Summary:

This project introduces an end-to-end framework for constructing integrated traffic network equilibrium models, which can serve as lightweight travel demand models, directly from multi-day aggregate traffic state observations. Unknown components on both the supply and demand sides are parameterized with computational graphs and embedded in a variational inequality to enforce user equilibrium conditions. The approach flexibly incorporates model-based, model-free (e.g., neural networks), or hybrid components and calibrates unknown parameters by minimizing discrepancies between observed and estimated traffic states.

The framework was validated through numerical experiments with synthetic networks and empirical data from the Ann Arbor network. It demonstrated strong predictive accuracy for link flows under network changes and resilience to incomplete or noisy data. In the Ann Arbor case study, the framework reduced link travel time prediction error from 83.6% in the benchmark model to 34.3% and successfully captured behavioral patterns such as reduced travel on weekends and snow days. The results also show that the framework has strong potential for prescribing optimal improvement plans to reduce congestion, as it integrates learning and optimization into a single data-to-decision pipeline.

Overall, the proposed end-to-end framework enables automated construction and calibration of transportation network equilibrium models using cross-source data and supports the evaluation and prescription of strategies such as capacity expansion or congestion pricing. In practice, the framework can improve the efficiency of transportation planning, reduce both capital and operational costs, and guide more effective resource allocation to maximize public benefit.

Publications:

• Final Report