Can simulated traffic conflicts from Markov game multi-agent models represent the risk of observed conflicts? An investigation using extreme value theory

Simulation-based safety analysis has become a key tool for evaluating road safety and gaining a good understanding of road user behavior in traffic interactions. However, the extent to which simulated models replicate real-world crash risk, particularly in vehicle–pedestrian interactions, remains largely unexplored. This study introduces a unified framework that combines Multi-Agent Adversarial Inverse Reinforcement Learning (MA-AIRL) with Extreme Value Theory (EVT) models to compare the crash risk of simulated and observed vehicle–pedestrian conflicts across multiple urban environments. Simulated conflicts were generated using MA-AIRL models, which had been developed using large-scale datasets from four different cities. Results indicate that the models based on simulated conflicts captured the same direction of the behavioral covariate effects as the models based on observed conflicts (i.e., the associations between minimum distance and pedestrian maximum speed with crash risk). However, the magnitude of these effects was different. Additionally, the results show that the crash risk estimates derived for both simulated and observed conflicts differ considerably across environments, with the EVT models with simulated conflicts mostly presenting an overestimation in the crash risk when compared to the models with observed conflicts. This research provides a probabilistic crash risk framework that enables conflict-based validation of simulation-based safety models against observed data. Moreover, it offers new insights into the role of EVT distributions and behavioral variability in shaping the accuracy of simulated safety assessments.