Abstract: Bayesian active learning algorithms are currently viewed as the state of the art for design and optimization of black-box functions, in large part because of their ability to compromise between exploration and exploitation before each function evaluation. We show, however, that the criteria commonly used for sample selection fall short when the black box has the ability to generate rare and extreme events, i.e., events that combine high-magnitude impact with low probability of occurrence. We introduce new criteria that guide the algorithm towards highly anomalous—and therefore highly relevant—regions of the search space by leveraging previously collected information in a mathematically elegant and computationally tractable manner. We demonstrate the proposed approach in several applications related to uncertainty quantification, extreme-event prediction, and environment monitoring.
Bio: Antoine Blanchard is a Postdoctoral Associate in the Department of Mechanical Engineering at MIT. He received a Ph.D. in Aerospace Engineering from UIUC in 2017. His research interests include dynamical systems theory, machine learning, extreme events, computational physics, and uncertainty quantification.