ROBUST UNCERTAINTY QUANTIFICATION USING PRECONDITIONED LEAST-SQUARES POLYNOMIAL APPROXIMATIONS WITH l1-REGULARIZATION

Abstract

We propose a non-iterative robust numerical method for the non-intrusive uncertainty quantification of multivariate stochastic problems with reasonably compressible polynomial representations. The approximation is robust to data outliers or noisy evaluations which do not fall under the regularity assumption of a stochastic truncation error but pertains to a more complete error model, capable of handling interpretations of physical/computational model (or measurement) errors. The method relies on the cross-validation of a pseudospectral projection of the response on generalized Polynomial Chaos approximation bases; this allows an initial model selection and assessment yielding a preconditioned response. We then apply a 1−penalized regression to the preconditioned response variable. Nonlinear test cases have shown this approximation to be more effective in reducing the effect of scattered data outliers than standard compressed sensing techniques and of comparable efficiency to iterated robust regression techniques.

Cardiovascular Modeling With Adapted Parametric Inference

“…There exists different ways of deriving sample weights from the discrepancy measure, but the main idea is to make it inversely proportional with some numerical bounding capability, i.e. ρ j ∼ max( , ∆ j ) −1 [23]. If all ρ j are large, it means that the surrogate is quite accurate in the sampled approximate posterior region, implying in turn thatπ…”

Goal-oriented error control of stochastic system approximations using metric-based anisotropic adaptations

“…We have seen up until now the adaptive strategies and algorithms when solving the optimisation problems (28) or (33). However, as part of the motivation and novelty of this paper we wish to: (a) quantify the deterministic error at each sample of the parameters space and compute a local (sample-wise) optimization deterministic problem with a required complexity C x and to (b) be able to choose which error, stochastic or deterministic, dominates a computation and thus to solve the corresponding problem.…”

“…compressible airflow around an aircraft). Putting aside modeling error, sometimes leading to unpredictable outlying fluctuations [28], a key element of the computational pipeline is the mesh; it discretizes the geometry, acts as the support of the numerical method and needs to be tailored to the underlying non-linear equations governing the system physics. A poor mesh will fail to capture both geometrical and physical complexities and will thus yield inaccurate results [31].…”

“…A major task at hand, Uncertainty Quantification (UQ), consists in propagating this model parameter uncertainty throughout subsequent calculations to quantify the statistical behavior of the output QoI [47]. Sometimes, the parametric exploration of a given model -out of its reliability range -cause a deterioration of the prediction, leading to unpredictable outlying fluctuations that need robust handling [28]. In other situations, the model response to parametric change is less biased, but nevertheless remains strongly nonlinear and requires adaptation.…”

Polynomial Surrogates for Open-Channel Flows in Random Steady State

“…The least absolute shrinkage and selection operator (LASSO) algorithm [45] is in this case an attractive modification of the ordinary least-square formulation that constrains the sum of the absolute regression coefficients. Weighted versions exist that make the approximation even more robust [46]. Another related model selection algorithm, the least-angle regression (LAR), is also very efficient in our framework.…”