Research On Multi-intersection Traffic Signal Control Algorithm Based On Reinforcement Learning And Graph Neural Network

With the advent of smart city construction,traffic control methods using traffic lights have rapidly evolved.Traffic signals have transitioned from manual to automatic operation,signal cycles have shifted from fixed to variable,and system control modes have progressed from point control and line control to area control.This has led to the development of intelligent traffic control systems.However,most traditional traffic signal control methods rely on empirical observation and pre-defined rules.These methods are only suitable for fixed traffic patterns and cannot effectively address the complex and dynamic traffic problems present in modern intelligent transportation systems.To account for the impact of upstream and downstream intersections and achieve coordination linkage,multi-intersection traffic signal control has been been extensively researched.The multi-intersection traffic signal control algorithm based on reinforcement learning and graph neural network benefits from the advantages of graph neural network in aggregating information from multiple nodes and reinforcement learning in dealing with sequential decision problems,which realizes the intelligence control of multi-intersection traffic signal well and has a broad prospect in this field.However,existing methods limit the model’s training speed and performance due to insufficient attention to key information on decision-making in state representation and model aggregation.Additionally,existing reinforcement learning-based methods often lack generalization performance,resulting in significant performance degradation in new scenarios.In light of these limitations,this paper presents research on a multi-intersection traffic signal control algorithm that employs reinforcement learning and graph neural network.The main contributions and innovations of this work are as follows:(1)We propose a new state feature called NOV-LADLE to extend the road section for counting the number of vehicles from a fixed length to a dynamic length.This integration of traffic information in a smaller state space improves model performance.Additionally,we combine the graph attention mechanism with the residual structure to enhance the importance of its own intersection and efficiently aggregate information from neighboring intersections.This accelerates model convergence and alleviates oscillations.(2)We propose a data augmentation method based on the symmetry of road networks and intersections.This method can expand data by three times with different data distributions,enriching training data and improving model generalization performance while also increasing training speed and stability.(3)Experiments on simulated datasets and real datasets from Hangzhou,Jinan,and New York show that our multi-intersection traffic signal control algorithm based on NOV-LADLE has significant advantages in throughput and average travel time.The average travel time is reduced by up to 23.44%.Our data augmentation method improves model generalization performance,speeds up training,and alleviates convergence instability.