Decision-Embedded Learning

Deep learning models are increasingly used to predict passenger demand and other inputs for decision models used in transportation planning, design, operations, and maintenance. Our research asks when improvements in prediction accuracy improve decision quality, how errors and adversarial inputs affect downstream decisions, and how transportation optimization can be incorporated directly into learning. Supported by NSF’s Civil Infrastructure Systems (CIS) program, this project develops a new decision-embedded deep learning paradigm for transportation systems, using transit systems as an application.

Do Better Predictions Always Improve Decisions?

Our investigation began with a simple observation. Many existing studies focus on developing deep learning models that predict passenger demand and other transportation system parameters more accurately, assuming that better predictions will lead to better planning and operational decisions. But does this assumption always hold? To test it, we evaluated demand prediction models across three passenger datasets and used their predictions as inputs to a transit system design model. We then compared prediction errors with the resulting decision errors, measured by the increase in system cost relative to decisions made with accurate demand information. As shown in Figure 1, decision errors generally decrease as prediction errors decrease, until prediction error falls below a certain threshold. Beyond this point, further improvements in prediction accuracy yield only marginal improvements in decision quality. These findings motivate our work on understanding—and directly accounting for—the connection between prediction and decision-making.

Decision-Embedded Deep Learning

This project aims to develop and validate a Decision-Embedded Deep Learning (DDL) paradigm for transportation systems. As shown in Figure 2, DDL connects prediction and transportation decision-making within the learning process, allowing models to learn from the quality of the decisions they support rather than prediction accuracy alone. The research follows three interconnected thrusts: (i) understanding how prediction errors affect decision quality, (ii) examining how this relationship changes under adversarial inputs and other disturbances, and (iii) developing learning methods that directly account for downstream decisions. Together, these thrusts aim to provide a foundation for more effective and reliable uses of deep learning in transportation systems. These theoretical and methodological developments will be tested using two applications: transit timetable design and motion planning for automated buses.