Session X (May 17, 10:30am-12:00pm): Sequential Design, Active Learning, and Bayesian Optimization, organized by Qiong Zhang
Title: Bayesian Optimization with Human Feedback
Speaker: Peter Frazier, Cornell U
Abstract: Bayesian optimization streamlines optimization of time-consuming-to-evaluate functions. Traditionally a black-box approach, our recent work reveals significant performance gains by incorporating additional information, a "grey-box" approach. This talk reveals how grey-box BayesOpt supports human-in-the-loop scenarios, offering a novel approach for individuals struggling to define a single objective function. Instead of estimating Pareto frontiers, 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 experiments effectively, resulting in a set of solutions with likely near-maximum utility. This strategy outperforms traditional multi-objective methods by better leveraging user preferences, as demonstrated via an application to product design at Meta.