Bayesian probabilistic propagation of imprecise probabilities with large epistemic uncertainty

Pengfei Wei, Fuchao Liu, Marcos Valdebenito, Michael Beer

Research output: Contribution to journalArticlepeer-review

12 Scopus citations

Abstract

Efficient propagation of imprecise probability models is one of the most important, yet challenging tasks, for uncertainty quantification in many areas and engineering practices, especially when the involved epistemic uncertainty is substantial due to the extreme lack of information. In this work, a new methodology framework, named as “Non-intrusive Imprecise Probabilistic Integration (NIPI)”, is developed for achieving the above target, and specifically, the distributional probability-box model and the estimation of the corresponding probabilistic moments of model responses are of concern. The NIPI owns two attractive characters. First, the spatial correlation information in both aleatory and epistemic uncertainty spaces, revealed by the Gaussian Process Regression (GPR) model, is fully integrated for deriving NIPI estimations of high accuracy by using Bayesian inference. Second, the numerical errors are regarded as a kind of epistemic uncertainty, by analytically propagating them, the posterior variances are derived for indicating the errors of the NIPI estimations. Further, an adaptive experiment design strategy is developed to accelerate the convergence of NIPI by making full use of the information of “contribution to posterior variance” revealed by the GPR model. The performance of the proposed methods is demonstrated by numerical and engineering examples.

Original languageEnglish
Article number107219
JournalMechanical Systems and Signal Processing
Volume149
DOIs
StatePublished - 15 Feb 2021

Keywords

  • Active learning
  • Bayesian inference
  • Epistemic uncertainty
  • Gaussian process regression
  • Imprecise probabilities
  • Probabilistic integration
  • Uncertainty quantification

Fingerprint

Dive into the research topics of 'Bayesian probabilistic propagation of imprecise probabilities with large epistemic uncertainty'. Together they form a unique fingerprint.

Cite this