Research On Freeway Ramp Merging Control With Mixed-autonomy Traffic

Connected and autonomous vehicles(CAVs),different from traditional human driven vehicles(HVs),not only break the limitations of the physical performance of traditional vehicles,but can be driven by controllers outside via advanced sensors and controllers equipped.In the future,CAVs may become the actuators of traffic control strategies to influence the propagations of traffic flow,so as to achieve a safer,more comfortable and more efficient transportation system.This study considers mixed-autonomy traffic with CAVs and HVs,and proposes a cooperative ramp merging strategy(CRM)which controls the movements of CAVs on both the ramp and the mainstream,to improve the flow efficiency of freeway ramp merging areas,providing a reference for the practical application of CAVs in intelligent freeway.This thesis mainly includes the following tasks:1.To deal with the freeway ramp merging problem with mixed-autonomy traffic,a cooperative CAV ramp merging strategy is proposed,which is deployed in a typical freeway ramp merging area.Microscopic simulation,based on cellular automata model,is applied to simulate the mixed-autonomy traffic and the cooperative CAV control strategy.Results reveal that the proposed cooperative CAV control strategy can effectively improve the performance of mainstream traffic and reduce the average vehicle delay of the entire merging area.2.The Actor-Critic deep reinforcement learning algorithm is applied to optimize the proposed cooperative CAV ramp metering strategy,and the resulted single ramp control model is named RL-CRM model.The cooperative controlled duration of the ramp CAVs is regarded as the action,the traffic densities at the merging bottleneck and in the immediate upstream are regarded as the states,and the average vehicle speed of the mainstream vehicles is regarded as the reward.Results reveal that the RL-CRM model performs better than no control,ALINEA ramp control and the CRM model.3.The centralized Actor-Critic multi-agent deep reinforcement learning algorithm is applied to train the proposed cooperative CAV ramp merging strategy deployed at consecutive freeway ramps,and the resulted multi-ramp control model is named MARL-CRM model.Based on the settings in training the RL-CRM model,the variables of all the on-ramps are combined to form a joint state space,a joint action space and a mean reward function,to train the MARL-CRM model.Results reveal that the CRM strategy deployed at consecutive freeway ramps performs better than no control and ALINEA ramp control,and the performance of the strategy can be further improved by applying the centralized Actor-Critic multi-agent deep reinforcement learning algorithm.