Séance Séminaire

Séminaire de Probabilités et Statistique

The geometry of Gaussian processes and Bayesian optimization

Chaining is known for years to be a powerful tool to study the supremum of stochastic processes. Yet, it did not appear in statistical learning techniques since very recently. In this talk we analyse Bayesian optimization in light of the geometry of Gaussian processes for their canonical distances. We show how chaining can be translated in efficient adaptive algorithms. We will present theoretical guarantees on the regret of the obtained method as well as numerical experiments.