Simulation of human–vehicle interaction at right-turn unsignalized intersections: A game-theoretic deep maximum entropy inverse reinforcement learning method

Abstract

The safety of pedestrians in urban transportation systems has emerged as a significant research topic. As a vulnerable group within this transportation framework, pedestrians encounter heightened safety risks in complex urban road environments. Protecting this group and safeguarding their rights and interests in urban transportation has garnered attention from academia and industry. The objective of this study is to develop a reliable simulation model that represents pedestrian crossing behavior at unsignalized crosswalks. A data-driven human–vehicle interaction behavior modeling framework is proposed, describing the human–vehicle interaction process at right-turning unsignalized intersections as a standard Markov decision-making process. In this framework, pedestrians are treated as the primary agents, and human–vehicle interactions are described using game theory. The Deep Maximum Entropy Inverse Reinforcement Learning (DMIRL) approach, combined with game theory, is employed to identify a reward function that encapsulates these interactions. The Deep Q-network (DQN) algorithm is then designed to simulate pedestrian crossing behavior based on the derived reward function. Finally, a comparison with a baseline algorithm that does not account for the game dynamics validates the proposed framework’s effectiveness and feasibility.