Expert Systems

with Applications
Expert Systems with Applications 34 (2008) 205213
www.elsevier.com/locate/eswa

A causal analytical method for group decision-making under

fuzzy environment
Chi-Jen Lin
a

a,*

, Wei-Wen Wu

b,1

Department of Industrial Engineering and Management, Ta Hwa Institute of Technology, No. 1, Ta Hwa Road, Cyong-Lin,
Hsin-Chu County 307, Taiwan, ROC
b
Department of International Trade, Ta Hwa Institute of Technology, No. 1, Ta Hwa Road, Cyong-Lin,
Hsin-Chu County 307, Taiwan, ROC

Abstract
Causal analysis largely inuences the eectiveness of decision-making and the productivity of actions. The complex relationship
between cause and eect as well as the fuzzy nature of human life make the casual analysis dicult. In this paper, we develop a fuzzy
DEMATEL method for group decision-making to gather group ideas and analyze the causeeect relationship of complex problems in
fuzzy environments. Procedures of the fuzzy DEMATEL method are then proposed. Using the fuzzy DEMATEL procedures, the
involved criteria of a system (or subsystem) are separated into the cause and eect groups for helping decision-makers focus on those
criteria that provide great inuence. An empirical study applies the proposed fuzzy DEMATEL method to the R&D project selection
of a Taiwanese company. The result shows that, within the cause group, the criterion of probability of technical success is the most
important factor for R&D project selection, whereas the strategic t and potential size of market have the best eect on the other
criteria. By contrast, the net present value is the most easily improved of the eect group criteria.
 2006 Elsevier Ltd. All rights reserved.
Keywords: Causal analysis; Fuzzy theory; Group decision-making; DEMATEL

1. Introduction
The eectiveness of decision-making depends largely on
the ability of decision-makers to analyze the complex
causeeect relationship and take productive actions based
on the analysis. Cause and eect are two dierent concepts.
Causes tell the reason why something happened, whereas
eects are the results of that happening. Cause and eect
link situations and events together in time since they generally cause preceding eects. However, the relationships of
cause and eect are often complex and subtle. To capture the cause and eect relationship is thus not an easy
task. The fuzzy nature of human life makes the causeeect
*
Corresponding author. Tel.: +886 3 5927700x2754; fax: +886 3
5926848.
E-mail address: lcj@thit.edu.tw (C.-J. Lin).
1
Tel.: +886 3 5927700x2902; fax: +886 3 5925715.

0957-4174/$ - see front matter  2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2006.08.012

analysis more dicult. Hence, a method for inferring the

causal relationship in fuzzy environments is necessary.
The Decision-Making Trial and Evaluation Laboratory
(DEMATEL) method, a very popular method in Japan,
applies matrices and digraphs for visualizing the structure
of complicated causal relationships. Hence, the DEMATEL method can separate the involved criteria of a system
(or subsystem) into the cause and eect groups for helping
make decisions. The crisp DEMATEL method has been
successfully applied in many elds. In this paper, we
develop a fuzzy DEMATEL method to gather group ideas
and analyze the causeeect relationship of complex problems in fuzzy environments.
The rest of this paper is organized as follows. In Section
2, the fuzzy theory and crisp DEMATEL method are
reviewed. In Section 3, a methodology derived from the
DEMATEL method with fuzzy logic is proposed. Section
4 presents an empirical study to illustrate the procedure

206

C.-J. Lin, W.-W. Wu / Expert Systems with Applications 34 (2008) 205213

of the proposed fuzzy DEMATEL method and to demonstrate its usefulness and validity.
2. DEMATEL method and fuzzy group decision-making
The foundations of the fuzzy DEMATEL method for
group decision-making are discussed in the following.
2.1. Fuzzy theory
Fuzzy theory is very helpful to deal with the vagueness
of human thoughts and language in making decisions.
Decision-makers tend to give assessments according to
their past experiences and knowledge, and also their estimations are often expressed in equivocal linguistic terms.
However, to integrate various experiences, opinions, ideas,
and motivations of an individual decision-maker, it is better to convert the linguistic estimation into fuzzy numbers.
Thus, the problems of group decision-making in real world
have created a need to employ fuzzy logic. In the following,
we briey review some essential denitions of fuzzy logic
(Kaufmann & Gupta, 1991).
e is a subset of a universe of
Denition 2.1.1. A fuzzy set A
discourse X, which is characterized by a membership
function le x representing a mapping le : X ! 0; 1. The
A
A
e is called the membership value,
function value of le x A
A
which represents the degree of truth that x is an element of
e It is assumed that l x 2 0; 1, where
fuzzy set A.
eA
e while
le x 0 reveals that x belongs completely to A,
A
e
indicates that x does not belong to the fuzzy set A.
e of the universe of discourse
Denition 2.1.2. A fuzzy set A
X is convex if and only if
lekx1 1  kx2 P minlex1 ; lex2 ;
A
A
A
8x 2 x1 ; x2 ; k 2 0; 1:
e of the universe of discourse
Denition 2.1.3. A fuzzy set A
X is normal if
max lex 1:
A

e of the uniDenition 2.1.4. The a-cut of the fuzzy set A

verse of discourse X is dened as
e a fx 2 X jl x P ag;
A
e
A

where a 2 [0,1].

e can be
Denition 2.1.7. A triangular fuzzy number N
dened as a triplet (, m, u), and the membership function
le xis dened as:
N
8
0;
x<
>
>
>
< x  =m  ; 6 x 6 m
le x
N
>
>
> u  x=u  m; m 6 x 6 u
:
0;
x > u;
where , m, and u are real numbers and 6 m 6 u. See
Fig. 1.
e ; m; u be a triangular fuzzy numTheorem 2.1.1. Let N
ber and k > 0 a crisp number, then
e k; km; ku:
kN

e 1 1 ; m1 ; u1 and N
e 2 2 ; m2 ; u2
Theorem 2.1.2. Let N
be two triangular fuzzy numbers. The addition operations
e 1 and N
e 2 , denoted by N
e1  N
e 2 yield another triangular
of N
fuzzy number.
e1  N
e 2 1 2 ; m1 m2 ; u1 u2
N

e 1 1 ; m1 ; u1 and N
e 2 2 ; m2 ; u2
Theorem 2.1.3. Let N
e1
be two triangular fuzzy numbers. The multiplication of N
e 2 , denoted by N
e1  N
e 2 , is defined by the membership
and N
function le e x as follows (Li, 1999):
N 1N 2
8
0
>
>
p
>
2
>
>
< 21 2 1 m2 2 m1 1 m2 2 2 m1 1  4xm1 1 m2 2
2m1 1 m2 2
p
>
2u1 u2 u1 m2 u2 m1  u1 u2 m2 u2 u1 m1 2 4xu1 m1 u2 m2
>
>
>
2u1 m1 u2 m2
>
:
0

; x 6 1 2

; 1 2 6 x 6 m1 m2
; m1 m2 6 x 6 u1 u2
; x P u1 u2

3
e
e
It is obvious that N 1  N 2 is not a triangular fuzzy number. However, the following property provides an approxe1  N
e 2 as a triangular fuzzy
imation formula to regard N
number.

e 1 1 ; m1 ; u1 and N
e 2 2 ; m2 ; u2
Theorem 2.1.4. Let N
e
e 2 approxibe two positive triangular fuzzy numbers, N 1  N
mates a triangular fuzzy number (1 2, m1 m2, u1 u2)
(Laarhoven & Pedrycz, 1983), i.e.,
e1  N
e 2 1  2 ; m1  m2 ; u1  u2 :
N
See Fig. 2.

e is a fuzzy subset in
Denition 2.1.5. A fuzzy number N
the universe of discourse X, which is both convex and
normal.
e be a fuzzy number and N
e a be the
Denition 2.1.6. Let N
e . If the lower bound of N
e a > 0 for a 2 [0,1], then
a-cut of N
e is called a positive fuzzy number.
N

e.
Fig. 1. A triangular fuzzy number N

C.-J. Lin, W.-W. Wu / Expert Systems with Applications 34 (2008) 205213

ita (1997) analyzed the obstructive factors of welfare service with the DEMATEL method, and Hori and Shimizu
(1999) employed it to design and evaluate the software of
displaying-screen structure in analyzing a supervisory control system.
The essentials of the crisp DEMATEL method will be
reviewed below. Suppose that a system contains a set of criteria C = {C1, C2, . . ., Cn}, and the particular pair-wise relations are determined for modeling with respect to a
mathematical relation.

e1  N
e 2 and its approximation.
Fig. 2. Membership function of N

Denition 2.1.8. Linguistic variables are used as variables

whose values are not numbers but linguistic terms (Zadeh,
1975).
The linguistic-variable approach is usually employed by
the decision-makers to express their assessments, which is
very useful in dealing with ill-dened situations in traditional quantitative expressions. Linguistic values can be
represented with fuzzy numbers. In particular, the triangular fuzzy numbers are commonly used.
As that of most fuzzy model, we had to convert the nal
fuzzy data into a crisp value. Here, we suggest the CFCS
method proposed by Opricovic and Tzeng (2003) for
defuzzication. This method has the advantages of giving
a greater crisp value with greater membership function
and distinguishing two symmetrical triangular fuzzy numbers with the same mean.
Theorem 2.1.5 (The CFCS defuzzication method). Let
e k k ; mk ; uk ; k = 1, 2, . . ., n, be the positive triangular
N
e def denote its representing crisp value.
fuzzy numbers, and N
k
Computing L = min (k); R = max (uk); k = 1, 2, . . ., n, and
D = R  L, then

Denition 2.2.1. The pair-wise comparison scale may be

designated into four levels, where the scores of 0, 1, 2, and
3 represent No inuence, Low inuence, High
inuence, and Very high inuence, respectively.
Denition 2.2.2. The initial direct-relation matrix Z is a
n n matrix obtained by pair-wise comparisons in terms
of inuences and directions between criteria, in which zij
is denoted as the degree to which the criterion Ci affects criterion Cj. Accordingly, all principal diagonal elements zii of
matrix Z are set to be zero.
C1
Z C2
..
.
Cn

m  LD u  m R  u  L D m 

D m  D u  m2 R  u  LD m  2 D u  m

2.2. The DEMATEL method

The DEMATEL method, originated from the Geneva
Research Centre of the Battelle Memorial Institute (Fontela & Gabus, 1976; Gabus & Fontela, 1973), was aimed
at the fragmented and antagonistic phenomena of world
societies and searched for integrated solutions. It is especially practical and useful for visualizing the structure of
complicated causal relationships with matrices or digraphs.
The matrices or digraphs portray a contextual relation
between the elements of the system, in which a numeral
represents the strength of inuence. Hence, the DEMATEL method can convert the relationship between the
causes and eects of criteria into an intelligible structural
model of the system. The DEMATEL method has been
successfully applied in many elds. For examples, Tamura,
Akazawa, and Nagata (2002) used it to extract various
uneasy factors in the life, Yamazaki, Ishibe, and Yamash-

C1 C2    Cn
3
0 z12    z1n
6z
0 z2n 7
6 21 0
7
6 .
.. 7
.. . .
6 .
7:
4 .
. . 5
.
2

zn1

zn2



Denition 2.2.3. Let

s max

16i6n

e def L D
N
k


207

n
X

!
zij ;

j1

then the normalized direct-relation matrix X can be obtained

through the following formula .
X

Z
:
s

The DEMATEL
Pn method further assumes that at least
one i such that j1 zij < s. This assumption is satised in
almost all practical cases. Hence, matrix X just resembles
the sub-stochastic matrix obtained from an absorbing
Markov chain matrix by deleting all rows and columns
associated with the absorbing states. It was proved that
limitw!1Xw = O and limitw!1(I + X + X2 +    + Xw) =
(I  X)1, where O is the null matrix and I is the identity
matrix (Goodman, 1988; Papoulis & Pillai, 2002).
Denition 2.2.4. The total-relation matrix T can be
acquired by calculating
T limitw!1 X X2    Xw XI  X1 :
The total-indirect-relation matrix H can be obtained
through the following formula.
H limitw!1 X2 X3    Xw X2 I  X1 :

208

C.-J. Lin, W.-W. Wu / Expert Systems with Applications 34 (2008) 205213

Denition 2.2.5. Let tij (i, j = 1, 2, . . ., n) be the elements of

the total-relation matrix T, then the sum of rows and the
sum of columns, denoted as Di and Rj, respectively, can
be obtained through the following two formulas:
Di

n
X

tij

i 1; 2; . . . ; n;

tij

j 1; 2; . . . ; n:

j1

Rj

n
X
i1

Denition 2.2.6. A causal diagram can be acquired by

mapping the ordered pairs of (Dk + Rk, Dk  Rk), where
the horizontal axis (D + R), named Prominence, is made
by adding Rk to Dk, and the vertical axis (D  R), named
Relation, is made by subtracting Rk from Dk.
In the causal diagram, the horizontal axis Prominence
shows how important the criterion is, whereas the vertical
axis Relation may divide the criteria into the cause and
eect groups. When the value (Dk  Rk) is positive, the criterion belongs to the cause group. If the value (Dk  Rk) is
negative, the criterion belongs to the eect group. Hence,
causal diagrams can visualize the complicated causal relationships between criteria into a visible structural model
and provide valuable insight for problem solving. Furthermore, with the help of a causal diagram, we may make
proper decisions by recognizing the dierence between
cause and eect criteria.
3. The procedure of the proposed method
The analytical procedure of the proposed fuzzy DEMATEL method is briey described as follows.
Step 1: Set the decision goal and set up a committee.
Decision-making is the process of dening the decision goals, gathering relevant information, generating the broadest possible range of alternatives,
evaluating the alternatives for advantages and disadvantages, selecting the optimal alternatives, and
monitoring the results to ensure that the decision
goals are achieved (Hess & Siciliano, 1996; Opricovic & Tzeng, 2004). Thus, a decision goal is set up
rst. Subsequently, a committee is set up for gathering group knowledge for problem solving.
Step 2: Develop the evaluation criteria and design the
fuzzy linguistic scale.
In this step, it is needed to establish sets of criteria
for evaluation. However, evaluation criteria have
the nature of causal relationships and usually comprise several complicated aspects. To deal with the
ambiguities of human assessments, we discard the
comparison scale used in the crisp DEMATEL
method (see Denition 2.2.1) but adopt the fuzzy
linguistic scale used in the group decision-making
proposed by Li (1999). The dierent degrees of

inuence are expressed with ve linguistic terms

as {Very high, High, Low, Very low, No} and their
corresponding positive triangular fuzzy numbers
are shown in Table 1 and see Fig. 3.
Step 3: Acquire the assessments of decision-makers.
To measure the relationship between criteria
C = {Ciji = 1, 2, . . ., n}, a decision group of p
experts was asked to make sets of pair-wise comparisons in terms of linguistic terms. Hence, p fuzzy
e h2i , . . ., Z
e hpi , each corresponding to
e h1i , Z
matrices Z
an expert and with triangular fuzzy numbers as
e hki as:
its elements, were obtained. Denote Z
3
2
hki
hki
0 ~z12    ~z1n
6 hki
hki 7
6 ~z21
0    ~z2n 7
7
6
hki
e 6
; k 1; 2; . . . ; p;
Z
.. 7
..
..
7
6 ..
4 .
. 5
.
.
hki

hki

~zn1



~zn2



hki
hki
hki
hki
where ez ij ij ; mij ; uij . Without loss of generhki

ality, elements ez ii (i = 1, 2, . . ., n) will be regarded

as a triangular fuzzy number (0, 0, 0) whenever it
e hki is called the initial
is necessary. Fuzzy matrix Z
direct-relation fuzzy matrix of expert k.
Step 4: Acquire the normalized direct-relation fuzzy
matrix
hki
Let ~ai be the triangular fuzzy numbers,
~ahki
i

n
X

hki
~zij

j1

n
X

hki
ij ;

j1

n
X
j1

hki
mij ;

n
X

!
hki
uij

j1

and
r

hki

max

16i6n

n
X

!
hki
uij

j1

Table 1
The correspondence of linguistic terms and linguistic values
Linguistic terms

Linguistic values

Very high inuence (VH)

High inuence (H)
Low inuence (L)
Very low inuence (VL)
No inuence (No)

(0.75, 1.0, 1.0)

(0.5, 0.75, 1.0)
(0.25, 0.5, 0.75)
(0, 0.25, 0.5)
(0, 0, 0.25)

Fig. 3. Triangular fuzzy numbers for linguistic variables.

C.-J. Lin, W.-W. Wu / Expert Systems with Applications 34 (2008) 205213

The linear scale transformation is then used as a

normalization formula to transform the criteria
scales into comparable scales. The normalized direct-relation fuzzy matrix of expert k, denoted as
e hki , is given by
X
3
2
hki
~xhki
~xhki
   ~x1n
11
12
7
6 hki
hki
hki 7
6 ~x
~
~



x
x
21
22
2n
7
6
e hki 6
; k 1; 2; . . . ; p;
X
.. 7
..
..
7
6 ..
. 5
.
.
4 .
~xhki
~xhki
   ~xhki
n1
n2
nn
where
hki

~xhki
ij

~zij
hki
r

mn1
2

mn2
u12

6u
6 21
Xu 6
6 ..
4 .
un1

   m2n 7
7
.. 7
..
7;
. 5
.


   u1n

   u2n 7
7
.. 7
..
7:
. 5
.
 0
P

n
hki
hki

max
u
,
Since
r
16i6n
j1 ij
Pn k
1
u
6
1
and
then
get
k
j1 ij
r
!
p
n
X
1 X
k
u
6 p:
rk j1 ij
k1
0
..
.
un2

we

have

Furthermore, the sum of the ith row of

Xu

n
X

uij

j1

n
X

0P

uij
p
k1 rk A

j1

p
n
1 X 1 X
k

u
p k1 rk j1 ij

k
p
n X
uij
1 X

p j1 k1 rk

!
6 1:

And
of Xu
Pn Pnthe sum of all Pentries
n Pn
i1
j1 uij 6 n. However,
i1
j1 uij n will
not happen in real cases.

   ~xnn

e is called the normalized directThe fuzzy matrix X

relation fuzzy matrix. Here, we use the arithmetic
mean to aggregate, or pool all the data across the
experts after computing the normalized direct-relae hki . This approach can make the
tion fuzzy matrix X
dierences of individuals apparent, which is better
than the way of aggregating all the data of the experts right after obtaining the initial direct-relation
e hki .
fuzzy matrix Z
Step 5: Establish and analyze the structural model.
e we
To compute the total-relation fuzzy matrix T,
e
have to ensure the convergence of limw!1 X w 0
e w , we apply the
in advance. In computing X
approximation formula (4) for the multiplication
of two triangular fuzzy numbers. In fact, under
the linguistic values of this paper, formula (4)
almost coincides with formula (3). Hence, the elee w are also triangular fuzzy numbers.
ments of X
Let ~xij ij ; mij ; uij and dene three crisp matrices,
e as follows:
whose elements are extracted from X,
2
3
0 12    1n
6
7
6 21 0    2n 7
X 6
. 7
.. . .
6 ..
7;
4 .
. .. 5
.
n1 n2    0

0
..
.

   m1n

and

!
hki
hki
hki
ij mij uij
:
;
;
rhki rhki rhki

~xn2

m12

6m
6 21
Xm 6
6 ..
4 .

As that in crisp DEMATEL

Pn hkimethod, we assume at
least one i such that j1 uij < rhki . This assumption is well satised in practical cases. Formulas
(1) and (2) are then used to calculate the average
e of X
e h1i ; X
e h2i ; . . . ; X
e hpi , i.e.,
matrix, denoted as X,


e h2i      X
e hpi
e h1i  X
X
e
:
6
X
p
e as:
Denote X
2
3
~x11 ~x12    ~x1n
p
P
6 ~x
7
~xhki
ij
6 21 ~x22    ~x2n 7
k1
e 6 .
X
.. 7
.. . .
6 .
7; where ~xij p :
4 .
. . 5
.
~xn1

209

ew
The following theorem enables the computation of X
to be executed by the multiplication of crisp matrices.
Theorem 3.1. Let
3
2 w w
w
~x11 ~x12  ~x1n
6 w w
w 7
7
6 ~x21 ~x22  ~x2n
7
6
w
w
w w
ew 6
; where ~xij ij ;mij ;uij :
X
.. 7
.. . .
7
6 ..
4 .
. . 5
.
w
w
w
~xn1 ~xn2  ~xnn
We further define three matrices,
3
2 w w
w
11 12    1n
6 w w
w 7
7
6 21 22    2n
7
6
w
;
ij  6 .
.. 7
..
..
7
6 .
4 .
. 5
.
.
w

2
w
mij 

n1

m11

6 w
6 m21
6
6 .
6 .
4 .
w

mn1

n2






..
.


m12

m22
..
.
w
mn2

w
nn

m1n

w 7
m2n 7
7
;
.. 7
7
. 5
mw
nn

210

C.-J. Lin, W.-W. Wu / Expert Systems with Applications 34 (2008) 205213

and

Theorem 3.4. Let

2

u11
6 w
6 u21
6
w
uij  6 .
6 .
4 .

u12

un1

un2

w
u22

..
.
w

3
w



u1n


..
.


w
u2n

~t11 ~t12
6 ~t
6 21 ~t22
e 6
T
..
6 ..
4 .
.
~tn1 ~tn2

7
7
7
;
.. 7
7
. 5

uw
nn

where et ij 00ij ; m00ij ; u00ij , then

then ij  Xw , mij  Xwm , and uij  Xwu .

Proof. The proof is a straightforward consequence of formula (4) and matrix multiplication. h
Theorem 3.2. It holds that limitw!1 Xwu O
1
limitk!1 I Xu X2u    Xku I  Xu .

3
   ~t1n
   ~t2n 7
7
.. 7
..
7;
. . 5
   ~tnn

Matrix00ij  X  I  X 1 ;

1

Matrixm00ij  Xm  I  Xm ;
Matrixu00ij 

1

Xu  I  Xu :

9
10

and
Proof. By formula (7), Theorems 3.1 and 3.3, we have

Proof. Adding a row and a column to matrix Xu, we get

the augmented matrix Xaug as

Matrix00ij  limitw!1 X X2    Xw
limitw!1 X X X2    Xw
1

X  I  X :
Formulas (9) and (10) can be proved with analogous
procedures. h

where u01;n1 ; u02;n1 ; . . . ; u0n;n1 are suitable values that make

Xaug a stochastic
P P matrix.
Since ni1 nj1 uij < n, at least one of u01;n1 ; u02;n1 ; . . . ;
u0n;n1 must be positive. Elements of Xu, except the principal
diagonal elements, are positive because the right parts of
triangular fuzzy numbers of linguistic values are positive.
Hence, Xaug is a stochastic matrix of an absorbing Markov
chain and matrix Xu is the sub-stochastic matrix of Xaug.
Now Xu is a sub-stochastic matrix of absorbing Markov
chain (Goodman, 1988; Papoulis and Pillai, 2002), so that
we get limitw!1 Xwu O and limitw!1 I Xu X2u   
Xwu I  Xu 1 . h
Let q(Xu) denote the spectral radius of matrix Xu. Then
limitw!1 Xwu ! O is a sucient and necessary condition of
q(Xu) < 1 (Goodman, 1988; Papoulis and Pillai, 2002). This
property helps to prove the next theorem.
Theorem 3.3. The following properties holds:limitw!1
Xw O, limitw!1 Xwm O, limitw!1 I X X2   
Xw I  X 1 and limitw!1 I Xm X2m    Xwm
I  Xm 1 .
Proof. Since O 6 X 6 Xm 6 Xu, it will be that
q(X) 6 q(Xm) 6 q(Xu) < 1 (Papoulis and Pillai, 2002).
We, then, have proved the theorem. h
According to the crisp case, we dene the total-relation
e as
fuzzy matrix T


e limitw!1 X
ew :
e X
e2  X
T

e is acquired, it is easy to calculate D

ei R
ei
Now that T
e
e
e
e
and D i  R i , where D i , R i are the sum of rows and the
e respectively.
sum of columns of T,
To acquire the causal diagram, we apply the CFCS
method, i.e., formula (5), for defuzzication. The causal
diagram and analysis are then able to construct like the
ei R
e i and
crisp DEMATEL method just by replacing ( D
def
def
ei  R
e i with ( D
ei R
e i and ( D
ei  R
e i , respectively.
(D
4. Practical case of the proposed method
The sustainable competitive advantage of a company
largely depends on the ability of research and development
(R&D) that greatly aects the eectiveness and performance of developing new products and processes. In this
sense, a sound R&D investment is important for a companys future. However, the evaluation of R&D project is
a kind of fuzzy group decision-making problem involved
with multiple criteria (Mohanty, Agarwal, Choudhury,
and Tiwari, 2005). In this section, an empirical study shows
how a high-tech company applied the proposed solution to
develop a structural model for evaluating R&D projects.
The case Company H is a Taiwanese rm with more
than 2350 employees in 18 countries with a turnover in
excess of US$ 476 million. The company is performing as
a leading solution provider in the industrial automation
market, oering more than 420 products and solutions
ranging from system-integration hardware and software
to customer-driven service. In order to cope with challenges
from newcomers and to maintain continued competitiveness in terms of shorter lead-time, high quality, competitive
prices, and improved customer service, Company H
intended to rebuild an evaluation model for R&D project

C.-J. Lin, W.-W. Wu / Expert Systems with Applications 34 (2008) 205213

selection. To that end, Company H successfully utilized the

proposed solution with the ve-step procedure to achieve
its problem-solving.
Step 1: Set the decision goal and set up a committee.
For profoundly evaluating R&D projects, Company H sets up an R&D investment committee
consisting of the general manager, and several
managers representing marketing, nancial, production, human resource, R&D, and informationtechnology departments.
Step 2: Develop evaluation criteria and design the fuzzy
linguistic scale.
First, it is necessary to gather the relevant criteria
in order to create a structural model. Through
the literature review and profound discussion, the
R&D investment committee nally adopted 11 criteria and fuzzy linguistic scale (see Table 1) from
Meade and Presley (2002). These 11 criteria were:
probability of technical success (C1), existence of
project champion (C2), existence of required competence (C3), availability of available resources
(C4), applicability to other products and processes
(C5), time to market (C6), probability of market
success of products (C7), potential size of market
(C8), products life cycle (C9), net present value

211

(C10), and strategic t (C11). All of these 11 criteria

were deemed to be signicant and indispensable.
Hence, they employed the fuzzy DEMATEL
method for capturing the complex relationships
among these evaluation criteria.
Step 3: Get the assessments of decision-makers.
Through the fuzzy linguistic scale (see Table 1), the
relationships between each pair of criteria were
measured and each individual assessment of
seven experts can be obtained. For example, the
assessment data of the General Manager is
shown in Table 2. Then, we get fuzzy matrices
e h1i ; Z
e h2i ; . . . ; Z
e h7i through the correspondence of
Z
Table 1.
Step 4: Calculate the normalized direct-relation fuzzy
matrix.
e h1i ; Z
e h2i ; . . . ; Z
e h7i , the
Based on the fuzzy matrices Z
e hki of
normalized direct-relation fuzzy matrix X
expert k; k = 1, 2, . . ., 7, are obtained. Formula (6)
e h1i ; X
e h2i ; . . . ;
is used to average fuzzy matrices X
h7i
e
X
and obtains the normalized direct-relation
e
fuzzy matrix X.
Step 5: Establish and analyze the structural model.
e was
Finally, the total-relation fuzzy matrix T
acquired by formulas (8)(10). Then, the values
e i, D
e i, D
ei R
e i and D
ei  R
e i can be acquired
of R

Table 2
The linguistic assessment data of the General Manager

C1
C2
C3
C4
C5
C6
C7
C8
C9
C10
C11

C1

C2

C3

C4

C5

C6

C7

C8

C9

C10

C11

L
VH
H
VL
VL
L
H
L
H
VH

H
H
H
VL
H
L
L
H
H

VL
VL

H
L
L
L
H
L
L
H

No
VL
No

VL
VL
L
L
VL
VL
H

VH
L
L
H

H
H
L
L
VL
VH

H
VL
L
L
L

L
H
L
L
H

H
H
VL
L
L
VH

VH
L
H
H

L
VL
VL
VL
L
VL
VH

L
VL
L

VH
H
VL
L
L
L
L
VH

VL
VH

H
VH
VL
VL
L
H
H
H
H

VH

L
H
L
L
VL
H
L
VL
L
L

Table 3
e i, D
e i, D
ei R
e i and D
ei  R
ei
The values of R
ei
R
C1
C2
C3
C4
C5
C6
C7
C8
C9
C10
C11

(0.605,
(0.555,
(0.422,
(0.314,
(0.526,
(0.567,
(0.610,
(0.397,
(0.414,
(0.597,
(0.355,

ei
D
1.576,
1.505,
1.301,
1.106,
1.466,
1.540,
1.608,
1.272,
1.290,
1.587,
1.179,

4.656)
4.659)
4.170)
3.857)
4.548)
4.570)
4.718)
4.108)
4.139)
4.705)
4.025)

(0.687,
(0.501,
(0.464,
(0.355,
(0.388,
(0.443,
(0.524,
(0.595,
(0.365,
(0.356,
(0.684,

ei R
ei
D
1.721,
1.436,
1.352,
1.199,
1.248,
1.319,
1.463,
1.583,
1.203,
1.192,
1.714,

4.924)
4.493)
4.264)
4.046)
4.051)
4.236)
4.512)
4.698)
4.018)
4.030)
4.885)

(1.292,
(1.057,
(0.886,
(0.670,
(0.914,
(1.010,
(1.134,
(0.992,
(0.779,
(0.953,
(1.039,

ei  R
ei
D
3.296,
2.942,
2.653,
2.305,
2.714,
2.859,
3.071,
2.856,
2.494,
2.779,
2.893,

9.580)
9.152)
8.434)
7.903)
8.599)
8.806)
9.230)
8.807)
8.157)
8.735)
8.910)

(3.969,
(4.158,
(3.706,
(3.502,
(4.193,
(4.182,
(4.275,
(3.584,
(3.544,
(4.340,
(3.669,

0.145, 4.319)
0.069, 3.938)
0.051, 3.843)
0.092, 3.732)
0.218, 3.525)
0.221, 3.668)
0.145, 3.902)
0.311, 4.301)
0.087, 3.605)
0.395, 3.433)
0.535, 4.530)

212

C.-J. Lin, W.-W. Wu / Expert Systems with Applications 34 (2008) 205213

0.6

C11

0.5
0.4

C8

0.3

C1

0.2

C4
0.1

C3

D-R
0
0

0.5

1.5

2.5

3.5

C9

4.5

C2

-0.1

C6

-0.2

C7

C5

-0.3
-0.4

C10

-0.5
D+R

Fig. 4. The causal diagram of criteria.

Table 4
e i def and D
ei  R
e i def
ei R
The values of D
C1
C2
C3
C4
C5
C6
C7
C8
C9
C10
C11

ei R
e i def
D

ei  R
e i def
D

4.723
4.383
3.991
3.626
4.076
4.225
4.478
4.218
3.810
4.155
4.281

0.165
0.096
0.063
0.107
0.295
0.245
0.173
0.343
0.009
0.434
0.465

(Table 3). The causal diagram (Fig. 4) could be

ei R
e i def ;
plotted by mapping a dataset of D
def
ei  R
e i (Table 4).
D
As shown in the causal diagram, the evaluation criteria
were visually divided into the cause group including C1, C3,
C4, C8, and C11, while the eect group was composed of
criteria C2, C5, C6, C7, C9 and C10. From the causal diagram, valuable cues are obtained for making profound
decisions. For example, if Company H wishes to reach a
high level of performance in terms of the eect group criteria, they must control and pay much attention to the cause
group criteria. Particularly, the criteria of strategic t
(C11) and potential size of market (C8) accumulate to
ei  R
e i def , these
70.7% of the whole positive values of D

two have the best eect on the other criteria and should
catch the rst concern. The criterion of probability of
ei R
e i def , is
technical success (C1), with the largest D
the most important factor for R&D project selection. By
contrast, the net present value (C10), with the most negei  R
e i def , is the most easily improved of
ative value of D
the eect group criteria. Additionally, these two cause
and eect groups may be further used to, respectively, serve
as causal criteria and eective criteria clusters in an
MCDM model such as the AHP or ANP method (Saaty,
1990, 1996) for selecting the optimal alternative.
5. Conclusions
Decision-makers look forward to a causal analytical
method which can do with the group decision-making
problem in the fuzzy environments of practical life. Hence,
we have developed a fuzzy DEMATEL method. With the
proposed methodology, the complex interactions between
criteria can be transformed into a visible structural model,
making it easier to capture the core of a problem, whereby
excellent decisions can be made.
The proposed fuzzy DEMATEL method can also be
applied to problems such as manufacture, environmental
engineering, nancial analysis, social science and many
other decision-making problems with multiple criteria in
a fuzzy environment. Further research may represent linguistic variables by trapezoidal fuzzy numbers or other
types of membership functions.

C.-J. Lin, W.-W. Wu / Expert Systems with Applications 34 (2008) 205213

References
Fontela, E., Gabus, A. (1976). The DEMATEL Observer, DEMATEL
1976 Report. Switzerland, Geneva, Battelle Geneva Research Center.
Gabus, A., Fontela, E. (1973). Perceptions of the World Problematique:
Communication Procedure, Communicating With Those Bearing
Collective Responsibility, DEMATEL Report No. 1, Battelle Geneva
Research Centre, Geneva, Switzerland.
Goodman, R. (1988). Introduction to stochastic models. Monlo Park,
California: Benjamin/Cummings Publishing Company.
Hess, P., & Siciliano, J. (1996). Management: Responsibility for performance. New York: McGraw-Hill.
Hori, S., & Shimizu, Y. (1999). Designing methods of human interface for
supervisory control systems. Control Engineering Practice, 7(11),
14131419.
Kaufmann, A., & Gupta, M. M. (1991). Introduction to fuzzy arithmetic
Theory and applications. New York: Thomson Computer Press.
Laarhoven, P. J. M. V., & Pedrycz, W. (1983). A fuzzy extension of
Saatys priority theory. Fuzzy Sets and Systems, 11, 229241.
Li, R. J. (1999). Fuzzy method in group decision making. Computers and
Mathematics with Applications, 38(1), 91101.
Meade, L. M., & Presley, A. (2002). R & D project selection using the
analytic network process. IEEE Transactions on Engineering Management, 49(1), 5966.

213

Mohanty, R. P., Agarwal, R., Choudhury, A. K., & Tiwari, M. K. (2005).

A fuzzy ANP-based approach to R&D project selection: A case study.
International Journal of Production Research, 43(24), 51995228.
Opricovic, S., & Tzeng, G. H. (2003). Defuzzication within a multicriteria decision model. Journal of Uncertainty, Fuzziness and Knowledgebased Systems, 11(5), 635652.
Opricovic, S., & Tzeng, G. H. (2004). Compromise solution by MCDM
methods: A comparative analysis of VIKOR and TOPSIS. European
Journal of Operational Research, 156(2), 445455.
Papoulis, A., & Pillai, S. U. (2002). Probability, random variables, and
stochastic processes. New York: McGraw-Hill.
Saaty, T. L. (1990). The analytic hierarchy process: Planning, priority
setting, resource allocation. Pittsburgh: RWS Publication.
Saaty, T. L. (1996). Decision making with dependence and feedback, the
analytic network process. Pittsburgh: RWS Publication.
Tamura, H., Akazawa, K., Nagata, H. (2002). Structural modeling of
uneasy factors for creating safe, secure and reliable society. SICE
System Integration Division Annual Conference, Japan. pp. 330340.
Yamazaki, M., Ishibe, K., & Yamashita, S. (1997). An analysis of
obstructive factors to welfare service using DEMATEL method.
Reports of the Faculty of Engineering, Yamanashi University, 48, 2530.
Zadeh, L.A. (1975). The concept of a linguistic variable and its application
to approximate reasoning. Information Science 8, 199249 (I), 301357
(II).