BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Date iCal//NONSGML kigkonsult.se iCalcreator 2.20.2//
METHOD:PUBLISH
X-WR-CALNAME;VALUE=TEXT:TSE
BEGIN:VTIMEZONE
TZID:Europe/Paris
BEGIN:STANDARD
DTSTART:20251026T030000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
RDATE:20261025T030000
TZNAME:CET
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20260329T020000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:calendar.138335.field_date.0@www.tse-fr.eu
DTSTAMP:20260510T205422Z
CREATED:20251006T161001Z
DESCRIPTION:Julien Grand-Clément (HEC Paris)\, “Separation of non-ergodic u
 niform convergence rates for regularized learning in games”\, MAD-Stat. Se
 minar\, Toulouse: TSE\, April 9\, 2026\, 11:00–12:15\, room Auditorium 5.
 \n\nSelf-play via online learning is a leading paradigm for solving large-
 scale games and has enabled recent superhuman performance (e.g.\, Go\, Pok
 er). This work clarifies that different convergence notions in self-play (
 last iterate\, best iterate\, and a randomly sampled iterate) can behave f
 undamentally differently. For a broad class of learning dynamics\, includi
 ng Optimistic Multiplicative Weights Update (OMWU)\, we prove a separation
 : even in two-player zero-sum games\, last-iterate convergence can be arbi
 trarily slow\, random-iterate convergence can be slower than any polynomia
 l\, while best-iterate convergence is polynomial. This departs from much p
 rior theory where these notions align\, and we attribute the gap to OMWU’s
  insufficient “forgetfulness\,” linking it to empirical behavior in practi
 cal game solving.
DTSTART;TZID=Europe/Paris:20260409T120000
DTEND;TZID=Europe/Paris:20260409T131500
LAST-MODIFIED:20260306T011001Z
LOCATION:Toulouse: TSE\, April 9\, 2026\, 11:00–12:15\, room Auditorium 5
SUMMARY:MAD-Stat. Seminar
URL;TYPE=URI:https://www.tse-fr.eu/seminars/2026-separation-non-ergodic-uni
 form-convergence-rates-regularized-learning-games
END:VEVENT
END:VCALENDAR