Artificial Intelligence Signal Control Method For Isolated Intersection With Dual-ring Phase Plan

The traditional traffic signal control at isolated intersections needs to establish logic rules based on explicit programming,or optimization models based on performance indicators artificially,then inject traffic signal controller to realize the design,adjustment and optimization of traffic signal timing plan.However,there are some bottleneck in artificial experience,at the same time,this type of open-loop control lacks feedback and effective application of performance evaluation,which ultimately limits the improvement of road users' performance.In recent years,the study of combining artificial intelligence with traffic signal control has shown a blowout trend,under the dual role of the rapid development of artificial intelligence and the emerging traffic data collection technology,Dual-ring phase plan is the most common method to assign the right-of-way for large signalized intersection,which can configure the green time of vehicle phase flexibly and efficiently.In this paper,an artificial intelligence signal control method at isolated intersection is developed by reinforcement learning technology for an isolated four-leg signal control intersection with dual-ring phase plan,which aims at making more vehicles to pass through the stop lines in a shorter time.Two agents are deployed for the leading phase of ring A and B respectively,and one agent is deployed for the lagging phases of ring A and B.The roadway space which locates 160 meters upstream of the stop line in each approach is divided into 13 areas,and the number of vehicles,the average instantaneous speed and the number of queued vehicles in each area are collected by the multi-target tracking radar as the traffic operational characteristics.They together with the signal operational characteristics form the state vector.The action is selected from 36 action spaces,which is set as the extended green time after the minimum green time at the current vehicle phase.The reward is set to the total number of vehicles passing through the stop line in each approach per second.According to the methods of state representation and action setting,a customized Double Deep Q-Network(DDQN)is established,a deep fully connected neural network is adopted to represent strategy,and a modified multi-step temporal-different method is used to update the action value.Agents are trained in the virtual road traffic environment by VISSIM.Taking the typical four-leg signal control intersection as the experimental object,under the road space conditions of three lanes in road section and the peak traffic demand,the simulation experiment results show that for the dual-ring phase scheme,compared with the traditional full actuated signal control at isolated intersection,the artificial intelligence signal control method at isolated intersection proposed in this paper is an effective signal control method which can reduce the average vehicle delay at the intersection by 17% and the average vehicle queue time by 22%.