BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Statistical Learning and Contextual Stochastic Optimization: Separ
 ate or Integrate?
DTSTART:20210921T170000
DTEND:20210921T190000
DTSTAMP:20260407T002748Z
UID:0dfa8bd8a88f34ddac3f5fb9984d82478c5ef9ed4475de1b020d1ef8
CATEGORIES:Conferences - Seminars
DESCRIPTION:Nathan Kallus\, Cornell University\nSeminar organized by the M
 anagement of Technology & Entrepreneurship Institute\n\nTitle\n"Statistica
 l Learning and Contextual Stochastic Optimization: Separate or Integrate?"
 \n\nSpeaker\nNathan Kallus\, Cornell University\n\nAbstract\nContextual st
 ochastic optimization (CSO) conditions on predictive observations availabl
 e at decision time\, which can reduce uncertainty and boost performance\, 
 but it also requires we learn a potentially complex predictive relationshi
 p between predictive observations and uncertain cost variables. While off-
 the-shelf ML methods can often be used to learn this relationship\, their 
 training loss ignores the downstream optimization task. Alternatively\, we
  can train the predictive model in an end-to-end fashion to directly opt
 imize the downstream costs of the decision policy it would induce. In this
  talk I will tackle the question\, which is better? Should we separate or 
 integrate the learning and optimization tasks?\n\nIn the first part of thi
 s talk I will focus on contextual linear optimization (CLO)\, where the co
 st function is bilinear in the decision and uncertain variables and the on
 ly relevant aspect of the predictive relationship is the conditional expec
 tation (aka regression function). Surprisingly\, I will show that the naiv
 e separated approach actually achieves regret convergence rates that are s
 ignificantly faster than any end-to-end method that directly optimizes dow
 nstream decision performance. I show this by leveraging the fact that spec
 ific problem instances do not have arbitrarily bad near-dual-degeneracy an
 d developing appropriate upper and lower bounds. This is overall positive 
 for practice: predictive models are easy and fast to train using existing 
 ML tools\, simple to interpret and reuse as a prediction\, and\, as shown\
 , lead to decisions that perform very well.\n\nIn the second part of this
  talk I will focus on the more general CSO problem\, where we must conside
 r the whole conditional probability model. As this object can be much high
 er dimensional than any decision policy\, it is better to integrate the ta
 sks and directly learn a policy. We adapt random forests to this integrate
 d task by searching tree splits to directly optimize downstream decisions\
 , rather than prediction accuracy. We solve this seemingly intractable pro
 blem by developing approximate splitting criteria that utilize optimizatio
 n perturbation analysis to eschew burdensome re-optimization for every can
 didate split\, so that our method scales to large-scale problems. We prove
  that our approximations are consistent and that our method is asymptotica
 lly optimal\, and we empirically validate its superior performance.\n\nThi
 s talk is based on the following papers:\nFast Rates for Contextual Linear
  Optimization\, with Y. Hu and X. Mao.\nhttps://arxiv.org/abs/2011.03030\n
 Stochastic Optimization Forests\, with X. Mao\nhttps://arxiv.org/abs/2008.
 07473\n\n\n\n 
LOCATION:ODY 0 16 https://plan.epfl.ch/?room==ODY%200%2016
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR