Anytime Backtrack Unimodal Bandits and Applications to Cloud Computing, - Equipe : Graphes, Algorithmes et Combinatoire Accéder directement au contenu
Communication Dans Un Congrès Année : 2019

Anytime Backtrack Unimodal Bandits and Applications to Cloud Computing,

Stephan Kunne
  • Fonction : Auteur
Lorenzo Maggi
  • Fonction : Auteur
Xu Xinneng
  • Fonction : Auteur

Résumé

Unimodal functions appear ubiquitously in resource allocation problems. They arise whenever the expected value f(x) of the metric to be optimized has a single peak as a function of the available control variable x. The exact shape of f is often not known in advance, hence it can only be learned via sampling. For this purpose, we exploit the formalism of multi-armed bandits to design a learning algorithm that i) converges in probability to the optimal arm maximizing f and ii) does not require to know the learning horizon in advance, i.e., it is anytime. Moreover, it adapts naturally to the scenario where iii) prior knowledge on the location of the peak is available, even if the prior is inaccurate. We also present an efficient heuristic that can use explicitly the prior belief on the location of the optimal arm. We demonstrate the applicability of our approaches to a basic resource allocation problem for cloud computing, where the orchestrator adapts the resources allocated to a client to its dynamic requests.
Fichier non déposé

Dates et versions

hal-03008647 , version 1 (16-11-2020)

Identifiants

Citer

Stephan Kunne, Lorenzo Maggi, Johanne Cohen, Xu Xinneng. Anytime Backtrack Unimodal Bandits and Applications to Cloud Computing,. IFIP Networking Conference (Networking), Jun 2020, Paris, France. pp.82-90, ⟨10.1007/s11276-017-1629-4⟩. ⟨hal-03008647⟩
73 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More