It’s hard to know what people want. This is a major challenge when tuning generative AI systems and when deploying Bayesian optimization methods for human decision makers. This talk considers the problem of optimizing inputs to a time-consuming-to-evaluate black-box oracle with the goal of producing outputs that satisfy one or more human users with unknown preferences. We focus on two examples: choosing hyperparameters in a generative AI pipeline to produce human-consumed content; and choosing algorithm parameters in an online platform to produce a panel of A/B test results evaluated by business leaders. We model human preferences with an (unknown) utility function that can be queried via user interactions. Quantifying our uncertainty about the user’s utility function via a Bayesian approach, we iteratively update our posterior as we learn more from the user. This allows us to prioritize computation effectively and produce a set of solutions with likely near-maximum utility. By better leveraging user preferences, this method outperforms traditional multi-objective methods. We describe how this method was used to improve a generative AI pipeline at a major social media platform and its use in ongoing work to schedule final exams at Cornell.
Peter Frazier is the Eleanor and Howard Morgan Professor of Operations Research and Information Engineering at Cornell University. His research is at the interface between machine learning, operations research, and chemistry, including Bayesian optimization and multi-armed bandits. During the pandemic, he led Cornell’s COVID-19 Mathematical Modeling Team, which helped design Cornell’s asymptomatic COVID-19 testing program and provided university leadership with science-based decision support. From 2015-2024, he worked as a scientist at Uber and designed pricing systems. He recently left Uber to found Saddlepoint Labs, a company using computer vision and sensor data to improve the reproducibility of wet lab experiments. He is the winner of best paper awards from the ACM Conference on Economics and Computation, the INFORMS Section on Auctions and Market Design, the INFORMS Applied Probability Society, the INFORMS Computing Society, and the Winter Simulation Conference.