Abonnement à la biblothèque: Guest
Portail numérique Bibliothèque numérique eBooks Revues Références et comptes rendus Collections
International Journal for Uncertainty Quantification
Facteur d'impact: 4.911 Facteur d'impact sur 5 ans: 3.179 SJR: 1.008 SNIP: 0.983 CiteScore™: 5.2

ISSN Imprimer: 2152-5080
ISSN En ligne: 2152-5099

Ouvrir l'accès

International Journal for Uncertainty Quantification

DOI: 10.1615/Int.J.UncertaintyQuantification.2016018697
pages 1-21

VPS: VORONOI PIECEWISE SURROGATE MODELS FOR HIGH-DIMENSIONAL DATA FITTING

Ahmad Rushdi
Sandia National Laboratories
Laura P. Swiler
Optimization and Uncertainty Quantification Department, Center for Computing Research, Sandia National Laboratories, Albuquerque, NM, 87123
Eric T. Phipps
Center for Computing Research, Sandia National Laboratories, Albuquerque, New Mexico 87185, USA
Marta D'Elia
Center for Computing Research, Sandia National Laboratories, Albuquerque, New Mexico 87185, USA
Mohamed S. Ebeida
Center for Computing Research, Sandia National Laboratories, Albuquerque, New Mexico 87185, USA

RÉSUMÉ

Surrogate models (metamodels) are indispensable for numerical simulations over high-dimensional spaces. They typically use well-selected samples of the expensive code runs to produce a cheap-to-evaluate model. We introduce a new method to construct credible global surrogates with local accuracy without dictating where to sample: Voronoi piecewise surrogate (VPS) models. The key component in our method is to implicitly decompose the parameter space into cells using the Voronoi tessellation around the sample points as seeds, via an approximate dual Delaunay graph. While explicit domain decompositions have storage and processing requirements that exponentially grow with dimension, VPS construction counts on the implicitness of Voronoi cells and the one-to-one mapping between seeds and cells, regardless of dimension, to avoid this curse of dimensionality. Each implicit cell can then use information provided by its neighbors to build its own local piece of the global surrogate. The piecewise locality breaks down the high-order approximation problem into a set of low-order problems, with better immunity against numerical oscillations. Domain points can be assigned to cells using a simple nearest seed search. Furthermore, a VPS model is naturally updated with the addition of new samples, can handle smooth and discontinuous functions, and can adopt a parallel implementation. We demonstrate the application of VPS models to numerical integration and probability of failure estimation problems.


Articles with similar content:

YIELD OPTIMIZATION BASED ON ADAPTIVE NEWTON-MONTE CARLO AND POLYNOMIAL SURROGATES
International Journal for Uncertainty Quantification, Vol.10, 2020, issue 4
Niklas Georg, Mona Fuhrländer, Sebastian Schöps, Ulrich Römer
STATISTICAL EXTRACTION OF PROCESS ZONES AND REPRESENTATIVE SUBSPACES IN FRACTURE OF RANDOM COMPOSITES
International Journal for Multiscale Computational Engineering, Vol.11, 2013, issue 3
S. P. A. Bordas, Timon Rabczuk, K. M. Schmidt, Pierre Kerfriden
ASYMPTOTICALLY INDEPENDENT MARKOV SAMPLING: A NEW MARKOV CHAIN MONTE CARLO SCHEME FOR BAYESIAN INFERENCE
International Journal for Uncertainty Quantification, Vol.3, 2013, issue 5
James L. Beck, Konstantin M. Zuev
BAYESIAN INFERENCE OF STOCHASTIC REACTION NETWORKS USING MULTIFIDELITY SEQUENTIAL TEMPERED MARKOV CHAIN MONTE CARLO
International Journal for Uncertainty Quantification, Vol.10, 2020, issue 6
Huy D. Vo, Thomas A. Catanach, Brian Munsky
A STOCHASTIC COLLOCATION METHOD FOR VOLTERRA INTEGRO-DIFFERENTIAL EQUATIONS WITH WEAKLY SINGULAR KERNELS AND RANDOM INPUTS
International Journal for Uncertainty Quantification, Vol.10, 2020, issue 2
Lijun Yi, Ling Guo