BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Time-Dependent Blackwell Approachability and Application to Absorb
 ing Games
DTSTART:20250403T090000Z
DTEND:20250403T101500Z
DTSTAMP:20260504T201300Z
UID:indico-event-12958@indico.math.cnrs.fr
DESCRIPTION:Speakers: Yijun WAN (PSL - Université Paris Dauphine)\n\nBlac
 kwell’s approachability is a general online learning framework where a D
 ecision Maker obtains vector-valued outcomes\, and aims at the convergence
  of the average outcome to a given “target” set. Blackwell gave a suff
 icient condition for the decision maker having a strategy guaranteeing suc
 h a convergence against an adversarial environment\, which ensures converg
 ence. Blackwell’s approachability has since been applied to numerous pro
 blems\, in regret minimization and game theory\, in particular. We extend 
 this framework by allowing the outcome function and the inner product to b
 e time-dependent. We establish a general guarantee for the natural extensi
 on to this framework of Blackwell’s algorithm. In the case where the tar
 get set is an orthant\, we present a family of time-dependent inner produc
 ts which yields different convergence speeds for each coordinate of the av
 erage outcome. We apply this framework to absorbing games (an important cl
 ass of stochastic games) for which we construct ε-uniformly optimal strat
 egies using Blackwell’s algorithm in a well-chosen auxiliary approachabi
 lity problem\, thereby giving a novel illustration of the relevance of onl
 ine learning tools for solving games. -- Paper\n\nhttps://indico.math.cnrs
 .fr/event/12958/
LOCATION:Auditorium 3 (Toulouse School of Economics)
URL:https://indico.math.cnrs.fr/event/12958/
END:VEVENT
END:VCALENDAR