An Independent Trajectory Advisory System in a Mixed-Traffic Condition: A Reinforcement Learning-Based Approach

Achieving smooth urban traffic flow requires reduction of sharp acceleration/deceleration and accordingly unnecessary stop-and-go driving behavior on urban arterials. Traffic signals at intersections, and induced queues, introduce stops along with increasing travel times, stress and emission. In this paper, an independent reinforcement learning-based approach is developed to propose smooth traffic flow for connected vehicles enabling them to skip a full stop at queues and red lights at urban intersections. Two reward functions, i.e., a fuzzy reward engine and an emission-based reward system, are proposed for the developed Q-learning scheme. Another contribution of this work is that the necessary information for the learning algorithm is estimated based on the vehicle trajectories, and hence, the system is independent. The proposed approach is tested in a mixed-traffic condition, i.e., with connected and ordinary vehicles, via a realistic traffic simulation with promising results in terms of flow efficiency and emission reduction.