Applied Soft Computing 9 (2009) 377–386

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

Fuzzy hierarchical TOPSIS for supplier selection

Jia-Wen Wang a,*, Ching-Hsue Cheng b, Huang Kun-Chengb
a
Department of Electronic Commerce Management, Nanhua University, 32, Chung Kcng Li, Dalin, Chiayi, 62248, Taiwan
b
Department of Information Management, National Yunlin University of Science and Technology, 123, section 3, University Road, Touliu, Yunlin 640, Taiwan

A R T I C L E I N F O A B S T R A C T

Article history: This study simplifies the complicated metric distance method [L.S. Chen, C.H. Cheng, Selecting IS
Received 25 April 2006 personnel using ranking fuzzy number by metric distance method, Eur. J. Operational Res. 160 (3) 2005
Received in revised form 14 April 2008 803–820], and proposes an algorithm to modify Chen’s Fuzzy TOPSIS (Technique for Order Preference by
Accepted 20 April 2008
Similarity to Ideal Solution) [C.T. Chen, Extensions of the TOPSIS for group decision-making under fuzzy
Available online 29 April 2008
environment, Fuzzy Sets Syst., 114 (2000) 1–9]. From experimental verification, Chen directly assigned
the fuzzy numbers 1̃ and 0̃ as fuzzy positive ideal solution (PIS) and negative ideal solution (NIS). Chen’s
Keywords:
method sometimes violates the basic concepts of traditional TOPSIS. This study thus proposes fuzzy
Decision analysis
Supplier selection
hierarchical TOPSIS, which not only is well suited for evaluating fuzziness and uncertainty problems, but
Fuzzy hierarchical TOPSIS also can provide more objective and accurate criterion weights, while simultaneously avoiding the
FAHP (Fuzzy Analytic Hierarchy Process) problem of Chen’s Fuzzy TOPSIS. For application and verification, this study presents a numerical
example and build a practical supplier selection problem to verify our proposed method and compare it
with other methods.
ß 2008 Elsevier B.V. All rights reserved.

1. Introduction (3) Outranking models

ELECTRE and PROMETHEE are two main families of method
Multi-criteria decision-making (MCDM) methods are formal in this group.
approaches to structure information and decision evaluation in
problems with multiple, conflicting goals. MCDM can help users AHP was first proposed by Saaty [20], and has been applied in
understand the results of integrated assessments, including several areas of social sciences and management, such as project
tradeoffs among policy objectives, and can use those results in a management [15], and military applications [6], etc. AHP integrates
systematic, defensible way to develop policy recommendations. experts’ opinions and evaluation scores, and devises the complex
MCDM methods have been widely used in many research fields. decision-making system into a simple elementary hierarchy system.
Different approaches have been proposed by many researchers, The evaluation method in terms of ratio scale is then employed to
including the Analytic Hierarchy Process (AHP) [20], Technique for perform relative importance pair-wise comparison among every
Order Preference by Similarity to Ideal Solution (TOPSIS) [13] and criterion. This method decomposes complicated problems from
MCDM [13,18]. According to Løken [16], existing MCDA methods higher hierarchies to lower ones.
can be classified into three broad categories: The concept of TOPSIS is that the most preferred alternative
should not only have the shortest distance from the positive ideal
(1) Value measurement models solution (PIS), but should also be farthest from the negative ideal
AHP and multiattribute utility theory (MAUT) are the best solution (NIS). Hwang and Yoon [13] also described the TOPSIS
known method in this group. concept, referring to the positive and negative ideal solutions as
(2) Goal, aspiration and reference level models the ideal and anti-ideal solutions, respectively. Numerous applica-
Goal programming (GP) and TOPSIS are the most important tions of TOPSIS exist, including airline performance evaluating [2]
methods that belong to the group. and optimal material selection [14].
AHP and TOPSIS possess advantages in that they are easy to
compute and easily understood, because the methods are directly
* Corresponding author. Tel.: +886 5 5342601x5312; fax: +886 5 531 2077.
giving a definite value by experts to calculate their final results.
E-mail addresses: [email protected] (J.-W. Wang), Though AHP is designed to capture expert knowledge/opinions,
[email protected] (C.-H. Cheng), [email protected] (K.-C. Huang). the conventional AHP does not reflect human thinking style. The

1568-4946/$ – see front matter ß 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2008.04.014
378 J.-W. Wang et al. / Applied Soft Computing 9 (2009) 377–386

linguistic expressions of fuzzy theory are regarded as natural If two metric spaces hX, ri and hY, vi form a new metric space
representations of preferences/judgments. Characteristics such as called Cartesian product X  Y, its point set is a set of X  Y = {hx, yi:
satisfaction, fairness and dissatisfaction indicate the applicability x 2 X, y 2 Y}, and its metric t is defined as follows:
of fuzzy set theory in capturing the preference structures of
1=2
decision makers, while fuzzy set theory aids in measuring the t ð < x1 ; y1 > ; < x2 ; y2 > Þ ¼ ½rðx1 ; x2 Þ2 þ vðy1 ; y2 Þ2  (2)
ambiguity of concepts associated with subjective human judg-
For convenience, the fuzzy number can be denoted using [a, b, c,
ments. Fuzzy MCDM theory thus can strengthen the comprehen-
d; 1], and the membership function (MF) f à of the fuzzy number
siveness and reasonableness of the decision-making process.
à ¼ ½a; b; c; d; 1; can be expressed as
Decision makers usually are more confident making linguistic
judgments than crisp value judgments. This phenomenon results 8 L
>
> f ðxÞ; a  x  b;
from inability to explicitly state their preferences owing to the fuzzy < Ã
1; b  x  c;
nature of the comparison process. Many studies have continually f à ¼ (3)
>
> f R ðxÞ; c  x  d;
: Ã
introduced the fuzzy concept to improve MCDM and solve linguistic 0; otherwise;
and cognitive fuzziness problems. For example, fuzzy theory and
AHP are combined to become the Fuzzy AHP (FAHP) method where fÃL : ½a; b ! ½0; 1 and fÃR : ½c; d ! ½0; 1. Since fÃL :
[3,4,6,19], which is a fuzzy extension of AHP, and was developed to ½a; b ! ½0; 1 is continuous and strictly increasing, the inverse
solve hierarchical fuzzy problems. FAHPs are systematic approaches function of fÃL exists. Similarly, fÃR : ½c; d ! ½0; 1 is continuous and
to the alternative selection and justification problem that use the strictly decreasing, the inverse function of fÃR also exists.
concepts of fuzzy set theory and hierarchical structure analysis. From Cheng [5], the inverse functions of fÃL and fÃR are denoted
FAHP can be applied to measure fuzzy linguistic cognition, and by gÃL and gÃR , respectively. Since fÃL : ½a; b ! ½0; 1 is continuous and
suffers form the disadvantage of unstable (i.e., non-unique) results strictly increasing, gÃL : ½a; b ! ½0; 1 is also continuous and strictly
being obtained by different defuzzification methods, and the increasing. Similarly, if fÃR : ½c; d ! ½0; 1 is continuous and strictly
ordering of alternatives will arise ranking reversion. decreasing, then gÃR : ½a; b ! ½0; 1 is also continuous and strictly
Chen [7] extended TOPSIS to fuzzy environments; this extended decreasing; gÃL and gÃR are continuous on a closed interval [0, 1] and
R1 R1
version used fuzzy linguistic value (represented by fuzzy number are integrable on [0, 1]. That is, both 0 gÃL ðyÞdy and 0 gÃR ðyÞdy
[11]) as a substitute for the directly given crisp value in grade exist.
assessment. This modified TOPSIS is a practical method and fits From Eq. (3), the metric can be adapted to fuzzy numbers, and
human thinking under actual environment. The criteria weighting can be redefined as:
( (
of Chen are directly provided by experts, and the expert weightings fÃL ðxÞ for x  m1 fB̃L ðxÞ for x  m2
are then averaged, also the calculated results sometimes violate f à ðxÞ ¼ ; f B̃ ðxÞ ¼
fÃR ðxÞ for x  m1 fB̃R ðxÞ for x  m2
the basic concept of TOPSIS. Therefore, this study proposes fuzzy
hierarchical TOPSIS, which uses simplified parameterized metric then
distance and FAHP to modify Chen’s Fuzzy TOPSIS to overcome the "Z #1=2
1 Z 1
above disadvantages. 2 2
DðÃ; B̃Þ ¼ ðgÃL  gB̃L Þ dy þ ðgÃR  gB̃R Þ dy (4)
The remainder of this paper is organized as follows: Section 2 0 0
briefly describes the metric distance method, and the parameters
of metric distance, linguistic variable and defuzzification. Next, denotes the distance of ÃðxÞ and B̃ðxÞ, and the inverse functions of
Section 3 simplifies the parameterized metric distance and fÃL and fÃR are gÃL and gÃR .
proposes an algorithm of fuzzy hierarchical TOPSIS, and illustrates Eq. (4) is the generalized ranking fuzzy number method, and
a numerical example to verify the proposed method. Section 4 then our method can be applied to the negative and positive ranking of
presents a practical problem involving supplier selection to verify fuzzy numbers. For convenience of ranking n positive fuzzy
the proposed method and compare it with other methods. The final numbers Ã1 ; Ã2 ;    ; and Ãn , let B̃ðxÞ ¼ 0 in Eq. (4), so that D can be
section presents conclusions. rewritten as:
"Z Z #1=2
1 1
2. Preliminary 2 2
DðÃi ; 0Þ ¼ ðgÃL Þ dy þ ðgÃR Þ dy (5)
i i
0 0
This section briefly describes the metric distance method,
and the parameters of metric distance, linguistic variable and The ranking of the fuzzy number increases with the value of
defuzzification. DðÃi ; 0Þ.

2.1. Metric distance method 2.2. The parameterized of metric distance

This study used the metric distance method of Chen and Cheng From the metric distance of Chen and Cheng [8], this study
[5,8] to calculate the dispersion between each point and FPIS and briefly describes the concept of minimum metric D and Eq. (10) to
FNIS. The method of Chen and Cheng thus is briefly reviewed calculate the parameter of fuzzy mean x0(m = x0) and fuzzy
below: standard deviation s (a = b = s) as follows:
A metric space hX, ri in which X is a nonempty and r is a real Let à denote is a generalized LR fuzzy number and let the
number function, both X and r are on the X  X, where x, y, and inverse functions of fÃL and fÃR be gÃL ¼ ðx0  s Þ þ s y and
0
z 2 X: gÃR ¼ ðx0 þ s Þ  s y. The symmetry function à s of S[x0,s] can be
8 obtained by immunizing this metric D, namely:
>
> i:rðx; yÞ  0;
< Z 1
ii:rðx; yÞ ¼ 0 if only if x ¼ y; 2
(1) DðÃ; S½x0; s Þ ¼ ðgÃL  S½x0 ; s L Þ dy
>
> iii:rðx; yÞ ¼ rðy; xÞ; 0
:
iv:rðx; yÞ  rðx; zÞ Z 1
2
þ ðgÃR  S½x0 ; s R Þ dy (6)
Then function r is called a metric. 0
 J.-W. Wang et al. / Applied Soft Computing 9 (2009) 377–386 379

Table 1
The MF of linguistic variables

Linguistic variable Rating

Very unsatisfied (VU) (1,1,3)

Unsatisfied (U) (1,3,5)
Fair (F) (3,5,7)
Satisfy (S) (5,7,9)
Very satisfy (VS) (7,9,9)

and Table 1 explains the numerical values of the membership

functions.
Fig. 1. The MF of linguistic variable.
2.4. Defuzzification

S[x0,s] is an LR type, while S[x0,s]L is the left membership Defuzzification is the conversion of practice quantities to fuzzy
function and S[x0,s]R is the right membership function. quantities. The centroid method (also called the center of area or
Then, setting @D/@x0 = 0 and, @D/@s = 0 in Eq. (6), we obtain: center of gravity method) is the most popular defuzzification
8 R1 R L R1 method [19]; it is given by the following algebraic expression:
>
< s ¼ 0 ðgà gà Þð1yÞdy/2 0 ð1yÞdy R
Z 1 (7) m ðzÞzdz
>
: x0 ¼ 1/2 ðgÃR þ gÃL Þdy z ¼ R C (10)
mC ðzÞdz
0

From Eq. (7) the ÃðxÞ ¼ ða; b; c; dÞ can be used to obtain an

approximate parameter of the symmetry fuzzy number: 3. Fuzzy hierarchical TOPSIS
8
> 2ðd  aÞ þ c  b This section is divided into two parts: (1) simplification of the
<s ¼
4 (8) parameterized metric distance method, (2) algorithm of fuzzy
:u ¼ a þ b þ c þ d
> hierarchical TOPSIS.
4
If b = c means this fuzzy number is a degenerative trapezoid 3.1. Simplifying parameterized metric distance
which means à is a triangular fuzzy number, the parameter of the
most approximate symmetry fuzzy number is: This section simplifies parameterized metric distance for
8 parameterized fuzzy numbers. For ease of understanding, trian-
> 2ðd  aÞ gular fuzzy number is used to simplify parameterized metric
<s ¼
4 (9) distance, an approach that can also be used to simplify other
: u ¼ a þ 2b þ d
>
parameterized fuzzy numbers. The simplified method can be
4
described as follows:
Eqs. (7)–(9) adapt both symmetry and non-symmetry fuzzy Let à denote an TFNs, ÃðxÞ ¼ ða; b; cÞ, and rewrite Eq. (11) in
numbers. The mean u is thus used to rank the fuzzy numbers. Lee parameterized form:
and Li rank fuzzy numbers based on two different criteria, namely, 8xa
>
fuzzy mean and fuzzy spread, and Lee and Li pointed out that < b  a; a  x  b;
>
human intuition favors fuzzy numbers with the following ˜
fà ¼ 1; x ¼ b; (11)
>
characteristics: high mean value and low spread. When the means : c  x; b  x  c;
>
of fuzzy numbers are equal, the method of Lee and Li can be used to cb
rank the fuzzy number by Eqs. (7)–(9). Then, the inverse functions of fÃL and fÃR are

2.3. Linguistic variable gÃL ¼ a þ ðb  aÞy (12)

A linguistic variable is ‘‘a variable whose values are words or gÃR ¼ c þ ðb  cÞy (13)
sentences in a natural or artificial language.’’ [25,26] For instance,
some matters are characterized by linguistic terms in nature, such From (4), (12) and (13), let two TFNs be ÃðxÞ ¼ ða; b; cÞ,
as unsatisfied, fair, and satisfy. Each linguistic variable can be B̃ðxÞ ¼ ðd; e; f Þ, and DðÃ; B̃Þ can be simplified as Eq. (14) via the
assigned one or more linguistic values, which are in turn connected p-norm metric method:

"Z Z 1= p
1 1
p p
DðÃ; B̃Þ ¼ ½ða þ ðb  aÞyÞ  ðd þ ðb  dÞyÞ dy þ ½ðc þ ðb  cÞyÞ  ð f þ ðe  f ÞyÞ dy (14)
0 0

to a numeric value through the mechanism of membership (1) When p = 1, (14) can be simplified as Eq. (15):
functions. From Miller [17] the number of terms that a person is  
 da f  c
able to discriminate is 7  2. This study adopts five fuzzy linguistic DðÃ; B̃Þ ¼ ða þ b þ cÞ  ðd þ e þ f Þ þ þ (15)
2 2 
terms by triangular fuzzy numbers to express un-quantified matters.
Fig. 1 shows the membership function of linguistic variables, (2) When p = 2, (14) can be simplified as Eq. (16):
380 J.-W. Wang et al. / Applied Soft Computing 9 (2009) 377–386

" #1=2
2 2
2 ½ðe  dÞ  ðb  aÞ 2 ½ðe  f Þ  ðb  cÞ
DðÃ; B̃Þ ¼ ðd  aÞ þ þ ðd  aÞ½ðe  dÞ  ðb  aÞ þ ð f  cÞ þ þ ð f  cÞ½ðe  f Þ  ðb  cÞ (16)
3 3

The proposed algorithm uses Eq. (15) to calculate the distance Am), and establish a hierarchical structure to obtain
between each point and FPIS and FNIS via the metric distance weight.
method presented in Section 3.2. Step 2: Use pair-wise comparison to get the degree of importance
of all criteria, and evaluate all of the alternatives under
3.2. An algorithm for fuzzy hierarchical TOPSIS each criterion, then ask experts to assign the alternatives
an appropriate fuzzy number based on the linguistic
During 2000, Chen extended the traditional TOPSIS to apply it to variable to form a fuzzy judgment matrix.
fuzzy environments, and renamed it the Fuzzy TOPSIS Method. Step 3: Use the Lambda-Max method [9] to calculate the fuzzy
This extended method suffered some weaknesses, as follows: weight (FAHP) of each criterion given by the experts. The
detailed procedure is shown in step 3 of Section 4.
(1) The need to assign an initial weight to each criterion. Step 4: Check the consistency index (C.I.).
(2) When fuzzy numbers 1̃, 0̃ are directly assumed to be the fuzzy The consistency index of the comparison matrix is
PIS and NIS, respectively, and when the weighted and graded C.I. = (lmax  n)/(n  1), where lmax denotes the largest
values are extremely small, then the distance between criterion eigenvalue of matrix Y, Y represents the reciprocal matrix
and the fuzzy PIS and NIS is increased. The result will lie outside of pair-wise comparisons yij, and yij is the importance of
the range [0,1]. alternative i over alternative j.
(3) The result sometimes does not conform to the basic conception Step 5: Through the geometry average method, integrate all the
that the best solution should be that nearest PIS and farthest expert opinions to obtain a fuzzy weight for every
from NIS. aggregative criterion.
1=k
To overcome above problems, this study proposes fuzzy w̃ j ¼ ðw̃1 w̃2 . . . w̃k Þ ;
(17)
hierarchical TOPSIS (as Fig. 2) for solving the method of Chen. k
where w̃ is a parameterized fuzzy number
The proposed model has four main components, as follows:
Step 6: Establish a normalized fuzzy performance matrix.
(1) FAHP uses a hierarchical structure to calculate the fuzzy weight Use the linguistic values, shown in Table 1, to evaluate
of each criterion. all of the alternatives under each criterion. After the
(2) TOPSIS uses the criterion characteristics to establish a normal- evaluation, apply the geometry average method to
ized fuzzy performance matrix and then multiplies all the integrate all of the opinions of experts and calculate
criterion weights to form a normalized weight performance them as follows:
matrix. 1=k
x̃i j ¼ ðx̃1i j x̃2i j . . . x̃kij Þ (18)
(3) Obtain FPIS and FNIS, and apply the metric distance method to
respectively calculate the dispersion between the alternative
where x̃i j ¼ ðai j ; bi j ; ci j Þ is the fuzzy evaluation value of
value under each criterion, and under FPIS and FNIS.
each alternatives i for each criterion j, which can be used
(4) Finally, apply Euclidean distance to aggregative the dispersions
to obtain the positive fuzzy performance matrix.
to judge and get a best ranking.
R̃ ¼ ½r̃i j mn (19)
The algorithm for fuzzy hierarchical TOPSIS is as follows:
Besides, B and C can be allowed to respectively represent
Step 1: Confirm the evaluation criteria and alternatives of the the profit criterion set and cost criterion set, respectively,
decision-making problem, and establish a hierarchical and Eq. (20) can be used to obtain a value r̃i j which is
structure. within [0,1].
Supposed that K experts analyze a problem involv- 8 !
>
> ai j bi j c i j
ing n criteria (C1,C2,. . .,Cn) and m alternatives (A1,A2,. . ., >
> r̃ ¼ ; ; ; j2B
< ij cj cj cj
! ; where
>
> cj cj cj
>
> r̃ ¼ ; ; ; j 2 C
: ij ci j bi j ai j (20)
8
< c j ¼ max ci j if j 2 B
i
: cj ¼ min ai j if j2C
i

Step 7: Get the weighted normalized fuzzy performance matrix.

Ṽ ¼ ½ṽi j mn ; i ¼ 1; 2; . . . ; m; j ¼ 1; 2; . . . ; n (21)

where ṽi j ¼ r̃i j ðÞw̃ j and ṽi j ; 8 i; j are normalized

positive TFNs and have values in the range [0,1].
Step 8: Determine FPIS and FNIS as follows:
For selecting ṽ0i j s max and min, first rank the orderings
Fig. 2. Fuzzy hierarchy TOPSIS.
of fuzzy number ṽi j by the metric distance method, then
 J.-W. Wang et al. / Applied Soft Computing 9 (2009) 377–386 381

calculate FPIS and FNIS by Eq. (22). Table 3

Comparison matrix (fuzzy number of degree of importance)
0
FPIS : A ¼ fðmax ṽi j j j 2 JÞ; ðmin ṽi j j j 2 J Þ; i ¼ 1; 2; . . . ; mg
FNIS : A ¼ fðmin ṽi j j j 2 JÞ; ðmax ṽi j j j 2 J0 Þ; i ¼ 1; 2; . . . ; mg D1 S T C F D2 S T C F D3 S T C F

(22) S 1 5 1/3 1/4 S 1 4 1/2 1/3 S 1 6 1/2 1/5

T 1/5 1 1/5 1/7 T 1/4 1 1/4 1/6 T 1/6 1 1/6 1/8
where J ¼ f j ¼ 1; 2; . . . ; nj j 2 Bg, J 0 ¼ f j ¼ 1; 2; . . . ; nj j 2 Cg, C 3 5 1 1/2 C 2 4 1 1/2 C 2 6 1 1/3
F 4 7 2 1 F 3 6 2 1 F 5 8 3 1
while FPIS and FNIS can be expressed as
A ¼ ðṽ1 ; ṽ2 ; . . . ; ṽ j ; . . . ; ṽn Þ, A ¼ ðṽ   
1 ; ṽ2 ; . . . ; ṽ j ; . . . ; ṽn Þ.
Step 9: Calculate the distance between each point and FPIS and
Table 4
FNIS by the metric distance method. Assessment grades given by decision-makers
Using Eq. (15) it is possible to obtain DðÃ; B̃Þ, and the
Size (S) Transportation Condition (C) Finance (F)
metric distance between the fuzzy evaluating value and (T)
FPIS and FNIS can be calculated by Eq. (23):
X Y Z X Y Z X Y Z X Y Z

hi j ¼ Dðṽi j ; ṽ j Þ; h
i j ¼ Dðṽi j ; ṽ j Þ (23)
D1 VS S VU U VU VS U S VS U VS F
Step 10: Apply the Euclidean distance method to aggregate all of D2 VS S VU F U VS U VS S F S F
D3 S F VU F VU VS S S F VS S
the criteria for each alternative.
According to ref. [1], the Euclidean distance method
can be applied to aggregate the metric distance of FPIS to Table 5
all criteria for every alternative, and similarly for FNIS. Normalized fuzzy performance matrix
The calculation formulas are as follows: Size Transportation Condition Finance
2 31=2
Xn
2
X (0.7,0.92,1) (0.23,0.47,0.7) (0.16,0.4,0.62) (0.23,0.47,0.7)
Si ¼ 4 ðhi j Þ 5 ; i ¼ 1; 2; . . . ; m (24) Y (0.47,0.7,0.92) (0.11,0.16,0.4) (0.62,0.85,1) (0.7,0.92,1)
Z (0.11,0.11,0.33) (0.78,1,1) (0.62,0.85,1) (0.4,0.62,0.85)
2 j¼1 31=2
X n
 4  25
Si ¼ ðhi j Þ ; i ¼ 1; 2; . . . ; m (25)
j¼1 Table 6
Weighted normalized fuzzy performance matrix
Step 11: Select the best alternative.
Size Transportation Condition Finance
(1) Rule 1: When Si is smaller, meaning Ai is nearer to the
best A*, i.e., Ai is the best solution. X (0.1,0.15,0.18) (0.01,0.02,0.04) (0.03,0.11,0.19) (0.09,0.24,0.37)
(2) Rule 2: When S i is bigger, meaning that the Ai is
Y (0.07,0.11,0.17) (0.01,0.01,0.02) (0.13,0.23,0.31) (0.28,0.47,0.53)
Z (0.02,0.02,0.06) (0.04,0.05,0.05) (0.13,0.23,0.31) (0.16,0.32,0.45)
farthest to the worst one A, i.e., Ai is the best solution.

3.3. Numerical example and comparison Table 7

FPIS and FNIS

Assume three decision-makers (D1, D2, D3) that plan to select Size Transportation Condition Finance
the best building from X, Y, Z as their office, where the selection
FPIS (0.1,0.15,0.18) (0.02,0.02,0.06) (0.01,0.01,0.02) (0.28,0.47,0.53)
criteria are size (S), transportation (T), condition (C) and finance (F). FNIS (0.02,0.02,0.06) (0.01,0.01,0.02) (0.03,0.11,0.19) (0.09,0.24,0.37)
These three decision-makers respectively compare the four criteria
and evaluate their degree of satisfaction with every building.
The compared and evaluated grades are shown in Table 3 (see (3) Use step 7 to compute the weighted normalized fuzzy
Tables 1 and 2 for the linguistic value and degree of importance). performance matrix, as shown in Table 6.
The procedure for selecting the best building by the proposed (4) Ranking Table 6 by the metric distance method to determine
algorithm is listed as follows: the maximum and minimum, then calculate FPIS and FNIS by
step 8, with the results being listed in Table 7.
(1) From Table 3 and steps 1–5 of the proposed algorithm, an agg- (5) From step 9, use (23) to calculate the distance between all of
regative fuzzy weight value can be obtained W̃ ¼ ½ð0:14; 0:16; the criterion and FPIS and FNIS, and then apply the Euclidean
0:18Þ; ð0:05; 0:05; 0:05Þ; ð0:21; 0:27; 0:31Þ; ð0:40; 0:51; 0:54ÞT . distance method in step 10 to aggregative all the criteria values.
(2) Use (18)–(20) of step 6 to calculate Table 4, and then list the (6) Finally, using rule 1 in step 11, choose building Y as the best
results in Table 5. solution, as listed in Table 8.

Table 2 3.3.1. Comparison with other methods

Fuzzy numbers for degree of importance Next, the proposed method is compared with other methods,
with Table 9 listing the results of the comparison. The proposed
Degree of importance Linguistic variable Positive TFN

1̃ Equal importance (1,1,1) Definite value

Table 8
2̃ Weak (1,2,4)
The result
3̃ Moderate importance (1,3,5)
4̃ Moderate plus (2,4,6) Fuzzy hierarchical TOPSIS (proposed) S* S Ordering
5̃ Strong importance (3,5,7)
6̃ Strong plus (4,6,8) X 0.47 0.23 3
7̃ Very strong importance (5,7,9) Y 0.1 0.5 1
8̃ Very very strong (6,8,9) Z 0.34 0.28 2
9̃ Extreme importance (7,9,9) a
Y is the best solution.
382 J.-W. Wang et al. / Applied Soft Computing 9 (2009) 377–386

Table 9 method and the Fuzzy TOPSIS Method [7] can solve problems in
Comparison with other method
uncertain environments. However, TOPSIS suffers some disadvan-
Aggregative value Ordering tages in terms of its calculation methods and results. Based on
*  review of the best solution calculated by Chen [7], although the
S S
result is the same, the dispersion between the calculated value and
Proposed method X 0.47 0.23 Y>Z>X FPIS is 4.81, compared to 1.35 for FNIS. Therefore Chen’s solution
Y 0.1 0.5
does not accord with the conception of TOPSIS, namely that the
Z 0.34 0.28
best solution should be that nearest to the FPIS and be farthest
Chen’s method X 5.22 0.95 Y>Z>X from FNIS. The results obtained by Chen thus are enlarged. Table 9
Y 4.81 1.35
Z 5.08 1.09
lists the results.
Finally, the comparison between previous methods [7,13,20]
TOPSIS X 0.29 0.14 Y>Z>X and the proposed fuzzy hierarchical TOPSIS method is summar-
Y 0.06 0.30
Z 0.22 0.17
ized. The comparison table is shown as Table 10.

AHP X 0.28 Y>Z>X

4. Selecting LI-BPIC supplier by fuzzy hierarchical TOPSIS
Y 0.36
Z 0.31
This section applies the fuzzy hierarchical TOPSIS to the
a
For comparison, the TOPSIS’s criteria weight use AHP’s criteria weight. lithium-ion battery protection IC (LI-BPIC) supplier selection
problem.
Table 10 Lithium-ion (Li-Ion) batteries are quickly replacing nickel-
Comparison table for different methods
hydride batteries due to advantages that include light weight, high
AHP TOPSIS Fuzzy Fuzzy hierarchical energy density, high cell voltage, and low self-discharge rate. These
[20] [13] TOPSIS [7] TOPSIS (proposed) characteristics make Li-Ion batteries perfectly suited to high
Use hierarchical structure _ _ performance portable products such as notebook computer, PDA,
Combine fuzzy theory _ _ smart phone, and so on.
To provide objective _ _ Li-Ion battery packs are constantly charged and discharged over
criteria’s weight
their life cycle. Four conditions occur during operation, as follows:
Comparison of idea _ _ _
solution
Ranking method _ _ _ _ 1. Normal condition: If the battery voltage (BV) ranges between
Be easy to understand _ _ _ _ the overdischarge detection voltage (ODDV) and the over-
charge detection voltage (OCDV), and the VM pin voltage

Fig. 3. The functional block diagram and the pin descriptions of a LI-BPIC.
 J.-W. Wang et al. / Applied Soft Computing 9 (2009) 377–386 383

Table 11a
The criteria weights given by expert I

Expert I Cost Key quality Service

characteristics

Cost 1̃ ¼ ð1; 1; 1Þ 2̃ ¼ ð1; 2; 4Þ 4̃ ¼ ð2; 4; 6Þ

Key quality / ¼ 1/4; 1/2; 1
1 2̃ 1̃ ¼ ð1; 1; 1Þ 2̃ ¼ ð1; 2; 4Þ
characteristics
Service / ¼ 1/6; 1/4; 1/2
1 4̃ / ¼ 1/4; 1/2; 1
1 2̃ 1̃ ¼ ð1; 1; 1Þ

Note: Expert I (E-I): financial manager.

Table 11b
The criteria weights given by expert II

Expert II Cost Key quality Service

Fig. 4. Supplier selection criteria.
characteristics

Cost 1̃ ¼ ð1; 1; 1Þ 1/7̃ ¼ 1/9; 1/7; 1/5 1/4̃ ¼ 1/6; 1/4; 1/2

Key quality 7̃ ¼ ð5; 7; 9Þ 1̃ ¼ ð1; 1; 1Þ 3̃ ¼ ð1; 3; 5Þ

ranges between the charger detection voltage and the characteristics
Service 4̃ ¼ ð2; 4; 6Þ / ¼ 1/5; 1/3; 1
1 3̃ 1̃ ¼ ð1; 1; 1Þ
overcurrent detection voltage (OCUDV) the current flowing
through the battery is equal or lower than a specified value, Note: Expert II (E-II): production manager.
both the charging and discharging control FETs are turned.
Charging and discharging can be freely performed in this
condition. Table 11c
The criteria weights given by expert III
2. Overcharge condition: If the BV exceeds the OCDV during
charging under normal condition and continues for the over- Expert III Cost Key quality Service
charge detection delay time or longer, the charging control FET characteristics
switches off to stop charging. Cost 1̃ ¼ ð1; 1; 1Þ 1/5̃ ¼ 1/7; 1/5; 1/3 1/3̃ ¼ ð1/5; 1/3; 1Þ

3. Overdischarge condition: If the BV falls below the ODDV during Key quality 5̃ ¼ ð3; 5; 7Þ 1̃ ¼ ð1; 1; 1Þ 2̃ ¼ ð1; 2; 4Þ
discharging under normal conditions and this situation con- characteristics
Service 3̃ ¼ ð1; 3; 5Þ / ¼ ð1/4; 1/2; 1Þ
1 2̃ 1̃ ¼ ð1; 1; 1Þ
tinues for the overdischarge detection delay time or longer, the
discharging control FET turns off and discharging stops. Note: Expert III (E-III): domain expert (consultant).
4. Overcurrent condition (load short-circuiting): If the discharging
current equals or exceeds a specified value (the VM pin voltage
is equal to or higher than OCUDV) during discharging under three sub-attributes of processing flexibility, on-time delivery,
normal conditions and this situation continues for the over- and response to changes. The attribute and sub-attribute perfor-
current detection delay time or longer, the discharging control mances of the four alternative suppliers are in Fig. 2. Conse-
FET switches off to stop discharging. quently, manufacturer supplier selection is driven by multi-
criteria decision-making.
Therefore, the OCDV, ODDV, and OCUDV of the LI-BPIC From the proposed algorithm of fuzzy hierarchical TOPSIS, it is
process are three crucial quality characteristics for product possible to compute the results of supplier selection as follows:
reliability performance, and significantly influence LI-BPICs
quality/reliability. The functional block diagram and the pin Step 1. Confirm the evaluation criteria and alternatives of
descriptions of a LI-BPIC are shown in Fig. 3. Additionally, the decision-making problem, and establish a hierarchical
management of this Just in Time (JIT) Lithium-ion battery structure as Fig. 4.
manufacturer must select from among six alternative suppliers Step 2. To obtain the degree of importance and the performance
of Lithium-ion battery protection IC and wishes to choose the of the four manufacturers for all criteria, ask experts to
best supplier in terms of not only quality but also in terms of assign appropriate fuzzy weighting and fuzzy scores for
price [10,23] and service. According to Ghoudsypour and O’Brien manufacturer performance based on the linguistic vari-
[12], service involves processing flexibility, on-time delivery, able (represented by a fuzzy number), then use pair-wise
and response to changes. The hierarchy structure of this comparison to form a fuzzy weighted matrix as listed in
supplier selection problem is shown in Fig. 4. Four alternative Tables 11a–11c.
suppliers are included in the evaluation process and their Step 3. Use the Lambda-Max method [9] to calculate the
three attributes of cost, key quality characteristics and service fuzzy weightings of each criterion given by the
capability, where key quality characteristics include three sub- experts. The detailed computational procedures are listed
attributes of OCDV, ODDV, and OCUDV, and service includes below.

(1) Let a = 1, then

2 3 2 3 2 3
1 2 4 1 1/7 1/4 1 1/5 1/3
Tm ¼ 4 1/2 1 2 5;
1 2
Tm ¼ 4 7 1 3 5; 1
Tm ¼ 45 1 2 5
1
/4 1/2 1 4 1/3 1 3 1/2 1
w1m ¼ ½0:57; 0:29; 0:14T ; w2m ¼ ½0:08; 0:66; 0:26T ; w3m ¼ ½0:11; 0:58; 0:31T
384 J.-W. Wang et al. / Applied Soft Computing 9 (2009) 377–386

(2) Let a = 0, then

2 3 2 3 2 3
1 1 2 1 1/9 1/6 1 1/7 1/5
Tl ¼ 1/4 1 1 5;
1 4 Tl2 ¼ 2 5 1 1 5;
4 Tl3 ¼ 3 1 15
4
2 /6 /4 13 2 4 /5 13 2 1 /4 13
1 1 1 1

1 4 6 1 1/3 1/2 1 1/3 1

Tu1 ¼ 4 1 1 4 5; Tu2 ¼ 4 9 1 5 5; Tu3 ¼ 4 7 1 6 5
1
/2 1 1 6 1 1 5 1 1
w1l ¼ ½0:56; 0:28; 0:16T ; w2l ¼ ½0:09; 0:57; 0:34T ; w3l ¼ ½0:13; 0:61; 0:26T
w1u ¼ ½0:55; 0:30; 0:15T ; w2u ¼ ½0:09; 0:61; 0:30T ; w3u ¼ ½0:12; 0:60; 0:28T

Ql1 ¼ 0:875; Ql2 ¼ 0:765; Ql3 ¼ 0:846

ð3Þ
Qu1 ¼ 1:036; Qu2 ¼ 1:082; Qu3 ¼ 1:107

w1l ¼ ½0:49; 0:25; 0:14T ; w2l ¼ ½0:07; 0:44; 0:26T ; w1l ¼ ½0:11; 0:52; 0:22T :
ð4Þ
w1u ¼ ½0:57; 0:31; 0:16T ; w2u ¼ ½0:10; 0:66; 0:32T ; w1l ¼ ½0:13; 0:66; 0:31T
(5) So, the fuzzy weights of the criteria are:
2 3 2 3 2 3
0:49 0:57 0:57 0:07 0:08 0:10 0:11 0:11 0:13
1 2 3
W̃ ¼ 4 0:25 0:29 0:31 5; W̃ ¼ 4 0:44 0:66 0:66 5; W̃ ¼ 4 0:52 0:58 0:66 5
0:14 0:14 0:16 0:26 0:26 0:32 0:22 0:31 0:31

Step 4. To check the consistency index. criterion. The fuzzy weight for each criterion can be
The consistency index of a comparison matrix is given calculated by Eq. (17), and the result is
by C.I. = (lmax  n)/(n  1, where lmax denotes the largest
eigenvalue of matrix Y, Y represents the reciprocal matrix W̃ ¼ ½ð0:16; 0:17; 0:19Þ; ð0:39; 0:48; 0:51Þ; ð0:20; 0:22; 0:25ÞT
of pair-wise comparisons yij, and yij is the importance of
alternative i over alternative j. The C.I. of matrix for each The fuzzy weights of the sub-criteria are then obtained
criterion is below 0.1. based on their ratio relative to the main criterion as
Step. 5 Using the geometric average method, integrate all expert assessed by expert opinion, and the results are listed in
opinions to obtain a fuzzy weight for each aggregative Table 12.

Table 12
Fuzzy weights of criteria and sub-criteria

Criteria Sub-criteria Fuzzy weight-1 Fuzzy weight-2

Cost (W1) (0.16,0.17,0.19)

Key quality characteristics (W2) (0.39,0.48,0.51)
OCDV (W21) (0.151,0.185,0.197)
ODDV (W22) (0.124,0.153,0.162)
OCUDV (W23) (0.115,0.142,0.151)

Service (W3) (0.20,0.22,0.25)

Processing flexibility (W31) (0.062,0.068,0.077)
On-time delivery (W32) (0.087,0.096,0.109)
Response to change (W33) (0.051,0.056,0.064)

Table 13
Normalized fuzzy performance matrix of four suppliers

Criteria and sub-criteria Supplier

S1 S2 S3

Cost (0.38,0.47,0.53) (0.05,0.11,0.14) (0.76,0.97,1.00)

Quality characteristics
OCDV (0.20,0.23,0.28) (0.01,0.09,0.13) (0.78,1.00,1.00)
ODDV (0.08,0.13,0.16) (0.00,0.05,0.11) (0.24,0.30,0.35)
OCUDV (0.06,0.11,0.14) (0.79,1.00,1.00) (0.71,0.80,0.88)
Service
Processing flexibility (0.77,0.92,1.00) (0.81,1.00,1.00) (0.21,0.25,0.28)
On-time delivery (0.59,0.67,0.73) (0.09,0.11,0.15) (0.38,0.44,0.51)
Response to change (0.29,0.33,0.36) (0.04,0.08,0.11) (0.72,0.78,0.85)
 J.-W. Wang et al. / Applied Soft Computing 9 (2009) 377–386 385

Table 14 Step 10. Apply the Euclidean distance method to aggregate all of
Weighted normalized fuzzy performance matrix of four suppliers
the criteria for each alternative. (The results of this step
Supplier Cost Quality characteristics Service are listed in Table 17.)
Step 11. Select the best alternative.
S1 (0.061,0.080,0.101) (0.047,0.078,0.102) (0.114,0.145,0.180)
S2 (0.008,0.019,0.027) (0.092,0.166,0.194) (0.060,0.083,0.100) The ordering of the four suppliers is S3 > S1 > S2 from
S3 (0.122,0.165,0.190) (0.229,0.345,0.387) (0.083,0.103,0.132) Table 17.
Step 12. Finally, compare with the results obtained using other
methods [7,13,19]. The results are listed in Table 18.
Table 15
FPIS and FNIS of four suppliers

Cost Quality characteristics Service Table 18 reveals that the results obtained using the proposed
method are similar to those obtained using other methods.
FPIS (0.122,0.165,0.190) (0.229,0.345,0.387) (0.114,0.145,0.180)
FNIS (0.008,0.019,0.027) (0.047,0.078,0.102) (0.060,0.083,0.100)
The proposed method and the Fuzzy TOPSIS Method [7] can be
applied to solve problems in uncertain environments. However,
based on the results, the method of Chen (see S* values) does
not discriminate clearly among suppliers. However, the pro-
Table 16 posed method can discriminate successfully and clearly among
The distance between each supplier to FPIS and FNIS suppliers.
Cost Quality character- Service
istics 5. Conclusions
hi1 h
i1 hi2 h
i2 hi3 h
i3
The fuzzy hierarchical TOPSIS method has proposed in this
S1 0.160 0.125 0.500 0 0 0.129 study, can not only resolve the inability of traditional methods to
S2 0.285 0 0.343 0.157 0.129 0
S3 0 0.285 0 0.500 0.082 0.047
measure fuzziness or uncertainty, but can also avoid the
disadvantage that of the calculated value being rendered unstable
by different methods of defuzzification and the assessment
factors being affected or the order is conversed due to only the
Table 17 FAHP method applied. The most important characteristic of this
The result of propose method new method is that it has revised and improved the idea of Chen
Supplier S* S Ordering that Fuzzy TOPSIS violates the fundamentals of TOPSIS and
enables more accurate and meaningful explanations. This study
S1 0.5247 0.1793 2
S2 0.4638 0.1570 3
proposed fuzzy hierarchical TOPSIS as a method of analyzing the
S3 0.0820 0.5770 1 LI-BPIC supplier selection problem. The analytical results demon-
strate that the proposed method is consistent with other methods.
Moreover, the proposed method is more reasonable than other
methods. Future studies could apply the proposed method to
Step 6. Establish a normalized fuzzy performance matrix (see other areas of decision-making or the computation of weights of
Table 13). other objects.
Step 7. Obtain a weighted normalized fuzzy performance matrix
References
(see Table 14).
Step 8. Rank Table 14 by the metric distance method to [1] Y.P. Cha, M. Jung, Satisfaction assessment of production schedules using extended
determine the maximum and minimum, then calculate TOPSIS, in: Proceedings of 2001 KORS/KIIE Joint Spring Conference, Korea, April
FPIS and FNIS as follows: (see Table 15) 27–28, (2001), pp. 556–559.
[2] Y.H. Chang, C.H. Yeh, Evaluating airline competitiveness using multi-attribute
Step 9. Calculate the distances between each point and FPIS and
decision making, Omega 29 (2001).
FNIS by metric distance method. (The results of this step [3] C.H. Cheng, D.L. Mon, Evaluating weapon system using fuzzy AHP, Proc. First Natl.
are listed in Table 16.) Conf. Sci. Technol. Natl. Defense (1992) 531–538.
[4] C.H. Cheng, Evaluating naval tactical missile system by fuzzy AHP based on the
grade value of membership function, Eur. J. Operational Res. 96 (1997) 343–
Table 18 350.
Comparison with other method [5] C.H. Cheng, A new approach for ranking fuzzy numbers by distance method, Fuzzy
Sets Syst. 95 (1998) 307–317.
Aggregative value Ordering [6] C.H. Cheng, L.L. Yang, C.L. Hwang, Evaluating attack helicopter by AHP based on
linguistic variable weight, Eur. J. Operational Res. 116 (1999) 423–435.
S* S
[7] C.T. Chen, Extensions of the TOPSIS for group decision-making under fuzzy
Proposed method S1 0.5247 0.1793 S3 > S1 > S2 environment, Fuzzy Sets Syst. 114 (2000) 1–9.
S2 0.4638 0.1570 [8] L.S. Chen, C.H. Cheng, Selecting IS personnel using ranking fuzzy number by
metric distance method, Eur. J. Operational Res. 160 (3) (2005) 803–820.
S3 0.0820 0.5770
[9] R. Csutora, J.J. Buckley, Fuzzy hierarchic analysis: the Lambda-Max method, Fuzzy
Chen’s method S1 2.6983 0.3099 S3 > S1 > S2 Sets Syst. 120 (2001) 181–195.
S2 2.7513 0.2592 [10] G.W. Dickson, An analysis of vendor selection systems and decisions, J. Purchasing
S3 2.4193 0.5959 2 (1) (1996) 5–17.
[11] D. Dubios, H. Prade, Operations on fuzzy numbers, Int. J. Syst. Sci. 9 (1978) 613–626.
TOPSIS S1 0.3745 0.1429 S3 > S1 > S2 [12] S.H. Ghoudsypour, C.O. O’Brien, A decision support system for supplier selection
S2 0.3823 0.0615 using an integrated analytic hierarchy process and linear programming, Int. J.
S3 0.0497 0.4220 Prod. Econ. 56–57 (1–3) (1998) 99–212.
[13] C.L. Hwang, K. Yoon, Multiple Attribute Decision Making Methods and
AHP S1 0.35303 S3 > S1 > S2 Application, Springer-Verlag, New York, 1981.
S2 0.274257 [14] D.H. Jee, K.J. Kang, A method for optimal material selection aided with decision
S3 0.692228 making theory, Mater. Design 21 (2000) 199–206.
[15] M. Kamal, Al-Subhi Al-Harbi, Application of the AHP in project management,
Note: For comparison, the TOPSIS’s criteria weight use AHP’s criteria weight. Project Manage. 19 (2001) 19–27.
386 J.-W. Wang et al. / Applied Soft Computing 9 (2009) 377–386

[16] E. Løken, Use of multicriteria decision analysis methods for energy planning [20] T.L. Saaty, The Analytic Hierarchy Process, McGraw-Hill, New York, 1980 .
problems, Renewable Sustainable Energy Rev. 11 (2007) 1584–1595. [23] C.A. Weber, J.R. Current, W.C. Benton, Vender selection criteria and methods, Eur.
[17] G.A. Miller, The magical number seven or minus two: some limits on our capacity J. Operational Res. 50 (1991) 2–18.
of processing information, Psychology 63 (1956) 81–97. [25] L.A. Zadeh, The concepts of a linguistic variable and its applications to approx-
[18] T.J. Ross, Fuzzy Logic with Engineering Applications (International ed.), McGraw- imate reasoning (I), Inf. Sci. 8 (1975) 199–249.
Hill, New York, 2000. [26] L.A. Zadeh, The concepts of a linguistic variable and its applications to approx-
[19] X. Ruoning, Z. Xiaoyan, Extensions of the Analytic Hierarchy Process in Fuzzy imate reasoning (II), Inf. Sci. 8 (1975) 301–357.
Environment, Fuzzy Sets Syst. 52 (1992) 251–257.