ARTICLE IN PRESS

Ocean Engineering 37 (2010) 966–977

Contents lists available at ScienceDirect

Ocean Engineering
journal homepage: www.elsevier.com/locate/oceaneng

Robust optimization for ship conceptual design

Matteo Diez , Daniele Peri
INSEAN—The Italian Ship Model Basin, Resistance and Optimization Department, Via di Vallerano 139, 00128 Rome, Italy

a r t i c l e in f o a b s t r a c t

Article history: The paper presents an approach for the robust optimization of a bulk carrier conceptual design, subject
Received 31 December 2009 to uncertain operating and environmental conditions. The uncertainties involved in the optimization
Accepted 22 March 2010 process are addressed and a general formulation for robust design is given. Specifically, the
uncertainties involved in the decision making process are taken into account by means of their
Keywords: probabilistic distributions. The expected values and the standard deviations of the relevant quantities
Robust design optimization (RDO) are assessed and included in the optimization objectives, whereas the constraints are evaluated in the
Ship conceptual design worst case. This leads to a robust design able to keep a good performance in the whole probabilistic
Particle swarm optimization (PSO) operating scenario. A particle swarm optimization algorithm is used for the global minimization
process, minimizing the expectation and the standard deviation of the unit transportation cost.
& 2010 Elsevier Ltd. All rights reserved.

1. Introduction are fundamental analytical and computational obstacles that

must be overcome before optimization can make a widespread
Over the years, optimization has been playing an increas- impact on the practice of ship design. Furthermore, robust and
ingly important role in engineering. Advanced modeling and automated grid generation and manipulation has proved to be a
algorithms in optimization constitute now an essential part in the serious challenge as well as the need to account for complex, real-
design and in the operations of complex aerospace (Hicks and industrial geometrical and functional constraints, together with
Henne, 1978; Sobieszczanski-Sobieski and Haftka, 1997; Alexan- the difficulty of generating the objective functions and their
drov and Lewis, 2002; Willcox and Wakayama, 2003; Morino derivatives automatically and robustly when these functions are
et al., 2006; Iemma and Diez, 2006) and automotive (Baumal computed by solving systems of partial differential equations
et al., 1998; Kodiyalam and Sobieszczanski-Sobieski, 2001) (PDE). The potential benefits and pay-offs of the impact of the
applications, when, for example, it is by all means important to optimization on the ship design process are so great, however,
reduce costs and shorten time of development. In the design of that despite the damping effects of reality on the immediate
large and complex systems, the use of efficient optimization expectations, research on optimization has continued, yielding
tools leads to better product quality and improved functionality. promising results and revealing specific new challenges and
The challenge is typically to model complex and frequently non- directions of research.
linear systems and structures to emulate and simulate possible The present paper presents an application of optimization for
operating conditions under a wide range of scenarios in order to ship conceptual design. The standard deterministic formulation
allow for more intelligent design and optimization of these for design optimization is extended to take into account the
technologies. uncertainty related to the design variables, to the operating
The recent years have seen some progress in optimization for conditions, and to the computational results of the simulations. In
ships too (Ray et al., 1995; Peri and Campana, 2003, 2005; Parsons the standard approach to the conceptual optimal design, all the
and Scott, 2004; Pinto et al., 2004; Campana et al., 2007, 2009; relevant quantities and models involved in the process are taken
Papanikolaou, 2009). However, arguably, the initial excitement into account from a deterministic viewpoint. The operating
that accompanied the emergence of these techniques has conditions are statically specified, a mathematical programming
diminished somewhat over the years, due to the fact that these problem is defined, and the final solution is obtained within this
methods are not as generally accepted or widely used in practical rigid framework. This means that objectives and constraints are
ship design as the optimization community initially hoped. The evaluated through a ‘‘static’’ or deterministic assessment of the
explanation is not straightforward. It is certainly true that there problem. As a consequence, no information is available about the
performance in off-design conditions. Moreover, the design tools
are not considered as a potential source of uncertainty, and the
 Corresponding author. Tel.: + 39 06 50299314; fax: + 39 06 5070619. accuracy of the analysis is not assessed during the optimization
E-mail addresses: [email protected] (M. Diez), [email protected] (D. Peri). process. Furthermore, the effects of small variations of the design

0029-8018/$ - see front matter & 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.oceaneng.2010.03.010
 ARTICLE IN PRESS
M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977 967

variables, potentially caused by manufacturing errors, are not It may be noted that, in naval applications (Diez and Peri,
addressed in the optimization procedure. In other words, within a 2009), as well as in aeronautical problems (Padula et al., 2006;
deterministic framework, design variables, environmental and Diez and Iemma, 2007), the operating and environmental
operating conditions, merit factors and constraints are assessed conditions may be considered as ‘‘intrinsic’’ stochastic functions,
disregarding any kind of uncertainty involved in the analysis. This whose expected values and standard deviations can neither be
yields an optimal solution which depends on the particular influenced by the designer nor by the manufacturer. Conversely,
deterministic model used for the analysis and on the environ- uncertainties related to design variables and functions evaluation
mental and operating conditions assumed. When using different are connected to the available knowledge and technology,
analysis models or working in off-design conditions (manufac- and—theoretically—may be reduced by improving modeling,
tured products deviate from designed products—actual products computing and manufacturing processes. A standard ‘‘determi-
must perform under a variety of different operating conditions), nistic’’ approach to manage different operating conditions in
the deterministic-optimal solution performance might drastically design optimization is to take into account an aggregate objective
drop. In this respect, Marczyk (2000) states that, in a deterministic function (AOF) as a linear combination of the system perfor-
engineering context, optimization is associated to specialization mance, evaluated in different operating points. Differently, in the
and, therefore, is the opposite of robustness. The aim of robust RDO approach, the attention is focused on the uncertain variation
design optimization (RDO) is to overcome this deficiency, finding of the operating conditions, addressed from a stochastic point of
an optimal solution capable of keeping a good performance in a view. The goal is that of minimizing the effects of the
wide range of variation of the uncertain parameters involved in uncertainties involved in the system design, without suppressing
the design process (such as operating conditions, manufacturing their causes.
errors, etc.). RDO methods developed to improve product quality In the next section, an overview of a general optimization
and reliability in industrial engineering, are about preventing the problem affected by uncertainty is presented. Specific attention
effects of uncertainty on the final product performance. Specifi- is paid to the different kinds of uncertainty generally involved
cally, the uncertain parameters are taken into account by means in the optimization task. In Section 3, the formulation of the
of their probabilistic distribution and included, some how, in the present optimization problem is given. Section 4 presents a brief
definition of the optimality criteria used for the optimization of discussion on the difference between a robust design problem and
the design (Beyer and Sendhofff, 2007; Park et al., 2006; Zang a multi-point design optimization procedure. The numerical
et al., 2005). results are presented and commented upon in Section 5, whereas
In this context, the goal of RDO is to interpret differently the final remarks are given in Section 6.
concept of optimization, taking into account the uncertainty
analysis and redefining ‘‘optimality’’ in terms of ‘‘robustness.’’ In
other words and in general, while looking for a robust-optimal 2. Optimization problems affected by uncertainty
solution, one considers flexibility about uncertainties rather than
specialization. In this section, an overview of an optimization problem
Following this approach, designers have become increasingly affected by uncertainty is presented. In this context, the designer
concerned with managing uncertainty. Generally, the sources of concern is that of finding an optimal configuration able to keep a
uncertainty may be categorized into two main types: external and good performance in a wide range of variation of a number of
internal (Du and Chen, 2000a). External uncertainties are related uncertain parameters. In order to achieve such an optimal
to the analysis models input, such as design variables tolerance or solution, an optimality criterion, based on robustness of the final
uncertain usage and operating conditions. Internal uncertainties design, has to be defined. In this context, the term ‘‘robust’’ always
are related to the system output and are associated with the deals with something which is uncertain. Therefore, attention to
accuracy in computing; in other words, internal sources of robustness always involves attention to some kind of uncertainty.
uncertainty are related to those stochastic variations which are A number of authors in the literature give different meaning to
not conditional on the input uncertainties. robustness depending on the application, and different kinds of
The tools of statistical decision theory, specifically Bayes uncertainty are addressed. The interested reader is referred to
criteria (De Groot, 1970), may provide a sound framework to Beyer and Sendhofff (2007), Park et al. (2006), and Zang et al.
manage the above uncertainties and formulate the problems for (2005).
RDO (Trosset et al., 2003). Specifically, throughout the design In order to define the context of the present work, the
process, the uncertain parameters are not defined by a ‘‘simple’’ following optimization problem may be formulated:
unique value, but are varying within a given range with a
specified probability density function, and the expectations of the minimize f ðx,yÞ given y ¼ y^ A B
xAA
merit factors, with respect to these variations, are taken into
^ r0, n ¼ 1, . . . ,N
subject to gn ðx, yÞ
account. In addition, the standard deviation of the relevant
^ ¼ 0, m ¼ 1, . . . ,M
and to hm ðx, yÞ ð1Þ
quantities may be also considered to improve the insensitiveness
of the final solution to the uncertain parameters variations (Zang
et al., 2005). where x A A is the design variables vector (which represents the
Accordingly, in the present work the expected values and the designer choice), y^ A B is the design parameters vector (which
standard deviations of the relevant quantities are assessed and collects those parameters independent of the designer choice, e.g.,
included in the optimization objectives, whereas the design environmental or usage conditions defining the operating point),
constraints are evaluated in the worst case. The paper presents and f ,gn ,hm : Rk -R, are, respectively, the optimization objective
numerical applications to a bulk carrier conceptual design, where and the inequality and equality constraint functions. While
the numerical noise due to environmental uncertainty is taken handling the above problem, the following uncertainties may occur.
into consideration. A particle swarm optimization algorithm Uncertain design variable vector: When translating the designer
(Kennedy and Eberhart, 1995) is used for the multi-objective choice into the ‘‘real world,’’ the design variables may be affected
global minimization process. The optimization objectives are the by uncertainty due to manufacturing tolerances or actuators
expectation and the standard deviation of the unit transportation precision. These uncertainties may be classified as external, when
cost. associated to the analysis models input (Du and Chen, 2000a).
 ARTICLE IN PRESS
968 M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977

Assume a specific designer choice x and define as n A X the error With respect to the uncertainties outlined above, different
or tolerance related to this choice.1 We may assume n be a approaches may be followed for the re-definition of the
stochastic process with probability density function pðnÞ; by optimization problem. Specifically, the optimization task may be
R
definition it is X pðnÞ dn ¼ 1. The expected value of x is, by defined in terms of:
definition,
Z  minimization of the variance, or of the standard deviation,
pffiffiffiffi
x :¼ mðx þ nÞ ¼ ðx þ nÞpðnÞ dn ð2Þ s :¼ V , of f: this leads to a robust design in a strict
X
sense—e.g., Taguchi (1986) methods;
It is apparent that if the stochastic process n has zero expectation,  minimization of the expectation of f: if f represents a general
i.e. loss in the performance, than the expected value of f can be
Z seen as a risk—Bayesian approach as in statistical decision
n :¼ mðnÞ ¼ npðnÞ dn ¼ 0 ð3Þ theory (De Groot, 1970; Trosset et al., 2003);
X  minimization of f in the worst possible case; this is the most
we obtain x ¼ x . It may be noted that, in general, the probability conservative approach—‘‘minmax’’ approach, see again Trosset
density function pðnÞ depends on the specific designer choice x . et al. (2003);
Uncertain environmental and usage conditions: In real life  assessing probabilistic constraints in the minimization of the
applications, environmental and operational parameters may objective function (Tu et al., 1999; Sues et al., 2001; Du and
differ from the design conditions y^ (see Problem 1), involving a Chen, 2000b; Agarwal, 2004; Agarwal and Renaud, 2004).
source of external uncertainty. The design parameters vector may
be assumed as a stochastic process with probability density With respect to the previous approaches, different definitions
function pðyÞ and expected value or mean may be found in the literature—the interested reader is referred,
Z again, to Beyer and Sendhofff (2007):
y :¼ mðyÞ ¼ ypðyÞ dy ð4Þ
B
 Robust design (RD): process of defining the robust design in the
Note that, in this formulation, the uncertainty on the usage strict sense (e.g., Taguchi methods). The attention in this case
conditions is not related to the definition of a specific design is mainly on variance or standard deviation.
point. Environmental and operating conditions are treated as  Robust optimization or robust design optimization (RDO):
‘‘intrinsic’’ stochastic processes in the whole domain of variation optimization process considering uncertainty in the evaluation
B, and the designer is not requested to pick a specific design point of the objective function; expected value, variance, worst case,
in the usage parameters space. For this reason, we do not define etc. may be taken into account.
an ‘‘error’’ in the definition of the usage conditions, preferring the  Reliability-based design optimization (RBDO): the attention is
present approach which identifies the environmental and opera- focused on the statistical feasibility of the design (i.e., on the
tional parameters in terms of their probabilistic distributions in constraints). The constraints are treated as probabilistic
the whole domain of variation. inequalities and give a statistical feasible region.
Uncertain evaluation of the function of interest: The evaluation of
the functions of interest (objectives and constraints) may by While RD and RDO are mainly focused on expectation and
affected by uncertainty due to inaccuracy in modeling or compu- variance of a cost function (Zang et al., 2005), the RBDO
ting—internal uncertainty as in Du and Chen (2000a). Collect concentrates on the probabilistic handling of the constraints (Tu
objective and constraints in a vector f :¼ ½f ,g1 , . . . ,gN ,h1 , . . . ,hM T , et al., 1999; Sues et al., 2001; Du and Chen, 2000b; Agarwal, 2004;
and assume that the assessment of f for a specific ‘‘deterministic’’ Agarwal and Renaud, 2004). These are treated as probabilistic

^ is affected by a stochastic error u A F.
design point, f :¼ fðx , yÞ, inequalities (Nocedal and Wright, 1999) and the n-th determi-

Accordingly, the expected value of f is nistic constraint of the type gn ðx,yÞ r0 is treated using the general
Z probabilistic statement
  
f :¼ mðf þ uÞ ¼ ðf þ uÞpðuÞ du ð5Þ
F PS :¼ P½gn ðx,yÞ r 0 Z P0 ð8Þ

Note that, in general, the probability density function of u, pðuÞ, where PS is the probability of success, PðAÞ denotes the probability

^
depends on f and, therefore, on the design point ðx , yÞ. of the event A and P0 is a given target probability. Note that
Combining the above uncertainties, we may define the the probability of failure, PF, equals (1 PS). In the following, the
expected value of f as constraints of the optimization problem will be defined in the
Z Z Z worst possible case, choosing a conservative approach ðP0 ¼ 1Þ.
f :¼ mðfÞ ¼ ½fðx þ n,yÞ þ upðn,y, uÞ dn dy du ð6Þ The issues connected to the probabilistic handling of the
X B F constraints are beyond the scope of the present work and will
where pðn,y, uÞ is the joint probability density function associated be not further addressed here.
to n,y, u. It is apparent that f ¼ f ðx Þ; in other words, the
expectation of f is a function of the only designer choice.
3. The present RDO problem
Moreover, the variance of f with respect to the variation of
n,y, u is
As already mentioned, while handling an RD or RDO problem,
Z Z Z
the focus in mainly on the expected value of the function of
VðfÞ :¼ s2 ðfÞ ¼ f½fðx þ n,yÞ þ uf ðx Þg2 pðn,y, uÞ dn dy du
X B F interest and on its variance or standard deviation (Zang et al.,
ð7Þ 2005). Specifically, in this work we will use as objectives the
expected value and the standard deviation of a suitably defined
resulting, again, a function of only the designer choice variables. cost function. It may be noted that, according to statistical
decision theory, the expectation of a general loss in performance
1
The symbol * is used in the present formulation to denote a specific designer represents the risk with respect to the variation of the
choice. While x represents all the possible choices in A, x defines a specific one. probabilistic parameters (De Groot, 1970). Therefore, adopting
 ARTICLE IN PRESS
M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977 969

Table 1
Test problems for robust optimization of a bulk carrier conceptual design.

Problem id. Uncertain parameter y Unit Distribution type Lower bound Upper bound Optimization objective(s)

#1 Port handling rate ton/day Uniform 1000 11,000 mðf Þ

#2 Round trip nm Uniform 1000 5000 mðf Þ
#3 Fuel price GBP/ton Uniform 50 150 mðf Þ
#4 Port handling rate ton/day Uniform 1000 11,000 mðf Þ, sðf Þ
#5 Round trip nm Uniform 1000 5000 mðf Þ, sðf Þ
#6 Fuel price GBP/ton Uniform 50 150 mðf Þ

f, unit transportation cost.

the loss expectation as the objective function implies the scheme for the integration of the objective f over the usage
minimization of the risk, and the identification of the Bayes variation domain, the following approximation holds:
decision against the stochastic variation of the uncertain Z X
m
parameters (De Groot, 1970; Trosset et al., 2003). The adoption mðf Þ :¼ f ðx,yÞpðyÞ dy  vh f ðx,yh Þpðyh Þ ð11Þ
B
of the standard deviation as objective function will lead, h¼1

conversely, to a robust design in a strict sense, with minimum where the vh and the yh represent, respectively, the Gaussian
sensitivity with respect to the variation of the probabilistic weights and abscissas, specifically assessed for the integration
parameters. In this work, a bulk carrier conceptual design over the domain B. It is apparent that the latter provides a
problem is addressed, assuming as the general loss in the system particular multi-point AOF, with uniquely defined points and
performance, the difference f f^ between the actual unit trans- weights. Assuming n ¼m, comparing Eqs. (10) and (11) gives
portation cost (f) and a reference target value ðf^ Þ. This implies
wk ¼ vk pðyk Þ, k ¼ 1, . . . ,n ð12Þ
that, for optimization purposes, the cost function reduces to the
unit transportation cost (f). with yk corresponding to the k-th Gaussian abscissa. It may be
Here, the focus is on the probabilistic variation of the ship observed that the performance expectation mðf Þ, evaluated
environmental and operating conditions, defined by the vector y. through quadrature, gives a uniquely defined AOF in the context
The latter is therefore taken into account as a stochastic process of multi-point optimization.
and the robust optimization problem is formulated as follows: It is worth noting that the evaluation of expected value (and
minimize f1 ðxÞ :¼ m½f ðx,yÞy A B and f2 ðxÞ :¼ s½f ðx,yÞy A B variance as well) needs the computation of an integral over the
xAA stochastic parameters domain. Having in mind that during the
subject to supfgn ðx,yÞg r0, n ¼ 1, . . . ,N optimization task the evaluation of the objective function may be
yAB
required hundreds or thousands of times, it is apparent that the
and to m½hm ðx,yÞy A B ¼ 0, m ¼ 1, . . . ,M ð9Þ
choice of a reliable and effective algorithm for the computation of
the integrals is a crucial issue. A convergence analysis (shown in
the numerical results, Section 5) reveals that, to the aim of the
As mentioned, the above multi-objective minimization pro- present work, a five-point Gauss–Legendre quadrature scheme
blem is solved using a particle swarm optimization technique represents a good compromise between accuracy and computa-
(Kennedy and Eberhart, 1995), whose formulation is presented in tional effort.
Appendix B.

5. Numerical application to a bulk carrier conceptual design

4. Robust design optimization versus multi-point
optimization In this section, the RDO formulation is applied to the con-
ceptual design of a bulk carrier, subject to uncertain usage
The aim of this section is to provide the reader with a conditions. Specifically, the test problems addressed are summar-
comparison between the RDO approach—handling probabilistic ized in Table 1, whereas the design variables used are indicated in
usage—and the multi-point optimization approach. Table 2. It may be noted that, in the proposed problems, the
A standard deterministic method to manage different operat- uncertain parameters are treated separately from each other.2
ing conditions within an optimization problem is the use of an The model used for the bulk carrier conceptual design analysis
aggregate objective function (AOF) combining the same objective, is that proposed by Sen and Yang (1998), as used by Parsons and
evaluated for different operating points—multi-point optimiza- Scott (2004). The same analysis tool has been also applied by Hart
tion, e.g. Reuther et al. (1999a, b), Mirzaei et al. (2007), and and Vlahopoulos (2009) within a particle swarm optimization
Jameson et al. (2007). The latter may be expressed as problem. The cost function addressed in this work is the unit
X
n transportation cost, whereas the design constraints pertain to
fAOF ðxÞ :¼ wk f ðx,yk Þ ð10Þ geometry, static stability, and model validity. The dimension
k¼1 constraints refer to the Capesize category (see again Table 2). The
If, on one hand, this approach is able to take into account different model by Sen and Yang (1998) as modified by Parsons and Scott
operating scenarios, on the other hand, the definition of the (2004) is briefly recalled in Appendix A. Specifically, in the
objective function in terms of yk and wk depends on designer numerical results, the stochastic variation of port handling rate,
experience and sensitivity, resulting in an arbitrary choice. round trip distance and fuel price, respectively, are considered
Moreover, no information about performance variation is avail- one by one (as uncorrelated random parameters). Moreover, for
able to the optimizer.
To compare the above approach to an RDO process managing 2
We assume that the random parameters involved in the process are
uncertain usage, advantage is taken of the numerical evaluation of independent from each other (uncorrelated) and the robust optimization task is
the integral of Eq. (6). Specifically, using a Gaussian quadrature conducted using their marginal density functions.
 ARTICLE IN PRESS
970 M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977

Table 2 uncertain port handling rate

Design variables used in this work.
1 deterministic solution

design variable non-dimensional value

robust solution
# Variable Unit Lower bound Upper bound

cruise speed
block coeff.
1 Length m 100 600
0.5
2 Beam m 10 100

length

beam

depth

draft
3 Depth m 5 30
4 Draft m 5 30
5 Block coefficient 0.63 0.75 0
6 Cruise speed knots 14 18

-0.5
Table 3
Robust vs. deterministic approaches for single objective optimization.

Approach Objective Constraints -1

Deterministic f ½x, mðyÞ supy A B fgn ðx,yÞg, n¼ 1,y,N 1 2 3 4 5 6

Robust m½f ðx,yÞ supy A B fgn ðx,yÞg, n¼ 1,y,N design variable id.

f, unit transporation cost; y, uncertain parameter. Fig. 1. Problem #1: non-dimensional optimal variables for deterministic and
robust case.

Table 4 24
Optimal parameters for problem #1 (uncertain port handling rate). deterministic solution
robust solution
22
# Variable/function Unit Deterministic solution Robust solution unit transportation cost [GBP/ton]
20
1 Length m 182.79 165.70
2 Beam m 30.60 27.75
18
3 Depth m 15.89 14.25
16
4 Draft m 11.90 10.76
5 Block coefficient 0.67 0.65 14
6 Cruise speed knots 14.00 14.00
12
f ½xopt , mðyÞ GBP/ton 7.22 7.31 10
m½f ðxopt ,yÞ GBP/ton 8.63 8.50
8

s½f ðxopt ,yÞ GBP/ton 3.47 2.93 6

Optimization objectives highlighted in bold.

4
2000 4000 6000 8000 10000
port handling rate [ton/day]
the sake of simplicity and without loss of generality, a uniform Fig. 2. Problem #1: performance parametric analysis for optimal solutions.
distribution is assumed for all the random parameters involved
(see again Table 1). If, on one hand, an accurate modeling of the
input distributions from observed data would likely lead to a 0.5
different definition of the probability density functions, on the
unit transportation cost [GBP/ton]

other hand, this does not represent a limitation for the application 0
of the present formulation. Furthermore, assessing probability
density functions from observed data is outside the scopes of the
-0.5
present work and, therefore, no further addressed.
In problems #1, #2 and #3, the expectation of the unit
transportation cost is taken as the only optimization objective, -1
whereas the constraints are treated in the worst case. A deter-
ministic particle swarm optimization (DPSO) algorithm coupled -1.5
with a penalty function method (see Appendix B) is used to solve
the constrained minimization problem. The solution is compared
with that obtained through a ‘‘deterministic’’ standard approach -2
difference between robust and deterministic
(the two approaches are summarized in Table 3). solution performances
Table 4 contains the optimal solutions for problem #1 -2.5
(uncertain port handling rate), whereas Fig. 1 depicts the non- 2000 4000 6000 8000 10000
dimensional optimal variables for the same problem. The port handling rate [ton/day]
dimensionless values are set from  1 to + 1. Specifically,  1
corresponds to the dimensional variables lower bound, whereas Fig. 3. Problem #1: difference between deterministic and robust solution
+ 1 stands for the upper bound. Fig. 2 presents a comparison performance.

between ‘‘robust’’ and ‘‘deterministic’’ solution performance in

terms of unit transportation cost as a function of the uncertain performance as a function of the port handling rate. It may be
parameter (port handling rate). Moreover, Fig. 3 shows the noted that, as the uncertain parameter has a uniform density
difference between ‘‘robust’’ and ‘‘deterministic’’ solution function, the integral of the latter function represents the
 ARTICLE IN PRESS
M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977 971

difference between ‘‘robust’’ and ‘‘deterministic’’ solution confirmed by the extremely close numerical solutions obtained
expectation of the transportation cost. through the optimization process.
As mentioned before, for the evaluation of the integral of The results of problem #3 (uncertain fuel price) are summarized
Eqs. (6) and (7) a five points Gauss–Legendre quadrature scheme in Table 6 and depicted in Figs. 8 (again,  1 represents the vari-
is used. Fig. 4 shows a convergence analysis for the integral ables lower bound, and +1 the upper bound) and 9. The same
evaluations, varying the number of Gaussian abscissas. Moreover,
a convergence analysis for the final optimal solution of a single-
objective minimization process is shown in Fig. 5. In this context, Table 5
Optimal parameters for problem #2 (uncertain round trip).
the choice of five Gaussian abscissas represents a good
compromise between accuracy and computational effort. # Variable/function Unit Deterministic solution Robust solution
Table 5 presents the optimal solutions for problem #2
(uncertain round trip distance, see also Fig. 6, where  1 1 Length m 182.79 182.95
corresponds to the variables lower bound, and + 1 to the upper 2 Beam m 30.60 30.63
3 Depth m 15.89 15.91
bound). Fig. 7 shows a performance parametric analysis of the 4 Draft m 11.90 11.92
solutions with respect to the round trip measure. It has to be 5 Block coefficient 0.67 0.67
noted that the dependance of the cost on the uncertain parameter 6 Cruise speed knots 14.00 14.00
is linear, therefore m½f ðyÞ ¼ f ½mðyÞ and the ‘‘robust’’ approach
theoretically coincides with the ‘‘deterministic’’ one. This is f ½xopt , mðyÞ GBP/ton 7.22 7.22
m½f ðxopt ,yÞ GBP/ton 7.22 7.22

numerical evaluation of expected value and s½f ðxopt ,yÞ GBP/ton 1.13 1.13
standard deviation
Optimization objectives highlighted in bold.
0.1
expected value
standard deviation

0
uncertain round trip miles
1 deterministic solution
design variable non-dimensional value

-0.1 robust solution

relative error

-0.2

cruise speed
0.5

block coeff.
length

-0.3
beam

depth

draft
0
-0.4

-0.5 -0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
1/n (Gauss-Legendre)
-1
Fig. 4. Convergence analysis for integral evaluations through Gauss–Legendre
quadrature: relative error for expected value and standard deviation as a function 1 2 3 4 5 6
of 1/n (n, number of Gaussian abscissas).
design variable id.

Fig. 6. Problem #2: non-dimensional optimal variables for deterministic and

robust case.
numerical evaluation of expected value
(effect on optimal solution)
0.12
length
beam 9.5
0.1 depth
draft 9
unit transportation cost [GBP/ton]

block coefficient
cruise speed
0.08 8.5
relative error

0.06 8

7.5
0.04
7
0.02
6.5
0 6

-0.02 5.5
deterministic solution
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 robust solution
5
1/n (Gauss-Legendre)
1000 1500 2000 2500 3000 3500 4000 4500 5000
Fig. 5. Convergence analysis for integral evaluations through Gauss–Legendre round trip [nm]
quadrature: relative error for the optimal point coordinates as a function of 1/n
(n, number of Gaussian abscissas). Fig. 7. Problem #2: performance parametric analysis for optimal solutions.
 ARTICLE IN PRESS
972 M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977

Table 6 uncertain port handling rate

Optimal parameters for problem #3 (uncertain fuel price). true pareto front (numerical, 500,000*n f.evals)
MODPSO (penalty function)
NSGA-II-aJG (penalty function)
# Variable/function Unit Deterministic solution Robust solution

standard deviation of unit transp.

NSGA-II-v1.1 real coded (constraints handling)

1 Length m 182.79 182.79

2 Beam m 30.60 30.60

cost [GBP/ton]
3 Depth m 15.89 15.89
5,000*n function evaluations
4 Draft m 11.90 11.90
5 Block coefficient 0.67 0.67 1,000*n function evaluations
6 Cruise speed knots 14.00 14.00

200*n function evaluations

f ½xopt , mðyÞ GBP/ton 7.22 7.22
m½f ðxopt ,yÞ GBP/ton 7.22 7.22

s½f ðxopt ,yÞ GBP/ton 0.17 0.17

Optimization objectives highlighted in bold. expected value of unit transp. cost [GBP/ton]

Fig. 10. Problem #4: multi-objective optimization solution—a comparison

uncertain fuel price between different algorithms, varying the number of function evaluations.

1 deterministic solution
design variable non-dimensional value

robust solution
uncertain port handling rate
true pareto front (numerical)
cruise speed

standard deviation of unit transp. cost

0.5 (1) MODPSO solution (penalty function)
block coeff.

3 NSGA-II-aJG solution (penalty function)

NSGA-II-v1.1 real coded (constraints handling)
length

beam

depth

draft

0 2.5
[GBP/ton]

(3)

-0.5 2

-1 1.5

1 2 3 4 5 6 (2)
design variable id. 1
8 9 10 11 12 13 14 15 16
Fig. 8. Problem #3: non-dimensional optimal variables for deterministic and
robust case.
expected value of unit transp. cost [GBP/ton]

Fig. 11. Problem #4: multi-objective optimization solutions (1000n evaluations, n,

number of variables)—(1) min. m, (2) min. s, (3) min. m þ s.
7.6
unit transportation cost [GBP/ton]

7.5 as the expectation of the unit transportation cost and as its

standard deviation, with respect to the uncertain parameters
7.4 variation. Again, the uncertain parameters are considered one at
time and the constraints are treated in the worst case.
7.3 Fig. 10 shows a comparison between the solutions for problem
#4 (uncertain port handling rate), obtained using different
7.2 multi-objective optimization algorithm and assuming a different
number of objective evaluations. Specifically, the MODPSO algo-
rithm of Appendix B is compared with an implementation of the
7.1
NSGA-II-aJG algorithm (Deb et al., 2002), managing the same
objectives and including the same penalty factor. The results
7
obtained using the NSGA-II real-coded implementation of the
deterministic solution
robust solution algorithm, able to manage explicitly the constraints—see again
6.9
Deb et al. (2002), are also depicted and compared with the
60 80 100 120 140
true Pareto front obtained using the numerical procedure of
fuel price [GBP/ton] Appendix C. In the results that follow, 1000n function evaluations
Fig. 9. Problem #3: performance parametric analysis for optimal solutions. are assumed, where n ¼6 is the number of design variables, and
the MODPSO algorithm is assumed to give an acceptable set of
solutions within this computational effort. Specifically, Fig. 11
considerations made for the previous case (problem #2) apply, being shows the Pareto front obtained using 1000n function evalua-
the dependence of the transportation cost on the fuel price linear. tions. Three solutions are chosen among the others. Particularly,
In problems #4, #5 and #6, two objectives are taken into the configurations, respectively, characterized by minimum cost
consideration and a multi-objective minimization algorithm is expectation (1), by minimum cost standard deviation (2), and by
applied. Specifically, the two objectives are, respectively, defined minimum sum of expectation and standard deviation (3), are
 ARTICLE IN PRESS
M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977 973

indicated. The latter configuration represents a compromise are summarized in Table 7 and depicted in Fig. 12, where the
between minimum cost expectation (Bayesian solution) and non-dimensional values are normalized from  1 to +1. Fig. 13
minimum standard deviation (robust solution in a strict sense). shows the unit transportation cost as a function of the port
Generally, the compromise between Bayesian and robust solu- handling rate for the three different solutions.
tion may be identified by minimizing the linear combination Similarly, the results for problem #5 (uncertain round trip
mðf Þ þ ksðf Þ,kA N. In this work, a value of k¼1 is used. The optimal distance) are summarized in Table 8 and depicted in Figs. 14, 15,
design variables corresponding to configurations (1), (2) and (3) and 16. The results for problem #6 (uncertain fuel price) are
shown in Table 9, and in Figs. 17, 18, and 19.
Table 7
Optimal parameters from Pareto-front choices for problem #4 (uncertain port Table 8
handling rate). Optimal parameters from Pareto-front choices for problem #5 (uncertain round
trip).
# Variable or Unit min. mðf Þ min. sðf Þ min. mðf Þ þ sðf Þ
function solution (1) solution (2) solution (3) # Variable or Unit min. mðf Þ min. sðf Þ min. mðf Þþ sðf Þ
function solution (1) solution (2) solution (3)
1 Length m 168.66 100.54 140.82
2 Beam m 27.40 13.06 22.70 1 Length m 191.45 404.50 198.34
3 Depth m 13.96 6.65 11.63 2 Beam m 31.85 61.64 32.82
4 Draft m 10.54 5.28 8.87 3 Depth m 16.19 29.58 16.91
5 Block coefficient 0.64 0.64 0.64 4 Draft m 12.13 21.51 12.53
6 Cruise speed knots 14.24 14.01 14.09 5 Block coefficient 0.67 0.70 0.67
6 Cruise speed knots 14.37 15.38 14.30

m½f ðxopt ,yÞ GBP/ton 8.60 15.06 8.98

s½f ðxopt ,yÞ GBP/ton 2.92 1.13 2.20 m½f ðxopt ,yÞ GBP/ton 7.33 13.88 7.36
s½f ðxopt ,yÞ GBP/ton 1.10 0.47 1.04

uncertain port handling rate

uncertain round trip miles
design variable non-dimensional value

1 min. mean solution (1)

standard deviation of unit annual transp.

min. std. deviation solution (2) 1.3

min. sum solution (3) true pareto front (numerical)
MOPSO solution (penalty function)
1.2 NSGA-II-aJG (penalty function)
(1) NSGA-II-v1.1 real coded (constraints handling)
cruise speed
block coeff.

0.5 1.1 (3)

cost [GBP/ton year]
length

beam

depth

1
draft

0 0.9
0.8

-0.5 0.7
0.6
0.5 (2)
-1
0.4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
design variable id. expected value of unit annual transp. cost [GBP/ton year]
Fig. 12. Problem #4: non-dimensional optimal variables. Fig. 14. Problem #5: multi-objective optimization solutions (1000n evaluations, n,
number of variables)—(1) min. m, (2) min. s, (3) min. m þ s.

22 uncertain round trip miles

min. mean solution (1)
min. std. deviation solution (2)
min. sum solution (3) 1 min. mean solution (1)
design variable non-dimensional value

20
unit transportation cost [GBP/ton]

min. std. deviation solution (2)

min. sum solution (3)
cruise speed

18
block coeff.
length

beam

0.5
16

14
0
12

10 -0.5
depth

draft

8
-1
6
2000 4000 6000 8000 10000 1 2 3 4 5 6
port handling rate [ton/day] design variable id.

Fig. 13. Problem #4: performance parametric analysis for optimal solutions. Fig. 15. Problem #5: non-dimensional optimal variables.
 ARTICLE IN PRESS
974 M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977

min. mean solution (1) uncertain fuel price

16 min. std. deviation solution (2)
min. sum solution (3)
unit transportation cost [GBP/ton]

1 min. mean solution (1)

design variable non-dimensional value

min. std. deviation solution (2)
min. sum solution (3)
14

cruise speed
block coeff.
0.5
12

length

beam
10 0

8
-0.5

depth

draft
6
-1
1000 1500 2000 2500 3000 3500 4000 4500 5000
1 2 3 4 5 6
round trip [nm]
design variable id.
Fig. 16. Problem #5: performance parametric analysis for optimal solutions.
Fig. 18. Problem #6: non-dimensional optimal variables.

Table 9
Optimal parameters from Pareto-front choices for problem #6 (uncertain fuel min. mean solution (1)
price). 12 min. std. deviation solution (2)

unit transportation cost [GBP/ton]

min. sum solution (3)

# Variable or Unit min. mðf Þ min. sðf Þ min. mðf Þþ sðf Þ

function solution (1) solution (2) solution (3) 11

1 Length m 193.88 349.511 196.85

2 Beam m 32.46 57.99 33.06 10
3 Depth m 16.62 27.60 16.87
4 Draft m 12.49 20.01 12.56
5 Block coefficient 0.65 0.67 0.65 9
6 Cruise speed knots 14.05 14.10 14.07

8
m½f ðxopt ,yÞ GBP/ton 7.27 11.41 7.29
s½f ðxopt ,yÞ GBP/ton 0.18 0.07 0.16
7

uncertain fuel price 60 80 100 120 140

0.2 fuel price [GBP/ton]
standard deviation of unit transp. cost

true pareto front (numerical)

(1) MODPSO solution (penalty function)
0.18 NSGA-II-aJG solution (penalty function)
NSGA-II-v1.1 real coded (constraints handling) Fig. 19. Problem #6: performance parametric analysis for optimal solutions.
(3)
0.16
the fuel price are considered, the minimum cost-standard-
0.14
[GBP/ton]

deviation carriers are significantly bigger than those characterized

0.12 by minimum cost expectation (Tables 8 and 9). This is essentially
due to the fact that, throughout the analysis model used, the
0.1 overall transportation cost of the two bigger ships has a small
sensitivity to the travel cost (which is directly related to fuel price
0.08 (2) and round trip distance). Moreover, the compromise solutions are
0.06 very close to the minimum cost-expectation configurations. This
is due to the fact that the standard deviation is small with respect
0.04
to the mean value. The latter has, therefore, a greater impact
7 8 9 10 11 12 13 14
when choosing the compromise solutions.
expected value of unit transp. cost [GBP/ton] It is also worth noting that the minimum cost-expectation
Fig. 17. Problem #6: multi-objective optimization solutions (1000n evaluations, n, solutions of the multi-objective problems #4, #5, and #6 (Tables 7,
number of variables)—(1) min. m, (2) min. s, (3) min. m þ s. 8, and 9) are theoretically the same of problems #1, #2, and #3
(robust solutions, see Tables 4, 5, and 6), respectively. The
differences occur because of the numerical implementation and
consequent convergence of the single- and multi-objective
Looking at the optimal configurations, some consideration can
optimization algorithms.
be made. When the uncertainty related to the port handling rate
is taken into account (problem #4), the minimum cost-standard-
deviation carrier is smaller than that ensuring the minimum
expectation of the cost (Table 7). Moreover, the ‘‘compromise’’ 6. Concluding remarks
solution (minimum sum of cost expectation and standard
deviation) has an average size between the latter configurations. A formulation for robust design optimization (RDO) has been
When the uncertainties related to the round trip distance and to presented and applied to the conceptual design of a bulk carrier
 ARTICLE IN PRESS
M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977 975

under uncertain usage. A particle swarm optimization (PSO) metacentric radius ¼ BMT ¼(0.085CB  0.002)B2/(TCB)
technique has been used to solve the single-objective minimiza- vert. center of gravity ¼ KG ¼ 1.0 + 0.52D
tion problem, being the expectation of the unit transportation
cost the function to minimize. An extension of the PSO algorithm Constraints
L=B Z 6
to handle multi-objective optimization problems (MODPSO) has
been also applied, being the expectation and the standard devia- L=D r15
L=T r 19
tion of the transportation cost the functions to minimize.
The numerical results show how the present approach may be T r 0:45 DWT0.31
T r 0:7D þ0:7
used to consider the uncertainty in the design process. It is also
shown how the robust formulation for optimal design leads to 25,000 rdwt r 500,000
final solutions with minimum value for the expected cost and/or Fn r0:32
minimum standard deviation from the mean. gmt ¼ kbþ bmtkg r0:07B

Acknowledgement
Appendix B. Particle swarm optimization
This work has been partially supported by the U.S. Office of
Naval Research, NICOP Grant no. N00014-08-1-0491, through
Particle swarm optimization (PSO) is an evolutionary algo-
Dr. Ki-Han Kim.
rithm introduced by Kennedy and Eberhart (1995). A swarm is
simulated in terms of the social behavior of the individuals,
Appendix A. Bulk carrier conceptual design tool sharing information among them while exploring the design
variables space. The particles have an adaptable velocity, and
In this appendix, the bulk carrier conceptual design tool by investigate the design space analyzing their own flying experi-
Sen and Yang (1998) is recalled in the form presented by Parsons ence, and that of all the other particles in the swarm.
and Scott (2004). PSO techniques are not (dramatically) influenced by the non-
smoothness of the objective functions and by the Pareto front
Independent variables shape, do not require derivatives and are easy to implement
L¼ length (m) and to parallelize. Moreover they are able to find the global
B ¼ beam (m) minimum in a given domain. Since its first implemen-
D¼ depth (m) tation, several enhancements have been proposed (Venter and
T ¼ draft (m) Sobieszczanski-Sobieski, 2003; Campana et al., 2009), demon-
CB ¼ block coefficient strating its applicability to solve a wide range of engineering
Vk ¼ speed (knots) problems.
In the following subsections, the reader is provided with an
Model
overview of single- and multi-objective deterministic particle
annual cost ¼ capital costs + running costs + voyage costs
swarm optimization techniques.
capital costs ¼ 0.2 ship cost
ship cost ¼ 1.3 (2000W0.85
s + 3500Wo + 2400P0.8)
1:7 0:7 0:4 0:5
steel weight ¼ Ws ¼ 0:034L B D CB B.1. Single-objective deterministic particle swarm optimization
0:8 0:6 0:3
outfit weight ¼ Wo ¼ L B D CB0:1
0:9 The single objective deterministic PSO algorithm (SODPSO)
machinery weight ¼ Wm ¼ 0:17P
may be summarized as follows. A set of particles is distributed
displacement ¼ 1.025LBTCB
into the design variable domain, traveling with a personal speed
power ¼ P ¼ displacement2/3V3k /(a + bFn)
vk each, where k represents the iteration index. Within single
Froude number ¼ Fn ¼ V/(gL)0.5
objective optimization problems, each particle remembers the
V ¼ 0.5144Vk m/s; g ¼ 9.8065 m/s2
best position ever visited pkj and the corresponding objective
a ¼ 4977.06C2B  8105.61CB + 4456.51
function value. All the particle share the information on their best
b ¼ 10,847.2 C2B + 12,817CB  6960.32
positions and objective values, and the current overall best
running costs ¼ 40,000 DWT0.3
position pg is known by all the particles. Based on this
deadweight ¼ DWT ¼ displacement  light ship weight
information, each particle updates its personal speed:
light ship weight ¼ Ws þ Wo þ Wm
voyage costs ¼ (fuel cost + port cost) RTPA
fuel cost ¼ 1.05 daily consumpt.  sea days  fuel price vjk þ 1 ¼ w½wk vkj þ cj rj ðpkj xkj Þ þcg rg ðpkg xkj Þ ðB:1Þ
daily consumption ¼ 0.19P 24/1000 + 0.2
sea days ¼ round trip miles/24Vk xkj þ 1 ¼ xkj þvkj þ 1 ðB:2Þ
round trip miles ¼ 5000 (nm)
fuel price ¼ 100 (GBP/ton) where w,wk ,cj ,rj ,cg ,rg A R þ . In the original proposal of the PSO
port cost ¼ 6.3 DWT0.8 method (Kennedy and Eberhart, 1995), rj and rg are random
round trips per year ¼ RTPA ¼ 350/(sea days + port days) coefficients. In the present deterministic implementation of the
port days ¼ 2[(cargo deadweight/handling rate) + 0.5] algorithm it is rj ¼rg ¼1. Moreover, in the present work, we use
cargo deadweight ¼ DWT—fuel carried—miscellaneous DWT w ¼ 1:0,wk ¼ 0:729,8k, and cj ¼cg ¼1.494, with a number of 24
fuel carried ¼ daily consumption (sea days + 5) particles and 250 maximum iterations.
miscellaneous DWT ¼ 2.0 DWT0.5 Furthermore, half of the particles is initially distributed inside
handling rate ¼ 8000 (ton/day) the domain following a Hammersley sequence (Kuipers and
annual cargo capacity ¼ DWT  round trips per year Niederreiter, 1974), whereas the other half is spread on the
unit transportation cost ¼ annual cost/annual cargo capacity variables domain bounds (following again an Hammersley
vert. center of buoyancy ¼ KB ¼ 0.53 T distribution). The initial speed is null for every particle.
 ARTICLE IN PRESS
976 M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977

B.2. Multi-objective deterministic particle swarm optimization Considering the inequality constraints of the minimization
problem in Eqs. (9), the following penalty parameters are defined:
In the multi-objective deterministic PSO method (MODPSO), 8
< supfgðx,yÞg if supfgðx,yÞg 4 0
the original formulation of the algorithm is revised. In Pinto et al. pn ðxÞ ¼ y A B yAB ðB:3Þ
(2004) modifications to the basic algorithm were introduced :0 otherwise
involving initialization of the particles and random coefficients.
Furthermore, a new formulation able to cover the Pareto optimal thus defining the total penalty function of the q-th order
front with a reduced computational effort is presented in 1X
N

Campana et al. (2007). In order to solve the optimization problem ptot ðxÞ ¼ pqn ðxÞ ðB:4Þ
en¼1
covering the Pareto optimal front, MODPSO adopts a swarm
subdivision technique which defines different guides for the where e A R þ and q A N. In this work, e ¼ 102 and q¼2, hence
particles of each small sub-swarm. The number of sub-swarms is defining a quadratic penalty function.
then linked to the number of Pareto optimal solutions, and Finally, the minimization problem is solved taking into account
potentially it is equal to P  . In fact, no limit on the archive size is the merit factors (or pseudo-objectives) f^ 1 ðxÞ :¼ f1 ðxÞ þ ptot ðxÞ and
imposed, so that every Pareto optimal solution is a possible guide f^ 2 ðxÞ :¼ f2 ðxÞ þ ptot ðxÞ, see Section 3.
for a sub-swarm. The initial swarm S is hence decomposed into
smaller sub-swarms, each one following a proper guide. The main
steps of the sub-swarm decomposition are: Appendix C. Numerical evaluation of the true Pareto front

 distance evaluation: the i-th particle evaluates its distance, in In this appendix, the procedure used to evaluate numerically
the objective functions space, from all the N Pareto points the ‘‘true’’ Pareto front is briefly presented. The algorithm starts
known at the current iteration; from a parameter space investigation—PSI, Statnikov and Matu-
 guide selection: the i-th particle selects the closest Pareto sov (1995)—and then ‘‘refines’’ the sampled Pareto set to improve
solution as its guide. the solution. Unfeasible points are disregarded in the analysis.
A similar approach may be found in Peri and Campana (2003)
where a gradient-based method is used to refine the solution.
By applying this procedure at each iteration the original swarm
Here, a derivative-free approach, based on the evaluation of the
S is divided into n new smaller sub-swarms Sn, where n rN, with
objectives in the vertices and facet-centroids of a hypercube built
N the current number of the Pareto optimal solutions. Therefore,
around the Pareto set sample, is used instead. Specifically, the
each sub-warm Sn has a guide which is different from all the other
overall procedure can be summarized as follows.
n  1 sub-swarms guides, as illustrated in Fig. 20. Once all the sub-
swarms have selected their own guide, each sub-swarm follows Feasible Pareto front refining
its own guide by using the classical equations for updating both Begin
velocity and position of each particles. Details are given in Pinto 1. Sample the variables domain A and find 0 o K rKmax
et al. (2004) and Campana et al. (2007). feasible Pareto points, xk ;
2. If K 4Kmax , then chose Kmax Pareto points (sub-sampling);
update K :¼ Kmax and update the Pareto points set, xk ;
B.3. Penalty function 3. for k ¼1,y,K do
for i¼1,y,n do
In this work, the constrained minimization problem is handled
fDjðiÞ g ¼ fedij g
using a penalty function method. It may be noted that no equality
constraints are involved in the present numerical problem (see x ¼ xk 7 DðiÞ
Appendix A). evaluate fðxÞ and store f ¼: f h , if x ¼: xh is feasible
end
end;
4. Evaluate the Pareto front from the feasible points xh , and
update the current set of Pareto points, xk ;

f2 P (f1,max; f2,max)
s1 a
a
a a

s2

s3
s4
s5

f1
Fig. 20. Sub-swarm decomposition: pn are the Pareto optimal solutions, i is the
particle index, S is the initial swarm and the Sn are the sub-swarms. Fig. 21. Sub-sampling for the Pareto front.
 ARTICLE IN PRESS
M. Diez, D. Peri / Ocean Engineering 37 (2010) 966–977 977

5. Update (if due) Kmax and e; Kennedy, J., Eberhart, R., 1995. Particle swarm optimization. In: Proceedings of the
6. Repeat actions 2 through 5 until convergence; IEEE International Conference on Neural Networks, Perth, Australia.
Kodiyalam, S., Sobieszczanski-Sobieski, J., 2001. Multidisciplinary design optimi-
zation—some formal methods, framework requirements, and application to
Stop.
vehicle design. International Journal of Vehicle Design 25 (1–2) (special
issue).
The Pareto front sub-sampling procedure of action 2 is shown Kuipers, L., Niederreiter, H., 1974. Uniform Distribution of Sequences. John Wiley
in Fig. 21. Specifically, the point P has coordinates (f1,max;f2,max) and Sons, Inc., New York, USA.
Marczyk, J., 2000. Stochastic multidisciplinary improvement: beyond optimiza-
and the angle a is defined as a :¼ p=2ðKmax 1Þ. The Pareto points tion. In: AIAA Paper 2000-4929, Proceedings of 8th AIAA/USAF/NASA/
chosen for the sub-sampling are those with the minimum ISSMO Symposium on Multidisciplinary Analysis and Optimization, Long
distance from the straight lines sk. Beach, USA.
Mirzaei, M., Hosseini, S.N., Roshanian, J., 2007. Single and multi-point optimization
In this work, the above procedure is accomplished varying Kmax of an airfoil using gradient method. Aircraft Engineering and Aerospace
(from 10 to 100) and e (from 0:1Jxupp xlow J to 0:01Jxupp xlow J), Technology 70 (6).
and using a total of about 3,000,000 evaluations for each function. Morino, L., Bernardini, G., Mastroddi, F., 2006. Multi-disciplinary optimization for
the conceptual design of innovative aircraft configurations. CMES—Computer
Modeling in Engineering and Sciences 13 (1).
References Nocedal, J., Wright, W., 1999. Numerical Optimization. Springer-Verlag, New York.
Padula, S.L., Gumbert, C.R., Li, W., 2006. Aerospace applications of optimization
under uncertainty. Optimization and Engineering 7.
Alexandrov, N.M., Lewis, R.M., 2002. Analytical and computational aspects of Papanikolaou, A., 2009. Holistic ship design optimization. CAD Computer Aided
collaborative optimization for multidisciplinary design. AIAA Journal 40 (2). Design, article in press, doi:10.1016/j.cad.2009.07.002.
Agarwal, H., 2004. Reliability based design optimization: formulations and Park, G.J., Lee, T.H., Wang, K.H., 2006. Robust design: an overview. AIAA Journal 44 (1).
methodologies. Ph.D. Thesis, University of Notre Dame, South Bend, IN. Parsons, M.G., Scott, R.L., 2004. Formulation of multicriterion design optimization
Agarwal, H., Renaud, J., 2004. Reliability based design optimization using response problems for solution with scalar numerical optimization methods. Journal of
surfaces in application to multidisciplinary systems. Engineering Optimization Ship Research 48 (1).
36 (3). Peri, D., Campana, E.F., 2003. Multidisciplinary design optimization of a naval
Baumal, A.E., McPhee, J.J., Calamai, P.H., 1998. Application of genetic algorithms to surface combatant. Journal of Ship Research 47 (1).
the design optimization of an active vehicle suspension system. Computer Peri, D., Campana, E.F., 2005. High-fidelity models in global optimization. In:
Methods in Applied Mechanics and Engineering 163 (1–4). Lecture Notes in Computer Science, vol. 3478.
Beyer, H.G., Sendhofff, B., 2007. Robust optimization—a comprehensive survey. Pinto, A., Peri, D., Campana, E.F., 2004. Global optimization algorithms in naval
Computational Methods in Applied Mechanics and Engineering 196. hydrodynamics. Ship Technology Research 51 (3).
Campana, E.F., Peri, D., Pinto, A., 2007. Multiobjective optimization of a containership Ray, T., Gokarn, R.P., Sha, O.P., 1995. A global optimization model for ship design.
using deterministic particle swarm optimization. Journal of Ship Research 51 (3). Computers in Industry 26 (2).
Campana, E.F., Liuzzi, G., Lucidi, S., Peri, D., Piccialli, V., Pinto, A., 2009. New global Reuther, J.J., Jameson, A., Alonso, J.J., Rimlinger, M.J., Saunders, D., 1999a.
optimization methods for ship design problems. Optimization and Engineering Constrained multipoint aerodynamic shape optimization using an adjoint
10 (4). formulation and parallel computers, part 1. Journal of Aircraft 36 (1).
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multiobjective Reuther, J.J., Jameson, A., Alonso, J.J., Rimlinger, M.J., Saunders, D., 1999b.
genetic algorithm: Nsga-II. IEEE Transactions on Evolutionary Computation 6 (2). Constrained multipoint aerodynamic shape optimization using an adjoint
De Groot, M.H., 1970. Optimal Statistical Decisions. McGraw-Hill, New York. formulation and parallel computers, part 2. Journal of Aircraft 36 (1).
Diez, M., Peri, D., 2009. Global optimization algorithms for robust optimization in Sen, P., Yang, J.B., 1998. Multiple Criteria Decision Support in Engineering Design.
naval design. In: 8th International Conference on Computer Applications and Springer, London, UK.
Information Technology in the Maritime Industries, COMPIT 2009, Budapest, Sobieszczanski-Sobieski, J., Haftka, R.T., 1997. Multidisciplinary aerospace design
Hungary. optimization: survey of recent developments. Structural Optimization 14 (1).
Diez, M., Iemma, U., 2007. Robust optimization of aircraft life-cycle costs including Statnikov, R.B., Matusov, J.B., 1995. Multicriteria Optimization and Engineering.
the cost of community noise. In: AIAA Paper 2007-3668, Thirteenth AIAA/CEAS Chapman & Hall, New York.
Aeroacoustics Conference, Rome, Italy. Sues, R.H., Cesare, M.A., Pageau, S., Wu, J.Y.-T., 2001. Reliability-based optimization
Du, X., Chen, W., 2000a. Methodology for managing the effect of uncertainty in considering manufacturing and operational uncertainties. Journal of Aerospace
simulation-based design. AIAA Journal 38 (8). Engineering 14 (4).
Du, X., Chen, W., 2000b. Towards a better understanding of modeling feasibility Taguchi, G., 1986. Quality Engineering through Design Optimization. Krauss
robustness in engineering design. ASME Journal of Mechanical Design 122 (2). International Publications, White Plains, NY.
Hart, C.G., Vlahopoulos, N., 2009. An integrated multidisciplinary particle swarm Trosset, M.W., Alexandrov, N.M., Watson, L.T., 2003. New methods for robust
optimization approach to conceptual ship design. Structural and Multi- design using computer simulations. In: Proceedings of the Section on Physical
disciplinary Optimization article in press, doi:10.1007/s00158-009-0414-0. and Engineering Sciences, American Statistical Association.
Hicks, R.M., Henne, P.A., 1978. Wing design by numerical optimization. Journal of Tu, J., Choi, K.K., Park, Y.H., 1999. A new study on reliability-based design
Aircraft 15 (7). optimization. Journal of Mechanical Design 121 (4).
Iemma, U., Diez, M., 2006. Optimal conceptual design of aircraft including Venter, G., Sobieszczanski-Sobieski, J., 2003. Particle swarm optimization. AIAA
community noise prediction. In: Collection of Technical Papers—12th AIAA/ Journal 41 (8).
CEAS Aeroacoustics Conference, 5. Willcox, K., Wakayama, S., 2003. Simultaneous optimization of a multiple-aircraft
Jameson, A., Leoviriyakit, K., Shankaran, S., 2007. Multi-point aero-structural family. Journal of Aircraft 40 (4).
optimization of wings including planform variations. In: 45th Aerospace Zang, C., Friswell, M.I., Mottershead, J.E., 2005. A review of robust optimal design
Sciences Meeting and Exhibit, Reno, NV. and its application in dynamics. Computers and Structures 83.