Data Driven Optimization Models and Algorithms
summary
Describes tractable DRO constructions (moment, likelihood and Wasserstein distance sets) that reduce to semidefinite programs and give finite‑sample confidence guarantees; quantifies the price of correlation (submodular objectives have a small gap — about e/(e−1) — while other cases blow up exponentially); and analyzes online packing with provable competitive ratios, showing minimum inventory B must scale with log M/ε^2 to get 1−ε, while dynamic learning via updating shadow prices drives that sample requirement down toward the ε^2 optimal.
AI-written summary · about
chapters (5)
- 0:00Introduction and logistics
- 2:08Distributionally robust opt.
- 22:51Price of correlation
- 37:19Online learning algorithms
- 1:05:37Q&A and concluding remarks
related talks
35mOnline Optimization and Learning Under Long-Term Convex Constraints and Objectives
55mZeroth-Order Methods for Convex-Concave Minmax Problems: Learning from Strategically Generated Data
15mFundamental Limits Of Learning In Data-Driven Problems
45mCharting the Landscape of Memory/Data Tradeoffs in Continuous Optimization: ...
01h 12mTutorial: Federated Optimization, Part I
01h 01mRobustness, Stochastics, Uncertainty 2
56mCompetitive Analysis of Online Algorithms (Part 1)
52mConstrained Subset Selection Problems under Diversity Measures