Date and Time
27 March 2019
3:00 pm - 4:30 pm


Venue
NUS Business School, Mochtar Riady Building, BIZ1, Seminar Room 3-2, 15 Kent Ridge Drive, Singapore 119245
View Map


Organised by: Analytic-operations



Abstract

Wasserstein distance based Distributionally Robust Optimization (DRO) has recently emerged as a popular tool towards tackling “optimizer’s curse”, while performing optimization under uncertainty in large scale. Unlike traditional stochastic optimization methods where the optimal decision choices minimise expected risk with respect to a fixed probability model, Wasserstein DRO aims to identify decisions that guarantee a uniform level of performance for any choice of probability distribution that is within a fixed radius (quantified by suitable Wasserstein distance) from a baseline model.

With data-driven Wasserstein DRO formulations finding widespread applications in Operations Research and Machine Learning, the need to understand the fundamental structural and statistical properties of its solutions is of importance to appropriately guide its utility in data-driven settings. To this end, the talk aims to provide an overview of recent developments on

  1. structural properties such as strong convexity and comparative statics; and
  2. qualitative statistical properties such as solution selection properties, rate of convergence of selected solutions and asymptotic normality.

We use the resulting insights to guide solving issues of practical relevance such as designing fast iterative schemes, optimally selecting radius of the Wasserstein ball and construction of associated confidence regions.