Bayesian Optimization of Composite Functions With Application to Computationally Expensive Inverse Problems

Abstract

We propose a novel Bayesian optimization (BayesOpt) algorithm for calibrating black-box derivative-free expensive-to-evaluate computer models. Our approach finds model parameters x that minimize f(x)=g(h(x)), where h(x) is the model’s vector-valued prediction and g(h(x)) is the sum of squared errors. Standard BayesOpt models f directly. By modeling h instead and leveraging knowledge of g, our approach outperforms standard BayesOpt by several orders magnitude on test problems.

Date
Jul 4, 2019 4:00 PM — 4:30 PM
Event
Applied Inverse Problems Conference 2019
Location
University of Grenoble Alpes
621 Avenue Centrale, Grenoble, 38400, France