Which Gaussian Process for Bayesian Optimization ?
Résumé
Bayesian Optimization (BO) is a popular approach to the global optimization of costly non-convex functions in moderate dimension. BO is based on Gaussian processes that are iteratively learned and serve as a model to control the exploration-exploitation trade-off through an acquisition criterion. Much of the recent research on BO has focused on new acquisition criteria and the specialization to specific problems (e.g., uncertain or multi-fidelity contexts).
In this talk, on the contrary, we consider a standard black-box single objective problem and a standard acquisition criterion, the expected improvement. The focus is the Gaussian process (GP) and how it can be modified to improve the optimization.
Three directions for progress are discussed: the trend of the GP, the adaptation to higher dimension by a linear embedding, and the collaboration between a global and a local GP through a trust-region mechanism.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|