Longitudinal control of connected and automated vehicles among signalized intersections in mixed traffic flow with deep reinforcement learning approach

Trajectory optimization for connected automated vehicles (CAVs) is an effective method to improve the overall performance of urban traffic. At the same time, the emergence of deep reinforcement learning (DRL) enables agents to learn the optimal strategy in a complex environment by constantly interacting with the environment. In a connected environment with signalized intersections, the optimal control of CAV can significantly reduce fuel consumption and improve driving comfort under the premise of ensuring both traffic efficiency and travel safety. Thus, a DRL-based longitudinal trajectory control approach is proposed to optimize the motion of the CAV in a mixed traffic flow with signalized intersections. By designing a reasonable reward function comprising fuel consumption reward, safety reward, traffic light reward, and travel time reward, the CAV can learn the optimal policy and reach the destination adhering to traffic light and safety. Moreover, the potential-based reward is carefully designed to provide prior knowledge for agents. Eventually, the performances of the CAV with the proposed strategy are testified under different scenarios. Simulation results show that our proposed DRL-based optimization method can effectively save fuel consumption of the CAV and the whole platoon, as well as improve driving comfort in all scenarios while satisfying the rules of traffic lights and travel safety.