Multiagent modeling of pedestrian-vehicle conflicts using Adversarial Inverse Reinforcement Learning

Abstract

There is a need for a better understanding of the collision avoidance behavior of road users in near misses. Recently, several models of road user behavior in near misses have been proposed. However, despite the multiagent nature of road user interactions, most of these studies modeled their behavior using a single-agent approach. However, this approach is unrealistic and can limit the models’ accuracy. Therefore, this study proposes the Markov-Game (MG) framework for modeling pedestrian-vehicle interactions and their collision avoidance mechanisms. Pedestrian-vehicle conflicts in a mixed traffic environment in China are extracted using computer-vision algorithms. Pedestrian and vehicle reward functions are recovered via the Multiagent Adversarial Inverse-Reinforcement-Learning approach. Road user optimal policies and collision avoidance mechanisms are predicted using multiagent Actor-Critic deep-reinforcement-learning. The results demonstrate the superiority of the multiagent modeling approach in predicting road user behavior, their collision avoidance mechanisms, and the Post-Encroachment-Time (PET) compared to a baseline single-agent model.