Three assistant professors in the School of Operations Research and Information Engineering have been named winners of the National Science Foundation Faculty Early Career Development Program (CAREER) Award. Paul Gölz, Ziv Scully, and Soroosh Shafiee were chosen on the strength of their research proposals in addition to their “potential to serve as academic role models in research and education and to lead advances in the mission of their department or organization.”
The CAREER Award is the National Science Foundation’s most prestigious award in support of early-career faculty. Awards vary in amount, but generally total at least $400,000 and cover a five-year period.
Gölz’s CAREER proposal (“Responsiveness to Heterogeneous Preferences in AI Alignment and Deployment”) explores a challenge that has become increasingly important as artificial intelligence systems are used by millions of people with different values, priorities and opinions. Most current AI systems are trained by combining feedback from many users into a single set of preferences, effectively treating everyone as though they share the same views. Gölz’s research asks whether AI systems can do a better job of recognizing and responding to the fact that people often disagree about what they want from technology.
Drawing on ideas from economics, mathematics and social choice theory—the study of how groups make collective decisions—Gölz will develop new methods for designing AI systems that balance competing preferences more fairly. He aims to create algorithms that can account for the needs of different groups of users while providing stronger guarantees that AI systems reflect a broad range of viewpoints. The research could help improve the alignment of large language models such as ChatGPT and inform the design of AI tools used in areas such as online communities and public decision-making.
Scully’s CAREER proposal (Efficient Scheduling for Machine Learning Training and Inference via the Gittins Index) focuses on one of the hidden challenges behind modern artificial intelligence: how to use computing resources more efficiently. As AI systems become larger and more widely used, they require enormous amounts of energy and computing power, both when they are being trained and when they are answering questions or completing tasks. Scully’s research aims to develop new mathematical tools that help computer systems decide how to allocate those resources more effectively, reducing costs and improving performance.
At the center of the project is a concept known as the Gittins index, a mathematical method for deciding how to prioritize tasks when their outcomes are uncertain. Scully will use this approach to improve two key stages of machine learning: training new AI models and running those models in real time. His goal is to create scheduling systems that can make better decisions about which computing jobs should receive priority, helping data centers and AI systems operate faster and more efficiently while consuming less energy.
Shafiee’s CAREER proposal (Robust Learning via Optimal Transport) addresses a growing concern in artificial intelligence: reliability. While AI systems often perform well under controlled conditions, they can make serious mistakes when confronted with unfamiliar situations, changing environments, or deliberately manipulated inputs. Shafiee’s research seeks to make AI systems more robust so that they can continue to perform safely and accurately when conditions differ from those encountered during training.
To accomplish this, Shafiee will develop new mathematical frameworks that help AI systems anticipate and withstand uncertainty. His work combines ideas from optimization, probability, and game theory to model a contest between an AI system and a hypothetical adversary attempting to expose its weaknesses. By better understanding this interaction, Shafiee hopes to create AI methods that are more dependable in high-stakes applications such as healthcare, finance, autonomous vehicles, and national security. The project also aims to provide stronger guarantees that AI systems will behave as expected even when faced with unexpected or challenging real-world conditions.
It is no surprise that all three proposals are focused on AI in one way or another. Research and teaching in the school are adjusting to recognize the ubiquitous role AI has very quickly come to play in society. In addition to research programs like those of Gölz, Scully, and Shafiee, the school has launched a new focused elective – Data, Decisions and AI – that brings together courses in machine learning, reinforcement learning, data mining, causal inference, and ethics, while explicitly connecting them to the core principles of operations research.
David Shmoys, the Laibe/Acheson Professor of Business Management & Leadership Studies, sees this focus on AI as a natural fit. “At its core, operations research is about building quantitative models that guide intelligent decision-making,” Shmoys said. “That process has always involved several interlocking components: defining what ‘better’ means through optimization, modeling uncertainty using probability combined with statistical tools, and developing algorithms to achieve it.
Those same components now underpin modern AI systems.
“When you go under the hood and look at how AI does what it does, the prediction mechanisms, the learning, the reasoning – it’s all using algorithms, methods and models that have been central to the operations research community for decades,” Shmoys said.