Research On Intelligent Traffic Signal Timing Optimization Based On Deep Reinforcement Learning

With the deepening of urbanization construction in China,a large number of people and vehicles flow into the city,and urban traffic becomes extremely congested.The increasing congestion has brought serious negative impact on the economic development and residents' life of the city,and also restricted the development of the urbanization construction process in China.There are many factors that cause traffic congestion,such as traffic signal control,traffic route guidance,the growth of the number of cars,and the imperfect traffic infrastructure.Aiming at the problem of traffic signal control,this paper uses clustering algorithm and deep reinforcement learning to realize adaptive traffic signal control decision-making,focusing on solving the potential congestion problem in high saturation road network.First of all,because of the correlation between the intersections in the traffic network,in order to avoid the congestion diffusion problem caused by the disadvantageous dredging of individual important intersections,this study takes the overall road network status as the perception range of convolution neural network,and in view of the large-scale and irregular structure problems existing in the actual traffic network,uses clustering algorithm to cluster the road network,and Establish multiple related receptive fields.At the same time,considering the excellent performance of deep learning in image perception,this study takes the normalized receptive field as the input,and senses the spatial characteristics of road topology through convolution neural network,so as to better perceive the state of road network.In this study,the intersection is used as an agent in the implementation of deep reinforcement learning strategy,and the deep reinforcement learning model is used to predict the actions that should be taken at each intersection.Through continuous interaction and trial and error with the road network environment,the intersection learns the knowledge that is conducive to the global road network signal decision-making from the traffic data,establishes the mechanism of deep reinforcement learning,accumulates learning experience,and achieves improvement The purpose of vehicle capacity and traffic congestion alleviation in the road network.In this study,the sumo platform is used to simulate the urban traffic network.Under the two indexes of the average waiting time and the total number of vehicles in the network,the proposed strategy is compared with the control of single intersection based on reinforcement learning and the control of collaborative intersection based on reinforcement learning.The experimental results show that the improved method based on deep reinforcement learning is better than the above two methods in high saturation road network,and can better alleviate traffic congestion.