Available online at www.sciencedirect.

com

Advances in Engineering Software 39 (2008) 505–513

www.elsevier.com/locate/advengsoft

Structural reliability analysis using Monte Carlo simulation

and neural networks
João B. Cardoso *, João R. de Almeida, José M. Dias, Pedro G. Coelho
Faculty of Science and Technology, New University of Lisbon, 2829-516 Caparica, Portugal

Received 10 September 2003; received in revised form 13 March 2007; accepted 30 March 2007
Available online 22 May 2007

Abstract

This paper examines a methodology for computing the probability of structural failure by combining neural networks (NN) and
Monte Carlo simulation (MCS). MCS is a powerful tool, simple to implement and capable of solving a broad range of reliability prob-
lems. However, its use for evaluation of very low probabilities of failure implies a great number of structural analyses, which can become
excessively time consuming. The proposed methodology makes use of the capability of a NN to approximate a function for reproducing
structural behavior, allowing the computation of performance measures at a much lower cost. This approach seems very attractive, and
its main challenge lies in the ability of a NN to approximate accurately complex structural response. In order to assess the validity of this
methodology, a test function and two structural examples are presented and discussed. The second example is also used to show how this
methodology can be used to perform reliability-based structural optimization.
 2007 Elsevier Ltd. All rights reserved.

Keywords: Reliability-based optimization; Structural reliability; Monte Carlo simulation; Neural networks

1. Introduction • Approximate probabilistic or level 2 methods such as

The structural designer must verify, within a prescribed
safety level, the serviceability and ultimate conditions,
commonly expressed by the inequality: Sd < Rd, where Sd
represents the action effect and Rd the resistance. The
intrinsic random nature of material properties and actions
is actually considered by the Eurocodes [1,2], which classify
the methods available to deal with this randomness in three
levels:
• Semi-probabilistic or level 1 methods, the most used in
common practice, where the probability of failure is
indirectly considered through the definition of charac-
teristic values and the application of partial safety
indexes.
*
Corresponding author. Tel.: +351 212948567; fax: +351 212948531.
E-mail address: jbc@fct.unl.pt (J.B. Cardoso).
the first order or second order reliability methods
(FORM/SORM) where the probability of failure is
based on the reliability index b [3].
• Exact probabilistic or level 3 methods, where the prob-
ability of failure is computed from the joint probability
distribution of the random variables associated with the
actions and resistances.
Theory and methods for structural reliability have been
developed substantially in the last few years and they are
actually a useful tool for evaluating rationally the safety
of complex structures or structures with unusual designs.
Recent developments allow anticipating that their applica-
tion will gradually increase, even in the case of common
structures.
Monte Carlo simulation (MCS) is a simulation method
of level 3 that presents the following characteristics: it can
be applied to many practical problems, allowing the direct
consideration of any type of probability distribution for the

0965-9978/$ - see front matter  2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.advengsoft.2007.03.015
506 J.B. Cardoso et al. / Advances in Engineering Software 39 (2008) 505–513

random variables; it is able to compute the probability of
failure with the desired precision; it is easy to implement.
However, despite the advantages it presents, the use of
this method is not widespread in structural reliability
because it is not efficient when compared to level 2 meth-
ods. In fact, MCS requires a great number of structural
analyses, one for each sample of the set of random vari-
ables. The number of analyses needed to evaluate the prob-
ability of failure of a structure with a prescribed precision
depends on the order of magnitude of that probability.
As the values of the probability of failure associated to
the ultimate limit states vary normally between 104 and
106, the number of analyses to be performed for ensuring
a 95% likelihood that the actual probability be within 5%
of the computed one must be at least 1.6 · 107 to
1.6 · 109, according to Shooman [4].
These analyses are frequently performed with the help of
finite elements codes. Therefore, the computation time can
be prohibitively high, especially when the structure exhibits
non-linear behavior or the numerical model is rather
complex.
To eliminate this drawback, it is proposed here the use
of neural networks (NN) to approximate structural
response. Once properly trained, a NN allows the determi-
nation of the structural performances with a very small
number of operations and at a fraction of the cost of the
corresponding structural analysis. This methodology
allows the application of MCS to practical cases of great
complexity where the direct use of this method would not
be feasible.
To illustrate the techniques proposed, three examples
are presented, considering respectively a mathematical
function and two different structures. In the first example,
a non-linear analytical function is approximated by means
of NN. In the second, a linear elastic steel frame is designed
according to the Eurocodes 1 and 3 [1,2], and subsequently
the probability of failure of the frame is computed by a
Monte Carlo approach in which NN are used for reproduc-
ing the structural response. In the third, a genetic algorithm
is used in combination with NN approximation and Monte
Carlo method to perform reliability-based optimization of
a steel truss.
where (x1, x2, . . ., xn) are values of the random variables and
fX 1 ;X 2 ;...;X n ðx1 ; x2 ; . . . ; xn Þ is the joint probability density
function.
The Monte Carlo method allows the determination of
an estimate of the probability of failure, given by
1 XN
pf ¼ IðX 1 ; X 2 ; . . . ; X n Þ ð2Þ
N i¼1
where I(X1, X2, . . ., Xn) is a function defined by

1 if gðX 1 ; X 2 ; . . . ; X n Þ 6 0
IðX 1 ; X 2 ; . . . ; X n Þ ¼ ð3Þ
0 if gðX 1 ; X 2 ; . . . ; X n Þ > 0
According to (2), N independent sets of values
x1, x2, . . ., xn are obtained based on the probability distribu-
tion for each random variable and the failure function is
computed for each sample. Using MCS, an estimate of
the probability of structural failure is obtained by
NH
pf ¼ ð4Þ
N
where NH is the total number of cases where failure has
occurred.
3. Neural networks
NN are numerical algorithms inspired in the functioning
of biological neurons. This concept was introduced by
McCulloch and Pitts [6], who proposed a mathematical
model to simulate neuron behavior. Use of NN has become
widespread in several fields of engineering, such as struc-
tural mechanics [7,8] and structural reliability [9]. Papad-
rakakis et al. [10] presented an approach in which a NN
was associated to MCS in order to obtain the probability
of structural failure. A similar approach is proposed herein.
The model adopted is shown in Fig. 1, representing
the neuron m that receives an input signal vector
x = [xi, . . ., xL]T from L input channels. It then computes
the weighted sum of the components of x, multiplying each
component xk by a coefficient wmk reflecting the impor-
tance of the input channel k. The neuron m activation,
am, is given by the expression

2. Monte Carlo simulation X

L
am ¼ wmk xk þ bm ð5Þ
k¼1
A reliability problem is normally formulated using a
failure function, g(X1, X2, . . ., Xn), where X1, X2, . . ., Xn are
random variables. Violation of the limit state is defined x1
by the condition g(X1, X2, . . ., Xn) 6 0 and the probability wm1
of failure, pf, is expressed by the following expression [5]: x2
wm2
f (am) sm
pf ¼ P ½gðX 1 ; X 2 ; . . . ; X n Þ 6 0 x3 wm3
Z Z Z .
.
wmL
¼ ... fX 1 ;X 2 ;...;X n ðx1 ; x2 ; . . . ; xn Þ dx1 dx2 . . . dxn .

xL bm
gðX 1 ;X 2 ;...;X n Þ60

ð1Þ Fig. 1. Artificial neuron.

 J.B. Cardoso et al. / Advances in Engineering Software 39 (2008) 505–513 507

It should be noticed that Eq. (5) includes a corrective
term bm, called bias, allowing for the possibility of having
a non-negative activation am when all xk are zero.
The output signal of neuron m, sm, is the numerical
value that results from the computation of an activation
function, f(am). In the present work, the logistic function
is adopted. This function is obtained making a = 1 in the
expression of the sigmoid function, which has the following
form:
1 respectively, the computed and the exact value of the func-
sm ¼ f ðam Þ ¼
1 þ eaam
Several neurons can be assembled to form a network.
Different network architectures for such an assembly have
been studied, together with algorithms to perform the
determination of coefficients wmk and bm, cf. Rosenblatt
[11] and Rumelhart et al. [12]. In this work, it was decided
to adopt a well-known network arrangement, the multi-
layer perceptron [10], which is composed of several neurons
arranged in layers, as shown in Fig. 2.
Fig. 2 represents a NN with three layers: an input layer
composed by three neurons, in which no transformations
are performed; one hidden layer with four neurons; an out-
put layer with three neurons. It is possible to prove [13]
that this type of network with sigmoid activation functions
in the hidden and output layers can approximate any con-
tinuous function with satisfactory precision, provided that
it has enough neurons in the hidden layer.
The process of obtaining the unknown coefficients wmk
and bm required to approximate the prescribed function
is called training and it is a somewhat challenging task.
The most common training process, called supervised
training [12], consists of proposing some initial values to
the coefficients and then adjusting those values in order
to minimize the error between the predicted output pro-
duced by the NN and the exact value of the function.
Therefore, to perform this training, it is necessary to know
the exact value of the function for a prescribed set of input
values, named as the training set.
The error for the training set is evaluated using the fol-
lowing equation:
1X t
1X r
E¼ ðsij  oij Þ2 ð7Þ
t i¼1 r j¼1
where t is the number of elements of the training set, r is the
number of neurons in the output layer, sij and oij represent,
ð6Þ
tion for the j neuron and i element of the set.
Due to the network architecture, the error (7) to be min-
imized often presents several local and global minima.
Unlike gradient-based algorithms, which may be trapped
in local minima, genetic algorithms are generally successful
for performing multimodal optimization. In this work, a
hybrid algorithm is used, combining gradient-based and
genetic algorithms. Minimization with genetic algorithms
is performed at an initial phase and the solution obtained
is used subsequently as a starting point by a conjugate gra-
dient algorithm.
As the coefficients to be determined, wmk and bm, are
non-integer, the genetic algorithm is adapted to work with
this type of values. The genetic algorithm uses binary chro-
mosomes where each coefficient associated with hidden and
output layers is represented by eight genes, implying that it
can assume one of 28 = 256 different values. The real value
r, corresponding to any combination of the eight genes
expressed by the integer z, depends on the following trans-
formation between z and r:
 
rmax  rmin
r ¼ rmin þ z ð8Þ
28  1
where rmin and rmax are the lower and upper limits of r,
respectively. In order to perform the optimization of a
function of real variables, the strategy followed consists

f(a11)
w111
x1 w112 s12
= w121w 1 1 f(a12)
13 b1
w122
w211 w132 w 2
b12
1
f(a21) 14

w22
w212
x2 w231 w222 s22
= b2 1
f(a22)
w311 w232 2
1 w24
w32
f(a31) 2 b22
w331 w32w
2 31

x3 w332 s32
w421 w41
1
b31
= f(a32)
w342
w431
f(a41)
b32

b 41

Fig. 2. Multi-layer perceptron.

508 J.B. Cardoso et al. / Advances in Engineering Software 39 (2008) 505–513
of restarting successively the genetic algorithm, modifying
progressively the domain bounds rmin and rmax in order
to circumscribe the optimization to a closer vicinity of
the optimal solution obtained in the previous iteration.
4. Example 1: test function
To evaluate the capability of a NN to approximate accu-
rately a non-linear function, an analytical test function was
first considered:
F ðx; yÞ ¼ 0:3 þ ð2 sinðxÞ cosðyÞ þ sinðxyÞÞ=6
with x 2 [3.5, 5.5] and y 2 [2.0, 4.0]. The function is repre-
sented graphically in Fig. 3. It is observed that the test
function value in the prescribed domain is always within
the range of the logistic function [0, 1], and so normaliza-
tion is not needed. It may be assumed that non-linear struc-
tural response is generally described by functions simpler
than this one. Consequently, if a good approximation of
(9) is achieved through NN, this would suggest that this ap-
proach can be applied successfully to structural problems.
Several different NN were considered, keeping constant
the number of neurons in the input and output layers, s0
and s2, respectively equal to 2 and 1, and varying the num-
ber of neurons in the hidden layer, s1. It is worth noticing
that a greater number of elements of the training set
increases both the accuracy of the response and the training
time of the NN, so its size must result from a compromise
between time and quality of the approximation. For this
example, a set of 196 different values of the function were
obtained, considering a uniformly spaced grid of 14 points
for each variable, and then presented to the network for
training. After training, the performance of each network
was evaluated with the help of the following error measure:
 
sij  oij 
e ¼ max    for i ¼ 1; . . . ; t and j ¼ 1; . . . ; r ð10Þ
oij 
where all symbols have the same meaning as in Eq. (7).
Since a large number of elements in the test set does not
represent an inconvenient, a test set of 1024 function values
was considered, using a uniformly spaced grid of 32 points
Z structural response and MCS is performed for computing
Y the probability of failure.
X
3.5
4 3.5
4.5 3
5 2.5
5.5 2
Fig. 3. Test function F(x, y).
Table 1
Errors obtained with NN approximation of function F(x, y)
s1 Training error, E Max error with
test set, e (%)
1 1.55 · 102 114
6 7.15 · 104 32
12 9.22 · 105 26
18 5.49 · 106 2.5
24 1.79 · 106 1.9
30 1.19 · 106 1.5
36 1.97 · 107 0.93
ð9Þ
for each variable. Such large number of test points was
adopted in order to get a good assessment of the precision
achieved throughout the domain. The training error, com-
puted with Eq. (7), and the testing error, given by Eq. (10),
are presented in Table 1.
Fig. 4 presents graphically the approximation to the
function F(x, y), computed by each one of the first four net-
works considered in Table 1. It can be seen that, as the
number of hidden neurons increases, the approximation
gradually converges to the exact function. Given the fine
match achieved, the 14 · 14-point training set adopted
seems adequate. Hence, no further effort was made in order
to refine the training set.
5. Example 2: single story steel frame
As an example of the methodology proposed for com-
puting the probability of failure, the single bay steel frame
represented in Fig. 5 is considered and analysed using the
methodology currently adopted in Europe for structural
design. Applied actions (dead load, live load and wind)
are defined according to the Eurocode 1 [1], and the steel
structure is designed according to the Eurocode 3 [2]. The
design must satisfy safety conditions related to serviceabil-
ity and ultimate limit states. This is done according to a
level 1 methodology and the structural response is deter-
mined using a finite element code. After choosing the
appropriate rolled-steel shapes, safety conditions for ulti-
mate states are again checked, but now following a level
3 methodology in which NN are used to approximate the
5.1. Structural design according to Eurocodes (level 1)
The frame considered is part of an industrial warehouse
0.8 with dimensions 20 · 10 · 4 m. The structure of the ware-
0.6 house consists of five independent and identical frames
0.4
spaced 5 m apart from each other. The geometry of each
4 frame is shown in Fig. 5. Applied actions are defined
according to the Eurocode 1, and their characteristic values
are:
• Dead load – 0.5 kN/m2.
• Live load – 2 kN/m2.
 J.B. Cardoso et al. / Advances in Engineering Software 39 (2008) 505–513 509

Z Z

Y Y

X X

0.8 0.8
0.6 0.6
0.4 0.4

3.5 4 3.5 4

4 3.5 4 3.5

4.5 3 4.5 3

5 2.5 5 2.5

5.5 2 5.5 2

s1 =1 s1 =6

Z
Z
Y
Y

X
X

0.8 0.8
0.6 0.6
0.4 0.4

3.5 4 3.5 4
4 3.5 4 3.5
4.5 3 4.5 3
5 2.5 5 2.5
5.5 2 5.5 2

s1 = 12 s1 = 18

Fig. 4. Approximations of test function obtained with NN.

B C functions considered, according to the Eurocode 3. In this

table, NSd and My,Sd are the design axial force and bending
moment; A and Wpl,y are the cross-section area and the
L2 = 4 m
plastic bending modulus; fy is the nominal yield stress of
the material (steel S235); vz, vmin and vLT are reduction
coefficients for the following cases: out-of-plane buckling,
A D minimum between coefficients for in-plane and out-of-
plane buckling, and lateral torsional buckling.
L1= 10 m
For the cross-sections of the elements, standard rolled-
steel shapes are adopted. It can be concluded that a frame
Fig. 5. Steel frame. built with HEA 260 columns and a HEA 300 beam verifies
safety for all limit state functions. The corresponding dis-
placements and internal forces, obtained using a linear
• Wind – For the region where the structure is going to be
elastic finite element code developed by the authors, are:
built (Portuguese seashore), the prescribed dynamic
pressure of the wind is equal to 0.456 kN/m2. From this
pressure, the wind loading is obtained following the pro- Table 2
Limit state functions for steel frame
cedure prescribed in the Eurocode 1 [1].
Limit function Description
Load combination is performed according to the level 1 dH 6 L2/150 Maximum horizontal displacement of the
methodology followed in Eurocode 1, where the character- column
dV 6 L1/300 Maximum vertical displacement of the beam
istic values for the loads are multiplied by combination N Sd 1;5M y;Sd
vmin Afy =1;1 þ W pl;y fy =1;1 61 Resistance to bending and compression of the
coefficients (to reflect the non occurrence of simultaneous column
N Sd 1;5M y;Sd
loads) and by partial safety coefficients (to ensure, indi- vz Afy =1;1 þv
LT W pl;y fy =1;1
61 Resistance to bending and compression of the
rectly, the appropriate safety level for each limit state). beam, considering also lateral torsional
Table 2 shows the serviceability and the ultimate limit state buckling
510 J.B. Cardoso et al. / Advances in Engineering Software 39 (2008) 505–513
• dH = 1.32 mm.
• dV = 16.8 mm.
• NSd(COLUMN) = 98.6 kN (on top of right column – sec-
tion C).
• My,Sd(COLUMN) = 122.5 kN m (on top of right column –
section C).
• NSd(BEAM) = 46.7 kN (on right end of beam – section C).
• My,Sd(BEAM) = 122.5 kN m (on right end of beam – sec-
tion C).
5.2. Probabilistic model
Subsequently, the probability of failure of the frame
considered above is evaluated. For that purpose, statistic
models must be defined for each random variable involved.
Hence, normal distributions are assumed for the Young’s
modulus and dead load, whereas for the live load, wind
load and yield stress lognormal distributions are consid-
ered. The probabilistic parameters adopted for each
variable are chosen appropriately, according to the charac-
teristic percentiles prescribed in the Eurocode 1 and
making the resulting characteristic values coincide with
the previously used when level 1 design was carried out.
This way, both methodologies (levels 1 and 3) can be
compared.
The statistic parameters adopted for the random vari-
ables are shown in Table 3. In this table, the characteristic
values for live load and wind load are the percentile 98 of
the probability distributions used. With respect to the yield
stress, its characteristic value corresponds to the percentile
5. For the Young’s modulus and the dead load, their char-
acteristic values are considered equal to the mean values of
the respective probability distributions.
In order to determine the ultimate limit state functions,
it is necessary to know the values of the axial force, NSd,
and bending moment, My,Sd, both in the critical section
of the columns (top of the right column) and in the critical
section of the beam (right end). Since the maximum bend-
ing moments in the column and in the beam are identical,
only three internal forces are required for computing those
functions.
5.3. Neural networks
The NN that approximates the structural response has
input, hidden and output layers. The number of neurons
in each layer is designated by s0, s1 and s2, respectively.
The number of input neurons is equal to the number of
Table 3
Random variables for steel frame example
Variable Distribution Mean
Young’s modulus (GPa) Normal 210
Dead load (kN/m2) Normal 0.50
Live load (kN/m2) LogNormal 1.06
Wind pressure (kN/m2) LogNormal 0.241
Yield stress (MPa) LogNormal 280
random variables that influence structural response. As a
linear stress–strain relation is assumed in this example, only
the random loads and the Young’s modulus affect the
structural response (the yield stress is considered only in
the limit state functions) and so s0 = 4. The number of neu-
rons in the output layer is equal to the number of displace-
ments and internal forces needed to compute the limit state
functions, that is, s2 = 5.
A first set of NN was trained and subsequent Monte
Carlo simulations were performed. Results revealed that
the probability of failure for the serviceability limit states
was insignificant compared to the allowable one (a value
of 1 · 108 was obtained and the admissible Pf is of the
order of magnitude of 102). Therefore, the two service-
ability limit state functions considered could be disregarded
and the problem could be simplified. This was expected,
since the values of the horizontal and vertical displace-
ments, dH, dV, obtained in Section 5.1 were much lower
than their limiting values shown in Table 2. Knowing also
that the internal forces in a linear elastic structure made of
a single material do not depend on the material Young’s
modulus, this variable is removed from the probabilistic
model since it does not have any influence on the ultimate
limit state functions considered.
The simplified networks with s0 = s2 = 3 are trained
with a training set representative of the structural response.
More relevance is given to values situated in the upper
extreme of the distributions, because it is important to
obtain a precise approximation for those values, as they
influence greatly the probability of failure. A set of 216 ele-
ments is considered, combining six possible values for each
applied load, related to the mean, l, and the standard devi-
ation, r, of the probabilistic distribution for the load. The
values adopted are: l  2r, l  r, l + r, l + 2r, l + 3r,
l + 4.2r. In addition to these elements, the design point
previously considered in level 1 methodology is also used.
A test set containing 12 possible values for each load, total-
ing 123 = 1728 elements, is also created, to measure the
precision of the approximation obtained with the trained
networks.
Five networks were trained, each with a different number
of neurons in the hidden layer, s1. Table 4 presents the val-
ues of the quadratic mean error for the networks trained,
obtained by Eq. (7) and the time consumed in the process.
The values of the maximum relative error, computed for
the test set by Eq. (10), are presented in Table 5 for the five
networks. It can be observed that the NN approximate the
structural response with very good precision.
Standard deviation Coefficient of variation Characteristic value
10.5 0.05 210
0.05 0.10 0.50
0.366 0.35 2.0
0.084 0.35 0.456
28 0.10 235
 J.B. Cardoso et al. / Advances in Engineering Software 39 (2008) 505–513 511

Table 4 Table 7
Quadratic mean error and training time for NN (steel frame example) Probability of failure pf and reliability index b for steel frame
s1 – number of neurons Error obtained Error obtained Time (s) Limit MCS + NN MCS FORM SORM
in hidden layer with training set with test set function conventional
4 6.28 · 106 2.97 · 106 95 Column pf 1.277 · 105 1.255 · 105 1.229 · 105 1.285 · 105
6 3.85 · 107 3.44 · 107 342 failure b 4.21 4.21 4.22 4.21
8 1.01 · 107 8.27 · 108 890
10 3.20 · 109 2.20 · 109 1912 Beam pf 8.630 · 106 8.607 · 106 8.275 · 106 8.650 · 106
12 1.77 · 109 1.53 · 109 3959 failure b 4.30 4.30 4.31 4.30

Table 5
Maximum relative error for the test set obtained with NN (steel frame
adopted for the random variables are in general greater
example) than what is generally suggested by other authors [14,15].
s1 – number e (%)
If smaller coefficients of variation were used, this would
of neurons in certainly result in a smaller probability of failure.
NSd(COLUMN) NSd(BEAM) My,Sd(COLUMN,BEAM)
hidden layer For each limit state function considered, the corre-
4 1.42 2.14 1.11 sponding probability of failure was compared to that
6 0.771 1.11 0.711 obtained using conventional MCS (without NN), with
8 0.545 0.516 0.503 computation of internal forces by finite elements for each
10 0.105 0.107 0.071 sample of the set of random variables, and using 108 sam-
12 0.067 0.069 0.067
ples. These probabilities were also evaluated using FORM
and SORM methods, built in the program COMREL-TI
5.4. Results and discussion [16]. The results for all alternative techniques are presented
in Table 7 and show a very good agreement between the
Considering, for the probabilistic model presented, NN methodology proposed by the authors and the other proce-
with s1 = 12, several Monte Carlo simulations are per- dures considered.
formed. The corresponding values of the probability of For this example, the application of a conventional MC
failure and computation time are shown in Table 6. Failure approach performing 108 computer analyses lasted 6524 s,
is supposed to take place if at least one of the ultimate limit being much slower than the NN-based approach, as seen
state functions is violated. It should be noted that the MCS by the corresponding times shown in Table 6. However, it
enables to conclude that beam failure always occurs simul- is important to stress that the most substantial amount of
taneously coupled with column failure; hence the probabil- time involved in a NN-based procedure is usually spent in
ity of failure of the frame, pf, is equal to the probability of the training phase, as shown by the NN training and simu-
violating the ultimate limit state function associated with lation times for this example, given respectively in Tables 4
column failure. Denoting column failure by event C with and 6. Hence, a comparison based uniquely on the time con-
probability pf(C) and beam failure by event B with proba- sumed during the simulation phase can be misleading.
bility pf(B), it was detected that pf = pf(C [ B) = pf(C)
because pf(C \ B) = pf(B). The possibility of making this 6. Example 3: six element truss
type of system reliability analysis can be seen has a major
advantage of MCS [5]. The next structure considered is a six-bar simply sup-
The probability of failure of the structure, pf, is found to ported truss with random load and random material yield
be 1.277 · 105, and the corresponding reliability index b is stress, also examined in Murotsu et al. [17] and Burton and
equal to 4.21. It is therefore clear that the level 1 method- Hajela [18]. The initial geometry is shown in Fig. 6. The x
ology prescribed in the Eurocodes produces, in this case, a and y coordinates of node 4 always stay equal and are con-
design safer than the limit b = 3.8 recommended for com- sidered as the first design variable, nx. All the six-bars have
mon structures [1]. It is important to outline that the struc- the same cross-sectional area, and this area is the second
tural shapes were chosen considering a design point located design variable, A. Limits on these variables are specified
very close to the infeasible domain, and so this safety mar- in Table 8. The mass of the truss is optimized considering
gin cannot be justified by a conservative initial design. It two random variables, a load P applied at node 3 and the
should also be pointed out that the coefficients of variation material yield stress, Ty. The statistic parameters adopted
are shown in Table 9. A Young’s modulus E = 206 GPa
Table 6 and a specific weight q = 7800 kg/m3 are considered. Six fail-
Probability of failure of steel frame computed with MCS + NN consid- ure functions (one for each bar) are defined, gi = Ty  |Ti|,
ering s1 = 12
where Ti is the stress in bar i shown in Fig. 6. Structural col-
Sample size Probability of failure Time (s) lapse is supposed to take place when at least one bar fails.
6 5
10 1.233 · 10 4 The optimization is performed assuming that the probability
107 1.368 · 105 42 of failure should not exceed pf = 0.001, corresponding to a
108 1.277 · 105 424
reliability index b = 3.090, as was used in [18].
512 J.B. Cardoso et al. / Advances in Engineering Software 39 (2008) 505–513

P 3 5 4

3m 2 4

1 2
1

3m

Fig. 6. Six-bar truss.

The design space for the problem is shown in Fig. 7. To

obtain this figure, a mesh of 41 · 41 equally spaced values
of the design variables is considered, and for each point the
probability of failure of all six-bars is computed by MCS
and the truss mass is also calculated. Using graphic soft-
ware and through interpolation it is then possible to have
a discrete approximation for the six probability functions
and for the structural mass over the prescribed domain.
The isolines for the mass are represented by dashed lines
and the lines corresponding to the probability of failure
pf = 0.001 for each bar are represented by solid lines and
named according to the related failure function gi. It can
be seen that the design space is not convex, having both
a local minimum and a global minimum; hence classic opti-
mization methods experience difficulties to handle it, which
is not the case when genetic algorithms are adopted. How-
ever, such algorithms require the computation of the fitness
for a great number of points of the design space, and in this
case the probability of failure for each point must be com-
puted using Monte Carlo method. This can be an excessive
computational task unless neural networks are used to
approximate the response.
Fig. 7. Design space for the six-bar truss.
for each design variable and one for the random load P
(as before, the yield stress is considered only for the limit
state functions). The number of neurons in the output layer
is equal to the number of failure functions plus the mass,
but, as the stresses in elements 1 and 2 always remain equal
and the same happens in elements 4 and 5, only five neu-
rons are needed for the output layer.
A training set of 83 = 512 elements and a test set con-
taining 143 = 2744 elements are created. Four networks
were trained, and Table 10 presents the values of the qua-
dratic mean error obtained by Eq. (7). The values of the
maximum relative error, computed for the test set by Eq.
(10), are presented in Table 11. As before, it can be
observed that the NN approximate the structural response
with very good precision.

6.1. Neural networks 6.2. Mass optimization

The NN has a number of neurons in each layer desig- As the probability of failure is of order of magnitude
nated by s0, s1 and s2, respectively, for the input, hidden 103, a Monte Carlo simulation with 105 independent sets
and output layers. Three input neurons are needed, one is sufficient. A simple genetic algorithm [19] with 24 genes
for each chromosome (12 for each design variable) and a
Table 8 population of 48 chromosomes was used. Probabilities of
Design variables for the six-bar truss
crossover and mutation were taken, respectively, equal to
Design variable Lower limit Upper limit 0.8 and 0.008. After evolving for 48 generations, results
Node 4 coordinates, nx (m) 1.5 4.0 very close to [18] were obtained, as shown in Table 12.
Cross-sectional area, A (m2) 1.0 · 104 4.0 · 104

Table 10
Table 9 Quadratic mean error and training time for NN (steel frame example)
Random variables for the six-bar truss s1 – number of neurons Error obtained with Error obtained
Variable Distribution Mean Standard Coefficient of in hidden layer training set with test set
deviation variation 8 7.4621 · 106 4.4976 · 106
Load, P (kN) Normal 30 3 0.10 16 4.1077 · 107 2.9092 · 107
Yield stress, Ty Normal 172 8.6 0.05 24 5.3745 · 108 4.6667 · 108
(MPa) 32 1.7546 · 108 3.2032 · 108
 J.B. Cardoso et al. / Advances in Engineering Software 39 (2008) 505–513 513

Table 11 efficiently implemented in such scheme. Consequently, the

Maximum relative error for the test set obtained with NN (six-bar truss corresponding computing time was kept within very rea-
example)
sonable limits.
s1 – number of neurons e (%) From the research presented herein, it can be concluded
in hidden layer
eM e1,2 e3 e4,5 e6 that the application of NN in conjunction with MCS seems
8 1.84 4.71 6.22 5.58 4.58 very promising and appears to offer great potential for
16 0.67 2.19 1.58 1.80 1.35 structural reliability and optimization.
24 0.29 0.64 0.51 0.59 0.81
32 0.23 0.63 0.53 0.42 0.57
Acknowledgements

Table 12
Reliability-based optimization results for the six-bar truss example
Design variable/ GA + NN Burton and Difference (%)
objective function Hajela [18]
nx (m) 2.359 2.366 0.28
A (m2) 1.563 · 104 1.557 · 104 0.36
Mass (kg) 22.512 22.450 0.28
The number of Monte Carlo simulations performed was
482 = 2304, requiring a total computational time of 329 s
in a parallel cluster with six nodes composed of Pentium
3.0 GHz PC’s.
7. Conclusions
Whenever it is possible to define and model adequately
the random variables that influence structural behavior,
the use of level 3 reliability methods such as MCS allows
a more rigorous design approach than the semi-probabilis-
tic level 1 methods presented in the Eurocodes. However,
the amount of time required by conventional MCS makes
its application not practical.
In recent years, NN have become a successful means for
performing function approximation. An example in which
a non-linear function was approximated by NN confirmed
the accuracy of this technique.
The application of NN to structural problems, while
leading to satisfactory precision, enables to speed up signif-
icantly the computation of structural response. This feature
is particularly relevant in reliability analysis, where a very
substantial computing effort is normally required in order
to evaluate accurately the probability of failure.
In the second example considered (reliability analysis of
a single-storey steel frame), MCS was performed using NN
for describing structural behavior. The results were similar
to those obtained by conventional MCS and the computing
time was reduced.
In the third example (reliability-based optimization of a
six-bar truss), NN were used in conjunction with genetic
algorithms and MCS. In spite of the problem analysed
being non-convex, this methodology made possible to get
a solution for the optimum structural design with reliability
constrains. A parallel computing scheme was adopted, hav-
ing been found that genetic algorithms can be easily and
The present work was conducted in the framework of
the Research Project POCTI/ECM/36055/99, financed by
the Portuguese Foundation for Science and Technology.
This support is gratefully acknowledged.
References
[1] Eurocode 1: basis of design and actions on structures. European
Committee for Standardization, Brussels; 1994.
[2] Eurocode 3: design of steel structures. European Committee for
Standardization, Brussels; 1992.
[3] Hasofer AM, Lind NC. Exact and invariant second-moment code
format. ASCE, J Eng Mech Div 1974;100(EM1):111–21.
[4] Shooman ML. Probabilistic reliability: an engineering
approach. McGraw-Hill; 1968.
[5] Mahadevan S. In: Cruse TA, editor. Monte Carlo simulation:
reliability-based mechanical design. Marcel Dekker; 1997. p. 123–46.
[6] McCulloch WS, Pitts W. A logical calculus of ideas immanent in
nervous activity. Bull Math Biophys 1943;5:115–33.
[7] Berke L, Hajela P. Applications of artificial neural nets in structural
mechanics. Structural optimization, vol. 4. Springer-Verlag; 1992, p.
90–8.
[8] Bento JP. Modeling mechanical behavior without mechanics. In:
Tasso C, de Arantes e Oliveira ER, editors. Development of
knowledge-based systems for engineering. Springer-Verlag; 1998. p.
37–58.
[9] Hurtado JE. Neural networks in stochastic mechanics. Arch Comput
Meth Eng 2001;8(3):303–42.
[10] Papadrakakis M, Papadopoulos V, Lagaros ND. Structural reliabil-
ity of elastic–plastic structures using neural networks and Monte
Carlo simulation. Comp Meth Appl Mech Eng 1996;136:145–63.
[11] Rosenblatt F. The perceptron: a probabilistic model for information
storage and organization in the brain. Psychol Rev 1958;65:386–408.
[12] Rumelhart DE, Hinton G, Williams R. Learning internal represen-
tations by error propagation. PDP Research Group. Cambridge:
MIT Press; 1986.
[13] Cybenko G. Approximation by superposition of sigmoidal functions.
Math Control Signal Syst 1989;2(3):303–14.
[14] Thoft-Christensen P, Baker MJ. Structural reliability theory and its
applications. Springer-Verlag; 1982.
[15] Neves LAC, Cruz PJS, Henriques AAR. SFEM reliability of steel
frames. In: Proceedings of the Eurosteel conference; 2002. p. 1531–40.
[16] COMREL-TI Users Manual, RCP GmbH, Munich; 1992.
www.strurel.de.
[17] Murotsu Y, Shao S, Kogiso N, Tomioka H. Optimal shape of truss
structure based on reliability. In: Proceedings of the US–Japan joint
seminar on structural optimization; 1996. p. 145–56.
[18] Burton SA, Hajela P. A variable-complexity approach to second-
order reliability-based optimization. Struct Multidiscip Optimization
2003;25:237–50.
[19] Goldberg DE. Genetic algorithms in search, optimization and
machine learning. Addison-Wesley; 1989.