The optimization of expensive black-box functions appears in many situations.
Bayesian optimization methods have been successfully applied to solve these prob-
lems using well-known single-point acquisition functions. Nowadays, the develop-
ments in technology allow us to perform evaluations of some of these expensive
function in parallel. Therefore, there is a need for batch infill criteria to consider
this parallelism.
In this paper, a novel batch infill criterion is presented. Our algorithm, at each
batch step, restricts the search space into a finite candidate pool and, from it, selects
the batch that maximizes the mutual information with respect to the objective
function. Depending on the total number of observations available, the candidate
pool will consist of a Latin Hypercube Sampling (exploration) or a Pareto Sampling
(trade-off between exploration and exploitation). We compare our strategy with
some of the state-of-the-art UCB-based batch approaches for different benchmark
objectives, outperforming or obtaining very similar results.
Citation
A novel UCB-based batch strategy for Bayesian optimization, C.Domínguez-Bravo and Jose A. Lozano, 2019