You are on page 1of 23

Applied Soft Computing 6 (2006) 221–243

www.elsevier.com/locate/asoc

An integrated multi-objective optimisation and fuzzy


multi-attributive group decision-making technique for
subdivision arrangement of Ro–Ro vessels
A.İ. Ölçer *, C. Tuzcu, O. Turan
The Ship Stability Research Centre, Department of Naval Architecture and Marine Engineering,
Universities of Glasgow and Strathclyde, 100 Montrose Street, Henry Dyer Building, Glasgow G4 0LZ, Scotland, UK
Received 6 November 2003; received in revised form 7 December 2004; accepted 18 January 2005

Abstract

This paper focuses on developing a new multiple criteria decision-making (MCDM) methodology for subdivision
arrangement of Ro–Ro vessels, which integrates the multi-objective optimisation with a fuzzy multi-attributive group
decision-making (FMAGDM) technique. The study concentrates on the task of finding and then evaluating (or ranking) the
finite number of pareto-optimal design alternatives (PODAs). A genetic algorithm based multi-objective optimisation technique,
namely MOGA, is employed for optimisation purpose in terms of chosen design parameters such as damage stability,
survivability, and cargo capacity. MOGA is the methodology where the solution space is searched for a set of PODAs, from
which experts can express their opinions and choose the best PODA. The subjectiveness and imprecision of the ranking process
is modelled as linear trapezoidal fuzzy numbers by means of linguistic terms. An attribute based aggregation technique for
homogeneous and heterogeneous groups of experts is employed and used for dealing with the fuzzy opinion aggregation.
Finally, a real case study on the subdivision arrangement of a Ro–Ro vessel is conducted to illustrate the effectiveness of the
approach.
# 2005 Elsevier B.V. All rights reserved.

Keywords: Multi-objective optimisation; Genetic algorithms; Pareto-optimal designs; Fuzzy sets; Group decision-making; Aggregation;
Consensus

1. Introduction designs, knowing that, nearly all types of ships are


subject to the risk of sinking or capsizing if they lose
Naval architects or ship designers face the problem their watertight integrity, whether by collision,
of providing an adequate level of safety for their grounding or an internal accident such as fire or
explosion. To this end the most effective protection is
* Corresponding author. Tel.: +44 141 5484047;
provided by means of the internal subdivision
fax: +44 141 5484784. arrangement of the vessel. It is intuitive that the
E-mail address: olcer@na-me.ac.uk (A.İ. Ölçer). subdivision is a safeguard for any possible damage

1568-4946/$ – see front matter # 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2005.01.004
222 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

that may breach the watertight integrity of the hull. capable of finding the global optima when the search
Even though the ship might lose her sailing functions space has local extremes. In addition, as they can just
as a result of an accident, she should remain floating in locate a single optimum, several runs are required to
reasonably high sea states and should continue to generate the set of PODAs. On the contrary, recent
accommodate safely the people on board, as well as studies on optimisation techniques based on genetic
preventing the leakage of environmentally hazardous and evolutionary algorithms have shown that these
substances into the sea. methods can be used efficiently to eliminate most of
There is an increasing awareness of the safety the difficulties of classical methods because of their
aspect as well as a developing will to implement the flexibility, ease of operation, minimal requirements
procedures by which the safety aspect can be assessed and global perspective [4]. The detailed recent reviews
and be implemented at the earliest possible design on multi-objective optimisation using many evolu-
stages. The study presented in this paper sets out a tionary algorithms can be found in [5,6].
good example for this as the internal subdivision All solutions lying on the pareto-optimal set are
arrangement is directly linked with damage stability potentially preferred by the ship designers and the
and survivability performance of the ships. choice of one solution over the other entails additional
There are a few empirical studies concerning knowledge of the problem such as designer prefer-
subdivision arrangement of a vessel. Sen and Gerigk ences. These preferences may be elicited a priori,
[1] dealt with some aspects of a knowledge-based interactively or a posteriori. Multiple attribute
system for intelligent arrangement of internal spaces decision-making (MADM) techniques, which are to
for ship safety at the preliminary stage of design. Ölçer determine the best design alternative with the highest
et al. [2] studied the subdivision arrangement problem degree of desirability with respect to all relevant
and evaluated PODAs in a totally crisp environment attributes, are usually required in posterior evaluation
where all the parameters are deterministic. Turan et al. of PODAs to choose one of them for implementation.
[3] presented a case based reasoning approach for The classical (or crisp) MADM techniques assume
internal subdivision problem of a ROPAX vessel. all performance ratings are crisp numbers. However,
In this study, internal subdivision arrangement of these ratings might involve subjectiveness, impreci-
Ro–Ro vessels is considered as an MCDM problem in sion and vagueness. For example, they can be
which PODAs are conducted by using a Genetic linguistic terms such as good, medium, fair, etc. or
Algorithms (GAs) based multi-objective optimisation verbal assessments such as ‘‘the sixth PODA is much
technique and ranking of PODAs is evaluated in an better than the eleventh PODA’’. Moreover, in order to
FMAGDM environment where available information find the best compromise PODA amongst competing
is imprecise and multiple experts are involved. attributes of the subdivision arrangement problem of
In reality, the internal hull subdivision arrangement Ro–Ro vessels, involvement of multiple experts and
of Ro–Ro vessels entails the achievement of several their subjective opinions are also required. They often
different objectives, which are often conflicting and cannot have a consensus with respect to an attribute
non-commensurable, such as improving survivability due to different preferences and their opinions usually
and increasing cargo capacity. This makes the subdivi- conflict.
sion arrangement process suitable for optimisation by Fuzzy multiple attribute decision-making
using multi-objective methods, which yield a family of (FMADM) methods have been developed due to the
non-dominated solutions (also named non-inferior or lack of precision in assessing the performance ratings
admissible) called pareto-optimal set. The concept of of alternatives with respect to an attribute. A large
non-dominance refers to the solutions for which no number of articles on the decision-making analysis
objective can be improved without worsening at least have addressed the FMADM methods [7–11]. Ölçer
one of the other objectives. Thus, the non-dominated and Odabasi [11] reviewed and analysed the most of
solutions are superior to the others with respect to all the known FMADM methods according to their group
objectives, but comparatively good among themselves. decision-making ability and introduced a new
In dealing with multi-objective optimisation FMADM method, which allows the involvement of
problems, classical optimisation methods are not multiple experts.
A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243 223

This paper proposes a new MCDM methodology the feasible design space. The objective functions are the
(see Fig. 1) to take into account the above-mentioned quantities that the designer wishes to minimise,
considerations for the subdivision arrangement of Ro– maximise, or attain a certain value. This problem can
Ro vessels and is organised as follows: The subsequent be formulated as follows:
section outlines the general features of the pareto- Minimise : FðYÞ ¼ f f1 ðYÞ; f 2ðYÞ; . . . ; ft ðYÞg
optimal concept and formulation of the FMAGDM
problem, and then Section 3 presents a new MCDM Subject to : p inequality constraints
approach, which consists of multi-objective optimisa- (1)
tion and ranking of PODAs stages. The paper gd ðyÞ  0; d ¼ 1; . . . ; p
continues with an illustrative example, which demon- q equality constraints hf ðyÞ ¼ 0; f ¼ 1; . . . ; q
strates the application of the new MCDM approach on
the subdivision arrangement of a Ro–Ro vessel. (2)
Finally, the last section details the conclusions and a where Y = [ y1, y2, . . ., yn] is the vector of decision
summary of the characteristics of the proposed variables. In other words, we wish to determine the
system. particular set y1 ; y2 ; . . . ; yn , which yields the optimum
values of all the objective functions among the set of
all points that satisfy Eqs. (1) and (2).
2. Problem formulation and pareto-optimal The most common method of handling multiple
concept objectives, with or without employing evolutionary
algorithms, is to aggregate the multiple objectives and
A general multi-objective optimisation problem evaluate the performance based on a single fitness
(also called multiple criteria optimisation or vector function. In single objective optimisation, optimality
optimisation) is to find the designvariables that optimise can be defined when a solution is able to give the best
a vector objective function ( F(Y) = { f 1, . . ., f t}) over (highest or lowest) fitness value achieved so far.

Fig. 1. Flow chart of the proposed methodology.


224 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

However in solving multi-objective problems, we are are fuzzy assessments, which can be different or
interested in obtaining a set of pareto-optimal solution identical, then this attribute is called ‘‘subjective
points rather than a single solution point. attribute’’.
For a multi-objective optimisation problem, any Subjective and objective attributes can also be
two solutions y1 and y2 can have one of two divided into two classes. The first class is of ‘cost’ (or
possibilities: one dominates the other or none ‘input’) nature (the larger, the less preference). The
dominates the other. In a minimisation problem, second class is of ‘benefit’ (or ‘output’) nature (the
without loss of generality, a solution y1 dominates y2 larger, the more preference), Parkan and Wu [12].
if the following two conditions are satisfied: The general FMAGDM model in which PODAs are
evaluated is described as follows:
8 g 2 f1; 2; . . . ; tg : fg ðy1 Þ  fg ðy2 Þ (3)
The main aim of FMAGDM techniques is to deal
1 2
9 l 2 f1; 2; . . . ; tg : fl ðy Þ < fl ðy Þ (4) with imprecise information and weighted multiple
If any of the above conditions is violated, the solution experts’ fuzzy opinions, which are expressed for
y1 does not dominate the solution y2. If y1 dominates subjective attributes of the ranking problem. Let
the solution y2, y1 is called the non-dominated solu- X = {X jj j = 1, . . ., N} be a finite set of alternatives
tion. The solutions that are non-dominated within the (courses of action, candidates, scenarios, strategies),
entire search space are denoted as pareto-optimal and A = {Aiji = 1, . . ., K} be a finite set of attributes
constitute the pareto-optimal set or pareto-optimal according to which the desirability of an alternative is
frontier. to be judged and E = {Ekjk = 1, . . ., M} be a finite set
Multiple attribute decision analysis of PODAs of experts. And let R = {(Rij)kji = 1, . . ., K; j = 1, . . .,
starts with the generation of the attributes. Determina- N; k = 1, . . ., M} be the K  N decision matrix, where
tion of the attributes is critically important to the final Rij is the performance rating of alternative X j with
ranking and thus should be done very carefully and the respect to attribute Ai. An FMAGDM problem can be
attributes should be complete and exhaustive, contain expressed in a matrix format for each expert as
mutually exclusive items. follows:

There are basically two types of attributes for a The weighting vectors w (for the attributes) and
ranking problem, namely ‘subjective’ and ‘objective’ we (for the experts) are given directly by the
attributes. The main assumption of this research is that moderator (or manager) and are represented as
if an assessment for an alternative with respect to an w = {wi ji = 1, . . ., K} and we = {wekjk = 1, . . ., M}.
attribute is crisp, this kind of attribute is called an Given the decision matrices and the weighting
‘‘objective attribute’’. Such ratings, which are usually vectors, the main aim of the FMAGDM problem is
provided by suppliers and manufacturers, do not to rank PODAs by using a stepwise ranking
change from one expert to the other. When experts’ procedure to determine an overall PODA perfor-
opinions for an alternative with respect to an attribute mance score with respect to all attributes.
A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243 225

3. An integrated multi-objective optimisation and solutions, which are close to each other on the grid.
FMAGDM technique The idea of crossover is that two designs (parents) can
be combined together to form another two solutions
An integrated multi-objective optimisation and (children), which are better than their parents. If this is
FMAGDM technique is proposed for the subdivision the case the children usually replace the parents but
arrangement of Ro–Ro vessels. The working proce- not always. As the search progresses from generation
dure of the proposed methodology is given in Fig. 1. to generation the population slowly converges towards
The proposed method has the following two major the trade-off surface or pareto-optimal frontier. The
stages. In the first stage, a multi-objective optimisation output from the MOGA optimiser is thus a set of non-
model is used as a means to determine the PODAs; in dominated solutions (or a set of PODAs) [15].
the second stage, an FMAGDM methodology is
utilised to rank the PODAs according to both 3.2. Fuzzy multi-attributive group decision-making
predetermined objective and subjective attributes. stage

3.1. Multi-objective optimisation stage This stage uses a new methodology, which was
developed by Ölçer [17]. The main reason of
In this stage, multi-objective design optimisation employing this method is its ability of overcoming
software, namely FRONTIER [13], is used. The the problems mentioned in Section 1 of this article. It
FRONTIER software uses GAs to carry out a multi- allows designers to express their fuzzy opinions for the
objective search. GAs were first pioneered by Holland PODAs with respect to each subjective attribute.
[14] and since then have been widely studied, These opinions are then aggregated with respect to
experimented with and applied in many fields of the each subjective attribute by taking into account
engineering world. GAs are evolutionary optimisation attribute based expert weights. Finally the PODAs
approaches, which simulate a natural evolution are ranked according to their overall aggregated
process based on the Darwinian theory, in which performance ratings. This stage consists of the
the fittest species survive and propagate while the less following three major phases:
successful ones tend to disappear. They are most
appropriate for complex non-linear models where the 3.2.1. Rating phase
location of the global optimum is a difficult task. GAs In this phase, in order to establish the decision
also differ from many other optimisation methods in matrix for each expert, experts express their opinions
the sense that they only use the objective function, not (or performance ratings) for each PODA with respect
derivatives, to identify possible solutions. to each subjective attribute. This can be carried out by
The multi-objective search method used in FRON- questionnaires, which are used for soliciting expert
TIER is an enhanced version of the simple GAs, called opinions for each alternative with respect to each
the Multiple Objective Genetic Algorithm (MOGA) subjective attribute. Because these ratings are gen-
[15,16]. The MOGA randomly generates a population erally in fuzzy data form, fuzzy numbers can model
of individuals. Each individual represents one possible better this kind of qualitative data. This phase aims to
design. The individuals can be assessed on several convert fuzzy data into standardised positive trape-
criteria, which can be specified based on the problem zoidal fuzzy numbers.
objectives. These criteria are maintained separately at If the decision matrix of the problem contains fuzzy
all times, allowing simultaneous optimisation of all data, which may be expressed in linguistic terms, these
criteria during the search. However, instead of finding terms must first be transformed into fuzzy numbers by
a single anear-optimal solution, the MOGA finds a set using the appropriate conversion scale. The linguistic
of solutions, which define the trade-off surface set S, S = {None, Very Poor (VP), Between Very Poor
between these search criteria. The system works by and Poor (P-VP), Poor (P), Fairly Poor (Fa.P), More or
initially generating a random set of solutions and Less Poor (Mo.P), Medium, More or Less Good
evaluating them on all criteria. The solutions are (Mo.G), Fairly Good (Fa.G), Good (G), Between
arranged on a grid. Crossover then occurs between Good and Very Good (G-VG), Very Good (VG),
226 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

Excellent}, is used to systematically convert linguistic by expert Ek is translated into standardised trapezoidal
terms to their corresponding fuzzy numbers. The fuzzy number Rk (k = 1, 2, . . ., M), where
linguistic terms used in this conversion scale, their  
 ak bk c k dk
corresponding representations of fuzzy numbers and Rk ¼ ; ; ; ¼ ðak ; bk ; ck ; dk Þ (5)
membership functions are given in Table 1. Engineers m m m m
and ship designers can employ this conversion system and 0  ak < bk  ck < dk  1. Where m is the max-
because it is simple enough to be understood by the imum value of non-standardised trapezoidal fuzzy
decision maker, and easy to use. numbers given by experts for the same attribute.
The concept of linguistic variable is very useful in
dealing with situations, which are too complex or too 3.2.2. Attribute based aggregation phase
ill defined to be reasonably described in conventional In the second phase, an attribute based aggregation
quantitative expressions [18,19]. A linguistic variable method for homogeneous and heterogeneous (homo/
is a variable whose values are not numbers but words heterogeneous) groups of experts under each sub-
or sentences in a natural or artificial language [20]. jective attribute is employed. Sometimes, one may
If the decision matrix contains non-standardised admit that the various experts are not equally
fuzzy numbers, which may be expressed in such important (or reliable). In such a case, it is called
phrases as ‘‘approximately equal to 11’’, they must be heterogeneous (non-homogeneous) group of experts’
converted to standardised fuzzy numbers. problem and, otherwise, homogeneous group of
Experts’ fuzzy opinions are represented as trape- experts’ problem is considered.
zoidal fuzzy numbers. The reason of using trapezoidal When more than one expert is involved in the
fuzzy numbers is that they are intuitively easy for selection problems, each expert might have a different
decision makers to use. However, the method is weight. For instance, an expert, who is very
independent of the type of membership functions experienced in ship design, may not give as good
being used. Some other membership functions, for an assessment of operational aspects as ship design
example triangular membership functions, are also aspects. Therefore, attribute based expert weighting is
applicable. a necessity and a good method of aggregating multiple
Let U be the universe of discourse, U = [0, m]. A expert opinions must consider the degree of impor-
trapezoidal fuzzy number can be defined by a tance of each expert in the aggregation procedure. One
quadruplet A = (a1, a2, a3, a4). The membership way of modelling this aspect is to consider the
function is existence of a moderator who assigns a weight to each
8 expert. Because the aggregation is based on each
>
> ðx  a1 Þ=ða2  a1 Þ for a1 < x  a2
< subjective attribute, expert weighting is determined
1 for a2  x  a3
mA ðxÞ ¼ separately for each subjective attribute. This yields
>
> ða  xÞ=ða  a Þ for a3  x < a4
: 4 4 3
more accurate and reliable decision-making models.
0 otherwise
The purpose of this phase is to establish an
with a1 < a2  a3 < a4 bearing in mind that if a2 = a3, algorithm to combine homo/heterogeneous groups of
A is a triangular fuzzy number. The x in interval [a2, experts’ opinions to form a group consensus opinion.
a3] gives the maximal grade of mA(x), i.e. mA(x) = 1; it After the weights of attributes and the degree of
is the most probable value of the evaluation data. a1 importance of experts are assigned, under each
and a4 are the lower and upper bounds of the available subjective attribute all performance ratings are
area for the evaluation data, which are used to reflect aggregated for each alternative.
the fuzziness of the evaluation data. In this phase of the proposed approach, the method
Assume that each expert Ek (k = 1, 2, . . ., M) presented in [11,17] is used for dealing with fuzzy
constructs a positive trapezoidal fuzzy number opinion aggregation for homo/heterogeneous groups
Rk = (ak, bk, ck, dk) to represent the performance of experts. The main advantage of using this method is
rating for an alternative with respect to a subjective that the method doesn’t require experts’ assessments
attribute, where 0  ak < bk  ck < dk  m. Each to have a common intersection at a-level cuts, where
trapezoidal fuzzy number Rk = (ak, bk, ck, dk) given a 2 (0,1] [21].
Table 1
Linguistic terms and their corresponding fuzzy numbers and membership functions used in the proposed approach

A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243


227
228 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

Assume that the degree of importance of expert Ek (b) Construct the agreement matrix (AM), after all the
PM 1, 2, . . ., M) is wek, where wek 2 [0,1] and
(k = agreement (or similarity) degrees between experts
k¼1 wek ¼ 1. are measured,
In some cases, the relative importance of experts is 2 3
1 S12 . . . S1v . . . S1M
widely different for each attribute of the problem. 6 ... ...
6 ... ... 77
Therefore, the relative importance and weight of each
AM ¼ 6 6 S u1 S u2 . . . S uv . . . S 7
uM 7
expert is considered. First, the most important person 4 ... ... ... ... 5
is selected among experts and weight one is assigned
SM1 SM2 . . . SMv . . . 1
to him/her, that is, rek = 1. Then the lth expert is
compared with the most important person and a where Suv = S(Ru, Rv ), if u 6¼ v and Suv = 1, if
relative weight for the lth expert rel, l = 1, 2, . . ., M, is u = v. By the definition of S(Ru, Rv ), the diagonal
obtained. So we have max {re1, re2, . . ., reM} = 1 and elements of AM are unity.
min {re1, re2, . . ., reM} > 0. Finally, the degree of (c) Calculate the average degree of agreement
importance wek is defined as: AA(Eu) of expert Eu (u = 1, 2, . . ., M) by using
the AM of the problem, where
rek
wek ¼ PM (6)
1 XM
k¼1 rek AAðEu Þ ¼ SðRu ; Rv Þ (8)
If the importance of each expert is equal, then M  1 v¼1;v 6¼ u
we1 = we2 = . . . = weM = 1/M.
(d) Calculate the relative degree of agreement RA(Eu)
The aggregation algorithm for homo/heteroge-
of expert Eu (u = 1, 2, . . ., M), where
neous groups of experts is presented as follows:
AAðEu Þ
RAðEu Þ ¼ PM : (9)
(a) Calculate the degree of agreement (or degree of
u¼1 AAðEu Þ
similarity) Suv (Ru, Rv ) of the opinions between
each pair of experts Eu and Ev , where Suv (Ru, (e) Calculate the consensus degree coefficient
Rv ) 2 [0,1], 1  u  M, 1  v  M, and u 6¼ v. CC(Eu) of expert Eu (u = 1, 2, . . ., M), where
A method introduced by Chen and Lin [22] is CCðEu Þ ¼ bweu þ ð1  bÞRAðEu Þ (10)
used for measuring the degree of similarity
where b (0  b  1) is a relaxation factor of the
between trapezoidal fuzzy numbers. According
proposed method. It shows the importance of the
to this approach, let A and B be two standardised
weu over RA(Eu). When b = 0, a homogeneous
trapezoidal fuzzy numbers,
group of experts problem is considered. The
A ¼ ða1 ; a2 ; a3 ; a4 Þ and B ¼ ðb1 ; b2 ; b3 ; b4 Þ consensus degree coefficient of each expert is a
where 0  a1  a2  a3  a4  1 and 0  b1  good measure for evaluating the relative worthi-
b2  b3 b4  1. Then the degree of similarity ness of each expert’s opinions.
between the standardised trapezoidal fuzzy (f) Finally, the aggregation result of the fuzzy
numbers A and B can be measured by the opinions is RAG as
similarity function S, RAG ¼
ja1  b1 j þ ja2  b2 j þ ja3  b3 j CCðE1 ÞR1 CCðE2 ÞR2     CCðEM ÞRM ;
þ ja4  b4 j (11)
SðA; BÞ ¼ 1 
4
where operators  and  are the fuzzy multi-
(7)
plication operator and the fuzzy addition operator,
where S(A, B) 2 [0,1]. Larger the value of S(A, B), respectively. The fuzzy multiplication and addi-
greater the similarity between the standardised tion of trapezoidal fuzzy numbers are also
trapezoidal fuzzy numbers A and B. It should be trapezoidal fuzzy numbers [23]. Some important
noted that S(A, B) = S(B, A). properties (e.g. agreement preservation) and their
A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243 229

proofs of the aggregation method can be found in 3.2.3.2. Ranking sub-phase. In this sub-phase, clas-
[21,24,25]. sical MADM methods can be utilised to determine the
ranking order of the PODAs. The TOPSIS method,
3.2.3. Selection phase which gives cardinal order of the alternatives, is used
The selection phase consists of two sub-phases: in this study because of its general and broad accept-
defuzzification, and ranking sub-phases. In order to ability in many problem domains. Hwang and Yoon
rank PODAs, all fuzzy elements of the aggregated [26] developed the TOPSIS method based on the
decision matrices for homo/heterogeneous groups of intuitive principle that the chosen alternative should
experts are deffuzzified in the deffuzzification sub- have the shortest distance from the positive-ideal
phase so that all components of the aggregated solution (PIS) and the longest distance from the
decision matrix are all crisp numbers and any classical negative-ideal solution (NIS).
MADM method can be used. The result of this sub- TOPSIS is quite effective in identifying the best
phase is a decision matrix, which contains only crisp alternative quickly. The underlying logic premise of
(or non-fuzzy) data. Then the PODAs of the problem the TOPSIS method is that an alternative, which is
are ranked by technique for order preference by more like an ideal alternative (the best that could be
similarity to ideal solution (TOPSIS) method in imagined) and more unlike a negative-ideal alternative
ranking sub-phase. (the worst that could be imagined) should be preferred.
In the TOPSIS method, the ideal alternative is
3.2.3.1. Defuzzification sub-phase. This sub-phase constructed out of exclusively the best attribute values
allows fuzzy numbers to be translated to crisp values attainable and therefore it is usually an invented
early on, so that the arithmetic process of the proposed alternative. The negative-ideal alternative is also
approach becomes easy. In this sub-phase, all the usually an invented alternative that is constructed
aggregated fuzzy numbers are transformed into out of exclusively the worst attribute values attainable.
numeric ratings using a fuzzy scoring approach [9] The relative closeness (similarity) of each alternative
as explained below. The reason of using this method is to the ideal alternative is rated on the basis of its
that it is intuitive and easy to implement. The result of distances from both the ideal and the negative-ideal
this sub-phase is a decision matrix, which contains alternatives simultaneously. Finally, the preference
only crisp data. According to the fuzzy scoring method order of the alternatives is obtained by their rank on a
developed in [1], the fuzzy maximising set and descending order of those ratings. The computational
minimising set should be first obtained, which are procedure of the TOPSIS method is quite straight-
defined as: forward.
 According to TOPSIS, the following steps are to be
x; for 0  x  1
mmax ðxÞ ¼ (12) performed:
0; otherwise

1  x; for 0  x  1 (a) Calculate normalised ratings: This step tries to
mmin ðxÞ ¼ (13)
0; otherwise transform various attribute dimensions into the
Then, the right score of fuzzy number B can be non-dimensional attribute, which allows compar-
determined using ison across the attributes. The vector normal-
isation technique is used for computing the
mR ðBÞ ¼ sup½mB ðxÞLmmax ðxÞ (14) element (r ji) of the normalised decision matrix,
x
which is given as
The left score of B can be determined using
x ji
mL ðBÞ ¼ sup½mB ðxÞLmmin ðxÞ (15) r ji ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN 2ffi ;
x
j¼1 x ji (17)
Given the left and right scores of B, the total score of B
can be computed using j ¼ 1; 2; . . . ; N; i ¼ 1; 2; . . . ; K:
mR ðBÞ þ 1  mL ðBÞ where x ji is the value of alternative j with respect
mT ðBÞ ¼ (16) to attribute i.
2
230 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

(b) Calculate weighted normalised ratings: A set of (d) Calculate separation measures: Separation (dis-
attribute weights assessed from the decision tance) between alternatives can be measured by
maker is accommodated to the normalised the n-dimensional Euclidean distance. Separation
decision matrix in this step. The weighted of each alternative from the positive-ideal solution
normalised decision matrix can be calculated by is then given by
multiplying each row of the normalised decision vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u K
matrix with its associated attribute weight wi . An uX
element of the weighted normalised decision S j ¼ t ðv ji  vi Þ2 j ¼ 1; 2; . . . ; N

(22)
i¼1
matrix is calculated as
Similarly, separation from the negative-ideal
v ji ¼ wi r ji ; solution is then given by
(18)
j ¼ 1; 2; . . . ; N; i ¼ 1; 2; . . . ; K vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u K
uX
where wi is the weight of the ith attribute. S j ¼ t ðv ji  v

i Þ
2
j ¼ 1; 2; . . . ; N (23)
There are many methods for assigning attribute i¼1
weights, such as the weighted evaluation techni- (e) Calculate similarities to positive-ideal solution:
que (WET), the eigenvector method, the entropy Relative closeness (or similarity) of A j with
method, and so forth. In the proposed method, respect to A* is defined as
WET is used for finding the attribute weights.
According to WET, the moderator begins by rank Sj
C j ¼ ;
ordering attributes and attribute relative impor- Sj þ Sj (24)
tances are assigned on a 0–100 scale. The attribute
perceived as most important is assigned a weight 0 < Cj < 1; j ¼ 1; 2; . . . ; N
of 100; all other attribute relative importances When C j
is close to 1, the alternative is regarded
are assigned relative to that. The final step of the as ideal; and when Cj is close to 0, the alternative
weighting procedure is to normalise the relative is regarded as non-ideal.
importances {r1, r2, . . ., rK}, to obtain the (f) Rank preference order: Choose an alternative
weights {w1 , w2 , . . ., wK }. The standard normal- with the maximum C j or rank alternatives
isation is according to C j in descending order. It is clear
ri that an alternative A j is closer to A* than to A as
wi ¼ PK ; i ¼ 1; 2; . . . ; K; (19)
C j approaches 1.
i¼1 ri
P
where 0  wi  1 and K i¼1 wi ¼ 1. The following steps required by the proposed appr-
(c) Identify positive-ideal and negative-ideal solu- oach are given to facilitate the presentation of the case
tions: Let the positive-ideal solution, A*, and the study in Section 4.
negative-ideal solution, A, be defined in terms of
the weighted normalised values: Step 1. Constitute optimisation parameters and

A ¼ fv1 ; v2 ; . . . ; vi ; . . . ; vK g constraints.
Step 2. Employ MOGA to generate the feasible
where
design space in multi-objective optimisa-
vi ¼ fmax v ji ; i 2 J1 ; min v ji ; i 2 J2 g (20) tion environment.
j j

A ¼ fv1 ; v2 ; . . . ; vi ; . . . ; v
   Step 3. Carry out design filtering and select
K g;
PODAs using the conditions for domina-
where tion given in Eqs. (3) and (4).
v Step 4. Form a committee of experts (or decision
i ¼ fmin v ji ; i 2 J1 ; max v ji ; i 2 J2 g (21)
j j makers), then identify the selection attri-
where J1 is the set of benefit attributes and J2 is the butes with types (cost or benefit) of them
set of cost attributes. and list all possible PODAs.
A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243 231

Step 5. Collect each expert opinion for each PODA Step 10. Construct the normalised ratings, and
with respect to a subjective attribute and weighted normalised ratings of the deffuz-
establish a decision matrix for each expert. zified matrices by using Eqs. (17) and (18).
Step 6. Transform the fuzzy data (linguistic Step 11. Calculate positive-ideal and negative-ideal
expressions or fuzzy assessments) into solutions, separation measures, and simila-
standardised positive trapezoidal fuzzy rities of each PODA by using Eqs. (20)–
numbers attribute by attribute by using (23), and (24).
Eq. (5). Step 12. Order or rank the PODAs according to the
Step 7. Assign the relative importance of experts Overall Alternative Ranking (OAR) values
and attributes, and then calculate the (or Cj values) and select the PODA with
weights of them by using Eqs. (6) and the maximum OAR value as the best
(19) respectively. PODA.
Step 8. Under each subjective attribute, aggregate
all experts’ fuzzy opinions for each PODA
by using the Eqs. (7)–(10), and (11). This 4. Case study
step gives us aggregated matrices for
homo/heterogeneous groups of experts. The proposed approach outlined in Section 3 is
Step 9. Deffuzzify these matrices by applying applied to the subdivision arrangement of a Ro–Ro
Eqs. (12)–(15), and (16) on every fuzzy vessel, in the sense that damaged stability require-
number in question. Up to this point, the ments are satisfied for better operational flexibility
aggregated decision matrices with fuzzy and more over, survivability performance is im-
elements have been transformed into ones proved for the safety consideration and the cargo
with crisp numbers. capacity is maximised for higher revenue. Car deck

Table 2
Optimisation variables and objectives
No. Variables Type Bounds
Discrete Continuous Lower Upper Increment
1 Car deck height (m) H 9.6 9.9
2 Side-casing width (m) H 1 2 0.5
3 Lower-hold height (from car deck) (m) H 2.6 5.2 2.6
4 Transverse Bulkhead 02 H 25 29 1
5 Transverse Bulkhead 03 H 37 41 1
6 Transverse Bulkhead 04 H 49 53 1
7 Transverse Bulkhead 05 H 61 65 1
8 Transverse Bulkhead 06 H 79 83 1
9 Transverse Bulkhead 07 H 97 101 1
10 Transverse Bulkhead 08 H 115 119 1
11 Transverse Bulkhead 09 H 127 131 1
12 Transverse Bulkhead 10 H 139 143 1
13 Transverse Bulkhead 11 H 151 155 1
14 Transverse Bulkhead 12 H 163 167 1
15 Transverse Bulkhead 13 H 175 179 1
16 Transverse Bulkhead 14 H 187 191 1

No. Objectives Type Description


Bounds for transverse bulkheads are given in frame numbers
1 HS value Maximisation For the worst two compartment damage case
2 KG limiting value Maximisation For the worst two compartment damage case
3 Cargo capacity value Maximisation Expressed in car lanes
232 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

height, lower-hold height, side-casing width, and The Stockholm Agreement can be complied with
layout of the watertight subdivision, which are either by prescriptive water on deck calculations or via
critical for the above considerations, are chosen as model tests to determine the survivability of a damaged
optimisation parameters. These optimisation vari- ship in a given sea-state. For this study, with respect to
ables, which are all discrete types of variables except survivability performance the Static Equivalent Method
for car deck height, and the objectives of the case (SEM) is utilised to determine HS value for the worst
study are given in Table 2. In this Table, increment SOLAS’90 damage determined from damage stability
values assigned for the discrete type of variables are calculations. The SEM (Vassalos [28]) is an empirical
introduced to describe a set of discrete intervals for capsize model for Ro–Ro ships that can predict with
the decision maker to take in the optimisation reasonable accuracy the limiting sea-state for specific
process. damage conditions. The SEM for Ro–Ro ships
The tragic accidents of the Herald of Free Enterprise postulates that the ship capsizes quasi-statistically, as
in 1987 and the Estonia in 1994 initiated a significant a result of an accumulation of a critical mass of water on
surge of research related to the capsizing of Ro–Ro type the vehicle deck or on the weather deck, the height of
ships. This effort culminated in significant develop- whichabove themeanseasurfaceuniquelycharacterises
ments that helped the ferry industry to raise safety the ability of the ship to survive in a given critical sea
levels to demanding new heights cost-effectively, in state. This method was developed following observa-
response to strict new regulations. These regulations tions of the behaviour of damage ship models in waves
require all Ro–Ro vessels to comply with SOLAS’90 and it was validated using several model experiments
standards, as well as with new water on deck standards and a large number of numerical simulations.
known as the Stockholm Agreement IMO [27], which
determines the limiting significant wave height (HS) at 4.1. Multi-objective optimisation stage calculations
which a Ro–Ro vessel survives in a damaged condition. (Steps 1, 2, and 3)
Therefore the main aim of this case study is to
maximise the survivability and damaged stability After the optimisation problem has been estab-
standards as well as to improve the cargo capacity. lished, the optimisation work was modelled in the

Fig. 2. Evolution of the KG limiting value.


A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243 233

Fig. 3. Evolution of the HS value.

FRONTIER software environment. Firstly a para- experiment (or design layout) with respect to each
metric and topological Ro–Ro vessel definition was optimisation parameter. For each design experiment,
created in NAPA [29] (Naval Architectural Software) the relevant adjustments of KG, Draught and
so that the vessel can be modified for each design Displacement have been made during the optimisation

Fig. 4. Evolution of the Cargo capacity.


234 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

Fig. 5. Decision hierarchy of multiple attribute evaluation for PODAs.

process. In addition to these, a computer program problem were articulated and then a new FMAGDM
written in C++ has performed bulkhead distribution. method was employed to find the ranking of the
All the SEM calculations as well as SOLAS’90 PODAs.
calculations were carried out using NAPA. The selection decision is made on the basis of three
Since the problem has three objectives to max- objective and three subjective attributes, which are all
imise, a multi-objective optimiser was selected with benefit type of attributes. These attributes, which are
50 generations. Directional crossover probability was all mutually exclusive and critical for the selection of
0.5, selection probability was 0.05, and mutation best PODA, are the following: A1: Cargo capacity, A2:
probability was 0.1 for the last run, which took HS value, A3: KG limiting value, A4: producibility, A5:
approximately 7 h in a PC (Pentium IV, 1.4 GHz) ease of maintenance and repair, and A6: loading/
environment. At the end of this run, 1269 different unloading efficiency of arrangement. It should be
designs were obtained in design space with 543 of noted that subjective attributes have been introduced
them being unfeasible designs, which do not comply in this stage in order to be able to incorporate
with optimisation constraints. Firstly, all 726 designers’ fuzzy preferences for the PODAs. The
(=1269  543) feasible designs were filtered in design hierarchical structure showing the overall objective,
space to obtain only designs that belong to a pareto- the attributes and alternatives is shown in Fig. 5.
optimal set. This filter marked 78 different good
designs and 6 PODAs. While filtering, convergence 4.2. Rating phase calculations (Steps 4, 5, and 6)
figures of the objectives and other scatter plot
diagrams of the optimisation parameters were taken The PODAs of the case study are evaluated by a
into account (see Figs. 2–4). group of experts (production engineer (E1), designer
After 6 PODAs out of 78 good designs were (E2), and operator (E3)) with respect to three
specified using the conditions for domination given in subjective attributes as shown in Table 3. Experts’
Eqs. (3) and (4), the next step was to rank them and to linguistic assessments for the attributes A4, A5 and A6
select the best design. To achieve this, attributes of the are transformed into fuzzy numbers by using the scale
Table 3
Experts’ evaluations of six PODAs under the three subjective attributes and their corresponding fuzzy numbers
X1 X2 X3
E1 E2 E3 E1 E2 E3 E1 E2 E3

A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243


A4 Expert opinions Very Good Excellent Good–Very Good Very Good Excellent Good–Very Good Fairly Good Fairly Good More or Less Good
Corresponding (0.8, 0.9, 0.9, 1) (0.9, 1, 1, 1) (0.7, 0.8, 0.8, 0.9) (0.8, 0.9, 0.9, 1) (0.9, 1, 1, 1) (0.7, 0.8, 0.8, 0.9) (0.5, 0.6, 0.6, 0.7) (0.5, 0.6, 0.6, 0.7) (0.5, 0.55, 0.55, 0.6)
fuzzy numbers

A5 Expert opinions Good Very Good Fairly Good Good Very Good Fairly Good Fairly Good Good More or Less Good
Corresponding (0.5, 0.7, 0.7, 0.9) (0.8, 0.9, 0.9, 1) (0.5, 0.6, 0.6, 0.7) (0.5, 0.7, 0.7, 0.9) (0.8, 0.9, 0.9, 1) (0.5, 0.6, 0.6, 0.7) (0.5, 0.6, 0.6, 0.7) (0.5, 0.7, 0.7, 0.9) (0.5, 0.55, 0.55, 0.6)
fuzzy numbers

A6 Expert opinions Fairly Good Fairly Good Good Fairly Good Fairly Good Good Good Good–Very Good Fairly Good
Corresponding (0.5, 0.6, 0.6, 0.7) (0.5, 0.6, 0.6, 0.7) (0.5, 0.7, 0.7, 0.9) (0.5, 0.6, 0.6, 0.7) (0.5, 0.6, 0.6, 0.7) (0.5, 0.7, 0.7, 0.9) (0.5, 0.7, 0.7, 0.9) (0.7, 0.8, 0.8, 0.9) (0.5, 0.6, 0.6, 0.7)
fuzzy numbers

X4 X5 X6
E1 E2 E3 E1 E2 E3 E1 E2 E3
A4 Expert opinions Fairly Good Fairly Good More or Less Good More or Less Good More or Less Good Medium Medium More or Less Poor Fairly Poor
Corresponding (0.5, 0.6, 0.6, 0.7) (0.5, 0.6, 0.6, 0.7) (0.5, 0.55, 0.55, 0.6) (0.5, 0.55, 0.55, 0.6) (0.5, 0.55, 0.55, 0.6) (0.3, 0.5, 0.5, 0.7) (0.3, 0.5, 0.5, 0.7) (0.4, 0.45, 0.45, 0.5) (0.3, 0.4, 0.4, 0.5)
fuzzy numbers

A5 Expert opinions Fairly Good Good More or Less Good More or Less Good More or Less Good More or Less Good Medium More or Less Poor Poor
Corresponding (0.5, 0.6, 0.6, 0.7) (0.5, 0.7, 0.7, 0.9) (0.5, 0.55, 0.55, 0.6) (0.5, 0.55, 0.55, 0.6) (0.5, 0.55, 0.55, 0.6) (0.5, 0.55, 0.55, 0.6) (0.3, 0.5, 0.5, 0.7) (0.4, 0.45, 0.45, 0.5) (0.1, 0.3, 0.3, 0.5)
fuzzy numbers

A6 Expert opinions Good Good Fairly Good More or Less Good Fairly Poor More or Less Poor More or Less Good Fairly Poor Fairly Poor
Corresponding (0.5, 0.7, 0.7, 0.9) (0.5, 0.7, 0.7, 0.9) (0.5, 0.6, 0.6, 0.7) (0.5, 0.55, 0.55, 0.6) (0.3, 0.4, 0.4, 0.5) (0.4, 0.45, 0.45, 0.5) (0.5, 0.55, 0.55, 0.6) (0.3, 0.4, 0.4, 0.5) (0.3, 0.4, 0.4, 0.5)
fuzzy numbers

235
236 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

Table 4
Attributes’ properties and weights of attributes and experts
Attributes Type of Type of attribute Relative w E1 E2 E3
assessment importance R.I. we1 R.I. we2 R.I. we3
A1 (the most important) Crisp Benefit Objective 100 0.22
A2 (the moderate important) Crisp Benefit Objective 75 0.17
A3 (the most important) Crisp Benefit Objective 100 0.22
A4 (the least important) Linguistic Benefit Subjective 50 0.11 1 0.4 1 0.4 0.50 0.2
A5 (the least important) Linguistic Benefit Subjective 50 0.11 0.30 0.15 0.70 0.35 1 0.5
A6 (the moderate important) Linguistic Benefit Subjective 75 0.17 0.40 0.2 0.60 0.3 1 0.5

described in Section 3.2.1. These transformed fuzzy and weights of attributes and experts are given in
numbers are also shown in Table 3. Table 4. Attributes’ properties such as type of attributes
and type of assessments are also summarised in Table 4.
4.3. Attribute based aggregation phase calculations During thewhole process of the aggregation phase, b,
(Steps 7 and 8) which indicates the moderator’s dominance on the
problem, is set to 0.4. As an example, detailed
In this phase, all ratings are aggregated under each aggregation calculations for the sixth attribute (A6)
subjective attribute by taking into account the attribute aregiven in Table 5. Experts’ opinionsfor the fifth PODA
based expert weights. Before aggregation, it is with respect to the sixth attribute and their aggregation
necessary to identify the weights of attributes and results for homo/heterogeneous groups of experts are
experts. Therefore, the relative importance of attributes shown in Fig. 6. Please note that there are 18 (=3
and experts are assigned according to importance subjective attributes X 6 PODAs) aggregation calcula-
observed through interview with moderator. Then tions in this case study. After aggregation calculations,
weights for attributes and experts are determined such aggregation matrices for homo/heterogeneous groups of
that the sum of all weights is 1. The relative importance experts are constructed as shown in Table 6.

Fig. 6. Aggregation of experts’ opinions for the X5 with respect to A6.


Table 5
Aggregation calculations for the sixth attribute (A6)

A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243


X1 X2 X3 X4 X5 X6
E1 (0.5, 0.6, 0.6, 0.7) (0.5, 0.6, 0.6, 0.7) (0.5, 0.7, 0.7, 0.9) (0.5, 0.7, 0.7, 0.9) (0.5, 0.55, 0.55, 0.6) (0.5, 0.55, 0.55, 0.6)
E2 (0.5, 0.6, 0.6, 0.7) (0.5, 0.6, 0.6, 0.7) (0.7, 0.8, 0.8, 0.9) (0.5, 0.7, 0.7, 0.9) (0.3, 0.4, 0.4, 0.5) (0.3, 0.4, 0.4, 0.5)
E3 (0.5, 0.7, 0.7, 0.9) (0.5, 0.7, 0.7, 0.9) (0.5, 0.6, 0.6, 0.7) (0.5, 0.6, 0.6, 0.7) (0.4, 0.45, 0.45, 0.5) (0.3, 0.4, 0.4, 0.5)

S12 1.000 1.000 0.900 1.000 0.850 0.850


S13 0.900 0.900 0.900 0.900 0.900 0.850
S23 0.900 0.900 0.800 0.900 0.950 1.000

AA(E1) 0.950 0.950 0.900 0.950 0.875 0.850


AA(E2) 0.950 0.950 0.850 0.950 0.900 0.925
AA(E3) 0.900 0.900 0.850 0.900 0.925 0.925

RA(E1) 0.3393 0.3393 0.3462 0.3393 0.3241 0.3148


RA(E2) 0.3393 0.3393 0.3269 0.3393 0.3333 0.3426
RA(E3) 0.3214 0.3214 0.3269 0.3214 0.3426 0.3426

CC(E1) 0.2836 0.2836 0.2877 0.2836 0.2744 0.1889


CC(E2) 0.3236 0.3236 0.3162 0.3236 0.3200 0.2056
CC(E3) 0.3929 0.3929 0.3962 0.3929 0.4056 0.2056

RAGHM (0.500, 0.632, 0.632, 0.764) (0.500, 0.632, 0.632, 0.764) (0.568, 0.702, 0.702, 0.836) (0.500, 0.668, 0.668, 0.836) (0.400, 0.467, 0.467, 0.534) (0.368, 0.451, 0.451, 0.534)
RAGHT (0.500, 0.639, 0.639, 0.779) (0.500, 0.639, 0.639, 0.779) (0.565, 0.693, 0.693, 0.821) (0.500, 0.661, 0.661, 0.821) (0.396, 0.462, 0.462, 0.528) (0.357, 0.443, 0.443, 0.528)

237
238
Table 6

A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243


Aggregated matrices for homo/heterogeneous groups of experts
X1 X2 X3 X4 X5 X6
A1 Homo 8 8 10 10 12 14
Hetero 8 8 10 10 12 14

A2 Homo 5.62165 5.25369 5.16295 5.07325 5.05024 5.42632


Hetero 5.62165 5.25369 5.16295 5.07325 5.05024 5.42632

A3 Homo 14.07130 14.13550 14.07720 14.08030 14.06260 13.94820


Hetero 14.07130 14.13550 14.07720 14.08030 14.06260 13.94820

A4 Homo (0.801, 0.901, 0.901, 0.968) (0.801, 0.901, 0.901, 0.968) (0.5, 0.584, 0.584, 0.668) (0.5, 0.584, 0.584, 0.668) (0.435, 0.534, 0.534, 0.633) (0.333, 0.451, 0.451, 0.569)
Hetero (0.808, 0.908, 0.908, 0.973) (0.808, 0.908, 0.908, 0.973) (0.5, 0.586, 0.586, 0.673) (0.5, 0.586, 0.586, 0.673) (0.445, 0.536, 0.536, 0.628) (0.336, 0.455, 0.455, 0.573)

A5 Homo (0.594, 0.729, 0.729, 0.865) (0.594, 0.729, 0.729, 0.865) (0.5, 0.615, 0.615, 0.729) (0.5, 0.615, 0.615, 0.729) (0.5, 0.55, 0.55, 0.6) (0.265, 0.418, 0.418, 0.571)
Hetero (0.598, 0.726, 0.726, 0.853) (0.598, 0.726, 0.726, 0.853) (0.5, 0.613, 0.613, 0.726) (0.5, 0.613, 0.613, 0.726) (0.5, 0.55, 0.55, 0.6) (0.253, 0.404, 0.404, 0.555)

A6 Homo (0.5, 0.632, 0.632, 0.764) (0.5, 0.632, 0.632, 0.764) (0.568, 0.702, 0.702, 0.836) (0.5, 0.668, 0.668, 0.836) (0.4, 0.467, 0.467, 0.534) (0.368, 0.451, 0.451, 0.534)
Hetero (0.5, 0.639, 0.639, 0.779) (0.5, 0.639, 0.639, 0.779) (0.565, 0.693, 0.693, 0.821) (0.5, 0.661, 0.661, 0.821) (0.396, 0.462, 0.462, 0.528) (0.357, 0.443, 0.443, 0.528)
Table 7
Defuzzified aggregated values, (weighted) normalised ratings for homo/heterogeneous groups of experts

A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243


239
240 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

Table 8
Values of separation measures and relative closeness to the positive-ideal solution for homo/heterogeneous groups of experts
X1 X2 X3 X4 X5 X6
Homo Hetero Homo Hetero Homo Hetero Homo Hetero Homo Hetero Homo Hetero
Si 0.0516 0.0514 0.0518 0.0516 0.0403 0.0404 0.0407 0.0408 0.0399 0.0398 0.0433 0.0439
Si 0.0399 0.0412 0.0393 0.0407 0.0350 0.0354 0.0320 0.0326 0.0356 0.0359 0.0513 0.0513
Ci(OAR) 0.4362 0.4452 0.4314 0.4405 0.4646 0.4673 0.4403 0.4440 0.4713 0.4737 0.5425 0.5389
Ranking 5 4 6 6 3 3 4 5 2 2 1 1

4.4. Selection phase calculations The positive-ideal solution is determined by taking


(Steps 9, 10, 11 and 12) the largest element for each benefit attribute and the
smallest element for each cost attribute. The negative-
Up to this phase, experts’ linguistic assessments ideal solution is just the opposite formation of the PIS.
have been transformed into standardised trapezoidal Positive and negative ideal solutions are also indicated
fuzzy numbers and then aggregated under each as superscripts in Table 7 for homo/heterogeneous
subjective attribute. In order to rank the alternatives, groups of experts. Table 8 shows the values of
aggregated matrices’ fuzzy elements should be separation measures and relative closeness to the
defuzzified. Defuzzified aggregated values for positive-ideal solution for homo/heterogeneous
homo/heterogeneous groups of experts are shown in groups of experts.
Table 7. Finally, PODAs are ranked on the basis of OAR
The TOPSIS procedure is applied to the six PODAs values. For homogeneous groups of experts, according
to obtain their OAR values and ranking orders. First, to the descending order of OAR, the preference order
the defuzzified aggregated values are normalised. is X6 > X5 > X3 > X4 > X1 > X2, where the sixth
Table 7 shows the normalised values of each attribute alternative is the leader and X5 and X3 are ranked
for homo/heterogeneous groups of experts. Weighted second and third. Similarly, for heterogeneous group
normalised ratings of each attribute are calculated by of experts (for b = 0.4), ranking of alternatives is given
multiplying each attribute with its associated weight: as X6 > X5 > X3 > X1 > X4 > X2.
(w1 ; w2 ; w3 ; w4 ; w5 ; w6 ) = (0.22, 0.17, 0.22, 0.11, Properties and performance ratings of the original
0.11, 0.17). Table 7 also shows the weighted normal- and best designs are given in Table 9. It can be
ised values for homo/heterogeneous groups of experts. concluded that there has been a significant improve-

Fig. 7. Sensitivity analysis for b coefficient.


A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243 241

Table 9
Properties and performance ratings of the original and best designs
No. Optimisation variables Original design Best design
1 Car deck height 9.7 m 9.8 m
2 Side-casing width No side-casing 1m
3 Lower-hold height (from car deck) 2.6 m 5.2 m

No. Watertight transverse bulkheads In frame numbers In frame numbers


4 Transverse bulkhead 02 27 26
5 Transverse bulkhead 03 39 41
6 Transverse bulkhead 04 51 52
7 Transverse bulkhead 05 63 65
8 Transverse bulkhead 06 81 83
9 Transverse bulkhead 07 99 97
10 Transverse bulkhead 08 117 115
11 Transverse bulkhead 09 129 128
12 Transverse bulkhead 10 141 140
13 Transverse bulkhead 11 153 151
14 Transverse bulkhead 12 165 164
15 Transverse bulkhead 13 177 178
16 Transverse bulkhead 14 189 190
Performance ratings of designs
1 Cargo capacity 8 8 + (4 more car lanes)
2 HS value 4.641 m 5.42632 m
3 KG limiting value 13.845 m 13.9482 m

ment in terms of performance criteria, particularly in PODAs in a FMAGDM environment. The proposed
cargo capacity. According to the latest configuration, approach is suitable and effective in dealing with
internal hull space arrangement of the Ro–Ro vessel is experts’ fuzzy opinions in an imprecise environment,
improved by increasing lower-hold height and car since conventional approaches tend to be less
deck height and by adding side-casing with 1 m width. effective in dealing with fuzziness and vagueness.
Thus the Ro–Ro vessel has the ability to carry more Based on the stepwise algorithm given above, various
cargo by making the above-suggested changes, which mathematical and commercial software such as
will allow approximately 250 more cars to be carried NAPA, Microsoft Excel and the FRONTIER opti-
in the lower-hold. misation environment have been integrated to
It is obvious that the developed methodology is develop the state of the art tool, which encapsulates
sensitive to weighting vectors, b coefficient and deterministic and performance based safety assess-
experts’ fuzzy assessments of the problem. Therefore, ments in a way suitable to perform internal hull
sensitivity analysis is performed to see the b effect on subdivision arrangement of Ro–Ro vessels in the
the OAR values (see Fig. 7). ship design.
The proposed approach provides satisfactory
conclusions, the important ones of which include:
5. Conclusions

It is a robust MCDM model, which combines multi-
This paper presents a new MCDM approach, objective optimisation methodology with a
which is an integrated, interactive and multi-stage FMAGDM technique.
process. It is proposed to investigate the impact of
It enables ship designers to search PODAs over a
POD propulsion systems on the internal hull feasible design space and to find the best one among
subdivision of a Ro–Ro vessel, where considerable them by the ability of incorporating their weighted
emphasis has been put on evaluation and ranking of fuzzy (or subjective) preferences.
242 A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243

It allows ship designers to deal with the fuzzy [7] H.J. Zimmermann, Fuzzy sets, decision-making and expert
opinions of homo/heterogeneous groups of experts. systems, Kluwer Academic Publishers, Dordrecht, 1987.

It allows FMAGDM problems to take data in the [8] L. Mikhailov, P. Tsvetinov, Evaluation of services using a
fuzzy analytic hierarchy process, Appl. Soft Comput. 5 (2004)
forms of linguistic terms, fuzzy numbers, and/or
23–33.
crisp numbers which yields more realistic, accurate [9] S.J. Chen, C.L. Hwang, Fuzzy Multiple Attribute Decision-
and reliable decision models than the existing ones. making: Methods and Applications, Springer-Verlag, New

Fuzzy numbers and linguistic variables are used to York, 1992.


quantify the mental perception of the expert. [10] R.A. Ribeiro, Fuzzy multiple attribute decision-making: a
review and new preference elicitation techniques, Fuzzy Sets

The ranking algorithm of the proposed approach is


Syst. 78 (1996) 155–181.
also easily coded into a computer program due to its [11] A.I. Ölçer, A.Y. Odabasi, A new fuzzy multiple attributive
stepwise description. group decision making methodology and its application to
propulsion/manoeuvring system selection problem, Eur. J.
Although the proposed MCDM methodology has Operat. Res. 166 (2005) 93–114.
[12] C. Parkan, M.L. Wu, Process selection with multiple objective
been successfully applied to a Ro–Ro vessel, it is also and subjective attributes, Production Plann. Control 9 (1998)
applicable to a broader area of MCDM problems in 189–200.
engineering and operational research as well as in [13] modeFRONTIER, Version 3.0.0, ESTECO, http://www.esteco.
naval architecture. it/, Trieste, Italy.
[14] J. Holland, Adaptation in Natural and Artificial Systems,
University of Michigan Press, Ann Arbor, 1975.
[15] P. Sen, D. Todd, Multiple criteria genetic algorithms: a cata-
Acknowledgements maran design study, University of Newcastle upon Tyne,
Department of Marine Technology, EDC, UK, 1999.
This multidisciplinary research is a part of FP5 EU- [16] C. Poloni, M. Fearon, N.G. D., Parallelisation of genetic
OPTIPOD project (optimal design and implementa- algorithm for aerodynamic design optimisation, Adaptive
Comput. Eng. Design Control 96 (1996) 59–64.
tion of azimuthing pods for the safe and efficient [17] A.I. Ölçer, Development of a new fuzzy multiple attribute
propulsion of ships) and is supported by the following decision making approach and its application to decision
grant and contract: Project No: GRD1-1999-10294, making in ship design and shipbuilding, Ph.D. Thesis, Istanbul
Contract No: G3RD-CT-1999-00017. Technical University, Institute of Science and Technology,
Istanbul, Turkey, 2001.
[18] L.A. Zadeh, The concept of a linguistic variable and its
application to approximate reasoning – I, Inform. Sci. 8
References (1975) 199–249.
[19] L.A. Zadeh, The concept of a linguistic variable and its
[1] P. Sen, M. Gerigk, Some aspects of a knowledge-based expert application to approximate reasoning – II, Inform. Sci. 8
system for preliminary ship subdivision design for safety, in: (1975) 301–357.
International Symposium PRADS’92, vol. 2, 1992, 1187– [20] H.J. Zimmermann, Fuzzy Set Theory and its Applications,
1197. second ed. Kluwer Academic Publishers, 1991.
[2] A.I. Ölçer, C. Tuzcu, O. Turan, Internal hull subdivision [21] S.M. Chen, Aggregating fuzzy opinions in the group decision-
optimisation of Ro–Ro vessels in multiple criteria decision making environment, Cybernetics Syst. 29 (1998) 363–376.
making environment, in: Proceedings of the 8th International [22] S.M. Chen, S.Y. Lin, A new method for fuzzy risk analysis, in:
Marine Design Conference, vol. 1, 2003, pp. 339–351. Proceedings of the 1995 Artificial Intelligence Workshop,
[3] O. Turan, B. Turkmen, C. Tuzcu, Case-based reasoning 1995, pp. 245–250.
approach to internal hull subdivision design, 4th International [23] A. Kaufmann, M.M. Gupta, Introduction to Fuzzy Arithmetic
Conference on Advanced Engineering Design, AED 2004, Theory and Application, Van Nortrand Reinhold, New York,
Glasgow, Scotland, UK, September 2004, pp. 5–8. 1991.
[4] V. Oduguwa, A. Tiwari, R. Roy, Evolutionary computing in [24] A. Bardossy, L. Duckstein, I. Bogardi, Combination of fuzzy
manufacturing industry: an overview of recent applications, numbers representing expert opinions, Fuzzy Sets Syst. 57
Appl. Soft Comput. 5 (2005) 281–299. (1993) 173–181.
[5] K. Deb, Multiobjective Optimization using Evolutionary [25] H.M. Hsu, C.T. Chen, Aggregation of fuzzy opinions under
Algorithms, Wiley, Chichester, UK, 2001. group decision making, Fuzzy Sets Syst. 79 (1996) 279–285.
[6] C.M. Fonseca, P.J. Fleming, Multi-objective optimisation, in: [26] C.L. Hwang, K.P. Yoon, Multiple Attribute Decision-making:
Handbook of Evolutionary Computation, IOP Publishing Ltd. Methods and Applications, Springer-Verlag, Berlin/Heidel-
and Oxford University Press, 1997. berg/New York, 1981.
A.I. Ölçer et al. / Applied Soft Computing 6 (2006) 221–243 243

[27] IMO, SLF 40/INF.14 (Annex 1), Ro–Ro passenger ship safety: [28] D. Vassalos, M. Pawlowski, O. Turan, A Theoretical Investiga-
guidance notes on Annex 3 of the agreement concerning tion on the Capsizal Resistance of Passenger/Ro–Ro Vessels
specific stability requirements for Ro–Ro passenger ships and Proposal of Survival Criteria, Final Report, Task 5, The
undertaking regular scheduled international voyages between North West European R&D Project, 1996.
or to or from designated ports in north west Europe and the [29] NAPA Release 2003.2, NAPA Oy, http://www.napa.fi/, Hel-
Baltic sea, July 1996. sinki, Finland.

You might also like