Bayesian Optimisation Beyond Modestly-dimensioned Continuous Problems

Abstract

Bayesian optimisation (BO) is an area within machine learning in which demand outstrips supply, driven by industrial interest in automated machine learning (AutoML). However, most “standard” endeavours in BO have focused on continuous-valued and modestly-dimensioned problems, although real-life problems can be considerably higher-dimensional, more heterogeneous and thus more “exotic”. This talk introduces some recent advances of BO in these challenging data structures that significantly differ from the “standard”: we propose NAS-BOWL for BO in a graph-like space for neural architecture search, and Casmopolitan for high-dimensional spaces with categorical variables. I hope these could be first steps towards broadening the scope of application of BO beyond its current limitations.