Learning-based game theoretical framework for modeling pedestrian motion

Pedestrians are active agents that undergo a repeated decision-making process while walking. These anisotropic, interactive, and feedback-oriented agents observe their surroundings, anticipate the future state of the network, and decide on their next movements accordingly, while ensuring a collision-free path toward their destination. Aiming at capturing these behavioral characteristics of human agents while walking, the present study puts forward a learning-based game theoretical approach for modeling pedestrian motion in dynamic environments. The proposed game structure provides a technical foundation to analyze optimal decision-making by pedestrians where the outcome of the game for each player's choice depends primarily on the strategies played by other players. This, in turn, ensures the frequently observed collision avoidance behavior of pedestrians while walking. The influence of nearby pedestrians on one's decision-making process and the feedback-oriented behavior of human agents are also captured via incorporating a learning structure. Optimum moving strategies are selected based on Nash equilibria calculations, where everyone is playing optimally given what all other players are playing. The validation results using real-world trajectories of pedestrians provide evidence of the model's capability in describing pedestrian motion and walking behaviors at microscopic as well as macroscopic levels.