Abstract

We consider robust optimization problems, where the goal is to optimize inthe worst case over a class of objective functions. We develop a reduction fromrobust improper optimization to Bayesian optimization: given an oracle thatreturns $\alpha$-approximate solutions for distributions over objectives, wecompute a distribution over solutions that is $\alpha$-approximate in the worstcase. We show that de-randomizing this solution is NP-hard in general, but canbe done for a broad class of statistical learning tasks. We apply our resultsto robust neural network training and submodular optimization. We evaluate ourapproach experimentally on corrupted character classification, and robustinfluence maximization in networks.

 

Quick Read (beta)