Beruflich Dokumente
Kultur Dokumente
Department of Mechanical Engineering, Graduate School, Hanyang University, 17 Haengdang-Dong, Seongdong-Gu, Seoul 133-791, South Korea
School of Mechanical Engineering, Hanyang University, 17 Haengdang-Dong, Seongdong-Gu, Seoul 133-791, South Korea
a r t i c l e
i n f o
Article history:
Received 31 May 2011
Accepted 15 April 2013
Available online 15 May 2013
Keywords:
Modied ant colony optimization (MACO)
Ant colony optimization (ACO)
Dynamic topology optimization for natural
frequencies
Finite element method
a b s t r a c t
A modied ant colony optimization (MACO) algorithm implementing a new denition of pheromone and
a new cooperation mechanism between ants is presented in this paper. The sensitivity of structural
response to the presence of each element included in the nite element (FE) model is evaluated. The
study aims to improve the suitability and computational efciency of the ant colony optimization algorithm in dynamic topology optimization problems. The natural frequencies of the structure must be
maximized yet satisfying a constraint on the nal volume. Optimization results obtained in three test
cases indicate that MACO is more efcient and robust than standard ACO in solving dynamic topology
optimization problems.
2013 Elsevier Ltd. All rights reserved.
1. Introduction
The ant colony optimization (ACO) algorithm is a metaheuristic
search method for global optimization. ACO was initially proposed
by Dorigo [1] to nd the optimal path in a graph. ACO mimics the
behavior of ants seeking a path between their colony and a food
source. This optimization technique was successfully applied to
many engineering problems including structural optimization
[24].
Topology optimization of structures underwent tremendous
development after the introduction of the homogenization method [5]. In addition, ad hoc optimization algorithms such as evolutionary structural optimization (ESO) [6,7], performance-based
optimization (PBO) [8,9], and level set theory [10,11] were successfully applied to topology optimization problems.
The homogenization method divides the design space in an innite number of microscale holes and optimal topology is obtained
by solving a material distribution problem. However, it may yield
an undesirable structure with innitesimal pores in the materials
that make the structure not realizable practically [12]. The
rationale of ESO and PBO is to remove gradually from the structure,
the inefcient elements hosting very low stress that hence do not
contribute much to the overall response of the structure: such an
operation is governed by the sensitivity number. Convergence
behavior of PBO depends on the performance of the structure. Bidirectional ESO (BESO) [13] was developed to improve the capability
Corresponding author.
E-mail address: syhan@hanyang.ac.kr (S.-Y. Han).
0045-7949/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.compstruc.2013.04.012
of adding or removing elements. However, the computational efciency of BESO depends on the previous positions of the elements
as well as on the area/volume of the element predened by the
meshing operation.
Topology optimization can be performed with metaheuristic
algorithms that mimic natural phenomena and physical processes.
Applications of genetic algorithms (GA) [14] and simulated annealing (SA) [15] to topology optimization problems were reviewed in
the paper written by Luh and Lin [16]. Also, ant colony optimization
was utilized in topology optimization problems [16,17]. For example, Luh and Lin [16] used the element transition rule instead of
node transition rule and connectivity analysis, pheromone updating rule and multiple-colony memories. Kaveh et al. [17] developed
a topology optimization technique to nd the stiffest structure with
a certain amount of material; the search criterion was based on the
contribution of each element to strain energy. It was found that
ACO can handle the topology optimization problem as an on-off
discrete optimization. However, there are not many other examples
of applying ACO to topology optimization problems documented in
literature.
Although dynamic topology optimization with respect to natural frequencies is of fundamental importance in aerospace and
automotive engineering, the number of papers published on this
subject is limited in comparison with the available literature on
static problems. For example, the homogenization method
[18,19] and modied SIMP (Solid Isotropic Microstructure with
Penalization) with a discontinuous function [2022] were successfully used to solve eigenvalue problems in topology design of
vibrating structures.
69
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
In this study, a modied ant colony optimization (MACO) algorithm is developed in order to improve computational efciency
and suitability of ACO in topology optimization problems dealing
with natural frequencies. An important improvement with respect
to classical ACO is the denition of a continuous variable, the element contribution signicance (ECS), which serves to evaluate the
effective contribution to structural response deriving from the
presence of each element. A mesh independent ltering scheme
[23] is adopted to prevent the formation of checkerboard patterns
in the optimization process. Optimal designs are compared with
those obtained with standard ACO and soft-kill BESO in order to assess the applicability and the efciency of the proposed MACO
algorithm in dynamic problems.
qxi xi q1
qxmin xmin q1
Exmin xmin E1
xmin xpmin
1 xpmin
8
Exi
Excessive vibration due to resonance occurs when the frequency of the dynamic excitation is close to a natural frequency
of the structure. Therefore, it is necessary to restrict the fundamental frequency or several of the lower frequencies of the structure to
a prescribed range in order to avoid severe vibration.
In the nite element method, the dynamic behavior of a structure
is represented by the following general eigenvalue problem [24];
K x
2
i Mfui g
f0g
where [K] is the global stiffness matrix, [M] is the global mass matrix, xi is the ith natural frequency and {ui} is the eigenvector corresponding to xi. The natural frequency xi and the corresponding
eigenvector {ui} are related by the Rayleigh quotient as follows;
x2i
fui gT Kfui g
fui gT Mfui g
Subject to : V
xi
N
X
V i xi 0
i1
1
xmin
if element is solid
if element is void
where Vi is the volume of an individual element, and V is the prescribed volume. N is the total number of elements in the structure.
The binary design variable xi denotes the density of the ith element;
xmin is set equal to a sufciently small value to denote a void
element.
2.2. Material interpolation scheme
To obtain the gradient information of the design variable, it is
necessary to interpolate the material between xmin and 1. A popular material interpolation scheme is the so-called power-law
penalization model (the SIMP model). For the solidvoid design,
the material density and Youngs modulus are functions of the design variable xi as [24];
qxi xi q1
Exi
xpi E1 0
1
8
>
x2i
1 T
>
>
ui ;
u
K
M
i
i
>
i
>
p
< 2xi
xi
>
uTi M i ui ;
>
>
>
2p
>
:
0 < xmin 6 xi 6 1
x 1
x xmin
10
where Mi is the mass matrix of each element, Ki is the stiffness matrix of each element, and ui is the element eigenvector which are related to the removed each element.
Maximize : xi
Dx i
1 xpi xpi E1
4
< xmin 6 xi 6 1
1
where q and E are the density and Youngs modulus of solid material, respectively. p is the penalty factor, usually used as 3.
70
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
si ta
ai t PN
a
j1 sj t
11
where a is a parameter that controls the relative weight of the pheromone trail, usually set equal to 1. N is the total number of nite
elements and t refers to an indication of the present cycle which
is analogous to the tth time of deploying our ants. It should be noted
that here the probability of an element being chosen by a typical ant
is the same as the decision index as dened in Eq. (11).
After completion of a cycle of designs by all ants, each ant k
deposits a quantity of pheromone Dski on each element based on
its relative natural frequency, as shown in Eq. (12). This is an index
of the performance of the element: a larger amount of pheromone
is deposited in purpose of improving the design.
Dski
D ki k
PN
k k
j1 D j
No
No
12
where the k exponent is a tuning parameter dened in order to optimize the performance and convergence behavior of the optimization algorithm. In static topology optimization, the usual value
k 2 is dictated by considerations on the convergence rate as the
strain energy used as pheromone is always positive. However, the
value k 1 should be considered in dynamic topology optimization
problems if some of Dxki s are not very small negative values, even
though the convergence rate may be slow. In other words, if the
magnitude of Dxki is a very small negative value (negligible) or it
is always positive, k 2 can also be used in dynamic topology optimization problems to accelerate the convergence rate.
The amount of pheromone in each element is due to the addition of new pheromones as well as evaporation which is implemented within the algorithm via the following rules,
si t 1 csi t Dsi
Pm
Yes
13
where Dsi k1 Ds and m is the number of ants used in each cycle. The evaporation coefcient c, usually set between 0.5 and 0.9,
takes into account the pheromone decay [17]. This serves to avoid
premature convergence to suboptimal designs. By increasing c, convergence behavior becomes stable but convergence rate decreases.
An initial amount of pheromone si(0) is introduced and a small positive constant value s(0) is considered for all elements in the rst
cycle [26].
The ACO process is summarized by the owchart of Fig. 1. The
algorithm includes the following steps:
k
i
objective function should hence be compared with each convergence error limit.
max :
si t si t 1
6e
s t
P 0 i
N
i1 xki1 xkN0 i1
error
PN 0
i1
xki1
14
6 econ
15
where e is the convergence limit set for the design vector (in this
study, e = 0.01%), x is the eigenfrequency that must be maximized
in the cost function, econ is the convergence limit set for the cost
function (in this study, econ = 0.1%), and N is the integer number
resulting in a stable frequency in at least 10 successive iterations.
In this study N was set set equal to 5: therefore natural frequency
must change marginally over the last 10 optimization iterations.
3.2. Modied ant colony optimization (MACO) algorithm for dynamic
topology optimization
As ants are randomly located in the elements of the structure
during the optimization process, their positions are expressed as
solid (1) or void (0) elements in the design domain. The best
topology found in each iteration includes only solid elements. This
allows us to obtain a stable topology if the target volume is large:
in such a case, symmetry of stiffness matrix can be preserved as
there are enough solid elements to represent the general topology
of the structure. However, this is not true if the target volume is set
to a small value. In such a case, computation efciency decreases
71
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
either in terms of stiffness matrix ill-conditioning (values of natural frequencies are not accurate) and high computation time.
In order to solve the above mentioned problems, design variables must be dened in a continuous distribution of density
[1820]. For this reason, a modied ant colony optimization algorithm (MACO) formulation was developed in this research. The
main difference between standard ACO and MACO is the introduction of a continuous variable, called Element Contribution Significance (ECS), in FEA. This new variable serves to replace the
positions of ants in a standard ACO. It serves to assess and monitor
the importance of each element in the optimization process. If an
element is always recognized as solid it contributes signicantly
to the structural response and concurs to drive the search process
towards the optimum design. Hence, it will always be included in
the population of ants used to update the distribution of material
in the structure.
For each element, the ECSi variable is dened as the ratio between the total number of ants passing through each element
and the inner loop number. That is:
PNiter
ECSi
iter1 Aiter i
Niter
A 1 or 0
16
Basically, the (Aiter)i term indicates if an ant was passing (if the
ant exists, A = 1; if the ant does not exist, A = 0) through the ith
element over the Niter optimization iterations completed up to this
moment. Convergence behavior depends on the ant decision index
No
Dxnew
ECSi Dxi
i
17
Dxnew
i
Min v alue
maxDxnew
18
No
Yes
72
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
Dx i
PNnode
wr ik Dxnode
k
PNnode
wr
ik
k1
k1
where, Dxnode
PNk
V Dxi element
Pi Nk
i1
19
i1
Vi
design and the number of optimization cycles for the rst variant
of test problem 1. The number of optimization cycles is indicated
between parentheses.
It can be seen that MACO is insensitive to the a exponent included in the ant decision index ai(t). This is because the a exponent does not affect the relative probability that an element will
be chosen by a typical ant. The k exponent plays the most important role. Based on the sign and magnitude of the changes in pheromone values, convergence behavior may signicantly depend on
the k exponent. As far as it concerns the evaporation coefcient
c, convergence behavior becomes stable but the number of optimization cycles increases as the evaporation coefcient increases
from 0.5 to 0.9 for a given value of k. In summary, the convergence
rate decreases as c increases. Through numerical experiments, the
best combination of MACO internal parameters was found for each
test problem based on the optimized design and the number of
optimization cycles.
The original BESO code, a soft-kill BESO implemented in MATLAB [25], along with the FEA routine built-in in BESO was utilized
in this research. Optimization runs were carried out on a standard
PC equipped with Core2 Quad Q8300 (2.5 GHz) and 2G of RAM
memory. Convergence analysis [23] in BESO was carried out for
all test problems in order to obtain mesh independent nite element solutions. It was performed for several mesh sizes and ER
with rmin = 3 for each example. After examining the optimized
topologies, the best combination of mesh size and ER for mesh
independent nite element solutions was determined.
The best combination of internal parameters for each problem
was selected in order to optimize convergence behavior of all optimization algorithms compared in this study. A lter scheme to prevent noise effects in optimized topology was utilized.
4.1. Test problem 1: short beam optimized for rst or second natural
frequency
The topology of the 5 m 1 m 0.01 m short beam clamped on
both sides shown in Fig. 4 must be optimized in order to maximize
the rst or second natural frequency. The volume of the structure
must be 90% of the original volume. Material properties are as fol-
Fig. 4. Schematic of the short beam under plane stress conditions optimized in test
problem 1.
Fig. 3. Sensitivity of optimized design to MACO internal parameters (rst variant of test problem 1).
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
73
74
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
Table 1
Optimization results for test problem 1.
Algorithm
First
frequency
(Hz)
Second
frequency (Hz)
ACO
BESO
MACO
7.439
8.371
7.947
10.311
10.611
10.855
77/28
22/23
56/18
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
75
ACO did not incur in this problem. Since MACO utilized the ECS
parameter as a continuous variable, it was able to complete the
optimization even for the tighter volume constraint (50%). Soft-kill
BESO also could complete the dynamic topology optimization
problem. It should be noted that ECS describes the tendency of ants
to cross an element. Standard ACO could not have accounted
Fig. 10. Convergence curves for test problem 2 (rst natural frequency).
76
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
properly for the real contribution of each element to the global dynamic response of the structure. This becomes more signicant as
the design domain is smaller.
The difference in convergence behavior of MACO and BESO is
due to the fact that BESO removes elements from the design domain always keeping the same elimination ratio (ER). Therefore,
the natural frequency computed by BESO changes monotonically
as the optimization progresses. Conversely, MACO searches for
effective elements to satisfy the target volume constraint until
the optimal topology is obtained. This leads to have non-monotonic convergence curves whose variation is also inuenced by
the considerable level of heuristics included in the ant colony
algorithm.
Optimal topologies obtained by the above mentioned two algorithms are presented in Fig. 11. Again, the upper part of each subgure shows the topology found in the rst optimization cycle
while the bottom part of each sub-gure shows the optimized
topology. Whilst the rst intermediate designs are totally different
(BESO did not change signicantly topology with respect to the initial design), optimal topologies are much more similar.
Table 2 lists the values of natural frequency corresponding to
the optimal topology found by each optimization algorithm and
the number of structural analyses required in the optimization
process. It can be seen that MACO designed the structure with
Table 2
Optimization results for test problem 2.
Algorithm
BESO
MACO
9.469
9.756
50
50
Fig. 12. Schematic of the clamped beam with a concentrated mass optimized in test problem 3.
Fig. 13. Convergence curves for test problem 3 (rst natural frequency).
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
77
mal topologies are almost the same. Table 3 lists the values of natural frequency corresponding to the optimal topology found by
MACO and BESO, and the number of structural analyses required
in the optimization process. It can be seen that BESO and MACO
had almost the same efciency in converging to the optimum design. In fact, MACO designed a structure with almost the same rst
natural frequency (0.0414 Hz of BESO vs. 0.0405 Hz of MACO).
4.4. Discussion
Optimization results indicate that MACO is considerably more
efcient than ACO and competitive with BESO. The superiority of
MACO over ACO derives from the fact that the present algorithm
searches for optimal topology by taking into account the real contribution of each single element included in the topology domain
to the overall structural response. This makes it is easier to eliminate the parts of the structure that do not provide stiffness where it
is necessary.
As far as it concerns the relative behavior of MACO and BESO, it
should be noted that the performance of meta-heuristic algorithms
can improve signicantly if a population of very efcient trial designs were created since the very beginning of the search process.
Furthermore, the pheromone parameter should always keep the
same sign as it happened for the second variant of test problem 1.
Conversely, even in presence of an initial population of very
efcient trial designs, convergence rate may be very low if pheromone parameter values vary between 1 and 1, and the evaporation coefcient is set to a large value. This was the case of the rst
variant of test problem 1. When a severe volume constraint is included in the design problem like in the case of test problems 2
and 3, MACO is fully competitive with BESO in terms of natural frequency and convergence rate provided that pheromone parameter
values do not change in sign and the evaporation coefcient is
small.
In summary, MACO is suited for dynamic topology optimization
as it could always maximize the natural frequency of the structure
Table 3
Optimization results for test problem 3.
Algorithm
BESO
MACO
0.0414
0.0405
42
43
This paper presented a modied ant colony optimization algorithm (MACO) for dynamic topology optimization of 2D structures.
MACO implemented a novel ant colony optimization formulation
where the concept of ant position is reformulated in terms of the
sensitivity of structural response to the presence of each element.
This effect is evaluated by introducing the ECS parameter.
Numerical tests demonstrated that MACO is a robust and stable
metaheuristic algorithm highly suited for topology optimization in
dynamic problems where the objective is to maximize natural frequencies. In particular, the proposed algorithm was absolutely
superior over standard ant colony optimization and very competitive with bilinear evolutionary structural optimization algorithms
in terms of convergence rate. The natural frequencies corresponding to the designs optimized by MACO are higher than those obtained by ACO, and almost the same or even slightly higher than
for BESO.
However, further research will be required in order to improve
the convergence speed of MACO that was much faster than ACO
but sometimes considerably slower than BESO. Natural frequencies
optimized by MACO were greater than those found by standard
ACO but sometimes slightly smaller than those found by BESO.
References
[1] Dorigo M, Learning and natural algorithms (in Italian), Ph.D thesis. Politecnico
di Milano, Italie, 1992.
[2] Dorigo M, Stutzle T. Ant colony optimization. Cambridge (MA), USA: MIT Press;
2004.
[3] Blum C. Ant colony optimization: introduction and recent trends. Phys Life Rev
2005;2:35373.
[4] Lamberti L, Pappalettere C. Metaheuristic design optimization of skeletal
structures: a review. Computational technology reviews, Vol. 4. UK: SaxeCoburg Publications; 2011. vol. 4, p. 132.
[5] Bensoe MP, Kikuchi N. Generating optimal topologies in structural design
using homogenization method. Comput Methods Appl Mech Eng
1988;71:197224.
[6] Xie YM, Steven GP. Evolutionary structural optimization. Springer; 1993.
78
K.-S. Yoo, S.-Y. Han / Computers and Structures 123 (2013) 6878
[7] Xie YM, Steven GP. A simple evolutionary procedure for structural
optimization. Comput Struct 1993;49:88596.
[8] Liang QQ. Performance-based optimization of structures: theory and
applications. London and New York: Spon Press; 2005.
[9] Liang QQ, Steven GP. A performance-based optimization method for topology
design of continuum structures with mean compliance constraints. Comput
Methods Appl Mech Eng 2002;191:147189.
[10] Belytschko T, Xiao SP, Parimi C. Topology optimization with implicit functions
and regularization. Int J Numer Methods Eng 2003;57:117796.
[11] Sethian J, Wiegmann A. Structural boundary design via level set and immersed
interface methods. J Comput Phys 2000;163:489528.
[12] Zhao H, Long K, Ma Z. Homogenization topology optimization method based
on continuous eld. Adv Mech Eng 2010:115561.
[13] Querin QM, Steven GP, Xie YM. Evolutionary structural optimization (ESO)
using a bidirectional algorithm. Eng Comput 1998;15:103148.
[14] Kane C, Schoenauer M. Topological optimum design using genetic algorithms.
Control Cybernet 1996;25:105988.
[15] Tagawa H, Ohsaki M. Simultaneous optimization of topology and geometry of
plane trusses with uniform cross-sectional area. J Struct Construct Eng AIJ
1999;521:7380.
[16] Luh GC, Lin CY. Structural topology optimization using ant colony optimization
algorithm. Appl Soft Comput 2009;9:134353.
[17] Kaveh A, Hassani B, Shojaee S, Tavakkoli SM. Structural topology optimization
using ant colony methodology. Eng Struct 2008;30:2259565.
[18] Tenek LH, Hagiwara I. Eigenfrequency maximization of plates by optimization
of topology using homogenization and mathematical programming. Jpn Soc
Mech Eng 1994;37:66777.
[19] Ma ZD, Cheng HC, Kikuchi N. Topological design for vibrating structures.
Comput Methods Appl Mech Eng 1995;121:25980.
[20] Kosaka I, Swan CC. A symmetry reduction method for continuum structural
topology optimization. Comput Struct 1999;47:4761.
[21] Pedersen NL. Maximization of eigenvalues using topology optimization. Struct.
Multidiscip. Optim. 2000;20:211.
[22] Du J, Olhoff N. Topological design of freely vibrating continuum structures for
maximum values of simple and multiple eigenfrequencies and frequency gaps.
Struct Multidiscip Optim 2007;34:91110.
[23] Huang X, Xie YM. Convergent and mesh-independent solutions for the bidirectional evolutionary structural optimization method. Finite Elem Anal Des
2007;43:103949.
[24] Huang X, Zuo ZH, Xie YM. Evolutionary topological optimization of vibrating
continuum structures for natural frequencies. Comput Struct 2010;88:35764.
[25] Huang X, Xie YM. Evolutionary topology optimization of continuum
structures. India: Wiley; 2009.
[26] Colorni A, Dorigo M, Maniezzo V, Distributed optimization by ant colony. In:
Proceedings of the rst European conference on articial life, USA 1991, p.
134142.
[27] Bullnheimer B, Hartl RF, Strauss CA, A New rank-based version of the ant
system: a computational study, Technical Report POM-03/97. Institute of
Management Science. University of Vienna, 1997.
[28] Shtovba SD. Ant algorithms: theory and applications. Program Comput Softw
2005;31:16778.
[29] Xie YM, Steven GP. Evolutionary structural optimization for dynamic
problems. Comput Struct 1996;58:106773.