Applied Mathematical Modelling 30 (2006) 920929

www.elsevier.com/locate/apm

A bulk arrival queueing model with fuzzy parameters

and varying batch sizes
Shih-Pin Chen *

Department of Business Administration, National Chung Cheng University, Ming-Hsiung, Chia-Yi 621, Taiwan

Received 1 November 2004; received in revised form 1 May 2005; accepted 27 June 2005
Available online 18 August 2005

Abstract

This paper develops a nonlinear programming approach to derive the membership functions of the
steady-state performance measures in bulk arrival queueing systems with varying batch sizes, in that the
arrival rate and service rate are fuzzy numbers. The basic idea is based on Zadehs extension principle.
Two pairs of mixed integer nonlinear programs (MINLP) with binary variables are formulated to calculate
the upper and lower bounds of the system performance measure at possibility level a. From dierent values
of a, the membership function of the system performance measure is constructed. For practice use, the
defuzzication of performance measures is also provided via Yager ranking index. To demonstrate the
validity of the proposed method, a numerical example is solved successfully.
2005 Elsevier Inc. All rights reserved.

Keywords: Bulk arrival; Queue; Fuzzy sets; Mixed integer nonlinear programming

1. Introduction

Bulk arrival queueing models have been widely applied to many practical situations, such as
production/manufacturing systems, communication systems, and computer networks [1,2].
For example, when the operation in a production/manufacturing system will not begin until a

*
Tel.: +886 5 2720411; fax: +886 5 2720564.
E-mail address: chensp@ccu.edu.tw

0307-904X/$ - see front matter 2005 Elsevier Inc. All rights reserved.
doi:10.1016/j.apm.2005.06.002
 S.-P. Chen / Applied Mathematical Modelling 30 (2006) 920929 921

specied number of raw materials are accumulated during an idle period, we often analyze this
system by a bulk arrival queueing model that provides a powerful tool for evaluating the system
performance.
In crisp environments, many articles on this topic have been published (for example, [37]).
Most of the related studies are based on traditional queueing theory, in that the interarrival
times and service times are assumed to follow certain probability distributions. However, in
practice there are cases that these parameters may be obtained subjectively [8]. Thus, fuzzy
queues are much more realistic than the commonly used crisp queues [810]. Several researchers
have discussed fuzzy queueing systems, e.g., [814]; however, few articles have been published on
the fuzzy bulk arrival queue with varying batch size. If the usual crisp bulk arrival queues can be
extended to fuzzy bulk arrival queues, bulk arrival queueing models would have wider
applications.
Relatively few articles have been published on the topic of fuzzy queues. Based on Zadehs
extension principle [1517] and fuzzy Markov chain [18], Li and Lee [12] proposed a general ap-
proach for analyzing fuzzy queues. A straightforward idea is to apply their method to the fuzzy
bulk arrival queueing problem. However, Negi and Lee [13] commented that their approach is so
complicated that generally unsuitable for computational purposes; it can hardly derive analytical
results for other more complicated queueing systems [19]. Consequently, it is very dicult and
unsuitable to apply Li and Lees approach to fuzzy bulk arrival queues. Negi and Lee [13] pro-
posed a procedure using two-variable simulation [11] and the a-cut concept to analyze fuzzy
queues. Unfortunately, their approach only provided crisp solutions. Kao et al. [19] investigated
four simple fuzzy queueing problems, namely, M/F/1, F/M/1, F/F/1, and FM/FM/1, where F de-
notes fuzzy time and FM denotes fuzzied exponential time. It seems that their approach can be
applied to the fuzzy bulk arrival queues. However, in bulk arrival queueing systems, customers
arrive according to a compound Poisson process with random arrival size [1,2], which results in
the fuzzy bulk arrival queues are much more complicated than the above four fuzzy queues,
the solution procedure for the fuzzy bulk arrival queues is unknown and deserves further
investigation.
Clearly, when the arrival rate or the service rate is fuzzy, the system performance measure of the
bulk arrival queue should be fuzzy as well; it should be described by a membership function. In
this paper, we develop a solution procedure that is able to provide fuzzy performance measures
for bulk arrival queues with varying batch sizes, fuzzied exponential arrival rate and service rate.
The membership functions of performance measures can be derived completely. The basic idea is
to apply Zadehs extension principle [1517]. Two pairs of mixed integer nonlinear programming
(MINLP) models are formulated to calculate the lower and upper bounds of the a-cut of the sys-
tem performance measure. The membership function of the system performance measure is de-
rived analytically.

2. Fuzzy bulk arrival queues with varying batch sizes

Consider a queueing system in which customers arrive at a single-server facility in batches as a

Poisson process with group arrival rate ~
k of all batches, where ~k is a fuzzy number, and all service
times are independent and identically distributed according to an exponential distribution with
922 S.-P. Chen / Applied Mathematical Modelling 30 (2006) 920929

fuzzy service rate l ~. The actual number of customers in any arriving module is stochastically
equivalent to a generic random variable K, which may take on any positive integer with probabil-
ity f(K). Customers are served according to a rst-come-rst-served (FCFS) discipline, and both
the size of calling population and the system capacity are innite. This model will hereafter be
denoted by FM[K]/FM/1.
In this model the group arrival rate ~ k and service rate l ~ are approximately known and can be
represented by convex fuzzy sets. Note that a fuzzy set A ~ in its universal set Z is convex if
lA~ /z1 1  /z2 P minflA~ z1 ; lA~ z2 g, where lA~ is its membership function, / 2 [0, 1], and
z1, z2 2 Z. Let l~k x and ll~ y denote the membership functions of the group arrival rate and ser-
vice rate, respectively. We have
k fx; l~k xjx 2 S~
~ kg; 1
~ fy; ll~ yjy 2 S~
l lg; 2
where S~k and S~l are the supports of ~ k and l ~, which denote the universal sets of the arrival
rate and service rate, respectively. Let p(x, y) denote the system performance measure of interest.
Clearly, when ~k and l ~ are fuzzy numbers, ~p~k; l
~ will be fuzzy as well. On the basis of Zadehs
extension principle [1517], the membership function of the performance measure ~p~k; l ~ is de-
ned as
l~p~k;~l z sup minfl~k x; ll~ yjz px; yg. 3
x2X ;y2Y

~ l
If the a-cuts of p~k; ~ at all a values degenerate to the same point, then the value of the system
performance measure is a crisp number. Otherwise, it is a fuzzy number. Without loss of gener-
ality, assume that the performance measure of interest is the expected number of customers in the
queue Lq. From the knowledge of traditional queueing theory [1,2], under the steady-state condi-
tion q = xE[K]/y < 1, where E[K] denotes the expectation of K, the expected number of customers
in the queue of a crisp queueing system with bulk arrival is

xfyEK 2  2xEK2  yEKg

Lq . 4
2yfy  xEKg
Following (3), the membership function for L ~q is
( )
2
xfyEK 2  2xEK  yEKg
lL~q z sup min l~k x; ll~ yjz . 5
x2X ;y2Y ;xEK<y 2yfy  xEKg

Membership functions for other performance measures can be obtained according to Littles for-
mula [2] in the same manner. For example, the membership function for the expected system
length L~ is
 
xfEK EK 2 g
lL~ z sup min l~k x; ll~ yjz 6
x2X ;y2Y ;xEK<y 2fy  xEKg
~ and the expected
and the membership functions for the expected sojourn time in the system W
~ q respectively are
waiting time in the queue W
 S.-P. Chen / Applied Mathematical Modelling 30 (2006) 920929 923
( )
yEK 2  2xEK2  yEK
lW~ q z sup min l~k x; ll~ yjz ; 7
x2X ;y2Y ;xEK<y 2yfy  xEKg
 
EK EK 2 
lW~ z sup min l~k x; ll~ yjz . 8
x2X ;y2Y ;xEK<y 2fy  xEKg
Theoretically, the membership functions in (5)(8) are correct; however, they are not in the usual
forms for practical use. Moreover, it is even very dicult to imagine their shapes.

3. The solution procedure

One approach to construct the membership function l~p~k;~l is on the basis of deriving the a-cuts
(or a-level sets) of l~p~k;~l . Denote the a-cuts of ~k and l ~ as
 
 L U
ka xa ; xa minfxjl~k x P ag; maxfxjl~k x P ag ; 9a
x2X x2X
 
 
la y La ; y U
a minfyjll ~ y P ag; maxfyjll ~ y P ag . 9b
y2Y y2Y

These intervals indicate where the group arrival rate and service rate lie at possibility level a. Note
that ka and la are crisp sets rather than fuzzy sets. By the concept of a-cuts, the imbedded fuzzy
Markov chain in the FM[K]/FM/1 can be decomposed into a family of ordinary Markov chains
with dierent transition probability matrices, which are also parameterized by a. The arrival rate
and service rate can also be represented by dierent levels of condence intervals [13,17]. Conse-
quently, the FM[K]/FM/1 queue can be reduced to a family of crisp M[K]/M/1 queues with dierent
a-level sets {ka j 0 < a 6 1} and {la j 0 < a 6 1}. These two sets cause nested structures for express-
ing the relationship between ordinary sets and fuzzy sets [20], and they represent sets of moveable
boundaries.
Without loss of generality, the fuzzy arrival rate ~k and fuzzy service rate l ~ are assumed to be
fuzzy numbers in this model, as crisp values can be represented by degenerated membership func-
tions which only have one value in their domain. By the convexity of a fuzzy number [17], the
bounds of these intervals are functions of a and can be obtained as xLa min l1 U
~k a; xa
1 L 1 U 1
max l~k a; y a min ll~ a; and y a max ll~ a, respectively. Clearly, as dened in (3), the
membership function of ~ p~
k; l~ is also parameterized by a. Consequently, we can use its a-cut
to construct the corresponding membership function.
Consider the membership function of the expected number of customers in the queue L ~q .
According to (5), lL~q z is the minimum of l~k x and ll~ y. We need either l~k x a
and ll~ y P a or l~k x P a and ll~ y a such that z = x{yE[K2] + 2x(E[K])2  yE[K]}/2y{y 
xE[K]} to satisfy lL~q z a. To nd the membership function lL~q z, it suces to nd the left
shape function and the right shape function of lL~q z, which is equivalent to nd the lower bound
Z La and the upper bound Z U a of the a-cuts of lL~q z. Since the requirement of l~k x a can be rep-
resented by x xLa or x xU a , this can be formulated as the constraint of x b1 xLa 1  b1 xU
a,
where b1 = 0 or 1; similarly, ll~ y a can be formulated as the constraint of y b2 y La
1  b2 y U
a , where b2 = 0 or 1. Moreover, from the denition of ka and la in (9a) and (9b), x 2 ka
924 S.-P. Chen / Applied Mathematical Modelling 30 (2006) 920929

and y 2 la can be respectively replaced by x 2 xLa ; xU L U

a  and y 2 y a ; y a . Consequently, considering
both of these two cases above, the membership function lL~q can be constructed via nding the
lower bound Lq La and upper bound Lq U ~
a of the a-cuts of Lq , in that we set Lq a
L
 L L U  U U
min Lq a 1 ; Lq a 2 and Lq a max Lq a 1 ; Lq a 2 , respectively, where

Lq La 1 min xfyEK 2  2xEK2  yEKg=2yfy  xEKg

x;y2R ;xEK<y

s:t: x t1 xLa 1  t1 xU
a;

y La 6 y 6 y U
a;

t1 0 or 1; 10a
2
Lq La 2 min 2
xfyEK  2xEK  yEKg=2yfy  xEKg
x;y2R ;xEK<y

s:t: y t2 y La 1  t2 y U
a;

xLa 6 x 6 xU
a;

t2 0 or 1; 10b
2
Lq Ua 1 max xfyEK 2  2xEK  yEKg=2yfy  xEKg
x;y2R ;xEK<y

s:t: x t3 xLa 1  t3 xU
a;

y La 6 y 6 y U
a;

t3 0 or 1; 10c
Lq Ua 2 max xfyEK 2  2xEK2  yEKg=2yfy  xEKg
x;y2R ;xEK<y

s:t: y t4 y La 1  t4 y U
a;

xLa 6 x 6 xU
a;

t4 0 or 1; 10d
where xU L
a < EK  y a . At least one x or y must hit the boundary of their a-cuts to satisfy lL ~q z a
as required by (5). From the knowledge of calculus, a unique minimum and a unique maximum of
the objective function of Models (10a), (10b), (10c), or (10d) are assured, which shows that the
L U ~q can be found by solving these four
lower bound Lq a and upper bound Lq a of the a-cuts of L
models. In fact, these four models are MINLP with 01 variables. There are several eective and
ecient methods for solving these problems [21]. Moreover, they involve the systematic study of
how the optimal solutions change as xLa ; xU L U
a ; y a ; and y a vary over the interval a 2 [0, 1]; they fall
into the category of parametric programming [22].
The crisp interval Lq La ; Lq U
a  obtained by solving Models (10a), (10b), (10c), and (10d) rep-
resents the a-cut of L~q . An attractive feature of the a-cut approach is that all a-cuts form a nested
structure with respect to a [20]. According to Zadehs extension principle, L ~q dened in (5) is a
fuzzy number that possesses convexity [17,20]. Therefore, for two values a1 and a2 such that
L L U U L
0 < a2 < a1 6 1, we have Lq a1 P Lq a2 and Lq a1 6 Lq a2 ; in other words, Lq a is nondecreas-
U
ing with respect to a and Lq a is nonincreasing with respect to a. This property assures the
 S.-P. Chen / Applied Mathematical Modelling 30 (2006) 920929 925

convexity of L ~q . Consequently, the membership function lL~ z can be obtained from the solutions
q
of Models (10a), (10b), (10c), and (10d).
If both Lq La and Lq U a are invertible with respect to a, then a left shape function LS z
L 1 U 1
Lq a  and a right shape function RS z Lq a  can be obtained. From LS(z) and RS(z),
the membership function lL~q is constructed as
8 L L
< LS z; Lq a0 6 z 6 Lq a1 ;
>
L U
lL~q z 1; Lq a1 6 z 6 Lq a1 ; 11
>
: U U
RS z; Lq a1 6 z 6 Lq a0 .
Otherwise, if the values of Lq La and Lq U
a cannot be obtained analytically, the numerical solu-
L U
tions for Lq a and Lq a at dierent possibility level a can be collected to approximate the shapes
nh i
o

of LS(z) and RS(z). That is, the set of intervals Lq La ; Lq U

a
a 2 0; 1 reveals the shape of lL~q ,
although the exact function is not known explicitly. Other membership functions of performance
measures can be derived in the similar manner.
Since the above performance measures are described by membership functions, they conserve
completely all of fuzziness of arrival rate and service rate. However, in practical point of view,
the management could prefer one crisp value for each performance measure rather than a fuzzy
set. To tackle this problem, in this paper we defuzzify the fuzzy value of performance measure into
a crisp one by one of the fuzzy number ranking methods, namely Yager ranking index [23]. Given
a convex fuzzy number ~t, the Yager ranking index based on area compensation is dened by
Z 1
1 L 
~
It ta tU
a da; 12
0 2
 
where tLa ; tU
a is the a-cut of ~t. It is a robust ranking technique that possesses the properties of
compensation, linearity, and additivity [24].

4. Numerical example

Consider a centralized parallel processing system in which jobs arrive at in batches. The prob-
ability mass function of the batch size random variable K is a geometrical distribution, which is
often studied in crisp bulk arrival queues [1,2,5], with expected value of 2; i.e., the probability mass
function is Pr(K = k) = 0.5 (1  0.5)k1, k = 1, 2, . . . Jobs arrive at this system in accordance with a
Poisson process, and the service times follow an exponential distribution. Both the group arrival
rate and service rate are trapezoidal fuzzy numbers represented by ~k 2; 3; 4; 5 and l ~
13; 14; 15; 16 per minute, respectively. The system manager wants to evaluate the performance
measures of the system such as the expected number of customers in the queue.
We have E[K] = 2 and it is easy to nd E[K2] = var[K] + {E[K]}2 = 6. Following (4), therefore
we have
xfyEK 2  2xEK2  yEKg 2x2x y
Lq .
2yfy  xEKg yy  2x
926 S.-P. Chen / Applied Mathematical Modelling 30 (2006) 920929

It is clear that in this example the steady-state condition q = 2x/y < 1 is satised, thus the
performance measures of interest can be constructed by using the proposed approach stated in
Section 3. Following (10a), (10b), (10c) and (10d), two pairs of MINLP models for deriving
the membership function of L~q can be formulated, whose solutions are as follows:
2a2 22a 40
Lq La 13a
3a2  20a 64
and
U 2a2  28a 115
Lq a . 13b
3a2 14a 13
The inverse functions of Lq La and Lq U ~q z as
a exist, which give the membership function lL
8  p
>
> 2 15z119 z2 6z1
>
< 3z2
; 5=12 6 z 6 14=15;
lL~q z 1; 14=15 6 z 6 44=21;
>
> p
>
: 21z2818 z2 6z1
3z2
; 44=21 6 z 6 230=39.
Fig. 1 depicts the shape of lL~q z. The membership functions for the expected system length lL~ ,
the expected queue waiting time lW~ q , and the expected system waiting time lW~ can be derived sim-
ilarly as follows:
8 12z8
< 3z4 ; 2=3 6 z 6 4=3;
>
lL~ z 1; 4=3 6 z 6 8=3;
>
: 203z
3z4
; 8=3 6 z 6 20=3;
8 p
>
> 130z 324z2 180z1
>
< 3z
; 5=24 6 z 6 14=45;
lW~ q z 1; 14=45 6 z 6 11=21;
>
> 121zp
>
: 324z2 180z1
3z
; 11=21 6 z 6 46=39;

~q .
Fig. 1. The membership function of the expected queue length L
 S.-P. Chen / Applied Mathematical Modelling 30 (2006) 920929 927

Table 1
The a-cuts of the performance measures at 11 a values
a Lq La ; Lq U
a  LLa ; LU
a W q La ; W q U
a  W La ; W U
a
0.0 0.4167 5.8974 0.6667 6.6667 0.2083 1.1795 0.3333 1.3333
0.1 0.4538 5.1913 0.7179 5.9394 0.2161 1.0595 0.3419 1.2121
0.2 0.4934 4.6061 0.7719 5.3333 0.2243 0.9596 0.3509 1.1111
0.3 0.5358 4.1138 0.8288 4.8205 0.2330 0.8753 0.3604 1.0256
0.4 0.5812 3.6944 0.8889 4.3810 0.2422 0.8031 0.3704 0.9524
0.5 0.6298 3.3333 0.9524 4.0000 0.2519 0.7407 0.3810 0.8889
0.6 0.6819 3.0196 1.0196 3.6667 0.2623 0.6863 0.3922 0.8333
0.7 0.7380 2.7448 1.0909 3.3726 0.2733 0.6383 0.4040 0.7843
0.8 0.7982 2.5024 1.1667 3.1111 0.2851 0.5958 0.4167 0.7407
0.9 0.8632 2.2873 1.2473 2.8772 0.2977 0.5579 0.4301 0.7018
1.0 0.9333 2.0952 1.3333 2.6667 0.3111 0.5238 0.4444 0.6667

and
8 43z1
>
< 3z ; 1=3 6 z 6 4=9;
lW~ z 1; 4=9 6 z 6 2=3;
>
: 43z
3z
; 2=3 6 z 6 4=3.
Applying the Yager ranking index method stated in (12), the above fuzzy performance mea-
sures can be transformed into crisp ones for practical use. For example, considering (13a) and
~q is
(13b), the Yager ranking index of the L
Z 1  2 
~ 2a 22a 40 2a2  28a 115
ILq 0.5 da  2.09435.
0 3a2  20a 64 3a2 14a 13
Table 1 lists the a-cuts of these four performance measures at 11 distinct a values: 0, 0.1,
0.2, . . . , 1.0. These a-cuts represent the possibility that these four performance measures will ap-
pear in the associated range. Specially, the a = 0 cut show the range that these four performance
measures could appear and the a = 1 cut shows these four performance measures that are most
likely to be. For example, while these four performance measures are fuzzy, the most likely value
of the expected queue length L ~q falls between 0.9333 and 2.0952, and its value is impossible to fall
outside the range of 0.4167 and 5.8974; it is denitely possible that the expected system waiting
time W ~ falls between 0.4444 min (or 26.7 s) and 0.6667 min (about 40.0 s) approximately, and
it will never fall below 0.3333 min (about 20 s) or exceed 1.3333 min (about 80 s) approximately.
The above information will be very useful for designing a queueing system.

5. Conclusion

Bulk arrival queueing models have wide applications in operations and service mechanism for
evaluating system performance. In fuzzy environments, few journal papers for analyzing bulk ar-
rival queues have been published. This paper develops a method to nd the membership function
of the system performance measure when the arrival rate and service rate are fuzzy numbers, and
928 S.-P. Chen / Applied Mathematical Modelling 30 (2006) 920929

the batch sizes vary. The idea is based on Zadehs extension principle to transform the bulk arrival
fuzzy queue to a family of crisp bulk arrival queues that can be described by two pairs of MINLP
models. Then from the obtained solutions, the a-cuts of the membership functions of the perfor-
mance measures can be found. Since the performance measure is expressed by the membership
function rather than by a crisp value, it maintains the fuzziness of input information, and the re-
sults can be used to represent the fuzzy system more accurately.
In numerical example section, the arrival rate and service rate are assumed to be trapezoidal
fuzzy numbers. Clearly, the proposed approach is not conned to trapezoidal fuzzy numbers.
Other cases with arrival rate and service rate being convex fuzzy sets such as LR-fuzzy numbers
are also applicable. Furthermore, in this study the batch size is only considered to be a random
variable. A referee points out that it is a meaningful contribution to study the case of fuzzy batch
sizes, which will be a more accurate practice. It seems that the proposed approach can be extended
to the case of fuzzy batch sizes which deserves further consideration.

Acknowledgements

This research is supported by the National Science Council of Taiwan (Republic of China)
under Contract NSC92-2416-H-194-010. The author is grateful to the assistance of Miss Yi-Ju
Hsueh (a graduate student at Department of Speech Communication, Shih Hsin University,
Taiwan) in collecting data.

References

[1] M.L. Chaudhry, J.G.C. Templeton, A First Course in Bulk Oueues, John Wiley & Sons, New York, 1983.
[2] D. Gross, C.M. Harris, Fundamentals of Queueing Theory, third ed., Wiley, New York, 1998.
[3] J.C. Ke, Bi-level control for batch arrival queues with an early setup and un-reliable server, Appl. Math. Modell. 28
(2004) 469485.
[4] R. Sharma, G.C. Sharma, A queueing model with the bulk arrival rate depending upon the nature of service, Appl.
Math. Modell. 20 (1996) 470475.
[5] G. Choudhury, An MX/G/1 queueing system with a setup period and a vacation period, Queueing Syst. 36 (2000)
2338.
[6] G. Choudhury, A batch arrival queue with a vacation time under single vacation policy, Comput. Oper. Res. 29
(2002) 19411955.
[7] H.W. Lee, J.G. Park, B.Y. Kim, S.H. Yoon, B.Y. Ahn, N.I. Park, Queue length and waiting time analysis of a
batch arrival queue with bilevel control, Comput. Oper. Res. 25 (1998) 191205.
[8] J.B. Jo, Y. Tsujimura, M. Gen, G. Yamazaki, Performance evaluation of network models based on fuzzy queuing
system, Jpn. J. Fuzzy Theor. Syst. 8 (1996) 393408.
[9] J.J. Buckley, Elementary queueing theory based on possibility theory, Fuzzy Set Syst. 37 (1990) 4352.
[10] J.J. Buckley, T. Feuring, Y. Hayashi, Fuzzy queueing theory revisited, Int. J. Uncertain. Fuzz. 9 (2001) 527537.
[11] S. Chanas, M. Nowakowski, Single value simulation of fuzzy variable, Fuzzy Sets Syst. 21 (1988) 4357.
[12] R.J. Li, E.S. Lee, Analysis of fuzzy queues, Comput. Math. Appl. 17 (1989) 11431147.
[13] D.S. Negi, E.S. Lee, Analysis and simulation of fuzzy queue, Fuzzy Sets Syst. 46 (1992) 321330.
[14] H.M. Prade, An outline of fuzzy or possibilistic models for queueing systems, in: P.P. Wang, S.K. Chang (Eds.),
Fuzzy Sets, Plenum Press, New York, 1980, pp. 147153.
[15] R.R. Yager, A characterization of the extension principle, Fuzzy Sets Syst. 18 (1986) 205217.
 S.-P. Chen / Applied Mathematical Modelling 30 (2006) 920929 929

[16] L.A. Zadeh, Fuzzy sets as a basis for a theory of possibility, Fuzzy Sets Syst. 1 (1978) 328.
[17] H.J. Zimmermann, Fuzzy Set Theory and Its Applications, fourth ed., Kluwer Academic, Boston, 2001.
[18] R.E. Stanford, The set of limiting distributions for a Markov chain with fuzzy transition probabilities, Fuzzy Sets
Syst. 7 (1982) 7178.
[19] C. Kao, C.C. Li, S.P. Chen, Parametric programming to the analysis of fuzzy queues, Fuzzy Sets Syst. 107 (1999)
93100.
[20] A. KaufmannIntroduction to the Theory of Fuzzy Subsets, vol. 1, Academic Press, New York, 1975.
[21] F.S Hillier, G.J. Lieberman, Introduction to Operations Research, seventh ed., McGraw-Hill, Singapore, 2001.
[22] T. Gal, Postoptimal Analysis, Parametric Programming, and Related Topics, McGraw-Hill, New York, 1979.
[23] R.R. Yager, A procedure for ordering fuzzy subsets of the unit interval, Inform. Sci. 24 (1981) 143161.
[24] P. Fortemps, M. Roubens, Ranking and defuzzication methods based on area compensation, Fuzzy Sets Syst. 82
(1996) 319330.