Toward Adaptive Transit Planning with Artificial Intelligence

Abstract

Transit systems continue to operate in an era of rapid technological change and evolving social norms, marked by the proliferation of platform-mediated mobility services, increasing behavioral heterogeneity, and persistent uncertainty in travel demand. Traditional transit planning approaches—largely dependent on static, infrequent survey-based demand representations—struggle to respond to these dynamics in a timely and behaviorally realistic manner. This dissertation proposes an AI-enabled adaptive transit planning framework that integrates diverse data sources, behavioral inference, and simulation-based evaluation to support responsive and context-aware transit decision-making. The framework is developed through three complementary analytical components. First, a simulation-based study of Mobility-as-a-Service-powered intermodal transit operations demonstrates the potential of adaptive control strategies with reinforcement learning to improve system performance across multiple policy objectives. These experiments reveal that operational outcomes are highly sensitive to assumptions about travelers’ activity needs and preference heterogeneity, motivating the need for richer behavioral representations. Second, a data fusion framework is introduced to infer trip purpose from large-scale automated fare collection data using supervised machine learning. By transferring behavioral insights from survey-based observations, this approach transforms operational data into behaviorally enriched inputs that support longitudinal monitoring of evolving travel patterns. Third, an artificial intelligence-integrated discrete choice modeling framework is developed to capture structured, interpretable preference heterogeneity. By embedding representation learning within an econometric framework, this approach aims to provide a deeper behavioral understanding while preserving policy interpretability. To unify these components, the dissertation ultimately plans a two-stage population synthesis concept using deep generative models to generate diverse, behaviorally consistent agent populations that respect structural constraints in travel behavior. This outlines a conceptual “closing” for the proposed adaptive transit planning framework with artificial intelligence by illustrating how enriched behavioral data, structured preferences, and synthetic populations can be jointly deployed to evaluate context-specific planning problems. Overall, this dissertation reframes transit planning as an adaptive, learning-oriented process in which behavioral understanding, preference heterogeneity, and operational experimentation are continuously linked. By positioning artificial intelligence as an enabling infrastructure rather than an end goal, the proposed framework suggests a pathway toward more resilient, equitable, and effective transit planning in the face of ongoing mobility transformation.