Computers and Structures 83 (2005) 1062–1075

www.elsevier.com/locate/compstruc

Design and sensitivity analysis of dynamical systems

subjected to stochastic loading
Hector A. Jensen *

Department of Civil Engineering, Federico Santa Maria University, Casilla 110-V, Valparaiso, Chile

Accepted 30 November 2004

Available online 26 February 2005

Abstract

The paper presents an efficient procedure which allows to carry out reliability-based optimization of linear systems
subjected to stochastic loading. The optimization problem is replaced by a sequence of approximate explicit sub-opti-
mization problems that are solved in an efficient manner. Approximation concepts are used to construct high quality
approximations of dynamic responses during the optimization process. The approximations are combined with efficient
simulation methods to generate explicit approximations of reliability measures in terms of the design variables. The
number of dynamic analyses required for the convergence of the design process is reduced dramatically. An efficient
sensitivity analysis with respect to the optimization variables and general system parameters becomes possible with
the proposed formulation. The sensitivity is evaluated by considering the behavior of the design when the parameters
vary within a bounded region. The analysis can identify the degree of robustness of the final design with respect to vari-
ations of selected system parameters. A numerical example in terms of a 26-story reinforced concrete building under
stochastic earthquake excitation exemplifies the proposed methodology.

2005 Elsevier Ltd. All rights reserved.

Keywords: Approximation concepts; Excursion probability; Linear systems; Optimization; Sensitivity analysis

1. Introduction the most difficult problems in the area of stochastic

dynamics. Early work on the first excursion problem
The optimization of structural systems subjected to has been focused on analytical and numerical solution
stochastic loading involves the calculation of design methods [1–6]. In general, these methods are limited in
parameters to meet some performance criteria with cor- application to simple systems of small size. Recently,
responding specified reliability over the service life. The efficient simulation techniques have been developed to
probability that design conditions are satisfied within a solve first excursion problems of more general systems
particular reference period provides a useful measure [7–10]. These methods are robust in application and they
of system performance. Such measure corresponds to a have shown promise to be suitable for a wide range of
first excursion probability problem, which is among dynamical problems. In an optimization environment,
reliability measures have to be evaluated several times
before a near optimal design can be obtained. The esti-
*
Corresponding author. Fax: +56 32 62 58 15. mates of the reliability measures at a given design are
E-mail address: [email protected] obtained by carrying out a simulation procedure. Then,

0045-7949/$ - see front matter
2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruc.2004.11.016
 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075 1063

the evaluation of the reliability measures for every subject to the design constraints
change of the optimization variables requires the evalu-
ation of dynamic responses of the structural system. In gq ðfhg; fsðfhgÞgÞ 6 0; q ¼ 1; . . . ; nc fhg 2 H; ð1Þ
general, these responses are nonlinear implicit functions
where c( Æ ) represents the design objective, {s({h})} de-
of the optimization variables and they are available only
notes the vector of considered reliability measures, nc
in a numerical way, for instance, by means of a finite ele-
is the number of constraints gq(Æ, Æ), and H is the set that
ment procedure. For systems of practical interest, the
contains the side constraints for the vector of design
repeated evaluation of dynamic responses can be extre-
variables {h}. In the present formulation, the objective
mely time consuming. Therefore, the use of direct opti-
function is assumed to be an explicit function of the de-
mization procedures is generally not applicable in the
sign variables only, that is, c({h}). On the other hand,
context of structural optimization of systems subjected
the probability that some stochastic dynamic responses
to stochastic excitation.
exceed within a specified time interval [0,T] certain crit-
In this paper an alternative method to the standard
ical upper bounds or fall below critical lower bounds is
direct optimization method is introduced. A local
used as reliability measure. Then, the optimization prob-
approximation strategy is used to approximate the sys-
lem (1) is rewritten as
tem responses as well as the reliability measures. System
responses and probability of failures are approximated Minimize cðfhgÞ
as functions of the optimization variables during the subject to the design constraints
optimization process. This strategy allows a formal sep-
aration of the system reliability analysis from the opti- [
nq
P Fq ðfhgÞ ¼ P f9t 2 ½0;T : rqi ðt; fhgÞ P rqi ðtÞ _ rqi ðt;fhgÞ
mization procedure. The proposed methodology can i¼1
be seen as a further development of the method pre- !
sented in Ref. [11]. In the present work, the use of local 6 rqi ðtÞg 6 P Fq q ¼ 1; .. . ;nc fhg 2 H; ð2Þ
approximations for the dynamic responses dramatically
reduces the number of exact structural analyses required
for the convergence of the optimal design process. Also, where rqi ðt; fhgÞ; i ¼ 1; . . . ; nq are the response functions
a very efficient sensitivity analysis with respect to the associated with the failure criterion q evaluated at the
dependencies of the final design on the variation of sys- design {h}, rqi ðtÞ is the critical upper bound, rqi ðtÞ is the
tem parameters becomes possible. This is due to the critical lower bound, and P Fq is the acceptable prescribed
availability of the quantities involved in the optimization or target failure probability for criterion q. It is noted
problem in explicit form. A parameter study can be car- that additional constraints related to general design
ried out without any considerable increase in the compu- requirements as well as objective functions dependent
tational effort. The respective influence on the final on reliability measures can also be included in the opti-
design can be determined easily. At the same time, the mal design problem. The methodology to be developed
degree of robustness of the final design with respect to in this work can be extended in a straightforward man-
variations of selected system parameters can be obtained ner for the more general case.
directly. In this context, the sensitivity information is
crucial to ensure the validity of the final design.
First, the definition of the optimization problem with 3. Response representation
reliability constraints is presented. The characterization
of the system responses as well as the stochastic loading The general matrix equation of motion for a damped
is then considered. Next, approximation concepts and linear structure is given by
efficient methodologies to estimate excursion probabili- ½M f€xðtÞg þ fRðfxðtÞg; f_xðtÞgÞg ¼ ½G ff ðtÞg; ð3Þ
ties are discussed. Finally, a building structure subjected
to a stochastic earthquake excitation is considered as an where {x(t)} is displacement response vector of dimen-
example problem to show the computational capability sion n, fRðfxðtÞg; f_xðtÞgÞg ¼ ½C f_xðtÞg þ ½K fxðtÞg is the
and performance of the proposed methodology. linear restoring force, [M], [C], and [K] are the mass,
damping and stiffness matrices of dimension n · n, {f}
is the excitation vector of dimension nf, and [G] is a
2. Problem definition
n · nf dimensional matrix that couples the excitation
components of the vector {f} to the degrees of freedom
The structural optimization problem of systems sub-
of the structure. In this paper a modal solution of the
jected to stochastic loading can be formulated as the
dynamic response problem will be used. In the modal
identification of the design vector {h}, hi, i = 1, . . . , nd
solution approach, it is assumed that the dynamic re-
such that
sponse can be represented by a linear combination of
Minimize cðfhg; fsðfhgÞgÞ mode shapes:
1064 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075

X
n
classical eigenvalue problem, which needs to be solved
fxðtÞg ¼ f/gr gr ðtÞ; ð4Þ
if the state representation of the equation of motion is
r¼1
used. It is also noted that Eq. (9) can be interpreted as
where gr(t), i = 1, . . . , n are the modal responses, and the modal superposition formula for the impulse
{/r}, r = 1, . . . , n are the eigenvectors associated with response functions. One advantage of this representa-
the eigenproblem of the undamped equation of motion. tion is that in general only a relatively small number
If the system is classically damped, the modal responses of modes will be needed in the dynamic analysis. In this
satisfy the differential equation case, the contribution of the remaining modes to the
f/gTr ½G ff ðtÞg structural response is ignored and the structural
gr ðtÞ þ 2fr xr g_ r ðtÞ þ x2r gr ðtÞ ¼
€ ; response ri(t) is written as
f/gTr ½M f/gr
nf Z
r ¼ 1; . . . n; ð5Þ X t
ri ðtÞ ¼ hij ðt sÞ f j ðsÞds; ð10Þ
0
where xr, r = 1, . . . , n are the natural frequencies of the j¼1

system, and fr, r = 1, . . . , n are the corresponding damp- with

ing ratios. The solution for the modal responses corre-
sponding to a unit impulse applied at the jth X
m n
fcgTi f/gr f/gTr fgj g 1 fr x r t
hij ðtÞ ¼ e sinðxdr tÞ;
component of the excitation vector at time t = 0 takes r¼1 f/gTr ½M f/gr xdr
the form
ð11Þ
f/gTr fgj g
1 fr x r t
gr ðtÞ ¼ T
e sinðxdr tÞ; where m is the number of modes considered in the dy-
x
f/gr ½M f/gr dr
namic analysis. The contribution of the remaining
r ¼ 1; . . . ; n; ð6Þ modes can also be considered in the formulation by
j} ffiffiffiffiffiffiffiffiffiffiffiffi
where {gq is the using the static solution of the structural modes of
ffi jth column of the [G] matrix, and
higher order [12]. In that case, the response consists of
xdr ¼ xr 1 f2r is the damped frequency. Then, the re-
a dynamic part and a static part. The consideration
sponse xi(t), (ith component of the displacement vector) of the static part of the solution was not critical in terms
can be written as of the accuracy of the system responses for the cases
nf Z t
X considered here. Therefore, the static contribution of
xi ðtÞ ¼ hij ðt sÞfj ðsÞds; ð7Þ the structural modes of higher order is not considered
j¼1 0
in this formulation. Finally, it is noted that in general
where the matrices [M], [C], [K], and [G] depend on the vector
of design variables and therefore any response function
X
n
/ir f/gTr fgj g 1
hij ðtÞ ¼ e fr x r t
sinðxdr tÞ ð8Þ ri(t) is also a function of {h}. The dependence of the re-
r¼1 f/gTr ½M f/gr xdr sponse functions on the vector of design variables is not
shown here for simplicity in notation. It is also noted
is the unit impulse response function for the response
that expressions similar to Eq. (9) can be obtained for
function xi(t) at time t due to a unit impulse applied at
the impulse response functions of other responses such
the jth input at time 0. In Eq. (7), zero initial conditions
as velocities and accelerations.
at time t = 0 have been assumed without loss of
generality.
Similar expressions can be derived if other response
functions are considered. For example, if the response 4. Stochastic loading
function of interest ri(t) is given as a linear combination
of the components of the displacement vector, that is, In the present formulation the stochastic loading is
ri ðtÞ ¼ fcgTi fxðtÞg, the corresponding impulse response assumed to be of seismic type. Hence, the loading vector
function is given by is assumed to be of the form [M][G]{a(t)}, where {a(t)}
is a discrete, stochastic earthquake acceleration vector
Xn
fcgTi f/gr f/gTr fgj g 1 fr x r t applied at the structures supports. The components of
hij ðtÞ ¼ e sinðxdr tÞ: ð9Þ
r¼1 f/gTr ½M f/gr xdr the acceleration vector are modelled as statistically inde-
pendent stochastic processes defined as filtered white
Note that since the present approach does not rely on noise. Each component aj(t) is defined as aj ðtÞ ¼
the state-space formulation of the equation of fbgTj fyðtÞg, where {b}j is a constant vector, and {y(t)}
motion, the evaluation of the impulse response functions denotes the state-vector of the filter which satisfies a
requires the solution of just one classical eigenvalue- first-order differential equation of the form
eigenvector problem. The solution of this problem is
computationally inexpensive compared with the non- fyðtÞg
_ ¼ ½B j fyðtÞg þ fbgj ej ðtÞwj ðtÞ; ð12Þ
 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075 1065

where [B]j denotes the system matrix of dimension n where

F · nF, {b}j is the distribution vector of the filter of
fbgTj fwgsj fvgTsj fbgj
dimension nF, ej(t) is a deterministic modulating time jjs ¼ ;
function, and wj(t) is a zero mean Gaussian white noise fvgTsj fwgsj
excitation. Then, the acceleration process aj(t) can be
xdr xdr
represented as 2 tanð tÞð1 þ fr xr Þ þ xdr tan2 ð tÞ xdr
cjrs ðtÞ ¼ 2 2
 ;
Z t ðk2sj þ 2ksj fr xr þ f2r x2r þ x2dr Þð1 þ tan2 xdr
2
t Þ
aj ðtÞ ¼ hfj ðt sÞej ðsÞwj ðsÞds; ð13Þ
0
and
where xdr
d jrs ¼ : ð20Þ
k2sj þ 2ksj fr xr þ f2r x2r þ x2dr
X
nF
fbgTj fwgsj fvgTsj fbgj
hfj ðtÞ ¼ eksj t ð14Þ Eqs. (19) and (20) provide an explicit expression for
s¼1 fvgTsj fwgsj
the impulse response functions hij ð; Þj¼1;...;nf . These
is the unit impulse response function for the process aj(t) functions are given in terms of the spectral properties
at time t due to a unit impulse applied at time t = 0. The of the structural system and the eigenvalues and eigen-
vectors {w}sj and {v}sj, s = 1, . . . , nF are the complex right vectors of the filters state equation.
and left eigenvectors associated with the right and left
eigenproblems of Eq. (12), and ksj, s = 1, . . . , nF are the
corresponding eigenvalues. Using the representation of 5. Excursion probability
the process aj(t) given by Eq. (14), the response function
ri(t) can be written as In practical applications, the input–output relation-
nf Z t
ship given by Eq. (15) is often written in a discrete man-
X ner. In that case, the response is evaluated at discrete
ri ðtÞ ¼ hij ðt sÞaj ðsÞds
j¼1 0 time steps by some numerical integration scheme using
XZnf t Z s the values of the signal input at the sampled time in-
¼ hij ðt sÞhfj ðs zÞej ðzÞwj ðzÞdz ds stants [13]. Let the sampling be uniform at time spacing
j¼1 0 0 Dt = T/(nT 1) where T is the duration of the excitation,
nf Z and nT the number of time points so that the sampling
X t
¼ hij ðt; sÞej ðsÞwj ðsÞds; ð15Þ times are tk = (k 1)Dt, k = 1, . . . , nT. Then, the dis-
j¼1 0
crete-time analog of the input–output relationship in
where Eq. (15) can be written as
Z t nf
X X
k
hij ðt; sÞ ¼ hij ðt zÞhfj ðz sÞdz ð16Þ ri ðtk Þ ¼ Dt l hij ðtk tl Þej ðtl Þwj ðtl Þ; ð21Þ
s
j¼1 l¼1
is the unit impulse response function for the system re-
sponse function ri(t) at time t due to a unit impulse ap- where l is a coefficient that depends on the particular
plied at the jth filter at time s. For the seismic load, the numerical integration scheme used in the evaluation of
impulse response function hij( Æ ) of the system response Eq. (15). If wj ( Æ ) is modelled as a band-limited Gaussian
ri(t), given by Eq. (11), becomes white noise process ffi with spectral density Sj, that is,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
wj ðtk Þ ¼ 2pS j =Dtnkj , where nkj ; k ¼ 1; . . . ; nT are inde-
X
m
1 pendent, identically distributed standard Gaussian ran-
fr x r t
hij ðtÞ ¼ aijr e sinðxdr tÞ; ð17Þ
r¼1
xdr dom variables, Eq. (21) becomes
nf
X X
k pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where ri ðtk Þ ¼ l hij ðtk tl Þej ðtl Þ 2pS j Dtnlj : ð22Þ
j¼1 l¼1
fcgTi f/gr f/gTr ½M fgj g
aijr ¼ : ð18Þ Hence, the excursion problem corresponding to crite-
f/gTr ½M f/gr
rion q can be formulated as
!
Integration of Eq. (16) yields [nq [nT
q q q q
P Fq ¼ P fri ðtk Þ P ri ðtk Þ _ ri ðtk Þ 6 ri ðtk Þg
hij ðt; sÞ ¼ hij ðt sÞ i¼1 k¼1
!
Xm X nF
 [
nq [
nT
fr xr ðt sÞ
¼ aijr jjs cjrs ðt sÞe þ d jrs eksj ðt sÞ ; ¼P F ik ;
r¼1 s¼1 i¼1 k¼1

ð19Þ ð23Þ
1066 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075

where the Gaussian response processes rqi ðÞ; i ¼ 1; . . . ; tion allows to describe any kind of Gaussian excitation
nq ; q ¼ 1; . . . ; nc are specified by an expression similar including filtered white noise, colored excitation, non-
to Eq. (15), and Fik is the elementary failure event that zero mean excitation, and non-stationary excitation.
the response rqi ðÞ at time tk exceeds the upper bound An input–output relationship similar to Eq. (22) can
rqi ðtk Þ or falls bellow the lower bound rqi ðtk Þ, that is, be derived in this case, and therefore the implementation
F ik ¼ frqi ðtk Þ P rqi ðtk Þ _ rqi ðtk Þ 6 rqi ðtk Þg. of the procedure to more general cases is direct.
From the last equation, it is clear that the probabil-
ity of failure corresponding to criterion q is given as the
probability of the union of a number of elementary fail-
ure elements. The elementary failure regions are defined 6. Approximation concepts
as the region in the random variables space which cause
a barrier crossing at instant tk due to the ith response The computation of the reliability measures has to be
function. Then, it is seen that the evaluation of PFq cor- carried out repeatedly during the optimization process.
responds to a reliability problem of a series system of For every change of the design variables {h}, due to
nq · nT failure elements. In this case, the boundary of directives of the optimizer, the reliability measures, that
the excursion domain consists of 2 · nq · nT plain is, PFq({h}), q = 1, . . . , nc have to be updated. From the
hyper surfaces. Efficient methodologies for solving first previous formulation, it is clear that the evaluation of
excursion problems are considered in this formulation these measures is completely determined by the impulse
[8,10]. The methods are adopted and integrated into response function of the system responses. For systems
the proposed optimal design process. The work re- of practical interest, the repeated evaluation of the im-
ported in [8] is based on the importance sampling tech- pulse response functions can be extremely time-consum-
nique. Using the linear relations between input and ing. Hence, in order to increase the computational
response, the design points (the nearest point to the ori- efficiency of the implementation, a local approximation
gin in the standard normal space) can be established in strategy is introduced here.
a straightforward manner which define uniquely the ele-
mentary failure regions Fik, i = 1, . . . , nq, k = 1, . . . , nT. 6.1. Local approximation of system responses
The elementary failure probabilities are known exactly
in this case, and therefore the optimal importance sam- It is seen from Eqs. (17) and (18) that the impulse
pling density for the crossing problem is available. To response function hij( Æ ) depends on the mode shapes
compute the first excursion probability, a weighted {/}r, r = 1, . . . , m and the natural frequencies xr,
sum of the elementary optimal importance sampling r = 1, . . . , m. These quantities are implicit functions of
densities is used in the approach. The details of the pro- the vector of design variables {h}, and they are avail-
cedure as well as the evaluation of its efficiency can be able only in a numerical way, for example, by means
found in reference [8]. On the other hand, the metho- of a finite element model. In this work, the mode fac-
dology presented in [10] is based on the concept of tors aijr and the natural frequencies of the structural
averaged excursion probability flow which allows to ac- system are locally approximated by using a convex lin-
count for the interaction of excursions at different in- earization [16]. The approximations are constructed in
stances in an efficient manner. The approach permits terms of intermediate design variables f#gð#i ; i ¼
to consider only a fixed time instead of a time interval nd Þ. These intermediate design variables are
1; . . . ; 
which reduces the complexity of the excursion problem introduced in order to enhance the quality of the
to a static one. Numerical results have shown that approximations [17,18]. In this context, it is assumed
excursion probabilities of linear systems subjected to that the intermediate design variables {#} are explicit
general Gaussian excitation are computed with high functions of the actual design variables {h}. For exam-
efficiency. ple, if the design variables are cross-sectional dimen-
Finally, it is noted that the previous formulation is sions, cross-sectional properties such as areas and
not restricted to filtered white noise processes. For gen- moments of inertia can be chosen as intermediate de-
eral discrete Gaussian second-order stochastic processes, sign variables. The approximation for the mode fac-
the Karhunen–Loéve (K–L) expansion can be used to tors and frequencies about a point f# gð#i ; i ¼
describe such processes [14,15]. In that case, the K–L nd Þ takes the form
1; . . . ; 
representation of each component aj(t) of the accelera-
tion vector can be written as aj ðtÞ ¼ a0j ðtÞ þ Xnd
oaijr ðf# gÞ
P na l l 0 l aijr ðf#gÞ ¼ aijr ðf# gÞ þ
~ Baijr ð#i Þ; ð24Þ
l¼1 aj ðtÞnj , where aj ðtÞ and aj ðtÞ denote the mean func- o#i
i¼1
tion and the jth K–L vector of the process aj(t), respec-
tively, nlj ; l ¼ 1; . . . ; na are independent, identically Xnd
oxr ðf# gÞ
distributed standard Gaussian random variables, and ~ r ðf#gÞ ¼ xr ðf# gÞ þ
x Bxr ð#i Þ; ð25Þ
na is the order of the series expansion. This representa- i¼1
o#i
 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075 1067

where tion techniques, such as the ones reported in references

8 ij [8,10], to approximate the excursion probabilities in
< #i #i ; if oar ðf# gÞ
P 0;
o#i the vicinity of the current design point, that is,
Baijr ð#i Þ ¼ ð26Þ
: ð#i Þ2 ð#1i 1
Þ; if
ij
oar ðf# gÞ
< 0; Pe Fq ðfhgÞ; q ¼ 1; . . . ; nc . Such approximations require
#i o#i
only one dynamic analysis.
and
8
< #i
#i ; if oxr ðf# gÞ
P 0; 6.2. Local approximation of excursion probabilities
o#i
Bxr ð#i Þ ¼   ð27Þ
: ð# Þ2 1 1
; if oxr ðf# gÞ
< 0; The computational efficiency of the optimization pro-
i #i #i o#i
cess is further increased by approximating the excursion
where {#*} = {#({h*})}, and {h*} is a point in the de- probabilities locally. The objective of these approxima-
sign space. The mode factors and frequencies are tions is to reduce the computational effort due to re-
approximated as a linear function of the intermediate peated excursion probability estimates and sensitivity
design variable #i if the partial derivative is positive. analyses with respect to the design variables during the
On the other hand, they are approximated as a linear optimization process. In the proposed formulation, the
function of the reciprocal of the intermediate design var- excursion probabilities are represented using approxi-
iable #i if the partial derivative is negative. In general, mate functions dependent on the design variables. The
convex linearization generates high quality first-order advantage of this representation is that during the opti-
approximations and it is frequently used in structural mization procedure the excursion probabilities can be
optimization. The analytical approximations of aijr ; r ¼ approximated by evaluating the explicit functions. The
1; . . . ; m, and xr, r = 1, . . . , m in terms of the actual de- excursion probability PFq({h}) is approximated locally
sign variables {h} are obtained directly since the inter- about a point {h*} as
mediate design variables #i ; i ¼ 1; . . . ; nd are explicit
functions of the variables hi, i = 1, . . . , nd. The evaluation P Fq ðfhgÞ ¼ expAq ðfhg;fh gÞ ; ð31Þ
of the partial derivatives used in the approximations
requires a classical eigenvalue-eigenvector sensitivity where Aq({h},{h*}) is a multivariable polynomial func-
analysis at the point {h*}. An efficient procedure for tion, and {h*} is a point in the design space. A simple
determining eigenvector derivatives is used in this imple- linear expansion for the multivariable polynomial func-
mentation [19]. In that procedure, the calculation of the tion is used in the implementation, that is Aq({h},
derivatives of a given eigenvector requires only eigen- {h*}) = a0 + {a}T({h} {h*}). The parameter a0 and
data associated with that eigenvector. The simplified the components of {a} are the polynomial coefficients
procedure is much more efficient than standard methods which have to be determined by solving a linear equa-
in which eigenvector derivatives are expressed in terms tion system. The system of equations is assembled by
of all eigenvectors. estimating excursion probabilities for specific design
Introducing the earlier approximations of the mode variables combinations in the vicinity of {h*}, and inter-
factors and frequencies in Eq. (17), yields an explicit polating the polynomial function Aq({h},{h*}). The
local approximation about the design point {h*}, for coefficient a0 is computed as a0 ¼ Lnð Pe Fq ðfhj gÞÞ, and
the unit impulse response function hij in terms of {h} as the other polynomial coefficients are obtained by mini-
X
n mizing the square of the error term LnðP Fq ðfhj gÞÞ
1
~
hij ðt; fhgÞ ¼ ~aijr e fr x
~rt
~ dr tÞ:
sinðx ð28Þ Lnð Pe Fq ðfhj gÞÞ; j ¼ 1; . . . ; nI , where {hj}, j = 1, . . . , nI are
~ dr
x
r¼1 appropriately selected interpolation points. The number
Then, the discrete input–output relationship in Eq. of interpolation points nI depends on the number of
(22) can be written as considered design variables. In order to minimize the
number of excursion probability estimates, only nd inter-
nf
X X
k pffiffiffiffiffiffiffiffiffiffiffiffiffiffi polation points are considered in the implementation.
~ri ðtk ; fhgÞ ¼ l ~
hij ðtk tl ; fhgÞej ðtl Þ 2pS j Dtnlj ; For the current purpose, the above approximation stra-
j¼1 l¼1
tegy provides sufficiently accurate results. It is noted
ð29Þ that the approximate excursion probability function
where Pe Fq ðfhj gÞ is evaluated using local approximations for
Z t the system responses, and therefore, the characterization
~
hij ðt s; fhgÞ ¼ ~hij ðt z; fhgÞhfj ðz sÞdz ð30Þ of P Fq ðfhj gÞ in the vicinity of {h*} requires just one dy-
s namic analysis. Approximations of excursion probabili-
represents the approximation of hij . Note that Eq. (29) ties in terms of exponential functions as well as other
provides an explicit approximation of the response pro- function types have already been used in the context
cess ri in terms of the design variables {h}. These of reliability based optimization problems [20,21]. The
approximations can be used in combination with simula- functions were used to represent the probability of
1068 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075

failure over the complete design space. For the class of Convergence
Optimization Process Control Stop
problems considered here, it is not possible to represent
with sufficient accuracy the behavior of the excursion Improved
probabilities over the entire design space. Therefore, Approximate
Trial Design Explicit Problem Design
Problem
the approximations are locally adjusted during the opti-
mization process in this work.
Approximate Problem Optimization
Generator Algorithm
Dynamic Analysis
Firstorder scheme
7. Optimization Sensitivity Analysis (NLPQL algorithm)
Construct explicit
response approximations
The solution of the optimization problem (2) pro-
ceeds by transforming it into a sequence of approximate First excursion estimates

subproblems having a simple explicit algebraic structure. Construct explicit excursion

probability functions
The input data required for the characterization of the
subproblems are the natural frequencies and mode
shapes of the structural system as well as their deriva- Fig. 1. Flow chart of the optimization process.
tives at the current design point. Using this information
together with the approximation concepts discussed ear- which adjust the move limits for each design variable
lier, the following subproblem is generated at each design cycle, have shown to be robust and they
are frequently used in the area of structural optimiza-
Minimize cðfhgÞ
tion. The new design point, produced at the end of the
subject to the design constraints optimization stage, is then analyzed at the beginning
of the next stage. This procedure is continued until the
P Fq ðfhgÞ 6 P Fq ; q ¼ 1; . . . ; nc fhg 2 H; ð32Þ
objective function changes less than some prescribed
where H is the set that contains the side constraints for amount for three consecutive stages. The corresponding
the design variables, in the vicinity of the current design flow chart of the design process is shown in Fig. 1.
point, where the approximation of the excursion proba-
bilities are expected to yield reasonable results. The ac- 8. Sensitivity analysis
tual optimization procedure applied to the subproblem
is executed without any dynamic response calculation The evaluation of how robust is the solution to
and reliability analysis, respectively. Therefore, the solu- changes in the various model parameters provides
tion of the subproblems can be obtained in an efficient valuable information into the nature of the so far final
manner. In principle, the optimization problem may be design. This information is important since the optimi-
solved by applying any optimization algorithm where zation process tends to push the system towards a design
the NLPQL algorithm [22] proved to be robust and effi- where one or more constraints are active. In this man-
cient for the cases considered here. ner, a design which even slight variability in the problem
As indicated earlier, the solution of the original prob- formulation could produce failure, is obtained. The sen-
lem is replaced by the solution of a sequence of approx- sitivity of the final design is evaluated by considering its
imate subproblems. The process is summarized as behavior when the system parameters vary within a
follows. At the beginning of each design stage, the sys- given region of the parameter space. The proposed for-
tem is analyzed to construct approximate optimization mulation allows to perform an efficient sensitivity ana-
problems. These problems are nonlinear but explicit, lysis with respect to the dependence of the final design
and therefore, they are inexpensive to solve compared on the variation of the final design formulation. For
with the solution of the actual problem. Each approxi- example, using the characterization of the excursion
mate optimization problem is solved by a sequential probability function given by Eq. (31), the behavior of
quadratic programming technique (NLPQL algorithm). the system reliability with respect to the design variables
The side constraints or move limits for the design vari- can be studied without performing a reanalysis. Simi-
ables in the approximate problems are changed at each larly, the behavior of the system reliability with respect
design cycle in order to protect the quality of the to other system parameters can be evaluated in exactly
approximations. In general, the size of the move limits the same way as the behavior with respect to the optimi-
depends on the problem and the quality of the approxi- zation variables. In that case, the excursion probability
mations. The more accurate approximations the larger functions are approximated in terms of the system
the move limits. Move limit strategies based on con- parameters by using approximate functions dependent
straint violations at the end of each design cycle, and on such parameters as
constraint and objective functions sensitivities have been p
suggested in the literature [23,24]. These strategies, P Fq ðfpgÞ ¼ expAq ðfpg;fp gÞ ; q ¼ 1; . . . ; nc ; ð33Þ
 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075 1069

where {p}, (pj, j = 1, . . . , np) denotes the vector of system 9. Example problem
parameters, fp g; ðpj ; j ¼ 1; . . . ; np Þ denotes the base or
nominal value of the vector of system parameters, and 9.1. Description
Apq ð; Þ is a multivariable polynomial function. As be-
fore, the construction of these approximations requires The 26-story reinforced concrete building shown in
just one dynamic analysis. In this manner, a parameter Figs. 2 and 3 is considered as an example problem.
study of the final design is carried out without any con- The structural model consists of a set of plane frames
siderable increase of the computational effort. The value connected by infinitely rigid diaphragms at story levels.
of the excursion probability corresponding to a given set Each of the 26 floors is supported by 24 columns of
of parameters is obtained directly and efficiently by eva- square cross section. The height of the first floor is
luating the approximate excursion probability function. 5.35 m and the height of rest of the floors is 3.10 m, lead-
After these evaluations are performed, it can be seen, for ing to a total height of 82.85 m. The objective of the
example, which active constraints have the largest viola- problem is to minimize the total volume of the column
tion and how close the inactive constraints are from elements. The properties of the reinforced concrete are
being active when system parameter variabilities are
considered. Also, acceptable ranges of variation of the
system parameters, given specific reliability tolerances, θ
can be determined directly from the expression of the θ
approximate excursion probability functions. 33.6m
The behavior of the excursion probability function
with respect to a system parameter pj at the final design

6.5m
can also be quantified by its dispersion about the nomi-
nal value through the following sensitivity metric
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
R ffi
1
ðP Fq ðpj Þ P Fq Þ2 dpj p y

6.6m
20.4m

lp j p j j x
cP Fq ;pj ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
R ffi ; ð34Þ
1
ðp p Þ2 dp P Fq
lp j pj j j j

6.5m
where lpj is the measure of the range of variation of the
system parameter pj, pj is the nominal value of the sys-
tem parameter, and P Fq is the nominal value of the 6.5m 6.6m 6.6m 6.6m 6.5m
excursion probability at the final design. In this context,
P Fq corresponds to the excursion probability when the Fig. 2. 26-story building: floor plan.
value of the system parameter is equal to its base value,
i.e., pj ¼ pj . The coefficient of sensitivity defined in Eq.
+82.85m
(34) can be considered as a generalization of the sensiti-
vity measure called elasticity, which is usually used in
standard sensitivity analyses. This coefficient is evalu-
ated numerically by using the characterization of the
approximate excursion probability P Fq ðfpj gÞ. The sensi-
tivity metric can identify the more critical system para-
meters and the less influential parameters at the final
design. This information can also be used to determine
25 × 3.1 = 77.5m

whether or not system parameter uncertainties should

be considered in the optimization process. If the sensiti-
vity of the final design with respect to an uncertain sys-
tem parameter is low, then the parameter can be fixed as
deterministic with the median as a measure. On the con-
trary, if the sensitivity is high, an unconditional design
process should be considered [25]. These measures can
also be used to determine the degree of functional cou-
pling in the set of system parameters with respect to dif-
+5.35m
ferent reliability constraints. With all this information, a
5.35m

deeper insight into the so far final design is obtained +0.00m

including information for further decisions concerning

additional analyses, if any. Fig. 3. 26-story building: elevation view.
1070 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075

set as follows: Youngs modulus E = 3.0 · 1010 N/m2, and hence the number of input random variables in
Poissons ratio m = 0.3, and mass density q = 2500 the discrete representation of the excitation in each
kg/m3. A classical damping is assumed in the model so direction is thus nT = 301.
that the first modes have 5% of critical damping. The To control service ability and minor damage, the de-
building is excited horizontally by earthquake excitation sign criterion considered in this example problem is de-
ax(t) and ay(t) which is assumed to act independently in fined in terms of the interstory drift ratio in the x and
the x and y direction, respectively. The stochastic ground y direction over all stories of the building, and the over-
acceleration in each direction is modelled by filtered all building drift at the roof level in the x and y direction.
white noise and it is characterized as in Eq. (12) with Therefore, the number of response functions involved in
0 1 the problem is 54 (27 in each direction), and the system
0 1 0 0
failure event F is defined as
B x2 2f1g x1g 0 0 C
B 1g C
½B ¼ B C; [
54 [
301 54 [
[ 301
B 0 0 0 1 C F ¼ fri ðtk Þ P ri ðtk Þ _ ri ðtk Þ 6 ri ðtk Þg ¼ F ik ;
@ A
i¼1 k¼1 i¼1 k¼1
2 2
x1g 2f1g x1g x2g 2f2g x2g
0 1 0 1 ð36Þ
2
0 x1g
B C B C where the upper and lower bound values ri ðtk Þ and ri(tk)
B eðtÞ C B 2f x C ð35Þ are assumed to be of the same magnitude, and they are
fbg ¼ B C; fbgT ¼ B 1g 1g C;
B 0 C B x22g C chosen to be 0.2% of the story height for the interstory
@ A @ A
drift ratio, and 0.1% of the building height for the over-
0 2f2g x2g
all drift. It is noted that these drift levels are related to
8 2
low level vibration where the structural response is ex-
< ðt=4Þ ;
> if t 6 4s
eðtÞ ¼ 1; pected to be dominated by linear elastic behavior. There-
if 4 6 t 6 10s
>
: ðt 10Þ2 fore, the use of the linear model considered in this work
e ; if 10 6 t 6 15s is adequate for the current example problem. The design
For the numerical example the values x1g = 15.6 constraint for the optimization problem is written as
rad/s, f1g = 0.6, x2g = 1.0 rad/s, and f2g = 0.9 have been P F 6 P F , where the target failure probability is taken
used. The frequencies x1g and x2g correspond to the equal to 10 3. The design variables are the dimensions
dominant and lower-cutoff frequency of the spectrum, of the square cross section of the column elements. Over
respectively, and f1g and f2g are the damping parameters the height of the building, the columns have seven differ-
associated with the dominant and lower-cutoff fre- ent cross-sections with initial design hi = 0.80 m,
quency, respectively. The white noise intensity S has i = 1, . . . , 7, and side constraints 0.50 m 6 hi 6 1.50 m.
been assumed to be 1.0 · 10 3 m2/s3. The corresponding This design is infeasible with a probability of failure
power spectral density of the stochastic ground acceler- equal to 200 · 10 3. The definition of the design vari-
ation is shown in Fig. 4. The sampling interval is as- ables over the height of the building is shown in Table
sumed to be Dt = 0.05 s and the duration of the 1. Column elements section properties (moments of iner-
excitation is T = 15 s. The total number of time points, tia) are taken as intermediate design variables during the
optimization process. For the dynamic analysis, 9 out of
the 78 classical modes are retained in the calculation.
The contribution of higher modes in the evaluation
of the response functions was negligible in this case.
The simplified model considered for the building is
now used in the optimal design process. It is noted that
no attempt has been made to consider a more realistic

Table 1
Design variables—linking detail
Optimization variable Design elements (floors)
h1 1–2
h2 3–5
h3 6–8
h4 9–12
h5 13–16
h6 17–21
Fig. 4. Power spectral density of the stochastic ground
h7 22–26
acceleration.
 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075 1071

representation of the structural system, because the

focus of the example problem is not on the representa-
tion of the structure, but to evaluate the performance
of the proposed methodology.

9.2. Optimization results

The degree of accuracy predicted by the local approx-

imation of the system responses is illustrated in Fig. 5.
This figure shows the natural frequency and the effective
mass of the first lower mode of the structure at the initial
design as a function of the intermediate design variable
I1 (moment of inertia). A parameter variability of 20
% with respect to its nominal value is considered in this
Fig. 6. Excursion probability function of the initial design as a
figure. The results of the proposed local approximations function of the intermediate design variables I1 and I2. Exact
are compared with those obtained using the exact re- and approximate responses.
sponse. It is found that approximate quantities fit very
well the exact variability over the range of variation
considered for the intermediate design variable. The The approximate excursion probability function fits
behavior of the other response functions involved in the exact variability of the probability of failure with
the problem is similar to these quantities in terms sufficient accuracy. It is noted that the accuracy of the
of the accuracy of the local approximations. Fig. 6 fit can be improved if a higher order expansion is used
shows the approximate excursion probability function in the exponent of Eq. (31). For example, if a quadratic
P F in the vicinity of the initial design, and it is depicted expansion is used, the approximate and exact excursion
in terms of the intermediate design variables I1 and I2. probabilities become undistinguishable in the vicinity of
the initial design in this case. However, for the current
purpose, the first-order expansion provides sufficiently
accurate results. It is emphasized that the construction
of the approximate excursion probability function re-
quires just one system dynamic analysis. The same re-
sults, regarding the accuracy of the approximations,
are found during the entire optimization process. From
these results and from further experience it is concluded
that the proposed local approximation strategy can be
applied with sufficient confidence during the optimiza-
tion process.
The final designs using direct optimization (exact)
and the proposed approach (approximate) are given in
Table 2. The direct solution corresponds to the final de-
sign obtained when the system dynamic analyses and
excursion probabilities are evaluated directly for every
change of the design variables during the optimization
process. It is seen that the approximate final design is al-
most identical to the direct or exact solution. The vol-
ume of the initial design, which is infeasible, is reduced
at the final design in about 20%. The iteration history
of the optimization process in terms of the reliability
constraint is shown in Fig. 7. It is observed that the en-
tire optimization process converges in just 10 design cy-
cles. This fast convergence leads to a small number of
dynamic analyses and excursion probability estimations
to be performed during the optimization process. It is
seen from Table 2 that the main difference between the
Fig. 5. Natural frequency (A) and effective mass (B) of the first approximate and direct solution is in the number of dy-
mode of the initial design as a function of the intermediate namic analyses required for convergence. The difference
design variable I1. Exact and approximate responses. is quite significant. Only 10 system dynamic analyses are
1072 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075

Table 2
Direct and approximate final designs
Design variable Initial design Direct Proposed
h1 (cm) 80.0 92.0 90.3
h2 (cm) 80.0 76.1 76.4
h3 (cm) 80.0 73.2 73.7
h4 (cm) 80.0 71.2 71.5
h5 (cm) 80.0 69.5 70.6
h6 (cm) 80.0 67.6 67.4
h7 (cm) 80.0 60.6 59.6
Objective function (m3) 1273 1022 1019
Number of excursion 1240 80
probability estimates
Number of dynamic 1240 10 Fig. 8. Relative importance of the response functions in
analyses contributing to the excursion probability of the initial design.

Fig. 7. Iteration history in terms of the excursion probability.

Fig. 9. Relative importance of the response functions in
contributing to the excursion probability of the final design.
required to obtain the final design in the approximate
case whereas 1240 analyses are required in the direct
case. This number is less than 1% of the total number Fig. 9 that the dominant response at the final design is
of analyses to be performed in the exact case. The effec- the overall building drift at the roof level. In this case,
tiveness of the proposed methodology in solving this the value of the excursion probability is obtained mainly
type of optimal design problems is evident from these from the contribution of the failure events related to the
results. overall building drift. This response function dominates
The relative importance of the response functions in the response of the system and therefore the building
contributing to the excursion probability at the initial drift response at the roof level can be considered to be
and final designs is presented in Figs. 8 and 9, respec- active at the final design.
tively. The relative importance of the response ri at time
tk is defined as the probability content of the elementary 9.3. Sensitivity
failure region Fik normalized by the probability content
of all elementary failure regions. It is observed from Fig. The sensitivity of the excursion probability with re-
8. that the interstory drift ratio response of the first floor spect to the optimization variables at the initial and final
is significantly more important that the other response designs is shown in Table 3. The sensitivity is measured
functions at the initial design. The value of the excursion by the sensitivity metric defined in Eq. (34) for a range of
probability is obtained almost exclusively from the con- variation of 5% of the optimization variables with re-
tribution of the failure events related to the interstory spect to their initial and final values, respectively. The
drift ratio of the first floor. Therefore, this response design variable h1 is the most significant at the initial de-
function can be considered to be active at the initial de- sign whereas the other design variables show a very little
sign. The interaction of the response functions changes influence on the excursion probability. Therefore, the
during the optimization process. In fact, it is seen from dimensions of the cross section of the column elements
 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075 1073

Table 3
Sensitivity of excursion probability
Design variable Sensitivity Normalized
measure sensitivity measure
Initial Final Initial Final
design design design design
h1 10.9 36.9 1.00 1.00
h2 0.17 4.96 0.016 0.13
h3 0.27 4.70 0.025 0.12
h4 0.51 5.93 0.047 0.16
h5 0.48 5.55 0.044 0.15
h6 0.48 6.79 0.044 0.18
h7 0.10 6.62 0.009 0.17

Fig. 10. Probability of failure of the final design as a function

corresponding to the first two levels show an important of the system parameter E.
influence on the excursion probability. This result is rea-
sonable because the interstory drift ratio of the first floor
is the active response function at the initial design and h1 level of variability of the corresponding system parame-
has a significant influence on this response. Contrarily to ters. For example, the interval showed in Fig. 10 illus-
the initial design, the design variables h2 to h7 show some trates the values of the modulus of elasticity for which
influence on the value of the excursion probability at the the value of the probability of failure of the final design
final design. This is to be expected because the active re- varies less than 10 % of its nominal value. Therefore, this
sponse function in this case is the overall building drift interval represents the acceptable range of variability of
at the roof level, and therefore the dimensions of the the global stiffness of the building for a 10 % tolerance in
cross section of the column elements corresponding to the failure probability.
all floors are important. This sensitivity information to- Next, the sensitivity of the excursion probability of
gether with the characterization of the approximate the final design with respect to the interstory stiffnesses
excursion probability can be used, for example, to per- ki, i = 1, . . . , 26 in the x and y direction is considered.
form a re-analysis of the system in the vicinity of the It is found that the interstory stiffness of the first floor
so far optimal design in a very efficient manner. is the most significant with respect to the probability
The variability of the excursion probability of the of failure. A coefficient of sensitivity equal to 3.5 is ob-
final design with respect to some system parameters is tained in this case for a parameter variability of 10 %
now investigated. Fig. 10 shows the excursion probab- with respect to its nominal value. This value is about five
ility function in terms of the modulus of elasticity E. A times greater than the values of the coefficients of sensi-
parameter variability of 20 % with respect to its nominal tivity corresponding to the other parameters, indicating
value is considered in the figure, and the values of the that this parameter has a significant influence on the
parameter and excursion probability are normalized by excursion probability of the final design. It is noted that,
their nominal values, respectively. The failure probabil- even though the interstory stiffness of the first floor
ity function is constructed using a quadratic expansion shows an important influence on the probability of fai-
in the exponent of the exponential function given by lure of the final design, the other interstory stiffnesses
Eq. (31). The probability of failure decreases as the stiff- also have some influence on the probability of failure.
ness property of the structure represented by the modu- This result is reasonable because, as stated before, the
lus of elasticity increases. This information can be useful final design is dominated by the overall building drift re-
to determine the degree of robustness of the final design sponse at the roof level, and therefore the value of the
with respect to the variability of the global stiffness of interstory stiffness of all floors have some importance
the structural system. For example, a parameter vari- in the determination of the excursion probability. This
ability of 5 % produces a variability of about 50 % in sensitivity information is useful to identify the para-
the value of the excursion probability of the final design. meters that should be determined with more accuracy
This value is ten times greater than the corresponding due to their impact on the overall behavior of the final
parameter variability. If this value is not acceptable, design. For example, an accurate description of the
the variability of the parameter should be considered parameters that determine the interstory stiffness of the
explicitly during the optimization process. The sensiti- first floor is required to ensure the validity of the final
vity information can also be used to implement the solu- design in this case. If such description is not possible,
tion of the inverse problem, that is, given an acceptable the variability of such parameters should be properly
tolerance of the probability of failure, determine the accounted for during the design process in order to
1074 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075

obtain a robust design. In that case, unconditional spect to selected system parameters can be obtained di-
excursion probabilities that account for the uncertainty rectly from the proposed formulation.
in the parameters should be computed. In general, the
unconditional quantities correspond to a multidimen-
sional probability integral that rarely, if ever, can be 10. Conclusions
integrated analytically. Methods such as asymptotic
expansion and accurate importance sampling techniques A methodology based on approximation concepts for
have been reported to be well suited for approximating efficient solution of reliability based optimization prob-
the value of this type of integrals [26–28]. It is empha- lems of linear systems subjected to stochastic loading
sized that the incorporation of unconditional quantities has been presented. The feasibility of the method is
during the optimization process increases the computa- based on the separation of system dynamic and reliabil-
tional effort significantly [25]. Then, it is seen that the ity analyses, and nonlinear programming techniques.
ability to quantify the influence of system parameters The reduction of the computational effort required for
variability is crucial in the determination of the degree reliability estimation and gradient calculations during
of robustness of the final design. the optimization procedure is crucial. This is achieved
The effect and interaction of system parameters vari- by application of approximation concepts which are
ability on the probability of failure of the final design used to approximate the system responses and the reli-
can also be illustrated by considering the behavior of ability measures during the optimization process. In
the excursion probability function in terms of such order to increase the accuracy of the approximations,
parameters. Fig. 11 shows some contours of the excur- the concept of intermediate design variables is used in
sion probability function in the stiffness-damping space, the proposed implementation. The use of approximation
where the values of the parameters are normalized by concepts dramatically reduces the number of exact dy-
their nominal values. In this case the stiffness of the namic analyses as well as reliability estimations required
structure is represented by the modulus of elasticity for convergence. Thus, this technique is expected to be
and the damping of the system is represented by the useful in the optimization of large and complex struc-
modal damping ratio. The figure also shows the accept- tural systems subjected to stochastic excitation. On the
able range of variation of the stiffness and damping other hand, an efficient sensitivity analysis with respect
parameters in the vicinity of the final design if a target to the optimization variables and general system para-
failure probability P F ¼ 10 3 is considered. The effect meters becomes possible with the proposed formulation.
of the variability of these parameters and their interac- The sensitivity is evaluated globally by considering the
tion is clear from the figure. Similar sensitivity analyses behavior of the design when the parameters vary within
can be performed with respect to other system parame- a bounded region. The analysis can identify the degree
ters. In this manner, valuable information regarding of robustness of the final design with respect to varia-
the sensitivity and robustness of the final design with re- tions of selected system parameters. This information
can be used to determine whether or not system param-
eters uncertainty should be considered explicitly in the
optimization process. The sensitivity information pro-
vides a deeper insight into the optimal design and it
can be used as a basis for decision making.

Acknowledgement

This research has been partially supported by CON-

ICYT under grant number 1030375 which is gratefully
acknowledged by the author.

References

[1] Rice OC. Mathematical analysis of random noise. Bell Syst

Tech J 1945;24:46–156.
[2] Crandall SH, Chandirmani KL, Cook RG. Some first
passage problems in random vibration. J Appl Mech,
ASME 1966;33:624–39.
[3] Vanmarcke EH. On the distribution of the first passage
Fig. 11. Contours of the excursion probability function at the time for normal stationary random processes. J Appl
final design in the stiffness-damping space. Mech, ASME 1975;42:215–20.
 H.A. Jensen / Computers and Structures 83 (2005) 1062–1075 1075

[4] Mason AB, Iwan WD. An approach to the 1st passage [16] Fleury C, Braibant V. Structural optimization: a new dual
problem in random vibration. J Appl Mech 1983;50:641–6. method using mixed variables. Int J Numer Methods Eng
[5] Langley RS. 1st passage approximation for normal 1986;23(3):409–28.
stationary random processes. J Sound Vib 1988;122(2): [17] Schmit LA. Structural synthesis—its genesis and develop-
261–75. ment. AIAA J 1981;19(8):1249–63.
[6] Lin YK, Cai GQ. Probabilistic structural dynamic: [18] Thomas HL, Sepulveda AE, Schmit LA. Improved
advance theory and applications. New York: McGraw- approximations for control augmented structural optimi-
Hill; 1995. zation. AIAA J 1992;30:171–9.
[7] Pradlwater HJ, Schueller GI. Assessment of low probabil- [19] Nelson RB. Simplified calculation of eigenvector deriva-
ity events of dynamical systems by controlled monte carlo tives. AIAA J 1976;14:1201–5.
simulation. Probab Eng Mech 1999;14:213–27. [20] Kanda J, Ellingwood B. Formulation of load factors based
[8] Au SK, Beck JL. First excursion probabilities for linear on optimum reliability. Struct Safety 1991;9(3):197–210.
systems by very efficient importance sampling. Probab Eng [21] Gasser M, Schueller GI. Reliability-based optimization of
Mech 2001;16(3):193–207. structural systems. Math Methods Oper Res 1997;46:
[9] Au SK, Beck JL. Estimation of small failure probabilities 287–307.
in high dimensions by subset simulation. Probab Eng Mech [22] Schittkowski K. NLPQL: a Fortran subroutine solving
2001;16(4):263–77. constrained non-linear programming problems. Ann Oper
[10] Pradlwater HJ, Schueller GI. Excursion probabilities of Res 1985;5:485–500.
linear and nonlinear systems. In: Zhu GQ, Zhang RC, [23] Fadel GM, Cimtaly S. Automatic evaluation of move-
editors. Advances in stochastic structural dynamics. Boca limits in structural optimization. Struct Optimiz 1993;6(4):
Raton FL., USA: CRC Press; 2003. 233–7.
[11] Jensen HA. Structural optimization of linear dynamical [24] Bloebaun CL. Variable Move Limit Strategy for Efficient
systems under stochastic excitation: a moving reliability Optimization. In: Procs. of the AIAA/ASME/ASCE/AHS/
database approach. Journal on Computer Methods in ASC 32nd Structures, Structural Dynamics and Material
Applied Mechanics and Engineering special issue on Conference, Baltimore, Maryland, 1991.
Computational Stochastic Mechanics and Reliability Anal- [25] Jensen HA. Reliability-based optimization of uncertain
ysis 2005;194(12–16):1757–78. systems in structural dynamics. AIAA J 2002;40(4):731–8.
[12] Bathe KJ. Finite elements procedures. Upper Saddle [26] Breitung K. Asymptotic approximations for multinormal
River, New Jersey: Prentice-Hall; 1996. integrals. J Eng Mech, ASCE 1984;110(3):357–66.
[13] Burden RL, Faires JD. Numerical analysis. Boston: PWS [27] Schueller GI, Stix R. A critical appraisal of methods to
Publishing Company; 1993. determine failure probabilities. Struct Safety 1987;4(4):
[14] Loéve M. Probability theory. Princeton, New Jersey: D: 293–309.
Van Nostrand Company Inc.; 1963. [28] Papadimitriou C, Beck JL, Katafygiotis LS. Asymptotic
[15] Ghanem R, Spanos P. Stochastic finite elements: a spectral expansions for reliability and moments of uncertain
approach. New York: Springer; 1991. systems. J Eng Mech 1997;123(12):1219–29.