BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:CDM-MTEI Seminars: Performance Optimization Programming: Methodolo
 gy\, Algorithms and Insights
DTSTART:20260501T113000
DTEND:20260501T130000
DTSTAMP:20260416T122854Z
UID:03e83353f9314bfa798b0877cbb0cd42063a98cc8235b60c034fb6f4
CATEGORIES:Conferences - Seminars
DESCRIPTION:Professor Bart VAN PARYS\n\nDutch National Institute for Mathe
 matics and Computer Science (CWI)\nAbstract:\nThe design of efficient opti
 mization algorithms has long relied on intuition and hand-crafted analysis
 . A newer paradigm -- Performance Optimization Programming (POP) -- refram
 es the question of algorithm analysis as a metaoptimization problem\, enab
 ling a computer-assisted\, systematic search for optimal methods. We intro
 duce Branch-and-Bound Performance Estimation Programming (BnB-PEP)\, a uni
 fied methodology for constructing certifiably optimal first-order methods 
 for both convex and nonconvex optimization. Prior PEP-based approaches for
 mulate worst-case performance as a convex semidefinite program\, but this 
 convexity is simultaneously a limitation: only a narrow class of problem s
 etups admit such tractable reformulations. BnB-PEP overcomes this barrier 
 by directly embracing nonconvexity. BnB-PEP is applied to several settings
  that were previously intractable: constructing the optimal gradient metho
 d without momentum\, deriving optimal methods for reducing gradient norms 
 of smooth nonconvex functions\, and designing efficient methods for weakly
  convex nonsmooth optimization. In each case\, the methodology yields new 
 state-of-the-art convergence bounds. Finally\, we show how the analysis ge
 neralizes to settings with gradient corruption and allows for the discover
 y of an entirely novel class of first-order optimization algorithms which 
 enjoy optimal corruption protection.\n\n\nReferences:\n\n	[Das Gupta\, van
  Parys\, Ryu: Branch-and-Bound Performance Estimation Programming](https:/
 /arxiv.org/pdf/2203.07305)\n	[Gösgens\, van Parys: Subgradient Methods fo
 r Nonsmooth Convex Functions with Adversarial Errors](https://arxiv.org/ab
 s/2510.03072)\n\n\n\nShort bio: Bart van Parys is a senior researcher in t
 he Stochastics groupt at the Dutch National Institute for Mathematics and 
 Computer Science (CWI). Bart has played a pioneering role in developing di
 stributionally robust optimization (DRO) as a practical tool for decision-
 making in uncertain environments. His doctoral work pioneered DRO formulat
 ions in stochastic control theory and power systems. Key innovations were 
 an extension of the inequality of Gauss from 1821 in terms of tractable co
 nvex optimization problems.\nMore recently\, Bart formalized the search fo
 r statistically efficient data-driven decisions by advancing a novel meta-
 optimization framework via large-deviations theory. Bart's current researc
 h is primarily methodological and revolves around fundamental concerns whe
 n making decisions based on wild data sets and the automated design of bet
 ter optimization algorithms.\n 
LOCATION:ODY 4 03 https://plan.epfl.ch/?room==ODY%204%2003 https://epfl.zo
 om.us/meeting/register/-K7fsTI-QaeQE6wowwJbtw
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR