Application and Evaluation of the Reinforcement Learning Approach to Eco-Driving at Intersections under Infrastructure-to-Vehicle Communications

Eco-driving behavior is able to improve vehicles’ fuel consumption efficiency and minimize exhaust emissions, especially with the presence of infrastructure-to-vehicle (I2V) communications for connected vehicles. Several techniques such as dynamic programming and neural networks have been proposed to study eco-driving behavior. However, most techniques need a complicated problem-solving process and cannot be applied to dynamic traffic conditions. Comparatively, reinforcement learning (RL) presents great potential for self-learning to take actions in a complicated environment to achieve the optimal mapping between traffic conditions and the corresponding optimal control action of a vehicle. In this paper, a vehicle was treated as an agent to select its maneuver, that is, acceleration, cruise speed, and deceleration, according to dynamic conditions while approaching a signalized intersection equipped with I2V communication. An improved cellular automation model was utilized as the simulation platform. Three parameters, including the distance between the vehicle and the intersection, signal status, and instant vehicle speeds, were selected to characterize real-time traffic state. The total CO 2 emitted by the vehicle on the approach to the intersection serves as a measure of reward policy that informs the vehicle how good its operation was. The Q-learning algorithm was utilized to optimize vehicle driving behaviors for eco-driving. Vehicle exhaust emissions and traffic performance (travel time, stop duration, and stop rate) were evaluated in two cases: (1) an isolated intersection, and (2) a medium-scale realistic network. Simulation results showed that the eco-driving behavior obtained by RL can not only reduce emissions but also optimize traffic performance.