Sie sind auf Seite 1von 8

EngineeringAnalysiswithBoundaryElements19 (1997) 129-136

© 1997 Elsevier Science Ltd. All rights reserved


Printed in Great Britain
PII: S0955-7997(97)000 14-3 0955-7997/97/$17.00
ELSEVIER

Shape optimization of continuum structures by


genetic algorithm and boundary element method
Eisuke Kita* & Hisashi Tanie
Department of Mechano-lnformatics & Systems, Nagoya University, Furo-cho, Chikusa-ku, Nagoya 464-01, Japan

This paper presents a new scheme for the shape optimization of continuum
structures by using genetic algorithm (GA) and boundary element methods
(BEM). The profiles of the objects under consideration are represented by the
free-form deformation (FFD) technique. The chromosomes for the profiles are
defined by considering as the genes the position vectors of the FFD control
points. The population, which is constructed of many chromosomes, is modified
by applying genetic operations such as the selection, the crossover and the
mutation for determining a better profile for the design objectives. Besides, the
objective functions are estimated by the BEM. Three schemes are presented in
order to improve the computational efficiency of the present scheme. Finally, a
cantilever beam under uniformly distributed load is considered as a numerical
example.

Key words: Shape optimization, BEM, genetic algorithm, FFD.

1 INTRODUCTION the crossover, the mutation and so on. Application of


the GA to shape optimization problems has already
In the shape optimization problems of large-scale struc- been presented by some researchers. 12 21 However,
tures and complicated machine elements, the objects most of them deal with the layout optimization of the
under consideration are divided into some small parts, truss structures. 14-1s On the other hand, the shape opti-
which are optimized individually by applying gradient mization schemes of the continuum structures by the
search schemes using first-order derivatives of the objec- GA are classified into the cell representation
tive functions, i.e. the design sensitivities. 1-7 This scheme 12'13'18'2° and the free curves/surfaces representa-
scheme, however, may not search the best profile satisfy- tion scheme. 21'22 This paper presents the free curves
ing the design objectives. Since the problems are formu- representation schemes using the Free-Form Deforma-
lated as multi-objective optimization problems with tion (FFD) technique. By the way, the conventional
complicated constraint conditions, the gradient search schemes employ the finite element method (FEM) for
schemes may search the local minima alone near the estimating the fitness function values. Since, however,
starting point of the search. Moreover, it is very difficult the FEM meshes are distorted by the successive shape
to calculate the design sensitivities. For solving such modification, the computational accuracy may be
problems effectively, many researchers focus on the reduced. For overcoming the difficulty, we employ the
genetic algorithm (GA). 8-9 In this paper, the GA is boundary element method (BEM). Since the BEM can
applied to the optimization of the boundary profiles of solve the problem by boundary discretization alone,
the continuum structures. As a preliminary, we present the distortion of the profiles during successive shape
here the shape representation technique which can con- modification is not so terrible that the computational
trol the profiles of the objects easily and is suitable for accuracy becomes worse.
the operations of the GA. In our study, the boundary profiles of the objects
The genetic algorithm is known as the optimization under consideration are expressed by the free-form
method simulating the evolutional process of animals deformation (FFD) technique. In this technique, a grid
and plants. The search points on the design space are of the F F D control points is imposed on the parallel-
expressed as the chromosomes, which are optimized by piped so as to include the objects. The objects are dis-
applying the genetic operators such as the selection, torted by moving the control points from the lattice
position. The position vectors of the control points are
*To whom correspondence should be addressed. considered as the genes and then the chromosomes for
129
130 E. K i t a , H . T a n i e

the profiles of the objects are defined by the position vec- determined according to the magnitudes of the physical
tors. The genetic operators are applied to the population values such as stress and strain, instead of the binary
of the chromosomes in order to determine the best pro- parameters specified for the cells. If the physical value
file. The F F D technique can represent the whole profiles of a cell is greater than the reference value, the cell is
of the objects by a relatively small number of control thought to be full of the material and if not so, the
points. However, the computational cost of the present cell is a void. The reference value, which is represented
scheme is expensive. We shall present some techniques by the binary number, is optimized by the GA. Since,
for improving the computational efficiency. Finally, a in this case, the chromosomes are independent of the
cantilever beam under uniformly distributed load is con- number of cells, the computational cost in the detailed
sidered as a numerical example in order to confirm the design is not increased.
validity of the present scheme. By using the square cell representation, the genetic
representation of the continuum structures can be
derived easily. Moreover, it is very appropriate for esti-
2 S H A P E O P T I M I Z A T I O N BY G E N E T I C mating the fitness functions by the finite element
ALGORITHM method. However, it has some difficulties. First, the pro-
files of the optimized objects and the computational
Application of the GA to the shape optimization pro- accuracy of the fitness function values are strongly
blems has been presented by Sandgren e t a / . , 12'13 dependent on the size and the number of the cells.
Haftka et al., 14'15 Lin and Hajela, j6 Sakamoto and Next, although the free curves/surfaces representation
Oda, 17 Chapman et al., 19 Hasegawa and Kawamo, 2° of the objects is necessary in the design and manufactur-
Watabe and Okino 2~ and Takahashi et al. 22 In this sec- ing processes, the optimized objects have zigzag-shaped
tion, the present scheme is compared with the ordinary profiles and therefore the data of the zigzag-shaped pro-
schemes. files must be translated to the data of the free curves/sur-
faces representation. For overcoming these difficulties,
2.1 Layout optimization of truss structures schemes using free curves/surfaces representation are
necessary.
Application of the GA to the layout optimization of
truss structures have been studied by Haftka et al., 14'~5 2.3 Free curves/surfaces representation scheme
Lin and Hajela, 16 Sakamoto and Oda 17 and so on.
The nodes are placed in the design spaces including Watabe and Okino 21 and Takahashi et al. 23 represent
the structures. The existence of the members between the profiles of the objects by the free curves and then
two arbitrary nodes are indicated by the binary para- the chromosomes of the profiles are defined by the coor-
meters. The parameter is specified as 1 if there exists a dinates of the control points of the curves. The values of
member and 0 if not so. The binary parameters are the fitness functions are estimated by the finite element
taken as the genes and then the chromosomes for the method (FEM). In the FEM, the objects under consid-
layouts are defined by the series of the parameters. eration are discretized by domain discretization. When
Since, in this case, the genetic representation of the the FEM is applied to the shape optimization problem,
truss layouts can be derived easily, many researchers the elements are distorted by the shape modification and
have been studying such problems. therefore the computational accuracy may become
worse. This problem can be overcome by applying the
2.2 Cell representation scheme automatic mesh generator. However, the performance
of the mesh generator is not necessarily good enough,
Sandgren et al., 12'13 Chapman et al. 19 and Hasegawa and especially for complicated profiles, and moreover the
Kawamo 2° presented square cell representation of the computational cost is expensive.
continuum structures. The objects under consideration
are represented by uniform square cells and then the 2.4 Present method
binary parameter is specified to each cell. If a cell is
full of the material, the parameter is specified as 1 and The present scheme is similar to Watabe and Okino 21
if it is a void, the parameter is 0. This scheme is the and Takahashi e t al. 23 The profiles of the objects are
direct extension of that to the truss structures. However, represented by the free form deformation (FFD) tech-
when the detailed design of the profiles is performed, nique. A grid of the F F D control points is specified so
many tiny cells are necessary for representing the as to include the objects and then the objects are dis-
objects. In this case, the chromosomes become very torted by moving the control points from the grid posi-
long and therefore the computational cost become tion. The chromosomes for the profiles of the objects are
very expensive. defined by the series of position vectors of the control
In order to overcome the difficulty, Hasegawa and points. While Watabe and Okino 2~ and Takahashi
Kawamo 2° present that the existence of the cells is e t a [ . 23 use the FEM for estimating the fitness function
Shape optimization of continuum structures by genetic algorithm and boundary element method 131

values, we employ the boundary element method controle points lie on the lattice position. The position
(BEM). When the objects under consideration are gov- vector of the arbitrary control point Pijk is given as:
erned by linear and homogeneous differential equations,
i JT ku
the BEM can solve the problem by boundary discretiza- P,jk=Xo+~S+m +n (5)
tion alone and therefore the distortion of the elements
by the successive shape modifications is smaller than where i, j and k are integer number taken as 0 < i < l,
with FEM. Besides, if the elements are distorted terribly 0 < j < m , 0 < k < n . These f o r m l + l planes in the S
by the large modifications, mesh generation of the BEM direction, m + 1 planes in the T direction and m + 1
is much easier and cheaper than the FEM. planes in the U direction, respectively.
The profiles of the objects under consideration are
deformed by moving the control points from the lattice
3 PRESENT S C H E M E S position. Assuming that X will move to XFVD, XFFD is
given as:
3.1 Shape representation by FFD
XFFD = Z Z Bi,t(s)Bj,m(t)Bk,.(u)Pijk (6)
The profiles of the objects under consideration are repre- i=0 j = 0 k = 0

sented by the free form deformation (FFD) technique,24 where


which is often employed in the field of computer gra-
phics. We shall explain the FFD technique for three- Bi,~(t) =n Ci(1 - t)n-it i (7)
dimensional objects. and nCi is the combination:
First, we impose the local coordinate system on the
parallelepiped region (Fig. 1). The positive vector of n!
.C, - - - (8)
an arbitrary point X is represented as: i!(n - i)!
X = Xo + sS + tT + uU (1) The FFD technique can represent the profiles of the
objects under consideration by a relatively small
where X0 denotes the position vector of the origin and S,
number of control points.
T and U the unit vectors in the S, T and U directions,
respectively. Here, s, t and u, which indicate the coordi- 3.2 Chromosome and genetic operations
nates of the position X in the local coordinate systems,
are calculated as follows: 3.2.1 Chromosome
T × U. ( X - Xo) The position vector of the control points Pijk are consid-
s= T × U-S (2) ered as the genes and then the chromosomes for the pro-
files of the objects under consideration are defined as:
S × U . ( x - Xo)
t= S × U-T (3) P000P001P010P100 • • • Ptmn (9)
S × T. ( X - Xo) The length of the chromosomes is ( l + 1) x (m + 1)
u= S × T-U (4) ×(n + 1), which is invariant during the optimization
process.
Next we impose a grid of control points on the paral-
The individual has the chromosome defined as in eqn
lelepiped. When the objects are not deformed, the
(9). The population, which is constructed of many indi-
viduals, is evolved by applying genetic operators for
finding better individuals, i.e. better profile for the
U design objectives.

3.2.2 Selection
The selection operation selects the parents from the
population according to the magnitudes of the fitness
function values of the individuals. The roulette wheel
selection is known to be the most popular. In this
.v_ T
case, the parents are selected in proportion to the mag-
nitudes of their fitness function values. However, when
the differences between the fitness function values of
the individuals are very large, the roulette wheel selec-
tion tends to select the best individual alone and there-
fore the local minima near the starting points are
searched. In order to overcome the difficulty, we
Fig. 1. stu-coordinates. employ the ranking procedure.
132 E. Kita, H. Tanie

The ranking procedure selects the parents according value of the movement is specified by the user in
to the ranks of the function values, instead of the mag- advance.
nitudes. The ranks of the individuals are specified While the usual mutation moves the control points
according to the magnitudes of their fitness values. randomly, the directed mutation moves them to the
While the highest selection rate is specified for the indi- function space where a better solution may exist. If the
vidual of the highest rank, the lowest rate is to the one of mutation improves the solution, the directions of the
the lowest rank so as to produce at least one offspring. movements of the genes are memorized. At the next gen-
The rates of the other individuals are determined by eration, the genes are moved slightly in the memorized
linear interpolation of them. In addition to the ranking directions and then randomly by the usual mutation.
procedure, the elitist model is employed in order to keep
the best individual at the next generation.
3.3 Boundary element method

3.2.3 Crossover We shall explain the BEM analysis in the two-


The crossover operation swaps some genes of the dimensional elastic problem. 25'26
selected parents in order to create the offsprings. The The governing equation without the body force and
one-point, two-point and uniform crossovers are com- the boundary conditions are given as:
pared in the numerical examples.
a0 J : 0 (in f~) (t0)
In the one-point crossover, the chromosomes of the
parents are divided at one point and then, the parts and
are swapped. For example, we consider the following
parents: Hi ~ ~i (on )
(11)
ti = li (on Ft)
PoooPool PoloPol 1PlooPIol P110P111
I / I I I / I where ui, ti and 0 0. denote the displacement, the traction
PoooPomPolol~ol 1PlooPtol Pl toP111
and the stress components in the two-dimensional coor-
Assuming that the chromosomes are divided between dinates, respectively, and ( j ) the derivative in t h e j direc-
the third and fourth genes and swapped, we have the fol- tion. Taking the Kelvin solutions as the weight
lowing offspring: functions, we have the weighted residual equation;

PoooPoo1PoloP011P'looP'lOlP'l loPtl 11
t t !
PoooPore Pm oPm l PlooPlol P1 loPl I t
J ' oii /U*kiekd[~ = 0 (12)

where ek denotes the base vectors. Applying the G a u s s -


In the two-point crossover, the chromosomes of the
Green formula, we have the boundary integral equation;
parents are divided at two points and swapped. Also,
in the uniform crossover, the genes of the parents are
swapped randomly.
Cui = (u /tj - t ju )dr (13)
r

3.2.4 Mutation where C is the parameter depending on the placement of


We shall apply, in addition to the usual mutation, the the source point; C = 1 when the point is inside the
directed mutation (Fig. 2). domain, C = 0 when it is outside the domain and
The usual mutation operation moves randomly arbi- C = 1/2 when it is on the smooth boundary; also u O.
trary genes of the chromosomes, i.e. the position vectors and t~ denote the fundamental solutions of the displace-
of the arbitrary control points. If, in this case, the mag- ments and the tractions, respectively. Discretizing eqn
nitude of the movement is large, the mutation destroys (13) with the boundary elements and collocating the
the profiles of the objects. Therefore, the maximum nodes on the elements, we have
Hu -- Gt (14)

where u and t denote the vectors of the nodal displace-


ments and the tractions, respectively and H and G are
their coefficient matrices.
Applying the boundary conditions to the above
system of equations, we have

(15)
(a) Nounal mutation (b) Directed mutation
where the subscripts u and t denote I', and l~t,
Fig. 2. Mutations. respectively. Assembling the system of equations, we
Shape optimization of continuum structures by genetic algorithm and boundary element method 133

have the profiles (Fig. 3). Left and upper segments of the pro-
files are fixed and therefore the shape modification is
[-Gunt]{~%): [-HuGt](~:l (16) performed by moving six control points which are not
taken on the segments.
and In the following numerical examples, each problem is
simulated five times from the differential initial popula-
Ax=b (17)
tions and then each result is shown as the average of the
The system of equations is solved for x to estimate the five results.
unknown nodal values.
4.2 Basic scheme
4 NUMERICAL EXAMPLE
By introducing the penalty function a, the optimization
4.1 Object under consideration problem with the constraint conditions is formulated as
follows:
A cantilever beam under a uniformly distributed load is
fitness = 1 - + a[H(amax, ac) + H(cross, 0)]
considered as a numerical example (Fig. 3). The design
objective is to minimize the area of the beam A by chan- (21)
ging the profile of the beam. As the constraint condi-
tions, we consider that the maximum stress O'max is where H is the step function and a is taken as 10000.
smaller than the allowable stress of the material a c The basic scheme, which is referred to as scheme 1,
and that the profiles of the objects do not cross. The employs the above fitness function and the process
objective function f and the constraint conditions ht shown in Fig. 4. Besides the roulette wheel selection,
and h2 are expressed as: the one-point crossover and the elitist model are
adopted. The total number of individuals of the popula-
A
m i n f = A-o (18) tion is 50. The crossover rate and the mutation rate are
specified as l and 0.02, respectively. These values are
subject to invariant during the computation. Figure 5 indicates
the object function values of the best individual at
hi = O'max - - O'c _~ 0 (19)
each generation. The performance of the best individual
h2 = cross <_ 0 (20) is improved during the process.

where A0 denotes the area of the initial profile; cross is


the parameter related to the crossing of the profile; 4.3 Improvement of population generation process
cross = 1 if the profile cross and cross = 0 if not.
Here, 12 control points are employed for controlling In scheme 1, the constraint conditions are not confirmed
after the offspring are generated. Therefore, the new
population includes the individuals which do not satisfy

START ]

I Birth of initialpopulation ]

I BEM analysis~ Calculationof fitness ft,nction]

I Selection ]

I Crossover I
[ Mu, tion 1
,......I.. ..g°.~.l m~.=~.oe

I Evaluati°n I

l E.D ]
Fig. 3. Object under consideration. Fig. 4. Scheme I.
134 E. Kita, H. Tanie

1.0 1.o - - 5000


i leestindividual: I i .."."
I . . . . . . . . . Scheme 11 !'","
0.9 0.9 - I Scheme 2 l ; , • ~.
- - 4000 z
IBEM analysis: I '
'\
• [ --- Scheme 1 1 " i 12r

0.8 0.8 -
~-.i I .... i Scheme 21 i 3000
O

,.. i .,',, i i u0
< •. . . . . < m
~r
: "'---,. ,' ,, i !
0.7 "7:",-.%.:.. i , 0.7- ! ;v~.:;.: i ! ! 2000
•-..........]......... '/i '"""" i i
"/ ! i ...... !
! ...........i.................. ",, i : ""!'. . . . .

0.6 :
i
i 0.6 - t000
,J" ,/

d
d !
i
0.5 i i i 0.5 - I/ I - 0
20 40 60 80 100 20 40 60 80 100
Generation Generation

Fig. 5. Results from scheme 1. Fig. 7. Results from scheme 2.

the constraint conditions. In order to construct the new respectively. Figure 7 indicates the object function
generation with the individuals satisfying the constraint values of the best individuals. Scheme 2 produces a
conditions alone, we introduce scheme 2 shown in Fig. 6. better individual than scheme 1. In scheme 2, the
When offspring are generated, the constraint conditions number of the BEM analyses is larger than scheme 1.
are confirmed. If an offspring does not satisfy the condi- This is because the BEM analyses are repeated until
tions, it is eliminated. The new population is constructed the population is filled with the individuals satisfying
by the individuals satisfying the constraint conditions the conditions alone.
alone. Since, in this case, all individuals satisfy the con-
ditions, we apply the fitness function not including the
conditions: 4.4 Improvement of crossover

A Both schemes 1 and 2 employ the one-point crossover.


fitness = 1 - - (22)
A0 We introduce the other crossover operations for improv-
ing scheme 2. The number of individuals is 50, the cross-
The number of individuals of the population is 50, the over and the mutation rates are 1 and 0.02, respectively.
crossover and the mutation rates are 1 and 0-02, Figure 8 indicates the object function values of the best
individuals at each generation. We notice that the uni-
form crossover can search the individual better than
START ] the other crossover operations.

[ Birth of initial population [


1.0 - - 5000
i 4 ";y
[ C'alculation of fitness function I
~-_ 0.9 - t Besl individual:
4000 Z
[ Selection [ ;:~, --- One point
",~ ......... Two p o i n t
0.8 i\ . . . . . Uniform 3000
I I ", ~
'~ ! /'
BEM analysis:
. . . . . One point
< m
'(~ ,i / - - - Two p o i n t
0.7 2000
"i',~ ..... s' i .... Uniform

F 0.6 100o ~
Z
•/
I Evaluati°n I
¢ 0.5 "/" - 0

20 40 60 80 100
Generation
Fig. 6. Scheme 2. Fig. 8. C o m p a r i s o n o f crossovers.
Shape optimization of continuum structures by genetic algorithm and boundary element method 135

4.5 Employment of directed mutation 1.0

The scheme using the algorithm shown in Fig. 6 and the


0.9
uniform crossover is referred to as scheme 3. We shall
improve scheme 3 by introducing the directed mutation.
The number of individuals is 50, the crossover and the 0.8
mutation rates are 1 and 0.02, respectively. Figure 9
<
indicates the objective function values of the best indivi- 0.7 ....

duals at each generation. We notice that the scheme


using directed mutation can search the individual
better than the scheme using the ordinary mutation. 0.6
Scheme 4 is referred to as scheme 3 using the directed
i : i
mutation, instead of the ordinary mutation. 0.5- : ~ =:
l I I
0 20 40 60 80 10l
4.6 Comparison of basic and final schemes Generation

Fig. 10. Comparison of schemes 1 and 4.


A comparison of the basic scheme (scheme l) and the
final scheme (scheme 4) is shown in Fig. 10. We notice
that the performance of scheme 4 is improved mainly. easily and, moreover, sensitivity analysis is not
The profile of the best individuals at each generation necessary.
are shown in Fig. l l. • The F F D technique can represent the boundary
profiles of the objects by a relatively small
number of genes.
• Since the BEM can solve the problem by boundary
5 CONCLUSION
discretization alone, the distortion of the mesh is
not terrible and therefore the computational accu-
This paper presented the application of the G A to the
racy is better than the FEM.
shape optimization of the continuum structures. The
profiles of the objects under consideration are repre-
sented by the F F D technique. The position vectors of
the F F D control points are considered as the genes
and then the chromosomes for the profiles are defined
by the series of genes. The population is constructed J
by many individuals of the chromosomes. Also, the fit-

1
ness functions of the individuals are estimated by the
BEM. The advantages of the present scheme can be
summarized as follows:
• The G A can deal with the multi-objective problems Generation=l Generation= 5

1.0 - -
/' • - 5000

0.9

0,8 \
. . . . . . . . .

!
i
i
il
i
l Ordinary I
I ......... DirectedJ .....
fl
I
/.

L. . . . . . . . . . ~ . . . i . . . . . . . . . . . . . .
[Best individual: [
4000

3ooo~
/
<= ~. i :: , BEM analysis: I o
< '%~ i ] ,/" [ ..... Ordinary I
/ .
--
0.7 k.J..~ ,"~ I, --- D,rected[, 2000~ /
'~.... / ! ! i
i/' ".., : i ! O~
0.6 ,./:~ • "....... :: ' .... 1000~'
/ ~ .......'-.i
/ ! i
f/
0.5
J J i o
0 20 40 60 80 100
/~ Generation= 10O
Generation= 10
Generation
Fig. 9. Comparson of mutations. Fig. 11. Results from scheme 4.
136 E. Kita, H. Tanie

The present scheme was applied to the shape optimiza- 12. Sandgren, E., Jensen, E. & Welton, J. Topological design
tion of the cantilever beam. The results were very satis- of structural components using genetic optimization
methods. In Sensitivity Analysis and Optimization with
factory. However, the computational cost is relatively
Numerical Methods. Proc. Winter Annual Meeting, ASME,
high. For overcoming the difficulty, we present the fol- Dallas, TX, 1990, pp. 31-43.
lowing techniques: 13. Sandgren, E. & Jensen, E. Automotive structural design
employing a genetic optimization algorithm. In SAE
• The algorithm is improved so the population is con- Technical Paper No. 920772. Proc. SAE Int. Cong. Exp.,
structed by the individuals satisfying the constraint Detroit. MI, 1992, 1992.
condition. 14. Nagendra, S., Haftka, R. T. & Gurdal, Z. Design of a
• Uniform crossover can search individuals better blade stiffened composite panel by a genetic algorithm. In
Proc 34th AIAA/ASME/ASCE/AHS/ASC Structures,
than the one-point and the two-point crossover
Structural Dynamics and Materials Conference, La Jolla,
operations. CA, 1993, pp. 2418-36.
• The directed mutation can search individuals better 15. Ponslet, E., Haftka, R. T. & Cudney H. H. Optimal
than the ordinary mutation. placement of tuning masses on truss structures by genetic
algorithm. In Proc 34th AIAA/ASME/ASCE/AHS/ASC
By adopting the above schemes, the computational effi- Structures, Structural Dynamics and Materials Conference,
ciency of the scheme was improved fairly. However, we La Jolla, CA, 1993, pp. 2448-57.
have to improve it still more. For this purpose, we are 16. Lin, C.-Y. & Hajela, P. Genetic search strategies in large
studying a new algorithm for accelerating the conver- scale optimization. In Proc 34th AIAA/ASME/ASCE/
AHS/ASC Structures, Structural Dynamics and Materials
gency property. Conference, Dallas, TX, 1992, pp. 2437-47.
17. Sakamoto, J. & Oda, J. A technique of optimal layout
design for truss structures using genetic algorithm. In Proc
REFERENCES 34th AIAA/ASME/ASCE/AHS/ASC Structures, Struc-
tural Dynamics and Materials Conference, 1993, pp.
1. Arora, J. S. & Cardoso, J. B. A design sensitivity analysis 2402-8.
principle and its implementation into ADINA. Comp. 18. Ohsaki, M. Genetic algorithm for topology optimization
Struct., 1989, 32(3/4), 691-705. of trusses, Comp. Struct., 1995, 57(2), 219-25.
2. Barone, M. R. & Yang, R. J. A boundary element 19. Chapman, C. D., Saitou, K. & Jakiela, M. J. Genetic
approach for recovery of shape sensitivities in three- algorithms as an approach to configuration and topology
dimensional elastic solids. Comp. Meth. Appl. Mech. design. J. Mech. Des., 1994, 116, 1005-12.
Engng, 1989, 74, 69-82. 20. Hasegawa, H. & Kawamo, K. A method for the shape and
3. Burczynski, T., Kane, J. H. & Balakrishna, C. Shape topology optimization of mechanical structures by using
design sensitivity analysis via material derivative-adjoint genetic algorithm. Trans. JSME, 1995, A61(581), 183-90
variable technique for 3-D and 2-D curved boundary (in Japanese).
elements. Int. J. Num. Meth. Engng, 1995, 38, 2839-66. 21. Watabe, H. & Okino, N. A study on genetic shape design.
4. Haftka, R. T. & Grandhi, R. V. Structural shape In Proc. 5th Int. Conf. Genetic Algorithm, ed. S. Forrest.
optimization--a survey. Comp. Meth. Appl. Mech. Morgan Kaufmann, San Mateo, CA, 1993, pp. 445 50.
Engng, 1986, 57, 91 106. 22. Furukawa, T. & Yagawa, G. Parameter identification
5. Haug, E. J., Choi, K. K. & Komkov, V. Design Sensitivity using an evolutionary algorithm and its performace under
Analysis of Structural Systems. Academic Press, New measurement errors. In Computational Mechanics'95 Proc.
York, 1986. ICES95, Hawaii, 1988, ed. S. N. Atluri, G. Yagawa, & T.
6. Sandgren, E. & Wu, S.-J. Shape optimization using the A. Cruse. Springer Verlag, Berlin, 1995, pp. 122-7.
boundary element method with substructuring. Int. J. 23. Takahashi, K., Takahashi, Y. & Watanabe, K. 2D shape
Num. Meth. Engng, 1988, 26, 1913 24. optimization by genetic algorithms. In Proceeding of 50th
7. Zhao, Z. Shape Design Sensitivity Analysis and Optimiza- Conference of Information Processing Society of Japan,
tion Using the Boundary Element Method. Lecture Notes in 1995, pp. 259-60 (in Japanese).
Engineering 62. Springer Verlag, Berlin, 1991. 24. Sederberg, T. W. & Parry, S. R. Free-form deformation of
8. Goldberg, D. E. Genetic Algorithms in Search, Optimiza- solid geometric models. In Computer Graphics (Proc.
tion and Machine Learning. Addison Wesley, Reading, SIGGRAPH'86), ed. D. C. Evens & R. J. Athay. ACM
MA, 1989. SIGGRAPH, 1986, pp. 151-60.
9. Davis, L. Handbook of Genetic Algorithms. Van Nostrand 25. Banerjee, P. K. & Butterfield, R. Boundary Element
Reinhold, New York, 1991. Method in Engineering Science. McGraw-Hill, New York,
10. Schwefel, H.-P. Numerical Optimization of Computer 1981.
Models. John Wiley & Sons, New York, 1981. 26. Brebbia, C. A. The Boundary Element Method Jor
11. Fogel, D. B. & Atmar, J. W. Proc. 1st Annual Conference Engineers. Pentech Press, London, 1978.
on Evolutionary Programming. Evolutionary Programming
Society, 1992.

Das könnte Ihnen auch gefallen