Abstract
Large-scale mobility data are transforming how we observe travel behavior, yet translating these observations into reliable and interpretable preference estimates remains challenging. Conventional discrete choice models provide a rigorous behavioral foundation, but assumptions about taste distributions can limit their ability to capture complex taste heterogeneity. Meanwhile, more flexible data-driven approaches, particularly neural-network-based models, often face challenges in interpretability, model identification, and estimation stability. This talk introduces an agent-based mixed logit (AMXL) framework that leverages inverse optimization to recover empirical distributions of taste parameters. The AMXL formulation decomposes taste heterogeneity into population-level tastes associated with traveler characteristics and agent-specific deviations learned from large choice data. Using simulated EV charging location choice data, the talk compares AMXL with mixed logit and TasteNet-logit in terms of predictive accuracy, taste parameter recovery, and robustness to noise in observed attributes. The talk further demonstrates the scalability of AMXL through a citywide microtransit revenue management application in Arlington, Texas. By integrating synthetic travel demand, a joint mode–subscription model, and simulation-based policy evaluation, the study examines how subscription pricing and targeted subsidies affect microtransit ridership, operator revenue, and traveler welfare.
