Exploratory distributions for convex functions
Mathematical Statistics and Learning · 2018
DOI: 10.4171/msl/1-1-3 ↗Kurz gefasst
Fachartikel in der Zeitschrift Mathematical Statistics and Learning, 2018. Verfasst von Sébastien Bubeck und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Given a Lipschitz and convex function f on a compact and convex domain in \mathbb R^n , we construct an exploratory distribution \mu of f in the following sense. Let g be a Lipschitz and convex function on the same domain, such that either g=f , or alternatively the minimum of g is \epsilon smaller than the minimum of f . Then \mu is such that \mathrm{poly}(n/\epsilon) noisy evaluations of g at i.i.d. points from \mu suffices to determine with high probability whether g=f or g \neq f . As an example of application for such exploratory distributions we show how to use them to estimate the minimax regret for adversarial bandit convex optimization.
Zitieren
APA
Bubeck, S., & Eldan, R. (2018). Exploratory distributions for convex functions. Mathematical Statistics and Learning. https://doi.org/10.4171/msl/1-1-3
Vancouver
Bubeck S, Eldan R. Exploratory distributions for convex functions. Mathematical Statistics and Learning. 2018. doi:10.4171/msl/1-1-3
BibTeX
@article{Bubeck2018,
author = {Bubeck, Sébastien and Eldan, Ronen},
title = {Exploratory distributions for convex functions},
journal = {Mathematical Statistics and Learning},
year = {2018},
doi = {10.4171/msl/1-1-3}
}
EUVIMED