Date : July 1, 2021, 3:30 p.m. - Room :Visio-conférence

Keynote 2 - Grey-box Bayesian Optimization

Peter FRAZIER - Cornell University (US)


Bayesian optimization (BayesOpt) 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, these new 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. When calibrating physics-based simulators to observational data, these new state-of-the-art methods use more information than classical methods: the residuals between each observational datapoint and the simulator's prediction, rather than just the overall quality of fit. We describe these new "grey box" BayesOpt methods that selectively exploit problem structure to deliver substantial performance improvements. We focus first on objective functions computed via a composition or network of individually-observable functions, describing methods for such problems in detail. We then survey a broader set of methods and research opportunities in this area.



Peter Frazier received a B.S. in Physics and Engineering/Applied Science from the California Institute of Technology in 2000, after which he spent several years in industry as a software engineer, working for two different start-up companies and for the Teradata division of NCR. In 2005, he entered graduate school in the Department of Operations Research and Financial Engineering at Princeton University, and received an M.A. in 2007 and a Ph.D. in 2009. He joined the faculty at Cornell in 2009 as an Assistant Professor in the School of Operations Research and Information Engineering, where he is now an Associate Professor.

His research is in sequential decision-making under uncertainty, optimal methods for collecting information, and machine learning, focusing on applications in simulation, e-commerce, medicine and biology. He is the recipient of a CAREER Award from the National Science Foundation and a Young Investigator Award from the Air Force Office of Scientific Research. He is currently on leave at Uber, where he is a Staff Data Scientist and Data Science Manager. At Uber, he worked on UberPOOL from 2015-17, and on broader pricing efforts from 2016-17. He now leads a data science team focused on pricing.

Plus d'informations sur sa page web: ou sur sa page dédiée à l'université


Lien de connexion: zoom -

Passcode: keynote