Multi-Vehicle Cooperative Optimization At Unsignalized Intersections In Intelligent Connected Environment

Intelligent connected technology combines the ubiquity of wireless communication and the universality of intelligent computing.It can construct a multi-vehicle system via vehicle-to-vehicle communication and realize vehicleinfra cooperative decision-making via vehicle-to-infrastructure communication.This technology improves traffic efficiency and makes multi-vehicle cooperative control possible at unsignalized intersections.However,the existing research is challenging to deal with the dynamic number of multi-vehicle cooperation under the constraint of unsignalized intersections,protect privacy while ensuring vehicle control performance,and define the multi-vehicle relationship in a mixed traffic environment.Given this,this thesis takes the multi-vehicle system as the research object and puts forward the multi-vehicle cooperative control constrained by unsignalized intersection,the cross-regional multi-vehicle cooperative control considering privacy protection,and the multi-vehicle cooperative control under mixed traffic conditions,which provides theoretical guidance and method support for the multi-vehicle cooperative optimization of unsignalized intersection in the intelligent connected environment.Due to the complexity of the road structure and the rapid changeing number of vehicles at unsignalized intersections,it brings risks to the safe and efficient passage of vehicles.This part studies the multi vehicle cooperative control method under the constraint of unsignalized intersection from the perspective of vehicle cooperative decision-making.This part models multi-vehicle passing at unsignalized intersections as a multi-agent Markov decision process.Considering the dynamic number of vehicles in the transportation system,each vehicle needs to select the necessary reference vehicles for decision-making.This process is a Partially Observation Markov Decision Process.Based on the Multi-Agent Deep Deterministic Policy Gradient(MADDPG)algorithm,this part proposes a Cooperative Multi-Agent Deep Deterministic Policy Gradient(CoMADDPG).By selecting nearby vehicles as reference vehicles,a stationary environment required for reinforcement learning is constructed to deal with the impact of dynamic vehicle number on decision-making.Considering the road characteristics of intersections,this part proposes a vehicle selection method by introducing the virtual lane method to assist the algorithm convergence.The proposed CoMADDPG algorithm can improve the traffic efficiency by 39.28%on the premise of ensuring safe driving.When the intersection scale is expanded to the urban level,the centralized processing of vehicle driving data becomes very difficult for the sake of privacy and security.This part studies the cross region multi vehicle cooperative control considering privacy protection from the perspective of data localization processing.Aiming at the scene of multiple unsignalized intersection,a "CloudEdge-End" three-layer federated learning framework is proposed,which includes three parts:Vehicle interactor,Edge trainer,and Cloud aggregator,to protect privacy on the premise of ensuring vehicle control performance.The framework contains a traffic-aware model aggregation algorithm and a ruleoriented imitation learning algorithm.Both algorithms are used to improve the performance of vehicle control.Furthermore,by introducing the method of"computing for communication",a loss-aware experience selection strategy is proposed to reduce the communication overhead in centralized training and distributed execution.The experimental results show that the imitation learning algorithm can obtain the same collision avoidance ability as the rule and reduce the driving instability by 55.71%.The traffic-aware model aggregation algorithm further reduces the instability by 41.37%.The proposed experience selection strategy can reduce the communication overhead by 12.80%while ensuring model convergence.Compared with the pure intelligent connected vehicle operating environment,the mixed traffic scenario contains a large number of uncontrollable human driven vehicles,which increases the difficulty of traffic coordination and control.Therefore,this part studies the multi vehicle cooperative control method under mixed traffic conditions from the vehicle-vehicle relationship structure of intelligent connected vehicles.This part puts forward the vehicle passing optimization problem under the condition of mixed traffic at an unsignalized intersection to promote the intelligent connected vehicle to give up the right of way at the appropriate time and improve the overall traffic capacity of the unsignalized intersection.This part proposes a Graph Reinforcement Learning for Unsignalized Intersection Vehicles(GRL-UIV)in the way of "Lane evaluation+Action generation." Firstly,the relationship among intelligent connected vehicles is established with the assistance of the lane conflict relationship.Secondly,the output of reinforcement learning is set as the lane evaluation value.Finally,an action generation method based on lane evaluation is proposed,which integrates lane evaluation value and lane conflict relationship to output a set of intersection vehicle actions.The experimental results show that the proposed GRL-UIV method can reduce the intersection deadlock and guide 29.55%more traffic scenes.