Inverse Learning and Intervention of Transportation Network Equilibrium

Описание: By 2035, nearly half of all new vehicles in the United States will be connected, generating unprecedented volumes of mobility data. Leveraging emerging connected mobility data, this talk establishes an AI-enabled inverse learning framework to transform the transportation network equilibrium modeling paradigm, which has been the foundation of system planning and management for over seventy years.

Traditional transportation network equilibrium models are time-consuming and costly to calibrate. This talk presents the inverse learning of user equilibrium as a novel framework for constructing nonparametric, context-dependent network equilibrium models directly from empirical travel patterns. We compare nonparametric and parametric approaches, mathematically clarifying the trade-offs among behavioral realism, data availability, and computational cost. The proposed neural-network-based nonparametric framework can automatically discover any well-posed network equilibrium model given
sufficient data, without relying on predefined behavioral assumptions. In contrast, the semi-parametric approach is more computationally tractable, as it simplifies the inverse learning problem into a sequence of convex optimizations.

Finally, we apply the inverse learning framework to a long-term network design problem for the city of Ann Arbor, Michigan. Using real-world crowdsourced data, we learned a context-dependent equilibrium model and introduce a certified, auto-differentiation-accelerated algorithm to solve the resulting distributionally robust bi-level network design problem under contextual uncertainty.