Abstract:

To facilitate safe and efficient interactions between Autonomous Vehicles（AVs） and pedestrians， this study employs the Multi-Agent Deep Deterministic Policy Gradient（MADDPG） algorithm to establish a pedestrian-vehicle interaction model in a mixed traffic context that includes both autonomous and human-driving vehicles. This model formulates interaction strategies enabling AVs to avert accidents without the necessity of direct inter-vehicle communication. In comparison with several benchmark algorithms， the proposed algorithm demonstrates substantial improvements in terms of training efficacy， collision frequency reduction， and traffic capacity. Additionally， the robustness of the proposed model is assessed across varied risk scenarios. Findings reveal that as the intensity of pedestrian behavioral randomness， or behavioral noise rises， the duration of interaction delays of both vehicle categories increases. Remarkably， the collision rate of AVs initially increases before declining， indicating an adaptive learning phase. Under conditions of elevated noise， AVs exhibit a superior capability for collision avoidance compared to human-driving vehicles， highlighting their enhanced resilience in chaotic urban traffic conditions. These outcomes underscore the potential of MADDPG-based frameworks to significantly contribute to safer， more efficient AV integration in mixed traffic scenarios.

Key words: engineering of transportation system, autonomous vehicle, interaction of vehicle and pedestrian, deep reinforcement learning, multi-agent system