Grey-box Bayesian Optimization by Peter Frazier
0A Google TechTalk, presented by Peter I. Frazier, 2021/06/08
ABSTRACT: Bayesian optimization is a powerful tool for optimizing time-consuming-to-evaluate non-convex derivative-free objective functions. While BayesOpt has historically been deployed as a black-box optimizer, recent advances show considerable gains by "peeking inside the box". For example, when tuning hyperparameters in deep neural networks to minimize validation error, state-of-the-art BayesOpt tuning methods leverage the ability to stop training early, restart previously paused training, perform training and testing on a strict subset of the available data, and warm-start from previously tuned network architectures. We describe new "grey box" Bayesian optimization methods that selectively exploit problem structure to deliver state-of-the-art performance. We then briefly describe applications of these methods to tuning deep neural networks, inverse reinforcement learning and calibrating physics-based simulators to observational data.
About the speaker: Peter Frazier is the Eleanor and Howard Morgan Professor of Operations Research and Information Engineering at Cornell University. He is also a Staff Data Scientist at Uber. He leads Cornell's COVID-19 Mathematical Modeling Team, which designed Cornell's testing strategy to support safe in-person education during the pandemic. His academic research during more ordinary times is in Bayesian optimization, incentive design for social learning and multi-armed bandits. At Uber, he managed UberPool's data science group and currently helps to design Uber's pricing systems.