Computers and Structures 151 (2015) 130–147

Contents lists available at ScienceDirect

Computers and Structures

journal homepage: www.elsevier.com/locate/compstruc

An adaptive quadratic approximation for structural and topology

optimization
Lei Li, Kapil Khandelwal ⇑
Dept. of Civil & Env. Engg. & Earth Sci., University of Notre Dame, United States

a r t i c l e i n f o a b s t r a c t

Article history: In dual sequential methods for structural and topology optimization, approximations based on diagonal
Received 20 May 2014 quadratic Taylor series expansion methods are frequently used to construct tractable sub-problems.
Accepted 13 January 2015 However, when approximating the second-order terms, some inconsistent enforcements are usually
Available online 18 February 2015
employed to ensure the convexity of the approximations (Groenwold et al. (2010)), which may cause
convergence problems in the optimization process. In this paper, an adaptive quadratic approximation
Keywords: (AQA) is proposed to improve robustness and convergence performance of the optimization process.
Structural optimization
Numerical results on representative structural and topology optimization problems show the efficiency
Topology optimization
Dual sequential approximation
of the new proposed method over other existing algorithms.
Diagonal quadratic approximation Ó 2015 Elsevier Ltd. All rights reserved.
Adaptive approximation

1. Introduction is carried out. Such approximations, by their construction, are

Approximation schemes are typically used in structural and
topology optimization problems, wherein the objective and the
constraint functions are approximated using appropriate schemes
at design points to construct an approximated sub-problem, which
is then solved using suitable optimization methods. Thus, the ori-
ginal difficult problem at each optimization step is replaced by a
tractable fast-to-compute approximated sub-problem. Since Sch-
mit [2–4] first introduced the approximation concept in structural
optimization in the late 1970s, extensive work has been done in
the area of developing high quality approximated sub-problems.
In early structural optimization studies [4–6], sub-problems were
created wherein the constraints were approximated using the
first-order Taylor series expansion in reciprocal variables since
such approximations can appropriately capture the behavior of
constraints. Fleury et.al [7] presented an approximation scheme
wherein first-order Taylor series expansion in both direct and reci-
procal variables are used to create a sub-problem. This method
yields a sub-problem that is convex and separable and can be effi-
ciently solved by dual methods [8,9]. This idea directly led to the
convex linearization (CONLIN) scheme [10], which is demonstrated
to be a successful algorithm in solving structural optimization
problems. The approximations in direct and reciprocal variables
are first order local approximations and are valid in a relatively
small vicinity near the design point at which the approximation
⇑ Corresponding author at: Dept. of Civil & Env. Engg. & Earth Sci., University of
Notre Dame, 156 Fitzpatrick Hall, Notre Dame, IN 46556, United States.
not very accurate away from the expansion point. All approxima-
tions of this kind are categorized into local approximation by Bar-
thelemy and Haftka [11].
To obtain better approximation, multi-point information can be
used to construct reliable mid-range approximations that are valid
in a relatively larger region near the expansion point. Haftka [12]
introduced modified reciprocal approximation and two-point pro-
jection method by enforcing derivative information at current and
previous design points. By linearizing the exponential intervening
variable, Fadel [13] proposed a two-point exponential approxima-
tion in which the exponent is calculated by matching the approx-
imate and exact function derivatives at two different design points.
To relax the conservatism in the CONLIN method and using the
previous point information, Svanberg [14] combined the modified
reciprocal approximation and CONLIN to develop the Method of
Moving Asymptotes (MMA). The moving asymptotes can adjust
the conservatism of MMA approximations by judging the oscilla-
tion behavior of the current and previous design points, so that
the approximation becomes adaptive and more efficient than
CONLIN.
The above methods have been shown to be successful in lot of
applications, but they are all first-order and monotonous schemes,
and therefore are not suitable to approximate non-monotonous
structural behaviors. Indeed, it is shown that these methods have
poor convergence properties or even fail to converge for some par-
ticular problems [15,16]. The limitation of first-order approxima-
tion is also discussed by Fleury et al. [9], and in the same study a
second order approximation based on Taylor series expansion is

http://dx.doi.org/10.1016/j.compstruc.2015.01.013
0045-7949/Ó 2015 Elsevier Ltd. All rights reserved.
 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147
shown to achieve the best compromise between conservativeness
and accuracy. Besides using the fully populated Hessian matrix to
construct the approximation, Fleury [17] showed that approxima-
tion formulated by using only a diagonal Hessian is a more effec-
tive way that is amenable to convex programming without much
accuracy loss by comparing with the sequential quadratic pro-
gramming (SQP) method. Intervening variables can also be
extended to second-order approximation, and the two point adap-
tive nonlinear approximations (TANA) schemes are of this type
[18,19]. TANA schemes add second-order correction term to expo-
nential approximation and are demonstrated to be effective for
optimizing truss problems. Similarly, a globally convergent MMA
(GCMMA) algorithm is presented by Svanberg [15,20] in which a
second-order term is added to the MMA approximation. To adjust
the second-order term, an additional inner loop is used in GCMMA
which increases the computational effort as compared to MMA.
Furthermore, to approximate the second-order term in GCMMA,
Bruyneel and Fleury [16,21] presented a family of MMA approxi-
mations and proposed a selection scheme for different conditions.
More recently based on the concept of using diagonal Hessian
approximation, Groenwold et al. [22] proposed a family of diagonal
quadratic approximations (DQA) based on reciprocal and exponen-
tial approximations. Also if dual methods are to be used, which are
efficient for large scale topology optimization problems, DQA can
always give the primal–dual relationship in a closed form, while
TANA and GCMMA approximations are not amenable to dual
methods as additional numerical analysis is required to construct
the dual sub-problem. The detailed application of DQA on struc-
tural and topology optimization problems is provided in the recent
studies by Groenwold et al. [1,23,24].
When solving large-scale structural/topology optimization
problems, the computation and storage of fully populated Hessian
becomes expensive; therefore, another benefit gained from using
a diagonal Hessian approximation is that the algorithm is computa-
tionally efficient. Furthermore, ignoring the off-diagonal entries of
Hessian makes an approximation separable, and then dual method
can be easily used with such approximations [17]. Hence, the sub-
problems based on the DQA approximations can be solved effi-
ciently by dual method to get the new starting design point for
the next iteration. This process continues sequentially until a con-
vergence criteria is satisfied, so it is also called dual method based
on sequential approximation (DSA) [25]. Although DQA proposed
by Groenwold [22] are quite suitable for DSA framework, they are
not problem-free approximations. In order to satisfy the convexity
and separability requirements of DSA algorithm, some spurious
enforcements in the DQA have been proposed [1]. These enforce-
ments ensure convexity of approximation where the derivative is
greater than zero by sacrificing the approximation accuracy and this
drawback is illustrated in Fig. 1. Under this enforcement, when the
gradient at a current design point xki is positive (where k is the iter-
ation number), the reciprocal or exponential approximation
becomes concave so the second derivative information they yield
becomes negative and, therefore the direct approximation without
any enforcement is concave. In this case, the enforced convex
approximation is constructed by changing the sign of derivatives
[1]; however, this is inaccurate and the enforced approximation
cannot represent the true behavior of the actual function anymore,
and the optimal point of the enforced approximation xkþ1 i is differ-
ent from the actual optimal value xi , as shown in Fig. 1. Thus,
although DQA family proposed by Groenwold [22] yields non-
monotonous approximations, the approximation may not be accu-
rate under certain conditions, and this may eventually result in slow
convergence. Alternatively, if these enforcements are relaxed or
improved by an adaptive approximation, the approximation accu-
racy will increase and this will finally speed up the convergence.
131
Enforced
Approximation
(Convex)
Current Expanding
Point (“+” Gradient)
Direct
Actual Function Approximation
(Concave)
0 ∗
Fig. 1. Illustration of enforcement for positive gradient.
In this paper, an adaptive quadratic approximation (AQA)
approach is proposed based on the DQA to overcome the difficul-
ties mentioned above by selecting proper approximations adap-
tively instead of direct enforcements. The AQA approximation is
shown to yield better results for structural and topology optimiza-
tion problems than the enforced DQA methods. The representative
test problems included in this paper are: truss optimization prob-
lems, Fleury’s weight like optimization problem, and topology
optimization for minimum compliance and compliant mechanisms
design.
The paper is arranged as follows: In Section 2, the outline of DSA
algorithm is provided. Then the various monotonous and non-
monotonous approximations together with their advantages and
restrictions are discussed in Section 3. In order to address the
restrictions described in Section 3, a new adaptive quadratic
approximation (AQA) scheme is proposed in Section 4. Perfor-
mance evaluation of the new proposed algorithm is carried out
by the comparison of various algorithms via numerical experi-
ments in Section 5. Finally, the important concluding remarks are
offered in Section 6.
2. Dual sequential approximation (DSA) algorithm
2.1. Nonlinear optimization problem
The primal structural/topology optimization problem can be
described as:
min f 0 ðxÞ
x2D
Subject to :
ð1Þ
f j ðxÞ 6 0; j ¼ 1; 2; . . . ; m
D ¼ fxi jxli 6 xi 6 xui g; i ¼ 1; 2; . . . ; n
where x 2 D  Rn is the design variable; f 0 ðxÞ is the objective func-
tion which to be minimized; f j ðxÞ are the inequality constraints, and
th
xli and xui are the upper and lower boundaries of the i design var-
iable xi , respectively. In most cases, f 0 ðxÞ and f j ðxÞ represent the lin-
ear or nonlinear responses of a structure, typically total weight,
displacement at certain nodes, and stiffness or compliance of a
structural system, and are assumed to be at least once continuously
differentiable. In structural optimization problems, the design vari-
ables typically represent area of cross section or any other geomet-
ric property of structural members while in topology optimization
132 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147

they represent density of material within the design domain. The ~f k ðkÞ ¼ inf Lk ðx; kÞ ð5Þ
d
x2D
optimization problem defined in Eq. (1) can be solved using: (a) pri-
mal methods, for e.g. active set and gradient projection methods; In DSA method, the sub-problem approximations ~f k0 ðxÞ and ~f kj ðxÞ
(b) penalty and barrier methods; (c) Lagrange multiplier methods
are combinations of convex and linear functions so that the
and (d) dual methods; see Ref. [26] for further details. For large
Lagrangian Lk ðx; kÞ is a strictly convex function. Therefore, for a
scale structural and topology optimization problems, however, the
optimization methods based on dual sequential approximations ~ ¼ arg inf x2D Lk ðx; kÞ can be obtained using the first
given k 2 Rm , x
are typically employed [27]. This is due to the fact that in some order stationary conditions as follows:
cases, e.g. topology optimization, the number of design variables X
m
far exceeds the number of constraints, and therefore working in r~f k0 þ kj r~f kj ¼ 0 ð6Þ
the dual space leads to computational efficiency [8]. Moreover, as j¼1

most of the computational cost in a large structural/topology opti-
mization problem is usually associated with the total number of
finite element analyses performed, it is desirable to keep the num-
ber of analyses required to a minimum, and therefore the optimiza-
tion methods that require direct estimation of Hessian are not
preferable as multiple finite element analyses are required con-
struct to the Hessian at each iteration.
 be the solution of Eq.(6) with no constraint enforcement;
Let x
at this stage the box constraints are enforced together with ‘‘trust
region’’ concept in DSA [29] that avoids expensive line search. In
‘‘trust region’’ methodology, considering the fact that the approxi-
mations used in DSA are usually only valid in a small region Di near
the current design point xki , the following updates are made to
obtain x ~ :

2.2. Approximated sub-problem ~xi ¼ maxðxki  Di ; xli ; minðxki þ Di ; xui ; ~xi ÞÞ ð7Þ

In dual sequential approximation algorithm, at each optimiza- where Di denotes the ‘‘trust region’’ and is also called ‘‘move limit’’
tion iteration, the original primal problem is replaced by a tractable [30,31]. Thus by using Eq. (7), the maximum change in the design
approximated sub-problem [27]. For structural and topology opti- variable is restricted in a region defined by the move limit and
mization problems, such approximations are typically constructed the box constraints. In a traditional DSA scheme a constant value
using the derivatives and function values of the objective and con- of move limit Di is used for all the design variables and for all the
straints, which are easy to compute from the finite element analy- optimization steps [23]. In this study we also propose a scheme
sis. A typical approximated sub-problem is given by Eq. (2) as wherein different move limits are used for different variables, and
follows: also the move limits are adaptively updated. Finally, the dual func-
tion can be obtained by simple back substitution as follows:
min ~f k0 ðxÞ
x2D
k
X
m
Subject to : ~ ; kÞ ¼ ~f k0 ðx
f d ðkÞ ¼ Lk ðx ~ Þ þ kj ~f kj ðx
~ Þ ð8Þ
ð2Þ j¼1
~f k ðxÞ 6 0; j ¼ 1; 2; . . . ; m
j

D ¼ fxi jxli 6 xi 6 xui g; i ¼ 1; 2; . . . ; n

2.4. Dual problem
where ~f k0 ðxÞ and ~f kj ðxÞ are the approximations of original objective
and constraint functions, respectively, at the current design point Now the dual optimization problem is given by Eq. (9):
xk . This sub-problem is formulated with special features that are max ~f kd ðkÞ
unique to structural/topology optimization problems such as using k

first order approximations in exponential and reciprocal variables Subject to : ð9Þ

[28]. Thus, at each iteration step, the original complex problem kP0
(Eq. (1)) is replaced by a tractable sub-problem (Eq. (2)). Finally,
in the DSA method, the sub-problem (Eq. (2)) is efficiently solved The solution of this problem is expressed as:
by dual method and the new starting point for the next iteration ^k 2 arg max ~f k ðkÞ
d ð10Þ
is determined using trust region strategy, and this process continue kP0
sequentially until a convergence criterion is satisfied [25].
Since in DSA the approximations ~f k0 ðxÞ and ~f kj ðxÞ are a combina-
2.3. Dual of the approximated sub-problem tion of convex and linear functions, there is no duality gap, and the
^ Þ of primal and dual sub-problems satisfy the
^ ; k
solution pair ðx
The global minimum of Eq. (2) is expressed as: following optimality conditions [26]:

^ 2 arg
x min ~f k ðxÞ
0 ð3Þ x^ 2 D ~f j ðx
^ Þ 6 0 ðPrimal feasibilityÞ
x2D
^k P 0 ðDual feasibilityÞ
~f k ðxÞ 6 0; j ¼ 1; . . . ; m
j
x^ 2 arg min Lk ðx; ^k Þ ðLagrangian optimalityÞ ð11Þ
x2D
^ , of the sub-prob-
The goal here is to find the global minimum, x
lem (Eq. (2)), and in the DSA algorithm the dual method is used to ^k~f j ð^x Þ ¼ 0 ðComplementary slacknessÞ
j
find this optimal solution. To this end, the Lagrangian function j ¼ 1; 2; . . . ; m
Lk ðx; kÞ : Rn  Rm ! R of Eq. (2) is defined by:
Moreover, as the optimization problem in Eq. (9) is a unilater-
X
m
ally constrained problem, any gradient based optimization method
L ðx; kÞ ¼ ~f k0 ðxÞ þ
k
kj ~f kj ðxÞ ð4Þ
^ Þ [26]. The
^ ; k
can be used to solve for the optimal solution pair ðx
j¼1
obvious benefit of the DSA method is that the number of variables
where kj , j ¼ 1; 2; . . . ; m are the Lagrange multipliers corresponding in the dual problem equals the number of constraints, and thus
to the constraints. Using the Lagrangian, the dual function, dual method is efficient when the number of design variables far
~f k ðkÞ : Rm ! R, is defined as: exceeds the number constraints, i.e. when n  m.
d
 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147 133

3. Approximations used in structural/topology optimization where U ki and Lki are the moving asymptotes at step k that control
the curvature of the approximation. The moving asymptotes are
3.1. Monotonous approximations adjusted using information at previous steps to get optimal conver-
gence rates for the current step. The MMA is a popular approxima-
3.1.1. Linear approximation tion and has been widely used in solving various kinds of structural
The linear approximation is based on the first-order Taylor ser- optimization problems [20].
ies expansion at the current design point xk and is given as follows:
  !
X
n
@f 3.1.5. Exponential approximation
^f k ðxÞ ¼ f ðxk Þ þ ðxi  xki Þ ð12Þ
L
@xi xi ¼xk To control the curvature of the reciprocal approximation and
i¼1 a
i
make it more adaptive, exponential intervening variable yi ¼ xi i
 
where @f
is the first-order derivative w.r.t xi at xki . Linear are sometimes used to generate the exponential approximation
@xi
xi ¼xki as follows:
approximation has no curvature because its second-order deriva- 02 3 1
!a k ! 
tive is zero. Xn i
xki
^f k ðxÞ ¼ f ðxk Þ þ @4 xi 1 5 @f A ð17Þ
E
3.1.2. Reciprocal approximation i¼1
xki aki @xi xi ¼xk
i

In some statically determinate structures, stresses and displace-

ments are exact linear functions of the reciprocal of design vari- Note that when aki ¼ 1 the linear approximation is recovered and
ables [27]. So linearizing in the reciprocal intervening variable, when aki ¼ 1 the reciprocal approximation is recovered. The expo-
yi ¼ 1=xi , yields a better approximation for solving such structural nential is usually determined through two point enforcement as
optimization problems. The reciprocal approximation can be discussed in Section 3.3.1.
expressed as follows: !
X
n  k  
^f k ðxÞ ¼ f ðxk Þ þ x @f 3.2. Non-monotonous approximations
R ðxi  xki Þ i ð13Þ
i¼1
xi @xi xi ¼xk
i
In order to capture the curvature information in the sub-prob-
Note that the reciprocal approximation is convex if lems, the second-order derivative terms are added to the monoto-
 
@f
<0 8xi . nous approximations, and the second order terms make the
@xi
xi ¼xki
approximation non-monotonous. Some of the important non-
monotonous approximations are presented in the following
3.1.3. Convex linearization (CONLIN) subsections.
The CONLIN approximation proposed by Fleury [10] ensures
convexity, and employs a combination of linear and reciprocal
3.2.1. Globally convergent MMA (GCMMA)
approximation as follows:
  Although the MMA has been used for solving a large class of
X @f
^f k ðxÞ ¼ f ðxk Þ þ ðxi  xki Þ structural optimization problems, there still exist some problems
CL
þ
@xi xi ¼xk for which MMA has poor convergence speed or even fails to con-
i

X  k   verge [15,32]. To overcome this issue, Svanberg [15,20] presented

k xi @f
þ ðxi  xi Þ ð14Þ an improved MMA approximation which is called globally conver-

xi @xi xi ¼xk
i gent MMA (GCMMA) and is given as follows:
where the symbol Rþ ðRÞ means ‘‘summation over the terms for !
  X
n pkij qkij
^f k k
GCMMA ðxÞ ¼ f ðx Þ  þ
@f
which @x is positive (negative). Since linear, reciprocal and
i k xi ¼xi i¼1 U ki  xki xki  Lki
CONLIN approximations are valid only within a vicinity of the
!
X
n pkij qkij
expansion point xki , they are also called local approximations [11]. þ þ ð18Þ
i¼1 U ki  xi xi  Lki
3.1.4. Method of Moving Asymptotes (MMA) where
The CONLIN uses a combination of reciprocal and linear approx- 0 1
 3 !
imation and there is no direct control on the convexity of the ðU ki  xki Þ @f @ 2
f
approximation. To adjust the convexity of CONLIN approximation, pki¼  @2 k k
þ ðxi  Li Þ A
2ðU ki  Lki Þ @xi xi ¼xk @x2i
Svanberg [14] introduced the methods of moving asymptotes i xi ¼xki
0 ! 1
(MMA) by extending the CONLIN scheme. The MMA approximation 3   ð19Þ
k ðxki  Lki Þ @ @f k k @2f A
is given by: ! qi ¼  2 þ ðU i  xi Þ
X
n 2ðU ki  Lki Þ @xi xi ¼xk @x2i
^f k ðxÞ ¼ f ðxk Þ  pki qki i k xi ¼xi
MMA þ
i¼1 U ki  xki xki
 Lki Lki < xki < U ki
!
X
n
pki qki
þ þ ð15Þ In GCMMA, pki and qki can be simultaneously nonzero, which leads to
k
i¼1 Ui  xi xi  Lki the non-monotonous behavior of the approximation. The GCMMA
approximation is constructed by adding a second-order term to
where
! MMA scheme, and in this way by using different schemes to
 
@f 2 approximate the second order derivative term, the non-monoto-
pki ¼ max  ðU ki ;0 xki Þ nous GCMMA family can be obtained [16]. However, one of the
@xi xi ¼xk
 
i
! main drawbacks of the GCMMA approximation is that an explicit
2 @f ð16Þ primal–dual relationship cannot be obtained and this approxima-
qki ¼ max ðxki  Lki Þ ;0
@xi xi ¼xk tion is not appropriate for DSA framework described in Section 2.
i
Moreover, this algorithm requires inner-loops to control the curva-
Lki < xki < U ki ture by approximating second order derivatives, and therefore,
134 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147

resulting in a greater number of finite element analyses at each 3.3. Remarks on approximations
iteration.
The accuracy and convergence speed of a DSA algorithm depend
on the quality of the approximation schemes. If the approxima-
3.2.2. Two-point adaptive nonlinear approximation (TANA)
tions are good representations of primal functions, the DSA algo-
Xu and Grandhi [19] proposed a two point adaptive nonlinear
rithm is fairly efficient [25]. In the DSA framework, for a general
approximation (TANA) for solving truss optimization problems.
nonlinear programming problem, determining an appropriate
TANA approximation can be expressed as:
approximation for constructing a sub-problem is not an easy task.
02 !a k 3 !  1 Furthermore, the approximated sub-problem should be convex
X n
x
i
x k
@f
^f k ðxÞ ¼ f ðxk Þ þ @4 i
 15 k i A and separable, so that its dual can be easily evaluated; else the dual
TANA
i¼1
xki ai @xi xi ¼xk framework loses its attractiveness and will not be more efficient
i
" # than other competing methods. Specifically, there are four require-
1 X n
aki
k 2
k ai ments that are needed to be satisfied for the approximations to
þ e ððxi Þ  ðxi Þ Þ ð20Þ
2 i¼1 work effectively in DSA framework.

where 3.3.1. Convexity

Approximations should be chosen such that the Lagrangian in
H
e ¼ P 2 Pn 2
 ð21Þ Eq. (4) is convex, and thus only have one global minimum. Also,
n aki ak aki ak
i¼1 ððxi Þ  ðxk1
i Þ iÞ þ i¼1 ððxi Þ  ðxk2
i Þ iÞ in this case there is no duality gap and the solution of dual sub-
problem also yields the solution of the primal sub-problem [26].
Among the first order approximations discussed above the linear
and H is defined as follows:
approximation is non-convex, reciprocal approximation is convex
2 2 3 3  
!a k !  only if @x@f
< 0, and CONLIN and MMA approximations are
Xn
x k1 i
x k
@f k i xi ¼xi
H ¼ 24f ðx k1
Þ  f ðx Þ  k 4 i
 15 k i 5 ð22Þ
i¼1
xki ai @xi xi ¼xk always convex. For exponential approximation the convexity is
i
determined by unknown exponents aki and this approximation is
In TANA approximation, a correction term is added to the expo- convex if the following conditions are satisfied:
nential approximation Eq. (17), in order to correct the approxima-  
tion error, by matching the function value of the approximation @f
aki < 1 if <0
with the exact value at the previous point. Additionally, the square @xi xi ¼xk
a   i ð26Þ
term of the intervening variable yi ¼ xi i is also considered in the @f
second order term of Taylor series expansion. TANA has been aki > 1 if >0
@xi xi ¼xk
i
proved to be efficient and accurate for solving truss problems
[18,19], but this approximation does not yield an explicit primal– In practice aki
is estimated from the following two point gradi-
dual relationship and so its not suitable for DSA method. ent enforcement [13]:
!  
@ ^f kE @f
3.2.3. Diagonal quadratic approximation (DQA) ¼ ð27Þ
@xi @xi xi ¼xk1
Based on the Fleury’s [17] concept of using diagonal Hessian, xi ¼xik1 i

Groenwold et al. [1,22,23] recently proposed an approximated

approximation based on the second-order Taylor series expansion, Eq. (27) leads to the following expression for the exponent:
    
which has the form: @f @f
ln @xi @xi
X   xi ¼xik1 xi ¼xki
^f k ðxÞ ¼ f ðxk Þ þ
n
@f 0 1Xn
2 aki ¼ 1 þ   ð28Þ
QD ðxi  xki Þ þ ck ðxi  xki Þ ð23Þ xik1
i¼1
@xi xi ¼xk 2 i¼1 i ln xki
i

where the cki denotes the diagonal elements in Hessian, i.e. Note that in Eq. (28), two possibilities exist that may cause
cki ¼ ð@ 2 f =@x2i Þxi ¼xk . Finally, cki ’s are evaluated using reciprocal and numerical problems: (a) the values within logarithm may be neg-
i ative when the gradients at two successive points have different
exponential approximations to yield T2:R and T2:E approximations signs, and since in structural optimization problem the design vari-
as follows: ables x are generally positive, this will lead to an imaginary value
for aki ; (b) the aki calculated from Eq. (28) may not satisfy the con-
(a) Reciprocal Approximation (T2:R) – in this case the cki ’s are vexity requirement shown in Eq. (26). To circumvent the difficul-
estimated from the reciprocal approximation (Eq. (13)): ties mentioned above, following enforcements are proposed by
 
@ 2 ^f R 2 @f Groenwold [1]:
cki ¼ 2
¼ k ð24Þ  
@xi xi @xi xi ¼xk @f
i
aki ¼ minðaki ; 1Þ if <0
@xi xi ¼xk
i
 
(b) Exponential Approximation (T2:E) – in this case the cki ’s are @f ð29Þ
aki ¼ maxðaki ; 1Þ if >0
estimated from the exponential approximation (Eq. (17)): @xi xi ¼xk
  i

@ 2 ^f kE 1  aki @f If aki 2 Im; then aki ¼ 1

cki ¼ 2
¼ k
ð25Þ
@xi xi @xi xi ¼xk
i Furthermore, for some non-convex problems, e.g. the compliant
Above approximations together with exponential, MMA, mechanisms, the following enforcements are found to be superior
GCMMA, TANA and DQA are more reliable than local approxima- to the enforcements shown in Eq. (29) which constrain aki within a
tions and so they are also called mid-range approximations [11]. certain range [1,23]:
 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147 135

   
@f @f approximations presented in Section 3.1 have good convergence
¼ speed while the non-monotonous approximations may be too con-
@xi xi ¼xki @xi xi ¼xki
ð30Þ servative as shown in Fig. 2(a). In this figure, xik1 means the start-
aki ¼ max½minðaki ; Þ; 3 th
ing point for the k iteration, xk k
nm;i and xm;i are the optimal point
If aki 2 Im; then aki ¼ 1 obtained from non-monotonous and monotonous approximation,
where  is a prescribed small positive number. Using the enforce- respectively. On the other hand, when the primal function has
ment in Eq. (30), the convexity of T2:E and T2:R approximations non-monotonous behavior (Fig. 2(b)), the monotonous approxima-
can be ensured as follows [1]: tions may have poor performance since the solution may hit the
move limit boundary xui , which will result in poor convergence
 
2 @f speeds and oscillation behavior especially when approaching the
cki ¼ final solution.
xki @xi xi ¼xki
  ð31Þ
@ 2 ^f kE 1  aki @f
cki ¼ ¼
@x2i xki @xi xi ¼xki 4. Adaptive quadratic approximation (AQA)

where again aki is determined from Eq. (28). Clearly, these enforce- In the structural optimization problems the choice of approxi-
ments are inconsistent in the cases where the gradients change the mation depends on the problem features. For some structural opti-
sign or in cases where the gradient is positive (Fig. 1). This inconsis- mization problems, it can be easily figured out the best type of
tency is addressed in the AQA presented in the next section. approximations that may be used, e.g. linear approximation is
the best for the total weight function of truss, and reciprocal
3.3.2. Separability approximation is suitable for representing stress and displacement
Separability condition implies that the approximations in Eq. constraints in simple trusses. However, in the more difficult situa-
(2) can be written as follows: tions, it is not easy to find the a priori approximation scheme to
use. Moreover, the type of approximation for the same problem
X
n
~f k ðxÞ ¼ ~f k ðx Þ for r ¼ 0; 1; 2 . . . ; m ð32Þ may change as the optimization proceeds as the behavior around
r ir i
i¼1 the design point may change. In this section, a new approximation
scheme based on adaptively selecting different DQA schemes is
With the separability property the first order minimization con-
ditions (Eq. (6)) becomes:

d~f ki0 ðxi Þ X m d~f kij ðxi Þ Actual

þ kj ¼0 ð33Þ Non-monotonous
dxi j¼1
dxi Function
Approximation
Therefore, with the separability assumption, a coupled system
of n nonlinear equations given by Eq. (6) is reduced to n – uncou-
pled nonlinear equations: i.e. the calculation of dual function is
decomposed into n simpler minimizations. This is an important
requirement of the DSA framework, as it leads to a simple evalua-
tion of the dual function; otherwise, the DSA framework may not Monotonous
be superior to other optimization methods as evaluation of the Approximation
Move Limit
dual function may not be trivial.

3.3.3. Closed form of primal–dual relationship

Another desirable property of the approximation is that Eq. (33)
can be analytically solved, i.e. the primal–dual relationship can be (a)
expressed in closed form. Although the Newton–Raphson method
can be used to solve one dimensional optimization problems (Eq.
(33)) such a scheme will become prohibitively expensive for large
scale optimization problems. For this reason the TANA [19] and the
GCMMA [20] approximations are not suitable for the DSA frame-
Non-monotonous
work as an explicit primal–dual relationship cannot be obtained
Approximation
in these cases, and extra numerical analyses are required to obtain
Actual
the primal–dual relationship at each iteration. Based on this con- Function
sideration the MMA and DQA family of approximations are well
suited for DSA framework and are further considered in this study.

3.3.4. Accuracy
The accuracy of the approximation has a significant impact on Monotonous
Move Limit
the effectiveness and convergence speed as well as on the quality Approximation
of final results. An optimization method with the fixed approxima-
tion scheme may not be very efficient as the behavior of the prob-
lem (objective and constraint functions) may be different at
different design points. So the ability of approximation schemes
(b)
to adjust their accuracy adaptively is an important and active field
in modern structural optimization research. When the primal Fig. 2. Monotonous and non-monotonous approximations: (a) f j ðxÞ is monotonous
function itself is monotonous, the first order monotonous (b) f j ðxÞ is non-monotonous.
136 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147

introduced in order to preserve the approximation accuracy irre- (c) If (N i > 1 and Di > 1) or if (0 < N i < 1 and Di < 1), then
spective of whether the derivative at two successive design points aki > 1 and T2:E is not applicable. In this case T2:R is used
have the same sign or not. As shown in Eq. (23), the accuracy of the instead, as shown in Fig. 3(c) and (e) with:
approximation is actually determined by parameter C ki in the sec-
 
ond order term, so selection of the appropriate scheme to calculate 2 @f
cki ¼  >0 ð39Þ
C kiis the main idea of the proposed AQA algorithm. All of the k @x
xi i xi ¼xk
i
choices need to be based on the available information, for example,
function values and gradient at the current iteration and previous
iterations. (d) If (N i < 0 and Di > 1), then Eq. (37) yields an imaginary aki
(Fig. 3(f)); in this case the quadratic approximation to finite
4.1. Primal–dual relationship difference approximation (T2:F) is used as follows:
   
As discussed in Section 2, the dual function of the DQA sub- @f j @f j
@xi
 @xi
problem Eq. (2) can be written as: xi ¼xki xi ¼xik1
cki ¼ >0 ð40Þ
X   xki  xk1
k
n
@f 0 1X n
2
i
f d ðkÞ ¼ f 0 ðxk Þ þ ðxi ðkÞ  xki Þ þ ck0i ðxi ðkÞ  xki Þ
i¼1
@x i x i ¼x k 2 i¼1
i
  ! (e) If (N i < 0 and Di < 1), then aki is imaginary and T2:F does
X m X n
@f j 1X n
2
k k
þ kj f j ðx Þ þ ðxi ðkÞ  xi Þ þ ckji ðxi ðkÞ  xki Þ not yield convex approximation. T2:R is again used for this
@xi xi ¼xk 2 i¼1
j¼1 i¼1 i case as shown in Fig. 3(g) with cki given in Eq. (36).
ð34Þ

The primal–dual relationship is obtained using stationary con-

ditions (Eq. (6)) which yields: 4.2.2. Positive gradient at a design point
     Now consider the positive gradient at the current step, i.e.
@f 0 Pm @f j  
@xi
þ j¼1 kj @xi
@f
> 0. In this case the following strategy is used for select-
xi ¼xki xi ¼xki @xi k
xi ¼xi
xi ðkÞ ¼ xki  Pm ð35Þ
ck0i þ k
j¼1 kj c ji ing the approximation subclass.

The important idea of the AQA scheme is to choose appropriate

(a) For the first step (k ¼ 1), there is no previous point informa-
ck0i and ckji based on the available information. tion to construct the approximations. In this case the qua-
dratic approximation to reciprocal approximation (T2:R) is
4.2. Approximation selection used with cki given by (see Fig. 3(h)):

4.2.1. Negative gradient at a design point  

2 @f
First, consider the negative gradient of the current step cki ¼ >0 ð41Þ
  xki @xi xi ¼xk
@f i
@xi k
< 0 (where f  f r and r ¼ 0; 1; . . . m), and there is mixed
xi ¼xi
Now starting from the second step (k P 2), the AQA algo-
strategy for the sub-cases.
rithm automatically selects the different schemes by using
the history information.
(a) For the first step (k ¼ 1), there is no previous point informa-
tion to construct the approximations. Since the reciprocal
  (b) If (N i > 1 and Di > 1) or if (0 < N i < 1 and Di < 1), then
@f
approximation is always convex when @x i k
< 0, the aki > 1 and the quadratic approximation to exponential
xi ¼xi

quadratic approximation to reciprocal approximation approximation (T2:E) is used as shown in Fig. 3(i) and (k),
(T2:R) is used, as shown in Fig. 3(a). In this case we have: and the second order term is given by:

   
2 @f aki  1 @f
cki ¼  >0 ð36Þ cki ¼ >0 ð42Þ
xki @xi xi ¼xk xki @xi xi ¼xk
i i

i.e. aki ¼ 1 in this case (Eq. (25)). Starting from the second
step (k P 2), the AQA algorithm automatically selects the (c) If (N i > 1 and Di < 1) or if (0 < N i < 1 and Di > 1), then
different schemes by using the history information. Let aki < 1 and T2:E is not applicable. In this case T2:R is used
   
@f @f instead, as shown in Fig. 3(j) and (l) with:
N i , ¼ @x @x
and Di , ¼ xk1
i =xki , then using
ik1 xi ¼xi k i xi ¼xi  
Eq. (28):
2 @f
cki ¼ k @x
>0 ð43Þ
xi i xi ¼xk
ln Ni i
aki ¼1þ ð37Þ
ln Di
(d) If (N i < 0 and Di < 1), then Eq. (37) yields an imaginary aki
(b) If (N i > 1 and Di < 1) or if (0 < N i < 1 and Di > 1), then (Fig. 3(m)); in this case the quadratic approximation to
aki < 1 and the quadratic approximation to exponential finite difference approximation (T2:F) is used as follows:
approximation (T2:E) is used as shown in Fig. 3(b) and    
@f j @f j
(d), and the second order term is given by: @xi
 @xi
xi ¼xki xi ¼xk1
  cki ¼ i
>0 ð44Þ
ak  1 @f xki  xk1
cki ¼ i k >0 ð38Þ i
xi @xi xi ¼xk
i
 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147 137

T2:R T2:E

( )
( )
( )
) )
)

(a) (b)

T2:R T2:E
( ) ( )

) )
) )
( ) ( )

(c) (d)
( ) T2:R T2:F
( )
) ( )
)
)
)
( )

(e) (f)
T2:R
T2:R
( )
( ) ( )
)
) )

(g) (h)
Fig. 3. Adaptive approximation schemes in proposed AQA algorithm.

(e) Finally, if (N i < 0 and Di > 1), then aki is imaginary and T2:F Dki ¼ sdecr Dk1
i ; if ðxki  xik1 Þðxik1  xk2
i Þ<0
does not yield a convex approximation. T2:R is again used ð46Þ
for this case as shown in Fig. 3(n) with cki given in Eq. (43). Dki ¼ sincr Dk1
i ; if ðxki  xik1 Þðxik1  xk2
i ÞP0
4.3. Move limits The main idea of the above scheme is that when
ðxki  xik1 Þðxik1  xik2 Þ < 0, oscillation occurs in the variable so
Instead of using the constant move limit Di, the move limit in this the ‘‘trust region’’ is decreased by sdecr . Else if
study is determined using an adaptive scheme as follows: for the ðxki  xik1 Þðxik1  xik2 Þ P 0, the design variable is monotonically
first two iterations, i.e. k = 1 and 2 the move limits are set to: increasing or decreasing and the ‘‘trust region’’ is increased by
Di ¼ s0 ðxui  xli Þ ð45Þ sincr . In our implementation we use: s0 ¼ 0:1, sdecr ¼ 0:8 and
sincr ¼ 1:4. Using this method, each design variable xi has its own
where s0 is the initial move limit coefficient. Then starting from the move limit Di , and this is different from the traditional move limit
third step. ðk P 3Þ, the following strategy is adopted: scheme, which has the same Di for every design variable xi [1]. This
138 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147

T2:E ( ) T2:R

) ( ) )

)
( ) ) ( )

(i) (j)
( ) T2:E T2:R
( )
) ( ) )
( )
)
)

(k) (l)
T2:F T2:R
( )
( )
)
( )
) )
( )
)

(m) (n)
Fig. 3 (continued)

move limit scheme is adapted from the scheme used in updating
the moving asymptotes in the MMA method presented by Svan-
berg [14].
5. Numerical experiments
The proposed AQA approximation scheme is tested and com-
pared with other approximation methods on several selected
5.1. Problem-1: Two bar truss (shape and sizing)
First, the two-bar truss problem proposed by Svanberg [14] is
considered as shown in Fig. 4(a). In this problem sizing design
(x1 ) and shape design (x2 ) are simultaneously considered, and the
objective is to minimize the weight of the truss subjected to stress
constraints. The problem can be expressed analytically as follows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
min f 0 ðxÞ ¼ c1 x1 1 þ x22
x

structural and topology optimization problems that exhibit differ- Subject to :

ent nonlinear behaviors. The approximations that are tested here qffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
8 1
include: the quadratic approximation to the reciprocal approxima- f 1 ðxÞ ¼ c2 1 þ x22 þ 160 ð47Þ
x1 x1 x2
tion (T2:R), the quadratic approximation to the exponential qffiffiffiffiffiffiffiffiffiffiffiffiffiffi  
8 1
approximation (T2:E), the Method of Moving Asymptotes (MMA) f 2 ðxÞ ¼ c2 1 þ x22  160
x1 x1 x2
[14] and the proposed adaptive quadratic approximation (AQA).
For T2:E, the enforcement Eq. (31) is used and T2:R can be recov- 0:2 6 x1 6 4:0; 0:1 6 x2 6 1:6
ered by choosing aki ¼ 1 in T2:E. Algorithms are assumed to con-
verge if kxk  xk1 k2 6 ex where ex is the prescribed tolerance. All where c1 ¼ 1:0 and c2 ¼ 0:124, and starting point is xð0Þ ¼ ð1:5; 0:5ÞT

the finite element analyses and optimization iterations are carried and optimal solution is f 0 ðxÞ ¼ 1:5086524. In this problem, we set
out in the MatlabÒ computing environment. To have a fair compar- ex ¼ 1010 to test the convergence of T2:R, T2:E, MMA and AQA
ison, all the design variables, objective and constraint functions are methods and the detailed numerical results are shown in Table 1.
identically scaled and all of the algorithms except MMA employ As shown in Table 1, the DQA family of schemes (T2:R and T2:E)
the adaptive move limit strategy described in Section 4.3 since perform better than MMA and take fewer steps to converge to the
MMA has its own move limit updating scheme. optimal solution. Meanwhile the solutions are more feasible than
 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147 139

1.7

1.6

1.5

0
Objective f
1(m)
1.4

1.3
T2:R
T2:E
1.2 MMA
AQA

0 5 10 15
Iteration

(a) (b)
0 0
10 10
T2:R T2:R
−2 T2:E T2:E
10
MMA MMA
AQA c ∞ −5 AQA
−4
10 10
Tolerance ||f ||
x
Tolerance ε

−6
10
−10
−8 10
10

−10
10
−15
10
−12
10
0 5 10 15 0 5 10 15
Iteration Iteration
(c) (d)
Fig. 4. Convergence history for two-bar truss: (a) truss configuration; (b) objective value convergence; (c) variable tolerance convergence; (d) maximum constraint violation.

X
n
Table 1 min f 0 ðxÞ ¼
x
qi li xi
Convergence results for the two-bar truss problem. i¼1

Approximations Number of steps to 

f0 kf c k1 Subject to :
converge  j 6 0; ð48Þ
f j ðxÞ ¼ uj ðxÞ  u j ¼ 1; 2; . . . m
T2:R 11 1.5086524 8.29225577e12
T2:E 13 1.5086524 8.61533067e14
f k ðxÞ ¼ rk ðxÞ  r
 k 6 0; k ¼ 1; 2; . . . p
MMA 17 1.5086524 6.62755293e08 KðxÞuðxÞ ¼ P
AQA 10 1.5086524 3.86357612e14 0 < xli 6 xi 6 xui

MMA since the constraint violations are smaller as shown by the
constraint norm kf c k1 , ¼ max fjf j jgj¼1;...;m . As shown in Fig. 4,
AQA performs better than other algorithms as both positive and
negative gradients are encountered in this problem and AQA uses
an adaptive scheme to handle these cases.
5.2. Problem-2: Minimum weight design of truss problems
Next the minimum weight design of truss problems is consid-
ered. Unlike the two bar truss, the objective function is a linear
function of design variables in this case. The problem behavior is
reciprocal or exponential-like which is known to be non-mono-
tonic [1]. The problem can be expressed as:
where the objective function f 0 ðxÞ is the weight of the truss (equal
to the sum of all the members weight); qi denotes the material den-
th
sity of the i member with length li . The constraints include node
displacement constraints f j ðxÞ and element stress constraints
 j and r
 k represent the prescribed allowable displace-
f k ðxÞ, while u
ments and stresses, respectively. The displacement constraints are
applied in all directions and the stress constraints are applied both
in tension and in compression. The problem is solved using finite
element method, and KðxÞ denotes the stiffness matrix, while uðxÞ
and P represent the displacement and force vectors, respectively.
The design variable, xi , is the cross-sectional area of an element
group, which has the lower and upper boundaries (xli and xui ). A
nested solution approach is used, wherein the state equation
KðxÞuðxÞ ¼ P is solved at each optimization iteration [33]. Five
benchmark truss problems are selected and studied in this paper:
140 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147

(a) Three bar truss: Geometry of the three bar truss is shown in
Fig. 5. For this problem the material density is q ¼ 2768 kg=m3 ,
Young’s modulus is E ¼ 68:947 GPa, allowable stresses in each ele-
ment are r 1 ¼ r  3 ¼ 34:47 MPa, r  2 ¼ 137:90 MPa in both tension
and compression, displacement limits are u  4x ¼ u
 4y ¼ 0:127 mm.
The starting point is set to x0i ¼ 6:4516 cm2 , lower and upper
boundaries of the design variables are xl ¼ 1:935 cm2 and
xu ¼ 64:516 cm2 , respectively.
(b) Ten bar cantilever truss: Geometry of this truss is shown in
Fig. 6. Material density is q ¼ 2768 kg=m3 and the Young’s modu-
lus is E ¼ 68:947 GPa and the allowable stresses in each element
are r k ¼ 172:36 MPa (k ¼ 1; 2 . . . ; 10) in both tension and compres-
sion, the displacement limits at node 2 and 4 are
x ¼ u
u  y ¼ 50:8 mm. The starting point is set to x0i ¼ 6:4516 cm2 ,
lower and upper boundaries of the design variables are
xl ¼ 3:23 cm2 and xu = 193.55 cm2, respectively. Fig. 6. Ten bar truss geometry.
(c) Twenty-five bar truss: Geometry of the 25-bar 3-D truss is
shown in Fig. 7. The material density q ¼ 2768 kg=m3 , and the
Young’s modulus E ¼ 68:947 GPa. The members are grouped into
seven distinct groups and within each group the members have
the same area. Therefore, there are seven design variables and
member grouping is as follows: x1 = {A1}, x2 = {A2, A3, A4, A5},
x3 = {A6, A7, A8, A9}, x4 = {A10, A11, A12, A13}, x5 = {A14, A15, A16, A17},
x6 = {A18, A19, A20, A21} and x7 = {A22, A23, A24, A25}, where Ai is the
th
area of the i member. Two load cases are considered and are
shown in Table 2. The stress constraints are r  k ¼ 172:36 MPa
(k ¼ 1; 2; . . . ; 25) in both tension and compression, and displace-
ment constraints are u x ¼ u
y ¼ u z ¼ 50:8 mm for all nodes. All ini-
tial design value are x0i ¼ 6:4516 cm2 and the design variables are
constraints within xl ¼ 0:0645 cm2 and xu ¼ 64:516cm2 .
(d) Thirty-six bar truss: Configuration of 36-bar truss is shown in
Fig. 8. Truss member nodal coordinates and nodal connectivity
information, member grouping, load condition, constraints and
material properties are shown in Table 3. The lower and upper lim-
its of cross-sectional areas are xl ¼ 0:645 cm2 and xu = 64.516 cm2,
Fig. 7. Twenty-five bar truss geometry.
respectively.
(e) Two hundred bar truss: The 200-bar truss configuration is
shown in Fig. 9. The material density is q = 7833 kg/m3 and the Table 2
Young’s modulus is E = 206.84 GPa. The design variables are Load conditions for 25-bar truss.
divided into 29 groups, the grouping details are given in Table 4. Load case Node Direction of load (kN)
The stress constraints are r  k ¼ 206:84 MPa (k ¼ 1; 2; . . . ; 200) in
x y z
both tension and compression and displacement limit is
x ¼ u
u  y ¼ 12:7 mm for all nodes. The initial value for members 1 1 2.224 0 0
2 2.224 0 0
A13, A15, A18, A20, A23, A25 is set as 45.61 cm2, for member
3 4.448 44.482 22.241
A27, A28, A29 is set as 91.22 cm2 and for rest of the members are 4 0 44.482 22.241
all set as 14.45 cm2. There are two loading conditions, the first
2 3 0 88.964 22.241
one is 44.48 kN acting in negative y-direction at node points 1, 2, 4 0 88.964 22.241
3, 4, 5, 6, 8, 10, 12, 14, 15, 16, 17, 18, 19, 20, 22, 24, . . ., 71, 72,
73, 74, 75. The second one is 4.449 kN acting in positive x-direction

at nodes 1, 6, 15, 20, 29, 34, 43, 48, 57, 62, 71. The lower and upper
limits of cross-sectional areas are xl ¼ 0:645 cm2 and xu = 129.032 -
0.254m 0.254m cm2, respectively.
All the numerical results for the truss examples discussed above
are presented in Table 5. Here we show the number of iterations
needed for the four algorithms to converge to the tolerance values
2
of ex ¼ f102 ; 104 ; 106 ; 108 g. As shown in Table 5, DQA approx-
1 3 imations T2:E and T2:R does not necessarily perform better than
MMA in every case. This is due to the inaccurate enforcement
63kN within T2:E and T2:R approximations. However, the AQA perform
x better than T2:E T2:R and MMA in all cases for solving these rep-
322kN resentative truss problems. These examples again demonstrate
y the efficiency of the adaptive approximation schemes used in AQA.
To evaluate the effect of proposed adaptive move limit strategy
Fig. 5. Three bar truss geometry. an additional study with fixed and adaptive move limits is carried
 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147 141

and optimization problem. Moreover, these results also demon-

strate that the benefits of AQA scheme are due to the associated
adaptivity in the AQA algorithm and not due to the adaptive move
limit strategy. The move limit strategy therefore is of less impor-
tance than the previous section suggests.

5.3. Problem-3: Fleury’s weight minimization-like problem

This problem was initially proposed by Fleury [34] and is also

studied by Groenwold [1,35,36]. This so-called Fleury’s weight like
minimization problem is then extended by Siddiqui et al. [37] for
an arbitrary given number of design variables as follows:
X
n
min f 0 ðxÞ ¼ xi
x
i¼1

Subject to :
X
0:95n
1 1 X
n
1
Fig. 8. Thirty-six bar truss geometry. f 1 ðxÞ ¼ þ n60
xi n2 xi ð49Þ
i¼1 i¼0:95nþ1

X
0:95n
1 1 X
n
1
out for two selected truss problems: 36 and 200 bar truss. Table 6 f 2 ðxÞ ¼   0:9n 6 0
i¼1
xi n2 i¼0:95nþ1
xi
shows the iteration numbers needed to reach ex ¼ 108 for differ-
1
ent constant move limits – 0.1, 0.2, 0.3, 0.4 and 0.5 – as well as 6 xi 6 n2 ; i ¼ 1; 2; . . . ; n
for the proposed adaptive move limit strategy. As shown in this n2
table, for a given algorithm, the change in the value of a constant where n is the number of design variables, the starting point is cho-
move limit will alter the convergence steps and there is no obvious sen as x0i ¼ 2 and ex ¼ 108 . The optimal solution is xi ¼ 1:0 for i = 1,
rule to determine which move limit is suitable for a certain algo- 2,. . ., 0.95 n and xi ¼ n12 for i = (0.95 n + 1), . . ., n. The results for dif-
rithm and problem. However, the adaptive move limit strategy ferent algorithms are shown in Table 7 for n ¼ 102, 103, 104, and
that automatically chooses move limit value at each iteration lead 105. In this problem, the gradient of the constraint f 2 ðxÞ can be both
to decent convergence speeds, and thereby obviate an additional positive and negative. As shown in Table 7, for n = 102 MMA takes
step of determining optimal move limits for a given algorithm fewer steps to converge than T2:R and T2:E. But when the number

Table 3
Nodal coordinates, member connectivity, member groups and load conditions for 36-bar truss.

Node x(m) y(m) z(m) Member Node 1 Node 2 Member group

1 0.0 0.0 0.0 1 1 4 1
2 0.0 4.572 7.919 2 4 7 2
3 0.0 4.572 7.919 3 7 10 3
4 9.144 0.0 0.0 4 3 6 4
5 9.144 4.572 7.919 5 6 9 5
6 9.144 4.572 7.919 6 9 12 6
7 18.288 0.0 0.0 7 2 5 4
8 18.288 4.572 7.919 8 5 8 5
9 18.288 4.572 7.919 9 8 11 6
10 27.432 0.0 0.0 10 4 6 7
11 27.432 4.572 7.919 11 7 9 8
12 27.432 4.572 7.919 12 10 12 9
Material property 13 5 6 10
14 8 9 11
E 68.95 GPa 15 11 12 12
q 2768 kg/m3 16 4 5 7
17 7 8 8
Load cases and constraints 18 10 11 9
19 1 6 13
20 3 4 14
21 4 9 15
Load case 1: P1 = 444.82 kN, negative z-direction at node 10 22 6 7 16
23 7 12 17
24 9 10 18
25 3 5 19
Displacement limit: 5.08 cm in negative z-direction at node 10 26 2 6 19
27 6 8 20
28 5 9 20
29 9 11 21
Load case 2: P2 = 444.82 kN, y-direction at node 10 30 8 12 21
31 2 4 13
32 1 5 14
33 5 7 15
Displacement limit: 5.08 cm in y-direction at node 10 34 4 8 16
35 8 10 17
36 7 11 18
142 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147

Fig. 9. Two hundred bar truss geometry.

of variables increases, MMA becomes unstable and even fails to
converge while T2:R and T2:E still converge. However, when using
AQA algorithm, the convergence is achieved in around 7 steps for all
cases, and the constraint violation norm kf c k1 is also smaller for all
DQA algorithms as compared to MMA. Again, AQA algorithm shows
superior convergence performance over other algorithms.
5.4. Problem-4: Topology optimization for minimum compliance
In this section topology optimization for minimum compliance
is considered, and the discrete density-based optimization formu-
lation with design variables defined at an element level in the finite
element setting is given by:
 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147 143

Table 4 Table 6
Member grouping of 200-bar truss. Iteration numbers needed to reach ex = 108 for different values of the move limit D.

Group Member numbers Move limit D 36-bar truss 200-bar truss

number
T2:E T2:R AQA T2:E T2:R AQA
1 1,2,3,4
0.1 89 148 15 55 118 34
2 5,8,11,14,17
0.2 86 148 15 58 122 34
3 19,20,21,22,23,24
0.3 80 148 15 60 124 32
4 18,25,56,63,94,101,132,139,170,177
0.4 81 148 15 62 125 43
5 26,29,32,35,38
0.5 199 148 18 63 125 44
6 6,7,9,10,12,13,15,16,27,28,30,31,33,34,36,37
Adaptive 85 148 15 56 120 35
7 39,40,41,42
8 43,46,49,52,55
9 57,58,59,60,61,62
10 64,67,70,73,76
11 44,45,47,48,50,51,53,54,65,66,68,69,71,72,74,75
12 77,78,79,80 In Eq. (50) the objective function f 0 ðxÞ represents the structural
13 81,84,87,90,93 compliance and the constraint function f 1 ðxÞ limits the total volume
14 95,96,97,98,99,100 of the material within the domain to the desired volume fraction V f .
15 102,105,108,111,114
16 82,83,85,86,88,89,91,92,103,104,106,107,109,110,112,113
There are n elements within the domain and each element has the
17 115,116,117,118 density xi . The lower boundary of design variables is set to
th
18 119,122,125,128,131 xli ¼ 0:001. The volume of i element is v i , and V s is the total vol-
19 133,134,135,136,137,138
ume of the domain. Note that the nested formulation is employed
20 140,143,146,149,152
21 120,121,123,124,126,127,129,130,141,142,144,145,147,148,150,151 here, and the displacement vector uðxÞ is obtained by solving the
22 153,154,155,156 equilibrium equation KðxÞuðxÞ ¼ P [33], where KðxÞ is the stiffness
23 157,160,163,166,169 matrix, P and uðxÞ are the applied force and displacement vector,
24 171,172,173,174,175,176 respectively.
25 178,181,184,187,190
26 158,159,161,162,164,165,167,168,179,180,182,183,185,186,188,189
As an illustrative example, a bridge under uniformly distributed
27 191,192,193,194 load is chosen for minimum compliance design. The problem setup
28 195,197,198,200 and relevant parameters are shown in Fig. 11. The domain is dis-
29 196,199 cretized into 1600  200 bilinear quadrilateral elements, and due
to the symmetry, only half of the domain (800  200) is analyzed.
Density of the bridge deck, which is the top 1=25 of the height, is
fixed to 1.0. The starting density for each element equals to the
Table 5
prescribed volume fraction x0i ¼ V f ¼ 0:25. The Simple Isotropic
Numerical results for the truss problems: number of iterations required to reach the
prescribed termination criteria. Material with Penalization (SIMP) [38,39] interpolation function
is used to interpolate the material stiffness and the stiffness matrix
Problem Algorithm Tolerance ex
is given by:
102 104 106 108
n
3-bar truss T2:E 10 14 18 23 KðxÞ ¼ A xpi K e1i ð51Þ
T2:R 9 16 23 30 i¼1
MMA 7 10 12 15
AQA 9 10 11 12 where A is the standard finite element assembly operator and K e1i is
10-bar truss T2:E 11 37 97 161 the element stiffness matrix with unit Young’s modulus. Here,
T2:R 15 62 177 307 p P 1 is the material penalization coefficient and note that for
MMA 19 35 43 60 p ¼ 1, the intermediate densities are not penalized and the objec-
AQA 8 14 20 28
tive function is convex in this case [33]; however, as p increases
25-bar truss T2:E 5 11 22 33 the intermediate densities are penalized but the objective function
T2:R 7 15 29 54
becomes non-convex. Also, we used the following continuation
MMA 26 31 38 41
AQA 4 9 17 23 scheme for avoiding being trapped into a local minimum:
p ¼ 1 þ 0:1ðk  1Þ starting from iteration k ¼ 1 until p ¼ 3
36-bar truss T2:E 16 30 57 85
T2:R 7 29 87 148 (Fig. 10) and for every p twenty-five iterations are carried out. Sen-
MMA 5 8 10 14 sitivity filter [40] is used to regularize the problem and the filter
AQA 5 8 11 15 radius is 0.02 (4  mesh size); further implementation details can
200-bar truss T2:E 7 23 39 56 be found elsewhere [25,41].
T2:R 14 50 85 120 Note that constraint in this problem is linear, so exact linear
MMA 20 35 43 –
approximation is used for all of the algorithms except MMA since
AQA 10 18 26 35
it convexifies all of the functions. The topologies results from dif-
‘‘–’’ Means no convergence in 500 iterations. ferent algorithms are shown in Fig. 12. As shown in this figure,
MMA algorithm converges to a different minimum which has sub-

min f 0 ðxÞ ¼ P T uðxÞ optimal performance with f 0 ¼ 4752:5 as compared to other algo-
x2D rithms. In T2:⁄ family, the final solutions are similar between T2:R
Subject to : and T2:E, and T2:E solution is little better than T2:R. In minimum
!
1 X n compliance problem sensitivity of the objective is always negative
f 1 ðxÞ ¼ xi v i  Vf 6 0 ð50Þ and the constraint is strictly linear, however, MMA convexifies
V s i¼1
both of the objective and constraint functions, so it has inferior
KðxÞuðxÞ ¼ P performance than T2:⁄ family as expected. Finally, the AQA algo-
x 2 D ¼ fxi jxli 6 xi 6 1; i ¼ 1; . . . ; ng rithm still improves the T2:⁄ family. Although there is little
improvement in this case due to monotonic behavior of the
144 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147

Table 7
Convergence properties for the Fleury-like weight problem.

Number of variables n = 102 n = 103 n = 104 n = 105


Optimal value f0 95.0005 950.00005 9500.000005 95000.0000005
Iteration number T2:R 18 24 31 38
T2:E 6 9 15 21
MMA 14 420 – –
AQA 6 6 7 7
kf c k1 T2:R 9.7913e12 1.0652e10 1.8190e12 2.8056e08
T2:E 1.2648e12 5.9117e12 8.7311e11 2.2803e08
MMA 8.9808e08 4.6668e05 – –
AQA 2.8422e14 5.5707e12 1.1642e10 7.36327e09

‘‘–’’ Means no convergence in 1000 iterations.

behavior. A problem that is much more difficult than the minimum

3.5 compliance problem is the compliant mechanism design problem
which is not self-adjoint so that sensitivities of the objective func-
tion may take both positive and negative values and so this prob-
3
lem is not monotonous anymore. The compliant mechanism design
problem is expressed as:
2.5
min f 0 ðxÞ ¼ LT uðxÞ
Penalization (p)

x2D
2 Subject to :
!
1 X n ð52Þ
1.5 f 1 ðxÞ ¼ xi v i  Vf 6 0
V s i¼1
x 2 D ¼ fxi jxli 6 xi 6 1; i ¼ 1; . . . ; ng
1

0.5
0 20 40 60 80 100
Iteration number (k)

Fig. 10. Continuation scheme for materila penalization parameter p.

Fig. 11. Bridge problem setup and parameters.

compliance function, the result demonstrates the accuracy of adap-

tive scheme of the approximations that are used in AQA, and AQA
is able to choose the ‘‘best’’ scheme that is suitable for a given
problem.

5.5. Problem-5: Topology optimization of compliant mechanisms

Fig. 13. Force inverter problem setup and parameters.

Minimum compliance problem is self-adjoint and sensitivity of
the objective function is always negative resulting in monotonous

Fig. 12. Topologies of bridge problem with different algorithms (1600  200 mesh, filter radius R = 0.02).
 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147 145

(a)

0.5
T2:R
T2:E
0 AQA
MMA
0

−0.5
Objective f

T2:E
−1
T2:R

−1.5
MMA
AQA
−2
0 20 40 60 80 100 Fig. 16. Gripper problem setup and parameters.

Iteration number
(b)
Fig. 14. Convergence history of force inverter problem for different algorithms, un-
penalized case with 120  60 mesh and filter radius R = 4.0: (a) topology; (b)
convergence history.
where the vector L 2 Rn is all zeros except for the output degree of
freedom, where it is set to 1. So in compliant mechanism design
problem the objective is to maximize the displacement in the neg-
ative direction for given applied force and other given parameters.
The lower boundary of the design variables is set to xli ¼ 0:001
and the nested formulation is again adopted with SIMP interpola-
tion function (Eq. (51)). In order to ensure the convexity of approx-
imations when sensitivities are positive, T2:E and T2:R are used
with enforcement given by Eq. (30) and Eq. (31), while MMA can
handle both positive and negative sensitivities. However, AQA
relaxes the enforcements adopted by T2:E and T2:R, so that it’s
expected to have better performance when solving compliance
mechanism problem than T2:E and T2:R.
The first compliant mechanism problem is the design of force
inverter problem shown in Fig. 13 [42]. The domain is discretized
with 120  60 mesh, and first an unpenalized problem with pen-
alty parameter p ¼ 1 is considered and the filter radius is R ¼ 4:0.
In this case the final topology has grey areas (Fig. 14(a)) as the
design variables are not penalized and the convergence history of
objective function for first 100 iterations is shown in Fig. 14(b).
As shown in this figure, MMA has better performance for the first
20 iterations; however, the performance then deteriorates and
the objective reaches to a relatively non-optimal value at iteration
100. Also, AQA does not show its advantage over other DQA algo-
rithms (T2:E and T2:R) in the beginning but it improves faster
starting from 10th iteration. This typically due to the relatively

Fig. 15. Final topologies for force inverter problem with different algorithms and continuation scheme (240  120 mesh with filter radius R = 2.0).
146 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147

highly non-convex case, MMA and T2:E converges to a different

local minima with a better optimal value than T2:R. However,
AQA have a different topology and the AQA result again outper-
forms the optimal value of all other algorithms.
The second problem is the design of gripper shown in Fig. 16
[43]. The gripper surface is fixed to a density of 1.0. The final topol-
ogy for the unpenalized case with 96  60 + 24  30 mesh and fil-
ter radius of 4.0 after 100 iterations is shown in Fig. 17(a) and
(a) convergence results are shown in Fig. 17(b). In this case also,
0.4 AQA have better performance after first 10 steps, and it converges
T2:R faster and eventually to a more optimal value than other algo-
0.2 T2:E rithms. The penalized topologies with continuation scheme are
AQA
shown in Fig. 18 for 192  120 + 48  60 mesh and with filter
MMA
0 radius of 2.0. This result shows that MMA converge to a difference
local minima while T2:E and T2:R have the similar topologies.
Objective f0

−0.2 Obviously, T2:⁄ do not have the better topologies then MMA, but
AQA has the similar topology as compared to MMA with best opti-
−0.4 
mal value of f 0 ¼  0.9016 among all of the algorithms.

−0.6 T2:R
MMA 6. Conclusions and discussion
T2:E
−0.8
AQA In this paper, we have discussed different approximations
−1 employed in dual sequential approximation algorithm for solving
0 20 40 60 80 100 both structural and topology optimization problems. The diagonal
Iteration number quadratic approximation (DQA) is demonstrated to be able to exhi-
bit non-monotonous behavior together with the ability to generate
(b) primal–dual relationship in closed-form. It is shown that inconsis-
tent enforcements made to ensure the convexity of T2:E and T2:R
Fig. 17. Convergence history of gripper problem for different algorithms, un-
penalized case with 96  60 + 24  30 mesh and filter radius R = 4.0: (a) topology; approximations can make the approximations inaccurate and
(b) convergence history. result in slow convergence. To address this issue, an adaptive qua-
dratic (AQA) approximation is proposed by carefully selecting the
approximation for the second-order term based on DQA. In this
low accuracy of T2:F approximation in AQA at the beginning where adaptive approach, the AQA method could automatically select
the difference jxik1  xki j may be large; however as the optimization the most suitable approximation for the particular problem
proceeds the T2:F approximation becomes more accurate. Finally, depending on the gradients and history information. The effective-
the penalized case is considered and Fig. 15 shows the final topol- ness of this approach is shown for various cases. It is demonstrated
ogies when adopting the continuation scheme shown in Fig. 10 that when solving the non-monotonous problems where the gradi-
with 25 iterations at each p and mesh size of 240  120. For this ent signs switches, like truss problems and complaint mechanism

∗ ∗
(a) T2:R, = −0.8850 (b) T2:E, = −0.8888

∗ ∗
(c) MMA, = −0.9009 (d) AQA, = −0.9016

Fig. 18. Final topologies of gripper problem for different algorithms with continuation scheme (192  120 + 48  60 mesh with filter radius R = 2.0).
 L. Li, K. Khandelwal / Computers and Structures 151 (2015) 130–147
topology optimization problems, T2:E and T2:R may not be better
than MMA approximation. However, AQA has the better perfor-
mance than T2:E and T2:R and it also shows its benefits over MMA.
Acknowledgements
The presented work is supported in part by the US National Sci-
ence Foundation through Grant CMS-1055314. Any opinions, find-
ings, conclusions, and recommendations expressed in this paper
are those of the authors and do not necessarily reflect the views
of the sponsors.
References
[1] Groenwold AA, Etman LFP, Wood DW. Approximated approximations for SAO.
Struct Multidiscip Opt 2010;41:39–56.
[2] Schmit LA, Farshi B. Some approximation concepts for structural synthesis.
AIAA J 1974;12:692–9.
[3] Schmit LA. Approximation concepts for efficient structural synthesis. NASA CR-
2552; 1976.
[4] Haftka RT, Shore CP. Approximation methods for combined thermal/structural
design; 1979.
[5] Prasad B. Explicit constraint approximation forms in structural optimization
Part 1: Analyses and projections. Comput Methods Appl Mech Eng
1983;40:1–26.
[6] Prasad B. Explicit constraint approximation forms in structural optimization.
Part 2: Numerical experiences. Comput Methods Appl Mech Eng 1984;46:15–38.
[7] Fleury C, Braibant V. Structural optimization: a new dual method using mixed
variables. Int J Numer Meth Eng 1986;23:409–28.
[8] Falk JE. Lagrange multipliers and nonconvex programs. SIAM J. Control
1969;7:534–5.
[9] Smaoui H, Fleury C, Schmit L. Advances in dual algorithms and convex
approximation methods. In: AIAA/ASME/ASCE/AHS 29th structures, structural
dynamics and materials conference, Williamsburg, VA; 1988. p. 1339–47.
[10] Fleury C. CONLIN: an efficient dual optimizer based on convex approximation
concepts. Struct Opt 1989;1:81–9.
[11] Barthelemy JFM, Haftka RT. Approximation concepts for optimum structural
design—a review. Struct Opt 1993;5:129–44.
[12] Haftka RT, Nachlas JA, Watson LT, Rizzo T, Desai R. Two-point constraint
approximation in structural optimization. Comput Methods Appl Mech Eng
1987;60:289–301.
[13] Fadel GM, Riley MF, Barthelemy JM. Two point exponential approximation
method for structural optimization. Struct Opt 1990;2:117–24.
[14] Svanberg K. Method of moving asymptotes – a new method for structural
optimization. Int J Numer Meth Eng 1987;24:359–73.
[15] Svanberg K. A globally convergent version of MMA without linesearch. In:
Proceedings of the first world congress of structural and multidisciplinary
optimization: Goslar, Germany; 1995. p. 9–16.
[16] Bruyneel M, Duysinx P, Fleury C. A family of MMA approximations for
structural optimization. Struct Multidiscip Opt 2002;24:263–76.
[17] Fleury C. Efficient approximation concepts using second order information. Int
J Numer Meth Eng 1989;28:2041–58.
[18] Wang L, Grandhi RV. Improved two-point function approximations for design
optimization. AIAA J 1995;33:1720–7.
[19] Xu S, Grandhi RV. Effective two-point function approximation for design
optimization. AIAA J 1998;36:2269–75.
147
[20] Svanberg K. A class of globally convergent optimization methods based on
conservative convex separable approximations. SIAM J Optim 2002;12:
555–73.
[21] Bruyneel M, Fleury C. Composite structures optimization using sequential
convex programming. In: ECT and CST, September 1, 2000 – September 1,
2000, Leuven, Belgium: Elsevier Ltd.; 2002. p. 697–711.
[22] Groenwold AA, Etman LFP, Snyman JA, Rooda JE. Incomplete series expansion
for function approximation. Struct Multidiscip Opt 2007;34:21–40.
[23] Groenwold AA, Etman LFP. A quadratic approximation for structural topology
optimization. Int J Numer Meth Eng 2010;82:505–24.
[24] Groenwold AA, Etman LFP. On the equivalence of optimality criterion and
sequential approximate optimization methods in the classical topology layout
problem. Int J Numer Meth Eng 2008;73:297–316.
[25] Li L, Khandelwal K. Two-point gradient-based MMA (TGMMA) algorithm for
topology optimization. Comput Struct 2014;131:34–45.
[26] Bertsekas D. Nonlinear programming: Athena Scientific; 1999.
[27] Beckers M. Dual methods for discrete structural optimization problems. Int J
Numer Meth Eng 2000;48:1761–84.
[28] Haug EJ, Arora JS. Applied optimal design: mechanical and structural
systems. New York: Wiley; 1979.
[29] Alexandrov NM, Dennis Jr JE, Lewis RM, Torczon V. A trust-region framework
for managing the use of approximation models in optimization. Struct Opt
1998;15:16–23.
[30] Wujek B, Renaud J. New adaptive move-limit management strategy for
approximate optimization, Part 1. AIAA J 1998;36:1911–21.
[31] Thomas H, Vanderplaats G, Shyy Y. A study of move limit adjustment
strategies in the approximation concepts approach to structural synthesis. In:
Proc 4th AIAA/USAF/NASA/OAI symposium on multidisciplinary analysis and
optimization, Cleveland, OH, AIAA-92-4839; 1992.
[32] Zhang WH, Domaszewski M, Fleury C. New mixed convex approximation
method with applications for truss configuration optimization. Struct Opt
1998;15:237–41.
[33] Christensen PW, Klarbring A. An introduction to structural
optimization. Berlin: Springer Verlag; 2009.
[34] Fleury C. Structural weight optimization by dual methods of convex
programming. Int J Numer Meth Eng 1979;14:1761–83.
[35] Groenwold AA, Etman LFP. On the conditional acceptance of iterates in SAO
algorithms based on convex separable approximations. Struct Multidiscip Opt
2010;42:165–78.
[36] Groenwold AA, Etman LFP. Sequential approximate optimization using dual
subproblems based on incomplete series expansions. Struct Multidiscip Opt
2008;36:547–70.
[37] Siddiqui S, Azarm S, Gabriel S. A modified Benders decomposition method for
efficient robust optimization under interval uncertainty. Struct Multidiscip
Opt 2011;44:259–75.
[38] Bendsøe MP. Optimal shape design as a material distribution problem. Struct
Opt 1989;1:193–202.
[39] Rozvany GIN, Zhou M. Applications of the COC algorithm in layout
optimization. In: Proceedings of the international conference on engineering
optimization in design process, September 3, 1990 – September 4, 1990.
Karlsruhe, Ger: Publ by Springer-Verlag, Berlin; 1991. p. 59–70.
[40] Sigmund O. Morphology-based black and white filters for topology
optimization. Struct Multidiscip Opt 2007;33:401–24.
[41] Tovar A, Khandelwal K. Topology optimization for minimum compliance using
a control strategy. Eng Struct 2013;48:674–82.
[42] Li L, Khandelwal K. Volume preserving projection filters and continuation
methods in topology optimization. Eng Struct 2015;85:144–61.
[43] Wang F, Lazarov B, Sigmund O. On projection methods, convergence and
robust formulations in topology optimization. Struct Multidiscip Opt
2011;43:767–84.