Iterative importance sampling for estimating expectation bounds under partial probability specifications.

Abstract

In this paper, we explore and enhance importance sampling techniques for calculating lower and upper expectations with respect to sets of probability distributions. We formalize an iterative algorithm that we proposed in earlier work, by formulating the algorithm as a procedure for identifying a fixed point. We show how the algorithm can break down under poor coverage of the sampling distribution, and explore simple methods to increase coverage and thereby improve the algorithm.