BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Dmitriy Drusvyatskiy (University of Washington) DTSTART:20201123T143000Z DTEND:20201123T153000Z DTSTAMP:20260423T021004Z UID:OWOS/23 DESCRIPTION:Title: St ochastic Optimization with Decision-dependent Distributions\nby Dmitri y Drusvyatskiy (University of Washington) as part of One World Optimizatio n seminar\n\n\nAbstract\nStochastic optimization problems often involve da ta\ndistributions that change in reaction to the decision variables. For\n example\, deployment of a classifier by a learning system\, when made\npub lic\, often causes the population to adapt their attributes in order\nto i ncrease the likelihood of being favorably labeled---a process\ncalled ``ga ming''. Even when the population is agnostic to the\nclassifier\, the deci sions made by the learning system (e.g. loan\napproval) may inadvertently alter the profile of the population (e.g.\ncredit score). Recent works hav e identified an intriguing solution\nconcept for such problems as an ``equ ilibrium'' of a certain game.\nContinuing this line of work\, we show that typical stochastic\nalgorithms---originally designed for static problems- --can be applied\ndirectly for finding such equilibria with little loss in efficiency.\nThe reason is simple to explain: the main consequence of the \ndistributional shift is that it corrupts the algorithms with a bias\ntha t decays linearly with the distance to the solution. Using this\nperspecti ve\, we obtain sharp convergence guarantees for popular\nalgorithms\, suc h as stochastic gradient\, clipped gradient\, proximal\npoint\, and dual a veraging methods\, along with their accelerated and\nproximal variants.\n\ nJoint work with Lin Xiao (Facebook AI Research)\n\nThe address and passwo rd of the zoom room of the seminar are sent by e-mail on the mailinglist o f the seminar one day before each talk\n LOCATION:https://researchseminars.org/talk/OWOS/23/ END:VEVENT END:VCALENDAR