Statistics and Data Science Seminar: Jack Buckingham (University of Warwick)

Title: Bayesian Optimisation under Uncertainty
Speaker: Jack Buckingham (University of Warwick)
Location: Harrison 170

Abstract: Bayesian optimisation (BO) is a sample-efficient, black-box optimisation algorithm. It makes use of a probabilistic surrogate for the objective, usually a Gaussian process, to trade off between exploring regions of the parameter space with high uncertainty and exploiting regions already known to contain good values. In this talk, I will introduce BO and specifically discuss three applications, each containing an additional source of uncertainty: two-stage stochastic programming, reliability maximisation and (time permitting) decoupling objectives in multi-objective problems. I will formulate the so-called ‘acquisition functions’ used to quantify the value-of-information of different potential samples, and outline computational strategies used to optimise them.