,traffic signal control methods in existing studies usually only consider traditional traffic parameters such as traffic flow and lane occupancy for signal timing scheme optimization,without making full use of key information of traffic status;in addition,there are fewer studies on coordinated control of signals at multiple intersections.Therefore,in order to alleviate traffic congestion and improve road transportation efficiency,as well as to enhance the efficiency and reliability of traffic signal control,this paper will study the signal intersection intelligent control method based on deep reinforcement learning.First,this paper summarizes and analyzes the current situation of deep reinforcement learning applications in traffic signal control,based on which a traffic simulation platform for signal intersection intelligent control experiments is built.The platform uses SUMO,a free open-source microscopic traffic simulation software,as the underlying architecture,and uses the programming language Python to call the interface TraCI for secondary development,and adopts Open AI’s Stable Baselines library as the framework for implementing deep reinforcement learning algorithms,and then completes the simulation platform from three aspects: environment configuration,platform architecture and functional modules.The simulation platform is built in terms of environment configuration,platform architecture and functional modules,which lays the foundation for the later study.Second,for single-intersection intelligent signal control,this paper proposes a method based on improved deep Q-learning.The method describes the traffic environment of single intersection in detail and precisely designs the three elements of signal intelligences(state is defined as the set containing average queue length,vehicle position and vehicle speed,action is defined as the green light duration of different phases within each cycle duration,and reward is defined as the total system delay time);meanwhile,considering the training inefficiency and Q deviation of traditional DQN model,the corresponding improved The model architecture and discount factor are improved accordingly,and the algorithm is verified on the simulation platform.The effectiveness and superiority of the proposed method are proved by simulation experiments under different traffic conditions.Finally,for multi-junction intelligent signal control,a deep reinforcement learning signal control method combined with game theory is proposed in this paper.The method designs an intelligent body control structure containing IA-MADRL mode and CAMADRL mode based on a single intersection by introducing game theory;IA-MADRL control is used when the traffic demand is not saturated and CA-MADRL mode control is used when the traffic demand is saturated,and the switching between the two is decided by the control mode detection module;the adopted regular form game effectively coordinates the joint actions of each intersection and enhances the real-time interaction.The reliability and scientific validity of the proposed method are proved by simulation experiments under different traffic demands.