BEGIN:VCALENDAR
PRODID:-//eluceo/ical//2.0/EN
VERSION:2.0
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:www.tcs.tifr.res.in/event/1733
DTSTAMP:20260610T113313Z
SUMMARY:Constrained Online Convex Optimization: A Simple Algorithm and Slat
 er-Slack Reductions
DESCRIPTION:Speaker: Aakash Ghosh (TIFR)\n\nAbstract: \n\nOnline convex opt
 imization with adversarial\, online-revealed constraints (COCO) is a natur
 al model for safe sequential decision-making. However\, adversarial constr
 aints rule out simultaneous sublinear regret and per-round feasibility\, s
 o one instead controls aggregate violation. Two main approaches handle thi
 s: the Neely and Yu drift-plus-penalty method attains O(1) violation under
  a Slater condition but requires solving a convex program every round\, wh
 ile the recent work of Sinha and Vaze attains O(√T) regret and Õ(√T) 
 strict cumulative violation requiring only a single projected-gradient ste
 p. After surveying this landscape\, I ask a safe-learning question: if the
  learner additionally holds a baseline action that is strictly feasible by
  a margin\, what benefits does this provide? I present a black-box reducti
 on that wraps any base algorithm carrying a regret and violation certifica
 te. I treat three regimes: a known point and margin\, a known point with a
 n unknown margin\, and an unknown point with known margin lower bounds\, c
 oncluding with a discussion of the tradeoffs involved.\n
URL:https://www.tcs.tifr.res.in/web/events/1733
DTSTART;TZID=Asia/Kolkata:20260611T163000
DTEND;TZID=Asia/Kolkata:20260611T173000
LOCATION:A-201 (STCS Seminar Room)
END:VEVENT
END:VCALENDAR