BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Adapting to Conflict: Equilibrium structure and adaptive learning 
 in harmonic games
DTSTART:20260409T111500
DTEND:20260409T120000
DTSTAMP:20260430T083121Z
UID:c02b603eb23c8c2289ff33924507ea3cea40a7ce885dce5b7bf588df
CATEGORIES:Conferences - Seminars
DESCRIPTION:Davide Legacci\, PhD candidate in Computer Science at Universi
 té Grenoble Alpes\, CNRS and Inria\,\nAbstract\nWe propose an adaptive\, 
 “parameter-agnostic” algorithm for learning in harmonic games\, a clas
 s of finite games characterized by strictly opposed\, conflicting interest
 s. In contrast to zero-sum games—which do not always capture conflict\, 
 even between two players—harmonic games may exhibit highly non-convex eq
 uilibrium configurations. In particular\, if the game’s ambient dimensio
 n is odd\, we show that the set of totally mixed Nash equilibria is a (typ
 ically non-convex) real algebraic variety of dimension at least one\, whic
 h extends all the way to the boundary of the game’s strategy space. Desp
 ite the difficulties incurred by this complicated equilibrium structure\, 
 we provide an extrapolation-based variant of “follow-the-regularized-lea
 der” (FTRL) which converges to Nash equilibrium with order-optimal regre
 t guarantees: O(1) in self-play\, and O( sqrt(T) ) against arbitrary play.
 \n\n\nBiography\nDavide Legacci is a PhD candidate in Computer Science at 
 Université Grenoble Alpes\, CNRS and Inria\, within the GHOST team\, supe
 rvised by Panayotis Mertikopoulos and Bary S. R. Pradelski. His research l
 ies at the intersection of game theory and online learning\, with a focus 
 on the geometric structure of competitive interactions and the design of n
 o-regret learning dynamics for adversarial environments. His work has been
  presented at leading machine learning venues including ICML and NeurIPS. 
 He is expected to defend his PhD in September 2026.\n 
LOCATION:ME B0 374 https://plan.epfl.ch/?room==ME%20B0%20374
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR