Priority Sequential Inference: Improving Safety for Efficient Autonomous Highway Driving Using MARL

Recent advancements in multiagent reinforcement learning (MARL) have shown significant progress in the field of autonomous driving. However, the key issue of guaranteeing safe decision-making and control of autonomous vehicles (AVs) in complex and dynamic traffic environments, while avoiding collisions, remains a critical challenge. To further enhance the safety and reliability of autonomous driving systems, this article presents a novel MARL approach called priority sequential inference proximal policy optimization (PSIPPO) to enhance the decision-making capabilities and safety of AVs in mixed traffic highway environments. The proposed method introduces a sequential decision-making framework that enables AVs to make informed decisions based on the actions and states of preceding agents, effectively managing the order of decisions and prioritizing critical ones. Moreover, PSIPPO incorporates a future reasoning mechanism, allowing AVs to predict and respond to potential collision risks by simulating road scenarios several time steps ahead. An adaptive risk avoidance strategy (RAS) is also introduced to assess the appropriateness and safety of lane changes, considering factors, such as intervehicle distance, relative velocity, and acceleration. Extensive experiments conducted in various highway driving scenarios demonstrate the superior performance of PSIPPO compared with several mainstream baseline MARL methods, achieving higher rewards, maintaining higher average speeds, and significantly reducing collision rates, especially in complex traffic conditions.