Adaptive method for indirect identification of the ...
Transcript of Adaptive method for indirect identification of the ...
HAL Id: hal-02373628https://hal-upec-upem.archives-ouvertes.fr/hal-02373628
Submitted on 21 Nov 2019
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Adaptive method for indirect identification of thestatistical properties of random fields in a Bayesian
frameworkGuillaume Perrin, Christian Soize
To cite this version:Guillaume Perrin, Christian Soize. Adaptive method for indirect identification of the statistical prop-erties of random fields in a Bayesian framework. Computational Statistics, Springer Verlag, 2020, 35,pp.111-133. 10.1007/s00180-019-00936-5. hal-02373628
Noname manus ript No.
(will be inserted by the editor)
Adaptive method for indire t identi ation of the
statisti al properties of random elds in a Bayesian
framework
Guillaume Perrin · Christian Soize
Re eived: date / A epted: date
Abstra t This work onsiders the hallenging problem of identifying the sta-
tisti al properties of random elds from indire t observations. To this end,
a Bayesian approa h is introdu ed, whose key step is the nonparametri ap-
proximation of the likelihood fun tion from limited information. When the
likelihood fun tion is based on the evaluation of an expensive omputer ode,
this work also proposes a method to sele t iteratively new design points to re-
du e the un ertainties on the results that are due to the approximation of the
likelihood. Two appli ations are nally presented to illustrate the e ien y of
the proposed pro edure: a rst one based on analyti data, and a se ond one
dealing with the identi ation of the random elasti ity eld of an heteroge-
neous mi rostru ture.
Keywords Bayesian framework · un ertainty quanti ation · statisti alinferen e · sto hasti pro ess · kernel density estimation
1 Introdu tion
Random eld analysis has be ome a major tool in many s ienti elds, su h
as un ertainty quanti ation, material s ien e, biology, medi ine, signal pro-
essing, quantitative nan e, et . However, in most of these appli ations, the
knowledge of these random elds, whi h we write X, is limited. Numeri al
methods are therefore needed to identify the probability distribution of X
from the available information.
G. Perrin
CEA/DAM/DIF, F-91297, Arpajon, Fran e
E-mail: guillaume.perrin2 ea.fr
C. Soize
Université Paris-Est, MSME UMR 8208 CNRS, Marne-la-Vallée, Fran e
E-mail: hristian.soizeu-pem.fr
2 Guillaume Perrin, Christian Soize
When the information is onstituted of dire t measurements of the eld
to model, several te hniques have been proposed to perform su h an identi-
ation. For instan e, the AutoRegressive-Moving-Average (ARMA) models
[Whittle, 1951,Whittle, 1983,Box and Jenkins, 1970, allow the des ription of
Gaussian stationary random elds as a parameterized integral of a Gaus-
sian white noise. When onsidering a priori non-Gaussian and nonstation-
ary random elds, the identi ation is generally based on a two-step pro e-
dure. The rst step is the approximation of the random eld by its proje tion
on a redu ed number of deterministi fun tions [Ghanem and Spanos, 2003,
Le Maître and Knio, 2010, using for instan e the proper orthogonal de ompo-
sition [Atwell and King, 2001, the proper generalized de omposition [Nouy, 2010,
or the Karhunen-Loève expansion [Williams, 2011,Perrin et al., 2014,Perrin et al., 2013.
The se ond step is the identi ation of general sto hasti representations of the
proje tion oe ients in high sto hasti dimension [Soize, 2010,Soize, 2011,
Perrin et al., 2012,Nouy and Soize, 2014,Soize and Ghanem, 2016,Perrin et al., 2018.
The main spe i ity of this work omes from the fa t that only indi-
re t observations are available for the identi ation, in the sense that the
experimental data is made of the transformations of a limited number of
independent realizations of X through a bla k-box time- onsuming nonlin-
ear mapping, denoted by g. To make this identi ation tra table, we as-
sume that the random eld to identify belongs to a known parametri lass.
Thus, identifying the distribution of X amounts to identifying the values of
these parameters, whi h are gathered in the ve tor z. A Bayesian framework
is then onsidered [Marzouk and Najm, 2009,Stuart, 2010,Arnst et al., 2010,
Matthies et al., 2016,Emery et al., 2016: parameter z is supposed to be ran-
dom, and we sear h its posterior distribution given the available data.
Markov Chain Monte Carlo (MCMC) [Rubinstein and Kroese, 2008,Tian et al., 2016
is generally onsidered as a powerful tool to explore the posterior distribution
for these parameters. However it an be omputationally prohibitive when
ea h posterior evaluation requires evaluations of a omputationally expensive
ode, as it the ase here. To ir umvent this problem, a standard approa h is to
repla e the ode by a surrogate model, and to dire tly sample from the approxi-
mated posterior distribution asso iated with the modied likelihood using las-
si al MCMC pro edures. The surrogate model an be based on polynomial rep-
resentations [Marzouk and Najm, 2009,Marzouk and Xiu, 2009,Wan and Zabaras, 2011,
Li and Marzouk, 2014,Tsilis et al., 2017, Gaussian pro ess regression [Kennedy and O'Hagan, 2001,
Santner et al., 2003,Higdon et al., 2008,Bilionis and Zabaras, 2015,Sinsbe k and Nowak, 2017,
Damblin et al., 2013, or runs of the ode at dierent resolution levels [Higdon et al., 2003,
Chen and S hwab, 2015. Alternatively, the surrogate model an be used to
adapt the proposal distribution. In that ase, the number of expensive pos-
terior evaluations per MCMC step an be strongly redu ed, while sampling
asymptoti ally from the exa t posterior distribution (see [Rasmussen, 2003,
Fielding et al., 2011,Conrad et al., 2016,Conrad et al., 2018 for further details
about this approa h).
This work an be seen as an extension of these methods to the ase of
sto hasti odes. Indeed, for a given value of z, as X(z) is random, g(X(z))
Adaptive method for indire t identi ation 3
is also a random quantity. But if the distribution of X(z) is known on e
z is xed, the distribution of g(X(z)) is unknown, and its identi ation is
omputationally demanding. Therefore, instead of onstru ting a surrogate
model of the ode, we fo us on the approximation of the probability density
fun tion (PDF) of g(X(z)).To run a MCMC pro edure based on the asso iated approximated likeli-
hood in a reasonable omputational time, this approximation of the PDF of
g(X(z)) in any z has to be onstru ted from a xed number of already om-
puted ode evaluations. To this end, we rst propose to dire tly work on the
joint PDF of (g(X(z)), z). Then, we fo us on the Gaussian kernel density esti-
mation (G-KDE) [Wand and Jones, 1995,S ott and Sain, 2004,Perrin et al., 2018
for the PDF approximation. Indeed, this method is parti ularly interesting
for its ability to model non-Gaussian distributions with omplex dependen e
stru tures, but also be ause it allows an expli it derivation of the PDF of
g(X(z))|z on e the joint PDF is known. To onstru t relevant PDF ap-
proximations of this potentially high-dimensional random ve tor from a re-
du ed number of ode evaluations, we nally introdu e two adaptations of
the lassi al G-KDE formalism. First, an optimal partitioning of the om-
ponents of g(X(z)) is introdu ed, whi h onsists in de omposing the ran-
dom ve tor to model in well- hosen groups of omponents that an reason-
ably be onsidered as independent. Se ondly, a sequential strategy is proposed
to hoose the evaluations points on whi h the G-KDE relies. Starting from
a spa e-lling design, the obje tive is to sequentially add new ode evalu-
ations in the regions where the posterior distribution of the parameters is
high. We refer to [M Kay et al., 1979,Fang and Lin, 2003,Fang et al., 2006,
Dragulji¢ et al., 2012,Joseph et al., 2015 for the onstru tion of the initial
spa e-lling designs when the input spa es is an hyperre tangle, and to [Stinstra et al., 2003,
Stinstra et al., 2010,Auray et al., 2012,Dragulji¢ et al., 2012,Lekivetz and Jones, 2015,
Mak and Joseph, 2016,Perrin and Cannamela, 2017 for the general ase.
The outline of this work is as follows. Se tion 2 presents the theoreti al
framework of the proposed method. Se tion 3 rst illustrates the e ien y
of the method on an analyti al example, and then shows its potential for
the identi ation of the me hani al properties of an unknown heterogeneous
medium.
2 Indire t identi ation of the statisti al properties of random
elds
The obje tive of this se tion is to des ribe the adaptive pro edure we propose
for the identi ation of the statisti al properties of random elds when the
available information is a set of indire t observations.
2.1 Denitions and notations
Let (Ω,A,P) be a probability spa e. For dx, dy, dz ≥ 1,
4 Guillaume Perrin, Christian Soize
P(X,Rdx) denotes the spa e of all the se ond-order random elds dened
on (Ω,A,P), with values in Rdx, indexed by a ompa t and onne ted spa e
X;
L2(X,Rdx) is the spa e of all the square-integrable fun tions dened on X
with values in Rdx;
g is a nonlinear measurable mapping whose omputational ost an be
high:
g :
L2(X,Rdx) → R
dy
h 7→ g(h); (1)
X (Rdz ,Rdx) refers to a parti ular lass of random elds in P(X,Rdx), whosestatisti al properties are parameterized by a deterministi ve tor z ∈ R
dz.
For instan e, X (Rdz ,Rdx) an orrespond to the set of Gaussian random
elds, whose mean and ovarian e fun tions are parameterized by the same
dz oe ients.
For all z in Rdz, X(z) is an element of X (Rdz ,Rdx).
Let X⋆be a parti ular element of P(X,Rdx), whi h an belong or not to
X (Rdz ,Rdx), and Y ⋆be its transformation by g. By onstru tion, Y ⋆
is a
dy-dimensional random ve tor. For ea h realization of X⋆, whi h we denote
by X⋆(θ) with θ ∈ Ω, Y ⋆(θ) := g(X⋆(θ)) denes a parti ular realization of
Y ⋆.
Given N independent realizations of Y ⋆, gathered in the set
S(N) := Y ⋆(θn)1≤n≤N , θn ∈ Ω,
the purpose of this work is to propose a Bayesian formalism for the identi-
ation of z⋆, su h that the probability distribution of X(z⋆) is the losest to
the one of X⋆.
Remarks
As mentioned in Introdu tion, it is important to noti e that for ea h z ∈R
dz, g(X(z)) is random. This strongly limits the possibility of repla ing
mapping z 7→ g(X(z)) by a surrogate model, as it is lassi ally done when
solving inverse problems that invoke omputationally expensive models.
In the following, for the sake of simpli ity, we assume thatX⋆ ∈ X (Rdz ,Rdx).If it was not the ase, it ould be ne essary to introdu e an error term to
model the dieren e betweenX⋆andX(z) [Kennedy and O'Hagan, 2001.
2.2 Bayesian formulation of the problem
In this work, z⋆is modeled by the random ve tor Z, to take into a ount
the fa t that its value is unknown. Let fZ be the probability density fun tion
Adaptive method for indire t identi ation 5
(PDF) of Z, whi h is supposed to be known as a prior model. Hen e, identify-
ing z⋆amounts to sear hing the posterior PDF of Z | S(N), whi h we denote
by fZ|S(N). Using the Bayes theorem, it omes:
fZ|S(N)(z) =LS(N)(z)fZ(z)
E[LS(N)(Z)
] , z ∈ Rdz . (2)
There, E [·] is the mathemati al expe tation and LS(N) is the likelihood
fun tion. The elements of S(N) being statisti ally independent, it follows:
LS(N)(z) =N∏
n=1
fY (z)(Y⋆(θn)), z ∈ R
dz , (3)
in whi h fY (z) is the PDF of Y (z) := g(X(z)) for given z in Rdz, and is un-
known. To approximate fY (z), a rst possibility is to generateM independent
realizations of Y (z). Thus, based on this set, the value fY (z)(y) of fY (z) in
any point y in Rdy
an be approximated using any parametri or nonparamet-
ri statisti al learning te hnique. However, this means that fun tion g has to
be evaluated M ×Q times to evaluate fun tion LS(N) in Q points for z. This
qui kly be omes burdensome when the omputational ost for ea h evaluation
of g is relatively high (between several minutes to several hours CPU for the
onsidered appli ations). One possible approa h to ir umvent this problem is
to dire tly approximate the joint PDF of the (dy + dz)-dimensional random
ve tor (Y (Z),Z) [Soize and Ghanem, 2017. Indeed, M independent realiza-
tions of (Y (Z),Z) an be obtained from the following two-step pro edure:
we rst draw at randomM independent realizations of Z a ording to the
distribution fZ , whi h we denote by Z(ω1), . . ., Z(ωM ), where ω1, . . . , ωM
are in Ω;
for ea h value of z in Z(ω1), . . . ,Z(ωM ), we draw, at random and inde-
pendently the ones from the others, a parti ular realization of X(z), andwe dedu e a realization of Y (z) by evaluating g in this realization of X(z).
For the sake of simpli ity, we denote these realizations by Y (ωm), 1 ≤ m ≤M . Based on these realizations, the kernel estimator of fY ,Z is:
fY ,Z(y, z;H) :=det(H)−1/2
M
M∑
m=1
K(H−1/2 ((y, z)− (Y (ωm),Z(ωm)))
).
(4)
Here, det(·) is the determinant operator, K is any positive fun tion whose
integral over Rdy+dz
is one, and H is a ((dy + dz) × (dy + dz))-dimensional
positive-denite symmetri matrix, whi h is generally referred as the "band-
width matrix". In the following, we fo us on the ase where K is the Gaussian
multidimensional density, and where H is proportional to the empiri al esti-
mation of the ovarian e matrix of (Y (Z),Z), denoted by C:
H = h2C, h ∈ R. (5)
6 Guillaume Perrin, Christian Soize
The main interest of this hypothesis omes from the fa t that it strongly
redu es the number of parameters that need to be identied for the onstru -
tion of H , while generally leading to very interesting results for the modeling
of multivariate PDFs (see [Perrin et al., 2018 for more details). Other parsi-
monious parameterizations ould be proposed for H, su h as diagonal repre-
sentations, but for su iently high values ofM , the inuen e of this hoi e on
the identi ation results is expe ted to be small.
Hen e, the PDF of (Y (Z),Z) is approximated by a mixture ofM Gaussian
PDFs, for whi h the means are the available realizations of (Y (Z),Z) and the
ovarian e matri es are all parameterized by a unique s alar h:
fY ,Z(y, z;h) =1
M
M∑
m=1
φ((y, z); (Y (ωm),Z(ωm)), h2C
). (6)
There, for any Rd-dimensional ve tor µ and for any (Rd×R
d)-dimensional
symmetri positive-denite matrixC, φ(·;µ,C) is the PDF of any Rd-dimensional
Gaussian random ve tor with mean µ and ovarian e matrix C:
φ (x;µ,C) :=exp
(− 1
2 (x− µ)T C−1 (x− µ))
(2π)d/2√det(C)
, x ∈ Rd. (7)
In addition, the blo k de omposition of C is written as:
C =
[CY Y CY Z
CT
Y Z CZZ
]. (8)
For all (y, z) ∈ Rdy ×R
dz, the kernel approximation of fY (z)(y), whi h we
denote by fY (z)(y;h), an therefore be written as follows (see Appendix for
more details about this expression):
fY (z)(y;h) =fY ,Z(y, z;h)∫
Rdy fY ,Z(v, z;h)dv
=
M∑
m=1
γm(z;h)∑M
m′=1 γm′(z;h)φ (y;µm(z),Cm(h)) ,
(9)
γm(z;h) := exp
(−
1
2h2(z −Z(ωm))
TC
−1
ZZ (z −Z(ωm))
), (10)
µm(z) := Y (ωm) + CY ZC−1
ZZ(z −Z(ωm)), (11)
Cm(h) := h2(CY Y − CY ZC
−1
ZZCT
Y Z
). (12)
It follows that the posterior PDF of Z is estimated for ea h z in Rdz
by:
Adaptive method for indire t identi ation 7
fZ|S(N)(z) ≈LS(N)(z;h)fZ(z)
E
[LS(N)(Z)
] , LS(N)(z;h) :=N∏
n=1
fY (z)(Y⋆(θn);h). (13)
Remarks
One key step of these methods is the exploration of the whole spa e of the
input variables. To maximize this overing, it is generally worth hoosing
Z(ω1), . . . ,Z(ωM ) as a spa e lling design of experiments that preserves
good proje tion properties for ea h s alar input (see [Fang and Lin, 2003,
Fang et al., 2006,Perrin and Cannamela, 2017 for the onstru tion of su h
designs when prior density fZ is uniform or not).
Another ru ial aspe t of these Bayesian approa hes is the hoi e of prior
distribution fZ . Indeed, the more informative it is, the less measurements
we need to get a useful posterior distribution for Z. But if it is over on-
dent around values that are potentially biased, the un ertainty arried by
the posterior distribution may not be large enough to adequately apture
the true value of Z (see [Marin and Robert, 2007 for more details on the
onstru tion of this prior distribution).
In the standard ase, the M ode evaluations are generally used to on-
stru t a surrogate model of a omputationally expensive but determin-
isti ode. Hen e, depending on the dimension of the input spa e and
the regularity of the ode output with respe t to the inputs, interest-
ing approximations an be obtained using relatively small values of M
[Perrin et al., 2017. On the ontrary, in our ase, as z 7→ g(X(z)) is a
sto hasti simulator, the value of M is likely to be higher, as we want the
ode evaluations to allow a pre ise approximation of the dependen e stru -
ture between Y (Z) and Z in the onstru tion of their joint PDF. And the
higher dy + dz is, the higher value of M we may need. However, when on-
fronted to expensive simulators, the maximal number of ode evaluations
is generally limited (M must be less than 1000 for instan e). In that ase,
it is parti ularly important to work on methods that allow the most pre ise
identi ation of the parameters at the minimal ost. This is the obje tive of
the following se tions. In Se tion 2.3, we rst propose to de ompose Y (Z)in several groups to improve the relevan e of the nonparametri represen-
tation of PDF fY ,Z for a xed value of M . Then, sele tion riteria are
proposed in Se tion 2.5 to sequentially on entrate the ode evaluations in
the most likely regions for Z, and therefore redu e the un ertainties on its
posterior PDF fZ|S(N).
2.3 Optimal partitioning
As it is explained in [Perrin et al., 2018, when dy be omes high, separating
in dierent groups the omponents of Y (Z)|Z that ould reasonably be on-
sidered as independent an strongly improve the relevan e of fY (z) for a xed
8 Guillaume Perrin, Christian Soize
number of ode evaluations. Let b = (b1, . . . , bdy) be a parti ular group de-
omposition of Y (Z)|Z in the sense that:
if bi = bj , Yi(Z)|Z and Yj(Z)|Z are supposed to be dependent and there-
fore belong to the same blo k,
if bi 6= bj, Yi(Z)|Z and Yj(Z)|Z are supposed to be independent and they
an belong to two dierent blo ks.
To avoid redundan ies in this blo k by blo k representation, ve tor b an
be hosen in the set:
B(dy) :=
b ∈ 1, . . . , dy
dy | b1 = 1, 1 ≤ bj ≤ 1 + max1≤i≤j−1
bi, 2 ≤ j ≤ dy
.
(14)
Hen e, for any b in B(dy), we an dene
Max(b) as the maximal value of b,
Y (ℓ)(z, b) as the random ve tor that gathers all the omponents of Y (Z)|Z =z with a blo k index equal to ℓ,
y(ℓ)(y, b) as the ve tor that gathers all the omponents of y with a blo k
index equal to ℓ.
For all b in B(dy), z in Rdz
and h := (h1, . . . , hMax(b)) in R
Max(b), if
fY (ℓ)(z,b)(y
(ℓ)(y, b);hℓ) is the kernel estimator of the PDF of Y (ℓ)(z, b), it omes:
fY (z)(y) ≈ fY (z)(y;h, b) :=
Max(b)∏
ℓ=1
fY (ℓ)(z,b)(y
(ℓ)(y, b);hℓ), y ∈ Rdy , (15)
leading to another approximation of fZ|S(N)(z) for ea h z in Rdz:
fZ|S(N)(z) ≈ fZ|S(N)(z) :=LS(N)(z;h, b)fZ(z)
E
[LS(N)(Z)
] , (16)
LS(N)(z;h, b) :=N∏
n=1
fY (z)(Y⋆(θn);h, b). (17)
2.4 Estimation of the kernel parameters
To evaluate LS(N), the values of h and b have to be identied. This an be
done by solving the following optimization problem:
(hAIC, bAIC) ≈ arg minh∈]0,+∞[Max(b), b∈B(dy)
AIC
LOO(h, b), (18)
Adaptive method for indire t identi ation 9
AIC
LOO(h, b) := 2Max(b)−2 log
M∏
m=1
Max(b)∏
ℓ=1
f(−m)
Y (ℓ)(Z(ωm),b)(Y (ℓ)(Y (ωm), b);hℓ)
,
(19)
where f(−m)
Y (ℓ)(Z(ωm),b)is the kernel estimator of the PDF of Y (ℓ)(Z(ωm), b) that
is based on all the evaluations of g but the mth
one. Indeed, given Eq. (9), this
amounts to minimizing a "Leave-One-Out" version of the Akaike information
riterion (AIC) [Akaike, 1974 asso iated with the PDF of Y (Z)|Z (very lose
results would be obtained by onsidering another information riterion su h
as the Bayesian information riterion (BIC)). We refer to [Perrin et al., 2018
for more details about the solving of this optimization problem.
2.5 Adaptive strategy
By onstru tion, the pre ision of the estimation of z⋆depends on the num-
ber of experimental measurements, N , and the number of ode evaluations,
M . Classi ally, the value of N is xed, whereas it should be possible to im-
prove the a ura y of fY (z), whi h is dened by Eq. (15), by adding new
ode evaluations in the learning set. For instan e, Mnew
new points ould be
added to the learning set by evaluating the ode inMnew
independent realiza-
tions of Z|S(N) (we remind that no ode evaluations are required to hoose
these new points). However, as the kernel density estimator is based on the
post-pro essing of independent and identi ally distributed realizations of the
random ve tor to model, non onsistent results ould be obtained by mixing
realizations of Z|S(N) with realizations of Z. If su h a sele tion riterion was
hosen, this would mean that theM ode evaluations at the initial step should
not be used for the rening.
As an alternative, we propose to evaluate the fun tion
z 7→ f(z) := LS(N)(z;hAIC, bAIC)fZ(z)
in ea h value of Z(ω1), . . . ,Z(ωM ). For ea h 1 ≤ m ≤ M , let πm be the
following weights:
0 ≤ πm :=f(Z(ωm))
∑Mm′=1 f(Z(ωm′))
≤ 1. (20)
Without loss of generality, these weights are assumed to be sorted in de-
reasing order, π1 ≥ π2 ≥ . . . ≥ πM . Hen e, for 0 < α < 1, if we denote by Qα
the smallest integer su h that:
Qα−1∑
m=1
πm ≥ α, (21)
10 Guillaume Perrin, Christian Soize
the domain Zα := z ∈ Rdz | f(z) ≥ f(Z(ωQα
) an be seen as a onservative
α- redible set for Z|S(N), in the sense that the probability for Z|S(N) to
be in Zα is likely to be higher than α. Therefore, adding new realizations of
Z|Z ∈ Zα seems a good mean to enri h the set of points on whi h the kernel
density estimator is based. Indeed, the most likely values of z at the former
step are kept in the adaptive pro edure, while a good exploration of the input
domain is expe ted if the value of α is hosen su iently high.
Finally, hoosing fZ|Z∈Zαinstead of fZ for the prior distribution of Z,
and repeating several times this pro edure, it is possible to iteratively redu e
the un ertainties about z⋆.
Remarks
By adding new ode evaluations, the obje tive is to make fY (z) be as lose
to fY (z) as possible, su h that the approximate posterior fZ|S(N) is as lose
to the true (but unknown) posterior fZ|S(N) as possible. Choosing a value
of α that is stri tly inferior to one only aims at limiting the number of new
ode evaluations that will be in the region where true posterior fZ|S(N) is
almost zero. However, this value has not to be hosen too small, as it would
arti ially redu e the un ertainty asso iated with the estimation of z⋆by
utting too mu h the tails of the true posterior. Hen e, in the appli ations
that will be presented in Se tion 3, α is hosen equal to 0.99. A ording to Eq. (21), we deliberately add one to the value of Qα to be
onservative for the estimation of the α- redible set. This is parti ularly
important for ases when after the rst iteration, π1 ≈ 1. Indeed, even if
one value of z appears to be mu h more relevant than the others, we do
not want to fo us too mu h around a single mode.
3 Appli ations
The purpose of this se tion is to illustrate the method proposed in Se tion 2
on two appli ations.
3.1 Analyti al appli ation
In this rst appli ation, X(z) refers to the Gaussian random elds whose
mean is equal to t 7→ sin(2πz3t + z4), and whose ovarian e fun tion is equal
to (t, t′) 7→ z21 exp(− (t−t′)2
2z22
).
This lass of random elds is therefore parameterized by four quantities:
two parameters for the mean value, denoted by z3 and z4, and two parameters
for the ovarian e fun tion, denoted by z1 and z2. We then introdu e U(X(z))as the image of X(z) by the following nonlinear mapping:
Adaptive method for indire t identi ation 11
0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
PSfrag repla ements
t
U(t, ω1)
U(t, ω2)
U(t, ω3)
(a) Realizations of U(X(Z))
0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
PSfrag repla ements
t
U⋆(t, θ1)
U⋆(t, θ2)
U⋆(t, θ3)
(b) Realizations of U(X(Z))|Z = z⋆
Fig. 1 Comparison of independent realizations U(X(Z)) and U(X(Z))|Z = z⋆.
U(X(z)) := X(t; z) sin(X(t; z)), t ∈ [0, 1] . (22)
The value of z⋆is hosen equal to (0.3, 0.2, 2, 1), and it is a priori modeled
by a uniformly distributed over [0.1, 1]× [0.05, 1]× [1, 3]× [0, 2] random ve tor,
denoted by Z. To identify z⋆, the available information is made of N = 10
independent realizations of U(X(Z))|Z = z⋆, denoted by U⋆(θ1), . . . , U
⋆(θN ).To solve the inferen e problem, M = 500 independent realizations of Z have
been drawn, whi h we write Z(ω1), . . . ,Z(ωM ). For ea h 1 ≤ m ≤ M , we
then draw at random one realization of U(X(Z(ωm))), and we denote it by
U(ωm) for the sake of simpli ity. As an illustration, several realizations of
U(X(Z)) and U(X(Z))|Z = z⋆are ompared in Figure 1.
In prin iple, the Bayesian formulation an be applied to any multi-variate
output ode. But in pra ti e, it is generally very onvenient to ondense (if
it is possible) the statisti al ontent of the ode output in a low-dimensional
ve tor [Perrin, ress. In our ontext, it is even more important, as a key step of
the proposed method is the identi ation of the joint distribution between the
parameters to be identied and the asso iated ode output, whose omplexity
strongly in reases with the dimension of the ode output. In that prospe t, we
introdu e ψp, p ≥ 1 as the solutions of the following eigenvalue problem:
∫ 1
0
M∑
m=1
U(t, ωm)U(t′, ωm)ψp(t′)dt′ = λpψp(t), (23)
12 Guillaume Perrin, Christian Soize
λ1 ≥ λ2 ≥ · · · → 0,
∫ 1
0
ψp(t′)ψq(t
′)dt′ = δpq, (24)
where δpq is the Krone ker symbol that is equal to 1 if p = q and 0 otherwise. To
solve the inferen e problem, we nally introdu e Y (z) as the ve tor gatheringthe proje tion oe ients of U(X(Z)) on the former eigenfun tions asso iated
with the dy highest eigenvalues:
Y (Z) :=
(∫ 1
0
U(t;X(Z))ψ1(t)dt, . . . ,
∫ 1
0
U(t;X(Z))ψdy(t)dt
). (25)
The value of dy an then be hosen to guarantee a relevant representation
of the observations. To this end, we introdu e ε2 as the following quantity:
ε2(dy) :=
∑Nn=1
∫ 1
0
(U⋆(t, θn)− U⋆(t, θn; dy)
)2dt
∑Nn=1
∫ 1
0 (U⋆(t, θn))2dt
, (26)
U⋆(t, θn; dy) :=
dy∑
p=1
ψp(t)
(∫ 1
0
U⋆(t′, θn)ψp(t′)dt′
). (27)
As an illustration, Figure 2 shows the evolution of ε2(dy) with respe t to
dy, as well as the dieren e between U⋆(t, θ1) and U
⋆(t, θ1; dy) for three valuesof dy. For this appli ation, dy was hosen equal to 12, whi h orresponds to a
value of ε2 that is less than 1%.
Based on theseM realizations of (Y (Z),Z), and on these N realizations of
Y (z⋆) := Y (Z)|Z = z⋆, the adaptive Bayesian formalism presented in Se tion
2 is now applied. For this appli ation, the parameter α, whi h was introdu ed
in Se tion 2.5, is hosen equal to 0.99. At ea h iteration, new samples are
therefore added in the region where fZ|S(N) is not too small using a reje tion
approa h until we get a total of M points (in luding the points omputed at
the former iterations) in the α- redible set Zα, whose denition is also given
in Se tion 2.5. After 5 iterations, the total number of alls to the ode is
equal to 2300, whi h means that around 450 new points have been added at
ea h iteration. The results are summarized in Table 3.1 and Figures 3 and
4. As a rst omment, we verify that the identi ation of z⋆after only one
iteration is not very pre ise, in the sense that the predi tion un ertainties are
very high. This is not surprising, as we are trying to approximate the PDF
of a 16-dimensional random ve tor (dy = 12, dz = 4) on its whole denition
domain from only 500 realizations. Moving from M = 500 to M = 2300,that is to say spending the total budget at the rst iteration, does not really
help. Indeed, the results we get in terms of mean and varian e of Z|S(N)are approximatively the same. This is explained by the fa t that even if the
number of points is almost multiplied by ve, the overage of the denition
Adaptive method for indire t identi ation 13
0 10 20 30 40 50 60 70 80−12−11−10
−9−8−7−6−5−4−3−2−1
0
PSfrag repla ements
log10(ε
2)
(
%
)
dy
(a) Error onvergen e
0 0.5 1−0.5
0
0.5
1
PSfrag repla ements
t
U⋆(t, θ1)
U⋆(t, θ1; dy = 15)
U⋆(t, θ1; dy = 5)
U⋆(t, θ1; dy = 10)
(b) Redu ed basis representation
Fig. 2 Evolution of the proje tion error with respe t to dy .
Z1 Z2 Z3 Z4
Referen e 0.3 0.2 2 1
E [Z|S(N)], M = 2300, i = 1 0.25 0.57 2.00 1.02
E [Z|S(N)], M = 500, i = 1 0.24 0.59 2.00 1.02
E [Z|S(N)], M = 500, i = 2 0.35 0.31 2.00 1.04
E [Z|S(N)], M = 500, i = 3 0.34 0.23 2.00 1.04
E [Z|S(N)], M = 500, i = 4 0.34 0.24 2.00 1.04
E [Z|S(N)], M = 500, i = 5 0.29 0.19 2.00 1.04
Table 1 Evolution of the posterior mean with respe t to the iteration number.
domain stays very sparse. On the ontrary, adding iteratively around 450 new ode evaluations in the most likely region, whose volume is mu h smaller than
the initial volume, allows E [Z|S(N)] to tend to z⋆, and strongly redu es the
redible intervals. This onvergen e is qui ker for the mean parameters than
for the ovarian e parameters, whi h was also expe ted, as the mean fun tion
is generally easier to identify than the ovarian e. Fo using on Figure 4, it is
also interesting to noti e that the referen e value does not need to be in the
99%- redible ellipse asso iated with Z|S(N) at the rst iteration to be in the
99%- redible ellipses at the next iterations.
14 Guillaume Perrin, Christian Soize
1 2 3 4 50
1
2
3
PSfrag repla ements
V
a
l
u
e
s
o
f
Z1,Z2,Z3,Z4
Iteration
99%- redible interval of Z1|S(N)
Mean value of Z1|S(N)
99%- redible interval of Z2|S(N)
Mean value of Z2|S(N)
99%- redible interval of Z3|S(N)
Mean value of Z3|S(N)
99%- redible interval of Z4|S(N)
Mean value of Z4|S(N)
z⋆1
z⋆2
z⋆3
z⋆4
Fig. 3 Evolution of the 99%- redible intervals and of the mean values of the omponents
of Z|S(N) with respe t to the iteration number.
0 0.2 0.4 0.6 0.80
0.2
0.4
0.6
0.8
1
PSfrag repla ements
z2
z1
i = 1
i = 2
i = 3
i = 4
i = 5
Referen e point
Fig. 4 Evolution of the 99%- redible ellipses of the two rst elements of Z|S(N) with
respe t to the iteration number.
3.2 Appli ation to the identi ation of the me hani al properties of an
unknown anisotropi material
The se ond appli ation deals with the identi ation of the me hani al proper-
ties of an heterogeneous mi ro-stru ture, whi h is modeled by a random elasti
medium. To this end, several experimental tests are performed on a series of
spe imens made of the same material. To be oherent with the notations in-
trodu ed in Se tion 2, we denote by X the elasti ity eld hara terizing the
Adaptive method for indire t identi ation 15
Champ nu
SCAL
> 2.87E−01
< 3.15E−01
0.29
0.29
0.29
0.29
0.29
0.29
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.31
0.31
0.31
0.31
0.31
0.31
0.31
0.32
(a) Evolution of the Poisson ratio
Champ young
SCAL
> 2.04E+02
< 2.17E+02
2.05E+02
2.05E+02
2.06E+02
2.07E+02
2.07E+02
2.08E+02
2.08E+02
2.09E+02
2.10E+02
2.10E+02
2.11E+02
2.11E+02
2.12E+02
2.13E+02
2.13E+02
2.14E+02
2.14E+02
2.15E+02
2.16E+02
2.16E+02
2.17E+02
(b) Evolution of the Young modulus
Fig. 5 One potential spatial variation of the Young modulus and the Poisson ratio asso i-
ated with the sto hasti model X(z⋆).
Fig. 6 Representation of the studied me hani al phenomenon.
16 Guillaume Perrin, Christian Soize
me hani al properties of the material that onstitutes the spe imens. Several
sto hasti models have been proposed in the framework of the heterogeneous
anisotropi linear elasti ity [Soize, 2006,Soize, 2008,Clouteau et al., 2013,Guilleminot and Soize, 2013.
It should be noted that the elasti ity eld is not a real-valued random eld, but
a tensor-valued random eld, and that the dierent omponents of this ran-
dom eld annot be identied separately due to algebrai onstraints. For this
appli ation, the sto hasti model for the elasti ity eld is based on the model
proposed in [Soize, 2006 and [Guilleminot and Soize, 2013 in 2D plan stresses
for the sake of simpli ity. Hen e, the distribution of X is non-Gaussian, and
it is parameterized by a 5-dimensional deterministi ve tor z = (z1, . . . , z5),where:
z1 is a positive dispersion oe ient that ontrols the level of u tuations,
z2, z3 are two spatial orrelation lengths,
z4 is the mean value of the Young Modulus (×109 Pa); z5 is the mean value of the Poisson ratio.
We then assume that N = 100 ubi spe imens are available, whose re-
spe tive me hani al properties are hara terized by one parti ular realization
of X(z⋆), with z⋆ = (2000, 0.1, 0.15, 210, 0.3). As an illustration, Figure 5
shows, for one parti ular spe imen, the evolution of the Young modulus and
the Poisson ratio in ea h point of [0, 1]2. The same pressure eld fS = −fSe2is then imposed on the top of ea h spe imen, and we only have a ess to the
indu ed displa ement eld on the boundaries of these spe imens (see Figure
6 for an illustration of the experimental proto ol). Let U⋆(θ1), . . . ,U⋆(θN ) be
these measured displa ements.
Based on this set of measurements, the method des ribed in Se tion 2 ould
dire tly be applied to the identi ation of z⋆. To speed up this identi ation,
following the works a hieved in [Nguyen et al., 2015, we propose an alterna-
tive method, whi h is based on a two-step pro edure. First, z⋆4 and z⋆5 will
be identied by onfronting the measured displa ements to the homogeneous
ase. On e z⋆4 and z⋆5 have been found, a Bayesian formalism will be proposed
for the identi ation of the three remaining omponents of z⋆.
Indeed, if the spe imens were made of a homogeneous material, hara -
terized by its young modulus E and its Poisson ratio ν, it is well known
[Lai et al., 2010 that the indu ed displa ement in ea h point s ∈ [0, 1]2 wouldbe equal to uhomo(s) = (as1, bs2), with
(a
b
)=
[λ λ+ 2µ
λ+ 2µ λ
]−1(0
−fS
), µ =
E
2(1 + ν), λ =
2µν
(1 − 2ν). (28)
Hen e, as we are onsidering a lass of stationary random pro esses, the
values of z⋆4 and z⋆5 an be identied as the arguments that minimize the L2
distan e between the N measured displa ements and the asso iated homoge-
neous displa ements. In this two-step approa h, the Bayesian identi ation is
Adaptive method for indire t identi ation 17
no longer arried out in dimension 5, but in dimension 3. This strongly redu es
the number of ode evaluations that will be needed for a orre t identi ation
of (z⋆1 , z⋆2 , z
⋆3).
Thus, in the following, only z⋆1 , z⋆2 and z⋆3 are modeled by random quan-
tities. They are gathered in the ve tor Z, whose omponents are assumed
independent and distributed a ording to the following distributions:
log(Z1) ∼ U(4.6, 11.5), Z2 ∼ U(0.01, 0.3), Z3 ∼ U(0.01, 0.3), (29)
where for all a < b, U(a, b) is the uniform distribution over [a, b]. For a given
value of Z, it is possible to simulate independent realizations of X(Z), andto approximate (using the Finite Element Method) the displa ements indu ed
by the experimental for e eld, whi h we write U(X(Z)). Thus, for this se -ond appli ation, we rst hose at random M = 1000 values of Z a ording to
its prior distribution. For ea h of these values, a parti ular realization of the
elasti ity tensor was then generated over [0, 1]2, and the me hani al problem
that orresponds to the experimental proto ol was solved (using the software
Cast3M) to get the displa ements at the boundary of the ube. In the same
manner than in Se tion 3.1, we nally introdu e Y (Z) as the proje tion of
U(X(Z)) on the dy rst eigenfun tions asso iated with the empiri al estima-
tion of the ovarian e of U(X(Z)) based on the M ode evaluations. In the
same manner, we gather in S(N) the proje tion oe ients of ea h measured
displa ement U⋆(θn) on this redu ed basis. To hoose the value of dy, the
normalized error dened by Eq. (26) is on e again onsidered. For this appli-
ation, dy is hosen equal to 23 in order to orre tly represent most of the lo al
os illations of the displa ements. A ording to Figure 7, this orresponds to a
proje tion error that is less than 0.1%.
Following the framework proposed in Se tion 2, the PDF of Z|S(N) is
dedu ed from the kernel estimator of the PDF of (Y (Z),Z). An adaptive
pro edure (with α = 0.99) is moreover introdu ed to better on entrate the
distribution of Z|S(N) on the true value of z⋆. To be more pre ise, 900 new
ode evaluations were added between the two rst iterations, and 620 betweenthe two last iterations, leading to a total budget of 2520 ode evaluations. Therelevan e of this approa h is shown in Figure 8, where the blue ontinuous
lines orrespond to the 95%- redible ellipses asso iated with the distribution
of Z|S(N). After three iterations, the values of z⋆1 , z⋆2 and z⋆3 are indeed iden-
tied with a high pre ision. To emphasize the interest of the partitioning pre-
sented in Se tion 3.2, these results are ompared to the ase where there is no
optimization of the blo k stru ture (the ellipses in red dotted lines). Although
these two approa hes are based on the same information, there is no denying
that sear hing groups of independent omponents of Y (Z)|Z is really helpful.
This is espe ially true for the rst iteration, where 23 groups of independent
omponents were hosen, and for the se ond iteration, where 8 groups of in-
dependent omponents were hosen. For the third iteration, as only 4 groups
18 Guillaume Perrin, Christian Soize
0 20 40 60 80 100−5
−4
−3
−2
−1
0
PSfrag repla ements
log10(ε
2)
dy
(a) Error onvergen e
0 0.2 0.4 0.6 0.8 1−1.5
−1
−0.5
0
0.5
1
1.5x 10−4
PSfrag repla ements
U⋆ 1
s2
Measured data
Proje tion with dy = 10
Proje tion with dy = 20
Proje tion with dy = 30
(b) Redu ed basis representation of U⋆1 (1, s2; θ1)− uhomo(1, s2)
0 0.2 0.4 0.6 0.8 1−5
0
5x 10−5
PSfrag repla ements
U⋆ 2
s1
Measured data
Proje tion with dy = 10
Proje tion with dy = 20
Proje tion with dy = 30
( ) Redu ed basis representation of U⋆2 (s1, 1; θ1)− uhomo(0, 1)
Fig. 7 Inuen e of trun ation parameter dy on the representation of the measured data.
Adaptive method for indire t identi ation 19
6 8 10
0.05
0.1
0.15
0.2
0.25
0.3
PSfrag repla ements
z1
z2
, blo k stru ture
, no blo k stru ture
, no blo k stru ture
Referen e point
(a) (z1, z2), i = 1
6 8 10
0.05
0.1
0.15
0.2
0.25
0.3
PSfrag repla ements
z1
z2
, blo k stru ture
, no blo k stru ture
, no blo k stru ture
Referen e point
(b) (z1, z2), i = 2
6 8 10
0.05
0.1
0.15
0.2
0.25
0.3
PSfrag repla ements
z1
z2
, blo k stru ture
, no blo k stru ture
, no blo k stru ture
Referen e point
( ) (z1, z2), i = 3
6 8 10
0.05
0.1
0.15
0.2
0.25
0.3
PSfrag repla ements
z1
z3
, blo k stru ture
, no blo k stru ture
, no blo k stru ture
Referen e point
(d) (z1, z3), i = 1
6 8 10
0.05
0.1
0.15
0.2
0.25
0.3
PSfrag repla ements
z1
z3
, blo k stru ture
, no blo k stru ture
, no blo k stru ture
Referen e point
(e) (z1, z3), i = 2
6 8 10
0.05
0.1
0.15
0.2
0.25
0.3
PSfrag repla ements
z1
z3
, blo k stru ture
, no blo k stru ture
, no blo k stru ture
Referen e point
(f) (z1, z3), i = 3
0.1 0.2 0.3
0.05
0.1
0.15
0.2
0.25
0.3
PSfrag repla ements
z2
z3
, blo k stru ture
, no blo k stru ture
, no blo k stru ture
Referen e point
(g) (z2, z3), i = 1
0.1 0.2 0.3
0.05
0.1
0.15
0.2
0.25
0.3
PSfrag repla ements
z2
z3
, blo k stru ture
, no blo k stru ture
, no blo k stru ture
Referen e point
(h) (z2, z3), i = 2
0.1 0.2 0.3
0.05
0.1
0.15
0.2
0.25
0.3
PSfrag repla ements
z2
z3
, blo k stru ture
, no blo k stru ture
, no blo k stru ture
Referen e point
(i) (z2, z3), i = 3
Fig. 8 Evolution of the 95%- redible ellipses with respe t to the iteration number. Blue
ontinuous line: dy = 23 with optimization of the blo k stru ture. Red dotted line: dy = 23without optimization of the blo k stru ture. Green dashed line: dy = 5 without optimization
of the blo k stru ture.
20 Guillaume Perrin, Christian Soize
were hosen, introdu ing the partitioning does not make a big dieren e for
the PDF identi ation, whi h explains the similarities between the blue and
the red urves.
This set of gures also emphasizes the importan e of onsidering a high
value of dy , even if it ompli ates the PDF identi ation. For instan e, hoosing
dy = 5 leads to the results in green dashed lines, whi h are learly less relevant
than the results in blue that orrespond to dy = 23. Intermediate results were
obtained for values of dy between 5 and 23, whereas still in reasing dy did not
really hange the results.
In order to emphasize the e ien y of the proposed method to re over
the true underlying sto hasti ity, three additional bat hes of Q = 104 sim-
ulations are laun hed. These simulations are asso iated with the same ubi
system than in Figure 6, but with dierent boundary onditions (by hang-
ing the boundary onditions, we want to verify that the identied values of
Z are not dependent of a xed onguration). While the boundary ondi-
tions on the inferior side of the ube do not hange, the left and right sides
are now free of onstraints, and the displa ements on the superior side are
hosen equal to 0.002e1 − 0.01e2. We then denote by
X(1,q), 1 ≤ q ≤ Q
,
X(2,q), 1 ≤ q ≤ Q
and
X(3,q), 1 ≤ q ≤ Q
the elasti ity elds hara teriz-
ing the material properties of the dierent ubes of the three sets respe tively.
For all 1 ≤ q ≤ Q,
X(1,q)is an independent realization of the true elasti ity eld, X(z⋆);
X(2,q)is an independent realization of X((zq,prior, z⋆4 , z
⋆5)), where zq,prior
is a realization of Z, whose distribution is given by Eq. (29),
X(3,q)is an independent realization of X((zq,post, z⋆4 , z
⋆5)), where z
q,postis
a realization of Z|S(N) after the three formerly presented iterations.
For ea h simulation, we denote by U(X(i,q)), 1 ≤ i ≤ 3, the on atena-
tion of the verti al and horizontal displa ements that are indu ed on the left
and right sides of the ube. To ompare the statisti al information gathered
in these displa ements, we then ompute, for ea h 1 ≤ i ≤ 3, the eigenval-
ues
v(i)j , j ≥ 0
asso iated with the empiri al estimate of their ovarian e
matri es. In addition, we denote by σVM(X(i,q)) the maximum value over the
ubi domain of the Von Mises stress. This Von Mises riterion is ommonly
used to hara terize the resistan e of the system (see [Lai et al., 2010 for more
details). The de rease of these eigenvalues and the PDF of these Von Mises
riteria are nally ompared in Figure 9. Looking at these gures, we see that
the results asso iated with the posterior distribution of Z are very lose to the
ones asso iated with the true elasti ity eld, whi h is not true for the results
asso iated with the prior distribution of Z. This underlines the apa ity of the
proposed method to take into a ount indire t observations for the identi a-
Adaptive method for indire t identi ation 21
0 50 100 150 200 250 30010
−4
10−3
10−2
10−1
100
101
PSfrag repla ements
j
v(1)j
v(2)j
v(3)j
(a) Eigenvalues
0 5 10 15 20 25 30 35 400
0.02
0.04
0.06
0.08
0.1
0.12
0.14
PSfrag repla ements
σ
V
M
σVM(X(1,q))
σVM(X(2,q))
σVM(X(3,q))
(b) Von Mises riterion
Fig. 9 Representation of the eigenvalues de reases (a) and of the PDFs of the Von Mises
riteria (b) for the three ompared ongurations. In ea h gure, the blue ontinuous lines are
asso iated with the referen e ase, the red dashed line orrespond to the results asso iated
with the prior distribution of Z, when the bla k two-dashed lines orrespond to the results
asso iated with the posterior distribution of Z.
tion of the parameters hara terizing the distribution of an unknown random
pro ess of interest.
22 Guillaume Perrin, Christian Soize
4 Con lusion
The in reasing of the omputational resour es and the generalization of the
monitoring of me hani al systems have en ouraged many s ienti elds to
take into a ount random elds in their modeling. In that prospe t, this work
proposes an adaptive Bayesian framework to e iently identify the statisti al
properties of these random elds when the available information is a redu ed
set of indire t observations. Two examples based on simulated data are nally
presented to show the potential of this approa h.
Extending this approa h to the ases where the number of parameters to
identify and the number of observations are very high would be interesting for
future work.
Appendix
A.1. Proof of the equality of Eq. 9
Let A,B,D be the blo k de omposition matri es of C−1:
C−1
=
[A B
BT D
]. (30)
Using the S hur omplement, if follows that:
C−1
ZZ = D −BTA−1B,
(CY Y − CY ZC−1
ZZC
T
Y Z)−1 = A,
− CY ZC−1
ZZ= A−1B.
(31)
It omes
((y, z)− (Y (ωm),Z(ωm)))T(h2C)−1 ((y, z)− (Y (ωm),Z(ωm)))
=1
h2
((y − Y (ωm))TA(y − Y (ωm)) + 2(y − Y (ωm))TAA−1B (z −Z(ωm))
+ (z −Z(ωm))TD(z −Z(ωm))
)
=1
h2
(y − Y (ωm) +A−1B (z −Z(ωm)))TA(y − Y (ωm) +A−1B (z −Z(ωm)))
+ (z −Z(ωm))T(D −BTA−1B
)(z −Z(ωm))
= (y − µn(z))TC−1
n (y − µn(z)) +1
h2(z −Z(ωm))
TC−1
ZZ(z −Z(ωm))
(32)
This leads to the sear hed result.
Adaptive method for indire t identi ation 23
Referen es
Akaike, 1974. Akaike, H. (1974). A new look at the statisti al model identi ation. IEEE
Transa tions on Automati Control, 19 (6):716723.
Arnst et al., 2010. Arnst, M., Ghanem, R., and Soize, C. (2010). Identi ation of bayesian
posteriors for oe ients of haos expansions. Journal of Computational Physi s, 229
(9):31343154.
Atwell and King, 2001. Atwell, J. and King, B. (2001). Proper orthogonal de omposition
for redu ed basis feedba k ontrollers for paraboli equations. Math. Comput. Modell., 33
(1-3):119.
Auray et al., 2012. Auray, Y., Barbillon, P., and Marin, J. M. (2012). Maximin design on
non hyper ube domains and kernel interpolation. Statisti s and Computing, 22(3):703
712.
Bilionis and Zabaras, 2015. Bilionis, I. and Zabaras, N. (2015). Bayesian un ertainty prop-
agation using gaussian pro esses. In: Ghanem R., Higdon D., Owhadi H. (eds) Handbook
of Un ertainty Quanti ation. Springer.
Box and Jenkins, 1970. Box, G. and Jenkins, G. (1970). Time Series Analysis: Fore asting
and Control. Holden-Day, San Fran is o.
Chen and S hwab, 2015. Chen, P. and S hwab, C. (2015). Sparse-grid, redu ed basis
bayesian inversion. Computer Methods in Applied Me hani s and Engineering, 297:84
115.
Clouteau et al., 2013. Clouteau, D., Cottereau, R., and Lombaert, G. (2013). Dynami s of
stru tures oupled with elasti media - a review of numeri al models and methods. J.
Sound Vibr., 332:24152436.
Conrad et al., 2018. Conrad, P. R., Davis, A., Marzouk, Y. M., Pillai, N. S., and Smith,
A. (2018). Parallel lo al approximation MCMC for expensive models. SIAM/ASA J.
Un ertainty Quanti ation, 6(1):39373.
Conrad et al., 2016. Conrad, P. R., Marzouk, Y. M., Pillai, N. S., and Smith, A. (2016).
A elerating asymptoti ally exa t MCMC for omputationally intensive models via lo al
approximations. Journal of the Ameri an Statisti al Asso iation, 111:15911607.
Damblin et al., 2013. Damblin, G., Barbillonz, P., Keller, M., Pasanisi, A., and Parent, E.
(2013). Adaptive Numeri al Designs for the Calibration of Computer Codes. SIAM/ASA
J. Un ertainty Quanti ation, 6(1):151179.
Dragulji¢ et al., 2012. Dragulji¢, D., Santner, T. J., and Dean, A. M. (2012). Non ollapsing
Spa e-Filling Designs for Bounded Nonre tangular Regions. Te hnometri s, 54(2):169
178.
Emery et al., 2016. Emery, J., Grigoriu, M., and Jr., R. F. (2016). Bayesian methods for
hara terizing unknown parameters of material models. Applied Mathemati al Modelling,
13-14:63956411.
Fang et al., 2006. Fang, K., Li, R., and Sudjianto, A. (2006). Design and modeling for
omputer experiments. Chapman & Hall, Computer S ien e and Data Analysis Series,
London.
Fang and Lin, 2003. Fang, K. and Lin, D. (2003). Uniform experimental designs and their
appli ations in industry. Handbook of Statisti s, 22:131178.
Fielding et al., 2011. Fielding, M., Nott, D. J., and Liong, S. Y. (2011). E ient MCMC
s hemes for Computationally Expensive Posterior Distributions. Te hnometri s, 53(1):16
28.
Ghanem and Spanos, 2003. Ghanem, R. and Spanos, P. D. (2003). Sto hasti Finite Ele-
ments: A Spe tral Approa h, rev. ed. Dover Publi ations, New York.
Guilleminot and Soize, 2013. Guilleminot, J. and Soize, C. (2013). On the statisti al de-
penden e for the omponents of random elasti ity tensors exhibiting material symmetry
properties. Journal of Elasti ity, 111:109130.
Higdon et al., 2008. Higdon, D., Gattiker, J., Williams, B., and Rightley, M. (2008). Com-
puter model alibration using high-dimensional output. Journal of the Ameri an Statis-
ti al Asso iation, 103(482):570583.
Higdon et al., 2003. Higdon, D., Lee, H., and Holloman, C. (2003). Markov hain monte
arlo based approa hes for inferen e in omputationally intensive inverse problems.
Bayesian Statisti s, 7:181197.
24 Guillaume Perrin, Christian Soize
Joseph et al., 2015. Joseph, V. R., Gul, E., and Ba, S. (2015). Maximum proje tion designs
for omputer experiments. Biometrika, 102(2):371380.
Kennedy and O'Hagan, 2001. Kennedy, M. and O'Hagan, A. (2001). Bayesian alibration
of omputer models. Journal of the royal statisti al so iety, 63:425464.
Lai et al., 2010. Lai, W. M., Rubin, D., and Krempl, E. (2010). Introdu tion to Continuum
Me hani s. Elsevier, In .
Le Maître and Knio, 2010. Le Maître, O. and Knio, O. (2010). Spe tral Methods for Un-
ertainty Quanti ation. Springer.
Lekivetz and Jones, 2015. Lekivetz, R. and Jones, B. (2015). Fast Flexible Spa e-Filling
Designs for Nonre tangular Regions. Quality and Reliability Engineering International,
31(5):829837.
Li and Marzouk, 2014. Li, J. and Marzouk, Y. M. (2014). Adaptive onstru tion of surro-
gates for the bayesian solution of inverse problems. SIAM Journal on S ienti Comput-
ing, 36:A1163A1186.
Mak and Joseph, 2016. Mak, S. and Joseph, V. R. (2016). Minimax designs using lustering.
pages 124.
Marin and Robert, 2007. Marin, J. M. . and Robert, C. P. (2007). Bayesian ore. Springer-
Verlag, New York.
Marzouk and Najm, 2009. Marzouk, Y. M. and Najm, H. N. (2009). Dimensionality redu -
tion and polynomial haos a eleration of bayesian inferen e in inverse problems. Journal
of Computational Physi s, 228 (6):18621902.
Marzouk and Xiu, 2009. Marzouk, Y. M. and Xiu, D. (2009). A sto hasti ollo ation ap-
proa h to bayesian inferen e in inverse problems. Communi ations in Computational
Physi s, 6:826847.
Matthies et al., 2016. Matthies, H., Zander, E., Rosi, B., and Litvinenko, A. (2016). Pa-
rameter estimation via onditional expe tation: a Bayesian inversion. Adv. Model. and
Simul. in Eng. S i., 3(24).
M Kay et al., 1979. M Kay, M., Be kman, R., and Conover, W. (1979). A omparison
of three methods for sele ting values of input variables in the analysis of output from a
omputer ode. Te hnometri s, 21:239245.
Nguyen et al., 2015. Nguyen, M. T., Des eliers, C., Soize, C., Allain, J., and Gharbi, H.
(2015). Multis ale identi ation of random elasti ity eld at mesos ale of a heteroge-
neous mi rostru ture using multis ale experimental observations. Journal for Multis ale
Computational Engineering, 13 (4):281295.
Nouy, 2010. Nouy, A. (2010). Proper generalized de omposition and separated represen-
tations for the numeri al solution of high dimensional sto hasti problems. Ar hives of
omputational methods in engineering, 17:403434.
Nouy and Soize, 2014. Nouy, A. and Soize, C. (2014). Random elds representations for
sto hasti ellipti boundary value problems and statisti al inverse problems. Eur. J. Appl.
Math., 25:339373.
Perrin, ress. Perrin, G. (in press). Adaptive alibration of a omputer ode with time-series
output. Reliability Engineering and System Safety.
Perrin and Cannamela, 2017. Perrin, G. and Cannamela, C. (2017). A repulsion-based
method for the denition and the enri hment of opotimized spa e lling designs in on-
strained input spa es. Journal de la So iété Française de Statistique, 158(1):3767.
Perrin et al., 2012. Perrin, G., Soize, C., Duhamel, D., and Funfs hilling, C. (2012). Identi-
ation of polynomial haos representations in high dimension from a set of realizations.
SIAM J. S i. Comput., 34(6):29172945.
Perrin et al., 2013. Perrin, G., Soize, C., Duhamel, D., and Funfs hilling, C. (2013).
Karhunenloève expansion revisited for ve tor-valued random elds: S aling, errors and
optimal basis. Journal of Computational Physi s, 242:607622.
Perrin et al., 2014. Perrin, G., Soize, C., Duhamel, D., and Funfs hilling, C. (2014). A
posteriori error and optimal redu ed basis for sto hasti pro esses dened by a nite set
of realizations. SIAM/ASA J. Un ertainty Quanti ation, 2:745762.
Perrin et al., 2017. Perrin, G., Soize, C., Marque-Pu heu, S., and Garnier, J. (2017). Nested
polynomial trends for the improvement of gaussian pro ess-based predi tors. Journal of
Computational Physi s, 346:389402.
Adaptive method for indire t identi ation 25
Perrin et al., 2018. Perrin, G., Soize, C., and Ouhbi, N. (2018). Data-driven kernel repre-
sentations for sampling with an unknown blo k dependen e stru ture under orrelation
onstraints. Journal of Computational Statisti s and Data Analysis, 119:139154.
Rasmussen, 2003. Rasmussen, C. E. (2003). Gaussian pro esses to speed up hybrid monte
arlo for expensive bayesian integrals. Bayesian Statisti s, 7:651659.
Rubinstein and Kroese, 2008. Rubinstein, R. T. and Kroese, D. (2008). Simulation and the
Monte Carlo method. John Wiley and Sons, In ., Hoboken, New Jersey.
Santner et al., 2003. Santner, T. J., Williams, B., and Notz, W. (2003). The design and
analysis of omputer experiments. Springer, New York.
S ott and Sain, 2004. S ott, D. W. and Sain, S. R. (2004). Multidimensional Density Esti-
mation.
Sinsbe k and Nowak, 2017. Sinsbe k, M. and Nowak, W. (2017). Sequential Design of Com-
puter Experiments for the Solution of Bayesian Inverse. SIAM/ASA J. Un ertainty Quan-
ti ation, 5:640664.
Soize, 2006. Soize, C. (2006). Non-gaussian positive-denite matrix-valued random elds
for ellipti sto hasti partial dierential operators. Comput. Methods Appl. Me h. Eng.,
195:2664.
Soize, 2008. Soize, C. (2008). Tensor-valued random elds for meso-s ale sto hasti model
of anisotropi elasti mi rostru ture and probabilisti analysis of representative volume
element size. Probab. Eng. Me h., 23:307323.
Soize, 2010. Soize, C. (2010). Identi ation of high-dimension polynomial haos expansions
with random oe ients for non-Gaussian tensor-valued random elds using partial and
limited experimental data. Computer Methods in Applied Me hani s and Engineering,
199(33-36):21502164.
Soize, 2011. Soize, C. (2011). A omputational inverse method for identi ation of non-
Gaussian random elds using the Bayesian approa h in very high dimension. Computer
Methods in Applied Me hani s and Engineering, 200(4546):30833099.
Soize and Ghanem, 2016. Soize, C. and Ghanem, R. (2016). Data-driven probability on-
entration and sampling on manifold. Journal of Computational Physi s, 321(September
2015):242258.
Soize and Ghanem, 2017. Soize, C. and Ghanem, R. (2017). Probabilisti learning on man-
ifold for optimization under un ertainties. Pro eeding of Un e omp 2017, pages 115.
Stinstra et al., 2003. Stinstra, E., den Hertog, D., Stehouwer, P., and Vestjens, A. (2003).
Constrained maximin designs for omputer experiments. Te hnometri s, 45(4):340346.
Stinstra et al., 2010. Stinstra, E., den Hertog, D., Stehouwer, P., and Vestjens, A. (2010).
Uniform designs over general input domains with appli ations to target region estimation
in omputer experiments. Computational Statisti sand Data Analysis, 51(1):219232.
Stuart, 2010. Stuart, A. M. (2010). Inverse problems: A bayesian perspe tive. A ta Nu-
meri a, 19:451559.
Tian et al., 2016. Tian, M., Li, D., Cao, Z., Phoon, K., and Wang, Y. (2016). Bayesian iden-
ti ation of random eld model using indire t test data. Engineering Geology, 210:197
211.
Tsilis et al., 2017. Tsilis, P., Ghanem, R. G., and Hajali, P. (2017). E ient Bayesian
Experimentation Using an Expe ted Information Gain Lower Bound. SIAM/ASA J.
Un ertainty Quanti ation, 5:3062.
Wan and Zabaras, 2011. Wan, J. and Zabaras, N. (2011). A Bayesian approa h to mul-
tis ale inverse problems using the sequential Monte Carlo method. Inverse Problems,
27.
Wand and Jones, 1995. Wand, M. P. and Jones, M. C. (1995). Kernel Smoothing. En y-
lopedia of Statisti s in Behavioral S ien e, 60(60):212.
Whittle, 1951. Whittle, P. (1951). Hypothesis testing in time series, PhD thesis. University
of Uppsala.
Whittle, 1983. Whittle, P. (1983). Predi tion and Regulation by Linear Least-Square Meth-
ods. University of Minnesota Press.
Williams, 2011. Williams, M. (2011). The eigenfun tions of the karhunen-loeve integral
equation for a spheri al system. Propabilisti Engineering Me hani s, 26:202207.