Reproduce “Sinkhorn Distributionally Robust Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jie Wang, Rui Gao, Yao Xie
Sinkhorn Distributionally Robust Optimization (DRO) introduces a framework using Sinkhorn distance to address computational inefficiencies and modeling conservatism of Wasserstein DRO. It achieves this by establishing a general strong duality reformulation and proposing efficient stochastic optimization algorithms, which enable absolutely continuous worst-case distributions and lead to improved out-of-sample performance in diverse applications.
Key insights
- Strong duality for Sinkhorn DRO holds under remarkably general assumptions, encompassing arbitrary loss functions, transport costs, nominal distributions, and probability supports, extending the theoretical scope of DRO.
- The worst-case distributions in Sinkhorn DRO are proven to be absolutely continuous with respect to a reference measure, offering a more realistic and less conservative representation of uncertainty compared to the discrete worst-case distributions of Wasserstein DRO.
- The entropic regularization embedded in the Sinkhorn distance results in a smoothed dual objective function, which significantly enhances the computational tractability of the DRO problem.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.