Data-driven and active learning of variance-based sensitivity indices with Bayesian probabilistic integration

https://doi.org/10.1016/j.ymssp.2021.108106Get rights and content

Highlights

  • Estimation of variance-based sensitivity indices is treated by Bayesian inference.

  • Posterior means are analytically derived as mean estimates for sensitivity indices.

  • Posterior variances are analytically derived for quantifying the numerical errors.

  • A data-driven learning method is devised for realizing the inference from data.

  • An active learning method is developed for further reducing the computational cost.

Abstract

Variance-based sensitivity indices play an important role in scientific computation and data mining, thus the significance of developing numerical methods for efficient and reliable estimation of these sensitivity indices based on (expensive) computer simulators and/or data cannot be emphasized too much. In this article, the estimation of these sensitivity indices is treated as a statistical inference problem. Two principle lemmas are first proposed as rules of thumb for making the inference. After that, the posterior features for all the (partial) variance terms involved in the main and total effect indices are analytically derived (not in closed form) based on Bayesian Probabilistic Integration (BPI). This forms a data-driven method for estimating the sensitivity indices as well as the involved discretization errors. Further, to improve the efficiency of the developed method for expensive simulators, an acquisition function, named Posterior Variance Contribution (PVC), is utilized for realizing optimal designs of experiments, based on which an adaptive BPI method is established. The application of this framework is illustrated for the calculation of the main and total effect indices, but the proposed two principle lemmas also apply to the calculation of interaction effect indices. The performance of the development is demonstrated by an illustrative numerical example and three engineering benchmarks with finite element models.

Introduction

Nowadays, owing to the rapid development of computation power, scientific computation based on computer simulators (e.g., finite element models) has been widely utilized in both academic research and engineering practice for predicting the behavior of complex systems or structures and aiding the design of new products. However, due to the uncertainties of various sources, the researchers and practitioners have found it difficult to achieve accurate and robust predictions with the deterministic simulators, and performing uncertainty quantification to properly incorporate those uncertainties in the model predictions has been a common trend in scientific computing [1], [2], and especially, in structural dynamics. As an important sub-task of uncertainty quantification, Sensitivity Analysis (SA) plays an important role in model developments and refinement as it informs the main sources of model prediction uncertainties [3], [4], [5]. This information is extremely useful for directing the future data collection (with the target of effectively reducing the model prediction uncertainty), and for specifying the subset of influential model parameters to be calibrated in finite element (FE) model updating [6].

Specifically, SA aims at attributing the uncertainty present in the model output to the input variables, and in this way to measure the contribution of each input variable to the uncertainty of model outputs [7]. Three groups of SA methods have been developed, i.e., local SA, regional SA, and Global SA (GSA), one can refer to Refs. [4], [5] for comprehensive reviews and comparisons of these methods. The local method measures the sensitivity of each input variable using the local partial derivatives, and it is widely used in the area of structural reliability for measuring the effects of the distribution parameters of input variables on the failure probability [8], [9]. The regional SA aims at quantifying the effects/contributions of the subregions of the distribution support of each input variables to the uncertainty of model outputs, and it can be especially useful for reduction of epistemic uncertainty [10]. The GSA indices are usually defined as the expected change of the statistical features (e.g., variance and density function) of model response when the input variables are fixed over their full supports, thus summarize the overall contribution of the uncertainty present in the input variables to those of the model outputs.

Among the above three groups of methods, the GSA has received the greatest attention during the past few decades, and a plenty of GSA techniques/indices have been developed for different purposes. The screening methods have been developed for screening the non-influential variables in moderate to high dimensional problems [11], [12]. The variance-based sensitivity indices [13], [14], [15], rooted in the Random sampling-high dimensional model representation (RS-HDMR) [16], aim at measuring the relative importance of the input variables by attributing the model response variance to each input variable and their interactions. Considering the setting of uncertainty reduction, a modified versions of the variance-based sensitivity indices, called W-indices, has also been developed for quantifying the effects of reducing the input uncertainty on that of model output [17]. Given that the variance is not sufficient for characterizing the uncertainty, the moment-independent sensitivity indices have also been devised for investigating the effect of each input variable on the full probability distribution of the model response [18], [19], [20]. The derivative-based sensitivity indices have also been established to realize variable screening with lower computational cost than the variance-based ones [21], [22], [23]. The global reliability sensitivity indices have been developed in the area of structural reliability, based on the variance-based indices, for measuring the contribution of input variables to the failure probability of structures [24], [25], [26]. Despite the extensive GSA indices that have been developed, the variance-based ones continue to receive the greatest concerns of both researchers and practitioners owing to the elegant mathematical interpretations for both independent and dependent variables, as well as their ability of capturing different types of effects [7], [27], [28]. Developing efficient and robust algorithms for estimating variance-based indices is then one of the most relevant challenges for performing the GSA analysis.

The past few decades have witnessed a rapid development of numerical algorithms for variance-based sensitivity indices, and one can refer to Ref. [29] for a comprehensive review on these related developments. Generally, these methods can be divided into three classes, i.e., Fourier amplitude sensitivity test (FAST), (quasi-) Monte Carlo simulation (MCS), and surrogate models. The FAST method, developed in the area of computational chemistry [30], estimates the partial variance terms involved in the variance-based sensitivity indices based on periodic sampling and Fourier transformation, and it has been widely studied and substantially improved since its development (see e.g., Refs. [31], [32], [33], [34]). The MCS method involves first formulating the partial variance terms with multi-dimensional integrals, and then utilizing MCS, driven by simple random sampling or Latin Hypercube Sampling (LHS) [35] or Sobol’s low-discrepancy sequence [36], to estimate these integrals. Following this scheme, a multitude of MCS estimators have been developed (see e.g., Refs. [37], [38], [39], [40]). The surrogate models, such as state dependent regression [41], polynomial chaos expansion [42], support vector regression [43] and Kriging, also called Gaussian Process Regression (GPR) [44], [45], [46], [47], have also been investigated for estimating the sensitivity indices. In terms of reliability of estimation, MCS is the most competitive scheme as confidence intervals can be computed for the sensitivity indices from the MCS estimators, but it also suffers from the large number of required simulator calls, which make it not applicable to computationally expensive simulators.

In recent years, Bayesian numerical analysis [48] with its different variants, such as Bayesian probabilistic optimization [49], Bayesian Probabilistic Integration (BPI) [50], [51], and Bayesian probabilistic Partial Differential Equation (PDE) solution [52], has emerged as a cutting-edge method in scientific computation. The aim of this work is therefore to extend the BPI methods for inferring the variance-based sensitivity indices from data and computer simulators. This topic has also been investigated in Ref.[46] in a full Bayesian scheme and in Ref.[53] with the so-called Bayesian MCS scheme, but in both papers, only the posterior mean and the main effect indices are investigated. In this work, both the posterior means and posterior variances will be first investigated for both the main and total variance-based indices based on BPI, following which, a data-driven BPI approach and an adaptive BPI approach will be developed for efficiently estimating the sensitivity indices. To achieve this goal, two principle lemmas are first developed for realizing the Bayesian inference, and then the posterior means and variances are both analytically derived for the sensitivity indices, where the posterior variances summarize the discretization errors for estimating these sensitivity indices. These analytical results form the basis of the data-driven BPI approach, with which the posterior features of the sensitivity indices can be inferred from any supervised learning data. To further improve the efficiency of the algorithm for computationally expensive simulators, an adaptive experiment design strategy is ultimately introduced. The effectiveness of the proposed methods are demonstrated by numerical examples, and their applicability to real-world engineering problems as well as their engineering significance are illustrated by three engineering benchmarks with FE simulators.

The rest of this paper is organized as follows. Section 2 briefly reviews the variance-based sensitivity indices and the BPI approach, followed by the core developments in Section 3, which includes the Bayesian inference of the sensitivity indices and the data-driven BPI. The adaptive BPI approach is then developed in Section 4, followed by the numerical and engineering test examples in Section 5. Section 6 closes the paper with conclusions.

Section snippets

Brief review of related topics

Before the introduction of the main developments, it is helpful to briefly review two important topics to be studied/utilized in this article, i.e., the variance-based sensitivity indices and the BPI. The expectation and variance operators utilized in this paper are declared in Table 1 for avoiding confusion.

Bayesian inference of sensitivity indices

In the previous section, the details of the BPI approach for estimating M0 have been reviewed, and it is concluded that, given the GPR representation of the model function Mx, the posterior distribution of M0 is also Gaussian. Indeed, the induced probabilistic models for any orders of HDMR components (e.g., Mixi and Mijxij) are Gaussian as well [46], [58]. This provides a basis to infer the posterior features of the first-order partial variances Vi, the total partial variance VTi, and the total

Adaptive experiment design

Until now, we have generated the analytical expressions of the posterior means and posterior variances for all the (partial) variance terms following the Bayesian inference scheme based on the training data D. Based on these results, a data-driven method is established for estimating the variance-based sensitivity indices. However, in real-world applications, the sensitivity analysis may also be implemented for computer simulators such as finite element models, which makes it possible to design

An illustrative example

Considering a two-dimensional model with g-function formulated as:gx1,x2=i=14ciexp-αi1x1-βi12-αi2x2-βi22where α=23143241,β=-0.50.5-0.50.5-0.5-0.50.50.5,c=1-1.5-1.52,x1 and x2 are independent standard normal random variables. This is a highly nonlinear model with large interaction effects, and the variance-based sensitivity indices can be analytically derived to provide comparison.

For implementing the adaptive BPI, the stopping criteria is set to be σM0/μM00.2, and the algorithm stops only

Conclusions and discussions

The estimation of the variance-based sensitivity indices is regarded as an statistical inference problem in this work, and based on a set of supervised training data, the posterior features (including means and variances) for all the (partial) variance terms involved in the sensitivity indices are analytically derived following two newly developed first principles and the rationale of BPI. Although the posterior distributions of these (partial) variance terms are no longer Gaussian, these

CRediT authorship contribution statement

Jingwen Song: Methodology, Software, Validation, Visualization, Writing - original draft. Pengfei Wei: Conceptualization, Methodology, Investigation, Writing - review & editing, Funding acquisition. Marcos A. Valdebenito: Validation, Resources, Writing - review & editing. Matthias Faes: Validation, Resources, Writing - review & editing. Michael Beer: Supervision, Project administration.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

This work is supported by the National Natural Science Foundation of China under Grant No. 51905430, the Sino-German Mobility Program under Grant No. M-0175, the ANID (Agency for Research and Development, Chile) under its program FONDECYT, Grant No. 1180271, and the Research Foundation Flanders (FWO) under Grant No. 12P3519N. The first author is supported by the program of China Scholarships Council (CSC). The second to forth authors are all supported by the Alexander von Humboldt Foundation of

References (66)

  • E. Plischke et al.

    Global sensitivity measures from given data

    European Journal of Operational Research

    (2013)
  • I. Sobol et al.

    Derivative based global sensitivity measures

    Procedia-Social and Behavioral Sciences

    (2010)
  • M. Lamboni et al.

    Derivative-based global sensitivity measures: general links with sobol’ indices and numerical tests

    Mathematics and Computers in Simulation

    (2013)
  • P. Wei et al.

    Efficient sampling methods for global reliability sensitivity analysis

    Computer Physics Communications

    (2012)
  • L. Li et al.

    Global reliability sensitivity estimation based on failure samples

    Structural Safety

    (2019)
  • Z. Kala

    Global sensitivity analysis of reliability of structural bridge system

    Engineering Structures

    (2019)
  • W. Hao et al.

    A new interpretation and validation of variance based importance measures for models with correlated inputs

    Computer Physics Communications

    (2013)
  • T.A. Mara et al.

    Variance-based sensitivity indices for models with dependent inputs

    Reliability Engineering & System Safety

    (2012)
  • A. Saltelli et al.

    An alternative way to compute fourier amplitude sensitivity test (FAST)

    Computational Statistics & Data Analysis

    (1998)
  • C. Xu et al.

    Understanding and comparisons of different sampling approaches for the fourier amplitudes sensitivity test (FAST)

    Computational Statistics & Data Analysis

    (2011)
  • S. Tarantola et al.

    Random balance designs for the estimation of first order global sensitivity indices

    Reliability Engineering & System Safety

    (2006)
  • T.A. Mara

    Extension of the RBD-FAST method to the computation of global sensitivity indices

    Reliability Engineering & System Safety

    (2009)
  • J.C. Helton et al.

    A comparison of uncertainty and sensitivity analysis results obtained with random and Latin hypercube sampling

    Reliability Engineering & System Safety

    (2005)
  • I.M. Sobol

    Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates

    Mathematics and Computers in Simulation

    (2001)
  • A. Saltelli

    Making best use of model evaluations to compute sensitivity indices

    Computer Physics Communications

    (2002)
  • A. Saltelli et al.

    Variance based sensitivity analysis of model output. design and estimator for the total sensitivity index

    Computer Physics Communications

    (2010)
  • M. Ratto et al.

    State dependent parameter metamodelling and sensitivity analysis

    Computer Physics Communications

    (2007)
  • B. Sudret

    Global sensitivity analysis using polynomial chaos expansions

    Reliability Engineering & System Safety

    (2008)
  • M. Steiner et al.

    An adaptive sampling method for global sensitivity analysis based on least-squares support vector regression

    Reliability Engineering & System Safety

    (2019)
  • A. Marrel et al.

    Calculations of Sobol’ indices for the Gaussian process metamodel

    Reliability Engineering & System Safety

    (2009)
  • P. Wei et al.

    Adaptive experiment design for probabilistic integration

    Computer Methods in Applied Mechanics and Engineering

    (2020)
  • Y. Zhou et al.

    A Bayesian Monte Carlo-based method for efficient computation of global sensitivity indices

    Mechanical Systems and Signal Processing

    (2019)
  • R. Lebrun et al.

    Do Rosenblatt and Nataf isoprobabilistic transformations really differ?

    Probabilistic Engineering Mechanics

    (2009)
  • Cited by (16)

    View all citing articles on Scopus
    View full text