Surrogate-Guided Adaptive Importance Sampling for Failure Probability Estimation

TL;DR

This paper introduces KDE-AIS, a novel single-stage importance sampling method using Gaussian process surrogates and kernel density estimation for efficient failure probability estimation with limited oracle evaluations.

Contribution

The paper proposes a new single-stage approach combining Gaussian process surrogates with kernel density estimation for adaptive importance sampling, improving efficiency over existing methods.

Findings

KDE-AIS converges asymptotically to the optimal zero-variance IS density.

KDE-AIS achieves more accurate failure probability estimates with fewer oracle evaluations.

Empirical results outperform previous Gaussian process-based adaptive importance sampling methods.

Abstract

We consider the sample efficient estimation of failure probabilities from expensive oracle evaluations of a limit state function via importance sampling (IS). In contrast to conventional ``two stage'' approaches, which first train a surrogate model for the limit state and then construct an IS proposal to estimate failure probability using separate oracle evaluations, we propose a \emph{single stage} approach where a Gaussian process surrogate and a surrogate for the optimal (zero-variance) IS density are trained from shared evaluations of the oracle, making better use of a limited budget. With such an approach, small failure probabilities can be learned with relatively few oracle evaluations. We propose \emph{kernel density estimation adaptive importance sampling} (\texttt{KDE-AIS}), which combines Gaussian process surrogates with kernel density estimation to adaptively construct the IS proposal density, leading to sample efficient estimation of failure probabilities. We show that \texttt{KDE-AIS} density asymptotically converges to the optimal zero-variance IS density in total variation. Empirically, \texttt{KDE-AIS} enables accurate and sample efficient estimation of failure probabilities compared to the state of the art, including previous work on Gaussian process based adaptive importance sampling.