BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Separation of non-ergodic uniform convergence rates for regularize
 d learning in games
DTSTART:20260409T090000Z
DTEND:20260409T101500Z
DTSTAMP:20260607T052200Z
UID:indico-event-15001@indico.math.cnrs.fr
DESCRIPTION:Speakers: Julien Grand-Clément (HEC Paris)\n\nSelf-play via o
 nline learning is a leading paradigm for solving large-scale games and has
  enabled recent superhuman performance (e.g.\, Go\, Poker). This work clar
 ifies that different convergence notions in self-play (last iterate\, best
  iterate\, and a randomly sampled iterate) can behave fundamentally differ
 ently. For a broad class of learning dynamics\, including Optimistic Multi
 plicative Weights Update (OMWU)\, we prove a separation: even in two-playe
 r zero-sum games\, last-iterate convergence can be arbitrarily slow\, rand
 om-iterate convergence can be slower than any polynomial\, while best-iter
 ate convergence is polynomial. This departs from much prior theory where t
 hese notions align\, and we attribute the gap to OMWU’s insufficient “
 forgetfulness\,” linking it to empirical behavior in practical game solv
 ing.\nPaper 1 -- Paper 2\n\nhttps://indico.math.cnrs.fr/event/15001/
LOCATION:Auditorium 5 - 2nd floor (Toulouse School of Economics)
URL:https://indico.math.cnrs.fr/event/15001/
END:VEVENT
END:VCALENDAR