Autonomous intersection management: graph-reinforcement learning method based on road information enhancement

Although extensive efforts have been made in the field of Autonomous Intersection Management (AIM), the transition to a fully autonomous traffic environment will require considerable time. In mixed traffic, the behavioural uncertainty of human-driven vehicles (HVs) poses a major challenge to stable cooperation among autonomous vehicles. To address this issue, this paper proposes a dynamic cooperation framework for multiple autonomous vehicles based on integrating Graph Attention Networks (GAT) with Deep Reinforcement Learning (DRL). The framework employs graph-based representations to model interaction relationships within the scene and innovatively incorporates prior knowledge of intersection topology into the representation process, helping to mitigate agent inaction caused by redundant interaction information. Safety constraints are introduced to restrict excessive exploration of driving policies, thereby enhancing driving safety performance. The effectiveness of the proposed method is validated through efficiency and safety metrics, and the reasons for performance improvements are further analyzed via individual agent behavious.