BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260203T233037EST-06515vNvVU@132.216.98.100
DTSTAMP:20260204T043037Z
DESCRIPTION:TITLE / TITRE\n\nFollow-the-Perturbed-Leader with Between-Actio
 n Dependence\n\n \n\nABSTRACT / RÉSUMÉ \n\nOnline learning is a framework 
 for analysing sequential decision-making problems in adversarial environme
 nts. Many important problems in statistics and computer science can be red
 uced to variants of online learning\, including finding equilibrium strate
 gies for multiplayer games\, stochastic and online optimization\, adaptive
  density estimation\, sequential testing\, adaptive experimental design\, 
 and online change-point detection. Design of algorithms for online learnin
 g can\, thus\, be used to study the information theoretic and computationa
 l limits of sequential decision making under uncertainty in a variety of c
 ontexts with numerous applications.\n\nAn important class of online learni
 ng algorithms are 'follow-the-perturbed leader' (FTPL) methods\, which mak
 e decisions by randomly simulating the future of the online learning game 
 and then playing the best response to the simulation. These methods are ty
 pically among those with the most efficient implementation. However\, the 
 analysis of FTPL performance in previous work has largely been limited to 
 cases where some strong 'between-action' independence assumptions are impo
 sed on the simulation of the future. This limits the broad applicability o
 f these methods when such independence is not appropriate.\n\nWe present a
  framework for analyzing Gaussian FTPL algorithms for full-information onl
 ine learning problems when the perturbation distribution exhibits between-
 action dependence. Applications include FTPL algorithms for online learnin
 g for i) infinite action spaces when the adversary plays bounded Lipschitz
  reward functions\, where the perturbations are random functions sampled f
 rom a Gaussian process\; and ii) linear polyhedral games\, where the pertu
 rbation is a random linear function. We demonstrate how to tightly account
  for dependence between actions in the FTPL analysis and present an ansatz
  for the selection of the perturbation distribution based on a Bayesian pe
 rspective of FTPL as a variant of Thompson sampling.\n\nPLACE /LIEU \n	Hybr
 ide - CRM\, Salle / Room 5340\, Pavillon André Aisenstadt\n\n \n\nLien ZOO
 M Link\n
DTSTART:20260130T203000Z
DTEND:20260130T213000Z
SUMMARY:Jeffrey Negrea (University of Waterloo)
URL:/mathstat/channels/event/jeffrey-negrea-university
 -waterloo-370589
END:VEVENT
END:VCALENDAR