Professional Documents
Culture Documents
www.elsevier.com/locate/compstruc
Department of Structural Mechanics, Faculty of Civil Engineering, Czech Technical University in Prague,
Th
akurova 7, 166 29 Prague 6, Czech Republic
Department of Mechanics, Klokner Institute, Czech Technical University in Prague, Solnova 7, 160 00 Prague 6, Czech Republic
Received 6 December 2001; accepted 14 February 2003
Abstract
This paper presents comparison of several stochastic optimization algorithms developed by authors in their previous
works for the solution of some problems arising in civil engineering. The introduced optimization methods are: the
integer augmented simulated annealing (IASA), the real-coded augmented simulated annealing (RASA) [Comp. Meth.
Appl. Mech. Eng. 190 (1314) (2000) 1629], the dierential evolution (DE) in its original fashion developed by Storn
and Price [R. Storn, On the usage of dierential evolution for function optimization, NAPHIS, 1996] and simplied
real-coded dierential genetic algorithm (simplied atavistic dierential evolution, SADE) [O. Hrstka, A. Kucerov
a,
Search for optimization methods on multi-dimensional real domains, Contributions to Mechanics of Materials and
Structures, CTU Reports 4, 2000, pp. 87104]. Each of these methods was developed for some specic optimization
problem; namely the Chebychev trial polynomial problem, the so called type 0 function and two engineering problemsthe reinforced concrete beam layout and the periodic unit cell problem, respectively. Detailed and extensive
numerical tests were performed to examine the stability and eciency of proposed algorithms. The results of our experiments suggest that the performance and robustness of RASA, IASA and SADE methods are comparable, while the
DE algorithm performs slightly worse. This fact together with a small number of internal parameters promotes the
SADE method as the most robust for practical use.
2003 Civil-Comp Ltd. and Elsevier Ltd. All rights reserved.
Keywords: Optimization; Evolutionary methods; Genetic algorithms; Dierential evolution; Engineering tasks
1. Introduction
Nowadays, optimization has become one of the most
discussed topics of engineering and applied research.
Advantages coming from using optimization tools in
engineering design are obvious. They allow to choose an
optimal layout of a certain structure or a structural
component from the huge space of possible solutions
based on a more realistic physical model, while the
traditional designing methods usually rely only on some
0045-7949/03/$ - see front matter 2003 Civil-Comp Ltd. and Elsevier Ltd. All rights reserved.
doi:10.1016/S0045-7949(03)00217-7
1980
n
X
ai xi ;
i0
1981
pf i ;
0
wi Ui =Ui;max 2
if Ui 6 Ui;max ;
otherwise;
75
62.5
50
37.5
25
12.5
0
200 150 100 50
50
100
150
200
1982
N
A X
Ik r;
2
N k1
3. Applied methods
During last few years, we have developed and tested
several evolutionary optimization methods that are
based on both binary/integer and real-valued representation of searched variables. Each of them was primarily
applied to one particular optimization problem of the
four introduced above. These methods are (in order of
appearance):
1983
F2 chbest j t chij t;
where chpj and chqj are the jth coordinates of two randomly chosen chromosomes and chbest j is the jth coordinate of the best chromosome in generation t. F1 and F2
are then coecients usually taken from interval (0,1).
Fig. 6 shows the geometrical meaning of this operator.
The method can be understood as a stand-alone evolutionary method or it can be taken as a special case of the
genetic algorithm. The algorithmic scheme is similar to
the genetic algorithms but it is much simpler:
1
The determination of K is inuenced by the parameter
called cross-rate (CR), see [15].
1984
10
where n is the number of variables of the tness function. Then, the simplied dierential operator can be
written as
13
11
where chpj , chqj and chrj are the jth coordinates of three
randomly chosen chromosomes and CR is the so called
cross-rate. Due to its simplicity this operator can be
rewritten also in the vector form:
CHi t 1 CHp t CRCHq t CHr t:
12
2
Contrary to the classical tournament strategy this approach can ensure that the best chromosome will not be lost
even if it was not chosen to any tournament.
q0
q
1 1 qpop
size
u0; 1 6 e
15
18
14
where q is the probability of selecting the best individual in the population, r is the rank of the ith
individual with respect to its tness, and pop_size
is the population size.
4. Apply operators to selected parent(s) to obtain possible replacement(s).
4a. Look for an individual identical to possible replacement(s) in the population. If such individual(s) exists, no replacement is performed.
4b. Replace old individual if
F Iold F Inew =Tt
1985
16
If actual temperature Tt1 is smaller than Tmin , perform reannealingi.e. perform step #1 for one half
of the population.
7. Steps 26 are repeated until the termination condition is attained.
List of operators. The following set of real-valued
operators, proposed in [13], was implemented. In the
sequel, we will denote L and U as vectors of lower/upper
pchjl t 1 pchil t;
chjl t;
if l k
otherwise:
21
Whole arithmetic cross-over: Simple arithmetic crossover applied to all variables of CHi and CHj .
Heuristic cross-over: Let p u0; 1, j 1; n and
k 1; n such that j 6 k and set:
CHi t 1 CHi t pCHj t CHk t:
22
1986
24
1
1 eF Iold F Inew =Tt
25
In connection with the notation and principles mentioned in previous sections, integer operators within
IASA algorithm have the following form:
Dierential cross-over: This operator is inspired by
the DE. A new individual CHj t is created from three
randomly selected solutions CHp t, CHq t and CHr t
according to
CHj t 1 CHp t u0:0; CRCHq t CHr t:
27
Note that all vectors CHi are integer numbers and also
that the inuence of the dierence on the right-hand side
randomly varies between zero and cross-rate CR.
Mutation: Mutation operator is provided by modifying each variable in CHj t to
jchij t chpj tj
chij t 1 chij t N 0;
1 ;
2
28
where N ; is a random integer number derived from
the Gauss normal distribution and chpj t is the jth
variable of a randomly selected vector CHp t.
1987
If the number of successful runs is non-zero, the average number of tness calls of all successful runs is
also presented.
Further details of individual function settings and
methodology for results evaluation can be found in the
next sections.
4.1. Results for the Chebychev problem
The computations were performed for the Chebychev
problem with a degree of the polynomial expression
n 8 (the T8 problem), which corresponds to the dimension of the problem 9. The computation was terminated if the algorithm reached a value of the objective
function smaller then 105 or the number of function
evaluations exceeded 100,000. Upper bounds on individual coecients were set to 512, while lower bounds
were equal to )512. The results of individual algorithms
are shown in Table 1 and Fig. 8.
4.2. Results for the type 0 trial function
Test computations for the type 0 problem were performed for a wide set of problem dimensions, ranging
from 1 to 200. The upper bound on each variable was set
to 400, while the lower bound value was )400. For each
run, the position of the extreme was randomly generated
within these bounds and the height of the peak y0 was
generated from the range 050. The parameter r0 was set
to 1. The computation was terminated when the value of
the objective function was found with a precision higher
than 103 . The results are given in the form of the
growth of computational complexity with respect to the
problem dimension. For each dimension, the computation was run 100 times and the average number of tness
calls was recorded (see Fig. 7 and Table 2).
4.3. Results for the reinforced concrete beam layout
problem
The basic parameters subjected to optimization were
the beam width b, which was assumed to take discrete
values between 0.15 and 0.45 m with the step 0.025 m
and the beam height h ranging from 0.15 to 0.85 m with
the step 0.025 m. For each of the three parts of a beam,
the diameter and the number of longitudinal reinforcing
Table 1
Results for the Chebychev polynomial problem
Method
IASA
RASA
DE
SADE
Successful runs
Average number
of tness calls
100
10,342
100
47,151
100
25,910
100
24,016
Table 2
Average number of tness calls for the type 0 function
Problem
dimension
IASA
RASA
DE
SADE
10
30
50
100
140
200
246,120
611,760
926,100
2,284,590
3,192,800
4,184,200
13,113
74,375
183,882
526,492
793,036
1,220,513
39,340
653,600
N/A
N/A
N/A
N/A
46,956
171,539
304,327
663,084
948,197
1,446,540
1988
Table 3
Results for the reinforced concrete beam layout
Table 5
Overall performance of methods
Method
IASA
RASA
DE
SADE
Method
IASA
RASA
DE
SADE
Successful runs
Average number
of tness calls
100
108,732
100
131,495
100
196,451
100
185,819
Chebychev problem
Type 0 test function
Concrete beam layout
Periodic unit cell
1
3
1
3
4
1
2
1
3
4a
4
4
2
2
3
2
14
Table 4
Results for the periodic unit cell problem
Method
IASA
RASA
DE
SADE
Successful runs
Average number
of tness calls
100
13,641
100
12,919
100
93,464
100
55,262
5. Conclusions
Dierential evolution. The DE algorithm showed to
be very ecient and robust for moderate-sized problems, but its performance for higher dimensions deteriorated. Moreover, the small number of parameters is
another advantage of this method. However, the results
suggest that the absence of mutation-type operator(s) is
a weak point of the algorithm.
Simplied atavistic dierential evolution. The SADE
algorithm was able to solve all problems of our test set
with a high reliability and speed. Although it needed
Acknowledgements
We would like to thank an anonymous referee for his
careful revision and comments that helped us to substantially improve the quality of the paper. The nancial
support for this work was provided by the Ministry of
Education, project nos. MSM 210000003 and MSM
R grant 103/97/K003.
210000015 and by GAC
1989
Table 8
Parameter settings for RASA
Parameter
Beam
Others
pop_size
q
p_uni_mut
p_bnd_mut
p_nun_mut
p_mnu_mut
p_smp_crs
p_sar_crs
p_war_crs
p_heu_crs
b
T_frac
T_frac_min
T_mult
num_success_max
num_counter_max
num_heu_max
precision (step 4a)
64
0.04
0.525
0.125
0.125
0.125
0.025
0.025
0.025
0.025
0.25
102
104
0.9
10 pop_size
50 pop_size
20
See Section 4.3
32
0.04
0.05
0.05
0.05
0.05
0.15
0.15
0.15
0.35
2.0
1010
1014
0.9
10 pop_size
50 pop_size
20
104
Table 9
Parameter settings for IASA
Parameter
Chebychev
Type 0
Beam
PUC
OldSize
NewSize
T_max
T_min
SuccessMax
CounterMax
TminAtCallsRate
CrossoverProb
CR
80
5
105
107
1000
5000
19%
97%
0.5
900
600
105
1010
1000
5000
100%
92%
0.6
180
250
104
105
1000
5000
25%
60%
1.3
200
100
101
105
1000
5000
20%
90%
1.0
Appendix A
See Tables 69.
References
Table 6
Parameter settings for DE
Parameter
Chebychev, type 0
Beam
PUC
pop_size
F1 F2
CR
10 dim
0.85
1
11 dim
0.85
0.1
10 dim
0.75
1
Table 7
Parameter settings for SADE
Parameter
Chebychev
Type 0
Beam
PUC
pop_size
CR
Radioactivity
MR
10 dim
0.44
0
0.5
25 dim
0.1
0.05
0.5
10 dim
0.3
0.05
0.5
10 dim
0.2
0.3
0.5
1990
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
http://www.fmmcenter.ncsu.edu/fac_sta/joines/
papers.
Michalewicz Z, Hinterding R, Michalewicz M. Evolutionary algorithms. In: Pedrycz W, editor. Chapter 2 in fuzzy
evolutionary computation. Kluwer Academic; 1997.
Michalewicz Z, Logan TD, Swaminathan S. Evolutionary
operators for continuous convex parameter spaces. In:
Sebald AV, Fogel LJ, editors. Proceedings of the 3rd
Annual Conference on Evolutionary Programming. River
Edge, NJ: World Scientic Publishing; 1994. p. 8497.
Ripley BD. Modelling of spatial patterns. J Roy Statist Soc
1977;39B(2):17292.
Storn R. On the usage of dierential evolution for function
optimization, NAPHIS, 1996.
ejnoha M. Numerical evaluation of eective
Zeman J, S
elastic properties of graphite ber tow impregnated by
polymer matrix. J Mech Phys Solids 2001;49(1):6990.