Вы находитесь на странице: 1из 11

Computers and Structures 84 (2006) 787–797

www.elsevier.com/locate/compstruc

Minimum cost design of a welded orthogonally stiffened cylindrical shell


K. Jármai a, J.A. Snyman b,*
, J. Farkas a

a
Faculty of Mechanical Engineering, University of Miskolc, H-3515 Miskolc, Hungary
b
Department of Mechanical and Aeronautical Engineering, University of Pretoria, South Africa

Received 23 March 2005; accepted 18 January 2006

Abstract

In this study the optimal design of a cylindrical orthogonally stiffened shell member of an offshore fixed platform truss, loaded by axial
compression and external pressure, is investigated. Ring stiffeners of welded box section and stringers of halved rolled I-section are used.
The design variables considered in the optimization are the shell thickness as well as the dimensions and numbers of stiffeners. The design
constraints relate to the shell, panel ring and panel stringer buckling, as well as manufacturing limitations. The cost function includes the
cost of material, forming of plate elements into cylindrical shape, welding and painting. In the optimization a number of relatively new
mathematical optimization methods (leap-frog – LFOPC, Dynamic-Q, ETOPC, and particle swarm – PSO) are used, in order to ensure
confidence that the finally computed optimum design is accurately determined, and indeed corresponds to a global minimum. The con-
tinuous optimization procedures are adapted to allow for discrete values of the design variables to be used in the final manufacturing of
the truss member. A comparison of the computed optimum costs of the stiffened and un-stiffened assemblies, shows that significant cost
savings can be achieved by orthogonal stiffening, since the latter allows for considerable reduction of the shell thickness, which results in
large material and manufacturing cost savings.
 2006 Elsevier Ltd. All rights reserved.

Keywords: Welded stiffened shells; Stiffened cylindrical shell buckling; Manufacturing cost calculation; Structural optimization; Mathematical optimiz-
ation methods

1. Introduction computer therefore requires, not only refined and realistic


mathematical models of the structure, but also sufficiently
Structural optimization is the search for designs that ful- accurate, reliable and economic numerical methods and
fill certain prescribed requirements in an optimal manner. procedures to drive the structural design to an acceptable
The main requirements for modern engineering structures optimum configuration.
are that they should be safe with regard to load-carrying The particular problem of interest here is the minimum
capacity, fit for production, whilst at the same time being cost design of a welded stiffened steel shell. Cylindrical
economic. In an optimum design procedure, structural shells are used in many engineering load-carrying struc-
assemblies are sought which fulfill the prescribed design tures, for example, as columns, towers for wind turbines
and manufacturing constraints, and at the same time min- or water tanks, and in offshore and submarine structures.
imize an associated cost function. The cost function and They are also used in belt-conveyor bridges, and in pres-
the functions specifying the constraints may depend on sure vessels. The main structural characteristics of stiffened
many variables, and the behaviour of the functions may cylindrical shells are that they are subjected to axial com-
be highly non-linear. The solution of such problems by pression loads, bending, and external or internal pressures.
The shells therefore require orthogonal stiffening via ring
or stringer stiffeners. Typically the stiffener shapes are flat,
*
Corresponding author. rolled T- or L-profiles, cold-formed L-profiles, welded
E-mail address: jan.snyman@eng.up.ac.za (J.A. Snyman). T- or box profiles, or trapezoidal.

0045-7949/$ - see front matter  2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruc.2006.01.002
788 K. Jármai et al. / Computers and Structures 84 (2006) 787–797

In recent years a number of new optimization algo- A further important problem is to evaluate the economy
rithms have been developed, such as the leap-frog algo- of stiffening, i.e. how to achieve cost savings by using thin-
rithm [1–3], the successive approximation Dynamic-Q ner stiffened plates or shells instead of thicker un-stiffened
method [4,5], the ETOPC gradient-only method [6], and ones. A stiffened structure is economic, if the thickness
the Particle Swarm Optimization (PSO) approach [7]. In can be decreased in such a manner that, the cost saving
the present study these methods are all applied to the min- caused by this decrease is higher than the additional cost
imum cost design of the above mentioned welded stiffened of the stiffening material and the extra welding. Since in
cylindrical steel shell assembly. Since the authors have been cylindrical shells the cost of the forming of the plate ele-
intimately involved in the development of all of these algo- ments into cylindrical shapes, and that of the welding of
rithms, the associated optimization computer programs the shell elements, increases with increase in shell thickness,
were readily available and could all be implemented with the decrease of thickness of the shells can result in signifi-
ease and little additional effort. The successful application cant cost savings.
of all these methods, including the particle swarm global In previous studies it was found that the ring-stiffeners
method, ensured confidence in the accuracy of the com- are indeed economic to use for external pressure, since a
puted minima, and indicated with a very high probability cylindrical shell is very sensitive to the buckling by external
that the global optimum was indeed found. It also allows pressure. On the other hand their buckling strength against
for a comparison of the relative accuracy and economy axial compression or bending is high, so the use of stringers
of the different methods for structural problems. The con- in these cases is uneconomic, unless a lateral displacement
tinuous optimization procedures are adapted, to allow for constraint of the whole shell is active and the halved rolled
the discrete nature of the design variables to be used in I-section stringers are welded outside the shell.
the actual final manufacturing of the truss member. In the present study the combined load of axial com-
The outline of the remainder of this paper is as follows. In pression and external pressure is considered, as it acts on
the next section more detailed information is given regard- parts of the columns of a truss tower of a fixed offshore
ing the particular cylindrical shell truss member design platform (see Fig. 1). The cylindrical-shell member that is
problem to be addressed in this study. In Section 3, for the orthogonally stiffened by using ring stiffeners of box
orthogonally stiffened cylindrical-shell member, the buck- cross-section and stringers of halved rolled I-section (see
ling and manufacturing constraints imposed on the design
are presented. In Section 4, the construction of the associ-
ated cost function, which includes material, manufacturing
and painting costs, is dealt with. Section 5 discusses the cor-
responding constraints and cost function for the un-stiffened
shell, subjected to the same loads as the orthogonally stiff-
ened one. In Section 6, on formal mathematical optimiza-
tion, a re-formulation of the optimization problem in
general mathematical notation is given, which allows for a
brief introduction to the new optimization algorithms to
be used in computing the optimum designs. In Section 7,
detailed presentation and discussion of the numerical opti-
mization results is given, and in Section 8, some final conclu-
sions, drawn from the study, are presented.

2. Cylindrical shell truss member design problem

This study follows on previous work by two of the


authors on welded steel shells. A method for calculating
radial displacements due to shrinkage of circumferential
welds has been developed [8]. Also, a cost calculation
method, mainly for welded structures, has been proposed
[9]. Using this method it is possible to evaluate the econ-
omy of structural assemblies, which allows for significant
cost savings in the design stage. In particular, minimum
cost designs have been computed for a ring-stiffened shell
subject to external pressure [10], a ring-stiffened shell
(belt-conveyor bridge) subject to bending [11], and a strin-
ger stiffened shell (column) loaded by axial compression
and bending, with a displacement constraint at the top of Fig. 1. A part of a fixed offshore structure, the main columns are stiffened
the column [12]. cylindrical shells.
K. Jármai et al. / Computers and Structures 84 (2006) 787–797 789

p
t

Rc hr
tr
R yG
NF NF x
G
yE
hr
tr yf
E

t
Lr Lr Le

Fig. 2. Stringer and ring stiffened cylindrical shell with compression and external pressure.

Le 3.1. Shell (curved panel) buckling


t

The equivalent stress re must, according to the Det


Norske Veritas rules [13] already referred to in the previous
section, satisfy the constraint
G
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
zG

fy1
s1 re ¼ r2a  ra rp þ r2p 6 qffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð1Þ
1 þ k4s
tw

b
f
where fy1 ¼ 1:1y , and the stress due to axial compression is
h1
/2

given by
s1=min. 300
NF As 2Rp
ra ¼ ; where te ¼ t þ and s ¼ . ð2Þ
2Rpte s ns
tf

Fig. 3. Cross-section of the stiffened shell. Here fy is the yield stress, NF is the factored compression
force, R is the shell radius, t is the shell thickness, As is
Fig. 2), is to be optimized with respect to a cost function, the cross-sectional area of a stringer, ns is the number of
which includes material, manufacturing and painting costs. longitudinal stiffeners (stringers). Also appearing in the
In order to demonstrate the economy of the stiffening, above constraint definition is the stress due to external
both stiffened and un-stiffened assemblies are optimized pressure
and their costs are compared to each other (Section 7,
pF R AR
Tables 1a, 1b, 2–7). The design rules used here are formu- rp ¼ ; a¼ ;
lated according to the design rules of Det Norske Veritas tð1 þ aÞ Le0 t
[13]. The cost function is formulated in correspondence pffiffiffiffiffi L
to the manufacturing sequence. The specified constraints Le0 ¼ minðLr ; Ler ¼ 1:56 RtÞ and Lr ¼ . ð3Þ
nr  1
relate to shell buckling, panel stringer and panel ring buck-
ling, as well as to manufacturing limitations. Here pF is the factored external pressure intensity, Ar is the
In the case of the orthogonally stiffened assemblies, the cross-sectional area of a ring-stiffener, L is the shell length,
particular design variable to be considered are the shell nr is the number of ring-stiffeners, Lr is the distance
thickness (t), the number of longitudinal stiffeners (string- between rings. Also used in the definition above is k2s
ers) (ns), the number of ring-stiffeners (nr), the box height defined by
(hr) and the stringer stiffener height (h = h1 + 2tf) (see Figs.  
2 and 3). fy1 ra rp
k2s ¼ þ ; where
re rEas rEps
3. Constraints for the orthogonally stiffened cylindrical p2 E  t 2
shell member rEas ¼ C as ; ð4Þ
12ð1  m2 Þ s

Some of the quantities that appear in the definitions of E and m are the Young modulus and the Poisson ratio,
the constraints are given below, are also indicated in Fig. 2. respectively
790 K. Jármai et al. / Computers and Structures 84 (2006) 787–797
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2
q n s2 pffiffiffiffiffiffiffiffiffiffiffiffi2 Furthermore
C as ¼ was 1 þ as as ; was ¼ 4; Z as ¼ 1m ; vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u !2
was Rt  2 u
p2 E t t qpp npp
rEpp ¼ C pp ; C pp ¼ wpp 1 þ ; ð17Þ
nas ¼ 0:702Z as ; ð5Þ 10:92 Lr wpp
pffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffi
 0:5 npp ¼ 1:04 Z pp ; Z pp ¼ Z ap ; qpp ¼ 0:6; and wpp ¼ 2ð1 þ 1 þ cs Þ.
R p2 E  t 2
qas ¼ 0:5 1 þ ; rEps ¼ C ps ; ð6Þ ð18Þ
150t 10:92 s
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!2
u
u qps nps 3.3. Panel ring buckling
t
C ps ¼ wps 1 þ ; qps ¼ 0:6; ð7Þ
wps
" The ring-stiffeners are welded square box sections (see
 2 # 2
s pffiffiffiffiffiffiffi s Fig. 2), constructed from three plate elements of width br
and nps ¼ 1:04 Z ps ; Z ps ¼ Z as and wps ¼ 1 þ . and thickness tr. The relationship that must be satisfied
Lr Lr
between the width and thickness, is prescribed by the Euro-
ð8Þ code 3 [14] rule for compression plates against buckling,
and is given by
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3.2. Panel stiffener (stringer) buckling tr P dr hr ; 1=dr ¼ 42e; e ¼ 235=fy ;

In this case the equivalent stress re, again according to f y ¼ 355; dr ¼ 1=34. ð19Þ
the Det Norske Veritas rules [13], must satisfy the Assuming the buckling constraint, given by expression
constraint (19), to be active, it becomes an equality constraint and
fy1 the cross-sectional area of a ring-stiffener is then given by
re 6 qffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð9Þ
1 þ k4P AR ¼ 3hr tr ¼ 3dr h2r . ð20Þ
where The required cross-sectional area of ring-stiffener, exclud-
  2
 2 ing the effective shell width, is
fy1 ra rp pE t
k2P ¼ þ ; rEap ¼ C ap ;  
re rEap rEpp 10:92 Lr 2 L2r
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ARreq ¼ þ 0:06 Lr t; where Z ¼ 0:9539. ð21Þ
u !2 Z2 Rt
u qap nap
t
C ap ¼ wap 1 þ ; ð10Þ
wap Thus the constraint is
ARreq 6 AR . ð22Þ
L2r
qap ¼ 0:5; nap ¼ 0:702Z ap ; Z ap ¼ 0:9539; ð11Þ The effective pffiffiffiffiflange width is given by Le ¼
Rt ffi
sffiffiffiffi minðLr ; 2  1:56 RtÞ.
1 þ cs I Sef E The distance of the centroid of ring (point E in Fig. 2),
wap ¼ ; cs ¼ 10:92 and sE ¼ 1:9t . ð12Þ
1 þ sAest st3 fy including the effective shell flange width is given by

With respect to further quantities to be computed below, Le tðhr þ t=2Þ þ dr h3r


yE ¼ . ð23Þ
the following rule is applied: 3dr h2r þ Le t
if sE 6 s; se ¼ sE ; and if sE P s; se ¼ s. ð13Þ
The moment of inertia of the ring about x-axis is
ISef is the moment of inertia of stiffener including effective  2
shell plating se. In the case of a stiffener of halved UB sec- dr h4r hr
IR ¼ þ 2dr h2r  y E þ dr h2r y 2E
tion (flange width b, flange thickness tf, web height h1/2, 6 2
and web thickness tw as shown in Fig. 3), the distance of  t 2
the centre of gravity is þ Le t h r þ  y E . ð24Þ
2
    The required moment of inertia of a ring is given by
h1
tw h41 þ 2t þ btf h1 þtþt f

zG ¼ 2 2
ð14Þ
se t þ btf þ h1 tw =2 I Rreq ¼ I a þ I p ; ð25Þ
and ISef is given by where Ia is the required moment of inertia for axial com-
 3  2  2 pression, and Ip that for external pressure, respectively
h1 tw h1 tw h1 t h1 þ t þ tf
I Sef ¼ se tz2G þ þ þ  zG þ btf  zG given by
2 12 2 4 2 2
 
ð15Þ ra t 1 þ Asts R40
and As ¼ btf þ h1 tw =2. ð16Þ Ia ¼ ; R0 ¼ R  ðhr  y E Þ ð26Þ
500ELR
K. Jármai et al. / Computers and Structures 84 (2006) 787–797 791

and The volume of a shell segment is


2 3 V 1 ¼ 2RptLs ð32Þ
p RR2 Lr 3Ey d0
I p ¼ F 0 42 þ  E 5 with d0 ¼ 0:005R. ð27Þ and the volume of a ring-stiffener is given by
3E R 2 fy
 rp0 2
V R ¼ 2pdr h2r ðR  hr Þ þ 4pdr h2r ðR  hr =2Þ. ð33Þ
The constraint is The material cost is expressed as

I Rreq 6 I R . ð28Þ K M ¼ k M1 5qV 1 þ k M1 qnr V R þ k M2 qns As L; ð34Þ


where kM1 and kM2 are the respective cost factors for plates
3.4. Manufacturing limitations and rolled I-sections and q denotes the specific density of
steel.
In order to ensure that the welding of the webs of the The manufacturing cost components are as follows:
halved rolled I-section stringers into the shell is possible, K F0 ¼ 5k F Hel ;
the minimum distance between the stringer flanges should
l ¼ 6:8582513  4:527217t0:5 þ 0:009541996ð2RÞ0:5 ; ð35Þ
satisfy the following condition: pffiffiffiffiffiffiffiffiffiffiffi
K F1 ¼ 5k F ðH jqV 1 þ 1:3  0:1520  103 t1:9358  2Ls Þ;
2ðR  hr =2Þp
 b P 300 mm ð29Þ H ¼ 2; j ¼ 2; ð36Þ
ns pffiffiffiffiffiffiffiffiffiffiffiffiffi 3 1:9358
K F2 ¼ k F ðH 25qV 1 þ 1:3  0:1520  10 t  4  2RpÞ;
or ð37Þ
pffiffiffiffiffiffiffiffiffiffiffiffi
2ðR  hr =2Þp K F3 ¼ nr k F ½3 3qV R þ 1:3  0:3394  103 a2wr 4pðR  hr Þ; ð38Þ
ns 6 . ð30Þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
b þ 300 K F4 ¼ k F ½3 ðnr þ 1Þqð5V 1 þ nr V R Þ

Another limitation is related to the minimum value of fillet þ 1:3  0:3394  103 a2wr nr 4Rp; ð39Þ
welds connecting the stringer webs to the shell, and that for and
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
connecting the plate elements of ring-stiffeners. They are K F5 ¼ k F ½3 ðnr þ ns þ 1Þqð5V 1 þ nr V R þ ns As LÞ
respectively given by aws = 0.4tw, aws.min = 3 mm, and þ 1:3  0:3394  103 a2ws ns 2L. ð40Þ
awr = 0.4tr, awr.min = 3 mm.
Finally, the cost of painting is given by
4. Cost function for the orthogonally stiffened cylindrical K P ¼ k P ½2RpL þ 2RpðL  nr hr Þ þ 2nr phr ðR  hr Þ
shell member  
hr
þ 4pnr hr R  þ ns Lðh1 þ 2bÞ. ð41Þ
2
The cost function (K) includes material (KM) and man-
ufacturing costs (KFi), as well as the cost of painting (KP)
the final assembly, i.e. the total cost is given by 5. Constraint and cost function for the un-stiffened shell
X
K ¼ KM þ K Fi þ K P . ð31Þ 5.1. Constraint on shell buckling
i

The manufacturing sequence, determining the associ- The equivalent stress re must satisfy the constraint
ated total manufacturing cost, is as follows: qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi fy1
re ¼ r2a  ra rp þ r2p 6 pffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð42Þ
1 þ k4
(1) Form plate elements of Ls = 3 m length, into cylindri-  
fy fy1 ra rp NF p R
cal shapes (KF0). where f y1 ¼ ; k2 ¼ þ ; ra ¼ ; rp ¼ F ;
1:1 re rEa rEp 2Rpt t
(2) Weld shell segments of Ls = 3 m length, from 2
ð43Þ
curved plate elements, with 2 butt welds using sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2ffi
GMAW-C (Gas Metal Arc Welding with CO2) (KF1). p2 E  t 2 qa na
rEa ¼ C a ; C a ¼ wa 1 þ ; wa ¼ 1; ð44Þ
(3) Weld whole un-stiffened shell of L = 15 m length, 12ð1  m2 Þ L wa
from 5 shell segments with 4 circumferential butt  0:5
L2 pffiffiffiffiffiffiffiffiffiffiffiffi2 R
welds, using GMAW-C (KF2). Za ¼ 1  m ; na ¼ 0702Z a ; qa ¼ 0:5 1 þ ; ð45Þ
Rt 150t
(4) Weld nr ring-stiffeners from 3 plate elements with 2 vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!2 ffi
u
fillet welds, each using SMAW (Shielded Metal Arc p 2 E  t 2 u
t qp np
rEp ¼ C p ; C p ¼ wp 1 þ ; qp ¼ 0:6; ð46Þ
Welding) (KF3). 10:92 L wp
(5) Weld nr ring-stiffeners into the whole shell with 2 nr and np ¼ 1:04Z p ; Z p ¼ Z a ; wp ¼ 4. ð47Þ
circumferential fillet welds, using SMAW (KF4).
(6) Weld ns stringers into the shell with 2 ns fillet welds, The reader should note that expression (42) for the un-stiff-
using SMAW (KF5). ened case is different from Eq. (1) for the stiffened shell,
792 K. Jármai et al. / Computers and Structures 84 (2006) 787–797

since the definitions of k used in (42) and ks used in (1) gj(x) 6 0, j = 1, . . . , m, where in general m denotes the num-
differ. These differences and others such as te in (2) but t ber, here 5, of inequality constraints. Side constraints, i.e.
in (43), are carried over in the defining expressions that fol- upper and lower limits on the design variables, may also
low on (1) and (42), respectively. be specified. The general optimization problem to be con-
sidered here is therefore
5.2. Cost function for the un-stiffened shell T
minimize f ðxÞ; x ¼ ½x1 ; x2 ; . . . xn  2 Rn ; ð54Þ
w.r.t. x
The cost function includes cost of material, manufactur- subject to the inequality and equality constraints:
ing and painting
X gj ðxÞ 6 0; j ¼ 1; 2; . . . ; m; ð55Þ
K ¼ KM þ K Fi þ K P . ð48Þ
i hj ðxÞ ¼ 0; j ¼ 1; 2; . . . ; r;
The manufacturing sequence in this case is as follows:
and side constraints:
(1) Form plate elements of Ls = 3 m length, into cylindri- xli 6 xi 6 xui ; i ¼ 1; 2; . . . ; n;
cal shapes (KF0).
(2) Weld shell segments of Ls = 3 m length from 2 curved where f(x), gj(x) and hj(x) are scalar functions of the real
plate elements, with 2 butt welds, using GMAW-C column vector x. For generality equality constraints,
(Gas Metal Arc Welding with CO2) (KF1). hj(x) = 0, j = 1, 2, . . . , r are also specified, although they
(3) Weld whole un-stiffened shell of L = 15 m length, are not explicitly imposed in this study. The optimum solu-
from 5 shell segments with 4 circumferential butt tion is denoted by x* with associate optimum function
welds, using GMAW-C (KF2). value f(x*). Of course, the optimization of the un-stiffened
shell can also be stated in the general form (54).
The volume of a shell segment is The formulation of the optimization problem in the
V 1 ¼ 2RptLs . ð49Þ form (54), facilitates the discussion of the optimization
methods used in this study. In practice such problems are
The material cost is expressed as usually solved by means of numerical optimization algo-
K M ¼ k M1 5qV 1 . ð50Þ rithms. Four different algorithms are applied to the prob-
lem at hand. As stated in the Introduction, this is done in
where kM1 is the cost factor for plates. order to have confidence in the finally computed optimum
The manufacturing cost components for the un-stiffened design. It also allows for a comparison between the meth-
shell are the same as those given for the stiffened shell in ods with regard to accuracy and economy.
(35) and (37). For the sake of completeness they are The four algorithms used here are all relatively new.
restated here for the un-stiffened case They are the gradient-based methods LFOPC, Dynamic-
K F0 ¼ 5k F Hel ; Q, ETOPC and the evolutionary Particle Swarm Optimiza-
l ¼ 6:8582513  4:527217t0:5 þ 0:009541996ð2RÞ0:5 ; ð51Þ tion (PSO) algorithm. In solving the constrained problem
pffiffiffiffiffiffiffiffiffiffiffi
K F1 ¼ 5k F ðH jqV 1 þ 1:3  0:1520  103 t1:9358  2Ls Þ; (55), each of these methods performs the unconstrained
H ¼ 2; j ¼ 2; ð52Þ minimization of a so-called penalty function, which
pffiffiffiffiffiffiffiffiffiffiffiffiffi 3 1:9358 actually embodies the constraints. For example, if we only
K F2 ¼ k F ðH 25qV 1 þ 1:3  0:1520  10 t  4  2RpÞ;
consider inequality constraints in problem (55), the appro-
ð53Þ
priate penalty function is
kF = 1.0 $/min, kM1 = 1.0 $/kg. X m
The cost of painting is given by KP = 4RpLkP, F ðxÞ ¼ f ðxÞ þ bj g2j ðxÞ; ð56Þ
kP = 14.4 · 106 $/mm2. j¼1

where bj = 0 if gj(x) 6 0, and bj = l if gj(x) > 0 where l is a


6. Mathematical optimization very large number.

6.1. General formulation of design optimization problem 6.2. Optimization algorithms

In the case of the orthogonally stiffened cylindrical-shell The optimization algorithms used in this study are pre-
member the design variables are t, ns, nr, hr, and hs, and sented here. Since the relatively new gradient-based local
the cost function K, also called the objective function, optimization methods have all been well documented, only
clearly depends on these variables. In more general optimi- brief discussions of them are given here. A little more is
zation notation, the design vector is denoted by x = said about the global PSO algorithm. Although the latter
[x1, x2, . . . , xn]T = [t, ns, nr, hr, h]T, and the objective function method has been widely used over the past ten years, it is
by f(x) = K(x). The constraints, here (1), (9), (22), (28), surprising how poorly the essentials of the method are pre-
(30), may be written in the standard inequality form as sented in the general literature. It is therefore decided to
K. Jármai et al. / Computers and Structures 84 (2006) 787–797 793

include here, a short but more rigorously detailed presenta- 6.2.2. The Particle Swarm Optimization (PSO) algorithm
tion of the basic PSO algorithm. This may be of interest to Particle Swarm Optimization (PSO) techniques belong
the uninitiated, since it attempts to convey the essentials of to a relatively new class of evolutionary based search pro-
the method in a simple but precise pseudo-code, which can cedures that may be used to find the optimum solution x*
easily be implemented on a computer. of the general optimization problem. The original PSO
algorithm, proposed by Kennedy and Eberhardt in 1995
6.2.1. The gradient-based algorithms [7], was inspired by the modelling of the social behaviour
patterns of organisms that live and interact within large
(i) The LFOPC leap-frog optimization algorithm for groups. In particular, PSO incorporates swarming behav-
constrained optimization [3], uses a dynamic trajec- iours observed in flocks of birds, schools of fish, or swarms
tory technique for the unconstrained minimization of bees. A PSO algorithm is easy to implement in most pro-
[1,2] of the penalty function F(x) (see definition gramming languages, since the core of the program can be
(56)), in order to obtain the solution of the associated written in a few lines of code. It has been proven to be both
constrained optimization problem. The trajectory fast and effective, when applied to a diverse set of optimi-
seeks the optimum by simulating the motion of a par- zation problems. PSO algorithms are especially useful for
ticle in n-dimensional space, where F(x) represents the parameter optimization in continuous, multi-dimensional
potential energy of the particle. The generated tra- search spaces.
jectory is monitored and an intermittent damping In performing a search in the multi-dimensional space
procedure is applied to the particle to ensure its con- associated with the optimization problem of the form
vergence to a local minimum. An outstanding charac- (54), the PSO technique assigns direction vectors and veloc-
teristic of the basic method is that it uses only ities to each member (particle) of the swarm at their current
function gradient information, so that no explicit positions. Each particle then ‘‘moves’’ or ‘‘flies’’ through
function line searches are performed. It has been pro- the search space according to the particle’s assigned veloc-
ven to be extremely reliable and robust in solving ity vector, which may be influenced by the directions and
many practical engineering design problems [4,15,16]. velocities of other particles in its neighbourhood. These
(ii) The Dynamic-Q optimization method [5] solves con- localized interactions with neighbouring particles, propa-
strained optimization problems by applying the gate through the entire ‘‘swarm’’ of particles and results
LFOPC trajectory method to a sequence of very sim- in the swarm as a whole moving to regions of the space clo-
ple spherically quadratic sub-problems. The functions ser to the solution of problem (54). The extent to which a
used in the sub-problems are two-point approxima- particular particle influences other particles is determined
tions of the original objective and constraint func- by its so-called ‘‘fitness’’ along its trajectory of candidate
tions. In constructing the approximations, use is solution points. The ‘‘fitness’’ is a measure assigned to each
made of the value of the function and its gradient vec- potential solution, and it indicates how good a particular
tor at the current point, and its function value at the candidate solution is relative to all other solution points.
previous point. Because this method has been found Hence, an evolutionary idea of ‘‘survival of the fittest’’
to converge quickly after relatively few sub-problems (in the sense of Darwinian evolution) comes into play,
[17], it is very economic when applied to problems as well as a social behaviour component through a ‘‘follow
where the functions are extremely expensive to com- the local leader’’ effect and emergent pattern formation [9].
pute, i.e. where costly computer simulations are A more precise and detailed description of the particular
required in order to evaluate the functions. PSO algorithm used in this study now follows.
(iii) The ETOPC algorithm [6] is a new implementation of Basic PSO Algorithm
the conjugate gradient method (both the Fletcher–
Reeves and Polak–Ribiere versions) for solving (1) Given M, kmax, Nmax. Set (time) instant k = 0,
constrained problem. The essential novelty in this F bi ¼ F g ¼ F gbefore ¼ 1. Initialise a random popula-
implementation is the use of gradient-only line tion (swarm) of M particles (swarm members), by
searches, originally proposed by one of the authors assigning an initial random position x0i (candidate
[18] in 1985. This conjugate gradient algorithm that solution), as well as a random initial velocity v0i , to
uses gradient-only line searches, has been proven to each particle i, i = 1, 2, . . . , M. Then compute simulta-
be highly accurate and more reliable than standard neous trajectories, one for each particle, by perform-
implementations, when applied to the solution of con- ing the following steps.
strained problems via the sequential unconstrained (2) At instant k, compute the fitness of each individual
minimization technique (SUMT). Although not of particle i at discrete point xki , by evaluating F ðxki Þ.
particular importance in the current application, where With reference to the minimization (54), the lower
the functions can be computed very accurately, it has the value of F ðxki Þ, the greater the particle’s fitness.
been shown that this algorithm, when using finite dif- (3) For i = 1, 2, . . . , M:
ference gradients in the line searches, easily overcomes if F ðxki Þ 6 F bi then set F bi ¼ F ðxki Þ and pbi ¼ xki {best
the problem of severe numerical noise in the functions. point on trajectory i},
794 K. Jármai et al. / Computers and Structures 84 (2006) 787–797

if F ðxki Þ 6 F g then set F g ¼ F ðxki Þ and gb ¼ xki {best namely (tw, b, tf, h) are approximated by curve-fitting func-
global point}. tions (Table Curve 2D [19]) giving
(4) If F g < F gbefore then set N = 1, else set N =N + 1. pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
(5) If N > Nmax or k > kmax then STOP and set x* = gb; tf ¼ 33:20534 þ 6:701288  104 h2 ; ð59Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
else continue.
b ¼ 5851:785 þ 1:671844  102 h lnðhÞ; ð60Þ
(6) Compute new velocities and positions for instant qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
k + 1, using the rule: for i = 1, 2, . . . , M: tw ¼ 15:62577 þ 4:358947  105 h2 lnðhÞ; ð61Þ
vkþ1
i :¼ vki þ c1 r1 ðpbi  xki Þ þ c2 r2 ðgb  xki Þ ð57Þ h1 ¼ h  2tf . ð62Þ
xkþ1
i :¼ xki þ vkþ1
i ð58Þ Practical considerations restrict the final values that of the
where r1 and r2 are independently generated random design variables may assume. The thicknesses t may as-
numbers in the interval [0, 1], and c1, c2 are parame- sume any integer mm value. The height of the ring stiffeners
ters with appropriately chosen values. hr is restricted to steps of 10 mm. For the UB sections,
(7) Set k = k + 1 and F gbefore ¼ F g ; go to step 2. according to the ARBED catalogue, the only acceptable
values for h are 152, 203, 254, 305, 356, 406, 457, 533,
All four continuous optimization methods presented in 610, 686, 762, 838, 914 mm [20].
this section were used in an adaptive manner in order to
accommodate the discrete nature of the design variables 7.2. Results for the orthogonally stiffened cylindrical shell
to be used in the final manufacturing of the truss member. member
In particular, for the particle swarm method, the discreti-
zation procedure was as follows. Starting from the opti- 7.2.1. Results using the gradient-based algorithms
mum continuous values, the secondary search chooses In addition to the specific inequality constraints (1), (9),
the nearest discrete sizes for each continuous variable from (22), (28), (29), the following side constraints were also
the series of discrete values. The number of chosen discrete imposed: 4 6 t 6 40; 0 6 ns 6 40; 0 6 nr 6 40; 136 6 hr 6
sizes for one continuous variable can be two, three or 510; 152 6 h 6 914. All three gradient-based local optimi-
more. The possible variations can be obtained using zation algorithms, LFOPC, Dynamic-Q and ETOPC (see
binary, ternary or larger systems. In the example consid- Section 6.2.1), gave the identical continuous solution shown
ered here we use the binary system, two discrete sizes, in Table 1a. Indeed, throughout, the results for the three
upper and lower, associated with each continuous vari- methods are identical.
able. In the binary system the digit 0 is taken to indicate This solution satisfied all constraints. However, since ns
the upper discrete size, and the digit 1 indicates the lower and nr can only assume integer values, the continuous val-
value. The first 2n numbers in binary system give all the ues were rounded up to 27 and 9, respectively, and the min-
possible variations. Each possibility is tested for whether imization was repeated, but this time with variables ns = 27
the explicit and implicit constraints are satisfied, and the and nr = 9 fixed, and the others having starting values as
optimal values minimizing the merit function are thus listed in Table 1a. The results are given in Table 1b.
determined. Again this solution satisfies all the constraints. However,
Details of the discretization procedures for the other practical availability considerations further restrict the val-
algorithms are given together with the presentation of the ues that are acceptable. Thickness may assume any integer
optimization results in Section 7.2.1. mm value. Thus t is obviously rounded up to 14. The
height of the ring stiffeners is restricted to steps of
10 mm, and therefore hr is rounded up to 260. Since the
7. Numerical optimization results continuous solution, 207.7, is very close to 203, the height
h = 203 mm is selected. However, this solution slightly vio-
7.1. Numerical data lates the normalized constraint (9) with 7.4 · 103 and

The actual numerical values used in this study of the


quantities defined in Sections 4 and 5 are as follows: Table 1a
The specific density of steel is taken as q = 7.85 · 106 Results for the continuous problem
kg/mm3. The loading is specified by NF = 5.4 · 107 N and t (mm) ns nr hr (mm) h (mm) K ($)
pF = 1.5 MPa. The values of the remaining data used are:
13.8 26. 9 8.4 260.3 223.9 54444.4
L = 15 m, R = 1850 mm, fy = 355 MPa, E = 2.1 · 105
MPa, m = 0.3. The cost factor values are kM1 = kM2 = 1.0
$/kg, kF = 1.0 $/min and kP = 14.4 · 106 $/mm2.
Table 1b
As already stated in the previous section, the design Results for the discrete number of stiffeners
variables are x = [x1, x2, . . . , xn]T = [t, ns, nr, hr, h]T (rolled
t (mm) ns nr hr (mm) h (mm) K ($)
I-section). In order to do the computations with continu-
13.9 27 9 254.5 207.7 54511.4
ous values, the geometric characteristics of an UB section,
K. Jármai et al. / Computers and Structures 84 (2006) 787–797 795

Table 2 Table 4
Discrete optimization results for (ns, nr) = (27, 9) fixed Further discrete optimization results
t (mm) ns nr hr (mm) h (mm) K ($) t ns nr hr (mm) h (mm) K ($) Max constraint
(mm) violation (no)
14 27 9 270 203 55342.9
14 27 11 240 152 54403.7 0.07 (2)
14 27 11 240 203 55810.3 –
f* = K = 54808.6. If this design is rejected, the most reason- 14 26 11 240 152 54061.4 0.08 (2)
able other choice is to increase hr to the next allowable
value of 270. This candidate design satisfies all the con-
straints and its details are listed in Table 2. Note that the In computing the optimum solution the greatest expense
added cost due to discretization is 1.5% compared to the is in solving the initial continuous problem (see Table 1). In
cost of the continuous solution in Table 1b, and 0.5% for comparison to the continuous solution, the computational
the rejected discretized solution. Thus, if the latter is expense for the secondary discrete optimizations is small.
accepted it would represent a saving of 534$ per unit. Typically, starting at the point x = [20, 20, 20, 200, 200]T
Summarizing, the final discrete solution is obtained by the number of function and/or gradient vector evaluations,
starting with the overall continuous solution in Table 1a, for each of the methods are as follows:
and performing another continuous optimization but with LFOPC: Requires 3450 gradient evaluations of the
(ns, nr) = (27, 9) fixed. The resultant continuous solution is objective and constraint functions.
then rounded to the nearest allowable discrete values. Dynamic-Q: Requires 27 function and 27 gradient eval-
To investigate the existence of possible better solutions uations of the objective and constraint functions; i.e. 27
in the neighbourhood of the current candidate optimum simple spherically quadratic sub-problems are solved.
point, further computational experiments were performed. ETOPC: Requires 1845 gradient evaluations of the
Integer changes in the numbers of longitudinal stiffeners objective and constraint functions.
(stringers) ns and number of ring-stiffeners nr, will clearly
significantly affect the optimum solution if, with their val- 7.2.2. Results using the PSO algorithm
ues fixed, the optimization is carried out with respect to In this study the basic PSO algorithm is implemented
the other variables. A sensible local search is therefore: using the code written by Wood and Groenwold [21].
for each of the eight nearest integer neighbours to of the The cognitive learning coefficient used is c1 = 2. The social
point (ns, nr) = (27, 9) in the plane, determine the optimum learning coefficient is taken as c2 = 1.4. To check the effi-
continuous values of the remaining variables, and then ciency of the particle swarm global search, different popu-
round each to its nearest allowable discrete value. The lation sizes from M = 1 to M = 500 were used, with
results are summarized in Table 3. The last column lists starting points randomly generated in an appropriately
the maximum violation of the normalized constraints, large region. The no-improvement termination criterion
together with the corresponding constraint number in the (iterations) used is Nmax = 10. The maximum allowable
text. number of function evaluations is kmax, = 100 000. The ori-
The only new feasible solution found is listed in bold in ginal code was modified, by introducing a secondary dis-
the first line of Table 3. This design gives an almost identi- crete optimization procedure, in order to automatically
cal cost (0.03% lower) than for the first found candidate find discrete solutions. Secondary discrete optimization is
solution. The optimization procedure was also repeated described in [22].
for the choices (ns, nr) = (27, 11) and (26, 11), near the point The best continuous solution obtained, together with the
(27, 10) that gave the best design to date, but as shown in discrete solution are listed in Table 5. This discrete solution
Table 4, no further improved feasible design was found. corresponds exactly with the previous discretiszed solution
The final discrete optimum solution is therefore taken found (listed in Table 2), following on the continuous solu-
as: x* = [t = 14 mm, ns = 27, nr = 10, hr = 250 mm, tion obtained in Section 7.2.1 using the gradient-based
h = 203 mm]T with optimum K* = 55326.3 $. local optimizers.

Table 3
Discrete optimization results for various fixed (ns, nr) combinations
t (mm) ns nr hr (mm) h (mm) K ($) Max. constraint violation (no)
14 27 10 250 203 55326.3 –
15 27 8 250 203 56243.4 0.19 (28)
14 26 10 250 203 54930.9 0.02 (1)
14 26 9 250 203 53968.5 0.07 (28)
14 26 8 260 254 55027.3 0.06 (28)
19 28 10 240 152 64463.4 0.00 (30)
26 28 9 240 152 78866.1 0.00 (30)
38 28 8 260 152 104763.9 0.01 (30)
796 K. Jármai et al. / Computers and Structures 84 (2006) 787–797

Table 5 sponding value of the objective function, are as


Results for the 5 variables: continuous and discrete solutions (M = 500) experimentally found over a large number (some hundreds)
t (mm) ns nr hr (mm) h (mm) K ($) of different runs.
13.82 26.85 8.31 260.96 225.79 54444.62 (cont.) Table 8 shows that the efficiency and reliability of PSO
14 27 9 270 203 55342.9 (disc.) algorithm, increases gradually with population sizes. It is
remarkable that it could find a relatively good solution
when only one particle was used. In that case, however,
Table 6 the result cannot be taken to be reliable, as is evident
Solutions for 4 variables, with the number of stringer stiffeners respectively
fixed at ns = 23, 24, 25, 26
from the range of objective function values obtained. Com-
paring the best continuous solution in Table 1 (cost
t (mm) ns (fixed) nr hr (mm) h (mm) K ($)
54444.4)$ with the best continuous PSO solution in Table
14 26 9 260 254 56042.3 5 (cost 54444.6)$, shows that they are almost identical.
15 25 8 270 254 57142.7
15 24 9 260 203 55724.8
Comparing, however, the respective number of function
15 23 8 280 254 56705.9 evaluations required, as is respectively indicated at the
end of Section 7.2.1 and in Table 8 above (LFOPC 3450
gradient evaluations, Dynamic-Q: 27 function and 27 gradi-
Table 7
ent evaluations, ETOPC 1845 gradient evaluations, and
Solutions for 4 variables, with the number of ring stiffeners respectively PSO 60 000 function evaluations), it is clear that the gra-
fixed at nr = 7, 8, 10, 11 dient-based local algorithms are vastly superior with regard
t (mm) ns nr (fixed) hr (mm) h (mm) K ($) to computational economy. Nevertheless, the results for
the PSO algorithm are important, because they confirm
14 26 7 280 305 56764.1
14 26 8 270 254 55501.3 that the accurate local optimizers have indeed converged
in a globally optimal region.
14 27 10 250 203 55326.3 (disc.)
13.91 27.51 10.00 246.47 186.21 54701.7 (cont.) 7.3. Results for the un-stiffened shell
14 27 11 240 203 55810.2
Using the shell buckling constraint for the un-stiffened
shell as defined by expressions (42)–(47), it follows that
Further experiments were performed with respectively the constraint is satisfied if the shell thickness is
ns, and nr kept fixed during the optimization. The discreti- t = 50 mm. In this case the objective function, defined
zation results are as reported in Tables 6 and 7. by (48) is K = 112131.3 $. Thus, a cost saving of 51% can
The only marginally improved discrete solution be achieved by the orthogonal stiffening of the shell
obtained is listed in bold in the third row of Table 7. This member.
solution is obtained from discretizing the continuous solu-
tion listed in the fourth row. This solution corresponds 8. Conclusions
exactly to the best solution obtained via the gradient-based
local optimizers in Section 7.2.1, and thus confirms that The study shows that a practically viable and economi-
the computed global discrete optimum solution is: x* = cally advantageous design of a loaded shell member of a
[t = 14 mm, ns = 27, nr = 10, hr = 250 mm, h = 203 mm]T platform truss may be obtained, by coupling a realistic
with optimum K* = 55326.3 $. and accurate mathematical model of the member, to opti-
Typical computational cost in using the PSO algorithm mization algorithms. In particular, a minimum cost design,
is as indicated in Table 8. Due to the random nature of the from a material and manufacturing point of view, may be
starting positions of the particles, no run is identical to achieved, whilst ensuring safety and feasibility of assembly
another. The ranges given for the number of function eval- through the inclusion of stability and manufacturing con-
uations required per optimization run, and for the corre- straints in the optimization algorithms. The optimization
procedures allow for a choice to be made between a shell
member that is orthogonally stiffened by ring stiffeners
Table 8
and stringers, and an un-stiffened shell member.
Representative range of number of function evaluations and range of
optimum objective function value for continuous optimization found at The results show that if the member is stiffened, a cost
different number of particles saving of more than 50%, compared to that of the un-stiff-
Number of Number of function K ($) ened shell, may be obtained. This is due to the fact that
particles M evaluations stiffening allows the thickness of the shell to be reduced
500 26 184–100 000 54444.4–56384.3 to 14 mm, which represents a significant material saving,
250 11 666–43 044 54444.8–56384.5 compared to the un-stiffened shell that requires a minimum
32 2778–9861 54450.0–57487.0 thickness of 50 mm. This material saving overshadows the
16 2231–8378 54615.1–58057.2 increase in labour and welding costs associated with the
1 2031–2816 55384.8–77040.1
stiffening.
K. Jármai et al. / Computers and Structures 84 (2006) 787–797 797

This study emphasises the importance of using a multi- [6] Snyman JA. A gradient-only line search method for the conjugate
ple-algorithm approach to solving practical optimization gradient method applied to constrained optimization problems with
severe noise in the objective function. Int J Numer Methods Eng
problems. To ensure confidence that the computed contin- 2005;62:72–82.
uous optimal design is accurate, and indeed corresponds to [7] Kennedy J, Eberhardt R. Particle swarm optimization. Proc int conf
the global optimum, use was made of four conceptually dif- on neural networks. NJ, USA: Piscataway; 1995. p. 1942–8.
ferent optimization algorithms readily available and famil- [8] Farkas J. Thickness design of axially compressed unstiffened
iar to the authors. The first three algorithms are local cylindrical shells with circumferential welds. Weld World 2002;
46(11–12):26–9.
optimizers and gave identical and clearly very accurate [9] Farkas J, Jármai K. Economic design of metal structures. Rotter-
results. To ensure that the optimum obtained by these local dam: Millpress; 2003.
optimizers corresponds to the global optimum, many inde- [10] Farkas J, Jármai K, Snyman JA, Gondos Gy. Minimum cost design
pendent applications of the relatively new PSO global opti- of ring-stiffened welded steel cylindrical shells subject to external
mization algorithm was also performed and with increasing pressure. In: Lamas A, Simões da Silva L, editors. Proc 3rd European
conf steel structures, Coimbra, 2002, Universidade de Coimbra, 2002.
number of particles. Through these experiments the PSO p. 513–22.
algorithm indicated with a very high probability that the [11] Farkas J, Jármai K, Virág Z. Optimum design of a belt-conveyor
other, locally more accurate algorithms, did indeed con- bridge constructed as a welded ring-stiffened cylindrical shell. Weld
verge to a globally optimal region. World 2004;48(1–2):37–41.
[12] Farkas J, Jármai K. Optimum design of a welded stringer-stiffened
steel cylindrical shell subject to axial compression and bending 57th
Acknowledgements annual assembly of international institute of welding. Weld World
2005;49(5–6):85–9.
[13] Det Norske Veritas (DNV): Buckling strength analysis. Classification
The research work was supported by the Hungarian Sci- Notes No. 30.1. Hovik, Norway, 1995.
entific Research Foundation grants OTKA T38058, T37941. [14] Eurocode 3: prEN 1993-1-1:2002. Design of steel structures. Part 1-1:
The project was also supported by the Hungarian-South General structural rules. Brussels, CEN.
African Intergovernmental S&T co-operation program [15] Snyman JA, Heyns PS, Vermeulen PJ. Vibration isolation of a
DAK 7/2002. The Hungarian funding is from the Research mounted engine through optimization. Mech Mach Theory
1995;30:109–18.
and Technological Innovation Fund, and the South African [16] Snyman JA, Berner DF. The design of a planar robotic manipulator
partner is the National Research Foundation. for optimum performance of prescribed tasks. Struct Multidisc Optim
1999;18:95–106.
[17] Craig KJ, de Kock DJ, Snyman JA. Using CFD and mathematical
References optimisation to minimize stack pollution. Int J Numer Methods Eng
1999;44:551–66.
[1] Snyman JA. A new and dynamic method for unconstrained minimi- [18] Snyman JA. Unconstrained minimization by combining the dynamic
zation. Appl Math Modell 1982;6:449–62. and conjugate gradient methods. Quaest Math 1985;8:33–42.
[2] Snyman JA. An improved version of the original leap-frog dynamic [19] TableCurve 2D: Users’ manual. Systat Software Inc., 2003.
method for unconstrained minimization LFOP1(b). Appl Math [20] Profil Arbed. Sales program, Structural shapes, Arcelor Long
Modell 1983;7:216–8. Commercial 2001.
[3] Snyman JA. The LFOPC leap-frog method for constrained optimi- [21] Wood DW, Groenwold AA. Basic PSO global optimization code,
zation. Comput Math Appl 2000;40:1085–96. Department of Mechanical and Aeronautical Engineering, University
[4] Snyman JA, Stander N, Roux WJ. A dynamic penalty function of Pretoria, 2003.
method for the solution of structural optimization problems. Appl [22] Jármai K. Topology optimization of tubular structures. In: Jármai K,
Math Modell 1994;18:453–60. Farkas J, editors. Mechanics and design of tubular structures. Springer-
[5] Snyman JA, Hay AM. The Dynamic-Q optimization method: an Verlag; 1998. p. 225–84. ISBN: 3-211-83145-2 [chapter 5].
alternative to SQP? Comput Math Appl 2002;44:1589–98.

Вам также может понравиться