Artificial intelligence can help predict when a commuter might take a new transit route or how a homeowner might respond to changing electricity prices, but forecasting human behavior becomes especially challenging when individual decisions affect an entire system - and the system, in turn, affects people's decisions.
A three-year, $600,000 National Science Foundation grant will support Cornell research aimed at addressing that challenge, developing AI models that can better represent human decision-making in cyber-physical systems while remaining interpretable and explicit about uncertainty.
Ricardo Daziano, professor of civil and environmental engineering, at the University of New South Wales where he gave a talk about research related to a new National Science Foundation grant he is leading to improve AI models of human choices in complex systems.
The project, "CPS Small: Integrating Bayesian Deep Learning, Microeconometrics, and Graph-Based Data for Enhanced Interpretable Choice Models with Heterogeneous Network Interactions," is led by Ricardo Daziano, professor of civil and environmental engineering, with Anna Scaglione, professor of electrical and computer engineering based at Cornell Tech.
Cyber-physical systems increasingly combine physical infrastructure with digital information, automated recommendations, connected technologies and human decisions. They include transportation networks, electric-vehicle charging systems, grid infrastructure and other environments in which people's choices interact with technology and with one another.
"Once people are part of the system, prediction becomes much more complicated," Daziano said. "Their decisions depend on the physical environment, on information, on what other people are doing and on feedback from the system itself. Those decisions can then change the system for everyone else."
Consider the electric grid. It is shifting from a system dominated by a few large power plants to one that must coordinate millions of distributed resources while absorbing new demand from electric vehicles and data centers. AI can help predict how consumers and grid operators will respond, but those decisions can ripple through the network, changing demand and conditions elsewhere. The challenge is predicting not just what people will do, but how their collective choices will change the system itself.
"The grid is becoming too distributed, dynamic and complex to manage with yesterday's tools," Scaglione said. "That makes it an extraordinary laboratory for the next generation of AI - and for achieving the kind of autonomy and intelligence we are beginning to see in self-driving vehicles, but at the scale of an entire infrastructure."
In many AI applications, predictive accuracy is the main objective, but in economics and engineering, more is needed.
"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," said Daziano, whose Ph.D. is in economics but whose academic career has focused on engineering research into complex decisions involving technology, infrastructure and public policy.
The intellectual origins of the project trace back to the doctoral research of Daziano's former student Daniel Villarraga, Ph.D. '25, now an assistant professor of industrial engineering and operations research at Universidad de los Andes in Colombia. Villarraga's dissertation work with Daziano explored how deep-learning models for discrete choice could retain an interpretable behavioral core while allowing flexible nonlinear components and quantifying uncertainty.
"We started from a basic question," Daziano said, "Can we get the flexibility of deep learning without giving up the economic interpretation that makes a choice model useful? This NSF project expands that question substantially, especially toward networked systems in which individual behavior and system performance feed back into each other."
The research will apply these ideas to networked complex systems, including technology adoption, energy and mobility, and other infrastructure settings. A key objective is to develop models that can recognize when a prediction is weakly supported by available data.
"These systems have structure, laws and constraints, and the consequences of getting things wrong are real," Scaglione said. "I'm particularly interested in whether AI can learn representations that capture something fundamental about the underlying physics, so that a model trained on one network can work on a different configuration - or even a completely different network, at a different scale - without starting over. If we can understand when that is possible, we can move from AI that is impressive on a benchmark to AI that is genuinely useful in the physical world."
