Eco-driving Under Mixed Autonomy at Signalized Intersection:A Deep Reinforcement Learning Model
Dynamic programming model with eco-through constraint and safety constraint often causes computational inefficiency and even unfeasible solutions in mixed autonomy and heavy traffic conditions.This paper proposes an eco-driving-oriented and deep reinforcement learning based trajectory optimization model for Connected and Autonomous Vehicles(CAVs)in mixed autonomy.The model uses a compound reward reshaping and a twin delayed deep deterministic policy gradient algorithm to optimize CAV trajectories at the upstream of signalized intersection in mixed autonomy.The vehicular gap,speed difference,speed,distance to intersection,queue length,signal phasing and timing are selected as agent state to describe safety and driving mobility.The queue length is augmented in state representation to mitigate CAV halting possibility caused by queue of human driving vehicles.A multi-objective reward function is established based on agent state and anticipated arrival time at the intersection to optimize the CAV driving mobility,energy efficiency,comfortability,and safety.The proposed model performs better than the dynamic programing model in terms of decoupling the strong correlation between model constraints and computational complexity.The training and testing of the proposed model with simulation demonstrate that the vehicle delay at intersections significantly decreases with the increase of CAV penetration rate.Besides,the energy consumption relatively decreases by 5.47%,4.42%,and 2.91%,compared to uncontrolled scenarios,dynamic programming-based trajectory optimization model,and deep deterministic policy gradient-based trajectory optimization model.In addition,the proposed model can ensure the CAV to cross the signalized intersection without stopping,and also show robustness against traffic demand and signal cycle.
intelligent transportationtrajectory optimizationtwin delayed deep deterministic policy gradientsignalized intersectionconnected and autonomous vehicle
NETL
NSTL
万方数据
