Abstract
Batch Bayesian optimisation (BO) has shown to be a sample-efficient method ofperforming optimisation where expensive-to-evaluate objective functions can bequeried in parallel. However, current methods do not scale to large batch sizes-- a frequent desideratum in practice (e.g. drug discovery or simulation-basedinference). We present a novel algorithm, SOBER, which permits scalable anddiversified batch BO with arbitrary acquisition functions, arbitrary inputspaces (e.g. graph), and arbitrary kernels. The key to our approach is toreformulate batch selection for BO as a Bayesian quadrature (BQ) problem, whichoffers computational advantages. This reformulation is beneficial in solving BQtasks reciprocally, which introduces the exploitative functionality of BO toBQ. We show that SOBER offers substantive performance gains in synthetic andreal-world tasks, including drug discovery and simulation-based inference.