Research And Implementation Of Intelligent Traffic Signal Control System With Strong Generalization Ability

In recent years,the problem of traffic congestion has received widespread attention.Optimizing the traffic signal control system is an essential method to relieve traffic jam.With the development of science and technology,the traditional traffic signal control strategy can no longer meet today’s complex traffic environment,so more flexible traffic signal methods are needed to better meet the changing traffic conditions.Due to the development of Internet of Vehicles technology and the perfect combination of traffic signal control(TSC)and reinforcement learning(RL)process,researchers have proposed many high-quality traffic signal control algorithms based on reinforcement learning.Although reinforcement learning has achieved success in the field of traffic signal control,most of the methods based on reinforcement learning have only solved the traffic signal control problem of a single road network.In the face of the new road network environment,the generalization ability is insufficient.Moreover,due to the long training time of reinforcement learning,these methods require a large computational cost.In response to the above problems,this article puts forward a TSC method on account of grad centralized control Meta-reinforcement learning,extracts meta-knowledge from multiple meta-training tasks,and applies the accumulated meta-knowledge to the new task adaptation,which improves the training efficiency in the new road network environment and greatly improves model generalization.In addition,for the sake of solving the dimension disaster problem under centralized control,this thesis designs a special agent to decompose the search space by using the divide and conquer paradigm.So far,this thesis is the first to combine centralized control method with Metareinforcement learning method.so as to promote the generalization productivity of agents in the face of new road network tasks and reduce the computational cost,this thesis further proposes a Metareinforcement learning control method based on implicit context variables.A special embedded network is designed to represent the historical experience of the meta training task,which is called an implicit context variable.The context variable is used to help the agent make traffic phase decisions.In addition,considering the importance of spatiotemporal correlation for traffic network,this thesis puts forward a Metareinforcement learning framework based on spatiotemporal characteristics for adaptive traffic signal control.Specifically,this thesis project a framework with graph attention network(GAT)and long short-term memory network(LSTM)to obtain spatiotemporal information.The spatiotemporal information are carefully combined to assist the agent in traffic phase decision.The algorithm effectively integrates the spatiotemporal correlation unique to the road network task,and greatly improves the model in terms of generalization.In addition,this thesis compares the performance with other advanced traffic signal control methods on several real road network data,and proves that the ours method can be extended to the road network that has not been seen before.Mutual union baseline,the method in this thesis has huge superiority in control efficiency and generalization.At the same time,this thesis also carried out ablation experiments to demonstrate the indispensability of each part of the proposed method.Finally,based on the above algorithms,this thesis designs and completes the generalization expansion module of intelligent traffic signal control system.