Struct Multidisc Optim (2017) 55:63–75

DOI 10.1007/s00158-016-1471-9

RESEARCH PAPER

A self-adaptive modified chaos control method

for reliability-based design optimization
Behrooz Keshtegar 1 & Peng Hao 2 & Zeng Meng 3

Received: 17 October 2015 / Revised: 11 February 2016 / Accepted: 22 April 2016 / Published online: 2 May 2016
# Springer-Verlag Berlin Heidelberg 2016

Abstract In the reliability-based design optimization 1 Introduction

(RBDO), the Advanced mean value (AMV) method some-
times yields unstable results such as chaotic and periodic Engineering structures inherently have uncertainties in mate-
solutions for highly nonlinear probabilistic constraints. The rial properties, manufacturing tolerances as well as external
chaos control (CC), modified chaos control (MCC) and adap- loads, etc. Therefore, optimization considering uncertainties
tive chaos control (ACC) methods are more robust than the becomes a modern design methodology (Chen et al. 2013;
AMV but inefficient for some moderately nonlinear perfor- Li et al. 2015). Reliability-based design optimization
mance functions. In this paper, a self-adaptive modified chaos (RBDO) model is developed to evaluate the probabilistic con-
control (SMCC) method is developed based on a dynamical straints based on reliability analysis (Du and Chen 2004; Qu
control factor to improve the efficiency of MCC for reliability and Haftka 2004; Liang et al. 2008; Cheng et al. 2006), which
analysis and RBDO. The self-adaptive control factor is dy- can provide an improved design, and also a higher level of
namically computed based on the new and previous results. confidence.
The efficiency and robustness of the proposed SMCC are The evaluation of probabilistic constraints is very impor-
compared with the AMV, CC, MCC and ACC methods using tant in the RBDO model. Therefore, two different approaches
several nonlinear structural/mathematical performance func- have been developed, including the reliability index approach
tions and RBDO problems. The results illustrate that the (RIA) (Enevoldsen and Sørensen 1994) and the performance
SMCC is more efficient than CC, MCC, and ACC methods, measure approach (PMA) (Tu et al. 1999). It has been
and also more robust than AMV method for highly nonlinear demonstrated that the PMA has higher efficiency and robust-
problems. ness when evaluating the probabilistic constraints in compar-
ison with the RIA (Yang and Yi 2009). Typically, the ad-
vanced mean value (AMV) method is widely applied in
Keywords Reliability-based design optimization . Advanced RBDO problems due to its simplicity and efficiency (Cheng
mean value . Self-adaptive modified chaos control method . et al. 2006; Yang and Yi 2009). However, the iterative AMV
Dynamical self-adaptive control factor scheme exhibits unstable results, such as chaotic and periodic
solutions, when nonlinear performance functions are involved
* Behrooz Keshtegar (Youn and Choi 2004; Meng et al. 2015). Hence, several im-
[email protected] proved algorithms were proposed to enhance the robustness of
AMV, including the conjugate mean value (CMV) method
1
Department of Civil Engineering, University of Zabol, P.B.
(Youn et al. 2003), hybrid mean value (HMV) method
9861335-856, Zabol, Iran (Youn and Choi 2004), enhanced hybrid mean value
2
State Key Laboratory of Structural Analysis for Industrial
(EHMV) method (Youn et al. 2005a, b), chaos control (CC)
Equipment, Department of Engineering Mechanics, Dalian method (Yang and Yi 2009), adaptive chaos control (ACC)
University of Technology, 116023 Dalian, China method (Li et al. 2015), modified chaos control (MCC) meth-
3
Hefei University of Technology, Hefei University of Technology, od (Meng et al. 2015), and conjugate gradient analysis (CGA)
Hefei 230009, People’s Republic of China method (Ezzati et al. 2015).
64 B. Keshtegar et al.

The CMV and HMV methods show good performances 2 RBDO model
in comparison with AMV for concave performance mea-
sure functions. The HMV and EHMV methods improve the The mathematical model of RBDO can be formulated as
numerical instability of AMV-based MPTP search using (Youn et al. 2003; Youn and Choi 2004; Li et al. 2015)
the convexity information of performance functions.
find d ; min f ðd Þ
However, for highly nonlinear performance functions,
S:t: P f ðgi ðd; xÞ≤ 0Þ ≤ Pi;ft ði ¼ 1; 2; …; pÞ ð1Þ
these iterative methods (i.e., AMV, CMV, and HMV) fail
dL ≤ d ≤ dU
to converge due to the numerical difficulties in the forms of
periodic or chaotic solutions. The chaos feedback control where, d = [d1, d2, …, dN]T is the design variable vector,
was implemented to improve the robustness of AMV-based representing the deterministic physical quantities with upper
MPTP search for highly concave problems (Yang and Yi bound dU and lower bound dL; x = [x1, x2, …, xM]T is the ran-
2009), where the stability transformation method of chaos dom variable vector, representing the uncertain quantities;
control for PMA reveals the inherent reason of numerical gi(d, x) is defined as the ith performance function, and the
instabilities such as periodic and chaotic solutions in AMV probabilistic constraints are described by the performance
iterative scheme, and provides a new idea for convergence function gi subject to uncertainty X. gi(d, x) ≤ 0 denotes the
control with solid mathematical theory. Nevertheless, the failure domain. Pi,ft is the target failure probability, which
CC method is computationally inefficient for both concave can be computed as (Yang and Yi 2009; Tu et al. 1999):
and convex problems. Recently, the ACC and MCC Z
methods were proposed to enhance the efficiency of CC Ρft ¼ Ρ½g ðd; xÞ≤ 0 ¼ f X ðxÞdX ð2Þ
based on an adaptive line search in ACC (Li et al. 2015)
gðd;xÞ ≤ 0
and a modified line search in MCC (Meng et al. 2015) for
concave reliability problems. The MCC can be converged where, fX(x) is the joint probability density function of the
by implementing more iterations for convex or moderately basic random variables X. The first-order reliability method
nonlinear performance functions, and the ACC has almost (FORM) (Yang 2010; Keshtegar and Miri 2014a, b) has been
a same efficiency level as MCC for highly nonlinear prob- developed to estimate integration of (2), which is commonly
lems in RBDO. Although the ACC and MCC are robust recognized to have adequate accuracy and thus widely used in
methods to evaluate the probabilistic constraints, a large various RBDO problems (Hasofer and Lind 1974). Generally,
number of iterations are required to achieve stable con- the main goal of the FORM is to search for the most proba-
verged solutions for moderately nonlinear performance bility failure point (MPFP, i.e. β = ‖U*‖) by transforming the
functions. basic random variables from original space (X-space) into
In this paper, a self-adaptive control factor is introduced standard normal U-space (U is normal variable with zero
for concave and convex probabilistic performance func- means, unit variance and independent components) i.e.,
tions in RBDO, aiming to improve the efficiency of u = Φ− 1{FX(x)} (Yang and Yi 2009). The target failure proba-
MCC method. To be specific, this self-adaptive control bility can then be approximated as (Xiao et al. 2011)
factor is computed using the new and previous iteration   
results, which is limited by a threshold value that is adap- Ρft ≈1−Φðβ t Þ ¼ Φ −U *  ð3Þ
tively adjusted based on the gradient of performance func-
tion at the mean value. The modified chaos control with where, Φ is the standard normal cumulative distribution func-
self-adaptive control factor is named as the self-adaptive tion. Therefore, the probabilistic constraint of the RBDO mod-
MCC (SMCC) in this study. The SMCC is more efficient el in (1) can be rewritten by two alternative relations (Youn
than the CC, MCC and ACC with a constant control et al. 2003):
factor for highly nonlinear performance functions. This     
β i;t ≈−Φ−1 Pi;ft ≤ βi ðd Þ≈ −Φ−1 P f ðgi ðd; xÞ≤ 0Þ ð4Þ
paper is structured as follows: The basic RBDO model    
is reviewed in Section 2. The PMA is then defined for gi ðd Þ ¼ F −1
g i d; Φ −β i;t ≥0 ð5Þ
RBDO in Section 3. In Section 4, several methods for the
evaluation of probabilistic constraints including the AMV, where, βi,t and gi(d) are the safety reliability index and prob-
CC, MCC and ACC are summarized. After that, the pro- abilistic performance measure for the ith probabilistic con-
posed SMCC method is introduced in Section 5. The ac- straint, respectively. F gi is the cumulative distribution func-
curacy, robustness and efficiency of SMCC are compared tion of the performance function gi, which can be computed as
with the AMV, CC, MCC and ACC methods using sev-
Z
eral nonlinear mathematical/structural reliability and  
Ρ½g ðd; xÞ≤ 0 ¼ F gi ðd; 0Þ ¼ f X ðxÞdX ≈Φ −βi;t ð6Þ
RBDO examples in Section 6. Finally, conclusions are
drawn in Section 7. gðd;xÞ ≤ 0
A self-adaptive modified chaos control method 65

In general, two approaches including the RIA and PMA are been demonstrated that the HL–RF could converge to period-
applied to perform the reliability analysis in RBDO problems. ic, bifurcation or chaotic solutions (Yang 2010; Keshtegar and
The RIA is an iterative mathematical formula of first-order Miri 2013). However, the PMA is computationally more effi-
reliability method, and the inner loop of RBDO model in (1) cient and stable than the RIA (Youn et al. 2005b), in particular,
is the reliability analysis loop using e.g. Hasofer-Lind (1974) it has been found that the PMA has higher efficiency and
and Rackwitz-Fiessler (1978) (HL–RF) method. The main robustness in comparison with the RIA when evaluating the
propose of the RIA is to search for the MPFP in FORM, which probabilistic constraints in RBDO (Lee et al. 2002).
is the nearest point on the limit-state function to the origin in
the U-space. In the PMA, the probabilistic constraint is eval- 3.2 First-order reliability analysis for PMA
uated by searching for the MPTP on the target reliability sur-
face (Lee et al. 2002; Yang and Yi 2009; Meng et al. 2015). The RBDO mathematical model can also be defined using the
PMA by the following relation (Youn et al. 2005a; Li et al.
2015):
3 Reliability analysis in RBDO
find d ; min f ðd Þ
Structural reliability analysis is an important tool to estimate S:t: gi ≥ 0 ði ¼ 1; 2; …; pÞ ð9Þ
the failure probabilities in RBDO problems. The point U* is dL ≤ d ≤ dU
searched based on the reliability analysis loop of RBDO,
The first-order probabilistic performance measure is ob-
which can be obtained by three steps as follows (Keshtegar
tained by a nonlinear optimization formulation as (Youn
and Miri 2014b):
et al. 2003):
Step 1: Transform the original random variables into stan- given d and find U * ; min gðd; U Þ
ð10Þ
dard normal variables with zero means, unit variance S: t: kU k ¼ βt
and independent components.
Step 2: Find the MPTP or MPFT on the basis of an iterative where U* is the MPTP on the target reliability sphere surface
process. with prescribed reliability index βt. In the PMA, the iterative
Step 3: Determine the performance function i.e. g(d, U*) or algorithm to search for the MPTP is an essential key point to
reliability index i.e. β = ‖U*‖. the convergence of RBDO problems.

(4) can be employed to evaluate the probabilistic con-

straints in (1) using the RIA. And also, (5) can be applied to 4 MPTP search methods
describe the probabilistic constraints in (1) by the performance
measure function, which is referred to be the PMA. 4.1 Advanced mean value (AMV) method

3.1 First-order reliability analysis for RIA The AMV iterative algorithm is widely utilized for computing
the MPTP as:
The RBDO model is defined based on the RIA for evaluating  
the probabilistic constraints as (Lee et al. 2002; Youn et al. U AM
kþ1
V
¼ βt n uAM k
V

2005a):   ∇u g ðd; U k Þ ð11Þ

n uAM V
¼−
k
k∇u gðd; U k Þk
find d ; min f ðd Þ
S:t: β it ≤ βi ðd Þ ði ¼ 1; 2; …; pÞ ð7Þ where, n(uAMV
k ) stands for the normalized steepest descent
dL ≤ d ≤ dU direction. The AMV iterative formula is simple and efficient
(Tu et al. 1999), however, its sequential iterations may lead to
The reliability index βi(d) for the ith probabilistic constraint unstable solutions for complicated and highly nonlinear per-
is computed by the following optimization problem (Youn and formance measure functions, such as periodic, bifurcation and
Choi 2004; Keshtegar and Miri 2014a): chaotic solutions (Yang and Yi 2009).
 1=2
given d and find U * ; min β ¼ U T U ð8Þ
subjected t o g ðU Þ ¼ 0 4.2 Chaos control (CC) method

where, U is the standard normal variable vector. The iterative Based on the chaos control approach, Yang and Yi (2009)
HL–RF formula of FORM is widely used to MPFP search (see extended the stability transformation method (STM) to the
Yang 2010; Keshtegar and Miri 2013, 2014a, b), but it has MPTP search by the following relations
66 B. Keshtegar et al.

 
kþ1 ¼ U k
U CC þ λ C f ðuk Þ−U CC
CC
k descent direction of the performance function at the kth point.
∇u gðd; U k Þ ð12Þ This means that the control factor is considered to be 1 in the
f ðuk Þ ¼ U AM
kþ1 ¼ −β t
V
k∇u gðd; U k Þk ACC method in this case. When sign(ςk + 1) ≤ 0, the MCC
method is implemented in ACC, but the control factor λ in
where, C is the n × n dimensional involutory matrix, which (13) is updated by the following formulation:
only has one element in each row and each column to be 1 8
or −1, and the other elements are 0. For the purpose of sim- < 0:2 λ. 0:2 θk > θk−1
plicity, the unit matrix I is commonly utilized for involutory λ ¼ λθk−1 θk θk > θk−1 ≥ 0:2 θk ð16Þ
:
matrix C in practice (Yang and Yi 2009; Yang 2010). λ is the λ θk ≤ θk−1
chaos control factor that within the range of [0, 1], and a small
value of λ should be selected to achieve stabilization. If λ = 1 where the initial value of control factor λ is considered to be
and C = I, the STM is similar to the AMV method. The CC 0.5, θk − 1 is the angle between UACC
k − 1 , θk is the angle
and UkACC
ACC AMV
method would become inefficient when a small control factor between Uk and Uk + 1, which are utilized to update the con-
λ < < 1 is selected to achieve stable convergence for both con- trol factor λ. The ACC improves the efficiency of the MCC for
cave and convex performance measure functions. convex performance functions based on the adapted control
factor in (16).
4.3 Modified chaos control (MCC) method

The MCC method was proposed by Meng et al. (2015) to 5 A self-adaptive modified chaos control (SMCC)
improve the efficiency of the CC, and the formulation is
defined as According to the above section, the iterations of the CC and
  MCC methods are highly dependent on the control factor. The
~
nkþ1 ¼ U M
k
CC
þ λC f ðuk Þ−U M
k
CC
ð13Þ CC and MCC may fail for highly nonlinear performance func-
tions when a large control factor is selected. In addition, the
~
n
UM CC  kþ1 
kþ1 ¼ β t  ð14Þ CC and MCC require more iterations to achieve stabilization

~
nkþ1  for convex performance functions when a small control factor
is selected. In this paper, a self-adaptive control factor is in-
where troduced to improve the efficiency of MCC.

∇u g ðd; U k Þ
f ðuk Þ ¼ U AM
kþ1 ¼ −β t
V
ð15Þ 5.1 Iterative formula of SMCC
k∇u g ðd; U k Þk

where, ñk + 1 stands for the modified descent direction based The iterative formula of SMCC scheme can be expressed as
on the chaos control on performance measure functions. The S
MCC method updates the iterative point by extending it to the ~nkþ1
beta circle using (14). The MCC is more efficient than the CC, U SMCC
kþ1 ¼ βt 
 S 
 ð17Þ
~n 
and more robust than the traditional iterative algorithm i.e.  kþ1 
AMV and HMV to search for the MPTP. The sequential iter-
S  
ations of the MCC method is dependent on the control factor ~nkþ1 ¼ U SMCC
k þ λSk C f ðuk Þ−U SMCC
k ð18Þ
in 13. Consequently, a larger number of function evaluations
are required when a small control factor is selected for convex where, βt is the target reliability index, USMCC
k is the previous
or moderately nonlinear performance functions. point that is computed using iterative formula of SMCC, ñSk + 1
stands for the enhanced descent direction based on the self-
adaptive control factor (λSk ). The self-adaptive control factor is
4.4 Adaptive chaos control (ACC) method
more important to achieve stabilization with an efficient iter-
ation for iterative formula in (17). Therefore, a self-adaptive
The ACC method was proposed by Li et al. (2015) based on an
control factor (λSk ) is introduced to achieve stabilization by a
adaptive control factor, in which the control factor is adapted
dynamical control factor and then enhance the efficiency by a
according to the convexity type of the performance function.
threshold value as follows:
The control factor of MCC in (13) is selected in terms of the
iterative angle during the iterative process. In the iterative for- λSk ¼ minfλint ; λk g ð19Þ
mula of ACC, when the performance function is convex i.e.
sign(ςk + 1) > 0, the AMV method is used as UACC AMV
k + 1 = Uk + 1 , where, λk is a dynamical control factor, λint is the initial con-
where ςk + 1 = (nk + 1 − nk) ⋅ (nk − nk − 1), and nk is the steepest trol factor or maximum control factor, which is suggested as
A self-adaptive modified chaos control method 67

(   )
1 gðd; U 1 Þ
λint ¼ min 1− ;1 ð20Þ
δβ 2t gðd; U 0 Þ

where 0 < δ < 1. g(d, U0) and g(d, U1) are the performance
functions at the point U0 = μ and U 1 ¼ −β t k∇∇uu ggððd;μ Þ
d;μÞk . The
maximum (initial) control factor (λint) is adaptively computed
based on the initial points U0 = μ and U1 by the performance
values g(d, U1) and g(d, U0) in (20), where the nonlinearity
degree of performance function can be considered. It is evi-
dent from (20) that the self-adaptive control factor is limited
based on the initial control factor as λSk ≤ λint. The value of
maximum control factor is adjusted by the factor δ. For
δ = 1, the initial control factor would be less than 1 for a linear
or convex performance function. Consequently, a smaller val-
ue of δ can provide a larger initial control factor. This means Fig. 1 MPTP search for the SMCC method
that a large control factor can be selected for convex problems,
which can improve the efficiency of MCC for reliability anal- control factor λ in MCC for the initial iterations. Therefore,
ysis and RBDO problems. For highly nonlinear performance less iterations are needed using SMCC to achieve stable re-
functions, the initial control factor may be computed as a large sults for moderately nonlinear performance functions. The
value, thus the proposed iterative formula in (17) produces MPTP search based on iterative methods usually yields stable
unstable solutions. Therefore, the dynamical control factor solutions when the spectral radius of the Jacobian matrix is
 
can be used to achieve the stabilization of SMCC by the fol- smaller than 1, i.e. ρ J < 1 (Yang 2010). The spectral radius
lowing relation: for the Jacobian matrix is determined by the self-adaptive
 
"  # control factor as J ¼ I þ λSk ð½ J −I Þ where
2 g d; U SMCC
λk ¼ 1−  k
 k ≥1 ð21Þ ½ J  ¼ ∂ ∂u
f i ðuÞ
jU k . The self-adaptive control factor in SMCC
kDk k2 g d; U SMCC
k−1 j

can be dynamically adjusted based on the nonlinearity degree

where, g(d, USMCC
k ) and g(d, UkSMCC
− 1 ) are the performance of performance function as well as the results from the new
values at pervious point of SMCC. Dk is the line search vector, and previous iterations. Consequently, a suitable control factor
which is computed using the new point of AMV and previous is obtained as λSk = min{λint, λk} ≤ 1, thus the spectral radius of
point of SMCC as the Jacobian matrix can be improved based on λSk . Therefore,
the SMCC can achieve stable convergence even for highly
Dk ¼ f ðuk Þ−U SMCC ð22Þ  
k nonlinear performance functions (ρ J < 1 ) by selecting a
The self-adaptive control factor in (19) is dynamically ad- suitable parameter δ. Thus, the efficiency and robustness of
justed in each iteration by the control factor in (21). A small the SMCC method are dependent on the parameter δ.
control factor was computed at final iterations due to the small
difference between g(d, USMCCk ) and g(d, UkSMCC
− 1 ). The largest 5.2 Algorithm of the SMCC for MPTP search
SMCC
difference between g(d, Uk ) and g(d, UkSMCC ) is deter-
 − 1
gðd;U SMCC Þ The SMCC method would produce unstable results in the
mined at initial iterations, thus 1− g d;U kSMCC may be more
ð k−1 Þ form of periodic and chaotic solutions for nonlinear perfor-
than 1. This means that there is a significant difference be- mance functions, when the self-adaptive control factor is too
tween f(uk) and USMCC
k , thus a large value is computed for Dk. large. On the other side, if the selected control factor is too
Therefore, the dynamical step size is controlled for highly small, this method would be converged slowly. The self-
nonlinear performance functions. A dynamical control factor adaptive control factor has the largest value at the beginning
may be computed to be larger than 1 when a smaller Dk is of the iteration procedure, thus the SMCC method is con-
determined since f(uk) ≈ USMCC k . Consequently, the iterative verged faster than the MCC. On the other hand, the self-
formula yields unstable solutions (i.e., chaotic and periodic adaptive control factor becomes smaller at the final iterations
solutions), thus the self-adaptive control factor is limited by due to the fact that f(uk) ≈ USMCC
k , thus a fixed point is obtain-
the initial control factor in (20) to control the unstable results ed. The flowchart of the SMCC based on the self-adaptive
of SMCC. control factor is shown in Fig. 2. The iterative procedure of
The iterative procedure of SMCC is plotted in Fig. 1. The the SMCC method is described by the following steps to eval-
control factor in (19) is computed to be larger than the constant uate the performance function:
68 B. Keshtegar et al.

Fig. 2 Flowchart of the SMCC

method

Step 1: Define performance function g(d, X) and βt, μ and σ search can also be established based on an empirical formula
for random variables. Set k = 0, initialize ε (stopping as well as the ACC (Li et al. 2015).
criterion) and δ
Step 2: Transform random variables by u = Φ− 1{FX(x)}
Step 3: Compute a new point based on the iterative AMV
(f(uk) = UAMV
k+1) 6 Illustrative examples
Step 4: For k = 0, compute the initial control factor as
n h i o
The possibility of applying the SMCC to evaluate the
λint ¼ min δβ1 2 1− ggððd;U 1Þ
d;U 0 Þ ; 1
t probabilistic constraints in RBDO has been investigated
Step 5: For k ≥ 1 , compute the line search vector as through several nonlinear problems with concave and con-
Dk = f(uk) − USMCC
k vex performance functions. The efficiency of proposed
and determine the dynamical step size as
  SMCC method is studied using mathematical and
g ðd;U SMCC Þ structural/mechanical RBDO-based PMA problems. The
λk ¼ kD k2 1− g d;U SMCC
2 k
k ð k−1 Þ AMV, CC, MCC, ACC and SMCC methods are imple-
Step 6: Compute the self-adaptive control factor as mented by MATLAB 7.10 program. The numbers of com-
λSk = min{λint, λk} putations of ∇g(U) and performance value are compared.
Step 7: Determine ñSk + 1 = USMCC
k + λSk C[f(uk) − USMCC
k ]; The convergence criterion is set as ‖U k + 1 − U k ‖/
Obtain the new point based on SMCC as ‖Uk‖ < 10− 6 for all examples in the reliability analysis
k + 1 = βtñk + 1/‖ñk + 1‖
USMCC S S
loop.
Step 8: Set the next iteration as Xk + 1 = T(USMCCk+1 )
k + 1 − Uk
Step 9: If ‖USMCC SMCC
‖/‖USMCC
k ‖ < ε, then stop and
print results; Else, k = k + 1 and Go to Step 2. 6.1 Nonlinear performance examples

The major difference between the SMCC and other existing Five nonlinear performance measure functions are used to
reliability methods lies in the self-adaptive control factor in compare the efficiency of the AMV, CC, MCC, ACC and
(19). The self-adaptive control factor can be adaptively adjust- SMCC. To be specific, the control factor λ for the CC and
ed based on the results from the new and previous iterations MCC methods is set as 0.1, and the initial control factor is set
without line search rules. In addition, the self-adaptive line as 0.5 for the ACC method. The parameters δ is set as 0.5 for
search to improve the robustness of reliability-based MPTP the SMCC.
A self-adaptive modified chaos control method 69

Example 1: g = x31 + x32 − 18 in which, x1 ~ N(10, 5) (N(μ, σ)), Example 4, the SMCC method is significantly more efficient
x2 ~ N(9.9, 5) and βt = 3.0 (Yang and Yi 2009) than the MCC. For convex performance functions, the MCC
Example 2: g = [exp(0.8x1 − 1.2) + exp(0.7x2 − 0.6) − 5]/10 in is as efficient as the CC method, but the AMV and ACC can
which, x1 ~ N(4, 0.8), x2 ~ N(5, 0.8) and βt = 3.0 (Youn et al. converge to stable results more efficiently. As can be seen, the
2005a) SMCC is as robust as the ACC, but is converged about twice
Example 3: g = 0.3x21x2 − x2 + 0.8x1 + 1 in which, x1 ~ N(1.2, faster than the ACC method for Examples 3 and 5.
0.42), x2 ~ N(1.0, 0.42) and βt = 6.0 (Youn et al. 2005a) In addition, Example 4 has a small curvature at the MPTP.
Example 4: g = 10 − exp(x1 − 7) − x2 in which, x1, x2 ~ N(6, 0.8) λ = 0.1 is rigorous for the CC and MCC methods in these
and βt = 3.0 (Youn et al. 2003) problems with small curvature. The control factor should be
Example 5: g = 0.489x 3 x 7 + 0.843x 5 x 6 − 0.0432x 9 x 10 + increased properly to reduce the number of function evalua-
0.0556x 9 x 11 + 0.000786x 211 − 0.75, in which, i = 1 ~ 7 tions, which is appropriately computed using the self-adaptive
x i ~ N(1, 0.005), i = 8 ~ 9 x i = 9 ~ 10 x i ~ N(0, 10.0) and control factor in the SMCC method. Therefore, the proposed
βt = 3.0 (Yang and Yi 2009) iterative formula based on the SMCC method is robust and
efficient, which can be used to search for the MPTP of prob-
The converged value of the performance measure functions abilistic constraints with both large and small curvatures in
and iterations to obtain the stable solutions of all algorithms RBDO, accurately and rapidly.
(i.e. the AMV, CC, MCC, ACC and SMCC methods) are Moreover, the effects of parameter δ of the SMCC method
listed in Table 1. It can be found that the AMV method does are investigated, and the required number of iterations to
not converge to a stable MPTP for highly nonlinear perfor- achieve the stabilization are listed in Table 2. As can be seen,
mance functions 1–3 and 5 as chaotic, periodic-2, periodic-18, a large number of iterations are required when δ < 0.2 for
and periodic-2 solutions, respectively. Therefore, it can be highly nonlinear performance functions (Examples 1–3). For
concluded that the AMV is not a robust method for highly convex performance functions (Example 4), a less number of
nonlinear problems. The CC is converged to stable results iterations is needed when δ < 0.3. The results from Table 2
accurately, however, it is the most inefficient among all these show that the robustness and efficiency of the SMCC are
methods but more robust than the AMV. The CC is inefficient depended on the parameter δ for Example 5. A unstable result
for both concave (Examples 1–3) and convex (Example 4) as chaotic solutions is obtained when δ ≤ 0.2, and this means
performance functions. that the initial control factor is computed to be larger than the
For highly nonlinear concave performance functions (i.e. critical control factor, which is prone to obtain a unstable
Examples 1–3 and 5), the MCC is more efficient than the CC result. Therefore, this approach requires some modifications
method, while the ACC is slightly more efficient than the to achieve a global convergence, and a line search rule such as
MCC and CC methods. It is obvious that the proposed ap- Armijo line search (Keshtegar and Miri 2014b), merit function
proach based on the self-adaptive control factor is converged (Santos et al. 2012), Wolf condition (Keshtegar and Miri
faster than the other methods for highly nonlinear perfor- 2014a), or sufficient descent (Keshtegar 2016) can be imple-
mance functions in Examples 1, 3 and 5. Besides, the mented to guarantee global convergence, theoretically.
SMCC is more robust than the AMV and more efficient than The required number of iterations to obtain stable results
the MCC and CC methods for both concave and convex per- are also shown in Table 2 for the parameter δ in the range of
formance functions. 0.8–0.95. It is evident that the SMCC method performs well
The iterative histories of the MCC, ACC and SMCC with respect to the δ in the range between 0.4 and 0.6.
methods are plotted in Fig. 3 for Examples 1, 2 and 4. It is
shown that the SMCC is more efficient than the MCC and Example 6: a steel frame under lateral load P, and distributed
ACC for concave functions (Examples 1 and 2). For load W are plotted in Fig. 4. The performance

Table 1 Iterative results for

different PMA iterative methods Example Method

AMV CC MCC ACC Proposed SMCC

#1 Chaos −31.06605 (152) −31.06647 (28) −31.06647 (15) −31.06647 (10)

#2 Period-2 0.20383 (111) 0.20382 (45) 0.20382 (9) 0.20382 (9)
#3 Period-18 −2.22932 (113) −2.22934 (30) −2.22934 (36) −2.22934 (16)
#4 −0.35794 (12) −0.35792 (155) −0.35794 (140) −0.35794 (12) −0.35794 (15)
#5 Period-2 0.075285 (113) 0.075282 (84) 0.075282 (47) 0.075282 (27)
70 B. Keshtegar et al.

Fig. 3 Iterative histories of MPTP search based on the MCC and SMCC methods

function is expressed as (Keshtegar 2016): (L) and height (H). The statistical properties of
the random variables are listed in Table 3, and the
g ¼ 0:05−Δ ð23Þ target reliability index is given as βt = 3.0.

where, Δ is the displacement of frame roof, The converged results of performance value using the AMV,
which is computed based on a finite element CC, MCC, ACC and SMCC methods are compared in Table 4.
model under loads P and W, mechanical proper- It is obvious that the SMCC converges to stable results, while
ties E1 (elastic modulus of columns) and E2 (elas- the AMV, CC, MCC and ACC methods yield chaotic results.
tic modulus of beam), and dimensions i.e. cross- Therefore, it can be concluded that the proposed modified line
sections (A1–A3), moments of inertia (I1–I3), span search based on the self-adaptive control factor can provide an
A self-adaptive modified chaos control method 71

Table 2 Required number of iterations to achieve stabilization for the This example involves two random variables and three
SMCC method with various δ
probabilistic constraints. The random variables (x1, x2)
Example δ are statistically independent and follow normal (N),
Lognormal (LN), Weibull (W), and Gumbel (G) distribu-
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.95 tions. Two standard deviations, including 0.5 (βt = 3.0) and
1 81 21 17 8a 10 13 16 19 22 23
1.0 (βt = 2.5), are considered to verify the proposed SMCC
method.
2 24 21 17 7 9 12 15 17 20 21
The optimum results are captured based on the AMV, CC,
3 136 41 46 12 16 19 23 26 29 31
MCC, ACC and SMCC methods for each distribution func-
4 11 11 11 11 15 20 23 27 31 32
tion of random variables. The optimum design variables, op-
5 Chaos Chaos 83 33 27 38 44 49 54 57
timum objective function and probabilistic constraints g1 − g3
a
Bold numbers are the best iteration among different δ for each example are listed in Tables 5 and 6 for case 1 (βt = 3.0 and σ = 0.5) and
case 2 (βt = 2.5 and σ = 1.0), respectively. It can be seen that
appropriate solution without chaos in comparison with other the AMV approach in PMA to find the design variables yields
existing reliability-based MPTP search methods. unstable solutions in case 1 with Gumbel distribution and case
2 with all distribution types. The CC, MCC, ACC and SMCC
methods are accurately converged to same results in cases 1
6.2 RBDO problems and 2 with all distribution types. The CC, MCC and SMCC
methods are more robust than the AMV, and the proposed
In order to verify the accuracy and efficiency of the proposed SMCC method is slightly more efficient than the CC and
SMCC method, two numerical and structural examples are MCC methods for this example. The ACC method is more
used to test the RBDO-based PMA methods. The AMV, efficient than the MCC and CC but more inefficient than the
CC, MCC, ACC and SMCC methods are utilized to find the SMCC for almost all distribution types in cases 1 and 2. By
MPTP in PMA. The control factor λ is set as 0.1 for the CC comparing the ACC and SMCC methods, it can be found that
and MCC methods and 0.5 for the ACC method. The param- the constraint-3 is evaluated with less iterations to obtain the
eter δ of the SMCC method is set as 0.4 based on the results optimum based on the ACC, while the constraint-2 is effi-
from Table 2 for all examples. ciently evaluated based on the SMCC for almost all distribu-
tion types. The SMCC is converged about 7-times, 4-times
Example 7: A mathematical RBDO example is given as fol-
faster than the MCC method in cases 1 and 2 with Normal
lows (Cho and Lee 2010; Chen et al. 2013):
and Weibull distributions, respectively. Thus, it can be con-
Find d ¼ ½d 1 ; d 2  T
min f  ðd Þ ¼ d1 þ d2
cluded that the proposed SMCC can be converged accurately
x21 x2   and efficiently in comparison with other PMA approaches for
S : t : P f g 1 ¼ 1− >0 ≤ Φ −β1;t
" 20 # this RBDO problem.
ðx1 þ x2 −5Þ2 ðx1 −x2 −12Þ2  
P f g2 ¼ 1− − > 0 ≤ Φ −β 2;t
30 120
  Example 8: A conical structure under compressive axial load
80  
P f g 3 ¼ 1− 2 >0 ≤ Φ −β 3;t P, and bending moment M is shown in Fig. 5 (Keshtegar and
x1 þ 8x2 þ 5  
where 0≤ d i ≤ 10; xi eN ; LN ; W; and G d i ; σ2 f or i ¼ 1; 2 Miri 2013).
d 0
¼ ½5; 5 ; β 1;t ¼ β2;t ¼ β 3;t ¼ βt
ð24Þ Buckling is the failure mode of this structure, which can be
defined in the form of the following performance function.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
3ð1−v2 Þ P M
g ¼ 1−  þ ð25Þ
π E t 2 cos2 α 2γ η r1

where, P and M are the compressive axial load and bending

moment, respectively. The coefficients γ and η are used to
correct the P and M between theoretical and experimental
results, respectively, which are set to be 0.33 and 0.41 in this
example. v is the Poisson’s ratio that equals to 0.3. The cost
function is defined as

Fig. 4 A steel frame structure C ¼ πt ½2r1 þ sinðαÞ ð26Þ

72 B. Keshtegar et al.

Table 3 Statistical properties of the random variables for the steel frame example

Random variable Mean Standard deviation Distribution Random variable Mean Standard deviation Distribution

A1 (m2) 91 × 10−4 9.1 × 10−4 Lognormal E1 (kg/m2) 2.38 × 1010 1.9 × 109 Normal
I1 (m4) 80.9 × 10−6 8.09 × 10−6 Lognormal E2 (kg/m2) 2.17 × 1010 1.74 × 109 Normal
A2 (m2) 53.8 × 10−4 5.38 × 10−4 Lognormal L (m) 6 0.48 Normal
I2 (m4) 83.6 × 10−6 8.36 × 10−6 Lognormal H (m) 4.5 0.36 Normal
A3 (m2) 106 × 10−4 10.6 × 10−4 Lognormal W (kg/m) 4500 630 Gumbel
I3 (m4) 112.6 × 10−6 11.26 × 10−6 Lognormal P (kg) 2000 350 Gumbel

Table 4 Performance function

values of different reliability Iterations AMV CC MCC ACC SMCC
methods for the steel frame
example 1 0.02884962 0.02884962 0.02884962 0.02884962 0.02884962
2 0.00637888 0.02717726 0.01125746 0.00637888 0.00840632
3 0.00333165 0.02554878 0.00593714 0.00333165 0.00408176
4 0.002546 0.02397276 0.0052294 0.00253455 0.00360553
5 0.00249781 0.02245591 0.00475019 0.00248833 0.00332634
… … … … … …
170 0.00248842 0.00248968 0.00248505 0.00248569 0.00249161
171 0.00248681 0.00249039 0.00248496 0.0024855 0.00249161
172 0.00249087 0.00249013 0.00248499 0.0024854 0.00249161
173 0.00248566 0.00248984 0.00248504 0.0024853 0.00249161
174 0.00248565 0.00248996 0.00248499 0.00248545 0.00249161
175 0.0024859 0.00248953 0.00248499 0.0024855 0.00249161
176 0.00248837 0.0024894 0.00248514 0.00248557 0.00249161
… … … … … Converged
995 0.00248683 0.00248837 0.00248520 0.00248498
996 0.0024857 0.00248922 0.00248525 0.00248503
997 0.00248546 0.00248912 0.00248518 0.00248501
998 0.00249072 0.00248929 0.00248543 0.00248497
999 0.00248681 0.00248888 0.00248524 0.00248496
1000 0.00248571 0.00248858 0.00248533 0.00248498

Table 5 Optimum results of Example 7 with βt = 3.0 and σ = 0.5

Method Normal Lognormal

Design variables Objective Evaluations g1/g2/g3 (total) Design variables Objective Evaluations g1/g2/g3 (total)
AMV (3.79905, 4.0170) 7.81605 135/288/105 (528) (3.67505, 3.75918) 7.43423 75/699/207 (981)
CC (3.79905, 4.01699) 7.81604 1920/1680/1797 (5397) (3.67504, 3.75917) 7.43421 1665/1671/2349 (5685)
MCC (3.79905, 4.0167) 7.81605 1623/999/1497 (4119) (3.67505, 3.75918) 7.43423 1413/753/2127 (4293)
ACC (3.79905, 4.0167) 7.81605 135/336/105 (576) (3.67505, 3.75918) 7.43423 75/663/207 (945)
Proposed SMCC (3.79905, 4.0170) 7.81605 237/117/207 (561) (3.67505, 3.75918) 7.43423 168/195/342 (705)
Method Weibull Gumbel
Design variables Objective Evaluations g1/g2/g3 (total) Design variables Objective Evaluations g1/g2/g3 (total)
AMV (4.03186, 4.51489) 8.54674 168/165/192 (525) – – Not converged
CC (4.03186, 4.51487) 8.54673 2091/1698/1665 (5454) (3.52511, 3.47121) 6.99632 1665/1665/3393 (6723)
MCC (4.03186, 4.51489) 8.54674 2022/1218/930 (4170) (3.52511, 3.47122) 6.99633 1239/672/3204 (5115)
ACC (4.03186, 4.51489) 8.54674 168/444/681 (1294) (3.52511, 3.47122) 6.99633 486/834/642 (1962)
Proposed SMCC 4.03186, 4.51489 8.54674 306/150/132 (588) (3.52511, 3.47122) 6.99633 338/390/1200 (1928)
A self-adaptive modified chaos control method 73

Table 6 Optimum results of Example 7 with βt = 2.5 and σ = 1.0

Method Normal Lognormal

Design variables Objective Evaluations g1/g2/g3 (total) Design variables Objective Evaluations g1/g2/g3 (total)
AMV – – Not converged – – Not converged
CC (4.09604, 4.40392) 8.49996 5037/4032/4461 (13,530) (3.84334, 3.88354) 7.72688 3663/3666/5865 (13,194)
MCC (4.09603, 4.40393) 8.49996 4326/1896/3837 (10,059) (3.84335, 3.88355) 7.72689 3459/1143/5442 (10,044)
ACC (4.09603, 4.40393) 8.49996 612/1926/507 (3045) (3.84335, 3.88355) 7.72689 399/2433/813 (3645)
Proposed SMCC (4.09603, 4.40393) 8.49996 1335/480/1077 (2892) (3.84335, 3.88355) 7.72689 867/375/1791 (3033)
Method Weibull Gumbel
Design variables Objective Evaluations g1/g2/g3 (total) Design variables Objective Evaluations g1/g2/g3 (total)
AMV – – Not converged – – Not converged
CC (4.17715, 4.63435) 8.81150 5766/4707/4737 (15,210) (3.76086, 3.69138) 7.45224 3663/3666/6102 (13,431)
MCC (4.17714, 4.63436) 8.81150 5436/2292/3357 (11,085) (3.76086, 3.69139) 7.45225 3378/1098/5715 (10,191)
ACC (4.17714, 4.63436) 8.81150 756/2147/417 (3330) (3.76086, 3.69139) 7.45225 642/1659/852 (3153)
Proposed SMCC (4.17714, 4.63436) 8.81150 1656/597/894 (3147) (3.76086, 3.69139) 7.45225 801/762/1959 (3522)

Therefore, the RBDO model for this example is given as method is as robust as the MCC and ACC methods but slight-
follows: ly more efficient. The SMCC method has top performance
including both robustness and efficiency in comparison with
Find ½t; r1 ; α T
min C  ¼ π t ½2r1 þ sinðαÞ
other reliability methods (i.e. AMV, CC, MCC and ACC).

1:6523 P M
S:t: P f g ¼ 1−  þ > 0 ≤ Φð−βt Þ
πEt2 cos2 α 0:66
 0:41 r1   ð27Þ
where 0:001≤ t ≤ 0:005; te t i ; σ ; 0:8 ≤ r1 ≤ 1:0; r1 e r1i ; σ2 ;
2
 
π = 6 ≤ α ≤ π = 4 ; α e αi ; σ2 ; X ¼ P ; M ; E 7 Conclusions
½t; r1 ; α 0 ¼ ½0:0025; 0:9; 0:524 ; β t ¼ 4:5 ; 6:0

This example includes six independent basic random vari- A self-adaptive control factor is proposed to search for the
ables, whose statistical characteristics are listed in Table 7. Also, MPTP in reliability and RBDO problems. In the proposed
this example involves three design variables i.e. t, r1, α with two SMCC method, the self-adaptive line is computed using the
target reliability indexes i.e. βt = 4.5, 6 .0, which represents a new and previous iteration results. The SMCC can be applied
type of problem with both normal and non-normal random var- to both concave and convex probabilistic constraints in
iables and a complicated structural constraint function. RBDO problems, efficiently and robustly. The efficiency, ro-
The optimum design variables, optimum objective func- bustness and accuracy of the SMCC method are compared
tion, the number of function evaluation, the performance mea- with the AMV, CC and MCC methods using six nonlinear
sure value and MPTP on the beta circle for different RBDO- performance functions and two nonlinear RBDO examples.
based PMA methods are listed in Table 8 for Example 8. As is The SMCC is converged as accurately as the MCC for high
evident from Table 8, the AMV and CC methods converge to nonlinear performance problems, but it is more efficient than
unstable solutions for this examples with βt = 6, but the MCC, the CC and MCC methods. The proposed SMCC iterative
ACC and SMCC methods converge to stable results. scheme has a same efficiency with the AMV for convex per-
Therefore, it can be concluded that the proposed SMCC formance functions, but the SMCC is more robust than other

Table 7 Basic random variables for Example 8

M P Variables Description Distribution Mean Coefficient

of variation

E Young’s modulus Lognormal 70,000 0.05

t t
(MPa)
Thickness (m) Normal 0.0025 0.05
α α Slop angle (rad) Normal 0.524 0.02
r1 Internal radius (m) Normal 0.9 0.025
M Bending moment Gumbel 80,000 0.08
r1 r2 P
(N-m)
Axial load (N) Gumbel 70,000 0.08
Fig. 5 A conical structure
 74

Table 8 Converged results of optimization and reliability analysis for Example 8

Method βt = 4.5 βt = 6

AMV CC MCC ACC SMCC AMV CC MCC ACC SMCC

Design variables t (m) 0.00257 Not converged 0.00257 0.00257 0.00257 Not converged Not converged 0.00288 0.00288 0.00288
r1 (m) 0.80000 0.80000 0.80058 0.80015 0.80000 0.80000 0.80000
α (rad) 0.52360 0.52360 0.52360 0.52360 0.52412 0.57026 0.52360
Opt. results C 0.01696 0.01696 0.01697 0.01697 0.01898 0.01947 0.01897
g 0.46978 0.46978 0.46979 0.46978 0.57609 0.57322 0.57612
g-iterations 9472 387,564 167,688 6372 479,732 190,077 25,445
Converged results of the MPTP based on the reliability analysis
MPTP t (m) 0.00219 Not converged 0.00219 0.00219 0.00219 Not converged Not converged 0.00245 0.00245 0.00245
(X *) r1 (m) 0.78714 0.78715 0.78776 0.78730 0.78267 0.78267 0.78267
α (rad) 0.52698 0.52698 0.52697 0.52698 0.52843 0.53082 0.52791
E (MPa) 65494.48 65494.47 65382.95 65494.46 64158.11 64091.90 64145.30
M (N-m) 109384.99 109385.09 109218.56 109380.06 139892.98 134782.41 139879.23
P (N) 72130.69 72130.72 72131.55 72131.52 72312.83 73376.56 72312.67
B. Keshtegar et al.
A self-adaptive modified chaos control method 75

existing reliability analysis methods for moderately nonlinear Keshtegar B, Miri M (2014a) Introducing conjugate gradient optimiza-
tion for modified HL-RF method. Eng Comput 31(4):775–790
structural functions (Examples 6 and 8). Moreover, the SMCC
Keshtegar B, Miri M (2014b) Reliability analysis of corroded pipes using
can significantly reduce the required number of evaluations conjugate HL–RF algorithm based on average shear stress yield
for nonlinear RBDO problems (Examples 1, 3, 5). It can be criterion. Eng Fail Anal 46:104–117
concluded that the proposed SMCC method has higher robust- Lee JO, Yang YS, Ruy WS (2002) A comparative study on reliability-
ness and efficiency in comparison with other reliability-based index and target performance- based probabilistic structural design
optimization. Comput Struct 80:257–269
MPTP search methods (i.e. AMV, CC, MCC and ACC), Li G, Meng Z, Hu H (2015) An adaptive hybrid approach for reliability-
which can enhance the convergence rate and stability in an based design optimization. Struct Multidiscip Optim 51(5):1051–
efficient manner. 1065
Liang J, Mourelatos ZP, Tu J (2008) A single-loop method for reliability-
Acknowledgments This work was supported by the National Natural based design optimization. Int J Prod Dev 5(1):76–92
Science Foundation of China (11402049). Meng Z, Li G, Wang BP, Hao P (2015) A hybrid chaos control approach
of the performance measure functions for reliability-based design
optimization. Comput Struct 146:32–43
Qu X, Haftka RT (2004) Reliability-based design optimization using
References probabilistic sufficiency factor. Struct Multidiscip Optim 27(5):
314–325
Rackwitz R, Fiessler B (1978) Structural reliability under combined ran-
Chen ZZ, Qiu HB, Gao L, Su L, Li P (2013) An adaptive decoupling
dom load sequences. Comput Struct 9:489–494
approach for reliability-based design optimization. Comput Struct
117:58–66 Santos SR, Matioli LC, Beck AT (2012) New optimization algorithms for
Cheng G, Xu L, Jiang L (2006) sequential approximate programming structural reliability analysis. Comput Model Eng Sci 83(1):23–56
strategy for reliability-based structural optimization. Comput Struct Tu J, Choi KK, Park YH (1999) A new study on reliability-based design
84:1353–1367 optimization. J Mech Des ASME 121:557–564
Cho TM, Lee BC (2010) Reliability-based design optimization using a Xiao NC, Huang HZ, Wang Z, Pang Y, He L (2011) Reliability sensitivity
family of methods of moving asymptotes. Struct Multidiscip Optim analysis for structural systems in interval probability form. Struct
42:255–268 Multidiscip Optim 44:691–705
Du X, Chen W (2004) Sequential optimization and reliability assessment Yang D (2010) Chaos control for numerical instability of first order reli-
method for efficient probabilistic design. J Mech Des ASME 126: ability method. Commun Nonlinear Sci Numer Simul 15:3131–
225–233 3141
Enevoldsen I, Sørensen JD (1994) Reliability-based optimization in struc- Yang D, Yi P (2009) Chaos control of performance measure approach for
tural engineering. Struct Saf 15(3):169–196 evaluation of probabilistic constraints. StructMultidiscip Optim
Ezzati G, Mammadov M, Kulkarni S (2015) A new reliability analysis 38(1):83–92
method based on the conjugate gradient direction. Struct Multidiscip Youn BD, Choi KK (2004) An investigation of nonlinearity of reliability-
Optim 51(1):89–98 based design optimization approaches. J Mech Des ASME 126(5):
Hasofer AM, Lind NC (1974) Exact and invariant second moment code 403–411
format. J Eng Mech 100:111–121 Youn BD, Choi KK, Park YH (2003) Hybrid analysis method for reliability
Keshtegar B (2016) Stability iterative method for structural reliability based-design optimization. J Mech Des ASME 125(3):221–232
analysis using a chaotic conjugate map. Nonlinear Dyn 1–14. doi: Youn BD, Choi KK, Du L (2005a) Adaptive probability analysis using an
10.1007/s11071-016-2636-1 enhanced hybrid mean value method. Struct Multidisc Optim 29:
Keshtegar B, Miri M (2013) An enhanced HL-RF Method for the com- 134–148
putation of structural failure probability based on relaxed approach. Youn BD, Choi KK, Du L (2005b) Enriched performance measure approach
Civ Eng Infrastruct 1(1):69–80 for reliability-based design optimization. AIAA J 43(4):874–884