Abstract

Bayesian optimisation (BO) is a popular routine for optimising functions that are plagued by high evaluation costs. In BO, an acquisition function is used to sequentially focus evaluations into promising areas of the search space, allowing the identification of competitively good solutions within even heavily restricted evaluation budgets. A recent particularly intuitive and empirically effective class of acquisition functions has arisen based around information theory. During this talk, I will motivate these powerful entropy-based methods, before describing a general-purpose extension built around a simple approximation of information gain—an information-theoretic quantity central to solving a range of BO problems, including noisy, multi-fidelity and batch optimisations. Previously, these tasks have been tackled separately within information-theoretic BO, each requiring a different sophisticated approximation scheme, except for batch BO, for which no computationally-lightweight information-theoretic approach has previously been proposed. GIBBON (General-purpose Information-Based Bayesian OptimisatioN) provides a single principled framework suitable for all the above, out-performing existing approaches whilst incurring substantially lower computational overheads.