Enhancing Pedestrian Route Choice Models Through Maximum-Entropy Deep Inverse Reinforcement Learning With Individual Covariates (MEDIRL-IC)

Understanding pedestrian route choices is pivotal for deciphering individual behaviors and informing decisions in urban planning. Unlike motorized transportation, primarily influenced by the built environment of the origin and destination, pedestrian route choices are also shaped by individual characteristics. Traditional route choice models often employ multi-segment training based on these characteristics, limiting their ability to capture heterogeneity among pedestrians within a unified framework. Additionally, these models frequently rely on oversimplified utility assumptions and neglect the implications of Markov decision sequences, potentially resulting in a misinterpretation of genuine route choice preferences. Addressing these, we introduce the Maximum-Entropy Deep Inverse Reinforcement Learning with Individual Covariates (MEDIRL-IC) framework, which accounts for the stochastic nature of state transitions in Markov decision process. Building on MaxEnt-IRL, MEDIRL-IC uses deep neural networks to separately fit constant individual features and dynamic built environment features before integration, effectively capturing the non-linear reward function in pedestrian scenarios. We further enrich model interpretability using a graph-based causal discovery algorithm, offering insights into complex feature interactions. Validations on a pedestrian mobile signaling dataset affirm MEDIRL-IC’s superior balance between model fit and interpretability. Our contributions present urban planners with a robust analytical tool, facilitating the data-driven design of pedestrian-centric urban landscapes. For research transparency and reproducibility, all codes are accessible at: https://github.com/BoyangL1/Advanced_DeepIRL.