Sei sulla pagina 1di 22

Finite Elements in Analysis and Design 37 (2001) 381}402

Optimising frame structures by di!erent strategies


of genetic algorithms
D. Greiner*, G. Winter, J.M. Emperador
Centro de Aplicaciones Nume& ricas en Ingeniern& a (CEANI), Campus de Taxra Baja, 35017,
Universidad de Las Palmas de Gran Canaria, Spain

Abstract

Two frame structure optimisation problems are handled. Firstly, the constrained minimum-mass problem,
for which both continuous and discrete cases with ideal (without buckling e!ect and its own gravitational
load) and real (with buckling e!ect and its own gravitational load) models are solved with several types of
genetic algorithms. Results for the continuous case (ideal model) are compared with referenced solutions of
traditional optimisation methods. From state of the art, the chosen strategies include a generational genetic
algorithm, a steady-state genetic algorithm and the CHC algorithm. Moreover, we suggest one elitist strategy
linked to the non-dominated sorting genetic algorithm (NSGA) to solve the multiobjective problem of
minimising simultaneously the number of di!erent cross-section types and the mass with several con-
straints.  2001 Elsevier Science B.V. All rights reserved.

Keywords: Continuous and discrete frame optimisation; Multiobjective optimisation; Elitist NSGA; CHC algorithm;
Buckling e!ect

1. Preliminary considerations

It is well known that the optimisation problem of determining the minimum-mass frame is
a problem that has many local minima, so a global optimisation method is necessary to be applied.
Genetic algorithms are iterative adaptive global search algorithms, that di!er from more
standard search algorithms (e.g. gradient methods, controlled random search, hill-climbing,
simulated annealing, etc.) in that the search is conducted using the information of a population of
structures instead of a direction of a single structure.

* Corresponding author.
E-mail addresses: davidg@baal.ulpgc.es (D. Greiner), gwinter@cicei.ulpgc.es (G. Winter), emperador@cicei.ulpgc.es
(J.M. Emperador).

0168-874X/01/$ - see front matter  2001 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 8 - 8 7 4 X ( 0 0 ) 0 0 0 5 4 - 8
382 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

The way operators are de"ned, the parameter settings they take, and the representation of the
search space all contribute to the likely success of the algorithm. Many factors have in#uence on
the convergence rate, such as the population size, the parent selection scheme, crossover operator
applying, mutation rate, forbidding of replicates, scaling procedures, etc [17].
What is less well documented is the role the GA operators play in de"ning the nature of
a particular search space. Moreover, if by coincidence the structure imposed by the operators
happens to align with some underlying features in the "tness landscape, then the algorithm
performs well. Conversely, a mismatch between operators and "tness landscape can produce poor
results. This fact has been shown to be true for all search methods: NFL Theorem. This result goes
as far as to show that all search techniques are equivalent in terms of "nding a global optimum
when viewed across the entire space of possible "tness functions. Once more this result serves to
emphasise the importance of "nding a search algorithm that "ts a given search problem if good
results are to be achieved.
It is important to understand how each of these choices a!ects the genetic algorithm search, and
how each introduces some form of bias into the search trajectory favoured by a given algorithm.
Testing the search problem from a number of views, will it be possible to gauge the relative success
of results?
Anyway some general considerations have been commonly accepted by genetic community from
theoretical studies and from solving many problems in di!erent "elds. To continue, we highlight
some of them to select and propose e$cient strategies.
Firstly, about the encoding, if a good representation is chosen for the problem to be solved then
the algorithm e$ciency in terms of the search is improved dramatically. The most commonly used
representations among evolutionary algorithms that use bit representations are standard binary
and standard binary re#ected Gray encodings. A &Gray code' is any bit representation that has the
property that adjacent integers are also adjacent neighbours in the bit-neighbourhood hyper-cube
graph de"ned over the set of bit strings of length ¸, removing Hamming cli!s. Empirically, Gray
coding usually results in better performance than using a standard binary coding when applying
various forms of genetic algorithms and neighbourhood search algorithms using common test
functions [12].
About selection, elitist strategies are based on copying the best individual of each generation into
the next one, where some/all parents are allowed to reproduce; thus, reducing the possibility that
the best individual of the population may fail to produce o!spring in the next generation.
GENITOR [15] was the "rst of the steady-state genetic algorithms, using a (, 2) selection
strategy in terms of evolution strategies instead of (, ) of the canonical genetic algorithm.
O!spring replace the worse individuals in the population and "tness is assigned to rank, helping to
maintain a more constant selective pressure over the course of the search.
As a small population can quickly be dominated by one member, CHC [14] stands for
cross-generational elitist selection having very aggressive search, heterogeneous recombination
and cataclysmic mutation. In the context of using small population, CHC selection uses random
selection, but restrictions are imposed on which strings are allowed to mate. Strings with binary
encodings must be a certain Hamming distance away from one another before they are allowed to
reproduce. This form of &incest prevention' is designed to promote diversity. CHC explicitly
borrows from (#) strategy of evolution strategies: after recombination the  best unique
individuals are drawn from the parent population and o!spring population to create the next
D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 383

generation. CHC is used for small populations. Eshelman uses HUX recombination (as uniform
crossover with exactly half of the di!erent bits being swapped during crossover). At the point where
the population converges reproducing similar strings, CHC uses cataclysmic mutation, except that
the best string is preserved intact.
Uniform crossover is unbiased with respect to de"ning length; it is more disruptive than
one-point crossover. Their use with respect to one-point or two-point crossover is for small
population helping to overcome the limited information capacity of smaller populations and the
tendency for more homogeneity.
Finally, the roulette wheel selection (RWS) scheme is a noisy one, existing other schemes with
lower variance, but with the same or approximately expectation, such as stochastic remainder
selection with replacement or stochastic universal selection [16].

2. Proposed strategies of genetic algorithms

We propose the following three strategies to establish a comparative study of performance


according to the speci"c characteristic of the structural problem considered here. As a case of
single-objective algorithm, it is the optimisation problem of minimum-mass of frame structures. We
use a small-intermediate population size of 50 individuals and a mutation rate of 3% as global
parameters. An MWC [5] random generator of 32 bits is used. The "tness function is scaled
through ranking, in order to avoid premature domination of a possible superindividual and
premature stagnation and also to bene"t the evolution in the last generations. A relation between
the best and mean "tness of the population of two is done, according to Goldberg's proposal
[7,18,2,3].

2.1. A generational genetic algorithm

As the "rst developed, in the generational strategy, the parent population as a whole, is
substituted by their children. With elitism, the two best individuals of each generation are
conserved and inserted directly in the next generation. A one-point crossover is used, where the
cross-position in the bit chromosome is selected through a uniform random variable. Selection is
done by the roulette wheel selection according to the selection probability of the individuals. The
crossover probability is 0.8.

2.2. A steady-state genetic algorithm

In a steady-state genetic algorithm, only a few individuals are inserted in the population
substituting some of the parents } usually one or two } in each generation (two in this case
substituting the worst individuals). Duplicates are not allowed in the population, so they are
avoided in the initial population and also during the insertion in the evolving population across the
execution of the algorithm. A uniform crossover is used, where each bit position in the chromosome
has a uniform random probability to be crossed. Selection is done by the stochastic universal
selection, according to the selection probability of the individuals. The crossover probability is 1.0.
384 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

2.3. The CHC algorithm

In the CHC algorithm, in each generation a children population is generated through a random
selection of the parents to be crossed and without mutation. Afterwards, truncation selection is
applied (it can be viewed as a (#) selection of evolution strategies). The crossover includes
dynamic incest prevention, where it is allowed only if more than 25% of the bits of the parents are
di!erent, and can be diminished if the evolution is not satisfactory. A HUX crossover is used, where
exactly 50% of the reduced surrogated bits are crossed. A massive mutation is applied when the
population does not evolve, duplicating the best chromosome and 37% of the bits randomly
#ipped (even completely random if it does not evolve).

3. Numerical results for the single-objective optimisation problem

The "tness function is evaluated through the direct sti!ness method and the following con-
straints are applied (with 6% of structure mass penalty for each 10% violation, as Goldberg
recommends [7,18]): stresses of the bars, displacements of joints or middle points of bars, and
a compressive slenderness limit of 200 to include the buckling e!ect. The evaluation of the  factor,
is based on the Julian and Lawrence criteria applied in the EA-95 Spanish code.
To compare the performance of genetic algorithms with other classical methods of optimisation,
"rst we consider a continuous approximation of the problem, encoding the area of the bars and
extracting the rest of the problem variables through a least-square evaluation that "ts the following
relations of the IPE cross-section type de"ned in the EA-95 Spanish code [1]. (Interval encoding
between 7.64;10\ and 155;10\ m).
A continuous approximation of the relations for the IPE cross-section types among the area (A)
and moment of inertia (I), modulus of section (=), web area (A ) and the relation of beam height

(h1/h) has been considered using a least-squares approximation. The optimal are the exponential
ones except in the case of h1/h which is a logarithmic one, with the following correlations: 0.99986,
0.99987, 0.99887, respectively for I, =, A and h1/h: 0.95382 (versus the exponential approximation

correlation: 0.95161).
The obtained relations are the following
I"0.78824A , ="0.72515A , A "0.24454A ,

h1/h"0.68749#0.03353 ln(A), versus the exponential one:0.69509A .
The relations founded in the referenced optimisation problem [4] are similar, but the web area is of
the cross-section type that di!ers from a suitable one. Our newly obtained relations are generally
more unfavourable than the following referenced in [4]: I"0.7918A ; ="0.727A ;
A "0.0297A ; h1/h"0.69541A . The used coe$cients are "nally the last printed ones,

except for the web area, because it di!ers clearly from the own calculus. The referenced problem [4]
considered here is shown in Fig. 1.
In the "gure the distances are expressed in meters and there exists a constraint of maximal
displacement in the middle point of bar 2 of length/300"6.6 ;10\ m. Its referenced solution [4],
obtained after trying multiple starting points and considered as the optimal one is: by means of the
D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 385

Fig. 1. Test case.

Table 1

Bar Referenced FD Own FD Referenced YC Own YC

1 192.08 189.43 227.36 227.16


2 105.84 106.82 70.56 105.64
3 234.22 233.24 235.2 235.49
4 195.02 195.41 213.64 219.52

feasible direction method (FD), mass of 3217 kg. improved with the yield criteria (YC) to a mass of
3213 kg.
Where the comparing stresses have taken into account the axial and the shearing stresses by the
shear e!ort, supposed to be uniformly distributed along the cross-section type and also the bending
e!ort.
The referenced solution has the following areas and comparing stresses for each one of the bars,
showing them with the comparing stresses of our direct sti!ness evaluation method for 200
intermediate sections per bar (yield stress"235.2 MPa) (See Table 1).
The slight di!erences seen in the table could be caused by the least-squares approximation
coe$cients or by our stress hypothesis, where two critical points per section have been taken into
account: the intersection of the web and #ange planes; top face of the #ange and bottom face of the
#ange.
The values of medium span displacement obtained by our direct sti!ness method are:
6.6134;10\ m for FD and 6.7072;10\ m for YC (remember that the maximum limit of
l/300"6.6 ;10\ m.)
The best solution achieved, without considering buckling e!ect and without its own gravi-
tational load (called ideal model) and respecting all the imposed constraints by the genetic
algorithm optimisation is a mass of 3118.27 kg (3% below the previous solutions). The detailed
sections of each solution are described in Table 2.
The Figs. 2}6, show the statistic results over 90 di!erent runs.
The buckling e!ect has been considered taking into account the Spanish code EA-95 and based
on the calculations of Julian and Lawrence. Considering the frame sidesway permitted with
386 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

Table 2

Bar FD area YC area GA area Comp. stress Comp. stress Comp. stress
(10\ m) (10\ m) (10\ m) FD (MPa) YC (MPa) GA (MPa)

1 90.3 75.5 72.986 189.43 227.16 235.2


2 129.6 142.6 125.652 106.82 105.64 132.3
3 9.8 10.0 10.655 233.24 235.49 213.44
4 90.7 71.0 110.917 195.41 219.52 171.5
Mass (kg) 3217 (1.032 of *1) 3213 (1.030 of *1) 3118.3 (*1)
Control displace- 6.6134 6.7072 6.66666
ment (10\ m)

Fig. 2. Continuous optimisation, ideal model, mean of 90 cases (1).

buckling e!ect in all the bars, beams and columns, and considering its own gravitational load
(called real model) the reached solution slightly di!ers with respect to the ideal model, as shown in
Table 3.

3.1. Results for real model

Figs. 7}11 consider buckling e!ect and its own gravitational load (real model) in a sidesway
permitted frame and for all its bars.
D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 387

Fig. 3. Continuous optimisation, ideal model, mean of 90 cases (2).

Fig. 4. Continuous optimisation, ideal model, best of 90 cases (1).


388 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

Fig. 5. Continuous optimisation, ideal model, best of 90 cases (2).

Fig. 6. Continuous optimisation, ideal model, variance of 90 cases.


D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 389

Table 3

Bar Area ideal model Area real model Stress ideal model Stress real model
(10\ m) (10\ m) (MPa) (MPa)

1 72.986 79.45 235.20 235.10


2 125.652 127.87 135.34 148.67
3 10.655 19.315 220.40 233.14
4 110.917 118.15 168.27 171.40
Mass (kg) 3118.27 (0.951 of *2) 3278.7 (*2)
Control displacement 6.6666 6.6666
(10\ m)

Fig. 7. Continuous optimisation, real model, mean of 90 cases (1).

3.2. Discrete optimisation by means of real discrete cross-section types

To establish a comparison with continuous optimisation, we use IPE cross-section types and
coding them in the chromosome as numerical index (16 types per bar, so 4 bits per bar are
required). The optimum reached that satis"es all the constraints, in the ideal model and detected in
the three strategies handled is detailed in Table 4.
The optimum reached that satis"es all the constraints, in the real model and detected in the three
strategies handled is detailed in Table 5.
390 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

Fig. 8. Continuous optimisation, real model, mean of 90 cases (2).

Fig. 9. Continuous optimisation, real model, best of 90 cases (1).


D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 391

Fig. 10. Continuous optimisation, real model, best of 90 cases (2).

Fig. 11. Continuous optimisation, real model, variance of 90 cases.


392 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

Table 4

Bar Cross-section type Comparative stress (MPa)

1 IPE } 330 233.24 Mass (kg) 3175.95


2 IPE } 500 160.43 Control displacement (10\2 m) 6.632
3 IPE } 400 138.38
4 IPE } 450 181.20

Table 5

Bar Cross-section type Comparative Stress (MPa)

1 IPE } 330 230.99 Mass (kg) 3324.32 (*3)


2 IPE } 500 178.16 Control displacement (10\2 m.) 6.608
3 IPE } 450 132.59
4 IPE } 500 164.44

It should be remarked, that even this genetic algorithm discrete solution surpasses the continu-
ous one of traditional optimisation referenced, which means the "nding of a local optimum not
detected in those cases. Moreover, if we approximate the genetic algorithm continuous solution (i.e.
real model) with IPE discrete cross-section types by excess, it constitutes the 42nd discrete solution
that satis"es the constraints by mass ordering (among a global set of solutions of
65535"2  *   "2) with a mass of 3492.9 kg, (1.05 of *3) showing the need for an e$cient
discrete optimisation method.
Figs. 12}19 present the evolutionary comparison of the three strategies proposed, with and
without the use of Gray coding, for both ideal and real models.
Comparing the statistical graphs between continuous and discrete cases, as a general remark it
should be noted that in the continuous case (with a greater search space), the capacity of
exploration of the CHC algorithm beats the other ones at the "nal stage. However, in the discrete
case (where the search space is considerably smaller) the capacity of exploitation of the steady-state
algorithm, which reaches the optimum location "rst, wins.

4. Introduction to the multiobjective problem

The minimisation of the number of di!erent cross-section types in large structures is important
for constructive requirements. Because of that, we propose to solve under a multiobjective
context both the minimum-mass problem and the minimisation of the number of cross-section
types.
The interaction among di!erent objectives leads to a set of compromised solutions, largely
known as the Pareto-optimal solutions [8]. Our objective is to "nd solutions close to the true
Pareto-optimal solutions and to "nd solutions that are di!erent from each other.
D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 393

Fig. 12. Discrete optimisation, ideal model, mean of 90 cases (1).

Fig. 13. Discrete optimisation, ideal model, mean of 90 cases (2).


394 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

Fig. 14. Discrete optimisation, ideal model, best of 90 cases.

Fig. 15. Discrete optimisation, ideal model, variance of 90 cases.


D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 395

Fig. 16. Discrete optimisation, real model, mean of 90 cases (1).

Fig. 17. Discrete optimisation, real model, mean of 90 cases (2).


396 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

Fig. 18. Discrete optimisation, real model, best of 90 cases.

Fig. 19. Discrete optimisation, real model, variance of 90 cases.


D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 397

The main problem in using classical search and optimisation methods is that these algorithms
cannot "nd multiple solutions in a single run. However, evolutionary search algorithms have
proved in the last few years to be very e$cient due to the use of a population of solutions in their
search, thereby, multiple Pareto-optimal solutions can be found in one single run [11].
The Srinivas and Deb [6] non-dominated sorting genetic algorithm, suggested in [7], varies from
a simple genetic algorithm only in the way the selection operator is used. The crossover and mutation
operators remain as usual. The classical "tness function is substituted by a ranking value. To all the
non-dominated members, a rank of one is assigned and then removed from the population. The
non-dominated members in the remaining part of the population are then assigned a rank of two and
this process is continued until the entire population is ranked. To each individual is then assigned
a "tness value based on its non-domination ranking (a solution x is said to dominate the other
solution x if both the following conditions are true: (a) The solution x is no worse than x in all
objectives and (b) the solution x is strictly better than x in at least one objective).
At each ranked front a sharing process is applied to maintain diversity in the population. First,
for each i solution, a normalised phenotypic Euclidean distance measure is computed:

  
xG !xH 
d " . N N ,
GH N x !x
N N
where P is the number of variables in the problem and x and x are the upper and lower bounds of
N N
variable x . The sharing function is computed then by comparing d with a pre-speci"ed parameter
N GH
 :
 
 



d
1! GH if d ) ,
Sh(d )"  GH   .
GH  
0 otherwise.
This process is repeated until all the points in the front are taken into account and then, the niche
count for the ith solution is calculated as follows: m " LI Sh(d ). This count is used to rank the
G H GH
solutions of the same front. The process continues for each solution of the front and for each front.
The  parameter can be calculated as follows [9]:  "0.5/(q,
. where in our case q"4
   
(number of optima to obtain) and P"6;4"24 (number of phenotypic variables, six per bar),
then  "0.471937.
 
Inspired by Tsutsui and Fujimoto [10,13], the inclusion of some kind of elitism in the NSGA
algorithm is proposed. So the next generation is formed with the inclusion of the half best children
in terms of ranking domination in the parental population, keeping its half bests in terms of
ranking domination. Some kind of duplicates elimination in the process of populations merging is
also recommended. Moreover, the constraint term is multiplied by two in order to prevent the
optimal solutions from having non-zero constraints as happens in the previous case. The probabil-
ity of crossover is therefore 1.0.

5. Numerical results for the multiple optimisation problem

With all that, results are improved in the optimal front of solutions, achieving the Pareto front,
that is localised in its optimal points in all the "ve cases executed, both considering ideal or real
398 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

Table 6

Bar Cross-section type Comp. stress (MPa)

For N"4
1 IPE } 330 233.24
2 IPE } 500 160.43
3 IPE } 400 138.38
4 IPE } 450 181.20
Mass (kg) 3175.95
Control displacement (10\ m) 6.632
For N"3
1 IPE } 360 224.22
2 IPE } 500 157.88
3 IPE } 300 150.23
4 IPE } 500 159.94
Mass (kg) 3225.85
Control displacement (10\ m) 6.726
For N"2
1 IPE } 360 206.39
2 IPE } 500 153.27
3 IPE } 360 136.61
4 IPE } 500 155.33
Mass (kg) 3280.67
Control displacement (10\ m) 6.392
For N"1
1 IPE } 500 105.25
2 IPE } 500 176.11
3 IPE } 500 84.08
4 IPE } 500 145.43
Mass (kg) 3824.52
Control displacement (10\ m) 5.670

model, until the "nal population (without elitism the optimum if founded, could be lost during the
evolutionary process). The population size is 100 individuals, the "nal generation is number 100,
the probability of mutation is 0.8%, N being the number of di!erent cross-section types:
Optional Pareto front for ideal model (see Table 6).
Optional Pareto front for real model (see Table 7).
Final fronts from "ve executions for ideal and real models, without and with elitism are shown in
Figs. 20}23.

6. Conclusions

It has been shown through the statistical "gures (di!erent mean, best and variance results
from a set of independent executions of various evolutionary strategies), that the selection of
D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 399

Table 7

Bar Cross-section type Comp. Stress (MPa)

For N"3
1 IPE } 330 230.99
2 IPE } 500 178.16
3 IPE } 450 132.59
4 IPE } 500 164.44
Mass (kg) 3324
Control displacement (10\ m.) 6.611
For N"2
1 IPE } 330 219.62
2 IPE } 500 185.42
3 IPE } 500 115.64
4 IPE } 500 161.41
Mass (kg) 3405.33
Control displacement (10\ m) 6.390
For N"1
1 IPE } 500 113.39
2 IPE } 500 191.79
3 IPE } 500 92.61
4 IPE } 500 158.76
Mass (kg) 3824.52
Control displacement (10\ m.) 6.196

Fig. 20. Multicriteria optimisation, ideal model, "nal fronts from "ve independent runs without elitism.
400 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

Fig. 21. Multicriteria optimisation, real model, "nal fronts from "ve independent runs without elitism.

Fig. 22. Multicriteria optimisation, Ideal Model, "nal fronts from "ve independent runs with elitism.
D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402 401

Fig. 23. Multicriteria optimisation, real model, "nal fronts from "ve independent runs with elitism.

appropriated strategies of genetic algorithms according to the nature of the treated problem is
important in order to reach high performance. In the continuous monocriteria optimisation case,
with a larger search space, the high explorative capacity of the CHC algorithm in the "nal stages
beats the other tested strategies. In the discrete monocriteria optimisation case, with a smaller
search space, the intense exploitative capacity in the initial stages of the steady-state strategy beats
the other tested strategies.
Even in small frames, genetic algorithms are an alternative to the traditional optimisation
methods. The characterisation of genetic algorithms as global optimizers allows them to "nd
solutions that for traditional methods (local optimizers) could be di$cult to achieve, which is
demonstrated in the continuous ideal model case: the GA continuous solution is about 3% better
than the traditional ones.
Moreover, the discrete upper approximation of the evolutionary continuous solution frame is
not a good discrete solution (i.e. there are 41 better solutions in the discrete real model case). So, it
shows the need for a discrete optimisation method that the genetic algorithms can provide.
About the encoding of the structural problem, "gures reveal that the use of Gray code seems to
be an improvement with respect to standard binary code in the discrete case of frame optimisation,
allowing better convergence curves, particularly in its best strategy: the steady-state.
The nondominated sorting genetic algorithm (NSGA) has been tested, showing its suitability to
solve the structural discrete multicriteria problem treated (optimising constrained mass and
number of di!erent cross-section types). Its correct behaviour in a discrete problem has been
proved, forming multiple Pareto-optimal solutions in one single run. Results of the test cases
highlight how elitism with some duplicates elimination implies a drastic improvement tool in the
NSGA e$ciency.
402 D. Greiner et al. / Finite Elements in Analysis and Design 37 (2001) 381}402

We can conclude that the properties of genetic algorithms as global, discrete and multicriteria
frame optimizers (shown through the results of the present paper), make them a powerful tool for
the engineering and design of frame structures.

References

[1] D. Greiner, G. Winter, J.M. Emperador, Genetic algorithm application in plane frame optimisation problems,
P.F.C. E.T.S.I.Industriales ULPGC, 2000.
[2] M. Galante, Genetic algorithms as an approach to optimise real-world trusses, Int. J. Numer. Methods Eng. 39
(1996) 361}382.
[3] M. Galante, Asimple genetic algorithm for plane trusses optimisation, Rev. Int. Met. Num. Para Calc. Disen o Ing.
9 (2) (1993) 179}199.
[4] S. Hernàndez Ibàn ez, Methods for optimum design of structures, ColeccioH n Seinor. Colegio de Ingenieros de
Caminos, Canales y Puertos, Madrid, 1990.
[5] G. Winter, M. GalaH n, P. Cuesta, D. Greiner, Genetic algorithms: a stochastic improvement technique. Tools, skills,
pitfalls and examples, Genetic Algorithms in Engineering and Computer Science (EUROGEN 1995), Wiley, New
York.
[6] N. Srnivas, K. Deb, Multiobjective optimisation using non-dominated sorting in genetic algorithms, Evol. Compu.
2 (3) (1995) 221}248.
[7] D.E. Goldberg, Genetic algorithms for Search, Optimisation, and Machine Learning, Addison-Wesley, Reading,
MA, 1989.
[8] V. Chankong, Y. Haimes, Multiobjective Decision Making Theory and Methodology, North-Holland, New York,
1983.
[9] K. Deb, D.E. Goldberg, An investigation of niche and species formation in genetic function optimisation,
Proceedings of the Third International Conference on Genetic Algorithms, 1989.
[10] S. Tsutsui, Y. Fujimoto, Forking genetic algorithms with blocking and shrinking modes (fGA), Proceedings of the
"fth International Conference on Genetic Algorithms, Morgan Kaufmann Publishers, Inc San Mateo, 1993, (pp.
206}213).
[11] K. Deb, Evolutionary algorithms for multi-criterion optimisation in engineering design, Evolutionary Algorithms
in Engineering and Computer Science (EUROGEN 1999), Wiley, New York.
[12] D. Whitley, S. Rana, R. Heckendorn, Representation issues in neighbourhood search and evolutionary algorithms,
Genetic Algorithms and Evolution Strategies in Engineering and Computer Science (EUROGEN 1997), Wiley
New York.
[13] S. Obayashi, Pareto Genetic Algorithm for Aerodynamic Design Using the Navier-Stokes Equations Genetic
Algorithms and Evolution Strategies in Engineering and Computer Science (EUROGEN 1997), Wiley, New York.
[14] L. Eshelman, J.D. Scha!er, Preventing premature convergence in genetic algorithms by preventing incest, Proceed-
ings of the Fourth International Conference on Genetic Algorithms, 1991, pp. 115}121.
[15] D. Whitley, The GENITOR algorithm and selection pressure: why rank-based allocation of reproductive trials is
best, Proceedings of the Third International Conference on Genetic Algorithms, 1989, pp. 116}121.
[16] S. W. Mahfoud, Niching methods for genetic algorithms, Illegal Report no. 95001, 1995.
[17] L. Davis, Handbook of Genetic Algorithms, Van Nostrand Reinhold, New York, 1991.
[18] D. Goldberg, M. P. Santani, Engineering optimization via genetic algorithm, Proceedings of the Ninth Conference
On Electronic Computation, ASCE, New York, 1986, pp. 471}482.

Potrebbero piacerti anche