| As a significant part of urban road environment,the control of traffic lights is closely related to road efficiency.An excellent traffic light control scheme can maximize traffic flow on the road network and reduce the average travel time.Therefore,the research on multi-intersection traffic signal control is conducive to alleviating traffic congestion and laying a foundation for the development of intelligent transportation in the future.In recent years,With the great breakthrough of machine learning technology,more and more researchers learn and obtain new traffic signal control schemes based on reinforcement learning algorithm.Compared with the fixed traffic light control scheme,the method based on reinforcement learning can quickly adapt to the dynamic traffic flow.However,these studies tend to observe the traffic conditions at the target intersection independently,ignoring the dynamic changes in the status of traffic lights and traffic flow at neighboring intersections.In addition,mining traffic data information for temporal characteristics,most methods only pay attention to the historical traffic data information of the same intersection in the same time period,ignore that the traffic conditions of the road network at different historical times also have different effects on the traffic state of the current intersection in a time period.To overcome these limitations,this thesis suggests a new multi-intersection traffic signal control scheme based on attention mechanism and reinforcement learning.Specifically,it includes the following parts:（1）In view of the problem that some research work ignores the dynamic traffic conditions of neighbor intersections,or simply directly combines the traffic conditions of neighbor intersections,which makes the model unable to realize effective learning.This thesis proposes to build a spatial attention module,that is,using the attention mechanism to learn and specify different influence weights at different neighbor intersections,so as to realize the implicit cooperation of traffic signal control between intersections,avoid the non-stationary impact on the target intersection caused by the sudden change of traffic conditions at the neighbor intersection.（2）In view of the fact that the traffic conditions of the road network at different historical times have different effect on the current traffic conditions at the intersections,this thesis uses Long short-term memory LSTM network combined with attention mechanism to extract the time correlation characteristics of traffic state.Different from the spatial attention module,the temporal attention module focuses on the impact of important time steps on traffic conditions,while the spatial attention module focuses on the influence of the traffic conditions of important neighbor intersections in each time step.（3）Based on spatial attention module and temporal attention module,the model can not only capture the spatial characteristics of multi-intersection traffic information,but also capture the time correlation characteristics of target intersection and historical traffic information.Next,the thesis combines spatial and temporal features to predict the phase of traffic lights at the next time at each intersection based on the distributed Ape-X DQN framework.Finally,the thesis selects four public simulation synthetic data sets and two public real traffic data sets to carry out the experiment.Under different evaluation indexes,the effect differences between the thesis model and the other nine comparison models are compared.The experimental results show that the performance of the multi-intersection traffic signal control scheme proposed in this thesis is better than some existing mainstream methods in terms of average vehicle travel time and road network throughput.Among them,compared with Co Light method,the thesis model reduces the average travel time of vehicles by 1.9% on the simulation data set.On the real dataset,the model reduces vehicle average travel time by 2.5%. |
PDF Full Text Request