Numerical moment matching stabilized by a genetic algorithm for engineering data squashing and fast uncertainty quantification. (15th July 2018)

• Record Type: Journal Article
• Authors: Cho, In Ho; Song, Ikkyun; Teng, Ya Lu
• Is Part Of: Computers & structures. Volume 204(2018)
• Page Start: 31
• Page End: 47
• Publication Date: 2018-07-15
• Subjects: Structural engineering -- Data processing -- Periodicals; Electronic data processing -- Structures, Theory of -- Periodicals; 624.171
• DOI: 10.1016/j.compstruc.2018.04.002
• Languages: English
• ISSNs: 0045-7949

Abstract

Highlights: Generalization of the genetic algorithm tailored for numerical moment matching technique. Stabilization of the numerical moment matching for complex engineering data sets. Data squashing of general engineering data. Efficient uncertainty estimation in conjunction with a high-precision computer simulation tool. General tool for data-driven engineering research. Abstract: Numerical moment matching (NMM) technique is a point estimation method that holds significant applicability in the era of large-scale data. One can use NMM to create an extremely small set of representative samples of an engineering population or to facilitate a fast and robust uncertainty quantification of a complex structure. However, the previous NMM method based on the multivariate Newton-Raphson (mNR) scheme often suffers from severe numerical divergence and initial-value dependency. This study overcomes the aforementioned limitations by stabilizing NMM with a genetic algorithm (GA), giving rise to a highly stable and fast NMM (denoted as GA-NMM). Inheriting NMM's strengths, GA-NMM exhibits no restriction to irregular distributions, large sizes, or many variables of engineering data. This paper elaborates the formulations of GA-NMM along with a practical recommendation for setting optimal parameters. Validations encompass theoretical and practical cases. Simulations with the default setting of GA-NMM demonstrate successful performances in data-squashing of an engineering population and a fast, robust uncertainty quantification of a complex structure. All the developed programs are made publicly available for promoting data-driven research paradigm in broader engineering domains.