Design Of Traffic Control Systems Of Urban Road-Networks Based On Deep Reinforcement Learning

The urban transportation system is a huge system with characteristics of randomness,complexity,and uncertainty.At present,it is difficult for people to establish accurate mathematical models for traffic control objects.Traditional traffic control has a bottleneck.The traffic control brings dawn,but the domestic traffic characteristics are more complicated,and the adaptive control system fails to play its best role.With the rapid development of computer technology,detector technology and artificial intelligence technology,the application of artificial intelligence algorithms for traffic signal control is favored by a large number of researchers,especially based on deep reinforcement learning methods,which do not require modeling of traffic objects.The agent is designed to implement strategy learning.The agent aims to improve the traffic operation state,perceive the traffic state through deep learning,and apply reinforcement learning to control decisions to obtain the optimal control strategy.The main purpose of this thesis is to apply deep reinforcement learning methods to generate traffic signal control schemes.Therefore,the thesis first builds a deep reinforcement learning model for traffic control,designs a reinforcement learning environment based on traffic state prediction,and a signal control agent based on improved QMIX decisions.Then,based on the traffic signal control logic,the state and state space of the traffic environment,the action and action space of the agent,and the reward function that can characterize the state impression of the action are designed.The main research contents are as follows:First,by analyzing the historical traffic status data and the spatial location of the traffic network,a sub-region type traffic control object based on key intersections is constructed.Then,based on the multi-agent technology and using the relevant knowledge of game theory,on the basis of the existing traffic control technology,the traffic control hierarchy control object is modeled,the collaborative control agent is designed,and the collaborative multi-agent object design traffic control framework.In the framework of traffic control,with the goal of improving the overall control effect of the sub-region,a rolling control method of traffic signals based on DRL is proposed.This method fully considers the spatial relationship of traffic objects,uses speed prediction to establish the relationship between the sub-region control actions and the traffic speed state function,and designs an agent decision method that can select the action with the largest cumulative expectation by observing the local state.The method verification was carried out by designing related experiments,and the method feasibility was verified from the aspects of control scheme comparison and operation effect analysis.Finally,based on the above research content and in combination with the needs of users,a rolling control system for traffic control schemes was designed.The system automatically issued the generated control schemes,which assisted the traffic signal timing personnel well.