Hybrid Reinforcement Learning-Based Eco-Driving Strategy for Connected and Automated Vehicles at Signalized Intersections

  • 2022-01-19 19:31:12
  • Zhengwei Bai, Peng Hao, Wei Shangguan, Baigen Cai, Matthew Barth
  • 1

Abstract

Taking advantage of both vehicle-to-everything (V2X) communication andautomated driving technology, connected and automated vehicles are quicklybecoming one of the transformative solutions to many transportation problems.However, in a mixed traffic environment at signalized intersections, it isstill a challenging task to improve overall throughput and energy efficiencyconsidering the complexity and uncertainty in the traffic system. In thisstudy, we proposed a hybrid reinforcement learning (HRL) framework whichcombines the rule-based strategy and the deep reinforcement learning (deep RL)to support connected eco-driving at signalized intersections in mixed traffic.Vision-perceptive methods are integrated with vehicle-to-infrastructure (V2I)communications to achieve higher mobility and energy efficiency in mixedconnected traffic. The HRL framework has three components: a rule-based drivingmanager that operates the collaboration between the rule-based policies and theRL policy; a multi-stream neural network that extracts the hidden features ofvision and V2I information; and a deep RL-based policy network that generateboth longitudinal and lateral eco-driving actions. In order to evaluate ourapproach, we developed a Unity-based simulator and designed a mixed-trafficintersection scenario. Moreover, several baselines were implemented to comparewith our new design, and numerical experiments were conducted to test theperformance of the HRL model. The experiments show that our HRL method canreduce energy consumption by 12.70% and save 11.75% travel time when comparedwith a state-of-the-art model-based Eco-Driving approach.

 

Quick Read (beta)

loading the full paper ...