Combining time dependency and behavioral game: A Deep Markov Cognitive Hierarchy Model for human-like discretionary lane changing modeling

Abstract

Human drivers take discretionary lane changes when the target lane is perceived to offer better traffic conditions. Improper discretionary lane changes, however, lead to traffic congestion or even crashes. Considering its significant impact on traffic flow efficiency and safety, accurate modeling and prediction of discretionary lane-changing (LC) behavior is an important component in microscopic traffic analysis. Due to the interaction process and driver behavior stochasticity, modeling discretionary lane-changing behavior is a non-trivial task. Existing approaches include rule-based, utility-based, game-based, and data-driven ones, but they fail to balance the trade-off between modeling accuracy and interpretability. To address this gap, we propose a novel model, called Deep Markov Cognitive Hierarchy Model (DMCHM) which combines time dependency and behavioral game interaction for discretionary lane-changing modeling. Specifically, the lane-changing interaction process between the subject vehicle and the following vehicle in the target lane is modeled as a two-player game. We then introduce three dynamic latent variables for interaction aggressiveness, cognitive level, and payoffs based on the Hidden Markov Model. The proposed DMCHM combines time dependency together with cognitive hierarchy behavioral games while preserving model interpretability. Extensive experiments on three real-world driving datasets demonstrate that DMCHM outperforms other game-theoretic baselines and has comparable performance with state-of-the-art deep learning methods in time and location errors. Besides, we employ SHAP values to present the model interpretability. The analysis reveals that the proposed model has good performance in discretionary LC prediction with high interpretability. Finally, we conduct an agent-based simulation to investigate the impact of various driving styles on macroscopic traffic flows. The simulation shows that the existence of massive aggressive drivers can increase traffic capacity because of small gaps during car-following, but inversely decrease discretionary LC rates. A balanced mixing of conservative and aggressive driving styles promotes discretionary LC frequencies since conservative car-following behaviors provide more spaces for LC. The codes can be found at https://github.com/zeonchen/DMCHM.