Computers and Structures 80 (2002) 571–589

www.elsevier.com/locate/compstruc

Structural optimization using evolutionary algorithms

Nikolaos D. Lagaros, Manolis Papadrakakis *, George Kokossalakis
Institute of Structural Analysis & Seismic Research, National Technical University Athens, Zografou Campus,
Athens 157 80, Greece
Received 6 February 2001; accepted 9 January 2002

Abstract
The objective of this paper is to investigate the efficiency of various evolutionary algorithms (EA), such as genetic
algorithms and evolution strategies, when applied to large-scale structural sizing optimization problems. Both type of
algorithms imitate biological evolution in nature and combine the concept of artificial survival of the fittest with
evolutionary operators to form a robust search mechanism. In this paper modified versions of the basic EA are im-
plemented to improve the performance of the optimization procedure. The modified versions of both genetic algorithms
and evolution strategies combined with a mathematical programming method to form hybrid methodologies are also
tested and compared and proved particularly promising. The numerical tests presented demonstrate the computational
advantages of the discussed methods, which become more pronounced in large-scale optimization problems.
2002
Elsevier Science Ltd. All rights reserved.

Keywords: Structural optimization; Genetic algorithms; Evolution strategies; Handling of constraints; Sequential quadratic
programming

1. Introduction Both GA and ES imitate biological evolution in na-

ture and have three characteristics that differ from other
Computer algorithms based on the process of natural conventional optimization algorithms: (i) In place of the
evolution have been found capable to produce very usual deterministic operators, they use randomized op-
powerful and robust search mechanisms although the erators: mutation, selection and recombination. (ii) In-
similarity between these algorithms and the natural stead of a single design point, they work simultaneously
evolution is based on crude imitation of biological re- with a population of design points in the space of de-
ality. The resulting evolutionary algorithms (EA) are sign variables. (iii) They can handle, with minor modi-
based on a population of individuals, each of which fications, continuous, discrete or mixed optimization
represents a search point in the space of potential solu- problems. The second characteristic allows for natural
tions of a given problem. These algorithms have some implementation of GA and ES on a parallel computing
selection process based on the fitness of the individuals environment [7,8].
and some recombination operators. The best known EA In a gradient-based mathematical programming ap-
in this class include evolutionary programming (EP) proach the optimization algorithm proceeds with the
[1,2], genetic algorithms (GA) [3,4] and evolution strat- following steps: (i) Evaluation of displacements and
egies (ES) [5,6]. stresses. (ii) Computation of sensitivities by perturbing
each design variable by a small amount. (iii) Solution of
the optimization problem and update the design of the
*
Corresponding author. Tel.: +301-772-1694; fax: +301-772- structure. These steps are repeated until convergence is
1693. achieved. The most time-consuming part of this process
E-mail address: mpapadra@central.ntua.gr (M. Papadraka- is related to the sensitivity analysis phase, which is an
kis). important ingredient of all mathematical programming

0045-7949/02/$ - see front matter
2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 4 5 - 7 9 4 9 ( 0 2 ) 0 0 0 2 7 - 5
572 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589

optimization methods and may consume a large part of large-scale and computationally intensive optimization
the total computational effort [9]. On the other hand, the problems.
application of EA that are based on probabilistic
searching, such as GA and ES, do not need gradient
information and therefore avoid performing the com- 2. Genetic algorithms
putationally expensive sensitivity analysis step.
Mathematical programming methods, such as the GA are probably the best-known EA, receiving sub-
sequential quadratic programming (SQP) approach, stantial attention in recent years. The first attempt to use
make use of local curvature information derived from EA took place in the sixties by a team of biologists [10]
linearization of the original functions by using their and was focused in building a computer program that
derivatives with respect to the design variables. The would simulate the process of evolution in nature.
linearization is performed at points obtained in the However, the GA model used in this study and in many
process of optimization to construct an approximate other structural design applications refers to a model
model of the initial problem. These methods present a introduced and studied by Holland and co-workers [4].
satisfactory local rate of convergence, but they cannot In general the term genetic algorithm refers to any pop-
assure that the global optimum can be found. They do ulation-based model that uses various operators (selec-
although assure if the problem is strictly convex. On the tion–crossover–mutation) to evolve. In the basic genetic
other hand, EA are in general more robust and present a algorithm each member of this population will be a bi-
better global behaviour than the mathematical pro- nary or a real valued string, which is sometimes referred
gramming methods. They may suffer, however, from a to as a genotype or, alternatively, as a chromosome.
slow rate of convergence towards the global optimum
and do not guarantee convergence to the global opti- 2.1. The basic genetic algorithms
mum.
Structural optimization problems are characterized 2.1.1. The three main steps of the basic GA
by various objective and constraint functions, which are Step 0 initialization: The first step in the implemen-
generally non-linear functions of the design variables. tation of any genetic algorithm is to generate an initial
These functions are usually implicit, discontinuous and population. In most cases the initial population is gen-
non-convex. The mathematical formulation of structural erated randomly. In this study in order to perform a
optimization problems with respect to the design vari- comparison between various optimization techniques
ables, the objective and constraint functions depends on the initial population is fixed and is chosen in the neigh-
the type of the application. However, all optimization borhood of the initial design used for the mathematical
problems can be expressed in standard mathematical programming method. After creating an initial popula-
terms as a non-linear programming problem (NLP), tion, each member of the population is evaluated by
which in general form can be stated as follows: computing the representative objective and constraint
functions and comparing it with the other members of
min F ðsÞ the population.
subject to hj ðsÞ 6 0 j ¼ 1; . . . ; m ð1Þ Step 1 selection: Selection operator is applied to the
with sli 6 si 6 sui i ¼ 1; . . . ; n current population to create an intermediate one. In the
first generation the initial population is considered as
where, s is the vector of design variables, F ðsÞ is the the intermediate one, while in the next generations this
objective function to be minimized, hj ðsÞ are the be- population is created by the application of the selection
havioral constraints, sli and sui are the lower and the operator.
upper bounds of a typical design variable si . Equality Step 2 generation (crossover–mutation): In order to
constraints are rarely imposed in this type of problems create the next generation, crossover and mutation op-
except in some cases for design variable linking. erators are applied to the intermediate population to
Whenever they are used they are treated for simplicity as create the next population. Crossover is a reproduction
a set of two inequality constraints. operator, which forms a new chromosome by combining
In this work the efficiency of various EA is investi- parts of each of the two parental chromosomes. Muta-
gated in structural sizing optimization problems. Fur- tion is a reproduction operator that forms a new chro-
thermore, in order to benefit from the advantages of mosome by making (usually small) alterations to the
both methodologies, combinations of EA with SQP are values of genes in a copy of a single parent chromosome.
also examined in an attempt to increase further the ro- The process of going from the current population to the
bustness as well as the computational efficiency of the next population constitutes one generation in the evo-
optimization procedure. The numerical tests presented lution process of a genetic algorithm. If the termination
demonstrate the computational advantages of the dis- criteria are satisfied the procedure stops, otherwise, it
cussed methods, which become more pronounced in returns to step 1.
 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589 573

2.2. Micro genetic algorithms The methods based on the use of penalty functions are
employed in the majority of cases for treating constraint
The micro genetic algorithm (lGA) was introduced optimization problems with GA. In this study methods
by Krishnakumar [11] and applied to simple mathe- belonging to this category have been implemented and
matical test functions and to the wind shear optimal will be briefly described in the following section.
guidance problem. The main objective of this scheme is
to reduce the size of the population compared to the 3.1. Method of static penalties
basic one. This corresponds, in the case of structural
optimization problems discretized with finite elements, In the method of static penalties the objective func-
to less finite element analyses per generation. It is a tion is modified as follows:
known fact that GA generally exhibit poor performance
ðnÞ
with very small population due to insufficient informa- F ðsÞ; if s 2 F
F 0 ðsÞ ¼ ð2Þ
tion processed and premature convergence to non-opti- F ðnÞ ðsÞ þ p  violðnÞ ðsÞ; otherwise
mal results. A remedy to this problem, suggested by
Goldberg [12], could be to restart the evolution process where p is the static penalty parameter, violðnÞ ðsÞ is the
in case of nominal convergence with a new initial pop- sum of the violated constraints and F ðnÞ ðsÞ is the objec-
ulation which will include the best solution already tive function to be minimized, both normalized in [0,1],
achieved. Based on this suggestion Krishnakumar pro- while F is the feasible region of the design space.
posed the lGA which can be described by the following X
m

steps: violðsÞ ¼ hj ðsÞ ð3Þ

j¼1
Step 0 initialization: The first step generates a popu-
lation of size five either randomly or by generating four
The sum of the violated constraints is normalized before
strings randomly and by selecting one good string from
it is used for the calculation of the modified objective
any previous search, or according to the experience of
function. The main advantage of this method is its sim-
the designer.
plicity. However, there is no guidance on how to choose
Step 1 fitness evaluation: In this step the fitness of
the single penalty parameter p. If it is chosen too small
each individual is evaluated and the best string is de-
the search will converge to an infeasible solution, oth-
termined. The best string is labeled as string five and it
erwise, if it is chosen too large, a feasible solution may
is carried to the next generation (elitist strategy). In this
be located but it would be far from the global optimum.
way there is a guarantee that the information about
A large penalty parameter will force the search proce-
good strings are not lost.
dure to work away from the boundary where the global
Step 2 generation: According to the previous step the
optimum is usually located and divides the feasible re-
best individual of the current generation is carried out to
gion from the infeasible one.
the next one. The remaining four members of the next
generation are chosen according to the tournament se-
lection operator. After the selection operator is termi- 3.2. Method of dynamic penalties
nated the crossover operator is applied.
Step 3 convergence check: If the termination criteria is The method of dynamic penalties was proposed by
satisfied the process ends, otherwise check for nominal Joines and Houck [13] and applied to mathematical test
convergence which is measured by bit wise convergence functions. As opposed to the previous method, dynamic
in case of binary coding or by comparing the design penalties are implemented in this case. Individuals are
variables in case of real valued strings. If converged go evaluated (at the generation g) by the following formula
to step 0, else return to Step 1. F 0 ðsÞ ¼ F ðnÞ ðsÞ þ ðc  gÞa violðnÞ ðsÞ ð4Þ
A modified version of lGA is tested in this study,
where only feasible designs are accepted for the evolu-
X
m
tion process. This version, which resembles the death violðsÞ ¼ hbj ðsÞ ð5Þ
penalty treatment of the constraints adopted by ES, is j¼1
abbreviated to m lGA.
where c, a and b are constants. A reasonable choice for
these parameters was proposed as follows: c ¼ 0:5–2:0,
3. Methods for handling the constraints a ¼ b ¼ 1 or 2. For high generation number, however,
the ðc  gÞa component of the penalty term takes ex-
Although genetic algorithms were initially developed tremely large values which makes even the slightly
to solve unconstrained optimization problems, during violated designs not to be selected in subsequent
the last decade several methods have been proposed generations. Thus, the system has little chances to escape
for handling constrained optimization problems as well. from local optima. In most experiments reported by
574 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589

Michalewicz [14] the best individual was found in early are evaluated using the ph penalty parameter, while the
generations. remaining designs N þ 1; . . . ; 2N are evaluated using the
p‘ penalty parameter.
3.3. Augmented Lagrangian method Step 1 selection: An intermediate population of size
N is created by selecting the best individuals from the
The Augmented Lagrangian method (AL-GA) was two populations.
proposed by Adeli and Cheng [15,16]. According to Step 2 generation: Generate N offsprings using the
this method the constrained problem is transformed basic operators mutation and crossover. The parents are
to an unconstrained problem, by introducing two sets evaluated using the ph penalty parameter while the off-
of penalty coefficients c½ðc1 ; c2 ; . . . ; cMþN Þ and l½ðl1 ; springs using the p‘ . The process is then repeated by
l2 ; . . . ; lM þN Þ . The modified objective function, for the returning to Step 1.
generation g, is defined as follows:
8 This version was used in [18] for the minimal weight
<X N design problem of a composite laminated plate.
1 1 ðgÞ ðgÞ
F 0 ðs; c; lÞ ¼ F ðsÞ þ c ½ðqj
1 þ lj Þþ 2
Lf 2 : j¼1 j
"   !þ #2 9
XM  d  = 4. Evolution strategies
ðgÞ j
þ cjþN  a 
1 þ lðgÞ ð6Þ
d j  jþN
;
j¼1
4.1. Basic evolution strategies
where Lf is a factor for normalizing the objective func-
In the majority of cases ES were applied to contin-
tion; qj is a non-dimensional ratio related to the stress
uous optimization problems. In engineering practice the
constraints of the jth element group (see Eqs. (18) and
design variables are not continuous because usually the
(19)); dj is the displacement in the direction of the jth
structural parts are constructed with certain variation of
examined degree of freedom, while dja is the corre-
their dimensions. Thus design variables can only take
sponding allowable displacement; N, M correspond to
values from a predefined discrete set. For the solution of
the number of stress and displacement constraint func-
discrete optimization problems Thierauf and Cai [19]
tions, respectively. Furthermore
have proposed a modified ES algorithm. The basic dif-
ðgÞ ðgÞ
ðqj
1 þ lj Þþ ¼ maxðqj
1 þ lj ; 0Þ ð7Þ ferences between discrete and continuous ES are focused
on the mutation and the recombination operators. The
  !þ   ! multi-membered ES adopted in the current study uses
d  d 
 j 
1 þ lðgÞ ¼ max  j 
1 þ lðgÞ ð8Þ three operators: recombination, mutation and selection
dja  jþN dja  jþN ; 0
operators that can be included in the algorithm as fol-
lows:
The penalty coefficients are updated at each generation Step 1 (recombination and mutation): The population
ðgþ1Þ ðgÞ ðgÞ
according to the expressions cj ¼ b  cj and lj ¼ of l parents at gth generation produces k offsprings. The
ðgÞ ðgþ1Þ ðgÞ ðgÞ ðgÞ
lj =b, where lj ¼ lj þ max½conj;ave ;
lj and genotype of any descendant differs only slightly from
ðgÞ
conj;ave is the average value of the jth constraint function that of its parents. For every offspring vector a tempo-
for the gth generation, while the initial values of c’s and rary parent vector ~s ¼ ½~s1 ; ~s2 ; . . . ; ~sn T is first built by
l’s are set equal to three and zero, respectively. Coeffi- means of recombination. For discrete problems the
cient b is taken equal to ten as recommended by Be- following recombination cases can be used
legundu and Arora [17]. 8
>
> sa;i or sb;i randomly ðAÞ
>
>
3.4. Segregated genetic algorithm < sm;i or sb;i randomly ðBÞ
~si ¼ sbj;i ðCÞ ð9Þ
>
>
The basic idea of the segregated GA (S-GA) [18] is > sa;i or sbj;i randomly ðDÞ
>
:
to use, as in the method of static penalties, two static sm;i or sbj;i randomly ðEÞ
penalty parameters instead of one. The two values of the
penalty parameters are associated with two populations ~si is the ith component of the temporary parent vector ~s,
that have a different level of satisfaction of the con- sa;i and sb;i are the ith components of the vectors sa and
straints. Each of the groups corresponds to the best sb which are two parent vectors randomly chosen from
performing individuals with respect to the associated the population. The vector sm is not randomly chosen
penalty parameter. The S-GA can be described as fol- but is the best of the l parent vectors in the current
lows: generation. In case C of Eq. (9), ~si ¼ sbj;i means that the
Step 0 initialization: Random generation of 2N de- ith component of ~s is chosen randomly from the ith
signs. The objective functions of the designs 1; 2; . . . ; N components of all l parent vectors. From the temporary
 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589 575

parent ~s an offspring can be created following the mu- tion value and the mean value of the objective function
tation operator. values from all parent vectors in the current generation
Let as consider the temporary parent sðgÞp of the is less than a given value ec ð¼ 0:0001Þ, (iv) when the
generation g that produces an offspring sðgÞ
o through the ratio lb =l has reached a given value ed ð¼ 0:5–0:8Þ where
mutation operator as follows lb is the number of the parent vectors in the current
generation with the best objective function value.
sðgÞ ðgÞ
o ¼ sp þ z
ðgÞ
ð10Þ

ðgÞ ðgÞ T
4.2. Contemporary ES––the (l,k,h) evolution strategies
where zðgÞ ¼ ½z1 ; z2 ; . . . ; zðgÞ
n is a random vector. Mu-
tation is understood to be random, purposeless events, This is a more general ES version, which was pro-
which occur very rarely. The fact that the difference posed by Schwefel and Rudolph [20] for application in
between any two adjacent values can be relatively large continuous problems but has not been applied either
is against the requirement that the variance r2i should be to continuous or to discrete optimization problems [21].
small. For this reason it is suggested [19] that not all the The two schemes of the multi-membered evolution
components of a parent vector, but only a few of them strategy, namely the ðl þ kÞ-ES and the ðl; kÞ-ES, differ
(e.g. ‘), should be randomly changed in every genera- in the way the parents of a new generation are selected.
tion. This means that n
‘, components of the ran- So far, only empirical results have shown that the ‘plus’
domly changed vector zðgÞ will have zero value. In other version performs better in structural optimization
words, the terms of vector zðgÞ are derived from problems [7,19].

The ðl; kÞ-ES version is in danger to diverge because
ðgÞ ðj þ 1Þdsi for ‘ randomly chosen components the so far best position is not preserved within the gen-
zi ¼
0 for n
‘ other components eration cycle (the so-called non-elitist strategy). The
ð11Þ ‘comma’ version implies that each parent can have
children only once (duration of life: one generation or
where dsi is the difference between two adjacent values in one reproduction cycle), whereas in the ‘plus’ version
the discrete set and j is a random integer number, which individuals may live eternally if no child achieves a
follows the Poisson distribution better or at least the same improvement in the objective
function value. The contemporary ES (C-ES) introduce
ðcÞj
c a maximal life span of h P 1 reproduction cycles which
pðjÞ ¼ e ð12Þ gives the ‘comma’ scheme for h ¼ 1 and the ‘plus’ one
c!
for h ¼ 1. If l P 1 is the number of parents, k > l is the
number of offspring, then q with 1 6 q 6 l is the number
c is the standard deviation as well as the mean value of
of ancestors for each descendant. This ES version differs
the random number j. The choice of ‘ depends on the
in two points from the basic one: (i) free number of
size of the problem and it is usually taken as 1=5 of the
parents are involved in reproduction ranging from 1 to
total number of design variables. The ‘ components are
l, (ii) a finite number of reproduction cycles per indi-
selected using uniform random distribution in every
vidual is performed, not one (1) or infinite (1) for the
generation according to Eq. (11).
‘comma’ and the ‘plus’ schemes, respectively. The se-
Step 2 (selection): There are two different types of
lection, mutation and recombination operators used in
the multi-membered ES:
the C-ES are the same as described in the section of the
basic evolution strategies.
ðl þ kÞ-ES: The best l individuals are selected from a
temporary population of ðl þ kÞ individuals to form
the parents of the next generation. 4.3. Adaptive ES
ðl; kÞ-ES: The l individuals produce k offsprings
ðl 6 kÞ and the selection process defines a new popula- The handling of the constraints by the basic ES is
tion of l individuals from the set of k offsprings only. based on the death penalty approach [22], where every
infeasible design point is discarded. Thus the process is
For discrete optimization the procedure terminates directed to search only in the feasible region of the de-
when one of the following termination criteria is satis- sign space. Due to this approach many designs that are
fied: (i) when the best value of the objective function in examined by the optimizer during the search process and
the last 4nl=k generations remains unchanged, (ii) when are close to the acceptable design space are rejected
the mean value of the objective values from all parent leading to the loss of valuable information. The idea
vectors in the last 2nl=k generations has not been im- introduced in this work is to use soft constraints dur-
proved by less than a given value eb ð¼ 0:0001Þ, (iii) when ing the first stages of the search and as the search
the relative difference between the best objective func- approaches the region of the global optimum the
576 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589

constraints to become more severe until they reach their 8. Selection step: Selection of the next generation par-
real values. ents according to (l þ k) or (l; k) selection schemes.
The implementation of adaptive ES (A-ES) in a 9. Convergence check: If satisfied stop, else return to
structural optimization problem is straightforward and Step 3.
follows the same steps described in the section of the
basic ES. The ES optimization procedure starts with a 4.4. An academic example
population of parent vectors, while a level of violation of
the constraints is determined. If any of these parents As an example for explaining the process of ES we
corresponds to an infeasible design lying outside the consider the three-bar truss shown in Fig. 1, where the
extended design space then this parent is modified until minimum volume is required to support a force P .
it becomes ‘feasible’. Then the offsprings are generated This structure has been used as a test bed in the lit-
and they are also checked if they are in the ‘feasible’ erature of structural optimization [23] and must satisfy
region according the current level of violation. In every various constraints, such as member crushing, member
generation the values of the objective function are buckling and failure by excessive deflection of node 4.
compared between the parent and the offspring vectors The final design of the structure must be symmetric.
and the worst vectors are rejected, while the remaining Therefore, the following design variables are defined:
ones are considered to be the parent vectors of the new A1 ¼ cross-sectional area of material for members l and
generation. This procedure is repeated until the chosen 3, and A2 ¼ cross-sectional area of material for member
termination criterion is satisfied. 2. The relative merit of any design for the problem is
In this adaptive scheme a nominal convergence check measured in its material volume. Thus, the total material
is adopted for the determination of the level of violation volume for the structure serves as an objective function:
of constraints. Nominal convergence occurs when the pffiffiffi
mean value of the objective function of the designs of the volume ¼ Lð2 2A1 þ A2 Þ
current population is relatively close to the best design
achieved until the current generation, according to the where L is defined in Fig. 1.
expression To define the constraint functions for the problem,
stresses and deflections for the structure are calculated.
ðgÞ ðgÞ Using analysis procedures for statically indeterminate
F
Fbest
6 ead ð13Þ structures, horizontal and vertical displacements u and v,
ðgÞ
F respectively, of the node 4 are given by
pffiffiffi
ðgÞ
where F is the mean objective function value and Fbest
ðgÞ 2LPu
u¼
is the best objective function value of all parents in the A1 E
gth generation, where ead ¼ 0:05.
The A-ES steps can be stated as follows: pffiffiffi
2LPv
v¼ pffiffiffi
ðA1 þ 2A2 ÞE
1. Initialization step: Selection of si , (i ¼ 1; 2; . . . ; l) par-
ent vectors of the design variables and the percentage where E is the modulus of elasticity for the material,
of violation of the constraints v0 (usually taken be- while Pu and Pv are the horizontal and vertical compo-
tween 20% and 50%). nents of the load P, respectively: Pu ¼ P cos h, Pv ¼
2. Analysis step: Solve Kðsi Þui ¼ f (i ¼ 1; 2; . . . ; l),
where K is the stiffness matrix of the structure and f
is the loading vector.
3. Constraints check: All parent vectors become ‘‘feasi-
ble’’, within the prescribed level of constraints viola-
tion v0 .
4. Offspring generation: Generate sj , (j ¼ 1; 2; . . . ; k) off-
spring vectors of the design variables.
5. Analysis step: Solve Kðsj Þuj ¼ f (j ¼ 1; 2; . . . ; k).
6. Nominal convergence check: If nominal convergence
has occurred the level of violation vg becomes more
severe by reducing its value by the quantity b (usually
0.1 or 0.2).
7. Constraints check: If satisfied according to the current
level of violation vg continue, else change sj and re-
turn to Step 4. Fig. 1. Symmetric three-bar truss.
 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589 577

P sin h. Stresses r1 , r2 and r3 in members 1–3 under the Table 1

load P are computed from member forces as follows: Design data for the three-bar structure
" # Allowable stress Members 1 and 3:
1 Pu Pv r1a ¼ r3a ¼ 34:5 MPa
r1 ¼ pffiffiffi þ pffiffiffi ðaÞ Member 2: r2a ¼ 137:9 MPa
2 A1 ðA1 þ 2A2 Þ
Allowable displacements Du ¼ 0.0127 cm
pffiffiffi Dv ¼ 0.0127 cm
2Pv Modulus of elasticity E ¼ 69:0E þ 03 MPa
r2 ¼ pffiffiffi ðbÞ Constant (assuming the sec- b ¼ 1:0
ðA1 þ 2A2 Þ
tions to be thin)
" # No. of design variables n¼2
1 Pv Pu Lower limits of the design (5.07, 5.07) cm2
r3 ¼ pffiffiffi pffiffiffi
ðcÞ
2 ðA1 þ 2A2 Þ A1 variables
Upper limits of the design (101.4, 101.4) cm2
variables
Note from Eqs. (a) and (c) that r1 is always greater Parameter L ¼ 2:54 cm
than r3 . Therefore, we need to impose constraints only Load P ¼ 178 kN
on r1 and r2 . If ra is an allowable stress for the material, Angle h ¼ 45
then
r1 6 ra and r2 6 ra " #
1 Pv Pu p2 EbA1
Horizontal and vertical deflections of the node 4 must be
pffiffiffi pffiffiffi
6
2 ðA1 þ 2A2 Þ A 1 ‘21
within the specified limits Du and Dv, respectively. Thus
u 6 Du and v 6 Dv. We will describe now step by step the ES generation
To impose buckling constraints for members under of the first population from the initial one for the design
compression, the dependence of moment of inertia I on data shown in Table 1. The members of the truss are
the cross-sectional area of the members must be speci- taken to be tabular sections.
fied. A form with quite general applicability is I ¼ bA2 , The selection scheme used in this application is the
where A is the cross-sectional area and is a non-dimen- ð5 þ 5Þ which means l ¼ 5 parents produce k ¼ 5 off-
sional constant. This relation follows if the shape of the springs selected according to the ðl þ kÞ scheme. The
cross-section is fixed and all its dimensions are varied in current difference between two adjacent values dsi of
the same proportion. The axial force for the ith member Eq. (11) is taken 2.535 cm2 and ‘ ¼ 1 (Eq. (11)). The
is Fi ¼ Ai ri , where i ¼ 1, 2, 3 with tensile force taken as recombination type D of Eq. (9) is used according to
positive. Members of the truss are considered bars with which the best member of the parent population is re-
pin ends. Therefore, the buckling load for the ith combined with a randomly chosen member of the parent
member is population. Termination criterion (ii) of ES is imple-
p2 Ei Ii mented according to which the procedure is terminated
Pb ¼ when the mean value of the objective values from all
‘2i
parent vectors in the last four generation has not been
where ‘i is the length of the ith member. Buckling con- improved by less than the given value eb ¼ 0:001.
straints are expressed as
Initial population
p2 Ei Ii Member of the Design (A1 , A2 ) Volume (l)

Fi 6
‘2i population (cm2 )
The negative sign for Fi is used to make the left-hand 1 (50.7, 50.7) 24.0
side of the constraints negative when the member is in 2 (71.0, 65.9) 33.0
tension since there is no need to impose buckling con- 3 (71.0, 71.0) 33.7
straints for members in tension. Substituting various 4 (40.6, 60.8) 21.7
quantities, member buckling constraints take the form 5 (40.6, 35.5) 18.6
" #
1 Pu Pv p2 EbA1

pffiffiffi þ pffiffiffi 6 • Recombine member 5 with member 1 to obtain the

2 A1 ðA1 þ 2A2 Þ ‘21 first temporary offspring vector (40.6, 50.7). The mu-
tation operator is performed in one of the two design
pffiffiffi variable since ‘ ¼ 1.
2Pv p2 EbA2

pffiffiffi 6 To randomly select which one of the two design

ðA1 þ 2A2 Þ ‘22 variable is to be mutated the following procedure
578 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589

is followed: Generate a random number that fol- Second population

lows the Gauss distribution in the (0,1): Member of the Design (A1 , A2 ) Volume
r ¼ 0:71471. population (cm2 ) (l)
Transform of the random variable r to an integer
one: 1 (43.1, 50.7) 21.4
i1 ¼ intðr ðn þ 1ÞÞ þ 1 ¼ 3 2 (40.6, 38.0) 18.9
if ði1:ge:ðn þ 1ÞÞi1 ¼ n þ 1 3 (38.0, 60.8) 20.9
i1 ¼ i1
ðn þ 2Þ=2 ¼ 1 if ði1:gt:0Þxi ¼ xi þ zi else 4 (68.4, 35.5) 28.4
x i ¼ x i
zi 5 (40.6, 35.5) 18.6
i2 ¼ absði1Þ ¼ 1 which means that the first design
variable is to be mutated. The procedure is repeated until the termination criteria
Mutate the design variable A1 with j ¼ 0 [24] is satisfied, for the current test example this is achieved
(Eq. (12)) to produce the first feasible offspring for the total number of generations 14 and 86 total
(43.1, 50.7) with objective function value equal number of FE analyses. The best design achieved is
to 21.4 l. the (30.4, 22.8) with objective function value equal to
• Recombine member 5 with member 5 to obtain the 13.5 l.
second temporary offspring vector (40.6, 35.5). Mu-
tate the design variable A2 with j ¼ 0 to produce
the corresponding feasible offspring (40.6, 38.0) with
objective function value equal to 18.9 l. 5. Gradient-based sizing optimization
• Recombine member 5 with member 4 to obtain the
third temporary offspring vector (40.6, 60.8). Mutate 5.1. Sensitivity analysis
the design variable A1 with j ¼ 0 to produce the cor-
responding feasible offspring (38.0, 60.8) with objec- Sensitivity analysis is the most important and time-
tive function value equal to 20.9 l. consuming part of a gradient-based sizing optimization
• Recombine member 5 with member 3 to obtain the procedure. The methods for sensitivity analysis can be
fourth temporary offspring vector (71.0, 35.5). Mu- divided into discrete and variational methods [25]. In the
tate the design variable A1 with j ¼ 0 to produce discrete approach the derivatives or the sensitivities of
the corresponding feasible offspring (68.4, 35.5) with the objective and constraint functions are evaluated
objective function value equal to 28.4 l. using the finite element equations of the discretized
• Recombine member 5 with member 1 to obtain the structure. Generally the effort required for the computer
fifth temporary offspring vector (50.7, 35.5). Mutate implementation of the discrete methods is less than the
the design variable A2 with j ¼ 1 to produce the cor- effort required for the implementation of the variational
responding feasible offspring (50.7, 30.4) with objec- methods.
tive function value equal to 21.5 l. A further classification of the discrete methods is the
following [26]:
If an offspring happens to be infeasible then the re-
combination of the best member with a randomly cho- (i) Global finite difference method: A full finite element
sen member is performed followed by the mutation step analysis has to be performed for each design variable
until a feasible design is achieved. and the accuracy of the method depends strongly on
the value of the perturbation of the design variables.
The global finite difference method (GFD) scheme is
Offspring population usually sensitive to the accuracy of the computed per-
Member of the Design (A1 , A2 ) Volume turbed displacement vectors which is dependent on
population (cm2 ) (l) the magnitude of the perturbation of the design vari-
ables. The magnitude of this perturbation is usually
1 (43.1, 50.7) 21.4
taken between 10
3 and 10
5 .
2 (40.6, 38.0) 18.9
(ii) Semi-analytical method: The stiffness matrix of
3 (38.0, 60.8) 20.9
the initial finite element solution is retained during
4 (68.4, 35.5) 28.4
the computation of the sensitivities. This provides
5 (50.7, 30.4) 21.5
an improved efficiency over the finite difference meth-
od by a relatively small increase in the algorithmic
complexity. The accuracy problem involved in the
We then select among the (l þ k) population pool the numerical differentiation can be overcome by using
best l individuals. Thus, the second population is as the ‘‘exact’’ semi-analytical (ESA) method which
follows: needs more programming effort than the simple
 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589 579

semi-analytical one but it is computationally more ef- based on an SQP algorithm from the NAG library [32].
ficient [27,28]. The primal algorithm is divided into three phases: (i) the
(iii) Analytical method: The finite element equations, solution of the QP subproblem to obtain the search di-
the objective and constraint functions are differenti- rection, (ii) the line search along the search direction,
ated analytically. (iii) the update of the Hessian matrix H.
Once the direction vector is found a line search is
The decision on which method to implement depends performed, involving only the non-linear constraints,
strongly on the type of problem, the structure of the in order to produce a ‘‘sufficient decrease’’ to the merit
computer program and the access to the source code. function u. This merit function is an augmented La-
The implementation of analytical and semi-analytical grangian function of the form [31]
methods is more complex and requires access to the X 1X
source code, whereas when a finite difference method is u ¼ F ðsÞ
ki ðgi ðsÞ
ci Þ þ pi ðgi ðsÞ
ci Þ2
applied the formulation is much simpler and the sen- i
2 i
sitivity coefficients can be easily evaluated even with ð16Þ
general purpose commercial codes. In the present in-
vestigation both the global finite difference method and where ci are the non-negative slack variables of the in-
the semi-analytical method have been used. equality constraints derived from the solution of the QP
subproblem. These slack variables allow the active in-
5.2. Sequential quadratic programming equality constraints to be treated as equalities and avoid
possible discontinuities. Finally, pi are the penalty pa-
SQP methods are the standard general purpose rameters which are initially set to zero and in subsequent
mathematical programming algorithms for solving NLP iterations are increased whenever this is necessary in
optimization problems [29]. Such methods make use of order to control the violation of the constraints and to
local curvature information derived from linearization ensure that the merit function follows a descent path
of the original functions, by using their derivatives with [30].
respect to the design variables at points obtained in the To implement SQP in discrete optimization prob-
process of optimization. Thus a quadratic programming lems, a continuous SQP step is performed first at each
(QP) model (or subproblem) is constructed from the step. Then the current continuous design reached is
initial NLP problem. A local minimizer is found by projected to the nearest discrete values of the design
solving a sequence of these QP subproblems using a space. The current continuous design is considered as a
quadratic approximation of the objective function. Each lower bound of the projected discrete one [33].
subproblem has the following form
1 T
minimize 2
p Hp þ gT p 6. The hybrid approach
with Ap þ hðsÞ 6 0 ð14Þ
s‘ 6 p 6 su Hybrid methods which combine evolutionary com-
putation techniques with deterministic procedures for
where p is the search direction subjected to upper and lower numerical optimization problems have been recently
bounds, g is the gradient of the objective function, A is the investigated. Papadrakakis et al. [34] used evolution
Jacobian of the constraints, usually the active ones strategies with the SQP method, while Waagen et al. [35]
only (i.e. those that are either violated, or not far from combined EP with the direction set method of Hooke
being violated), s‘ ¼ s‘
s, su ¼ su
s and H is an ap- and Jeeves [36]. The hybrid implementation proposed in
proximation of the Hessian matrix of the Lagrangian [34] was found very successful on shape optimization
function test examples, while the method proposed in [35] was
Lðs; kÞ ¼ F ðsÞ þ khðsÞ ð15Þ applied to unconstrained mathematical test functions.
Myung et al. [37] considered a similar to Waagen
In Eq. (15) k are the Lagrange multipliers under the non- et al. approach, but they experimented with constrained
negativity restriction (k P 0) for the inequality con- mathematical test functions. Myung et al. combined a
straints. In order to construct the Jacobian and the floating-point EP technique, with a method-developed
Hessian matrices of the QP subproblem the derivatives by Maa and Shanblatt [38] applied to the best solution
of the objective and constraint functions are required. found by the EP technique. The second method iterates
These derivatives are computed during the sensitivity until the system defined by the combination of the ob-
analysis phase. jective function, the constraint functions and the design
There are two ways to solve this QP subproblem, variables reach equilibrium.
either with a primal [30], or with a dual [31] formulation. A characteristic property of the SQP based optimiz-
In the present study a primal algorithm is employed ers is that they capture very fast the right path to the
580 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589

nearest optimum, irrespective of its nature a local or 7. Numerical tests

global optimum. However, after locating the area of this
optimum it might oscillate until all constraints are sat- The optimization of two space frame structures is
isfied since it is observed that even small constraint vi- selected to illustrate the efficiency of the presented
olations often slow down the convergence rate of the methodologies in structural design problems. In both
method. On the other hand EA proceed with slower test examples the modulus of elasticity E is taken 200
rate, due to their random search, but the absence of GPa and the yield stress ry is 250 MPa. The cross-sec-
strict mathematical rules, which govern the convergence tion of each member is assumed to be a I-shape and for
rate of the mathematical programming methods, make each member two design variables are considered as
EA less vulnerable to local optima and therefore it is shown in Fig. 2. The values of b and h are selected from
much more likely to converge towards the global opti- an integer design space, while t and w are given as fol-
mum in non-convex optimization methods. These two lows: f ¼ 0:06h þ 0:10ðb
10Þ, w ¼ 0:625f . Those two
facts give the motivation to combine EA with MP expressions make sure that the web thickness is less than
methodologies. Between the two EA examined in this b, the opposite of which would have been not accept-
study the genetic algorithms seems to be faster than able. The objective function of the problems is the
evolution strategies since they do not always operate on weight of the structure. The constraints are the member
the feasible region of the design space as evolution al- stresses and the inter-storey drifts. For rigid frames in
gorithms. However, they are most often found unable to rolled I-shapes, under allowable stress design require-
converge to feasible designs. ments specified by Eurocode 3 [39], the stress constraints
In order to benefit from the advantages of both are defined by the non-dimensional ratio q of interaction
methodologies (EA and SQP) a hybrid approach is formulas
proposed, which combines the two methods in an effort
fa fby fbz fa
to increase the robustness and the computational effi- q¼ þ þ 6 1:0 if 6 0:15 ð18Þ
Fa Fby Fbz Fa
ciency of the optimization procedure. The optimization
process is divided into two separate phases. During the and
first phase, the evolution algorithm optimizes the ob-
jective function. After the termination of this phase, the fa fy fz fa
q¼ þ by þ bz 6 1:0 if > 0:15 ð19Þ
second phase starts by applying the SQP to the best 0:60ry Fb Fb Fa
solution found during the first phase.
The procedure of EA is first used in order to locate where fa is the computed compressive axial stress, fby , fbz
the region where the global optimum lies, then the SQP are the computed bending stresses for y- and z-axis, re-
is activated in order to exploit its higher rate of con- spectively. Fa is the allowable compressive axial stress,
vergence in the neighborhood of the optimum. The
switch from EA to SQP is performed when EA reaches
the vicinity of an optimum that is considered a good
one. This approach appears to be more rational in the
general case when more complex and non-convex design
problems are to be solved with many local optima. In
the case of GA-SQP the final design achieved by GA can
be tolerated to be infeasible since the SQP will eventually
locate a feasible design.
In the present study two hybrid methodologies are
examined which are based on the combination of GA
and ES with the SQP algorithm. The transition from one
algorithm to the other is performed when there is a small
difference between the best designs of two consecutive
generations

 
 fjþ1
fj 
 6e ð17Þ
 fj 

It was found in the test cases performed in this study and

in shape optimization problems [34] that a reasonable
value for e could be taken in the vicinity of 10%. Fig. 2. I-shaped cross-section design variables.
 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589 581

Fby , Fbz are the allowable bending stresses for y- and z- Table 2
axis, respectively, and ry is the yield stress of the steel. Test example 1––performance of the two selection schemes
The allowable inter-storey drift is limited to 1.5% of the (l þ k)-ES and (l, k)-ES
height of each storey. One load case is considered in all Optimizer Weight (kN) FE analyses Time (s)
examples. (5 þ 5)-ES 675 416 177
The termination criterion adopted for both GA and (5,5)-ES 677 430 183
ES is the same for comparison reasons. If no improve- (10 þ 10)-ES 668 847 363
ment of the best value of the objective function has oc- (10,10)-ES 661 834 357
curred in the last six generations, then the optimization
procedure is terminated.
Example 1: The first example is a six-storey space performance of the methods for handling the constraints
frame, first studied by Orbinson et al. [40], with 63 ele- is measured with two parameters: the weight achieved
ments and 180 degrees of freedom. The beams length is and the average level of violation of the constraint
L1 ¼ 7:32 m and the columns height is L2 ¼ 3:66 m. The functions. The average percentage violation of the con-
loads consisting of 17 kPa gravity load on all floor levels straint functions is abbreviated to Avg. Const. Violation
and a lateral load of 100 kN applied at each node in the (%). Fig. 4 depicts the performance of GA with the
front elevation in the z direction. The element members method of static penalties for handling the constraints.
are divided into five groups shown in Fig. 3 and the total It can be seen that the performance of the method is very
number of design variables is 10. The initial design used sensitive to the value of the penalty parameter, while
in this example was chosen away from the optimum there is not a rule of thumb on how to choose this single
corresponding to the weight of 2846 kN for every test. penalty parameter. If it is chosen too small the search
Table 2 shows the performance of the two types of the will converge to an infeasible solution otherwise if it is
basic multi-membered ES, namely (l þ k)-ES and (l; k)- chosen too large a feasible solution may be located but it
ES for l ¼ k ¼ 5 and l ¼ k ¼ 10. The best design would be far from the global optimum. Fig. 5 presents
achieved by these tests is used as the basis for compar- the performance of GA with the method of dynamic
ison in subsequent applications of the optimization penalties for handling the constraints. It appears that the
methods examined. constraint violation percentage in this case is equal to
In Figs. 4–7 various techniques for handling the zero for all cases considered, since for high generation
constraints of the genetic algorithms are presented. The number the ðc  gÞa component of the penalty term takes

Fig. 3. Six storey space frame.

582 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589

Fig. 6. Test example 1––performance of S-GA with ph , p‘

scheme.
Fig. 4. Test example 1––performance of GA with static pen-
alties.

Fig. 7. Test example 1––performance of S-GA with p‘ , ph

scheme.

Fig. 5. Test example 1––performance of GA with dynamic the system has little chances to escape from local op-
penalties (a ¼ 2, zero constr. violation).
tima.
Fig. 6 presents the performance of the segregated GA
large values, which makes even the slightly violated de- (S-GA) method for handling the constraints for different
signs not to be selected in subsequent generations. Thus, values of penalty parameters ph , p‘ . The results indicate
 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589 583

that the optimum designs achieved are much closer to

the global one, compared to the other constraint han-
dling techniques, but are still in many cases local optima.
A general comment that can be deduced from Figs. 4–6
for micro GA is that l-GA show premature convergence
to worse local optima compared to the corresponding
basic GA. Fig. 7 depicts the performance of the S-GA
method when the parents are evaluated using the p‘
penalty parameter and the offsprings using the ph .
Tables 3 and 4 contain the results of the Augmented
Lagrangian GA method (AL-GA). The constraint vio-
lation percentage is equal to zero for all tests considered.
Table 3 depicts the results of the method with the ter-
mination criteria suggested by Adeli and Cheng in Ref.
[15], while Table 4 includes the results obtained for the
termination criteria used for the rest of the methods. For
the adopted termination criterion we have examined
four different cases, according to the allowable number
of generations with no improvement of the objective
function. It can be seen that both termination criteria
converge to the same result which appears to be a local
minima compared to the minimum achieved by the ES Fig. 8. Test example 1––performance of the C-ES.
as shown in Table 2. It can also be seen that the Aug-
mented Lagrangian GA method is not affected by the
value of normalization parameter Lf .
Table 3 Figs. 8 and 9 show the performance of the two
Test example 1––performance of the AL-GA
modified versions of the evolution strategies, namely the
Lf Weight (kN) FE analyses Time (s) contemporary ES (C-ES) and the adaptive ES (A-ES),
100 1010 145 62 respectively. For this test example the reduction pa-
500 1010 145 62 rameter b used in the A-ES method is equal to 0.1, while
1000 1010 145 62 the parameters ai and af denote the initial and final level
Termination criterion of Ref. [15].

Table 4
Test example 1––performance of the AL-GA with the termi-
nation criterion adopted in this work
Lf Weight (kN) FE analyses Time (s)
Case a: 6l=k
100 1060 110 48
500 1060 110 48
1000 1060 110 48
Case b: 12l=k
100 1042 140 60
500 1042 140 60
1000 1042 140 60
Case c: 18l=k
100 1010 145 62
500 1010 145 62
1000 1010 145 62
Case d: 24l=k
100 1010 145 62
500 1010 145 62
1000 1010 145 62
Fig. 9. Test example 1––performance of the A-ES (b ¼ 0:1).
584 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589

of violation of the constraints. The initial parameter ai Comparing the performance of the two sensitivity anal-
for the normalized constraint functions (Eqs. (18) and ysis methods it can be seen that the ESA sensitivity
(19)) is taken between 1.2 and 1.6, which correspond to a analysis method is faster than GFD method. For this test
percentage violation of the constraints between 20% and example the ESA and GFD sensitivity analysis methods
60%. In the case of Fig. 8 all designs achieved are fea- are used to compute the sensitivities and magnitude of
sible, while in the case of Fig. 9 the designs achieved by perturbation is equal to 10
5 .
the optimizers 2 and 5 are not feasible. The results of Example 2: The second example is the 20-storey
Figs. 8 and 9 indicate that the two new versions of ES space frame, studied by Papadrakakis and Papadopou-
manage to converge to better designs than the basic ES los [41], with 1020 members and 2400 degrees of free-
for a number of different parameters used at a marginal dom. The loads considered here are uniform vertical
increase of computational effort. The C-ES method forces applied at joints equivalent to uniform load of 4.8
appears to be slightly more robust than the adaptive kPa and horizontal forces equivalent to uniform forces
one, which may converge to infeasible designs when the of 1.0 kPa on the largest surface. The element members
parameter af is not equal to one. The A-ES, however, are divided into 11 groups shown in Fig. 10 and the total
manage to converge to the best optimum design of 620 number of design variables is 22. The initial design used
kN in 470 FE analyses for ead ¼ 0:005 and ai ¼ 1:6. in this example was chosen away from the optimum
The performance of the hybrid approaches is depicted corresponding to the weight of 42,248 kN for every test.
in Table 5. The results indicate that despite the fact that Table 6 shows the performance of the two types of
all GA methods, except of AL-GA, converge to infeasi- the multi-membered ES, namely (l þ k)-ES and (l; k)-
ble designs (the letter ‘v’ denotes constraint violation) the ES for l ¼ k ¼ 5 and l ¼ k ¼ 10. The (10 þ 10) version
SQP optimizer which follows manages to reach a feasible of ES manages to converge to the best design, which is
one at a competitive time. The best design found by A- used as reference design. In Figs. 11 and 12 various
ES-SQP and AL-GA-SQP optimizer is 20% better than techniques for handling the constraints by the GA are
the design achieved by SQP requiring 35% and 25% less presented. The average percentage violation of the
computational effort, respectively. We also examined two constraint functions is abbreviated as previously to Avg.
EA hybrid methods by combining AL-GA with ES and Const. Violation (%). It can be seen that all feasible so-
vice versa. Both perform well in terms of the design lutions achieved appear to be local optima although for
achieved and the required computing time. Furthermore, this example the GA with dynamic penalties showed a
the modified version of the lGA, which excludes all the rather more robust behaviour.
infeasible designs gave the same solution as the (l; k)-ES Tables 7 and 8 contain the results of the Augmented
with 25% less computational effort, while the modified Lagrangian GA method (AL-GA). The constraint vio-
micro GA (m lGA) achieves better performance in terms lation percentage is equal to zero for all tests considered.
of computing time compared to the (l þ k)-ES method. Table 7 depicts the results of the method with the ter-

Table 5
Test example 1––combination GA-SQP and ES-SQP
Optimizer Initial design for Final design Sensitivity FE analyses Time (s)
second optimizer (kN) analysis
EA SQP EA SQP Total
SQP – 795 GFD – 272 – 340 340
SQP – 795 ESA – 271 – 204 204
S-GA-SQP 708v 672 GFD 110 86 47 107 154
S-GA-SQP 708v 672 ESA 110 86 47 66 113
S-GA-SQP 635v 675 GFD 120 155 51 191 242
S-GA-SQP 635v 675 ESA 120 156 51 113 164
AL-GA-SQP 1010 663 GFD 145 121 62 155 217
AL-GA-SQP 1010 663 ESA 145 121 62 95 157
C-ES-SQP 912 681 GFD 212 114 90 142 232
C-ES-SQP 912 681 ESA 212 114 90 88 178
A-ES-SQP 829 663 GFD 136 96 58 120 178
A-ES-SQP 829 663 ESA 136 96 58 74 132
AL-GA-ES 1010 663 – 145 þ 207 – 151 – 151
ES-AL-GA 829 675 – 136 þ 257 – 177 – 177
m lGA – 677 – 319 – 154 – 154
 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589 585

Fig. 10. Twenty storey space frame.

Table 6 of generations with no improvement of the objective

Test example 2––performance of the two selection schemes function value. It can be seen that both termination
(l þ k)-ES and (l; k)-ES criteria converge to the same result which appears to be
Optimizer Weight (kN) FE analyses Time (s) a local minimal compared to the minimum achieved by
(5 þ 5)-ES 6028 240 3708 the ES as shown in Table 6. It can also be seen, as in the
(5,5)-ES 6085 452 6984 previous example, that the Augmented Lagrangian GA
(10 þ 10)-ES 5819 422 6519 method is not affected by the value of normalization
(10,10)-ES 5877 727 11,230 parameter Lf .
Figs. 13 and 14 depict the performance of the two
versions of the evolution strategies, namely the con-
mination criteria suggested by Adeli and Cheng in Ref. temporary ES (C-ES) and the adaptive ES (A-ES), re-
[15], while Table 8 includes the results obtained for the spectively, all designs depicted in those two figures are
termination criteria used for the rest of the methods. For feasible. A comparison of the results of Table 6 and
the adopted termination criterion we have examined those depicted in Figs. 13 and 14 indicates that the two
four different cases according to the allowable number new methodologies outperform in most cases the basic
586 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589

Table 8
Test example 1––performance of the AL-GA with the termi-
nation criterion adopted in this work
Lf Weight (kN) FE analyses Time (s)
Case a: 6l=k
100 8073 120 1917
500 8073 120 1917
1000 8073 120 1917
Case b: 12l=k
100 7397 140 2231
500 7397 140 2231
1000 7397 140 2231
Case c: 18l=k
100 7015 195 3097
500 7015 195 3097
1000 7015 195 3097
Case d: 24l=k
100 7015 195 3097
500 7015 195 3097
1000 7015 195 3097

Fig. 11. Test example 2––optimizer 1–3: GA with static pen-

alties; optimizer (4–7) GA and dynamic penalties (a ¼ b ¼ 2).

Fig. 12. Test example 2––performance of S-GA with p‘ , ph

scheme.

Table 7
Test example 1––performance of the AL-GA Fig. 13. Test example 2––performance of the C-ES.
Lf Weight (kN) FE analyses Time (s)
100 7015 195 3097 version of evolution strategies in terms of the achieved
500 7015 195 3097 optimum weight at the expense of more computing time.
1000 7015 195 3097 The performance of the hybrid approaches is de-
Termination criterion of Ref. [15]. picted in Tables 9–11. Three initial designs are consid-
 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589 587

this test example the ESA and GFD sensitivity analysis

methods are used to compute the sensitivities and
magnitude of perturbation is equal to 10
5 .

8. Conclusions

In this work two versions of evolution strategies as

well as various methods for handling the constraints in
GA have been discussed and compared in structural siz-
ing optimization problems with the standard mathemat-
ical programming approach of SQP. The techniques for
handling the constraints in GA are based on the use of
penalty functions, which transform the constraint opti-
mization problem into an unconstraint one. Techniques
for handling the constraints based on static and dynamic
penalties as well as the micro GA (lGA) and segregated
GA (S-GA) are examined and tested. It appears that the
efficiency of GA in structural sizing optimization prob-
lems is sensitive to the values of the characteristic pa-
rameters for handling the constraints. The computational
effort that is required by GA is less than the corresponding
Fig. 14. Test example 2––performance of the A-ES (b ¼ 0:1). effort by ES but they are hindered by premature conver-
gence either to non-optimal or to infeasible designs. The
contemporary ES (C-ES) and the adaptive ES (A-ES), did
ered, one close, one medium and one away from the manage to improve the final design achieved by the basic
optimum. It can be seen that the computing time spent ES for a number of values of the characteristic parameters
by the optimizers is affected by the initial design, espe- at almost no increase of the computational effort.
cially in the case of the SQP approach. Furthermore, The proposed hybrid optimization approach proved
GA-SQP, ES-SQP and AL-GA-SQP optimizers manage to be a robust and efficient method for structural opti-
to converge to final designs with 10% less weight than mization. Both GA-SQP and ES-SQP manage to con-
the SQP optimizer with less computational effort. We verge to better designs than those achieved by ES or
have also examined a EA hybrid method by combining SQP alone at a reduced computational effort compared
AL-GA with ES, which perform well in terms of the to the SQP procedure. The combination of GA with
design achieved and the required computing time. For SQP is particularly promising in bad initial designs due

Table 9
Test example 2––hybrid methods
Optimizer Initial design for Final de- Sensitivity Time (s) Time (s)
second optimizer sign (kN) analysis
EA SQP EA SQP Total
SQP – 6427 GFD – 435 – 21,932 21,932
SQP – 6427 ESA – 438 – 13,655 13,655
S-GA-SQP 7928v 5764 GFD 65 116 1100 5166 6266
S-GA-SQP 7928v 5764 ESA 65 117 1100 3608 4708
AL-GA-SQP 7015 6427 GFD 195 99 3301 4415 7716
AL-GA-SQP 7015 6427 ESA 195 99 3301 3077 6378
C-ES-SQP 7132 5834 GFD 193 152 3266 6770 10,036
C-ES-SQP 7132 5834 ESA 193 152 3266 4687 7953
A-ES-SQP 6531 5713 GFD 107 111 1811 4943 6754
A-ES-SQP 6531 5713 ESA 107 110 1811 3392 5202
AL-GA-ES 7015 5819 – 195 þ 65 – 4401 – 4401
m lGA – 5772 – 395 – 6117 – 6117
Bad initial design.
588 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589

Table 10
Test example 2––hybrid methods
Optimizer Initial design for Final de- Sensitivity Time (s) Time (s)
second optimizer sign (kN) analysis
EA SQP EA SQP Total
SQP – 6119 GFD – 385 – 17,519 17,519
SQP – 6119 ESA – 387 – 12,097 12,097
S-GA-SQP 7669 5695 GFD 40 169 681 7391 8071
S-GA-SQP 7669 5695 ESA 40 169 681 5291 5972
AL-GA-SQP 6816 5695 GFD 155 98 2639 4361 7000
AL-GA-SQP 6816 5695 ESA 155 98 2639 3057 5696
C-ES-SQP 6835 5764 GFD 171 111 2901 4943 7844
C-ES-SQP 6835 5764 ESA 171 111 2901 3469 6370
A-ES-SQP 6319 5764 GFD 91 67 1552 2983 4535
A-ES-SQP 6319 5764 ESA 91 67 1552 2089 3641
AL-GA-ES 6816 5430 – 155 þ 73 – 3866 – 3866
m lGA – 5472 – 339 – 5243 – 5243
Medium initial design.

Table 11
Test example 2––hybrid methods
Optimizer Initial design for Final de- Sensitivity Time (s) Time (s)
second optimizer sign (kN) analysis
EA SQP EA SQP Total
SQP – 6119 GFD – 259 – 11,508 11,508
SQP – 6119 ESA – 259 – 8049 8049
S-GA-SQP 7669 5695 GFD 35 169 590 7391 7981
S-GA-SQP 7669 5695 ESA 35 169 590 5291 5881
AL-GA-SQP 6816 5695 GFD 115 98 1934 4361 6295
AL-GA-SQP 6816 5695 ESA 115 98 1934 3057 4991
C-ES-SQP 6835 5764 GFD 158 111 2685 4943 7628
C-ES-SQP 6835 5764 ESA 158 111 2685 3469 6154
A-ES-SQP 6275 5764 GFD 69 67 1198 2983 4181
A-ES-SQP 6275 5764 ESA 69 67 1198 2089 3287
AL-GA-ES 6816 5430 – 115 þ 73 – 3161 – 3161
m lGA – 5472 – 339 – 5243 – 5243
Good initial design.

to the fast convergence of GA towards the neigh- [2] Fogel DB. Evolving artificial intelligence. PhD thesis,
borhood of the optimum and the property of SQP University of California, San Diego, 1992.
to compute quickly the nearest optimum once in the [3] Goldberg DE. Genetic algorithms in search, optimization
neighborhood of the solution. However, the proposed and machine learning. Reading, Massachusetts: Addison-
Wesley Publishing Co., Inc.; 1989.
adaptive ES when coupled with SQP and the combina-
[4] Holland J. Adaptation in natural and artificial systems.
tion of the Augmented Lagrangian GA, as the first stage Ann Arbor, MI: University of Michigan Press; 1975.
optimizer followed by ES, proved to be the more effi- [5] Rechenberg I. Evolution strategy: optimization of techni-
cient optimization algorithms. cal systems according to the principles of biological
evolution. Stuttgart: Frommann-Holzboog; 1973 (in Ger-
man).
References [6] Schwefel HP. Numerical optimization for computer mod-
els. Chichester, UK: Wiley & Sons; 1981.
[1] Fogel LJ, Owens AJ, Walsh MJ. Artificial intelligence [7] Papadrakakis M, Lagaros ND, Thierauf G, Cai J.
through simulated evolution. New York: Wiley; 1966. Advanced solution methods in structural optimization
 N.D. Lagaros et al. / Computers and Structures 80 (2002) 571–589 589

based on evolution strategies. J Eng Comput [25] Bletzinger KU, Kimmich S, Ramm E. Efficient modelling
1998;15(1):12–34. in shape optimal design. Comput Syst Eng 1991;2(5/
[8] Adeli H, Cheng NT. Concurrent genetic algorithms for 6):483–95.
optimization of large structures. J Aerospace Eng ASCE [26] Hinton E, Sienz J. Aspects of adaptive finite element
1994;7(3):276–96. analysis and structural optimization. In: Topping BHV,
[9] Papadrakakis M, Tsompanakis Y, Hinton E, Sienz J. Papadrakakis M, editors. Advances in Structural Optimi-
Advanced solution methods in topology optimization and zation. Edinburgh: Civil–Comp Press; 1994. p. 1–26.
shape sensitivity analysis. J Eng Comput 1996;13(5):57–90. [27] Olhoff N, Rasmussen J, Lund E. Method of exact
[10] Barricelli NA. Numerical testing of evolution theories. numerical differentiation for error estimation in finite
ACT A Biotheoretica 1962;16:69–126. element based semi-analytical shape sensitivity analyses.
[11] Krishnakumar K. Micro genetic algorithms for stationary Special Report No. 10, Institute of Mechanical Engineer-
and non stationary function optimization. SPIE Proceed- ing, Aalborg University, Aalborg, DK, 1992.
ings Intelligent control and adaptive systems, vol. 1196, [28] Hinton E, Sienz J. Studies with a robust and reliable
1989. structural shape optimization tool. In: Topping BHV,
[12] Goldberg DE. Sizing populations for serial and parallel editor. Developments in computational techniques for
genetic algorithms. TCGA Report No. 88004, University structural engineering. Edinburgh: Civil-Comp Press;
of Alabama, 1988. 1995. p. 343–58.
[13] Joines J, Houck C. On the use of non-stationary penalty [29] Gill PE, Murray W, Wright WH. Practical optimization.
functions to solve non-linear constrained optimization London: Academic Press; 1981.
problems with GA. In: Michalewicz Z, Schaffer JD, [30] Gill PE, Murray W, Saunders MA, Wright MH. User’s
Schwefel H-P, Fogel DB, Kitano H, editors. Proceedings guide for NPSOL (Version 4.0): A Fortran Package for
of the First IEEE International Conference on Evolution- Non-linear Programming. Technical Report SOL 86-2,
ary Computation. IEEE Press; 1994. p. 579–84. Department of Operations Research, Stanford University,
[14] Michalewicz Z. Genetic algorithms, numerical optimiza- 1986.
tion and constraints. In: Eshelman LJ, editor. Proceedings [31] Fleury C. Dual methods for convex separable problems.
of the 6th International Conference on Genetic Algo- In: Rozvany GIN, editor. NATO/DFG ASI optimization
rithms. Los Altos, CA: Morgan Kaufmann; 1995. p. 151–8. of large structural systems. Berchtesgaden, Germany,
[15] Adeli H, Cheng NT. Integrated genetic algorithm for Dordrecht, Netherlands: #132;lands; 1993. p. 509–30.
optimization of space structures. J Aerospace Eng ASCE [32] NAG, Software manual. NAG Ltd, Oxford, UK, 1988.
1993;6(4):315–28. [33] Thanedar PB, Vanderplaats GN. Survey of discrete vari-
[16] Adeli H, Cheng NT. Augmented Lagrangian genetic able optimization for structural design. J Struct Eng ASCE
algorithm for structural optimization. J Aerospace Eng 1995;121(2):301–6.
ASCE 1994;7(1):104–18. [34] Papadrakakis M, Tsompanakis Y, Lagaros ND. Structural
[17] Belegundu AD, Arora JS. A computational study of shape optimization using evolution strategies. Eng Optim
transformation methods for optimal design. AIAA J 1999;31:515–40.
1984;22(4):535–42. [35] Waagen D, Diercks P, McDonnell J. The stochastic
[18] Le Riche RG, Knopf-Lenoir C, Haftka RT. A segregated direction set algorithm: a hybrid technique for finding
genetic algorithm for constrained structural optimization. function extrema. In: Fogel DB, Atmar W, editors.
In: Eshelman LJ, editor. Proceedings of the 6th Interna- Proceedings of the 1st Annual Conference on Evolutionary
tional Conference on Genetic Algorithms, 1995. p. 558–65. Programming. Evolutionary Programming Society; 1992.
[19] Cai J, Thierauf G. Discrete structural optimization using p. 35–42.
evolution strategies. In: Topping BHV, Khan AI, editors. [36] Hooke R, Jeeves TA. Direct search solution of numerical
Neural networks and combinatorial in civil and structural and statistical problems. J ACM 1961;8.
engineering. Edinburgh: Civil–Comp Ltd; 1993. p. 95–100. [37] Myung H, Kim J-H, Fogel D. Preliminary investigation
[20] Schwefel H-P, Rudolph G. Contemporary evolution strat- into a two-stage method of evolutionary optimization on
egies. In: Morgan F, Moreno A, Merelo JJ, Chacion P, constrained problems. In: McDonnell JR, Reynolds RG,
editors. Advances in artificial life, Proceedings of the Fogel DB, editors. Proceedings of the 4th Annual Confer-
Third European Conference on Artificial Life Granada, ence on Evolutionary Programming. Cambridge, MA:
Spain, 4–6 June. Berlin: Springer; 1995. p. 893–907. MIT Press; 1995. p. 449–63.
[21] Rudolph G. Personal communication, 1999. [38] Maa C, Shanblatt M. A two-phase optimization neural
[22] Back T, Hoffmeister F, Schwefel H-P. A survey of network. IEEE Trans Neural Networks 1992;3(6):
evolution strategies. In: Belew RK, Booker LB, editors. 1003–9.
Proceedings of the 4th International Conference on [39] Eurocode 3, Design of steel structures, Part1.1: General
Genetic Algorithms. Los Altos, CA: Morgan Kaufmann; rules for buildings, CEN, ENV 1993-1-1/1992.
1991. p. 2–9. [40] Orbinson JG, McGuire W, Abel JF. Yield surface appli-
[23] Schmit LA. Structural design by systematic synthesis. cations in non-linear steel frames analysis. Comput Meth
Proceedings of the second ASCE Conference on Electronic Appl Mech Eng 1982;33:557–73.
Computations, Pittsburgh, 1960. p. 105–22. [41] Papadrakakis M, Papadopoulos V. A computationally
[24] Naylor TH, Balintfy JL. In: Computer simulation tech- efficient method for the limit elasto plastic analysis of space
niques. New York: John Wiley & Sons; 1966. p. 114. frames. Computat Mech 1995;16(2):132–41.