Bayesian Optimization of Multiobjective Functions Using Multiple Information Sources

Abstract

Multiobjective optimization is often a difficult task owing to the need to balance competing objectives. A typical approach to handling this is to estimate a Pareto frontier in objective space by identifying nondominated points. This task is typically computationally demanding owing to the need to incorporate information of high enough fidelity to be trusted in design and decision-making processes. In this work, we present a multi-information source framework for enabling efficient multiobjective optimization. The framework allows for the exploitation of all available information and considers both potential improvement and cost. The framework includes ingredients of model fusion, expected hypervolume improvement, and intermediate Gaussian process surrogates. The approach is demonstrated on a test problem and an aerostructural wing design problem.