Robust Importance Sampling for Rare Events via Constrained Gaussian Mixtures
The study addresses estimating rare event probabilities where standard methods fail due to cold start and variance control issues. It proposes a two-stage framework: first, a coverage-first initialization to get a stable starting point, and second, constrained Gaussian Mixture Models (GMMs) refined using KL divergence objectives. This constraint ensures finite importance sampling variance, leading to highly robust and efficient estimators.