| Traffic signal control can improve the efficiency of the transportation system and alleviate traffic congestion by coordinating the right of way of vehicles at intersections.The adaptive signal control method based on deep reinforcement learning developed in recent years is a method that can effectively utilize various types of traffic information,adapt to different traffic scenarios and make real-time optimal signal control decisions.As a data-driven method,the key issue in its practical application lies in the acquisition of traffic information.The development of connected and automated vehicle technologies may provide favorable conditions for the field application of deep-reinforcement-learning-based signal control methods.Under connected and automated environment,utilizing the rich traffic information provided by connected and automated vehicles,deep-reinforcement-learning-based signal control methods can fully leverage its control advantages and better adapt to complex traffic system environments.This thesis utilizes deep reinforcement learning methods to train traffic signal control models for single intersection and grid network traffic scenarios under connected environment.Furthermore,considering that connected and automated vehicles may drive under control,a collaborative optimization method for signal timing and vehicle longitudinal trajectory is proposed.The main work of this thesis is summarized as follows:(1)Intersection-level signal control based on deep reinforcement learning under connected environment.Connected environment can provide a more comprehensive understanding of the traffic conditions at intersections,which is helpful to model training and realization of adaptive control.Taking into account many practical elements of signal control,such as signal phase structure,action granularity,and traffic flow dynamics,an improved deep reinforcement learning training method is proposed.The method applies more concise state representation and more general action design to train adaptive signal control models under complicated and dynamic traffic demand scenarios,so that the trained models may possess higher control resolutions and more stable control performances.(2)Network-level signal control based on multi-agent deep reinforcement learning under connected environment.This chapter extends the signal control method from single intersection to grid network,under connected environment.Given network structure and traffic demand scenario,a network-level signal coordination model is trained with multi-agent reinforcement learning.Compared with traditional network-level signal control method,method with parameter sharing and method with model transfer,the proposed method can produce control models that can effectively improved the network traffic flow efficiency.(3)Collaborative control of traffic signals and vehicle longitudinal trajectories based on deep reinforcement learning under connected and automated environment.Further extending the signal control methods,leveraging the controllability of connected and automated vehicles,another reinforcement learning method is developed to train collaborative control models for signal timings and vehicle trajectories,under connected and automated environment.A trajectory control method for connected and automated vehicles grouped by road sections is explored,which can substantially reduce computation burden,and effectively optimize the operation of mixed-autonomy traffic so as to improve traffic flow efficiency. |
