Hybrid Monte Carlo for Failure Probability Estimation with Gaussian Process Surrogates

Annie S. Booth, Ashwin Renganathan

Abstract:

We tackle the problem of quantifying failure probabilities for expensive computer experiments with stochastic inputs. The computational cost of evaluating the computer simulation prohibits direct Monte Carlo (MC) and necessitates a statistical surrogate model. Surrogate-informed importance sampling – which leverages the surrogate to identify suspected failures, fits a bias distribution to these locations, then calculates failure probabilities using a weighted average – is popular, but it is data hungry and can provide erroneous results when budgets are limited. Instead, we propose a hybrid MC scheme which first uses the uncertainty quantification (UQ) of a Gaussian process (GP) surrogate to identify areas of high classification uncertainty, then combines surrogate predictions in certain regions with true simulator evaluation in uncertain regions. We also develop a stopping criterion which informs the allocation of a fixed budget of simulator evaluations between surrogate training and failure probability estimation. Our method is agnostic to surrogate choice (as long as UQ is provided); we showcase functionality with both GPs and deep GPs. It is also agnostic to design choices; we deploy contour locating sequential designs
throughout. With these tools, we are able to effectively estimate small failure probabilities with only hundreds of simulator evaluations. We validate our method on a variety of synthetic benchmarks before deploying it on an expensive computer experiment of fluid flow around an airfoil.