Batch Bayesian Optimization via Particle Gradient Flows

  • 2022-09-10 19:10:15
  • Enrico Crovini, Simon L. Cotter, Konstantinos Zygalakis, Andrew B. Duncan
  • 20

Abstract

Bayesian Optimisation (BO) methods seek to find global optima of objectivefunctions which are only available as a black-box or are expensive to evaluate.Such methods construct a surrogate model for the objective function,quantifying the uncertainty in that surrogate through Bayesian inference.Objective evaluations are sequentially determined by maximising an acquisitionfunction at each step. However, this ancilliary optimisation problem can behighly non-trivial to solve, due to the non-convexity of the acquisitionfunction, particularly in the case of batch Bayesian optimisation, wheremultiple points are selected in every step. In this work we reformulate batchBO as an optimisation problem over the space of probability measures. Weconstruct a new acquisition function based on multipoint expected improvementwhich is convex over the space of probability measures. Practical schemes forsolving this `inner' optimisation problem arise naturally as gradient flows ofthis objective function. We demonstrate the efficacy of this new method ondifferent benchmark functions and compare with state-of-the-art batch BOmethods.

 

Quick Read (beta)

loading the full paper ...