Joint Estimation of Dynamic O-D Demand and Choice Models for Dynamic Multi-modal Networks: Computational Graph-Based Learning and Hypothesis Tests

Researchers from Carnegie Mellon University have developed a novel AI framework that simultaneously estimates dynamic travel demand and behavioral choice models for entire multi-modal transportation systems. The research, published on arXiv, addresses critical gaps in existing transportation modeling by integrating private driving and public transit data—including traffic counts, probe speeds, and transit ridership—into a unified computational graph-based approach. This allows for the first time the joint estimation of how many people are traveling (origin-destination demand) and why they choose specific routes and modes (behavioral models) as these patterns change throughout the day.

The technical breakthrough lies in using computational graphs—similar to those in deep learning—to model dynamic traffic assignment at scale, incorporating detailed travel time calculations for multiple modes and heterogeneous traveler preferences. The framework includes a hypothesis testing component that provides statistically rigorous validation of behavioral parameters, enabling transportation agencies to test policies like congestion pricing or transit investments before implementation. This represents a significant advancement over traditional static models that struggle with large networks and limited behavioral factors, offering cities a powerful tool for data-driven transportation planning and emerging mobility policy design.