Abstract
In physics-based engineering modeling, two primary sources of model uncertainty that account for the differences between computer models and physical experiments are parameter uncertainty and model discrepancy. One of the main challenges in model updating results from the difficulty in distinguishing between the effects of calibration parameters versus model discrepancy. In this chapter, this identifiability problem is illustrated with several examples that explain the mechanisms behind it and that attempt to shed light on when a system may or may not be identifiable. For situations in which identifiability cannot be achieved using only a single response, an approach is developed to improve identifiability by using multiple responses that share a mutual dependence on the calibration parameters. Furthermore, prior to conducting physical experiments but after conducting computer simulations, in order to address the issue of how to select the most appropriate set of responses to measure experimentally to best enhance identifiability, a preposterior analysis approach is presented to predict the degree of identifiability that will result from using different sets of responses to measure experimentally. To handle the computational challenges of the preposterior analysis, we also present a surrogate preposterior analysis based on the Fisher information of the calibration parameters.
This is a preview of subscription content, log in via an institution.
Buying options
Tax calculation will be finalised at checkout
Purchases are for personal use only
Learn about institutional subscriptionsReferences
Kennedy, M.C., O’Hagan, A.: Bayesian calibration of computer models. J. R. Stat. Soc. Ser. B 63(3), 425–464 (2001)
Higdon, D., Kennedy, M.C., Cavendish, J., Cafeo, J., Ryne, R.: Combining field data and computer simulations for calibration and prediction. SIAM J. Sci. Comput. 26(2), 448–466 (2004)
Reese, C.S., Wilson, A.G., Hamada, M., Martz, H.F.: Integrated analysis of computer and physical experiments. Technometrics 46(2), 153–164 (2004)
Bayarri, M.J., Berger, J.O., Paulo, R., Sacks, J., Cafeo, J.A., Cavendish, J., Lin, C.H., Tu, J.: A framework for validation of computer models. Technometrics 49(2), 138–154 (2007)
Higdon, D., Gattiker, J., Williams, B., Rightley, M.: Computer model calibration using high-dimensional output. J. Am. Stat. Assoc. 103(482), 570–583 (2008)
Chen, W., Xiong, Y., Tsui, K.L., Wang, S.: A design-driven validation approach using bayesian prediction models. J. Mech. Des. 130(2), 021101 (2008)
Qian, P.Z.G., Wu, C.F.J.: Bayesian hierarchical modeling for integrating low-accuracy and high-accuracy experiments. Technometrics 50(2), 192–204 (2008)
Wang, S., Tsui, K.L., Chen, W.: Bayesian validation of computer models. Technometrics 51(4), 439–451 (2009)
Drignei, D.: A kriging approach to the analysis of climate model experiments. J. Agric. Biol. Environ. Stat. 14(1), 99–112 (2009)
Akkaram, S., Agarwal, H., Kale, A., Wang, L.: Meta modeling techniques and optimal design of experiments for transient inverse modeling applications. Paper presented at the ASME International Design Engineering Technical Conference, Montreal (2010)
Huan, X., Marzouk, Y.M.: Simulation-based optimal Bayesian experimental design for nonlinear systems. J. Comput. Phys. 232(1), 288–317 (2013)
Loeppky, J., Bingham, D., Welch, W.: Computer model calibration or tuning in practice. Technical Report, University of British Columbia, Vancouver, p. 20 (2006)
Han, G., Santner, T.J., Rawlinson, J.J.: Simultaneous determination of tuning and calibration parameters for computer experiments. Technometrics 51(4), 464–474 (2009)
Arendt, P.D., Apley, D.W., Chen, W.: Quantification of model uncertainty: Calibration, model discrepancy, and identifiability. J. Mech. Des. 134(10) (2012)
Arendt, P.D., Apley, D.W., Chen, W., Lamb, D., Gorsich, D.: Improving identifiability in model calibration using multiple responses. J. Mech. Des. 134(10) (2012)
Ranjan, P., Lu, W., Bingham, D., Reese, S., Williams, B., Chou, C., Doss, F., Grosskopf, M., Holloway, J.: Follow-up experimental designs for computer models and physical processes. J. Stat. Theory Pract. 5(1), 119–136 (2011)
Williams, B.J., Loeppky, J.L., Moore, L.M., Macklem, M.S.: Batch sequential design to achieve predictive maturity with calibrated computer models. Reliab. Eng. Syst. Saf. 96(9), 1208–1219 (2011)
Tuo, R., Wu, C.F.J., Vu, D.: Surrogate modeling of computer experiments with different mesh densities. Technometrics 56(3), 372–380 (2014)
Maheshwari, A.K., Pathak, K.K., Ramakrishnan, N., Narayan, S.P.: Modified Johnson-Cook material flow model for hot deformation processing. J. Mater. Sci. 45(4), 859–864 (2010)
Xiong, Y., Chen, W., Tsui, K.L., Apley, D.W.: A better understanding of model updating strategies in validating engineering models. Comput. Methods Appl. Mech. Eng. 198(15–16), 1327–1337 (2009)
Liu, F., Bayarri, M.J., Berger, J.O., Paulo, R., Sacks, J.: A Bayesian analysis of the thermal challenge problem. Comput. Methods Appl. Mech. Eng. 197(29–32), 2457–2466 (2008)
Arendt, P., Apley, D.W., Chen, W.: Updating predictive models: calibration, bias correction, and identifiability. Paper presented at the ASME 2010 International Design Engineering Technical Conferences, Montreal (2010)
Chakrabarty, J.: Theory of Plasticity, 3rd edn. Elsevier/Butterworth-Heinemann, Burlington (2006)
Liu, F., Bayarri, M.J., Berger, J.O.: Modularization in Bayesian analysis, with emphasis on analysis of computer models. Bayesian Anal. 4(1), 119–150 (2009)
Joseph, V., Melkote, S.: Statistical adjustments to engineering models. J. Qual. Technol. 41(4), 362–375 (2009)
Rasmussen, C.E., Williams, C.K.I.: Gaussian Processes for Machine Learning. Adaptive Computation and Machine Learning. MIT Press, Cambridge (2006)
Qian, P.Z.G., Wu, H.Q., Wu, C.F.J.: Gaussian process models for computer experiments with qualitative and quantitative factors. Technometrics 50(3), 383–396 (2008)
McMillan, N.J., Sacks, J., Welch, W.J., Gao, F.: Analysis of protein activity data by gaussian stochastic process models. J. Biopharm. Stat. 9(1), 145–160 (1999)
Cressie, N.: Statistics for Spatial Data. Wiley Series in Probability and Statistics. Wiley, New York (1993)
Ver Hoef, J., Cressie, N.: Multivariable spatial prediction. Math. Geol. 25(2), 219–240 (1993)
Conti, S., Gosling, J.P., Oakley, J.E., O’Hagan, A.: Gaussian process emulation of dynamic computer codes. Biometrika 96(3), 663–676 (2009)
Conti, S., O’Hagan, A.: Bayesian emulation of complex multi-output and dynamic computer models. J. Stat. Plan. Inference 140(3), 640–651 (2010)
Williams, B., Higdon, D., Gattiker, J., Moore, L.M., McKay, M.D., Keller-McNulty, S.: Combining experimental data and computer simulations, with an application to flyer plate experiments. Bayesian Analysis 1(4), 765–792 (2006)
Bayarri, M.J., Berger, J.O., Cafeo, J., Garcia-Donato, G., Liu, F., Palomo, J., Parthasarathy, R.J., Paulo, R., Sacks, J., Walsh, D.: Computer model validation with functional output. Ann. Stat. 35(5), 1874–1906 (2007)
McFarland, J., Mahadevan, S., Romero, V., Swiler, L.: Calibration and uncertainty analysis for computer simulations with multivariate output. AIAA J. 46(5), 1253–1265 (2008)
Drignei, D.: A kriging approach to the analysis of climate model experiments. J. Agric. Biol. Environ. Stat. 14(1), 99–114 (2009)
Kennedy, M.C., Anderson, C.W., Conti, S., O’Hagan, A.: Case studies in gaussian process modelling of computer codes. Reliab. Eng. Syst. Saf. 91(10–11), 1301–1309 (2006)
Sacks, J., Welch, W.J., Mitchell, T.J., Wynn, H.P.: Design and analysis of computer experiments. Stat. Sci. 4(4), 409–423 (1989)
Rasmussen, C.E.: Evaluation of Gaussian Processes and Other Methods for Non-linear Regression. University of Toronto (1996)
Lancaster, T.: An Introduction to Modern Bayesian Econometrics. Blackwell, Malden (2004)
Arendt, P.D., Apley, D.W., Chen, W.: A preposterior analysis to predict identifiability in experimental calibration of computer models. IIE Trans. 48(1), 75–88 (2016)
Jiang, Z., Chen, W., Apley, D.W.: Preposterior analysis to select experimental responses for improving identifiability in model uncertainty quantification. Paper presented at the ASME 2013 International Design Engineering Technical Conferences & Computers and Information in Engineering Conference, Portland (2013)
Jiang, Z., Apley, D.W., Chen, W.: Surrogate preposterior analyses for predicting and enhancing identifiability in model calibration. Int. J. Uncertain. Quantif. 5(4), 341–359 (2015)
Berger, J.O.: Statistical Decision Theory and Bayesian Analysis. Springer Series in Statistics. Springer, New York (1985)
Carlin, B.P., Louis, T.A.: Empirical bayes: Past, present and future. J. Am. Stat. Assoc. 95(452), 1286–1289 (2000)
Wu, C., Hamada, M.: Experiments: Planning, Analysis, and Optimization. Wiley, New York (2009)
Montgomery, D.C.: Design and Analysis of Experiments, 7th edn. Wiley, Hoboken (2008)
Abramowitz, M., Stegun, I.A.: Handbook of Mathematical Functions. Dover, New York (1972)
Beyer, W.H.: CRC Standard Mathematical Tables, 28 edn. CRC, Boca Raton (1987)
Robert, C., Casella, G.: Monte Carlo Statistical Methods. Springer, New York (2004)
Smith, A., Gelfand, A.: Bayesian statistics without tears: a sampling-resampling perspective. Am. Stat. 46(2), 84–88 (1992)
Johnson, R., Wichern, D.: Applied Multivariate Statistical Analysis, 6th edn. Prentice Hall, Upper Saddle River (2007)
Kennedy, M.C., O’Hagan, A.: Supplementary Details on Bayesian Calibration of Computer Models, pp. 1–13. University of Sheffield, Sheffield (2000)
Billingsley, P.: Probability and Measure, Anniversary Edition. John Wiley & Sons, Inc., Hoboken (2011)
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Appendices
Appendix A: Estimates of the Hyperparameters for the Computer Model MRGP
To obtain the MLEs of the hyperparameters for the computer model MRGP model, the multivariate normal likelihood function is first constructed as:
where vec(⋅ ) is the vectorization of the matrix (stacking of the columns), ⊗ denotes the Kronecker product, R m is a N m × N m correlation matrix whose ith-row, jth-column entry is \(R^{m}((\mathbf{x}_{i}^{m},\boldsymbol{\uptheta }_{i}^{m}),(\mathbf{x}_{j}^{m},\boldsymbol{\uptheta }_{j}^{m}))\), and \(\mathbf{H}^{m} = [\mathbf{h}^{m}(\mathbf{x}_{1}^{m},\boldsymbol{\uptheta }_{1}^{m})^{T},\,\,\ldots,\mathbf{h}^{m}(\mathbf{x}_{N_{m}}^{m},\boldsymbol{\uptheta }_{N_{m}}^{m})^{T}]^{T}\). Taking the log of Eq. (4.18) yields:
The MLE of B m is found by setting the derivative of Eq. (4.19) with respect to B m equal to zero, which gives:
The MLE of \(\boldsymbol{\Sigma }^{m}\) is found using result 4.10 of Ref. [52], which yields:
Finally, the MLE of \(\boldsymbol{\upomega }^{m}\), denoted by \(\hat{\boldsymbol{\upomega }}^{m}\), is found by numerically maximizing Eq. (4.19) after plugging in the MLEs of B m and \(\boldsymbol{\Sigma }^{m}\).
Appendix B: Posterior Distributions of the Computer Responses
After observing Y m, the posterior of the computer response \(y_{i}^{m}(\mathbf{x},\boldsymbol{\uptheta })\) given Y m (and given \(\boldsymbol{\upomega }^{m}\) and \(\boldsymbol{\Sigma }^{m}\) and assuming a non-informative prior for B m) is Gaussian with mean and covariance:
where \(\mathbf{r}^{m}(\mathbf{x},\boldsymbol{\uptheta })\) is a N m × 1 vector whose ith element is \(R^{m}((\mathbf{x}_{i}^{m},\boldsymbol{\uptheta }_{i}^{m}),(\mathbf{x},\boldsymbol{\uptheta }))\). Using an empirical Bayes approach, the MLEs of the hyperparameters from Appendix A are plugged into Eqs. (4.22) and (4.23) to calculate the posterior distribution of the computer responses. Notice that Eqs. (4.22) and (4.23) are analogous to the single-response GP model results.
Appendix C: Estimates of the Hyperparameters for the Discrepancy Functions MRGP
To estimate the hyperparameters \(\boldsymbol{\upphi }^{\delta } =\{ \mathbf{B}^{\delta }\), \(\boldsymbol{\Sigma }^{\updelta }\), \(\boldsymbol{\upomega }^{\delta }\), \(\boldsymbol{\uplambda }\}\) of the MRGP model representing the discrepancy functions, the procedure outlined by Kennedy and O’Hagan [1] is used and modified to handle multiple responses. This procedure begins by obtaining a posterior of the experimental responses given the simulation data and the hyperparameters from Module 1, which has prior mean and covariance:
where \(\hat{\boldsymbol{\upphi }}^{m}\) are the MLEs of the hyperparameters for the computer model MRGP model. Since Eqs. (4.24) and (4.25) depend on the unknown true value of \(\boldsymbol{\uptheta }^{{\ast}}\), these two equations are integrated with respect to the prior distribution of \(\boldsymbol{\uptheta }(p(\boldsymbol{\uptheta }))\) via:
Kennedy and O’Hagan [53] provide closed form solutions for Eq. (4.26) under the conditions of Gaussian correlation functions, constant regression functions, and normal prior distributions for \(\boldsymbol{\uptheta }\) (for details, refer to Section 3 of [53] and Section 4.5 of [1]). In this chapter, similar closed form solutions are used except that a uniform prior distribution is assumed for \(\boldsymbol{\uptheta }\).
After observing the experimental data, Y e, one can construct a multivariate normal likelihood function with mean and variance from Eq. (4.26). The MLEs of \(\boldsymbol{\upphi }^{\delta }\) maximize this likelihood function. The MLE of B δ can found by setting the analytical derivative of this likelihood function with respect to B δ equal to zero (see Section 2 of Ref. [53]). However, there are no analytical derivatives with respect to the hyperparameters \(\boldsymbol{\Sigma }^{\delta }\), \(\boldsymbol{\upomega }^{\delta }\), and \(\boldsymbol{\lambda }\). Therefore, numerical optimization techniques are needed to find these MLEs.
Appendix D: Posterior Distribution of the Calibration Parameters
The posterior for the calibration parameters in Eq. (4.12) involves the likelihood function \(p(\mathbf{d}\vert \boldsymbol{\uptheta },\hat{\boldsymbol{\upphi }})\) and the marginal posterior distribution for the data \(p(\mathbf{d}\vert \hat{\boldsymbol{\upphi }})\). The likelihood function is multivariate normal with mean vector and covariance matrix defined as:
where \(\hat{\mathbf{B}}= \left (\mathbf{H}(\boldsymbol{\uptheta })^{T}\mathbf{V}(\boldsymbol{\uptheta })^{-1}\mathbf{H}(\boldsymbol{\uptheta })\right )^{-1}\mathbf{H}(\boldsymbol{\uptheta })^{T}\mathbf{V}(\boldsymbol{\uptheta })^{-1}\mathbf{d}\), which is calculated based on the entire data set (instead of using the estimates from Modules 1 and 2 for B m and B δ) as detailed in Section 4 of [53]. \(\mathbf{H}^{m}(\mathbf{X}^{e},\boldsymbol{\uptheta }) = [\mathbf{h}^{m}(\mathbf{x}_{1}^{e},\boldsymbol{\uptheta })^{T},\,\,\ldots,\mathbf{h}^{m}(\mathbf{x}_{N_{e}}^{m},\boldsymbol{\uptheta })^{T}]_{}^{T}\) and \(\mathbf{H}^{\delta } = [\mathbf{h}^{\delta }(\mathbf{x}_{1}^{e})^{T},\,\,\ldots,\mathbf{h}^{\delta }(\mathbf{x}_{N_{e}}^{\delta })^{T}]_{}^{T}\) denote the specified regression functions for the computer model and the discrepancy functions at the input settings X e. C m denotes the N m × N e matrix with ith-row, jth-column entries \(R^{m}((\mathbf{x}_{i}^{m},\boldsymbol{\uptheta }_{i}^{m}),(\mathbf{x}_{j}^{e},\boldsymbol{\uptheta }))\). \(\mathbf{R}^{m}(\mathbf{X}^{e},\boldsymbol{\uptheta })\) denotes the N e × N e matrix with ith-row, jth-column entries \(R^{m}((\mathbf{x}_{i}^{e},\boldsymbol{\uptheta }),(\mathbf{x}_{j}^{e},\boldsymbol{\uptheta }))\). R δ denotes the N e × N e matrix with ith-row, jth-column entries R δ(x e i , x e j ). Finally, I q and \(\mathbf{I}_{N_{e}}\) denote the q × q and N e × N e identity matrices.
The marginal posterior distribution for the data \(p(\mathbf{d}\vert \hat{\boldsymbol{\upphi }})\) is:
which can be calculated using any numerical integration technique. In this chapter, Legendre-Gauss quadrature is used for the low-dimensional examples. Alternatively, Markov chain Monte Carlo (MCMC) could be used to sample complex posterior distributions such as those in Eq. (4.12).
Appendix E: Posterior Distribution of the Experimental Responses
Since a MRGP model represents the experimental responses, the conditional (given \(\boldsymbol{\uptheta }\)) posterior distribution at any point x is Gaussian with mean and covariance defined as (assuming a non-informative prior on B m and B δ and using the empirical Bayes approach that treats \(\boldsymbol{\upphi }=\hat{\boldsymbol{\upphi }}\) as fixed):
where:
\(\mathbf{R}^{m}((\mathbf{X}^{m},\boldsymbol{\Theta }^{m}),(\mathbf{x},\boldsymbol{\uptheta }))\) is a N m ×1 vector whose ith entry is \(R^{m}((\mathbf{x}_{i}^{m},\boldsymbol{\uptheta }_{i}^{m}),(\mathbf{x},\boldsymbol{\uptheta }))\), \(\mathbf{R}^{m}((\mathbf{X}^{e},\boldsymbol{\uptheta }),(\mathbf{x},\boldsymbol{\uptheta }))\) is a N e ×1 vector whose ith entry is \(R^{m}((\mathbf{x}_{i}^{e},\boldsymbol{\uptheta }),(\mathbf{x},\boldsymbol{\uptheta }))\), and R δ(X e, x) is a N e × 1 vector whose ith entry is R δ(x e i , x).
To calculate the unconditional posterior distributions (marginalized with respect to \(\boldsymbol{\uptheta }\)) of the experimental responses, the conditional posterior distributions are marginalized with respect to the posterior distribution of the calibration parameters from Module 3. The mean and covariance of the unconditional posterior distributions can be written as:
where the outer expectation and covariance are with respect to the posterior distribution of the calibration parameters. Equations (4.34) and (4.35) are derived using the law of total expectation and the law of total covariance [54]. Due to the complexity of the posterior distribution of the calibration parameters, the marginalization requires numerical integration methods. For the examples in this chapter, Legendre-Gauss quadrature is used.
Rights and permissions
Copyright information
© 2017 Springer International Publishing Switzerland
About this entry
Cite this entry
Jiang, Z., Arendt, P.D., Apley, D.W., Chen, W. (2017). Multi-response Approach to Improving Identifiability in Model Calibration. In: Ghanem, R., Higdon, D., Owhadi, H. (eds) Handbook of Uncertainty Quantification. Springer, Cham. https://doi.org/10.1007/978-3-319-12385-1_65
Download citation
DOI: https://doi.org/10.1007/978-3-319-12385-1_65
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-12384-4
Online ISBN: 978-3-319-12385-1
eBook Packages: Mathematics and StatisticsReference Module Computer Science and Engineering