Exploration and Optimization of Black-Box Functions with Safety Guarantees

October 22, 2024 3:30 PM Singapore (Registration starts at 3:20 PM)

Abstract

The exploration and optimization of black-box functions is of broad interest in domains such as hyperparameter tuning, robotics, scientific discovery, and more. One of the prevailing techniques has been Gaussian Process methods, in which the function’s smoothness properties are encoded via a suitably-chosen kernel. In this talk, we are interested in safety considerations in this topic, in which certain parts of the search space are “unsafe” (e.g., drone parameters that may cause a crash, unsafe drug dosages in medical trials, etc.) and must be avoided.

In the first part of the talk, I will give an overview of the standard setting without safety requirements. Afterwards, for the safe setting, I will cover our recent paper highlighting the benefits of monotonicity assumptions: A single controllable “safety variable” controlling the degree of “cautiousness” can considerably help in avoiding unsafe regions, finding a global maximum, and attaining improved theoretical guarantees.

PAPERS:
https://proceedings.mlr.press/v238/losalka24a.html

About the Speaker

Jonathan Scarlett is an Associate Professor in the Department of Computer Science and Department of Mathematics, National University of Singapore. His research interests are in the areas of information theory, machine learning, signal processing, and high-dimensional statistics. He is a recipient of the Singapore National Research Foundation (NRF) fellowship.

Previously, Jonathan received the B.Eng. degree in electrical engineering and the B.Sci. degree in computer science from the University of Melbourne, Australia. From October 2011 to August 2014, he was a Ph.D. student in the Signal Processing and Communications Group at the University of Cambridge. From September 2014 to September 2017, he was post-doctoral researcher at LIONS, EPFL.