AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-26
Empirically Calibrated Differential Games for Simulating Driver Interactions
Game-theoretic models can describe vehicle interactions at intersections without hand-written priority rules, but their fit to human drivers after calibration has rarely been tested. The authors calibrate a general-sum differential game to recorded trajectories at four Washington, D.C. intersections and test it on held-out vehicles, where it is the most accurate of the compared models. Its position error is 1.00 m on 819 calibrated vehicles and 7.42 m on 96 held-out vehicles, and a stiffer contact barrier removes collisions in 43% of tight closed-loop conflicts.