INTERNATIONAL JOURNAL OF OPTIMIZATION IN CIVIL ENGINEERING

Int. J. Optim. Civil Eng., 2014; 4(1):1-26

NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF

OPTIMALLY SEISMIC DESIGNED MOMENT FRAMES

S. Gholizadeh*, †, V. Aligholizadeh and M. Mohammadi

Department of Civil Engineering, Urmia University, Urmia, Iran

ABSTRACT

In the present study, the reliability assessment of performance-based optimally seismic

designed reinforced concrete (RC) and steel moment frames is investigated. In order to
achieve this task, an efficient methodology is proposed by integrating Monte Carlo
simulation (MCS) and neural networks (NN). Two NN models including radial basis
function (RBF) and back propagation (BP) models are examined in this study. In the
proposed methodology, MCS is used to estimate the total exceedence probability associated
with immediate occupancy (IO), life safety (LS) and collapse prevention (CP) performance
levels. To reduce the computational burden of MCS process, the required nonlinear
responses of the generated structures are predicted by RBF and BP models. The numerical
results imply the superiority of BP to RBF in prediction of structural responses associated
with performance levels. Finally, the obtained results demonstrate the high efficiency of the
proposed methodology for reliability assessment of RC and steel frame structures.

Received: 10 December 2013; Accepted: 20 February 2014

KEY WORDS: Seismic reliability, performance-based design, structural optimization,

Monte Carol simulation, neural network

1. INTRODUCTION

Performance-based design (PBD) [1-3] of structures subject to sever ground motions is an

efficient design process and many researchers and engineers have proposed various
methodologies, which incorporated PBD concepts and criteria, to improve structural
performance. In essence, PBD process is based on the principle that a structure should meet
performance levels according to a set of specified reliabilities over its service life. The PBD

*
Corresponding author: S. Gholizadeh, Department of Civil Engineering, Urmia University, Urmia, Iran
†
E-mail address: s.gholizadeh@urmia.ac.ir
2 S. Gholizadeh, V. Aligholizadeh

approach aims to present structural configurations with a predictable and reliable

performance level which highly depends on structural seismic responses and capacity that
both of these parameters are inherently uncertain. In addition, material properties and
characteristics associated with ground motions are also uncertain parameters. Theory and
methods for structural reliability assessment have been developed substantially in the last
few years and they are actually useful tools for evaluating rationally the safety of structures
against the existing uncertainties. Thus, the reliability theory and PBD approach should be
simultaneously utilized to design reliable structures for earthquake loadings.
During the last years, some computational strategies have been proposed for seismic
reliability assessment of reinforced concrete (RC) and steel structures. Several researchers
[4–6] have followed different methods and strategies in developing fragility curves for RC
structures. Moller et. al. [7–8] proposed a methodology with variable uncertainties for
seismic vulnerability of RC frames. In this approach, seismic vulnerability was defined as
the conditional probability of exceeding different limit states within a performance
requirement, given by a hazard level. They used neural network techniques to evaluate the
required structural responses. In the most recent work, Khatibinia et. al. [9] proposed a
meta-model to reduce the computational cost in seismic reliability assessment of existing
RC structures considering soil-structure interaction (SSI) effects. Their proposed meta-
model was designed by combining weighted least squares support vector machine (WLS-
SVM) and a wavelet kernel function. Their study showed that SSI has the significant
influence on the seismic reliability assessment of structures. In the case of steel structures,
Lagaros et. al. [10] proposed NN-based strategies for solving reliability-robust design
optimization problems. They used standard BP model for predicting structural responses
during the optimization process.
Nonlinear structural analysis by finite element method (FEM) is a time consuming
process. Consequently, the required computational cost will be expensive when the
nonlinear structural responses are required for seismic reliability assessment of the structure
using MSC. The MCS method requires a large number of structural analyses in order to
obtain an acceptable confidence within probabilities of the order close to 10 -4 –10-6. For
such results, the number of analyses that will achieve a 95% likelihood for actual probability
must be between 1.6×106 and 1.6×109 [11]. One of the best candidates to reduce the
computational burden of evaluating nonlinear structural responses required for performing
seismic reliability assessment is soft computing-based approximation tools: neural network
(NN) and support vector machines (SVM). A comprehensive review of NNs and SVMs
applied for reliability assessment of structural systems has been presented in [9].
In the present study, the bat algorithm (BA) [12] are employed to find performance-based
seismic optimum deign of two RC frames and two steel frames. The BA is a newly
developed meta-heuristic optimization algorithm based on the echolocation behavior of
bats. The capability of echolocation of bats is fascinating, as these bats can find their prey
and discriminate different types of insects even in complete darkness. The superiority of BA
to some popular meta-heuristic algorithms such as genetic algorithm (GA) and particle
swarm optimization (PSO) for solving engineering optimization problems have been
demonstrated [13] and thus in this paper BA is utilized to implement optimization process.
In order to assess reliability of the optimally seismic designed structures for existing
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 3

uncertainties a combination of MCS method and NN models is employed. Radial basis

function (RBF) and back propagation (BP) neural networks, as two popular networks, are
employed to predict the seismic responses of the frame structures in the framework of MCS.
The numerical results indicate that the BP model in conjunction with MCS method provides
a powerful computational tool for reliability assessment of RC and steel frames.

2. OPTIMAL SEISMIC DESIGN

Many studies have shown that structures designed by modern seismic code procedures are
expected to undergo large cyclic deformations in the inelastic range when subjected to
severe earthquake ground motions. Nevertheless, most seismic design codes are still based
on elastic methods using equivalent static lateral design forces. This procedure can result in
unpredictable and poor response during severe ground motions with inelastic activity
unevenly distributed among structural members. To solve this problem, performance-based
design (PBD) procedure was developed. PBD procedures allow engineers to determine
explicitly performance of structures for a special seismic hazard level. PBD enables
designers to design structures having predictable and reliable performance against seismic
loadings. The designs obtained by PBD approach should meet performance objectives. A
performance objective is defined as a given level of performance for a specific hazard level.
To define a performance objective, at first the level of structural performance should be
selected and then the corresponding seismic hazard level should be determined. In the
present work, immediate occupancy (IO), life safety (LS) and collapse prevention (CP)
performance levels are considered according to FEMA-356 [1]. Each objective corresponds
to a given probability of being exceed during 50 years. A usual assumption [14] is that the
IO, LS and CP performance levels correspond respectively to a 20%, 10% and 2%
probability of exceedance in 50 year period. In this study, the mentioned hazard levels are
considered.

2.1 Nonlinear pushover analysis

The estimation of demands can be accomplished using a variety of available procedures.
Using the nonlinear static procedure, the inelastic behavior of the structure as a whole can
be captured by a static pushover curve. The pushover curve gives an accurate description of
the structural behavior, compared to the dynamic analysis, at least for a structure that has a
low number of participating modes. The advantage in utilizing a pushover analysis relies in
the fact that it can be used in most practical cases. On the other hand, dynamic inelastic time
history analyses are often difficult to implement. The practical objective of inelastic seismic
analysis procedures is to predict the expected behavior of the structure in future earthquake
shaking. This has become increasingly important with the emergence of performance-based
engineering (PBE) as a technique for seismic evaluation and design [15]. Among the various
methods of static nonlinear pushover analyses, the displacement coefficient method [1] is
adopted to evaluate the seismic demands on building frameworks under equivalent static
earthquake loading. In this method the structure is pushed with a specific distribution of the
lateral loads until the target displacement is reached. The target displacement can be
4 S. Gholizadeh, V. Aligholizadeh

obtained from the FEMA-356 as follows:

Te2
 t  C0C1C2C3Sa g (1)
4 2

where C0 relates the spectral displacement to the likely building roof displacement; C1
relates the expected maximum inelastic displacements to the displacements calculated for
linear elastic response; C2 represents the effect of the hysteresis shape on the maximum
displacement response and C3 accounts for P-D effects. Te is the effective fundamental
period of the building in the direction under consideration; Sa is the response spectrum
acceleration corresponding to the Te.
In this work, the OPENSEES [16] platform is utilized to conduct the pushover analyses.

2.2 Problem formulation

In a sizing structural optimization problem, the aim is usually to minimize the weight or cost
of the structure under some behavioural constraints. For a frame structure consisting of ne
members that are collected in n design groups, if the variables associated with each design
group are selected from a given database of sections, a discrete optimization problem can be
formulated as follows:

Minimize f ( X )
Subject to: g i (X )  0 i  1,2, , ng (2)
X  { x 1 x 2 ... x j ... x n }T

where xj is an integer value expressing the sequence numbers of sections assigned to jth
group; f represents the objective function of the frame; gi (X ) is the ith behavioral
constraint; ng is the number of constraints.
Objective function for a RC frame, f R , is usually taken as the cost of structure and can be
defined as follows [17]:
nb
f R    CC bb ,i hb ,i  C S AS,b ,i  C F ( bb ,i  2 hb ,i )  Li 
i 1 (3)
 C b hc ,j  C S AS,c ,j  2C F ( bc ,j  hc ,j )  H j
nc

C c ,j
j 1

where nb is the number of beams; bb,i , hb,i , Li and AS,b,i are the ith beam width, depth, length
and area of the steel reinforcement, respectively; nc is the number of columns; bc,j , hc,j , Hj
and AS,c,j are the jth column width, depth, length and area of the steel reinforcement,
respectively; CC, CS and CF are the unit cost of concrete, steel and framework, respectively.
As mentioned in [17], in the present work the following unit costs are also considered:
CC=105 $/m3, CS=7065 $/ m3, CF=92 $/ m2.
In the case of a steel frame, the weight of structure is usually considered as the objective
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 5

function, f S , and it can be defined as follows [18]:

n nm
fS   i Ai  L j (4)
i 1 j 1

where ρi and Ai are weight of unit volume and cross-sectional area of the ith group section,
respectively; nm is the number of elements collected in the ith group; Lj is the length of the
jth element in the ith group.
In this study, the constraints of the optimization problem are handled using the concept
of exterior penalty functions method (EPFM) [19]. In this case, the pseudo unconstrained
objective function is expressed as follows:

 nc
2
Φ( X , rp )  f ( X ).1  rp  max{0, gi ( X )}  (5)
 i 1 

where Φ and rp are the pseudo objective function and positive penalty parameter,
respectively.
In this work, two types of constraints are checked during the optimization process. The first
type includes the checks of each structural element for gravity loads. In this case, the
following load combination is considered:

QG1  1.2QD  1.6QL (6)

where QD and QL are dead and live loads, respectively.

Each structural element of RC and steel frames should satisfy the constraints specified
respectively by ACI 318-08 [20] and LRFD-AISC [21] codes for the non-seismic load
combination.
If the first type constraints are not satisfied then the candidate design is rejected, else a
nonlinear pushover analysis is performed in order to estimate the maximum inter-story drift
ratios at various performance levels. In nonlinear static pushover analysis, the lateral load
distribution in the height of the frame is defined as follows [1]:

Gs H sk (7)
Ps  Vb ( ns
)
G
m 1
m H k
m

where Ps = lateral load applied at story s; Vb = base shear; Hs, Hm = height from the base of
the building to stories s and m, respectively; Gs, Gm = seismic weight for story levels s and
m, respectively; k = 2 and this means that the lateral load pattern is parabolic.
The following component gravity force is considered for combination with the seismic
loads [1]:

QG 2  1.1(QD  QL ) (8)
6 S. Gholizadeh, V. Aligholizadeh

In this study, the lateral inter-story drift ratios are considered as the acceptance criteria. The
inter-story drift ratio constraints at various performance levels can be expressed as:

il
gil ( X )   1  0 , l  IO; LS; CP, i  1,2,..., ns (9)
 all
l

where  il and  all

l
are respectively the ith story drift and its allowable value associated with
lth performance level; ns is the number of stories.
In order to find performance-based seismic optimum deign of frames BA is employed.
The basic concepts of BA are defined as follows.

2.3 Bat algorithm

The BA meta-heuristic is inspired from the echolocation behaviour of microbats [22]. The
echolocation characteristics of microbats in BA are idealized as the following three rules [13].
The first rule is that all bats use echolocation to sense distance, and they also know the
difference between prey and background barriers in some magical way. As the second rule,
bats randomly fly with velocity Vi at position Xi with a fixed frequency fmin, varying
wavelength and loudness Ω 0 to search for prey. They can adjust the frequency of their emitted
pulses and adjust the rate of pulse emission r [0,1], depending on the proximity of their
target. The third rule says that although the loudness can vary in many ways, it is assumed that
the loudness varies from a large and positive Ω 0 to a minimum constant value Ωmin .
The position and velocity of each bat should be updated in the design space. As
optimization of RC and steel frames using the section databases is a discrete optimization
problem, the following equations are employed for updating position and velocity of ith bat:

fi  f min  ( f max  f min )ui (10)


Vi k 1  round Vi k  ( X ik  X * ) fi  (11)

X ik 1  X ik Vi k 1 (12)

where fmin and fmax are the lower and upper bounds imposed for the frequency range of bats.
In this study, fmin=0.0 and fmax=1.5 are used; ui  [0,1] is a vector containing uniformly
distribution random numbers; X * is the current global best solution;
A local search is implemented on a randomly selected bat from the current population as:

X k 1  X k  round ( j Ω k 1 ) (13)

where  j is a uniform random number in the range of [-1, 1] selected for each design variable
of the selected bat. Ω k 1 is the average loudness of all the bats at the current iteration.
The loudness Ωi and the rate ri of pulse emission have to be updated accordingly as the
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 7

iterations proceed. As the loudness usually decreases once a bat has found its prey, while the
rate of pulse emission increases, the loudness can be chosen as any value of convenience. In
this work, Ω0  1 and Ωmin  0 also, r0 = 0 and rmax = 1.

Ωik 1   Ωik (14)

ri k 1  ri0 (1  e .k ) (15)

where α and γ are constants. In this study, α = 0.5, and γ = 0.5.

3. SEISMIC RELIABILITY ASSESSMENT

Deterministic structural optimization without considering the uncertainties in structural

capacity and seismic demands results in an unreliable design and cannot balance both cost
and safety. The structural seismic responses can be affected by many uncertain variables.
Material properties and earthquake loading are considered as intervening uncertain variables
in this study. Therefore, the main aim of the present paper is to assess the seismic reliability
of optimally seismic designed RC and steel frames considering the mentioned sources of
uncertainties. To achieve this purpose, nonlinear pushover analysis of the structures is
implemented to obtain the structural seismic responses. The limit state functions associated
with each performance level are calculated using the structural seismic responses. Then, the
non-performance probability corresponding to each performance level is evaluated. MCS is
a powerful tool, simple to implement and capable of solving a broad range of reliability
problems. However, its use for evaluation of very low probabilities of failure implies a great
number of structural analyses which can become excessively time consuming especially
when the nonlinear analysis is involved. To reduce the computational burden of the MCS-
based reliability assessment process, RBF and BP NN models are employed to predict the
required nonlinear responses of the structures at IO, LS and CP performance levels. The
proposed methodology provides the possibility of reproducing structural behavior for
performance evaluation at a trivial computational cost. In order to obtain seismic reliability
assessment using the proposed methodology, random samples are generated and used to
train and test the RBF and BP NN models. Random combinations of intervening variables
and the corresponding desired structural seismic responses are respectively considered as
input and outputs of the NN models.

3.1 Input and output vectors

For structural elements of RC frames, Kent–Scott–Park model [23] is used as the confined
and unconfined concrete constitutive model. The concrete of cover and core in cross-section
of the columns is considered as unconfined and confined, respectively. As shown in Figure
1(a) fc, fu, ε0 and εu are the constitutive parameters of this model representing respectively
concrete peak strength in compression, residual strength, strain at peak strength, and
ultimate compressive strain. Constitutive behavior of the reinforcing steel shown in Figure
8 S. Gholizadeh, V. Aligholizadeh

1(b) is based on using the one-dimensional plasticity model with linear hardening. The
material parameters of E, fy and H which are respectively Young’s modulus, yield strength,
and hardening modulus define the plasticity model. In this study, the above constitutive
parameters fc, E and fy are considered as the random parameters. For structural elements of
steel frames, the constitutive behavior of Figure 1(b) is employed considering E and fy as the
random parameters. Beams and columns of the RC and steel moment frames are modeled
using force-based nonlinear beam-column element on the OPENSEES platform that
considers the spread of plasticity along the element’s length.

(a) (b)
Figure 1. Material constitutive behavior (a) concrete, (b) steel

In this study, response spectrum acceleration Sa associated with the mentioned triple
hazard levels are considered as the random parameters termed here as S aIO , S aLS and S aCP .
Therefore, the random variables vector for RC and Steel frames which are employed as the
input vector of the NN models can be represented as follows:

I NN, R  { f c E fy SaIO SaLS SaCP}T (16)

I NN, S  {E fy SaIO SaLS SaCP}T (17)

where I NN, R and I NN,S are input vectors of NN models for RC and steel frames, respectively.
It should be noted that in a structure including n design groups, a set of random
constitutive parameters is considered for each design group. In this case, constitutive
parameters fc, E and fy are vectors defined as follows:

fc  { fc1 fc 2  fc i  f c n }T (18)

E  {E1 E2  Ei  En }T (19)

f y  { f y1 f y2  f yi  f yn }T (20)
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 9

For seismic reliability assessment, maximum inter-story drift ratios at IO, LS and CP
performance levels are selected as the structural seismic responses. Thus, for both RC and
steel frames the output vector of the NN models, ONN, is as follows:

ONN  {max
IO
max
LS
max
CP T
} (21)

3.2 Monte carlo simulation

A reliability problem is normally formulated using a limit state function. Limit state
function for each performance level is defined using capacity and demand as follows:

G(Z )  RLIM  R(Z) (22)

where G is a limit state function; Z is the vector of random variables defined for RC and
steel frames by Eqs. (16) to (20); RLIM is the limiting value for a seismic response R(Z).
In similar studies such as [7-9] the limiting capacities RLIM were also considered random.
In the present study, these parameters for RC and steel frames are considered deterministic.
As in the present work only the maximum inter-story drift ratios at the performance levels
are selected as the structural seismic responses, the limit state functions considered for the
three performance levels are as follows:

Immediate Occupancy: GIO (Z )  LIM

IO
 max
IO
(Z) (23)

Life Safety: GLS (Z )  LIM

LS
 max
LS
(Z) (24)

Collapse Prevention: GCP (Z )  LIM

CP
 max
CP
(Z) (25)

where  LIM
IO
,  LIM
LS
and  LIM
CP
are the limiting values for maximum inter-story drift seismic
responses at the IO, LS and CP performance levels, respectively.
The non-performance probability, Pf, is defined as a function of the limit state functions
corresponding to a given performance level. Estimation of the non-performance probability
in the time-invariant domain requires the evaluation of the multiple integral over the failure
domain, G(Z) < 0, as follows [9]:

Pf     FZ ( Z )dZ (26)
G(Z )

where FZ (Z ) is the joint probability density function of Z.

The total exceedence probability, Pf E , for each performance level is defined as a series
system when one of the limit state functions fails:

 nl 
Pf E  P {Gi ( Z )  0} (27)
 i1 
10 S. Gholizadeh, V. Aligholizadeh

where nl is the number of the limit state functions for each performance level.
As in the present work only one limit state function is defined for each performance
level, Eq. (27) can be rewritten as follows:

Pf E , l  PGl (Z )  0 , l  IO, LS, CP (28)

Computation of total exceedence probability, Pf E , l , requires integration of a multi-normal

distribution function [9]. This integral can be estimated by the MCS method. In this study,
the MCS method is utilized simultaneously for all limit state functions of the performance
levels. The MCS method allows the determination of an estimate of Pf E , l , given by

1 N
Pf E , l   Il , i (Z ) , l  IO, LS, CP
N i1
(29)

1 if Gl ( Z )  0
Il (Z )   , l  IO, LS, CP (30)
0 if Gl ( Z )  0

where N is the number of independent samples generated based on the probability

distribution for each random variable for implementation MCS.
Implementation of the MCS requires a large number of structural nonlinear pushover
analyses. The MCS is a time consuming process because of high computational cost of
pushover analysis. To reduce the computational burden of MCS, the required structural
seismic responses are predicted by NN models.

4. NEURAL NETWORKS

The principal advantage of a properly trained NN is that it requires a trivial computational

burden to produce an approximate solution. Such approximations appear to be valuable in
situations where the actual response computations are intensive in terms of computing time
and a quick estimation is required. For such problems a NN model is trained utilizing
information generated from a number of properly selected analyses. The data from these
analyses are processed in order to obtain the necessary input and target pairs, which are
subsequently used to train the network [24]. In the present paper, two well known and
popular NN models are employed: Radial basis function (RBF) and Back-propagation (BP).

4.1 RBF model

RBF neural networks due to their fast training, generality and simplicity are popular. They
are two layers feed forward networks. The hidden layer consists of RBF neurons with
Gaussian activation functions. The outputs of RBF neurons have significant responses to the
inputs only over a range of values called the receptive field. The radius of the receptive field
allows the sensitivity of the RBF neurons to be adjusted. During the training, the receptive
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 11

field radius of RBF neurons is such determined as the neurons could cover the input space
properly. The output layer neurons produce the linear weighted summation of hidden layer
neurons responses. To train the hidden layer of RBF networks no training is accomplished
and the transpose of training input matrix is taken as the layer weight matrix [25].

W1RBF  ΛT (31)

where, W1RBF and ΛT are input layer weight and training input matrices, respectively.
In order to adjust output layer weights, a supervised training algorithm is employed. The
output layer weight matrix is calculated from the following equation:

W2RBF  Δ1 T (32)

where T is target matrix, Δ is the outputs of the hidden layer and W2RBF is the output layer
weight matrix.

4.2 BP model
Standard BP [26] is a gradient descent optimization algorithm, which adjusts the weights in
the steepest descent direction according to the following equation:

1  Wk  ηk Gk
WkBP BP
(33)

where WkBP and Gk are the weight and the current gradient matrices, respectively and ηk is
learning rate.
Levenberg-Marquardt (LM) [27] algorithm was designed to approach second-order
training speed without having to compute the Hessian matrix. In the LM algorithm the
weights updating is achieved as follows:

1 T
1  Wk  [ J J  μI ] J E
WkBP BP T
(34)

where J is the Jacobian matrix that contains first derivatives of the network errors with
respect to the weights; E is a vector of network errors; μ is a correction factor.
One of the techniques used to prevent overfitting is regularization [26] in which the
performance function of the network is modified by adding a term that consists of the mean
of the sum of squares of the network weights as:

1 m 2 (1  γ) nw
msereg  γ(  Ei )  nw 
m i 1 j 1
(W jBP ) 2 (35)

where γ and nw are the performance ratio and number of network weights, respectively; m is
the size of Ei.
12 S. Gholizadeh, V. Aligholizadeh

4.3 Prediction of seismic responses

NN models are trained to predict the structural seismic responses for implementation of
MCS spending a reasonable amount of time. The input and output vectors of the NN models
are represented for RC frames by Eqs. (16) and (21) and for steel frames by Eqs. (17) and
(21), respectively. The topology of NN models trained to predict the responses of RC and
steel frames are shown in Figures (2) and (3), respectively.
To evaluate the prediction accuracy of the trained NN models in testing mode, the
absolute percentage error (APE) between the nts number of predicted and the actual
responses and also mean absolute percentage error (MAPE) are computed as follows:

 max
i
  max
i

APE i actual predicted , i = IO, LS, CP , j = 1, …, nts (36)

j
 max
i

actual j

 100 APE 
nts
1
MAPE i  i
j
, i = IO, LS, CP (37)
nts j 1

Figure 2. Topology of NN models for prediction of seismic responses of RC frames

Figure 3. Topology of NN models for prediction of seismic responses of steel frames

 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 13

5. NUMERICAL RESULTS

The numerical results of the present study are presented for RC and steel ordinery moment
frames in this sections. The required computer programs for performing optimization
processes are coded in MATLAB [28]. In addition, a personal Pentium IV 3.0 GHz is used
for computer implementation.
For deterministic performance-based seismic design optimization of RC and steel
moment frames the allowable values of inter-story drift in low-rise frame, for IO, LS and CP
performance levels are taken according to HAZUS [29] provisions. These values in mid and
high rise frames should be multiplied to 2/3 and 0.5, respectively.
In order to generate dataset for training NN models for RC and steel frames, Latin
Hypercube Design (LHD) sampling [30] is employed.

5.1 RC frames
In order to illustrate the computational advantages of the proposed methodology in the case
of RC frames, a three-bay, six-story and a four-bay, nine-story 2-D RC frames are
considered. The geometry and element groups of the frames are shown in Figure 4.

(a) (b)
Figure 4. Geometry and element groups for (a) three-bay, six-story and (b) four-bay, nine-story
RC frames
14 S. Gholizadeh, V. Aligholizadeh

Uniform gravity loads are considered on beams as a dead load DL = 25 kN/m and a live
load LL = 10.0 kN/m. A semi-infinite set of member sizes and reinforcement arrangement
for beams and columns is reduced by constructing databases in practical range. The
databases information is sectional dimensions and number of reinforcing bars for the beams
and columns as show in Tables 1 and 2, respectively.

Table 1: Cross section database for beams

Number of bars (D22)
Beam NO. Width (cm) Depth (cm)
Top Bottom
1 30 45 2 2
2 30 45 3 2
    
496 45 90 9 8
497 45 90 9 9

Table 2: Cross section database for columns

Column NO. Width (cm) Depth (cm) Number of bars (D25)
1 30 30 4
2 35 35 6
   
43 90 90 38
44 90 90 40

Median response spectra for three hazard levels [31] employed for RC frames are shown
in Figure 5.

Figure 5. Median response spectra for three hazard levels [31]

In the RC frame examples deterministic parameters are as follows: ε0=0.002, fu=0.0,

εu=0.005 and H=0.01. The properties, probability density function, mean value and standard
deviation of each random parameter are given in Table 3.
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 15

Table 3: Properties of the random variables for RC frames

Random Variable Probability density function Mean value Standard deviation
fc Normal 28 MPa 0.1 fc
E Normal 200 GPa 0.1E
fy Normal 240 MPa 0.1 fy
S aIO Lognormal S aIO (Figure 5) 0.15 S aIO
S aLS Lognormal S aLS (Figure 5) 0.15 S aLS
CP
S a Lognormal S aCP (Figure 5) 0.15 S aCP

5.1.1 Three-bay, six-story RC frame

The six-story RC frame is designed for optimal cost using BA meta-heuristic. In the
optimization process a population of 30 bats is considered and the maximum number of
generations is limited to 1000. During the optimization process the lateral inter-story drifts
are checked at various performance levels as the design constraints. As this structure is a
mid-rise frame, the allowable values of inter-story drift at IO, LS and CP levels are 0.66%,
2% and 5.33%, respectively. The results of optimization are given in Table 4.

Table 4: Results of optimization for three-bay, six-story RC frame

Element Sectional dimensions Reinforcements
Type Group Width (cm) Depth (cm) Top Bottom
B1 35 60 3-D22 3-D22
Beam B2 35 55 4-D22 4-D22
B3 30 55 4-D22 3-D22
C1 50 50 12-D25
C2 50 50 12-D25
Column
C3 45 45 12-D25
C4 40 40 8-D25
Frame cost (fR) 33364 $

Inter-story drift profiles of the optimum design are shown in Figure 6.

Figure 6. Inter-story drift profiles of the optimum six-story RC frame at performance levels
16 S. Gholizadeh, V. Aligholizadeh

The vertical dashed lines denote drift limits for performance levels. It can observe that
optimum solution is feasible and the constraint of IO level dominates the design.
In order to assess the seismic reliability of the optimum six-story RC frame, a database of
random parameters should be generated. In this case, 8000 random vectors, i.e. INN,R, are
selected and the resulted structures are analyzed using nonlinear pushover analysis and the
maximum inter-story drifts, i.e. ONN, are saved. The generated data is divided into training
and testing sets including 7500 and 500 samples, respectively. RBF and BP models are
trained and the results reveal that the computational performance of BP is very better than
that of the RBF model. Thus, only the results of BP model are presented in this example.
Various numbers of hidden layer neurons are examined for BP model and the best accuracy
is achieved using 15 ones. Figure 7 shows the APE of the predicted  maxIO
,  max
LS
and  max
CP
.

(a)

(b)

(c)
Figure 7. APE of the predicted  IO
max
, LS
max
and  max
CP
by BP for the optimum six-story RC frame
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 17

The MAPEs of the predicted  max

IO
,  max
LS
and  max
CP
are respectively equal to 1.9319%, 1.8902%
and 4.3626%. These results demonstrate the good accuracy of the trained BP for predicting
the seismic responses of the structure. The trained BP model is effectively employed to
implement MCS. In this case, 107 vectors of INN,R are generated and their corresponding ONN
are predicted by the trained BP model. The values of PfE are calculated using Eqs. (29) and
(32) for various performance levels. These values are shown in Figure 8 for various numbers
of samples.

(a)

(b)

(c)
Figure 8. Total exceedence probability of limit state functions for the optimum six-story RC
frame at (a) IO, (b) LS and (c) CP performance levels

These results indicate that the optimum six-story RC frame is highly vulnerable against
the existing uncertainties at IO and LS performance levels. The results of Figure 8 reveal
18 S. Gholizadeh, V. Aligholizadeh

that for all performance levels, the difference between PfE values obtained by 5×104 and 107
samples is trivial. It is clear that for seek of efficiency in terms of computational cost the
application of 5×104 samples is the best choice.

5.1.2 Four-bay, nine-story RC frame

The BA is used to find performance-based optimal seismic design of the nine-story RC
frame. The number of bats and the maximum number of generations during optimization
process are 30 and 10000, respectively. This structure is considered as a high-rise frame and
the allowable inter-story drift at IO, LS and CP levels are 0.5%, 1.5% and 4%, respectively.
Optimization results are given in Table 5.

Table 5: Results of optimization for four-bay nine-story RC frame

Element Sectional dimensions Reinforcements
Type Group Width (cm) Depth (cm) Top Bottom
B1 40 60 5-D22 4-D22
Beam B2 40 60 3-D22 3-D22
B3 30 50 4-D22 4-D22
C1 60 60 16-D25
C2 60 60 16-D25
C3 55 55 14-D25
C4 55 55 14-D25
Column C5 50 50 12-D25
C6 45 45 10-D25
C7 45 45 10-D25
C8 45 45 10-D25
C9 45 45 10-D25
Frame cost (fR) 70315 $

Inter-story drift profiles of the optimum design at various performance levels are shown
in Figure 9.

Figure 9. Inter-story drift profiles of the optimum nine-story RC frame at performance levels
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 19

The vertical dashed lines denote drift limits for performance levels. The results imply
feasibility of optimum solution. It is observed again that the constraint associated with IO
level dominates the design.
For assessing the seismic reliability of the optimum nine-story RC frame, a database of
random vectors INN,R and their corresponding ONN including 12000 samples is generated and
11500 ones are used for training and 500 ones are employed for testing the NN models. As
well as the first example, RBF and BP models are trained and the results demonstrate the
superiority of BP to the RBF model. Therefore, in this example only the results based on
application of BP model are presented.

(a)

(b)

(c)
Figure 10. Total exceedence probability of limit state functions for the optimum nine-story RC
frame at (a) IO, (b) LS and (c) CP performance levels
20 S. Gholizadeh, V. Aligholizadeh

The numerical results indicate that the best results of BP model are obtained when 15
neurons are set in the hidden layer. The MAPEs of the predicted  max IO
,  max
LS
and  max
CP
are
respectively equal to 2.2304%, 2.2393% and 4.5951%. These results demonstrate the good
accuracy of the trained BP model. To implement MCS, 107 vectors of INN,R are generated
and their corresponding ONN are predicted by the trained BP model. The values of PfE are
calculated for various performance levels. These values are given in Figure 10.
Total exceedence probability of limit state functions for the optimum nine-story RC
frame at IO, LS and CP performance levels are 0.855, 0.346 and 0.004, respectively. These
obtained results imply the high vulnerability of the optimum nine-story RC frame against
the existing uncertainties at IO and LS performance levels. The results of Figure 10 reveal
that for all performance levels, the difference between PfE values obtained by 5×104 and 107
samples is trivial.

5.2 Steel frames

Two steel frame examples are presented to illustrate the computational performance of the
proposed methodology. These examples include a two-bay, three-story and a three-bay, ten-
story 2-D steel frames as shown in Figure 11.

(a) (b)
Figure 11. Geometry and element groups for (a) two-bay, three-story and (b) three-bay, ten-story
steel frames
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 21

Uniform gravity loads are considered on beams as a dead load DL = 25 kN/m and a live
load LL = 10.0 kN/m. In the present study, design variables of steel frames are selected from
W-shaped sections found in the AISC [21] design manual.
The spectral acceleration S ai can be calculated for each design spectrum i as follows:

 Fa S si (0.4  3T/T0 ) if 0  T  0.2 T0i


S ai   Fa S si if 0.2 T0i  T  T0i , i  IO, LS, CP (39)
F Si/ T if T  T0i
 v 1

Fv S1i
T0i  (40)
Fa S si

where T is the elastic fundamental period of the structure, which is computed here from
structural modal analysis; S si and S1i are the short-period and the first sec.-period response
acceleration parameters, respectively; T0i is the period at which the constant acceleration and
constant velocity regions of the response spectrum intersect; Fa and Fv are the site
coefficient determined respectively based on the site class and the values of the response
acceleration parameters S si and S1i , according to Table 6 [32].

Table 6: Performance level site parameters for site class of D

Performance Level Hazard Level Ss (g) S1 (g) Fa Fv
IO 20% / 50-years 0.658 0.198 1.27 2.00
LS 10% / 50-years 0.794 0.237 1.18 1.92
CP 2% / 50-years 1.150 0.346 1.04 1.70

In the steel frame examples deterministic parameter is H=0.03. The properties,

probability density function, mean value and standard deviation of each random parameter
are given in Table 7.

Table 7: Properties of the random variables for steel frames

Random Variable Probability density function Mean value Standard deviation
E Normal 210 GPa 0.1E
fy Normal 240 MPa 0.1 fy
S aIO Lognormal S aIO (Eq. (38)) 0.1 S aIO
S aLS Lognormal S aLS (Eq. (38)) 0.1 S aLS
S aCP Lognormal S aCP (Eq. (38)) 0.1 S aCP

5.2.1 Two-bay, three-story steel frame

The three-story steel frame is designed for optimal weight using BA meta-heuristic. In the
22 S. Gholizadeh, V. Aligholizadeh

optimization process a population of 40 bats is considered and the maximum number of

generations is limited to 300. During the optimization process the lateral inter-story drifts
are checked at various performance levels as the design constraints. As this structure is a
low-rise frame, the allowable values of inter-story drift at IO, LS and CP levels are 1.2%,
3% and 8%, respectively. The results of optimization are given in Table 8.

Table 8. Results of optimization for two-bay three-story steel frame

Design variables Optimal sections
C1 W16×26
C2 W18×40
B1 W16×26
Weight (fS) 2404.83 kg

Inter-story drift profiles of the optimum design are shown in Figure 12 for performance
levels. The vertical dashed lines denote drift limits for performance levels. It can be
observed that optimum solution is feasible and the constraint associated with IO level
dominates the design.

Figure 12. Inter-story drift profiles of the optimum three-story steel frame at performance
levels

For performing seismic reliability analysis of the optimum three-story steel frame, a
database including 10000 samples is generated. From the generated database 9000 and 1000
samples are employed for training and testing the NN models, respectively. As well as the
RC frame example, RBF and BP models are trained and the results demonstrate the
superiority of BP to the RBF model. Thus, in this example only the results of BP model are
presented. The numerical results indicate that the best results of BP model are obtained
when 5 neurons are set in hidden layer. The MAPEs of the predicted  max IO
,  max
LS
and  max
CP
are
respectively equal to 0.0043%, 0.1734% and 0.0994%. These results demonstrate the
excellent accuracy of the trained BP model. To implement MCS, 10 7 vectors of INN,S are
generated and their corresponding ONN are evaluated by the BP model. The values of PfE are
calculated for various performance levels and it is observed that for LS and CP performance
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 23

levels PfE =0. The values of PfE are given in Figure 13 for IO performance level.

Figure 13. Total exceedence probability of limit state functions for the optimum three-story
steel frame at IO performance level

It is clear that the optimum three-story steel frame is highly vulnerable against the
existing uncertainties at IO level. The results of Figure 13 show that the difference between
PfE values obtained by 5×104 and 107 samples is also trivial.

5.2.2 Three-bay, ten-story steel frame

The ten-story steel frame is designed for optimal weight using BA meta-heuristic. In the
optimization process a population of 40 bats is considered and the maximum number of
generations is limited to 500. The limits on the lateral inter-story drift at various
performance levels are considered as the design constraints of this example. The results of
optimization are given in Table 9.

Table 9. Results of optimization for three-bay ten-story steel frame

Design variables Optimal sections
1 W21×55
2 W21×48
3 W24×55
4 W21×44
5 W21×44
6 W24×55
7 W30×99
8 W24×55
9 W21×55
10 W27×94
11 W21×55
12 W30×90
13 W16×36
14 W16×26
15 W16×26
Weight (fS) 21033.91 kg
24 S. Gholizadeh, V. Aligholizadeh

As well as the previous examples, in this example also the constraint associated with IO
level dominates the design.
A database including 10000 samples is generated and from the generated database 9000
and 1000 samples are employed for training and testing the NN models, respectively. RBF
and BP models are trained and due to better accuracy of BP with 5 hidden layer neurons,
only the results of this model are discussed. The MAPEs of the predicted  max IO
,
 max
LS
and  max
CP
are respectively equal to 1.5033%, 1.6952% and 1.7395%. These results
demonstrate good accuracy of the BP model. To implement MCS, 107 vectors of INN,S are
generated and the BP model is employed to predict their corresponding ONN. The values of
PfE are calculated for various performance levels and it is observed that for LS and CP
levels PfE =0. The values of PfE are given in Figure 14 for IO level.

Figure 14. Total exceedence probability of limit state functions for the optimum ten-story steel
frame at IO performance level

It is clear that the optimum ten-story steel frame is highly vulnerable against the
uncertainties at IO level. The results of Figure 14 show again that the difference between PfE
values obtained by 5×104 and 107 samples is trivial.

6. CONCLUSIONS

The main aim of the present paper is to assess the seismic reliability of performance-based
optimally seismic designed RC and steel moment frames by a combination of MCS and NN
models. In order to achieve this purpose, two RC structures including six and nine story
frames and two steel structures including three and ten story frames are optimized based on
PBD criteria using BA meta-heuristic. During the optimization process the lateral inter-story
drifts are checked at IO, LS and CP performance levels as the design constraints. For all
examples, the constraint associated with IO level dominates the design. In order to assess
the seismic reliability of the optimum structures, a database is generated in the case of each
example. The generated data is divided into training and testing sets. RBF and BP models
are trained and the results reveal that for all examples the computational performance of BP
is better than that of the RBF model. The trained BP model is effectively employed to
implement MCS. In this case, 107 input vectors are generated and their corresponding
 NEURAL NETWORK-BASED RELIABILITY ASSESSMENT OF OPTIMALLY... 25

outputs are predicted by the trained BP model. These results indicate that the optimum RC
frames are highly vulnerable against the existing uncertainties at IO and LS performance
levels with the mean exceedence probability of about 0.8 and 0.3, respectively. For the
optimum steel frames, the mean exceedence probability for IO level is about 0.4 while good
safety is observed for these structures at LS and CP performance levels. The results indicate
that the optimum performance-based RC and steel moment frames are highly vulnerable
against existing uncertainties of structural capacity and seismic demands. Therefore,
implementation of seismic reliability-based optimization for these structures is necessary to
have seismically reliable and safe structures.

REFERENCES

1. FEMA-356. Prestandard and commentary for the seismic rehabilitation of buildings.

Washington DC: Federal Emergency Management Agency, SAC Joint Venture; 2000.
2. SEAOC Vision 2000 Committee. Performance based seismic engineering of buildings.
Sacramento, California, USA: Structural Engineers Association of California; 1995.
3. Applied Technology Council, ATC-40. Seismic evaluation and retrofit of concrete
buildings. California Seismic Safety Commission; 1997.
4. Kwon OS, Elnashai A. The effect of material and ground motion uncertainty on the
seismic vulnerability of RC structure, Eng Struct, 2006; 28: 289–303.
5. Ramamoorthy SK, Gardoni P, Bracci JM. Probabilistic demand models and fragility
curves for reinforced concrete frames, J Struct Eng, 2006; 132: 1563–72.
6. Buratti N, Ferracuti B, Savoia M. Response surface with random factors for seismic
fragility of reinforced concrete frames, Struct Safe, 2010; 32: 42–51.
7. Moller O, Foschi RO, Quiroz LM, Rubinstein M. Structural optimization for
performance-based design in earthquake engineering: applications of neural networks,
Struct Safe, 2009; 31: 490–9.
8. Moller O, Foschi RO, Rubinstein M, Quiroz LM. Seismic structural reliability using
different nonlinear dynamic response surface approximations, Struct Safe, 2009; 31:
432–42.
9. Khatibinia M, Fadaee MJ, Salajegheh J, Salajegheh E. Seismic reliability assessment of
RC structures including soil–structure interaction using wavelet weighted least squares
support vector machine, Reliab Eng Syst Safe, 2013; 110: 22–33.
10. Lagaros ND, Plevris V, Papadrakakis M. Neurocomputing strategies for solving
reliability-robust design optimization problems, Eng Comput, 2010; 27: 819–40.
11. Shooman ML. Probabilistic reliability: an engineering approach. McGraw-Hill; 1968.
12. Yang XS. A new metaheuristic bat-inspired algorithm. In: Gonzalez JR et al. (eds)
Nature inspired cooperative strategies for optimization (NISCO 2010). Studies in
Computational Intelligence, vol 284. Springer, Berlin, pp. 65–74, 2010.
13. Gandomi AH, Yang XS, Alavi AH, Talatahari S. Bat algorithm for constrained
optimization tasks, Neural Comput Appl, DOI 10.1007/s00521-012-1028-9.
14. Fragiadakis M, Lagaros ND. An overview to structural seismic design optimisation
frameworks, Comput Struct, 2011; 89: 1155–65.
26 S. Gholizadeh, V. Aligholizadeh

15. FEMA-440. Improvement of Nonlinear Static Seismic Analysis Procedure Federal

Emergency Management Agency Washington, D.C., 2005.
16. McKenna F, Fenves GL. The OpenSees Command Language Manual (1.2. edn). PEER,
2001.
17. Gholizadeh S, Aligholizadeh V. Optimum design of reinforced concrete frames using
bat meta-heuristic algorithm, Int J Optim Civil Eng, 2013; 3: 483–97.
18. Gholizadeh S, Kamyab R, Dadashi H. Performance-based design optimization of steel
moment frames, Int J Optim Civil Eng, 2013; 3: 327–43.
19. Vanderplaats GN. Numerical Optimization Techniques for Engineering Design: With
Application, McGraw-Hill, New York, 1984.
20. American Concrete Institute (ACI). Building code requirements for structural concrete
and commentary. ACI 318–08, 2008.
21. Manual of steel construction. Load and resistance factor design, Chicago, IL: American
Institute of Steel Construction, 2001.
22. Carbas S, Hasancebi O. Optimum design of steel space frames via bat inspired
algorithm, 10th World Congress on Structural and Multidisciplinary Optimization,
Florida, USA, 2013.
23. Kent DC, Park R. Flexural members with confined concrete, J Struct Div, ASCE, 1997;
97: 1969–90.
24. Gholizadeh S, Salajegheh J, Salajegheh E. An intelligent neural system for predicting
structural response subject to earthquakes, Adv Eng Softw, 2009; 40: 630−9.
25. Wasserman PD. Advanced methods in neural computing, New York: Prentice Hall
Company, Van Nostrand Reinhold, 1993.
26. Hagan MT, Demuth HB, Beal MH. Neural network design, PWS Publishing Company,
Boston, 1996.
27. Hagan MT, Menhaj M. Training feed-forward networks with the Marquardt algorithm,
IEEE Trans Neural Network, 1999; 5: 989−93.
28. The Language of Technical Computing, MATLAB. Math Works Inc, 2009.
29. FEMA-National Institute of Building Sciences, HAZUS-MH MR1, Multi-hazard Loss
Estimation Methodology Earthquake Model, Washington, DC, 2003.
30. Mckay MD, Beckman RJ, Conover WJ. A comparison of three methods for selecting
values on input variables in the analysis of output from a computer code, Tech, 1979;
21: 439–45.
31. Webpage: http://iranhazard.mporg.ir/PSHA.aspx.
32. Kaveh A, Farahmand Azar B, Hadidi A, Rezazadeh Sorochi F, Talatahari S.
Performance-based seismic design of steel frames using ant colony optimization, J
Constr Steel Res, 2010; 66: 566–74.