Bayesian optimisation in many dimensions with bespoke models

Опубликовано 20 мая 2017, 0:11

Bayesian optimisation (BO) is an optimisation method which incrementally builds a statistical model of the objective function to refine its search. Unfortunately, due to the curse of dimensionality, BO can fail to converge in problems with many dimensions. In this talk, I will show how better priors for BO can result in orders of magnitude improvements in convergence for problems with a known structure. I will introduce a class of models that are both useful and support inference at a reasonable computational cost.

See more on this video at microsoft.com/en-us/research/v...