See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/223171725

A novel multi-objective optimization method based on an approximation

model management technique

Article in Computer Methods in Applied Mechanics and Engineering · June 2008

DOI: 10.1016/j.cma.2007.12.014

CITATIONS READS

65 257

3 authors:

Pei-Gui Liu Xu Han

Chinese Academy of Sciences Hebei University of Technology
112 PUBLICATIONS 2,133 CITATIONS 523 PUBLICATIONS 16,474 CITATIONS

SEE PROFILE SEE PROFILE

Chao Jiang
Hunan University
157 PUBLICATIONS 6,140 CITATIONS

SEE PROFILE

All content following this page was uploaded by Xu Han on 07 July 2018.

The user has requested enhancement of the downloaded file.

 Available online at www.sciencedirect.com

Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731

www.elsevier.com/locate/cma

A novel multi-objective optimization method based

on an approximation model management technique
G.P. Liu, X. Han *, C. Jiang
State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body (Hunan University), Hunan University 410082, PR China

Received 13 April 2007; received in revised form 16 October 2007; accepted 29 December 2007
Available online 3 January 2008

Abstract

In this paper, a novel multi-objective optimization method is suggested based on an approximation model management technique. It
is a sequential approximation method, in which a multi-objective optimization with approximation models subject to design variable
move limits is iterated until convergence. In each iteration step, the approximation models are constructed by the response surface
approximations with the samples which are obtained from the design of experiments, and a Pareto optimal set predicted by the approx-
imations is identified through a multi-objective genetic algorithm. According to the prediction of the approximation models, a move lim-
its updating strategy is employed to determine the design variable move limits for the next iteration. At the end of each iteration step,
some uniform distributed points chosen from the predictive Pareto optimal frontier are verified by the high fidelity models and the
obtained actual Pareto optimal set is stored in an external archive. The high efficiency of the present method is demonstrated by four
different test functions and two engineering applications.
Ó 2008 Elsevier B.V. All rights reserved.

Keywords: Multi-objective optimization; Approximation model management; Micro multi-objective genetic algorithm; Trust region; Engineering opti-
mization

1. Introduction of methods only finds one particular Pareto optimal solu-

tion at a time and they have to be applied many times
Many engineering problems involve multiple objectives for identifying the Pareto optimal set. The other class of
and are treated as multi-objective optimization (MOO) methods is a generating method which tries to find multiple
problems. In MOO, there often exist a set of optimal solu- solutions of the Pareto optimal set in a single simulation
tions, no solutions from which can be said to be better than run. Genetic algorithms (GAs) seem to be the most success-
any other without any further information. This set is ful approach of this class of methods. Many different multi-
known as the non-dominated set or the Pareto optimal objective genetic algorithms (MOGAs) are proposed since
set and its corresponding tradeoff in objective space is 1990s [6–10]. These methods are validated to have good
known as the Pareto optimal frontier which are made up performances in different test functions, but they need a
of the Pareto optimal points [1]. In general, there are two large number of function evaluations to identify the Pareto
main classes of methods to deal with MOO. One is the pref- optimal set during the optimization process.
erence-based methods which convert multiple objectives For engineering MOO problems with expensive high
into a single one by the information about a relative pref- fidelity computational models such as complicated simula-
erence vector of the objectives [2–5]. Apparently, this class tions and detailed analyses, the standard multi-objective
optimization methods mentioned above are hard to per-
*
Corresponding author. Tel.: +86 0731 8823993; fax: +86 0731
form directly because of their high computational cost of
8822051. function evaluations. For this reason, the use of approxi-
E-mail address: [email protected] (X. Han). mation models, which are the corresponding low fidelity

0045-7825/$ - see front matter Ó 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2007.12.014
2720 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731

models, becomes more attractive in these problems. How- accuracy of the approximations. In each iteration step, a
ever, due to the loss of the accuracy, it cannot guarantee MOO problem with approximation models prohibited by
the same Pareto optimal set that could be obtained from the design variable move limits is to be dealt with. First,
the high fidelity models. Therefore, some efforts are made the approximation models are constructed by the response
to minimize the loss arising from using the approximation surface approximations (RSA), and the samples it used are
models [11–16]. Wilson et al. [11] employ surrogate approx- obtained from the design of experiments (DOE). Then a
imations (response surfaces and Kriging models) for com- standard multi-objective optimization method is employed
putationally expensive models to explore the design space to search the Pareto optimal set of the MOO problem with
and identify the Pareto optimal set. This approach mini- approximation models. At the end of each iteration step, a
mizes the loss of the accuracy by validating the approxima- move limits updating approach is adopted to determine the
tions before the optimization procedure. It does not update design variable move limits for the next iteration according
the approximation models according to the predictive to the predictions of the approximations to the actual ones.
information obtained from the optimization process, which The actual Pareto optimal set is verified by the high fidelity
makes it depend much more on the accuracy of the approx- models and stored in an external archive. The efficiency of
imations. If it is used to deal with more complicated prob- the present method is validated in four simulation tests.
lems, it would be more difficult even fail to minimize the Eventually, it is applied in two engineering MOO prob-
loss of the accuracy for approximations before the optimi- lems: one is the structural optimization of a ten-bar truss
zation, then it is hard to obtain the same Pareto optimal set and the other is the optimization of variable binder force
from the approximations as that obtained from the high in a shoe plate of a car front floor forming.
fidelity models. Nain and Deb [12] combine an artificial
neural network approximation with the NSGA-II to enable 2. Statement of the multi-objective optimization problem
the use of GAs on computationally expensive problems.
Cunha and Vieira [13,14] propose two more approaches Generally, an MOO problem can be described as
using neural networks following Nain and Deb’s approach. follows:
Yang et al. [15] propose a GA-based approach with
sequentially updated Kriging models. Recent work by Min ff1 ðxÞ; . . . ; fm ðxÞg;
Knowles [16] proposes an extended efficient global optimi- s:t: gi ðxÞ 6 0; i ¼ 1; . . . ; p;
zation algorithm (ParEGO) for MOO problems using the ð1Þ
hj ðxÞ ¼ 0; j ¼ 1; . . . ; q;
design and analysis of computer experiments (DACE)
T
models which are sequentially updated in the iteration pro- xðLÞ 6 x 6 xðUÞ ; x ¼ ½x1 ; . . . ; xn  ;
cedure. These five methods all make their efforts on captur-
ing the whole features of the real models during the where f, g and h stand for the objective function, inequality
optimization process. They manage the approximation and equality constraint function with the total number
models through a sequential approximation models updat- ofm, p and q, respectively. For an MOO problem, m is usu-
ing strategy in the iteration procedure, but verify the actual ally greater than one. x = [x1, . . . , xn]T stands for the vector
Pareto optimal set after optimization. Successfully applica- of the design variables with the total number of n. x(L) and
tions of these approaches in some engineering MOO x(U) are the lower and upper bounds of x. For engineering
problems are also performed. However, for some more MOO problems, the objective functions, inequality and
complicated MOO problem these approaches could fail equality constraint functions usually are expensive compu-
to capture the whole features of the high fidelity models tational models such as finite element analysis (FEA) and
during the optimization procedure, and then the obtained computational fluid dynamics (CFD) etc.
Pareto optimal set may not be the same one from the high
fidelity models. To a certain extent, their ability of finding 3. Formulation of the multi-objective optimization method
the actual Pareto set depends on the accuracy of the based on an approximation model management technique
approximation models a little more. For efficiently and
effectively solving engineering MOO problems with expen- In this section, the formulation of the multi-objective
sive computational models, a new approach with high optimization method based on an approximation model
efficiency less dependence on the accuracy of the approxi- management technique is presented. The approximation
mation models is required. model management technique is a sequentially updating
In this paper, a novel multi-objective optimization move limits approach which mainly contains three steps
method based on an approximation model management in each iteration, including of constructing the approxima-
technique is suggested to further reduce the dependence tion models, solving the MOO problem with approxima-
on approximations’ accuracy. It manages the approxima- tion models and updating the move limits. In the
tion models through a sequential move limits updating following parts, these steps are discussed in details one by
strategy in the iteration procedure. Efforts are made to one, and then the procedure of the suggested method is pre-
explore the interest region of the design space and find sented. In the last part, comparisons of the present method
actual Pareto optimal set not to minimize the loss of the to other methods are discussed.
 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731 2721

3.1. Constructing the approximation models Min ff~ 1 ðxÞ; . . . ; f~ m ðxÞg

s:t: g~i ðxÞ 6 0; i ¼ 1; . . . ; p;
There are many approximation techniques for con-
~hj ðxÞ ¼ 0; j ¼ 1; . . . ; q;
structing the approximation models, such as RSA, Kriging, n o n o ð6Þ
radial basis functions, artificial neural network etc. In our ðtÞ ðtÞ
max xðLÞ ; xðLÞ 6 x 6 min xðUÞ ; xðUÞ ;
formulation, the quadratic polynomial RSA technique
[17] is adopted to construct the approximation models. A x ¼ ½x1 ; . . . ; xn T ;
quadratic polynomial RSA model, ~rðzÞ, of the high fidelity
where f~ , g~ and ~h are the approximation models constructed
model r(z) can be described as ðtÞ ðtÞ
by the quadratic polynomial RSA. xðLÞ and xðUÞ are the
X
n n X
X n move limits for the tth optimization iteration. They are
~rðzÞ ¼ a0 þ ai z i þ aij zi zj ; ð2Þ the constraints in the design space, which is to ensure the
i¼1 i¼1 j¼1 accuracy of the quadratic approximations in the interest re-
gion of the design space where the Pareto optimal solution
where zi and zj are the design variables with the total num- may exist.
ber of n. a0, ai and aij are the unknown polynomial coeffi- In this paper, the micro multi-objective genetic algo-
cients. There are a total of nc = (n + 1)(n + 2)/2 unknown rithm (lMOGA) developed by Liu et al. [10] is employed
coefficients, so a necessary condition for the proper con- to solve the above problem. It is an efficient multi-objective
struction of this approximation model is that the total genetic algorithm with a small population. To classify the
number of sample points Ns is not smaller than nc. Under population into different non-dominated levels and keep
this condition, the estimation problem for these unknown the diversity in the population, a non-dominated sorting
coefficients could be given in a matrix form as and a crowded-comparison approach are employed. The
crowded-comparison procedure is performed in each non-
r ¼ Za; ð3Þ dominated level by using a crowding distance metric and
each solution in the population is assigned a crowding dis-
where r is the vector of the function values and Z is a ma- tance value. Then each non-dominated level is arranged in
trix given as follows: a descending order of magnitude of the crowding distance
2 3 values. All solutions from the first non-dominated level
ð1Þ ð1Þ 2
1 z1 z2  ðznð1Þ Þ form an external elite population. Once the small popula-
6 7 tion converges, an exploratory operator will be applied to
6 7
Z ¼ 6 ... ..
.
..
.
..
.
..
. 7: ð4Þ the external elite population to explore more non-domi-
4 5
ðn Þ ðn Þ 2 nated solutions near the current non-dominated set, and
1 z1 c z2 c    ðzðncÞ
n Þ
a restart strategy will be subsequently adopted to maintain
the genetic diversity.
Since Ns P nc, the vector of unknown coefficients a can be
determined by

~a ¼ ðZT ZÞ1 ZT r: ð5Þ x2

x2 U
( )
In the RSA approach mentioned above, one of the impor-
The trust region of iteration t+1
tant steps for successfully constructing the approximation
models is to obtain proper samples from the design space x2( U )
t +1
Po( )
t
( )
by DOE. There are many DOE techniques, such as orthog-
onal arrays, Latin hypercube design, and D-optimal design
etc. Latin hypercube design [18] is adopted in this paper Δ(
t +1)

for its property of flexibility. By LHD, when the search

region is changed, the evaluated sample points which are
Δ(
t +1)
projected into the new search region will be inherited.
(t +1)
It can cut down some computational cost of function x2 L
( )

evaluations.

3.2. Solving the multi-objective optimization problem with

Design Space
approximation models x2 L
( )

After the approximation models have been constructed, 0 x1( L )

t +1
x1( U )
t +1
x1 L
( )
x1 U x1
( ) ( )
an efficient method is needed to solve the MOO problem ( )

with approximation models. The sequential approximation Fig. 1. The move limits and the trust region for a two-variable MOO
form of the MOO problem formulated in Eq. (1) is given as problem.
2722 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731

 
3.3. Updating the move limits based on the trust region ðtÞ N PoðtÞ
q ¼  ðtÞ  ; ð7Þ
N Pa
The next step is to update the move limits according to
the prediction of the Pareto optimal set obtained by solv- where PðtÞ
a stands for the Pareto optimal set obtained by the
ing the MOO problem given in Eq. (6). There are many approximation models and PoðtÞ stands for the actual Pareto
move limits updating strategies which can be used to optimal set obtained from PðtÞ a after checking the Pareto
update the move limits [19–22]. In our formulation, the optimality of all
 evaluated
 ðtÞ solutions with the high fidelity
trust region method is used. If Pareto optimal solutions models. N PðtÞ a and N P o represent the number of the
obtained by the approximations are the current actual solutions in PðtÞa and P ðtÞ
o . The value of q(t) is between 0
(t)
Pareto optimal solutions, then the approximations can and 1. If q is close to zero, it means that current approx-
be said to have a good prediction of the current interest imation models have a bad prediction of the current inter-
region in design space, and the more actual Pareto opti- est region in design space, then the trust region radius will
mal solutions exist in the approximation obtained Pareto reduce. Otherwise means that they have a good prediction,
optimal set, the better prediction approximation models and then the trust region radius will enlarge. The trust re-
have. Therefore, a reliability index q(t) will be defined as gion radius is updated with the current ones according to
follows: [23]

Fig. 2. Flowchart of the present method.

 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731 2723
8 ð1Þ
>
> c DðtÞ if qðtÞ < R1 ; xðUÞ with Eqs. (9) and (10). Set t = 1 and the external
< 1
elite population Pe = /;
Dðtþ1Þ ¼ c2 DðtÞ if qðtÞ < R2 ; ð8Þ
>
> (2) Constructing the approximation models:
: ðtÞ
D otherwise; (a) If t = 1, generate Ns sample points in current trust
region of the design space by the optimal LHD,
where R1, R2 2 (0, 1) are updating criterion constants and else generate the sample points with LHD includ-
0 < c1 < 1, c2 > 1 are control constants. These constants ing the inherited points,
can be selected as different values for different problems. (b) Evaluate the new sample points using the high
After the next trust region radius is determined, the move fidelity models,
ðtþ1Þ ðtþ1Þ
limits xðLÞ and xðUÞ for the next iteration can be calcu- (c) Construct the approximation models ff~ 1 ðxÞ;
lated as . . . ; f~ m ðxÞg, f~g1 ðxÞ; . . . ; g~p ðxÞg, f~
h1 ðxÞ; . . . ;
~hq ðxÞg with the quadratic polynomial RSA;
ðtþ1Þ  
xðLÞ ¼ min xðtÞ ðtÞ
o jxo 2 Po
ðtÞ
 Dðtþ1Þ ; ð9Þ (3) Solving the sequential approximation of the MOO
ðtþ1Þ   problem by the lMOGA, and then Identifying a Par-
xðUÞ ¼ max xðtÞ ðtÞ
o jxo 2 Po
ðtÞ
þ Dðtþ1Þ : ð10Þ eto optimal set;
(4) Updating move limits based on the trust region:
For a MOO problem with a two dimensional design space, (a) Choose Na uniform distributing points from the
the move limits and the trust region of optimization itera- obtained Pareto optimal frontier for PðtÞ a and eval-
tion t + 1 is illustrated in Fig. 1. The initial trust region ra- uate them with the high fidelity models,
dius is given by the user beforehand. (b) Check Pareto optimality of all evaluated points
with the high fidelity models and set the obtained
3.4. Procedure of the present method Pareto optimal points to Pe,
(c) Update the trust region of the design space.
The procedure of the present method is outlined in the First, set PðtÞ a \ Pe to a temporary set Pt. If
following steps as illustrated in Fig. 2. Pt = ;, set previous Pareto optimal set Poðt1Þ
to the current one PðtÞ ðtÞ
o , else set Po ¼ Pt . Second,
(1) Initializing: Give the initial trust region radius D(1), calculate the reliability index q(t) and update
the initial point x(0) and constants R1, R2, c1, c2. Ini- the trust region radius D(t+1) according to Eqs.
ð1Þ
tialize Pareto optimal set PðtÞ
o . Then calculate xðLÞ and (7) and (8). Finally, calculate move limits

Sample design space Sample design space Sample in the interest

region of design space

Construct Construct
approximation models Construct
approximation models
approximation models

Validate approximation Optimize

models Optimize

Identify approximate
Identify approximate
Optimize Pareto optimal set
Pareto optimal set

Identify approximate Update approximation

Pareto optimal set Verify actual Pareto
models
optimal set

Verify actual Pareto Verify actual Pareto Update the

optimal set optimal set interest region

Actual Pareto optimal

set

(a) Simple illustration of the (b) Simple illustration of the (c) Simple illustration of
first kind of methods second kind of methods the present method

Fig. 3. Comparison of the present method with other two approximation MOO methods.
2724 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731

ðtþ1Þ ðtþ1Þ
xðLÞ and xðUÞ for next iteration with Eqs. (9) the total number of optimization iterations is greater
and (10); than a given number M, then stop, otherwise, set
(5) Judging the convergence. If D(t+1) is smaller than a t = t + 1, go to step 6;
given minimum of the trust region radius Dmin or (6) Inheriting the evaluated points which fall into the
new trust region, then go to step 2.

Table 1 3.5. Comparing with other approximation MOO methods

Initialization and numerical results of the test functions
Test Ns x(0) D(1) Total no. of point Numerical Most of current approximation MOO methods strive
functions evaluations results
  hard to minimize the loss of the approximation accuracy
1 6
Dx01
;...;
Dxn0
Dx0 80 Fig. 4a
2 2 4 and build sufficiently accurate approximation models,
Dx0 77 Fig. 4b
2 but such efforts lead them to depending much on the accu-
3Dx0 77 Fig. 4c
  4
0
racy of the approximations. The suggested method pro-
Dx01 Dx0n Dx
2 18 2 ;...; 2 4 224 Fig. 5a vides an alternative to directly put its emphasis on
Dx0 221 Fig. 5b exploring more Pareto optimal solutions on a sequentially
2
3Dx0 225 Fig. 5c updating interest region in design space. In this method,
  4
3 10
Dx01 Dx0
; . . . ; 2n Dx0 87 Fig. 6a approximation models with insufficient accuracy can still
2 4
Dx0 89 Fig. 6b obtain the actual Pareto optimal set efficiently, which
2
3Dx0 87 Fig. 6c means that it has much less dependence on the approxima-
  4
4 13
Dx01
;...;
Dx0n Dx0 240 Fig. 7a tion models.
2 2 4
Dx0 234 Fig. 7b
The comparison of the present method with the other
2
3Dx0 two kinds of methods is illustrated in Fig. 3. The procedure
4 242 Fig. 7c
of the first kind of methods which is shown in Fig. 3a tries

Fig. 4. Obtained Pareto optimal points of test function 1 with three different initial trust region radiuses.
 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731 2725

to minimize the loss of the approximation accuracy before is further reduced compared with the other two kinds of
the optimization process through a variety of validation methods.
means, i.e. Wilson et al. [11]. It depends much on the
approximations because of no extra information provided
during the optimization process. The simple process of 4. Numerical tests
the second kind of methods is illustrated in Fig. 3b sequen-
tially updates the approximations during the optimization In this section, the present method is tested with four
process according to the identical approximate Pareto opti- different test functions for its efficiency of exploring the
mal set, i.e. [12–16]. Its ability of finding the actual Pareto design space and identifying the Pareto optimal set. The
optimal set is better than the first kind of methods. How- test functions are taken from some significant literatures
ever, it still strives to build accurate approximation models [24–27]. The first two have convex and non-convex contin-
and verifies the actual Pareto optimal set after the optimi- uous Pareto optimal frontier respectively. The last two
zation process. For this reason, it still depends much on the have convex and non-convex non-smooth Pareto optimal
approximation models. The present method shown in frontier respectively. These test functions are described as
Fig. 3c employs an approximation model management follows:
technique to explore the design space and identify the Par-
eto optimal set, in which not only sequentially updates the (1) Test function 1 [24]
move limits in design space, but also verifies the actual Par-
eto optimal set during the iteration procedure. The evalu- Min f 1 ðx1 ; x2 Þ ¼ ðx1  2Þ2 þ ðx2  1Þ2
ated Pareto optimal solutions are stored in an external f 2 ðx1 ; x2 Þ ¼ x21 þ ðx2  6Þ
2
ð11Þ
archive and the archive keeps being updated at each itera-
tion step. Therefore the dependence on the approximations s:t: 0:4 6 x1 6 1:6; 2 6 x2 6 5:

Fig. 5. Obtained Pareto optimal points of test function 2 with three different initial trust region radiuses.
2726 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731

(2) Test function 2 [25] (4) Test function 4 [27]

! X
2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
3
1
2
Min f 1 ðx1 ; x2 ; x3 Þ ¼ 1  exp  xi  pffiffiffi Min f 1 ðx1 ; x2 ; x3 Þ ¼ 10 exp 0:2 x2i þ x2iþ1
i¼1 3 i¼1
3 
X 
! 0:8
X
3
1
2 ð12Þ f 2 ðx1 ; x2 ; x3 Þ ¼ jxi j þ 5 sin x3i
f 2 ðx1 ; x2 ; x3 Þ ¼ 1  exp  xi þ pffiffiffi i¼1
i¼1 3
s:t:  5 6 x1 ; x2 ; x3 6 5:
s:t:  4 6 x1 ; x2 ; x3 6 4: ð14Þ

(3) Test function 3 [26] In all test functions, the minimum of the trust region radius
Dmin is chosen as 1% of the design space. The maximum
Min f 1 ðx1 ; x2 Þ ¼ ðx1 þ x2  7:5Þ2 þ ðx2  x1 þ 3Þ2 =4 number of iteration M is selected for different test func-
tions. The values of the constants in Eq. (8) are selected
f 2 ðx1 ; x2 Þ ¼ ðx1  1Þ2 =4 þ ðx2  4Þ2 =2 as R1 = 0.4, R2 = 0.7, c1 = 0.25, and c2 = 2. The number
of chosen uniform distributing points Na is 10. Ns, D(1)
s:t: g1 ðx1 ; x2 Þ ¼ ðx1  2Þ3 =2 þ x2  2:5 6 0; and x(0) for each test function are given as those in Table
1, where Dxo = x(U)  x(L).
g1 ðx1 ; x2 Þ ¼ x2 þ x1  3:85  8ðx2  x1 þ 0:65Þ2 6 0;
As listed in Table 1, Fig. 4a–c show the Pareto points of
0 6 x1 6 5; 0 6 x2 6 3:
function 1 obtained by the present method with three dif-
ð13Þ ferent initial trust region radiuses after 80, 77 and 77 point

Fig. 6. Obtained Pareto optimal points of test function 3 with three different initial trust region radiuses.
 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731 2727

evaluations respectively, where M = 7. It can be found that point evaluations, respectively, where M = 14. The present
the actual Pareto frontier and two extreme points have method with the second initial trust region radius also
been found by the present method with all initial trust obtains better results than with the other two. Although
region radiuses. Fig. 5 shows the results of test function 2 for MOO problems with discontinuous Pareto optimal
which have two high order objective functions, where frontier, it is affected by the initial trust region radius, it
M = 11. The actual Pareto optimal frontier and two still can obtain good results with a careful chosen initial
extreme points also have been found by the present method trust region radius.
with three initial trust region radiuses after 224, 221 and The convergent performance and efficiency of the pres-
225 point evaluations, respectively. For function 1 and 2, ent method for high order test functions are shown in
the initial radius does not affect the results much. It demon- Table 2. The convergence metric proposed by Deb et al.
strates that the present method is efficient and effective to [28] measures the extent of convergence of the obtained
find the actual Pareto frontier with continuous Pareto opti- non-dominated set to the corresponding known Pareto
mal frontier.
The numerical results of function three which has two
discontinuous regions in the actual Pareto optimal frontier
Table 2
are shown in Fig. 6a–c which are obtained with three initial
The convergent performance of the present method for high order test
trust region radiuses after 87, 89, and 87 point evaluations, functions
respectively, where M = 8. With the second initial trust
Test Total no. of point evaluations Convergent metric
region radius the obtained Pareto optimal points seem to functions
The lMOGA The lMOGA
be better than with the other two. Test function 4 has
present without present without
two high order objective functions and three discontinuous method approximations method approximations
regions in the actual Pareto optimal frontier. Its numerical
2 221 3000 0.00473 0.00490
results are shown in Fig. 7a–c which are obtained with 4 234 2300 0.08998 0.09306
three initial trust region radiuses after 240, 234, and 242

Fig. 7. Obtained Pareto optimal points of test problem 4 with three different initial trust region radiuses.
2728 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731

optimal set. The present method’s convergent performance 5. Engineering applications

is as good as that of lMOGA, while its number of point
evaluations is almost ten times less than that of lMOGA. In this section, two engineering applications of the sug-
gested method are presented to demonstrate its efficiency.
One is the structure optimization of a ten-bar truss; the
other is the optimization of variable binder force in a shoe
plate of the car front floor forming.

5.1. The optimization of ten-bar truss

A well-known ten-bar truss structure with the node and

element numbering shown in Fig. 8 is adopted here. It is
used as an application by many researchers, e.g. [29–31].
The location of external load shown is F = 444.8 kN. The
modulus of the truss is E = 68948 MPa. There are two
objectives: one is to minimize the volume of the structure;
the other is to minimize the vertical displacement at node
6, which is expressed by Dy6. Cross-sectional area of trusses
(A1 to A10) is chosen as design variables. Suppose
Fig. 8. Ten-bar truss structure. A1 = A2 = A3 = A4 = A5 = A6, A7 = A8, and A9 = A10.
Thus, there are practically three design variables (A1, A7,
and A9). One constraint on the truss is taken as limiting
the maximum of the principal stress ri in each truss below
the allowable stress ra, which is set to 172.37 MPa. The sta-
tic analysis is done by ANSYS. The ten-bar truss of the
MOO problem is formulated as follows:
Min f 1 ðA1 ; A7 ; A9 Þ ¼ 6A1  9:144 þ 2A7  9:144
pffiffiffi pffiffiffi
 2 þ 2A9  9:144  2
f 2 ðA1 ; A7 ; A9 Þ ¼ Dy 6
maxfjri j; i ¼ 1; . . . ; 10g
s:t: gðA1 ; A7 ; A9 Þ ¼  1 6 0;
ra
6:45  105 6 A1 ; A7 ; A9 6 1:93548  102 :
ð15Þ
(1)
The parameters are set as: Ns = 12, D = [9.6451 
103, 9.6451  103, 9.6451  103], and x(0) = (9.7097 
Fig. 9. Obtained Pareto optimal points of ten-bar truss MOO problem.
103, 9.7097  103, 9.7097  103).

Fig. 10. The geometry of a shoe plate of a car front floor.

 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731 2729

The Pareto optimal frontier obtained through the pres- Min f 1 ðBFL ; BFH ; T s Þ ¼ maxðDhdÞ
ent method after 189 point evaluations and that obtained f 2 ðBFL ; BFH ; T s Þ ¼ maxðDha Þ
through lMOGA with the high fidelity models after 5000
s:t: 10 KN 6 BFL 6 100 KN; ð16Þ
point evaluations are given in Fig. 9. The present method
has found most part of the Pareto optimal frontier which 100 KN 6 BFH 6 200 KN:
found by lMOGA with the high fidelity models and it 0:5T e 6 T s 6 0:9T e ;
reduces about 26 times of point evaluations.
where max (Dhd) stands for the maximum reduction of
thickness and max (Dha) stands for the maximum increase
of thickness.
5.2. Application for the optimization of variable binder force
The stamping model of the shoe plate is shown in
Fig. 12. The material properties are: E = 206 GPa, m =
In the following part, the presented method is applied
0.3, l = 0.144, r = 524.6(0.007117 + ep)0.2589 MPa. The
to a more complicated engineering MOO problem. It is
punch velocity is 5000 m/s. To describe the blank, the Bely-
the optimization of variable binder force in a car sheet
tschko-Tsay element with five integration points in the
metal forming. The geometry of a shoe plate of a car front
thickness direction was used. The 3-parameter Barlat mate-
floor is shown in Fig. 10. The dimensions of mild steel
rial model was selected to show anisotropic elastic–plastic
blank are 1110 mm length, 482 mm width, and 1.2 mm
properties of the blank. Stamping process simulation was
thickness.
done by LS-DYNA.
Apparently, there are two main defects during the
The parameters are set as: Ns = 12, D(1) = [45, 50, 0.2],
forming process of this shoe plate: wrinkling and cracking.
and x(0) = (55, 150, 0.7). The Pareto optimal frontier
Applying reasonable variable binder force in the forming
obtained with the present method after 58 point evalua-
process is one of the useful ways to overcome these
defects. A step variable binder force curve shown in
Fig. 11 is applied [32]. It contains three characteristic
parameters which are chosen as the design variables: low Punch
binder force BFL, high binder force BFH and the timeTs
at which BF is changed. Te stands for the end time of Holder
the forming process. The variety of blank thickness may
determine the forming quality. Cracking and wrinkling
are affected by the reduction and increase of thickness, Blank
respectively. The maximum reduction and increase of
thickness may evaluate cracking and wrinkling, so they
are selected as two objectives of this optimization prob-
lem. There are many other ways of evaluating cracking Die
and wrinkling, however it is not the main content of this
paper, so it will not be discussed here.
Fig. 12. The shoe plate stamping model.
The MOO problem of the binder force in the shoe plate
forming is formulated as follows:

BF

BFH

BFL

Ts Te T
Fig. 13. Obtained Pareto optimal points of the binder force MOO
Fig. 11. The step variable binder force curve. problem in the shoe plate forming.
2730 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731

tions are given in Fig. 13. The optimization with the [7] N. Srinivas, K. Deb, Multiobjective optimization using nondominat-
numerical model does not perform here since the numerical ed sorting in genetic algorithms, IEEE Trans. Evolut. Comput. 2 (3)
(1994) 221–248.
model is time consuming. The present method can obtain a [8] E. Zitzler, L. Thiele, Multiobjective evolutionary algorithms: a
set of binder force curves which have the maximum reduc- comparative case study and the strength Pareto approach, IEEE
tion of thickness from 33% to 40% and maximum increase Trans. Evolut. Comput. 3 (4) (1999) 292–301.
of thickness from 8% to 15%. The preference of the two [9] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist
objectives can be given to choose the final best solution multiobjective genetic algorithm: NSGAII, IEEE Trans. Evolut.
Comput. 6 (2) (2002) 182–197.
from the obtained Pareto optimal set. The simulation [10] G.P. Liu, X. Han, A micro multi-objective genetic algorithm for
results demonstrate that the present method is also efficient multi-objective optimizations, in: The Fourth China-Japan–Korea
for complex engineering MOO problems. Joint Symp. on Optimization of Structural and Mechanical Systems,
Kunming, China, 2006.
[11] B. Wilson, D.J. Cappelleri, T.W. Simpson, M.I. Frecker, Efficient
6. Conclusions Pareto frontier exploration using surrogate approximations, in: 8th
AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Anal-
In this paper, a multi-objective optimization method ysis and Optimization, Long Beach, CA, 2000.
based on an approximation model management technique [12] P.K.S. Nain, K. Deb, A computationally effective multi-objective
search and optimization technique using coarse-to-fine grain model-
is presented. By the using of the trust region move limits
ing, IITK, Kanpur, India, Tech. Rep. Kangal Rep. 2002005, 2002.
updating strategy, it guarantees the accuracy of the [13] A.G. Cunha, A. Vieira, A multi-objective evolutionary algorithm
approximations in the interest region of the design space using approximate fitness evaluation, in: G. Bugeda, J.A. Dsidri, J.
where the actual Pareto optimal solutions may exist. It puts Periaux, M. Schoenauer, G. Winter, Eds., Proc. Int. Congr. Evol.
emphasis on obtaining the actual Pareto optimal set Methods for Design, Optimization and Control With Applications to
Industrial Problems EUROGEN, 2003.
instead of catching the main characteristics of the high
[14] A.G. Cunha, A.S. Vieira, A hybrid multi-objective evolutionary
fidelity models. Therefore the present method has much algorithm using an inverse neural network, in: Proc. Hybrid Meta-
less dependence on the accuracy of the approximation heuristics (HM 2004) Workshop at ECAI 2004, Valencia, Spain,
models than other two approximation methods. The simu- 2004, pp. 25–30.
lation results of four test functions demonstrate that the [15] B.S. Yang, Y.S. Yeun, W.S. Ruy, Managing approximation models in
multiobjective optimization, Struct. Multidiscipl. Optim. 24 (2) (2002)
present method can efficiently find the actual Pareto opti-
141–156.
mal set with insufficient accurate approximation models. [16] Joshua Knowles, ParEGO: a hybrid algorithm with on-line landscape
It should be mentioned that the initial trust region radius approximation for expensive multiobjective optimization problems,
must be carefully chosen for MOO problems with discon- IEEE Trans. Evolut. Comput. 10 (1) (2005) 50–66.
tinuous Pareto optimal frontier. Two engineering applica- [17] N.R. Draper, H. Smith, Applied Regression Analysis, Wiley, New
York, 1998.
tions have further demonstrated that this method is
[18] M.D. Morris, T.J. Mitchell, Exploratory designs for computational
efficient for engineering MOO problems, even for more experiments, J. Statist. Planning Inference 43 (3) (1995) 381–
complicated ones. 402.
[19] C.L. Bloebaum, W. Hong, A. Peck, Improved moved limit strategy
Acknowledgement for approximate optimization, in: Proc. of the Fifth AIAA/USAF/
NASA/ISSOM Symposium, Panama City, FL, 1994.
[20] T.Y. Chen, Calculation of the move limits for the sequential linear
This work is supported by the national 973 program un- programming method, Int. J. Numer. Methods Engrg. 36 (1993)
der the Grant no. 2004CB719402, the program for New 2661–2679.
Century Excellent Talents in University through Grant [21] G.M. Fadel, M.F. Riley, J.F.M. Barthelemy, Two point exponential
NCET-04-0766, and National Natural fund under the approximation method for structural optimization, Struct. Optim. 2
(1990) 117–124.
Grant no. 60635020.
[22] J.E. Dennis, R.B. Schnabel, Numerical Methods for Unconstrained
Optimization and Nonlinear Equations, Prentice-Hall, Engelwood
References Cliffs, NJ, 1983.
[23] N.M. Alexandrov, J.E. Dennis Jr., R.M. Lewis, V. Torczon, A trust-
[1] K. Deb, Multi-objective Optimization using Evolutionary Algo- region framework for managing the use of approximation models in
rithms, John Wiley & Sons Ltd., England, 2001. optimization, Struct. Optim. 15 (1998) 16–23.
[2] L. Zadeh, Optimality and non-scalar-valued performance criteria, [24] Y. Li, G.M. Fadel, M.M. Wiecek, Approximating Pareto curves
IEEE Trans. Automat. Control 8 (1963) 59–60. using the hyper-ellipse, in: 7th AIAA/USAF/NASA/ISSMO Sympo-
[3] Y.Y. Haimes, L.S. Lasdon, D.A. Wismer, On a bicriterion formu- sium on Multidisciplinary Analysis and Optimization, St. Louis,
lation of the problems of integrated system identification and system 1998.
optimization, IEEE Trans. Syst. Man Cybernet 1 (3) (1971) 296–297. [25] R.V. Tappeta, W.F. Rosenberger, Interactive multiobjective optimi-
[4] M. Zeleny, Compromise programming, in: J.L. Cochrane, M. Zeleny zation design strategy for decision based design, J. Mech. Des., Trans.
(Eds.), Multiple Criteria Decision Making, University of South ASME 123 (2) (2001) 205–215.
Carolina Press, Columbia, 1973. [26] C.M. Fonseca, P.J. Fleming, Multiobjective optimization and multi-
[5] H.W. Corley, A new scalar equivalence for Pareto optimization, ple constraint handling with evolutionary algorithms – Part II:
IEEE Trans. Automat. Control 25 (4) (1980) 829–830. application example, IEEE Trans. Syst. Man Cybern. A 28 (1) (1998)
[6] C.M. Fonseca, P.J. Fleming, Genetic algorithms for multiobjective 38–47.
optimization: formulation, discussion and generalization, in: Proc. of [27] F. Kursawe, A variant of evolution strategies for vector optimization,
the Fifth Int. Conf. on Genetic Algorithms, San Mateo, CA, 1993. in: Parallel Problem Solving from Nature, Springer-Verlag, 1990.
 G.P. Liu et al. / Comput. Methods Appl. Mech. Engrg. 197 (2008) 2719–2731 2731

[28] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist [31] F.T.K. Au, Y.S. Cheng, L.G. Tham, G.W. Zeng, Robust design of
multiobjective genetic algorithm: NSGAII, IEEE Trans. Evolut. structures using convex models, Comput. Struct. 81 (28) (2003) 2611–
Comput. 6 (2) (2002) 182–197. 2619.
[29] G.C. Luh, C.H. Chueh, Multi-objective optimal design of truss [32] G. Liu, Z.Q. Lin, W.L. Xu, Y.X. Bao, Variable blankholder force in
structure with immune algorithm, Comput. Struct. 82 (11) (2004) U-shaped part forming for eliminating springback error, J. Mater.
829–844. Process. Technol. 120 (1) (2002) 259–264.
[30] W.J. Roux, N. Stander, R.T. Haftka, Response surface approxima-
tions for structural optimization, Int. J. Numer. Methods Engrg. 42
(3) (1998) 517–534.

View publication stats