Sparsifying to optimize over multiple information sources: an augmented Gaussian process based algorithm

Abstract Optimizing a black-box, expensive, and multi-extremal function, given multiple approximations, is challenging task known as multi-information source optimization (MISO), where each has different cost the level of approximation (aka fidelity ) can change over search space. While most current approaches fuse Gaussian processes (GPs) modelling source, we propose to use GP sparsification select only “reliable” function evaluations performed all sources. These selected are used create an augmented process (AGP), whose name implied by fact that on expensive with reliable less A new acquisition based confidence bound, also proposed, including both next query location-dependent source. This estimated through model discrepancy measure prediction uncertainty GPs. MISO-AGP MISO-fused counterpart compared two test problems hyperparameter machine learning classifier large dataset.