Cornell wins $600,000 grant to make AI predictions of human choices more explainable

Cornell researchers have received a three-year, $600,000 National Science Foundation grant to develop AI models that explain human choices in complex infrastructure systems. Described in an October 19 Cornell announcement, the project targets a difficult feedback loop: people respond to a system’s conditions, then their decisions change those conditions for everyone else.

Ricardo Daziano, a Cornell professor of civil and environmental engineering, leads the project with Anna Scaglione, a professor of electrical and computer engineering at Cornell Tech. Their aim is not just to improve predictions. They want models whose reasoning remains interpretable and whose uncertainty is explicit—particularly when researchers ask about a new policy or technology.

When a prediction changes with everyone’s choices

The research concerns cyber-physical systems: physical infrastructure connected to digital information, automated recommendations and human decisions. Transportation networks, electric-vehicle charging and power grids all fit that description. A commuter’s route choice or a homeowner’s response to electricity prices is part of a larger network, rather than an isolated decision.

Daziano describes several influences acting at once: the physical environment, information, other people’s behavior and feedback from the system itself. Decisions made in response can then alter the system for others. The forecasting problem therefore includes both individual behavior and the collective changes that behavior produces.

The electric grid illustrates the challenge. Cornell describes a shift from a few large power plants toward millions of distributed resources, alongside new demand from electric vehicles and data centers. Consumer and grid-operator decisions can send effects through that network, changing demand and operating conditions elsewhere. Predicting a response at one point is only part of the task.

We need to understand the trade-offs behind a prediction, know how uncertain they are and determine whether the model is still reliable when we ask about a new policy or technology

Ricardo Daziano, Cornell professor of civil and environmental engineering

Keeping an explanation inside the model

The project will combine Bayesian deep learning, microeconometrics and graph-based data. Its intended models bring together flexible AI methods, economic analysis of individual choices and information about network connections. The goal is to represent differences in how people interact across a network while retaining a useful explanation of their choices.

That approach builds on doctoral research by Daziano’s former student Daniel Villarraga, now an assistant professor at Universidad de los Andes in Colombia. His dissertation explored deep-learning models for discrete choice—decisions among alternatives—that preserved an interpretable behavioral core. Flexible nonlinear components allowed more complex relationships, while the models also quantified uncertainty.

The new grant extends that question into networked settings, where individual behavior and system performance affect each other. Applications will include technology adoption, energy, mobility and other infrastructure. A key objective is for models to recognize when the available data provides weak support for a prediction, rather than treating every forecast as equally dependable.

Can the learning survive a different network?

Scaglione is also interested in whether AI can learn something fundamental about a system’s underlying physics. She wants to investigate whether a model trained on one network could work on a different configuration, or another network at a different scale, without starting over. That is a research question for the funded work, not a demonstrated capability.

Her distinction is between AI that performs well on a benchmark and AI that remains useful in physical systems with laws, structure and constraints. Understanding when knowledge can transfer between networks is part of that ambition. The grant funds an investigation into those conditions, alongside the work on human behavior.