European Journal of Operational Research 244 (2015) 277288

Contents lists available at ScienceDirect

European Journal of Operational Research

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

Decision Support

Generalized analytic network process

Bin Zhu a, Zeshui Xu b,, Ren Zhang c, Mei Hong c
a

School of Economics and Management, Tsinghua University, Beijing 100084, China

Business School, Sichuan University, Chengdu, Sichuan 610064, China
c
College of Meteorology and oceanography, PLA University of Science and Technology, Nanjing, Jiangsu 211101, China
b

a r t i c l e

i n f o

Article history:
Received 24 March 2014
Accepted 8 January 2015
Available online 19 January 2015
Keywords:
Decision support systems
Decision analysis
Distribution
Simulation

a b s t r a c t
The analytic network process (ANP) is a methodology for multi-criteria decision making used to derive
priorities of the compared elements in a network hierarchy, where the dependences and feedback within
and between the elements can be considered. However, the ANP is limited to the input preferences as crisp
judgments, which is often unfavorable in practical applications. As an extension of the ANP, a generalized
analytic network process (G-ANP) is developed to allow multiple forms of preferences, such as crisp (fuzzy)
judgments, interval (interval fuzzy) judgments, hesitant (hesitant fuzzy) judgments and stochastic (stochastic
fuzzy) judgments. In the G-ANP, a concept of complex comparison matrices (CCMs) is developed to collect
decision makers preferences in the multiple forms. From a stochastic point of view, we develop an eigenvector
method based stochastic preference method (EVM-SPM) to derive priorities from CCMs. The main steps of
the G-ANP are summarized, and the implementation of the G-ANP in Matlab and Excel environments are
given in detail, which is also a prototype for a decision support system. A real-life example of the piracy risk
assessment to the energy channels of China is proposed to demonstrate the G-ANP.
2015 Elsevier B.V. All rights reserved.

1. Introduction
In multi-criteria decision making, the analytic hierarchy process
(AHP) (Saaty, 1980) is a popular methodology to derive priorities of
compared elements (or objectives, alternatives etc.) to assist decision
makers (DMs) to make decisions. It has been successfully applied
in practice (Saaty, 1989, 2008). In the AHP, the DMs provide their
preferences over paired comparisons of elements by a 19 scale in
each level of a hierarchy. The hierarchy of the AHP is a linear top down
form with clear independent levels, where the elements in each level
are also independent.
In the AHP, prioritization methods are necessary to derive priorities of the compared elements in each level of the hierarchy, such
as the eigenvector method (EVM) (Saaty, 1977), the logarithmic least
squares method (LLSM) (Crawford & Williams, 1985), and the logarithmic goal programming method (GPM) (Bryson, 1995). Due to the
inherent difference of the DMs, their preferences can be inconsistent. As discussed by Saaty and Vargas (1987), only the EVM takes
consistency of the preferences into account. The use of other prioritization methods may lead to wrong decisions by reversing ranks.
Therefore, consistency checking and improving are signicant in the
AHP to guarantee meaningful results.

Corresponding author. Tel.: +86 18262603063.

E-mail addresses: binzhu@263.net (B. Zhu), xuzeshui@263.net (Z. Xu).

http://dx.doi.org/10.1016/j.ejor.2015.01.011
0377-2217/ 2015 Elsevier B.V. All rights reserved.

The AHP has a limitation that it cannot deal with interactions and
dependencies between the elements in the levels of the hierarchy.
For example, to predict the market share of cell-phone providers, the
elements that inuence the market share of a company can be costs
and services, where the services may also inuence the costs. To overcome this limitation, Saaty (2001) considered the dependences and
feedback of the elements, and then developed the analytic network
process (ANP).
In the ANP, the network allows clusters of elements inuence each
other, or has loops if the elements in the clusters have inner dependences. So the network spreads out in all directions and its cluster
of elements are not arranged in a particular order (Saaty, 2004). The
advantage of the ANP in dealing with dependences and feedback enables it to be very useful in many practical applications. For example,
the ANP has been used for the interdependent information system
project selection (Lee & Kim, 2000), the R&D project selection (Meade
& Presley, 2002), the logistics service provider selection (Jharkharia &
Shankar, 2007), the product mix planning (Chung, Lee, & Pearn, 2005),
the SWOT analysis (Yksel & Dagdeviren, 2007), the nancial-crisis
forecasting (Niemira & Saaty, 2004), and the multi-criteria analysis
(Wolfslehner, Vacik, & Lexer, 2005).
However, the preferences in the ANP are limited to crisp judgments based on the 19 scale. In many cases, the DMs may prefer to many other possible forms to represent their preferences. To
overcome this limitation of the ANP, Mikhailov and Singh (2003) developed the fuzzy analytic network process (F-ANP) which allows

278

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

multiple forms of preferences, such as crisp judgments, interval judgments, and fuzzy judgments. Moreover, the main difference between
the F-ANP and the ANP is that the F-ANP uses a fuzzy preference programming (FPP) method as an alternative prioritization method to
the EVM.
The FPP is a linear programming method that maximizes the DMs
satisfaction degree to consistency of the preferences, where a consistency index is also provided. However, this consistency index is a
relative index varying with a so called deviation parameter, which
cannot be used to measure whether the preferences are consistent or
not. Therefore, similar to the LLSM and the GPM, the FPP also does not
take consistency of the preferences into account. Although the FPP is
convenient to be used to derive priorities, it may result in ineffective
decisions in some cases if consistency of the preferences cannot be
guaranteed.
As complexity of the socio-economic environment is increased,
more uncertainties are experienced by the DMs apart from crisp judgments, interval judgments and fuzzy judgments. For example, the
DMs may be hesitant about several possible values for a judgment.
This concept of hesitance of being used for describing the preferences
in decision making is based on hesitant fuzzy sets originally introduced by Torra (2010). Based on a 0.10.9 scale, the judgments that
each is characterized by several possible values are called hesitant
fuzzy judgments. They have been widely used in decision making
problems (Xia & Xu, 2011; Xia, Xu, & Chen, 2013; Xu & Xia, 2011).
Based on the 19 scale, hesitant judgments were also proposed for
decision making (Xia & Xu, 2013; Zhu & Xu, 2014a).
The judgments can be also indicated by stochastic variables, which
have been considered by some researchers as stochastic judgments.
For example, Rosenbloom (1997) claimed that the subjective preferences with a continuous probability distribution is a standard requirement in the conventional decision analysis, so stochastic judgments
should be used rather than crisp judgments and interval judgments.
Moskowitz, Tang, and Lam (2000) recommended to use stochastic
variables to account for inconsistency and imprecision in the preferences. Hahn (2003) claimed that the deterministic methods in decision making are special cases of their stochastic counterparts, and
then proposed a stochastic formulation of the AHP with stochastic
judgments.
Based on the discussion above, so many possible forms of preferences are available for the DMs. In practice, different forms of preferences shall be useful in different situations. However, different forms
of preferences also require different prioritization methods. Is there
a general method can deal with all the possible forms of preferences
mentioned above and produce meaningful results? In this paper, we
develop a generalized analytic network process (G-ANP) method as
an extension of the ANP. In the G-ANP, the DMs can provide their preferences over paired comparisons of elements in the network as crisp
(fuzzy) judgments, interval (interval fuzzy) judgments, hesitant (hesitant fuzzy) judgments and stochastic (stochastic fuzzy) judgments,
etc.
Since the preferences in the G-ANP can be in multiple forms, we
dene them as complex judgments to construct complex comparison
matrices (CCMs). An expected index is developed to measure whether
or not a CCM is of the acceptable consistency. Then an automatic consistency improving method is proposed to repair inconsistent CCMs to
guarantee meaningful results. Consistency checking and improving of
the G-ANP also eliminate the drawbacks of the F-ANP that cannot take
consistency into account. To derive priorities from CCMs, we develop
an eigenvector method based stochastic preference method (EVMSPM) as a new prioritization method. Based on these new developed
methods, a step by step procedure of the G-ANP is proposed.
The rest of the paper is organized as follows. Section 2 introduces
all the possible forms of preferences in the G-ANP, and then denes complex judgments and CCMs. Section 3 focuses on consistency
checking and improving of CCMs. Section 4 develops the EVM-SPM. In

Table 1
The 19 scale.
Scale

Meaning

1
3
5
7
9
Other values between 1 and 9

Equally preferred
Moderately preferred
Strongly preferred
Very strongly preferred
Extremely preferred
Intermediate values used to represent
compromise

Section 5, the step by step implementation of the G-ANP is discussed

in detail. Then a real-life example of the piracy risk assessment to the
energy channels of China is given to illustrate the G-ANP in Section 6.
Section 7 gives some concluding remarks.
2. Complex judgments and complex comparison matrices
In this section, we introduce the judgments based on a 19 scale
and a 0.10.9 scale respectively, then dene complex judgments and
complex comparison matrices (CCMs).
2.1. Judgments based on the 19 scale
Based on the 19 scale (Saaty, 1980), shown in Table 1, the DMs
preferences can be clearly represented by crisp values to describe the
relationships between paired comparisons of elements, where the
preferences are called crisp judgments.
A reciprocal comparison matrix consisting of crisp judgments, A =
(aij )nn , can be shown as follows:

A=

.
.
.

a12
1
..
.
1/aij
...

a13
a23
1
..
.

...
...

..

a1n
a2n

..
.

..

(1)

where the judgments satisfy the reciprocal property as aij = 1/aji ,

i, j = 1, 2, . . . , n.
If a judgment is represented by an interval to indicate uncertainties, then it is called an interval judgment. An interval reciprocal
comparison matrix (Saaty & Vargas, 1987) constructed by interval
judgments, A = (a ij )nn , is given as follows:

A =

.
.
.

[al12 , au12 ]
1
..
.
[1/auij , 1/alij ]
...

[al13 , au13 ]
[al23 , au23 ]
1
..
.

...
...

..

[al1n , au1n ]

[al2n , au2n ]

..

..

(2)

where the judgments satisfy the reciprocal property as alij = 1/auji ,

auij = 1/alji and alij auij , i, j = 1, 2, . . . , n.
In accordance with Saaty and Vargas (1987) of considering discrete judgments within the interval, we also use the integers and
reciprocals of the integers between the upper and lower bounds of an
interval. For example, the interval judgment [1/4,2] has 1/4, 1/3 and
1/2 as reciprocals of integers, and 1 and 2 as integers, for a total of
ve judgments. So the reciprocal comparison matrix is a special case
of the interval reciprocal comparison matrix.
If a judgment includes several possible values, then it is called a
hesitant judgment, which can be represented by a set h = {a(l)|l =
1, . . . , |h|}. For example, a hesitant judgment {1/3, 1} has two possible values as 1/3 and 1. A hesitant reciprocal comparison matrix consisting of hesitant judgments, denoted by H = (hij )nn with

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

hij = {a(ijl)|l = 1, . . . , |hij |}, can be shown as follows:

H=

.
.
.

(|h

(|h

|)

{a(121), . . . , a12 12 }

|)

(|h

(|h23 |)

{a23 , . . . , a23

1
.
.
.
(|hij |)

|)

.
.
.

} . . . {a2n , . . . , a2n
.
.
.
..

...

.
.
.
1

(l)

where the judgments satisfy the reciprocal property as aij

(l)

(3)

p (percent)

(|h2n |)

(1)

{1/a(ij1), . . . , 1/aij

Table 3
The distribution table of F (k ) = pk .

{a(131), . . . , a13 13 } . . . {a(1n1), . . . , a1n 1n }

(1)

279

(l)

= 1/aji

1
65

2
28

3
3

4
2

5
2

formula as

aij = bij /bji ,

i, j = 1, 2, . . . n,

(4)

to transform B into a reciprocal comparison matrix A = (aij )nn .

is the th element in hij ), i, j = 1, 2, . . . , n.

The hesitant reciprocal comparison matrix can be also called a hesitant multiplicative preference relation (Zhu & Xu, 2014a). Since we
consider discrete values in interval judgments, the interval reciprocal
comparison matrix can be considered as a special case of the hesitant
reciprocal comparison matrix.
If a judgment is a stochastic variable following a probability function, then it is called a stochastic judgment. For example, let ij
be a stochastic variable satisfying ij hij and the reciprocal property as ij = 1/ji , and F (ij ) be the probability function of ij , then
Z = (ij )nn is called a stochastic reciprocal comparison matrix for all
i, j = 1, 2, . . . , n. As usual, the possible values in hesitant judgments
are assumed to be uniform distributed. So the hesitant reciprocal comparison matrix can be considered as a special case of the stochastic
reciprocal comparison matrix. According to practice, more possible
distributions can be used to describe the judgments.
Based on the discussion above, we conclude that the stochastic reciprocal comparison matrix encompasses the reciprocal comparison
matrix, the interval reciprocal comparison matrix, and the hesitant
reciprocal comparison matrix as special cases.

To encompass crisp (fuzzy) judgments, interval (interval fuzzy)

judgments, hesitant (hesitant fuzzy) judgments and stochastic
(stochastic fuzzy) judgments as special cases of complex judgments,
we dene a complex judgment as cp , where c represents the value(s)
of the judgment with a probability interpretation. The probability interpretation is to use a probability function to specify the distribution
of the values in c.
For example, assume c = [1, 5]p , then the probability function, denoted by F (k ) = pk (k c), can be shown as a distribution table in
Table 3.
Since crisp (fuzzy) judgments, interval (interval fuzzy) judgments
and hesitant (hesitant fuzzy) judgments do not require the probability
interpretations or the values follow uniform distribution, it is not
necessary to explicitly give the probability functions. So cp is written
as c in such cases.
Constructed by complex judgments, a CCM is dened as follows.

2.2. Fuzzy judgments based on the 0.10.9 scale

where cij is a complex judgment indicating the preference degree(s)

of the alternative xi over xj .

(aij

2.3. Complex judgments and complex comparison matrices

Denition 1. Let X = {x1 , x2 , . . . , xn } be a xed set, then a comp

plex comparison matrix (CMM) is represented by C = (cij )nn X X,
p

The 0.10.9 scale is uniformly and symmetrically distributed

around 0.5. Using the judgments based on the 0.10.9 scale to indicate the DMs preferences is another natural way to represent uncertainties, which incorporates the vagueness of human thinking. The
judgments represented upon the 0.10.9 scale are called fuzzy judgments. The judgments between the 19 scale and the 0.10.9 scale
have close relationships, which are shown in Table 2.
Similar to interval judgments, hesitant judgments and stochastic
judgments, we also have interval fuzzy judgments (Zadeh, 1975), hesitant fuzzy judgments (Torra, 2010) and stochastic fuzzy judgments
(not formally dened in previous literature). Particularly, fuzzy judgments are used to construct complementary comparison matrices or
called fuzzy preference relations (Tanino, 1984). With respect to interval fuzzy judgments and hesitant fuzzy judgments, we have interval
fuzzy preference relation (Xu, 2004b) and hesitant fuzzy preference
relation (Zhu & Xu, 2013) respectively.
With respect to Table 2, many transformation formulas have been
developed for the transformation of the judgments between the
two scales. For example, for a fuzzy preference relation B = (bij )nn
consisting of fuzzy judgments, Xu (2002) gave a transformation

It is clear that the reciprocal comparison matrix, the interval reciprocal comparison matrix, the hesitant reciprocal comparison matrix
and the stochastic reciprocal comparison matrix are special cases of
the CCM.
Generally, we assume that the complex judgments in a CCM
satisfy the reciprocal property associated with the forms of preferences. For example, if a complex judgment in C is an interval judgment, denoted by cij = [2, 4], then the corresponding diagonal element should be cji = [1/4, 1/2]; if a hesitant judgment cij = {3, 5},
then cji = {1/3, 1/5}.
Remark 1. In this paper, we transform fuzzy judgments into the judgments based on the 19 scale for the purpose of developing a new
prioritization method.
Remark 2. Sine we mainly deal with reciprocal comparison matrices, reciprocal comparison matrices are called comparison matrices
in brief in the rest of the paper. It is also the interval reciprocal comparison matrix, the hesitant reciprocal comparison matrix and the
stochastic reciprocal comparison matrix.
3. Consistency checking and improving

Table 2
The relationships of judgments between the two scales.
19 scale

0.10.9 scale

Meaning

1
3
5
7
9
Other values
between 1 and 9

0.5
0.6
0.7
0.8
0.9
Other values
between 0 and 1

Equally preferred
Moderately preferred
Strongly preferred
Very strongly preferred
Extremely preferred
Intermediate values used to represent
compromise

Consistency checking and improving are of vital importance for

comparison matrices to guarantee meaningful results. Saaty (1977)
developed an eigenvector method (EVM) to provide a consistency
index to measure the consistency degrees of comparison matrices.
Then the consistency thresholds are established to dene whether
the comparison matrices are with the acceptable consistency.
With respect to inconsistent comparison matrices, some consistency improving methods have been developed (Cao, Leung, & Law,
2008; Ergu, Kou, Peng, & Shi, 2011; Li & Ma, 2007; Xu & Wei, 1999).

280

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

Table 4
The reference values of RI for different values of n.

3.2. A stochastic consistency improving method

10

RI

0.52

0.89

1.12

1.26

1.36

1.41

1.46

1.49

For interval comparison matrices, hesitant comparison matrices and

stochastic comparison matrices, there are very few research results
focus on this topic. In this paper, to deal with CCMs, we develop
a new expected index for consistency checking and a stochastic
method for consistency improving, respectively. This index and the
stochastic method can also be used for interval comparison matrices,
hesitant comparison matrices and stochastic comparison matrices
respectively.
3.1. An expected index for consistency checking
The eigenvector method (EVM) (Saaty, 1977) gives a consistency
index, CR = CI/RI, to measure consistency of a comparison matrix
A = (aij )nn , where CI = (max n)/(n 1), max is the maximum
eigenvalue of A, and RI is a random index. The reference values of
RI for different values of n are shown in Table 4.
If CR < 0.1, then A is said to be with the acceptable consistency;
otherwise unacceptable.
For a comparison matrix A = (aij )nn , if

i
aij =
, i, j = 1, 2, . . . , n,
j

(5)

where = (1 , 2 , . . . , n ) is the priority vector of A, satisfying

n
i=1 i = 1, i > 0, i = 1, 2, . . . , n, then A is consistent (Saaty, 1980),
where its consistency index CRA = 0.
p
p
For a CCM C = (cij )nn , assume that the stochastic variables ij cij ,
and ij are specied by the probability function F (ij ), i, j = 1, 2, . . . , n.
Then following F (ij ), we can get a stochastic comparison matrix
Z = (ij )nn . According to the EVM, we obtain the corresponding consistency index denoted by CRZ ( ). Then in the judgment space of C,
denoted by C ( ), an expected index can be dened to measure consistency of C:

E(CRC ) =

C ( )

CRZ ( )d

(6)

Being in accordance with the consistency thresholds of CR, if

E(CRC ) < 0.1, then C is said to be with the acceptable consistency.
Otherwise, unacceptable.
The expected index can be obtained by Monte Carlo simulation
(Milton & Arnold, 1995). Since each iteration contributes to the result,
to achieve the accuracy d with 95 percent condence level for the
expected index, the required number of iterations is

K=

1.962
4d2

(7)

If the error limit is d = 0.01, then the number of iterations should be

9604.
Remark 3. We let d = 0.01, and the number of iterations be 10,000
in the rest of the paper to calculate the expected index.
Example 1. Assume a CCM with ve elements xi (i = 1, 2, . . . , 5) to
compare:

1
[2, 5]
{1, 3}
[2, 6]
[3, 5]

{2, 3, 5} [1/5, 1/3] [3, 7]

1

[1/2, 1/5]

C1 =
1
[5, 7]
4

{1, 1/3} {1/2, 1/3, 1/5}

[5, 3]
[1/5, 1/7]
1
[1/3, 5]
[1/2, 1/6]
[1/3, 1/5]
[1/3, 1/7]
1/4
[3, 1/5]
1
By Monte Carlo simulation operated in the Matlab environment,
we have E(CRC1 ) = 0.4778. Since E(CRC1 ) > 0.1, C1 is not with the
acceptable consistency.

In each iteration of Monte Carlo simulation of calculating the expected index, there exists a stochastically obtained comparison matrix A = (aij )nn and a corresponding consistency index, denoted by
CRA . If CRA > 0.1, we improve consistency of A until it is of the acceptable consistency. For the consistency improving method of A, we
refer to the iteration method proposed by Xu and Wei (1999). The
step by step algorithm is given in Algorithm 1
Algorithm 1.
Step 1: Let A(k) = (a(ijk))nn , and k = 0.

Step 2: Calculate the maximal eigenvalue max (A(k)) of A(k) and

the normalized principal right eigenvector (k) = (w(1k),
Step 3:
Step 4:

Step 5:
Step 6:
Step 7:

w(2k), . . . , w(nk))T .
Calculate CRA(k) .
If CRA(k) < 0.1, then go to step 6; otherwise, continue the
next step.
Let A(k+1) = (a(ijk+1)), where a(ijk+1) = (a(ijk))1 (i(k)/j(k)) ,
[0, 1]. Let k = k + 1 and return to step 2.
Output A(k), CRA(k) .
End.

In step 5, is the parameter indicating the accuracy in each iteration. The bigger the value , the more consistent the improved A,
but the less the original preferences are preserved. Generally, we set
= 0.01 to preserve the original preferences as many as possible.
In order to prove the convergence of Algorithm 1, Xu and Wei
(1999) modied CRA(k) = 0 in step 4 of Algorithm 1, which is called
Algorithm 2. Then they gave a theorem to prove the convergence of
Algorithm 2 as follows.
Theorem 1. (Xu & Wei, 1999) Let A = (aij )nn be an inconsistent comparison matrix. Let {A(k)} be the matrix sequence generated in Algok)
rithm 2 and (max
be the maximal eigenvalue of A(k). Then for each k,
(
k+1
)
max (A
) < max (A(k)) and limk max (A(k)) = n.
Based on Algorithm 1, for a CCM C = (cij )nn , we now develop a
stochastic consistency improving method shown in Algorithm 3 to
improve consistency of C until E(CRC ) < 0.1.
p

Algorithm 3.
s)
Step 1: Let C (s) = (cij )(nn
, and s = 0.
Step 2: Calculate E(CRC (s) ), if E(CRC (s) ) < 0.1, go to step 6; otherwise
go to the next step.
Step 3: Stochastically get a comparison matrix A = (aij )nn from
C (s), then let A(k) = A, and k = 0.
Step 4: Obtain the comparison matrix A(k) with the acceptable consistency by Algorithm 1.
Step 5: Return A(k) to C (s) to replace A(0). Let s = s + 1 and return to
step 2.
Step 6: Output C (s) and E(CRC (s) ).
Step 7: End.
p

In order to prove the convergence of Algorithm 3, we let E(CRC (s) ) =

0 in step 2 of Algorithm 3. Then the modied Algorithm 3 is called
Algorithm 4.
Theorem 2. Let C be a n n positive comparison matrix, and {E(CRC (s) )}
be the sequence of the expected index of C generated in Algorithm 4. Then,
for each s, we have E(CRC (s+1) ) < E(CRC (s) ) and limk {E(CRC (s) )} = 0.
The proof of Theorem 2 is shown in the Appendix.
Example 2. Continued with Example 1, by Algorithm 3, we output an
improved CCM of C1 , denoted by C1 , as

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

{0.5347, 0.5542, 0.4528, 0.2529}

{0.8991, 0.5198}
C1 =

{0.5489, 0.3760, 0.4059, 0.2176, 0.1667}

{0.3023, 0.2378, 0.1979}

{1.8702, 1.8045, 2.2083, 3.9548}

1
{0.5391, 0.9579, 0.6530}
{1.3755, 1.3047, 2.2896}
{0.3326, 0.2825, 0.3703, 0.1858, 0.1553}

{1.8217, 2.6594, 2.4634, 4.5965, 6}

{0.7270, 0.7664, 0.4368}
{3.2863, 2.4628, 2.4097}

{0.3043, 0.4060, 0.4150}

0.2875

{3.0067, 3.5393, 2.7006, 5.3810, 6.4386}

3.4782
,

{0.9571, 1.2283, 1, 2, 3.0158, 3.7978, 4.3584}

1
density function (Zhu & Xu, 2014b):

f ( ) =
4. An eigenvector method based stochastic preference method
The prioritization methods that aim to derive priorities from
comparison matrices and preference relations have been studied
extensively over last decades. Some related methods are listed in
Table 5.
Based on Table 5, for the comparison matrices and the preference
relations, the prioritization methods can be roughly divided into four
commons ones: (1) the classic EVM (Saaty, 1977), (2) the programming methods (Crawford, 1987; Gong, 2008; Mikhailov, 2000, 2004;
Wang, 2006 ; Wang & Elhag, 2007; Wang, Fan, & Hua, 2007; Wang,
Parkan, & Luo, 2008; Wang, Yang, & Xu, 2005a, 2005b; Xu, 2004a; Xu
& Chen, 2008; Xu & Da, 2005), (3) the aggregation methods (Barzilai,
1997; Escobar, Aguarn & Moreno-Jimnez, 2004; Xu, 2006; Zhu, Xu &
Xu, 2013), and (4) the stochastic methods (Hahn, 2003; Rosenbloom,
1997). Among these methods, the EVM, the programming methods,
and the aggregation methods are deterministic methods which are
the special cases of the stochastic methods.
Since all the comparison matrices and preference relations listed in
Table 5 can be considered as special cases of CCMs, we should develop
a new prioritization method for CCMs from a stochastic point of view.
In this section, we develop an eigenvector method based stochastic
preference method (EVM-SPM) to derive priorities from CCMs.
The EVM-SPM is motivated by a stochastic preference analysis
method developed by Zhu and Xu (2014b). By analyzing the judgment space of numerical preference relations, Zhu and Xu (2014b)
gave several outcomes to provides constructive suggestions for the
DMs. Following the stochastic preference analysis method, we conp
centrate on a CCM C = (cij )nn associated with a set of elements
X = {x1 , x2 , . . . , xn }.
p
Let Z = (ij )nn be a stochastic comparison matrix, where ij cij ,
and let f (ij ) be the density function that species ij . Then the joint
probability distribution of ij (i, j = 1, 2, . . . , n) can be specied by a

{1.1122, 1.9240}
{1.8550, 1.0439, 1.5314}

{3.3081, 4.2047, 5.0529}

{1.0488, 0.8142, 1, 0.5, 0.3316, 0.2633, 0.2294}

and output the expected index E(CRC  1 ) = 0.097.

281

fij (ij )

(8)

i<j

By the EVM, we can get the priority vector of the elements from Z,
denoted by = (1 , 2 , . . . , n ). For the element xi , a ranking function is dened to rank xi as follows:

ranki () = 1 +

(k > i ) = r

(9)

k=1

where (ture) = 1 and (false) = 0.

So the set of favorable judgments in Z, which makes xi rank r, can
be dened as

Zri

( ) = Z : ranki () = 1 +

(k > i ) = r

(10)

k=1

The acceptable possibility of judgments which makes xi rank r

is dened as a percentage of the favorable judgments to the overall
judgments as follows (Zhu & Xu, 2014b):

bri =

Zir ( )

f ( )d

(11)


where bri [0, 1], nr=1 bri = 1. The bigger bri , the higher acceptable
possibility of xi ranking r.
To obtain the overall acceptable possibility of xi , it requires an additional process to aggregate the acceptable possibilities for each possible rank of xi . For this aggregation process to discern the elements,
some possible so called meta-weights could be the linear weights
r = (n r)/(n 1), theinverse weights r = 1/r, and the centroid
weights r = ni=r 1/ i/ ni=1 1/ i, where the linear weights give more
weight to the mediocre ranks, while the inverse and centroid weights
emphasize the best ranks (Lahdelma & Salminen, 2001). The overall
acceptable possibility of xi is dened as follows:

oi =

r bri

(12)

Table 5
Prioritization methods for comparison matrices and preference relations.
Comparison matrix
Prioritization methods

Prioritization methods

Interval comparison matrix

Hesitant comparison matrix

EVM (Saaty, 1977); The geometric means (Barzilai, 1997; Linear and nonlinear programming methods (Wang
Escobar et al., 2004); Logarithmic least-squares
et al., 2005a); Two-stage logarithmic goal
method (Crawford, 1987); Fuzzy programming method
programming method (Wang et al., 2005b);
(Mikhailov, 2000); Linear programming method
Lexicographic goal programing method (Wang, 2006);
(Wang et al., 2008); Probability method (Rosenbloom,
Linear programming methods, (Wang & Elhag, 2007);
1997); Stochastic modeling method (Hahn, 2003).
Fuzzy programming method (Mikhailov, 2004).

None

Fuzzy preference relation

Interval fuzzy preference relation

Hesitant fuzzy preference

relation

Goal programming method (Xu, 2004a); Least deviation

method (Xu & Da, 2005); Chi-square method (Wang
et al., 2007); Least square method (Gong, 2008).

Goal programming method (Xu & Chen, 2008);

Aggregation method (Xu, 2006).

Hesitant goal programming

method (Zhu et al., 2013);
Aggregation method (Zhu
et al., 2013).

282

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

Table 6
The acceptable possibilities of the elements for each rank with respect to C1 .

5.1. Generalized analytic network process

Elements

b1

b2

b3

b4

b5

x1
x2
x3
x4
x5

0.9988
0
0.0012
0
0

0.0012
0.1167
0.8819
0.0002
0

0
0.7851
0.1166
0.0983
0

0
0.0982
0.0003
0.9015
0

0
0
0
0
1

Then we normalize the overall acceptable possibility of xi as

 r r
r bi
=  
r r
i oi
i
r bi

oi
pi = 

(13)

which can be referred as the nal priority of xi .

p
As discussed above, to deal with C = (cij )nn by the EVM-SPM, we
give the following algorithm based on Monte Carlo simulation.
Algorithm 5.
Step 1: Stochastically get a comparison matrix A = (aij )nn from
C, then calculate the acceptable possibilities bri (i, r =
1, 2, . . . , n) of the elements xi (i = 1, 2, . . . , n) for each possible rank by Eq. (11).
Step 2: Determine the meta-weights, then according to Eq. (12) and
bri (i, r = 1, 2, . . . , n), calculate the overall acceptable possibilities of xi (i = 1, 2, . . . , n).
Step 3: Calculate the nal priorities of xi (i = 1, 2, . . . , n) by Eq. (13).
We now give an example to illustrate the EVM-SPM.
Example 3. Continued with Example 2, according to Eq. (11), we get
the acceptable possibilities of the elements for each rank shown in
Table 6.
A three dimensional column chart is helpful to show all the acceptable possibilities of the elements for each rank in a visual way.
The results in Table 6 can be described in Fig. 1.
Furthermore, we choose the centroid weights which can avoid
zero weights to aggregate the acceptable possibilities. Then by
Eq. (13), the nal priorities for the ve elements can be obtained,
denoted by p1 = 0.4564, p2 = 0.1618, p3 = 0.2452, p4 = 0.0966 and
p5 = 0.04 respectively. So x1 is with the highest overall acceptable
possibility.
5. Generalized analytic network process and its implementation
In this section, we rst summarize the main steps of the generalized analytic network process (G-ANP), then give the detail of the
implementation of the G-ANP, which is a prototype for a decision
support system.

The generalized analytic network process (G-ANP) that includes

the developed new methods in this paper mainly has the following
steps.
Step 1: Construct a control hierarchy with control criteria which
serve as a basis for making judgments, and then construct
a network hierarchy with clusters of elements.
Step 2: Identify dependences and feedback within and between the
clusters and the clusters of elements.
Step 3: Construct CCMs in the network hierarchy according to the
dependences and feedback identied in step 2.
Step 4: Check consistency of the CCMs. If they are not with the
acceptable consistency, then we utilize Algorithm 3 to improve their consistency until acceptable.
Step 5: Derive priorities from the CCMs by the EVM-SPM.
Step 6: Construct an unweighted supermatrix.
Step 7: Adjust the unweighted supermatrix to a weighted supermatrix (or called a stochastic matrix).
Step 8: Calculate limiting priorities of all the elements from the
stochastic matrix.
Step 9: Calculate nal priorities of the elements according to the
limiting priorities.
The related methods in steps 3, 4 and 5 have been discussed in
Sections 2, 3 and 4 respectively. The other steps are similar to the
traditional analytic network process (ANP), which can be found in
(Saaty, 2001) in detail. When the uncertainties of judgments in GANP tend to zero, the G-ANP reduces to the ANP.
5.2. A step by step implementation
To implement the G-ANP step by step, Excel and Matlab are
adopted as the main development environments, because of their
good visualization capabilities, powerful computation capabilities,
and programming interfaces.
Step 1: (Implemented in Excel): Use a network table of clusters and
a network table of elements to list all the dependences and
feedback within and between the clusters and the clusters
of elements, respectively.
Step 2: (Implemented in Excel): According to the network table of
clusters and the network table of elements, we compare
the clusters and the clusters of elements to construct CCMs,
respectively.
Step 3: (Implemented in Excel and Matlab): Send all the CCMs to
Matlab by Excel Link, then check and improve (if necessary)
their consistency, and then return the results to Excel. The
owchart of step 4 is shown in Fig. 2.
Step 4: (Implemented in Excel and Matlab): Send the improved
CCMs with the acceptable consistency to Matlab by Excel

Excel Link
Send arguments
to Matlab

Matlab operations
Input complex
comparison
matrices

Stochastic consistency
improving method
N

Calculate the
expected index

Cons?

Output complex
comparison
matrices
Excel Link
Send results
to Excel

Fig. 1. Acceptable possibilities of the elements for each rank.

Fig. 2. Flowchart of step 4.

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

Step 5:

Step 6:

Step 7:

Step 8:

Link, then utilize the EVM-SPM to obtain priorities of the

clusters and priorities of the clusters of elements, and then
return the priorities to Excel by Excel Link.
(Implemented in Excel): Construct an unweighted supermatrix by the macros of Excel worksheets lled with the
priorities of the clusters of elements.
(Implemented in Excel and Matlab): Send the unweighted
supermatrix and the priorities of the clusters to Matlab
by Excel Link, then construct a stochastic matrix whose
columns sum to unity, and then return the stochastic
matrix.
(Implemented in Excel and Matlab): Send the stochastic
matrix obtained in step 7 to Matlab by Excel Link, then obtain limiting priorities of all the elements, and then return
the limiting priorities to Excel by Excel Link.
(Implemented in Excel): Calculate nal priorities of the elements by the macros of Excel worksheets based on the
limiting priorities.

6. A practical application
For a country, energy channels are signicant for its development
by transporting power, oil and gas etc. However, the security of the
energy channels is often challenged by hostile forces, where piracy is
one of them. Supported by Institute of Meteorology, PLA University
of Science and Technology, 100 military experts are invited to participate this project of the piracy risk assessment to the energy channels
of China.
Since this problem can be hierarchically constructed, and elements
in the levels of the hierarchy inuence each other, we use the G-ANP
to get nal priorities of the elements. Then by aggregating normalized
property values and the nal priorities of the elements, we can get the
risk scores of the energy channels to assess piracy risk of the energy
channels.
The military experts are required to determine the related elements that inuence the energy channels of China with respect to
piracy risk, and provide judgments over paired comparisons of the
elements. By analysis, it is concluded that there are ve main clusters of elements inuencing the energy channels, where some of the
elements also inuence each other resulting in a network.
Considering the large quantity of judgments provided by the military experts, we use a statistic method to collect these judgments,
which can better represent the original judgments without aggregation. So the judgments are with probability interpretations. Moreover,
the reason why we use the G-ANP instead of the ANP in this example is
that some military experts prefer to use the 0.10.9 scale as a basis to
provide their judgments, and many judgments cannot be aggregated
due to the divergent opinions among the experts. As required by the

military experts to allow more exible methods to represent their

preferences, and fully consider their original judgments, we develop
the G-ANP to enhance the modeling abilities of the ANP in practice.
As discussed by the military experts, the structure of this problem
is described by a network consisting of ve clusters regarding the
piracy risk assessment as the control criterion, where connections
between them are shown in Fig. 3. The connections of the clusters
indicate the ow of inuence between them.
The ve clusters, i.e. the energy channels, the sea control capability, the island chain blockade capability, the risk elements, and the
piracy, consist of several basic elements respectively. We list them
below.
(1) The energy channels: South Sea (E1), Strait of Malacca (E2),
Indian Ocean (E3), Gulf of Aden (E4), East Africa (E5), Cape of
Good Hope (E6), Arafura Sea (E7) and Okhotsk (E8).
(2) The sea control capability: Military bases (C1), the number of
garrison (C2), reconnaissance capabilities (C3), military support capabilities (C4) and logistical capacity (C5).
(3) The island chain blockade capability: Strategic position (B1),
geographical features (B2), stationed conditions (B3), diplomatic relations (B4) and support capabilities (B5).
(4) The risk elements: Geography (R1) economic conditions (R2)
political situation (R3) ethnic and religious (R4) international
checks and balances (R5).
(5) The piracy: The number of pirates (P1) weaponry (P2) event
location (P3) casualties/kidnapping (P4) and property damage
(P5).
Following the step by step implementation of the G-ANP in
Section 5.2, we give the detail below to illustrate the G-ANP of solving
this piracy risk assessment problem.
Step 1. The inuence relationships between the clusters of elements in the network are identied in a network table of elements
shown in Table 7, where the symbol  indicates that a top row
element is inuenced by a left-column element.
Similarly, the inuence relationships between the clusters in the
network can be shown in Table 8.
Step 2. The military experts provide their judgments over paired
comparisons of the elements that have relationships identied in
Table 7. There should be 75 CCMs according to Table 7, and ve CCMs
according to Table 8. For simplicity, we only give one CCM for illustration.
For example, according to Table 7, South sea (E1) is inuenced by
the cluster of elements as the sea control capability, i.e. C1, C2, C3, C4
and C5. With respect to the piracy risk assessment, and set E1 as the
control criteria for making paired comparisons of C1, C2, C3, C4 and C5,
the judgments provided by the military experts can be represented
p
by a CCM C = (cij )55 as follows:

Piracy risk assessment

Control hierarchy

Energy channels

Sea control capability

Piracy

Island chain
blockade capability

283

Risk elements

Fig. 3. The network of the piracy risk assessment.

Network hierarchy

284

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

Table 7
The network table of elements.
E1
E1
E2
E3
E4
E5
E6
E7
E8
C1
C2
C3
C4
C5
B1
B2
B3
B4
B5
R1
R2
R3
R4
R5
P1
P2
P3
P4
P5





























E2

E3

E4

E5

E6

E7

E8

C1

C2

C3

C4

C5

B1

B2

B3

B4

B5

R1

R2

R3

R4

R5

P1

P2

P3

P4

P5


















































































































































































































































































































































































































































































































































































E
C
B
R
P






















R


12

13




p13 (percent)

14




1/2
24
1
24
1/6
81
1/3
9
4
27
3
15
2
42
7
95
1/3
48
4
14

p12 (percent)

p14 (percent)

15

p15 (percent)

1
[2, 1/2]p

C = [1, 1/3]p

[6, 4]p
[3, 1/3]p

[1/2, 2]p
1
[1/4, 1/6]p
[1/3, 1/6]p
{1/2, 1/3}p

[1, 3]p
[4, 6]p
1
[1/7, 1/9]p
[3, 1]p

[1/6, 1/4]p
[3, 6]p
[7, 9]p
1
[1/4, 1/6]p

23

[1/3, 3]p
p
{2, 3}

[1/3, 1]p .

p
[4, 6]
1

p23 (percent)

24

p24 (percent)

25

p25 (percent)

34

p34 (percent)

According to the statistic method, the probability functions,

p
denoted by F (ijk ) = pijk (ij cij ; i, j = 1, 2, . . . , n), that specify the
distributions of the values in cij (i, j = 1, 2, . . . , n) are shown in Table 9.
Step 3. Follow the owchart in Fig. 2, all the CCMs are sent to
Matlab by Excel Link to check and improve consistency, and then
p
return improved CCMs to Excel. For C = (cij )55 , according to Eq. (6),
we get the expected index of C as E(CRC ) = 0.7618. Since E(CRC ) > 0.1,
we use Algorithm 3 to improve its consistency. Then we have the
improved C, denoted by C  as

1
{2.215, 1.3554, 0.9178}p

p
C  = {1, 0.8846, 0.4589}

{1.7151, 5, 3.3228}p
{3, 2, 0.763, 0.5, 0.4358}p

35

p35 (percent)

45

p45 (percent)

{0.4515, 0.7378, 1.0895}p

1

{0.5001, 0.2, 0.2445}p













1/2
5

1
32
2
58
1/5
9
1
50
5
28
4
79
3
58
8
2
1/2
31
5
27

2
44
3
28
1/4
10
2
28
6
45
5
3

9
3
1
21
6
59

{1, 1.1304, 2.1792}p

{1.9996, 5, 4.0901}p
1

{0.4877, 0.5562, 0.2, 0.1667}p {1.1868, 0.125, 0.1111}p

{0.3233, 0.3617}p
{2.0649, 1.2159, 0.5229}p

{0.5831, 0.2, 0.301}p

{2.0505, 1.7979, 5, 6}p
{0.8426, 8, 9}p
1

{0.25, 0.3825, 0.2727}p

where E(CRC  ) = 0.0873.







Table 9
The distribution table of F (ijk ) = pijk .
B













Table 8
The network table of clusters.
E







{0.3333, 0.5, 1.3106, 2, 2.2947}p

{3.0933, 2.7645}p

p
{0.4843, 0.8225, 1.9124}
,

{4, 2.6143, 3.6675}p

1

3
8

6
3

E2

0.00732
0
0.1298
0.15506
0.10113
0.02715
0.12069
0.45884
0.1253
0.14679
0.10565
0.46587
0.15639
0.55552
0.1246
0.0897
0.20118
0.029
0.27396
0.18723
0.03216
0.16175
0.3449
0.02119
0.35017
0.1832
0.13269
0.31275

E1

0
0.26485
0.20059
0.04684
0.05045
0.06675
0.22555
0.14497
0.1152
0.4525
0.1391
0.2432
0.0502
0.36082
0.10207
0.07922
0.26586
0.19203
0.57654
0.12898
0.11112
0.17743
0.00594
0.00606
0.36263
0.19584
0.35542
0.08005

Unweighted
supermatrix

E1
E2
E3
E4
E5
E6
E7
E8
C1
C2
C3
C4
C5
B1
B2
B3
B4
B5
R1
R2
R3
R4
R5
P1
P2
P3
P4
P5

0.40018
0.02064
0
0.13311
0.05532
0.28259
0.04086
0.0673
0.00847
0.40756
0.1391
0.13622
0.30865
0.04152
0.16368
0.62942
0.03084
0.13453
0.29996
0.10709
0.19838
0.18574
0.20883
0.19981
0.04675
0.51887
0.0259
0.20866

E3

Table 10
The unweighted supermatrix.

0.13282
0.37346
0.08093
0
0.00905
0.05488
0.11045
0.2384
0.43912
0.15049
0.19475
0.04564
0.17
0.27328
0.21459
0.12632
0.13251
0.25329
0.44907
0.15091
0.1734
0.19751
0.02911
0.12536
0.24157
0.06296
0.28968
0.28043

E4

0.14953
0.12472
0.03835
0.02175
0
0.01674
0.48194
0.16697
0.14207
0.28631
0.16493
0.13288
0.27381
0.22003
0.30443
0.07305
0.19486
0.20762
0.15347
0.05776
0.48912
0.25583
0.04382
0.01635
0.05467
0.63929
0.09264
0.19705

E5

C1

0.01949
0.10562
0.00928
0.08889
0.22586
0.08539
0.06069
0.40477
0
0
0
0
0
0.35107
0.45186
0.04634
0.14253
0.00819
0
0
0
0
0
0.24572
0.17442
0.05998
0.18885
0.33103

E8

0.08546
0.18097
0.03383
0.00084
0.33233
0.26082
0.10575
0
0.06213
0.56551
0.01892
0.00381
0.34962
0.06389
0.34098
0.07749
0.31645
0.20118
0.08871
0.30207
0.19715
0.34715
0.06492
0.20301
0.52608
0.0659
0.17958
0.02543

E7

0.34234
0.22396
0.0862
0.00201
0.04634
0.01969
0
0.27945
0.60324
0.2146
0.00991
0.1278
0.04445
0.62362
0.02336
0.2453
0.08862
0.0191
0.02539
0.28967
0.11761
0.38758
0.17975
0.06214
0.56379
0.11622
0.03539
0.22246

C1 = 0.0504,

E6

0.06924
0.02197
0.03094
0.11142
0.16905
0
0.09211
0.50527
0.21079
0.05459
0.21537
0.51243
0.00682
0.29724
0.06678
0.4876
0.12207
0.02631
0.21629
0.41918
0.04808
0.20773
0.10873
0.13149
0.49804
0.28066
0.00193
0.08788

C4

0.08779
0.19613
0.02999
0.03485
0.0502
0.31932
0.22637
0.05535
0
0
0
0
0
0.06017
0.4546
0.14203
0.22025
0.12294
0
0
0
0
0
0.44588
0.0359
0.04716
0.26924
0.20182

C3

0.10514
0.06972
0.01979
0.11946
0.41984
0.16828
0.04717
0.05059
0
0
0
0
0
0.43808
0.26689
0.01885
0.24429
0.03189
0
0
0
0
0
0.08008
0.02348
0.02764
0.38386
0.48494

C2

0.05374
0.21619
0.36695
0.08047
0.24303
0.01968
0.01033
0.00961
0
0
0
0
0
0.05437
0.30199
0.12796
0.42975
0.08592
0
0
0
0
0
0.12873
0.56791
0.15573
0.09295
0.05468

0.03628
0.24326
0.2777
0.03892
0.07321
0.11278
0.04628
0.17157
0
0
0
0
0
0.15105
0.04364
0.11632
0.16046
0.52852
0
0
0
0
0
0.2588
0.07389
0.15065
0.18551
0.33116

C5

B2

0.12993
0.0875
0.13865
0.04207
0.22354
0.02487
0.24682
0.10661
0.08694
0.04206
0.18874
0.23121
0.45106
0
0
0
0
0
0.2386
0.0838
0.09454
0.0171
0.56596
0
0
0
0
0

B1

0.06546
0.19259
0.04802
0.05469
0.07183
0.41639
0.1176
0.03341
0.30581
0.03107
0.05697
0.30476
0.30139
0
0
0
0
0
0.1486
0.65948
0.04495
0.1326
0.01437
0
0
0
0
0

0.00902
0.03694
0.37841
0.27944
0.0279
0.00139
0.22508
0.04182
0.5721
0.14513
0.00634
0.26918
0.00725
0
0
0
0
0
0.04936
0.04961
0.22845
0.34375
0.32883
0
0
0
0
0

B3

0.00756
0.01573
0.17835
0.36319
0.21605
0.04245
0.16563
0.01103
0.42699
0.47032
0.0157
0.06355
0.02344
0
0
0
0
0
0.201
0.07442
0.14418
0.07953
0.50086
0
0
0
0
0

B4

0.11005
0.18706
0.0613
0.00255
0.09249
0.00781
0.32001
0.21874
0.04796
0.2478
0.06353
0.47595
0.16477
0
0
0
0
0
0.04914
0.54785
0.26131
0.02838
0.11332
0
0
0
0
0

B5

0.00687
0.07212
0.35224
0.0294
0.02181
0.19263
0.02376
0.30118
0
0
0
0
0
0.37733
0.13753
0.39139
0.09273
0.00102
0
0
0
0
0
0.05842
0.37483
0.00871
0.39763
0.16041

R1

R3

0.53573
0.19142
0.03965
2.8E-05
0.1315
0.04856
0.0129
0.04022
0
0
0
0
0
0.41717
0.27596
0.19548
0.05129
0.0601
0
0
0
0
0
0.43329
0.08827
0.22443
0.12691
0.1271

R2

0.06612
0.18481
0.11395
0.26651
0.14615
0.03923
0.07073
0.1125
0
0
0
0
0
0.11758
0.08853
0.10732
0.41651
0.27005
0
0
0
0
0
0.32531
0.29295
0.12334
0.19404
0.06436

R5

0.36964
0.07732
0.36374
0.02319
0.09651
0.0084
0.04899
0.01222
0
0
0
0
0
0.165
0.07782
0.0455
0.16755
0.54414
0
0
0
0
0
0.01529
0.04683
0.34983
0.33871
0.24933

R4

0.45764
0.06446
0.23706
0.0862
0.05267
0.0332
0.04696
0.02181
0
0
0
0
0
0.3604
0.44444
0.01093
0.09804
0.08619
0
0
0
0
0
0.1833
0.0427
0.24301
0.18365
0.34734

P2

0.08394
0.23856
0.08496
0.08306
0.10145
0.02512
0.29742
0.08551
0.0627
0.08988
0.18454
0.10883
0.55406
0
0
0
0
0
0.11212
0.39009
0.23438
0.14789
0.11551
0
0
0
0
0

P1

0.17077
0.08276
0.23561
0.09793
0.09117
0.14494
0.13845
0.03839
0.15401
0.35506
0.15896
0.05764
0.27432
0
0
0
0
0
0.27214
0.02428
0.30904
0.30799
0.08655
0
0
0
0
0

0.29947
0.25384
0.04566
0.17684
0.01849
0.08471
0.01799
0.10299
0.13053
0.0898
0.53731
0.14386
0.09849
0
0
0
0
0
0.00549
0.15737
0.09946
0.26722
0.47045
0
0
0
0
0

P3

P5

0.09784
0.11358
0.03706
0.46018
0.03785
0.04251
0.06671
0.14425
0.32149
0.22977
0.33102
0.01757
0.10015
0
0
0
0
0
0.27853
0.00889
0.19328
0.1098
0.40951
0
0
0
0
0

P4

0.29971
0.18031
0.0034
0.06424
0.00956
0.12755
0.30706
0.00817
0.17702
0.20153
0.04421
0.54993
0.02731
0
0
0
0
0
0.0282
0.01067
0.46108
0.18212
0.31794
0
0
0
0
0

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

E
C
B
R
P

R1 = 0.0269,

R1 = 0.0343,

C2 = 0.0555,

285

Table 11
The priorities of the clusters.
E
C
B
R
P

0.1270
0.3877
0.1911
0.1076
0.1886
0.0975
0
0.5348
0
0.3676
0.2785
0.2684
0
0.4531
0
0.4854
0
0.4018
0
0.1128
0.1419
0.6584
0
0.1997
0

Steps 4 and 5. Utilize the EVM-SPM to get priorities of the clusters of

elements from the improved comparison matrices related to Table 7,
then construct an unweighted supermatrix, shown in Table 10. Particularly, the obtained priorities of C1, C2, C3, C4 and C5 against E1 are
as follows:

pE1
C1 = 0.1152,
pE1
C2 = 0.4525,

pE1
C4 = 0.2432,
and

pE1
C3 = 0.1391,

pE1
C5 = 0.0502

Similarly, the priorities of the clusters can also be obtained, shown

in Table 11, which are used to weight the clusters of elements in the
block of the unweighted supermatrix.
Step 6. According to the priorities of the clusters and the unweighted supermatrix, we can construct a stochastic matrix, shown
in Table 12.
Step 7. For a stochastic supermatrix, limiting priorities of all the
elements depend on its reducibility, primitivity, and cyclicity (Saaty,
2004). Since the stochastic matrix in step 7 is irreducibility, and the
eigenvalue = 1 is single, we develop an iterative method shown
in Algorithm 6 to output the limiting priorities from the stochastic
matrix.
Algorithm 6.

Step 1: Input the supermatrix M = (mij )nn , and let d = 0.00005 be

the accuracy.
Step 2: Let k = 0, output the initial eigenvector (k) by the EVM.
Step 3: Calculate (k+1) = M(k), then normalize (k+1) = (k+1)/
n
(k+1)
.
i=1 i
Step 4: If |i(k+1) i(k+1)| < d (i = 1, 2, . . . , n), output (k+1) and
then go to step 5; else let k = k + 1, and then go to step 2.
Step 5: End.

According to Algorithm 6, the limiting priorities of all the elements

can be obtained after 24 iterations listed as follows
E1 = 0.0350,
E2 = 0.0281,
E3 = 0.0345,
E4 = 0.0231,

E5 = 0.0246,
E6 = 0.0195,
E7 = 0.0266,

C3 = 0.0381,
C4 = 0.0502,
E8 = 0.0233

C5 = 0.0493,

B1 = 0.0546,
B2 = 0.0578,
B3 = 0.0307,
B4 = 0.0505,
B5 = 0.0415,

R2 = 0.0392,
R3 = 0.0267,
R4 = 0.0220,
R5 = 0.0443,

P1 = 0.0284,

P2 = 0.0329,

P3 = 0.0225,

P4 = 0.0295,

P5 = 0.0342.

Step 8. Apart from the energy channels, we normalize the limiting

priorities to get the nal priorities, denoted as follows:

C1 = 0.0642, C2 = 0.0706, C3 = 0.0485, C4 = 0.0639, C5 = 0.0628,

B1 = 0.0695, B2 = 0.0736, B3 = 0.0391, B4 = 0.0644, B5 = 0.0529,

P1 = 0.0362, P2 = 0.0419, P3 = 0.0287, P4 = 0.0376, P5 = 0.0435.

R2 = 0.0499, R3 = 0.0340, R4 = 0.0281, R5 = 0.0564,

With respect to each energy channel, the property values of the

elements are obtained by some public resources and National Administration of Surveying, Mapping and Geoinformation. Since the
process of data acquisition is not related to our topic, we directly list
the normalized property values in Table 13.

E6

E7

E8

C1

C2

C3

0.05082
0.00262
0
0.0169
0.00703
0.03589
0.00519
0.00855
0.00329
0.15801
0.05393
0.05281
0.11966
0.00793
0.03128
0.12028
0.00589
0.02571
0.03228
0.01152
0.02135
0.01999
0.02247
0.03728
0.00872
0.09682
0.00483
0.03894

0.01687
0.04743
0.01028
0
0.00115
0.00697
0.01403
0.03028
0.17025
0.05834
0.07551
0.01769
0.06591
0.05222
0.04101
0.02414
0.02532
0.0484
0.04832
0.01624
0.01866
0.02125
0.00313
0.02339
0.04508
0.01175
0.05405
0.05233

0.01899
0.01584
0.00487
0.00276
0
0.00213
0.06121
0.0212
0.05508
0.111
0.06394
0.05152
0.10616
0.04205
0.05818
0.01396
0.03724
0.03968
0.01651
0.00622
0.05263
0.02753
0.00471
0.00305
0.0102
0.11929
0.01729
0.03677

0.00879
0.00279
0.00393
0.01415
0.02147
0
0.0117
0.06417
0.08172
0.02117
0.0835
0.19867
0.00264
0.0568
0.01276
0.09318
0.02333
0.00503
0.02327
0.0451
0.00517
0.02235
0.0117
0.02454
0.09293
0.05237
0.00036
0.0164

0.04348
0.02844
0.01095
0.00026
0.00589
0.0025
0
0.03549
0.23388
0.0832
0.00384
0.04955
0.01723
0.11917
0.00446
0.04688
0.01694
0.00365
0.00273
0.03117
0.01265
0.0417
0.01934
0.01159
0.1052
0.02169
0.0066
0.04151

0.01085
0.02298
0.0043
0.00011
0.04221
0.03312
0.01343
0
0.02409
0.21925
0.00733
0.00148
0.13555
0.01221
0.06516
0.01481
0.06047
0.03845
0.00954
0.0325
0.02121
0.03735
0.00698
0.03788
0.09817
0.0123
0.03351
0.00475

0.0019
0.0103
0.0009
0.00867
0.02202
0.00833
0.00592
0.03947
0
0
0
0
0
0.18775
0.24165
0.02478
0.07623
0.00438
0
0
0
0
0
0.09033
0.06412
0.02205
0.06942
0.12169

0.00524
0.02108
0.03578
0.00785
0.0237
0.00192
0.00101
0.00094
0
0
0
0
0
0.02908
0.16151
0.06843
0.22983
0.04595
0
0
0
0
0
0.04732
0.20876
0.05725
0.03417
0.0201

In this paper, we have developed a generalized analytic network

process (G-ANP) as an important extension of the analytic hierarchy
process (ANP) to improve its modeling abilities in practice. In the
framework of the ANP, the contribution of this study is the development of a systematic concepts and methods to deal with multiple forms of preferences which are called complex judgments. Another contribution is the implementation of the G-ANP in Matlab
and Excel environments as a working prototype for a decision support system. The G-ANP also eliminate the drawbacks of the fuzzy
analytic network process (F-ANP) that cannot take consistency into
account.
In the G-ANP, as a new prioritization method, an eigenvector method based stochastic preference method (EVM-SPM) has
been developed to derive priorities from complex comparison
matrices (CCMs). As comparison matrices, interval comparison
matrices, hesitant comparison matrices, and stochastic comparison matrices can be considered as special cases of CCMs, the
EVM-SPM can straightforwardly be applied to them for deriving
priorities.
We have applied the G-ANP to a piracy risk assessment problem.
By a statistic method, the preferences provided by 100 military experts are collected and specied by probability functions. Following
a step by step procedure of the G-ANP, we show the detail to demonstrate the G-ANP in dealing with this problem. The G-ANP is specially
useful for decision making problems with large data sets, or even
big data. In the future work, more studies can focus on CCMs related
to consistency checking, consistency improving, and prioritization,
etc.

0.00093
0
0.01648
0.01969
0.01284
0.00345
0.01533
0.05827
0.04858
0.05691
0.04096
0.18062
0.06063
0.10616
0.02381
0.01714
0.03845
0.00554
0.02948
0.02015
0.00346
0.0174
0.03711
0.00395
0.06534
0.03419
0.02476
0.05836

0.00354
0.02372
0.02708
0.00379
0.00714
0.011
0.00451
0.01673
0
0
0
0
0
0.08078
0.02334
0.06221
0.08581
0.28265
0
0
0
0
0
0.09514
0.02716
0.05538
0.06819
0.12173

E5

0.00856
0.01912
0.00292
0.0034
0.00489
0.03113
0.02207
0.0054
0
0
0
0
0
0.03218
0.24312
0.07596
0.11779
0.06575
0
0
0
0
0
0.16391
0.0132
0.01733
0.09897
0.07419

E4

0.01025
0.0068
0.00193
0.01165
0.04093
0.01641
0.0046
0.00493
0
0
0
0
0
0.23429
0.14273
0.01008
0.13065
0.01706
0
0
0
0
0
0.02944
0.00863
0.01016
0.14111
0.17827

0
0.03364
0.02548
0.00595
0.00641
0.00848
0.02864
0.01841
0.04466
0.17543
0.05393
0.09429
0.01946
0.06895
0.01951
0.01514
0.05081
0.0367
0.06204
0.01388
0.01196
0.01909
0.00064
0.00113
0.06767
0.03654
0.06632
0.01494

E1
E2
E3
E4
E5
E6
E7
E8
C1
C2
C3
C4
C5
B1
B2
B3
B4
B5
R1
R2
R3
R4
R5
P1
P2
P3
P4
P5

E3

C5

E2

7. Conclusions

C4

E1

Stochastic
supermatrix

Table 12
The stochastic supermatrix.

0.01823
0.05364
0.01337
0.01523
0.02
0.11597
0.03275
0.00931
0.08208
0.00834
0.01529
0.0818
0.08089
0
0
0
0
0
0.06733
0.29881
0.02037
0.06008
0.00651
0
0
0
0
0

B1

0.03619
0.02437
0.03861
0.01172
0.06226
0.00693
0.06874
0.02969
0.02333
0.01129
0.05066
0.06206
0.12106
0
0
0
0
0
0.10811
0.03797
0.04284
0.00775
0.25644
0
0
0
0
0

B2

0.00251
0.01029
0.10539
0.07783
0.00777
0.00039
0.06269
0.01165
0.15355
0.03895
0.0017
0.07225
0.00195
0
0
0
0
0
0.02236
0.02248
0.10351
0.15575
0.14899
0
0
0
0
0

B3

B5

0.03065
0.0521
0.01707
0.00071
0.02576
0.00217
0.08912
0.06092
0.01287
0.06651
0.01705
0.12774
0.04422
0
0
0
0
0
0.02227
0.24823
0.1184
0.01286
0.05135
0
0
0
0
0

B4

0.00211
0.00438
0.04967
0.10115
0.06017
0.01182
0.04613
0.00307
0.1146
0.12623
0.00421
0.01706
0.00629
0
0
0
0
0
0.09107
0.03372
0.06533
0.03604
0.22694
0
0
0
0
0

R2

0.03209
0.08971
0.05531
0.12936
0.07094
0.01904
0.03433
0.05461
0
0
0
0
0
0.04724
0.03557
0.04312
0.16735
0.10851
0
0
0
0
0
0.03669
0.03304
0.01391
0.02189
0.00726

R1

0.00334
0.03501
0.17097
0.01427
0.01059
0.0935
0.01153
0.14619
0
0
0
0
0
0.15161
0.05526
0.15726
0.03726
0.00041
0
0
0
0
0
0.00659
0.04228
0.00098
0.04485
0.01809

0.26004
0.09291
0.01924
1.4E-05
0.06383
0.02357
0.00626
0.01952
0
0
0
0
0
0.16762
0.11088
0.07854
0.02061
0.02415
0
0
0
0
0
0.04887
0.00996
0.02532
0.01432
0.01434

R3

0.22214
0.03129
0.11507
0.04184
0.02556
0.01612
0.02279
0.01059
0
0
0
0
0
0.14481
0.17858
0.00439
0.03939
0.03463
0
0
0
0
0
0.02068
0.00482
0.02741
0.02072
0.03918

R4

0.17942
0.03753
0.17656
0.01125
0.04684
0.00408
0.02378
0.00593
0
0
0
0
0
0.0663
0.03127
0.01828
0.06732
0.21863
0
0
0
0
0
0.00172
0.00528
0.03946
0.03821
0.02812

R5

0.02423
0.01174
0.03343
0.0139
0.01294
0.02057
0.01965
0.00545
0.1014
0.23377
0.10466
0.03795
0.18061
0
0
0
0
0
0.05435
0.00485
0.06172
0.0615
0.01728
0
0
0
0
0

P1

0.01191
0.03385
0.01206
0.01179
0.0144
0.00356
0.0422
0.01213
0.04128
0.05918
0.1215
0.07165
0.36479
0
0
0
0
0
0.02239
0.0779
0.04681
0.02953
0.02307
0
0
0
0
0

P2

0.04249
0.03602
0.00648
0.02509
0.00262
0.01202
0.00255
0.01461
0.08594
0.05913
0.35377
0.09472
0.06484
0
0
0
0
0
0.0011
0.03143
0.01986
0.05336
0.09395
0
0
0
0
0

P3

P5

0.01388
0.01612
0.00526
0.0653
0.00537
0.00603
0.00947
0.02047
0.21167
0.15128
0.21794
0.01157
0.06594
0
0
0
0
0
0.05562
0.00177
0.0386
0.02193
0.08178
0
0
0
0
0

P4

0.04253
0.02559
0.00048
0.00912
0.00136
0.0181
0.04357
0.00116
0.11655
0.13269
0.0291
0.36207
0.01798
0
0
0
0
0
0.00563
0.00213
0.09208
0.03637
0.06349
0
0
0
0
0

286
B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288
Table 13
The normalized property values with respect to the energy channels.

C1
C2
C3
C4
C5
B1
B2
B3
B4
B5
R1
R2
R3
R4
R5
P1
P2
P3
P4
P5
E1
E2
E3
E4
E5
E6
E7
E8

0.91
0.92
0.5
0.5
0.3
0.84
0.6
0.82
0.6
0.4
0.93
0.7
0.23
0.3
0.6
0.2
0.54
0.3
0.3
0.5
0.56
0.6
0.67
0.5
0.3
0.71
0.6
0.76
0.5
0.4
0.65
0.5
0.5
0.63
0.3
0.73
0.55
0.54
0.15
0.15
0.4
0.5
0.7
0.7
0.5
0.5
0.48
0.2
0.4
0.3
0.4
0.4
0.7
0.5
0.7
0.4
0.5
0.6
0.1
0.65
0.6
0.6
0.9
0.75
0.5
0.6
0.4
0.41
0.3
0.6
0.4
0.2
0.9
0.85
0.9
0.4
0.65
0.3
0.45
0.8
0.4
0.6
0.3
0.6
0.5
0.7
0.75
0.4
0.3
0.4
0.4
0.2
0.75
0.8
0.7
0.35
0.47
0.3
0.5
0.5
0.05
0.1
0.2
0.1
0.06
0.1
0.15
0.05
0.14
0.1
0.1
0.2
0.1
0.05
0.05
0.1
0.05
0.1
0.05
0.1
0.55
0.15
0.3
0.1
0.4
0.4
0.3
0.2
0.5
0.3
0.3
0.1
0.45
0.3
0.1
0.15
0.5
0.7
0.2
0.14
0.1
0.1
0.05
0.1
0.1
0.1
0.05
0.2
0.05
0.15
0.2
0.05
0.1
0.2
0.2
0.1
0.15
0.1
0.04
0.25

Then by aggregating the normalized property values and the nal

priorities, we can get the risk scores of the eight energy channels,
denoted as follows:

S(E1) = 0.5812,
S(E2) = 0.5141,
S(E3) = 0.4863,
S(E4) = 0.5731,

S(E5) = 0.5051,
S(E6) = 0.1001,
S(E7) = 0.3025,
S(E8) = 0.1131.

The bigger value the risk scores, the higher level of risk
the energy channels have. We use regions in different colors to
show the risk levels of piracy for the eight energy channels in
Fig. 4.
It is clear that South Sea and Gulf of Aden have high risk levels of
piracy. So they deserve more protection from piracy.

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

287

Fig. 4. The risk levels of piracy for the eight energy channels.

Acknowledgments
The authors would like to thank the anonymous referees for their
insightful and constructive comments and suggestions that have led
to an improved version of this paper. The work was supported by the
Project Funded by China Postdoctoral Science Foundation.
Appendix
Proof of Theorem 2. (Convergence of Algorithm 4).
Lemma 1. (Saaty, 1980) Let A be a n n positive comparison matrix
and max (A) be the maximal eigenvalue of A. Then max n and equality
holds if and only if A is consistent.
For the CCM C = (cij )nn , since E(CRC (s) ) > 0, C is inconsistent. Folp

lowing the distributions of cij , i, j = 1, 2, . . . , n, we stochasticlly get the

comparison matrix A = (aij )nn (aij cij ).

Use Algorithm 2 to improve A. According to Theorem 1, we
have limk max (A(k)) = n. According to Lemma 1, A(k) is consistent, which means that CRA(k) = 0. Therefore, in the judgment space
of C, according to Eq. (6), we have E(CRC (s+1) ) < E(CRC (s) ).
p
Since the stochasticlly obtained comparison matrix A follows cij ,
p

and the improved A(k) is consistent, then A(k) is consistent everywhere in the judgment space of C. Therefore, let A(s) be the stochastically obtained comparison matrix in the sth iteration of Algorithm
4, then lims A(s) = Z, where Z is a consistent stochastic comparison matrix satisfying CRZ = 0. Further according to Eq. (6), we have
lims E(CRC (s) ) = 0.
This completes the proof of Theorem 2.
References
Barzilai, J. (1997). Deriving weights from pairwise comparison matrices. Journal of the
Operational Research Society, 48, 12261232.
Bryson, N. (1995). A goal programming method for generating priority vectors. Journal
of the Operational Research Society, 46, 641648.
Cao, D., Leung, L. C., & Law, J. (2008). Modifying inconsistent comparison matrix in
analytic hierarchy process: A heuristic approach. Decision Support Systems, 44, 944
953.
Chung, S. -H., Lee, A. H., & Pearn, W. -L. (2005). Analytic network process (ANP) approach for product mix planning in semiconductor fabricator. International Journal
of Production Economics, 96, 1536.
Crawford, G. (1987). The geometric mean procedure for estimating the scale of a judgement matrix. Mathematical Modelling, 9, 327334.

Crawford, G., & Williams, C. (1985). A note on the analysis of subjective judgment
matrices. Journal of Mathematical Psychology, 29, 387405.
Ergu, D., Kou, G., Peng, Y., & Shi, Y. (2011). A simple method to improve the consistency
ratio of the pair-wise comparison matrix in ANP. European Journal of Operational
Research, 213, 246259.
Escobar, M. T., Aguarn, J., & Moreno-Jimnez, J. M. (2004). A note on AHP group
consistency for the row geometric mean priorization procedure. European Journal
of Operational Research, 153, 318322.
Gong, Z. -W. (2008). Least-square method to priority of the fuzzy preference relations
with incomplete information. International Journal of Approximate Reasoning, 47,
258264.
Hahn, E. D. (2003). Decision making with uncertain judgments: A stochastic formulation
of the analytic hierarchy process. Decision Sciences, 34, 443466.
Jharkharia, S., & Shankar, R. (2007). Selection of logistics service provider: An analytic
network process (ANP) approach. Omega, 35, 274289.
Lahdelma, R., & Salminen, P. (2001). SMAA-2: Stochastic multicriteria acceptability
analysis for group decision making. Operations Research, 49, 444454.
Lee, J. W., & Kim, S. H. (2000). Using analytic network process and goal programming
for interdependent information system project selection. Computers & Operations
Research, 27, 367382.
Li, H. -L., & Ma, L. -C. (2007). Detecting and adjusting ordinal and cardinal inconsistencies
through a graphical and optimal approach in AHP models. Computers & Operations
Research, 34, 780798.
Meade, L. M., & Presley, A. (2002). R&D project selection using the analytic network
process. IEEE Transactions on Engineering Management, 49, 5966.
Mikhailov, L. (2000). A fuzzy programming method for deriving priorities in the
analytic hierarchy process. Journal of the Operational Research Society, 51, 341349.
Mikhailov, L. (2004). A fuzzy approach to deriving priorities from interval pairwise
comparison judgements. European Journal of Operational Research, 159, 687704.
Mikhailov, L., & Singh, M. G. (2003). Fuzzy analytic network process and its application
to the development of decision support systems. IEEE Transactions on Systems, Man,
and Cybernetics, Part C: Applications and Reviews,, 33, 3341.
Milton, J., & Arnold, J.C. (1995). Introduction to probability and statistics. New York:
McGraw Hill.
Moskowitz, H., Tang, J., & Lam, P. (2000). Distribution of aggregate utility using stochastic elements of additive multiattribute utility models. Decision Sciences, 31, 327
360.
Niemira, M. P., & Saaty, T. L. (2004). An analytic network process model for nancialcrisis forecasting. International Journal of Forecasting, 20, 573587.
Rosenbloom, E. (1997). A probabilistic interpretation of the nal rankings in AHP.
European Journal of Operational Research, 96, 371378.
Saaty, T. (1989). Group decision making and the AHP. The analytic hierarchy process
(pp. 5967). Berlin, Heidelberg: Springer.
Saaty, T. L. (1977). A scaling method for priorities in hierarchical structures. Journal of
Mathematical Psychology, 15, 234281.
Saaty, T. L. (1980). The analytic hierarchy process. New York: McGraw-Hill.
Saaty, T. L. (2001). The analytic network process: Decision making with dependence and
feedback. RWS Publications.
Saaty, T. L. (2004). Fundamentals of the analytic network processDependence and
feedback in decision-making with a single network. Journal of Systems Science and
Systems Engineering, 13, 129157.
Saaty, T. L. (2008). Decision making with the analytic hierarchy process. International
Journal of Services Sciences, 1, 8398.
Saaty, T. L., & Vargas, L. G. (1987). Uncertainty and rank order in the analytic hierarchy
process. European Journal of Operational Research, 32, 107117.

288

B. Zhu et al. / European Journal of Operational Research 244 (2015) 277288

Tanino, T. (1984). Fuzzy preference orderings in group decision making. Fuzzy Sets and
Systems, 12, 117131.
Torra, V. (2010). Hesitant fuzzy sets. International Journal of Intelligent Systems, 25,
529539.
Wang, Y. -M. (2006). On lexicographic goal programming method for generating
weights from inconsistent interval comparison matrices. Applied Mathematics and
Computation, 173, 985991.
Wang, Y. -M., & Elhag, T. (2007). A goal programming method for obtaining interval weights from an interval comparison matrix. European Journal of Operational
Research, 177, 458471.
Wang, Y. -M., Fan, Z. -P., & Hua, Z. (2007). A chi-square method for obtaining a priority vector from multiplicative and fuzzy preference relations. European Journal of
Operational Research, 182, 356366.
Wang, Y. -M., Parkan, C., & Luo, Y. (2008). A linear programming method for generating the most favorable weights from a pairwise comparison matrix. Computers &
Operations Research, 35, 39183930.
Wang, Y. -M., Yang, J. -B., & Xu, D. -L. (2005a). Interval weight generation approaches
based on consistency test and interval comparison matrices. Applied Mathematics
and Computation, 167, 252273.
Wang, Y. -M., Yang, J. -B., & Xu, D. -L. (2005b). A two-stage logarithmic goal programming method for generating weights from interval comparison matrices. Fuzzy Sets
and Systems, 152, 475498.
Wolfslehner, B., Vacik, H., & Lexer, M. J. (2005). Application of the analytic network
process in multi-criteria analysis of sustainable forest management. Forest Ecology
and Management, 207, 157170.
Xia, M. M., & Xu, Z. S. (2011). Hesitant fuzzy information aggregation in decision making.
International Journal of Approximate Reasoning, 52, 395407.
Xia, M. M., & Xu, Z. S. (2013). Managing hesitant information in GDM problems under
fuzzy and multiplicative preference relations. International Journal of Uncertainty,
Fuzziness and Knowledge-Based Systems, 21, 865897.
Xia, M. M., Xu, Z. S., & Chen, N. (2013). Some hesitant fuzzy aggregation operators
with their application in group decision making. Group Decision and Negotiation,
22, 259279.

Xu, Z. S. (2002). Two methods for priorities of complementary judgement matricesweighted least-square method and eigenvector method. Systems Engineering
Theory & Practice, 22, 7175.
Xu, Z. S. (2004a). Goal programming models for obtaining the priority vector of incomplete fuzzy preference relation. International Journal of Approximate Reasoning, 36,
261270.
Xu, Z. S. (2004b). On compatibility of interval fuzzy preference relations. Fuzzy Optimization and Decision Making, 3, 217225.
Xu, Z. S. (2006). AC-OWA operator-based approach to decision making with interval
fuzzy preference relation. International Journal of Intelligent Systems, 21, 1289
1298.
Xu, Z. S., & Chen, J. (2008). Some models for deriving the priority weights from interval
fuzzy preference relations. European Journal of Operational Research, 184, 266
280.
Xu, Z. S., & Da, Q. L. (2005). A least deviation method to obtain a priority vector of a
fuzzy preference relation. European Journal of Operational Research, 164, 206216.
Xu, Z. S., & Wei, C. P. (1999). A consistency improving method in the analytic hierarchy
process. European Journal of Operational Research, 116, 443449.
Xu, Z. S., & Xia, M. M. (2011). Distance and similarity measures for hesitant fuzzy sets.
Information Sciences, 181, 21282138.
Yksel, I ., & Dagdeviren, M. (2007). Using the analytic network process (ANP) in a SWOT
analysisA case study for a textile rm. Information Sciences, 177, 33643382.
Zadeh, L. A. (1975). The concept of a linguistic variable and its application to approximate reasoningI. Information Sciences, 8, 199249.
Zhu, B., & Xu, Z. S. (2013). Regression methods for hesitant fuzzy preference relations.
Technological and Economic Development of Economy, 19, S214S227.
Zhu, B., & Xu, Z. S. (2014a). Analytic hierarchy process-hesitant group decision making.
European Journal of Operational Research, 239, 794801.
Zhu, B., & Xu, Z. S. (2014b). Stochastic preference analysis in numerical preference
relations. European Journal of Operational Research, 237, 628633.
Zhu, B., Xu, Z. S., & Xu, J. P. (2013). Deriving a ranking from hesitant fuzzy preference
relations under group decision making. IEEE Transactions on Systems, Man, and
Cybernetics, Part B: Cybernetics, 44, 13281337.