HARL: A Novel Hierachical Adversary Reinforcement Learning for Automoumous Intersection Management

As an emerging technology, Connected Autonomous Vehicles are believed to pass intersections with greater efficiency, and related researches have been conducted for decades. However, compared to rule-based or model-based scheduling approaches, the distributed reinforcement learning has only begun to be applied in the field of autonomous intersection management since the past three years. The distributed learning-based method help vehicle make decisions more independently, and the adaptive adjustment of dynamic strategy in varied condition makes this method more suitable for the AIM at hybrid intersection. We propose a hierarchical decision framework with various receptive scope, action step length, and feedback period of reward, making use of information from different perspectives within different time spans and forming driving trends for vehicle to co-determine the action confrontationally at each step. The proposed model is proven effective in the experiment undertaken in a complicated intersection, and show better performance compared with baselines.