.

FUIZY
sets and systems
ELSEVIER Fuzzy Sets and Systems 10711999) 93-100
w w w . e l sev ier.com/Iocate/£',s

Parametric programming to the analysis of fuzzy queues

Chiang Kao a'*, Chang-Chung Li b, Shih-Pin Chen b
"Graduate School of Industrial Management, National Cheng Kung University, Tainan, Taiwan, ROC
bDepartment of bldustrial Engineering, National Tsing Hua UniversiO,, Hsinchu, Taiwan, ROC
Received June 1997; received in revised form August 1997

Abstract

This paper proposes a general procedure to construct the membership functions of the performance measures in
queueing systems when the interarrival time and service time are fuzzy numbers. The basic idea is to reduce a fuzzy queue
into a family of crisp queues by applying the ;,.-cut approach. A pair of parametric programs is formulated to describe that
family of crisp queues, via which the membership functions of the performance measures are derived. To demonstrate the
validity of the proposed procedure, four fuzzy queues, namely, M.F."I, I:'/M. I, F:I ~ I, and FM, F M / I . are exemplified.
The discussion of this paper is confined to systems with one and two fuzzy variables: nevertheless, the procedure can be
extended to systems with more than two fuzzy variables. ,' 1999 Elsevier Science B.V. All rights reserved

Ko,words." Queueing theory; Parametric programming; Membership functions

I. Introduction usual crisp queues are extended to fuzzy queues [7],

queueing models would have even wider applica-
Queueing models have wide applications in ser- tions.
vice organizations as well as manufacturing firms, Li and Lee [5] have derived analytical results for
in that various types of customers are serviced by two fuzzy queueing systems based on Zadeh's ex-
various types of servers according to specific queue tension principle [9,10]. the possibility concept,
disciplines [3]. Within the context of traditional and fuzzy Markov chains [8]. However, as c,mt-
queueing theory, the interarrival times and service mented by Negi and Lee [6]. their approach is
times arc required to follow certain distributions. fairly complicated and is generally unsuitable for
However, in many practical applications the arrival computational purposes. Furthermore, for other
pattern and service pattern are more suitably de- more cornplicated qucucing systems, it i'~ hardly
scribed by linguistic terms such :is fast, slow. or possible to derive analytical results. Ncgi and l,cc
moderate, rather than by probability distributions. [611. therefore, propose Iwo approaches, the ~-cut
Restated, lhe interarrival times and servtce times and 1wo-variablc simulation [I], t o ;.tnalyzc i'u~,zy
arc more possibilistic than probabilistic, if the queues, l,!nfortunalcly, lheir approac|l on[; pro-
rides possible numbers rather than intervals: izt
other ~ords. the membership functions of the per-
*( "Ol i'0Spollding ttHlhol'. formance ntcasures arc z~,ot completely described.

0!65-0114.99,~ - sec i'r*~J~t matter ( 1999 Flsexicr £clencc B.\. Ai; rich ly, rcsci "~,.'d
PII: S(II 6 5 - D I 1 4 1 9 7 1 0 0 2 9 5 -t)
94 C. Kao et aL /Fuzzy Sets and Systems 107 (1999) 93-100

Aiming at the goal of deriving the membership confidence and derive different performance
functions of the performance measures for fuzzy measures analytically for different possibility level
queues, this paper adopts the a-cut approach to e. For complicated queueing systems, analytical
decompose a fuzzy queue to a family of crisp derivation is replaced by numerical simulation. In
queues. As the e value varies, the parametric pro- this paper we use the parametric programming
gramming technique is applied to describe that technique to formulate the problem. From the solu-
family of crisp queues. The solutions from the para- tions, the membership functions of interests can be
metric programs derive the membership functions constructed. We discuss two models, one for the
of the performance measures. To demonstrate the cases of one fuzzy variable and the other for that of
validity of the proposed approach, four typical two fuzzy variables.
fuzzy queues, namely, M/F/1, F / M / I , F/F/1, and
F M / F M / 1 , where F denotes fuzzy time and F M 2.1. One fuzzy variable
denotes fuzzified exponential time, are exemplified,
and advantages over previous studies are discussed. Since the discussion that either the interarrival
time is fuzzy or the service time is fuzzy is similar,
we assume that the interarrival time is a fuzzy
2. Problem formulation number and the service time follows a crisp prob-
ability distribution.
Consider a general queueing system with one The e cut of ~ defined in (2a) is a crisp
server. The interarrival time ,~ and service time interval which can be expressed in the form
are approximately known and are represented by [min.a~x{all~(a) >>,e}, max.a~x{all~(a) >~ e}].
the following fuzzy sets: This interval indicates where the constant interar-
rival time lies at possibility e. By the convexity of
)[ = {(a, I~a(a))laeX}, (la) a fuzzy number, the bounds of the interval are
functions of a and can be obtained as IA~)=
,g = {(s, #g(s))]s ~ Y}, (lb) m i n . / ~ 1(a) and UA~,) = max. p j 1(e), respectively.
Let p(a, s) denote the system performance measure
where X and Y are the crisp universal sets of the
of interest. When ~ is fuzzy, p(,4, s) is also a fuzzy
interarrival time and service time, and ~i(a) and
number. According to Zadeh's extension principle,
l~g(s) are the corresponding membership functions.
the membership function of p(g, s) is:
The a-cuts or a-level sets of 2[ and g are:
/~p~,,)(z) = sup{#~i(a)[ z = p(a, s)} (3)
A(a) = {a ~ X I ~,i(a) ~> e}, (2a) aEX

S(e) = {s e Y I/~(s)/> a}. (2b) which is also parametered by e. Consequently, we

can utilize its a-cut to construct its membership
Note that A(e) and S(a) are crisp sets. Using a-cuts, function.
the interarrival time and service time can be repre- To find the bounds of the intervals for p(A, s) at
sented by different levels of confidence intervals possibility level e, a pair of mathematical programs
[11]. Consequently, a fuzzy queue can be reduced parametered by e can be formulated:
to a family of crisp queues with different a-level sets
{A(e)10 < a ~< 1} and {S(a)10 < e ~< 1}. These two lv(~) = min.p(a,s) s.t. IA~) <, a <. UA(~) (4a)
sets represent sets of movable boundaries, and they up~ = max.p(a,s) s.t. IA(~) <~ a <~ UA~). (4b)
form nested structures for expressing the relation-
ship between ordinary sets and fuzzy sets [4]. This pair of mathematical programs involve the
Suppose _~ and ~ are fuzzy numbers. Denoting systematic study of how the optimal solutions
their intervals of confidence as [IA(~),UA(a)] and change as the bounds la(~) and UA~,) vary over the
[ls(~),Us(j, the idea of Negi and Lee [6] is to interval e E (0, 1], they fall into the category of para-
assume a uniform distribution in the interval of metric programming [2].
 C. Kao et aL / Fuzzy Sets and Systems 107 (1999) 93-100 95

If both Ira) and urn) are invertible with respect to and for the latter case are:
a, then a left shape function L(z) = lp~,)
- and a right ¢t
shape function R ( z ) = upon) can be obtained, lp(,~ = min. p(a, s)
from which the membership function pp~i.~ is s.t. ls~ <~ s <~ Using, (7c)
constructed:
a • A(a),
L(z), Zl ~ z ~ z 2 , tt
urn) = max. p(a, s)
].lp(~t.s)(Z ) = l, Z 2 ~ Z ~ Z3, (5)
s.t. ls~,) <~ s <~ Us~), (7d)
R(z), z3 <~ z <<.z4,
a • A(a).
where z ~ < z 2 ~ z 3 ~ z , and L ( z O = R ( z , ) = O .
Otherwise, since an attractive feature of the a-cut According to the definition of A(a) and S(a) in (2),
approach is that all a-cuts form a nested structure a • A ( a ) and s•S(a) can be replaced by a •
with respect to a [11]; that is, given cq, az •(0, 1] [Ia ~), UA~ ] and s • [ls~,), Us~,~]. Therefore, (7a) and
w i t h a 1 > a2, we have [lpt~,l,Up(~,)] ~ [lpt~2),Up(~2)]. (7c), and (7b) and (7d) are the same, respectively,
The set of intervals {[lvt,), uvt,)] I a • (0, 1]} reveals which can be rewritten as:
the shape of the membership function, although the
exact function form is not known explicitly. lp~,~ = min. p(a, s)

s.t. IA~a) <~ a ~ UA(~), (8a)

2.2. Two fuzzy variables
lsl~ <~ s <~ Us~,
When both the interarrival time and service time
upt~ = max. p(a,s)
are fuzzy numbers, based on Zadeh's extension
principle, the membership function of the perfor- s.t. IA~O <. a <~ UA(~t), (8b)
mance measure p(/~, g) is defined as:
Is~l <~s <~ Us~.
ppl~,g~(z) = sup min.{#a(a),pg(s)lz = p(a,s)}.
aEX.s~Y F r o m { [ l p ~ ) , u r ~ J l a • ( 0 , 1]}, the membership
(6) function ~,~i.~j can be constructed following the
same procedure discussed in Section 2.1.
One approach to construct the membership func-
tion ~tp{j,g) is to derive the a cuts of #poi.g ). Accord-
ing to (6), we need either pa(a) = a and/~(s) >~ a or
3. Some fuzzy queues
V~i(a) >/a a n d / ~ ( s ) = a such that z = p(a, s) to sat-
isfy/~i.g)(z) = a. This can be accomplished via the
Several fuzzy queues have been studied in differ-
parametric p r o g r a m m i n g technique. F o r the for-
ent extent in the literature [5,6]. In this paper we
mer case, the corresponding parametric programs investigate four typical ones; namely, M/F/1,
for finding the lower and upper bounds of the a-cut
F/M/1, F/F/1, and F M / F M / 1 . All four queues
of p~aA~(z) are: adopt a first-come-first-served discipline and con-
lpl, ) = min. p(a, s) sider an infinite source population. The first system
has an exponentially distributed interarrival time
s.t. IA~,) <~ a <~ UA~,), (7a) with mean ~ and a fuzzy service time ~'. The second
s • S(a), system, on the contrary, has a fuzzy interarrival
time A and an exponentially distributed service
u~,~,) = max. p(a, s) time with mean a. These two systems have only one
fuzzy variable. For the third queue, both the in-
s.t. Ial~) ~ a <~ uat, ~, (7b)
terarrival time ~ and the service time g are fuzzy. In
s • S(a), the fourth queue, both the interarrival time and the
96 C Kao et al. l:uz2v Sets and 3)'Meres 107 (1999) 9.!-100

service time follow exponential distributions with the ~ cut of ~q is [min. p.( 1(~.), max. l¢,~ 1(~)]
rates ,;-. and fi, respectively. ~hich are fuzz) wtri- = [3:z. -4- 5, 16 - 4~.]. To lind tile m e m b e r s h i p func-
ables rather than crisp values. These four queues tion for p [ , , ~ t = / 7 , , two parametric p r o g r a m s
are very representative in fuzzy queues. Therefore, based on (4) are formulated:
to gain some insight into fuzzy queues, it suffices to
study these four systems. lu~ I = rain. s/20 + s2/(800 - 40s)

s.t. 3x + 5 ~< s ~< 1 6 - 4:~, (12a~

3.1. The M/F~1 queue
uL,,, = max. s/20 + s-/(800 - 40s)
F o r the M/F/1 queueing system. Li and Lee [5]
s.t. 3y. + 5 ~ s ~< 16 - 47.. (12b)
have derived the m e m b e r s h i p ftmctions for the ex-
pected n u m b e r of persons in the system /7, and the Intuitively, the shortest system length It,~,~ occurs
expected waiting time in the system if' based on (3). when the constant service time is at its lower bound
In our notation, they are: and the longest system length ut,~,~ occurs when the
l,£(z)= sup r ,, i constant service time is at its upper bound. As
a matter of fact, the objective function in (12t has
a positive slope for 0 ~ < s < r c = 2 0 . F r o m the
z = ['s(2 - s/n)] / ['2x(1 - s/n)]], (9a)
knowledge of calculus, L(~) attains its extrema at
the bounds. T h a t is,
p~(z)= sup (m(s) l
s E Y , ~ < r¢ 1L(,~ = (175 + 90~ -- 9e2)/(600 -- 120~), (13a)
z = I-s(2 - s/n)]/['2(1 - s/n)]}. (9b) u u , , = (24 -- 2~ -- ~2)/(10 + 10~). (13b)
These two m e m b e r s h i p functions are not in the T a k i n g the inverse of (13) derives the left and right
usual form and are very difficult to imagine their shape functions o f / ~ . Accordingly,
shapes. The a p p r o a c h of this p a p e r a d o p t s the
cuts to reduce the M/F/1 queue to a series of [(20z + 15) - 20(z 2 + 1)~/2]/3,
M/D~1 queues which are special cases of the M/G/1 7/24 ~< z ~< 8/15,
queue. According to the Pollaczek-Khintchine for-
m u l a [3], the expected system length is p£(z)= 1, 8/15 <~z ~< 21/20, (14)
--(5Z+ I ) + 5 ( Z 2 + 1) 1/2,
L = s/n + s2/['2rt2(1 -- s/g)] (10)
21/20 ~< z ~< 12/5.
which is finite provided s < n. A d o p t i n g L as the
p e r f o r m a n c e measure, i.e., p(a, s) = L, the p a r a m e t - In (9a) if s is represented in terms of z, we have
ric p r o g r a m m i n g technique described in the s = 20(z + 1) - 20(z 2 + 1) 1/2. Substituting s into
previous section can be applied to derive the m e m - (11), a m e m b e r s h i p function of exactly the same
bership function Ft£ in the usual form. form as (14) is derived. This confirms that the ap-
p r o a c h p r o p o s e d in this p a p e r is correct.
E x a m p l e 1. Consider the M/F/1 queue studied by By using simulation, Negi and Lee ['6] are only
Negi and Lee [6], where the interarrival time fol- able to find the possible number, rather than the
lows an exponential distribution with m e a n n = 20 b o u n d s / z , , , and ULI,~ of persons in the system for
minutes and the service time is fuzzy with each possibility level ~. O u r a p p r o a c h is obviously
= [5,8, 12, 16]. Thus, m o r e informative.

(s- 5)/3, 5 ~< s ~< 8, 3.2. The F/M/1 queue

p,~(s)= 1, 8~<s~< 12, (11)
F o r the F/M/I queue, when the interarrival time
(16--s)/4, 12~<s~<16, is fixed at a constant, it becomes a D/M/l queue,
 C. Kao et al. ,; Fuzzy Sets and Systems 107 (1999) 93-100 97

which is a special case of the G I / M / 1 queue. Ac- are:

cordingly, the existing formulae [3] can be utilized
to calculate W and L: lu~ ~ = min. 2/[a(1 - fl)]

w = s,'(1 - / ~ ) , L = s/[(.(l -/~)], (15) s.t. 4 + 2 ~ < a ~<8--~, (17a)

where fl is the solution of fl = e-a(l-~)/2,

f l = e -''11-~''~, fl~:l. (16) Uz~ = max. 2/[a(1 - fl)]

It has been shown [3] that if the m e a n service time s.t. 4 + 2 ~ ~<a-N< 8 - ~, (17b)
is smaller than the mean interarrival time, then
there exists a unique fle (0, I ) w h i c h satisfies (16). At f l : e-a(I -ill,'2
each possibility level ~, the fuzzy interarrival time is
a crisp value lying in the range of [Ia(~, lt4(~)], which Function f(/:0 = e -''(1 -t~)/2 has a positive slope for
reduces the F / M / I queue to a DIM~1 queue. The all fl > 0, which intersects function g(fl) = fl at two
p e r f o r m a n c e measures in (15) can then be sub- points as illustrated in Fig. 1. O n e point always
stituted into the p a r a m e t r i c p r o g r a m s in 14) to occurs at/3 = 1 and the other lies in (0, 1) [3]. As
derive the corresponding m e m b e r s h i p functions. a increases from its lower b o u n d (4 + 2~t) to its
upper b o u n d ( 8 - ~t), the intercept of f ( f l ) =
Example 2. Consider an F / M / 1 queue, where the e -a(x-#)/2 drops from e -t4+2~0/2 to e -(8-')/2, and
interarrival time is a trapezoidal fuzzy n u m b e r the intersection of g(fl) = fl and f(fl) = e -act-#)/2
= [4, 6, 7, 8] and the service time is exponentially decreases from fit to /J,. Obviously, the objective
distributed with m e a n a = 2. It is easy to find function 2/[a(1 - fl)] attains its m i n i m u m when a
[lace), UAC0t)'] = [min. p j 1(~), max. p j 1(~)] = [4 + 2~, reaches its upper b o u n d and its m a x i m u m when a
8 - ~]. The two p a r a m e t r i c p r o g r a m s to be solved reaches its lower bound.

g(fl) = fl
1

f ( f l ) = e °'"P~"

-I4*2a~
e
e-,,-o~ 3
ill, t~, 1
Fig. 1. The solution of fl = e -o~t -0}..2
98 C. Kao et al. / Fuzzy Sets and Systems 107 (1999) 93-100

Due to the complicated form of fl = e - a t 1 -fl)/2, Table 1

it is impossible to represent the optimal solution The ~-cuts of pr. in Example 2 at 11 ct values
lu, } and uu, ) in terms of ~. Consequently, a closed-
form membership function for/~ is not obtainable.
However, as the number of ~ values considered 1.0 0.29578 0.35443
increases, we will be able to see the shape of/~£. 0.9 0.29106 0.36950
Fig. 2 depicts a rough membership function con- 0.8 0.28651 0.38611
structed from one hundred ct values. The "rough" 0.7 0.28212 0.40451
0.6 0.27787 0.42504
function turns out rather fine, just like a continuous 0.5 0.27376 0.44811
function. Table 1 presents the at-cuts of kt£ at 11 0.4 0.26978 0.47424
distinct ~ values. 0.3 0.26593 0.50414
0.2 0.26219 0.53873
0.1 0.25857 0.57926
3.3. The F/F/1 queue 0.0 0.25506 0.62750

The F/F/1 queue involves two fuzzy variables,

which is a generalization of the M/F/1 queue and
the F/M/1 queue. Hence, it is more complicated
than the preceding two fuzzy queues. F r o m the [IA(~),uA(j and [Ist~,D,Us(~,l~]. Thus, the F/F/1
knowledge of crisp queueing theory, when both queue reduces to a D/D~1 queue, upon which the
the interarrival time and service time are constant, parametric programs of (8) can be applied to con-
the average system length is: struct the membership function for L.

L={s/a if s~<a, (18) Example 3. Consider the F/F/1 queue studied

otherwise. by Negi and Lee [6], where the fuzzy interarrival
time and fuzzy service time are described by
Based on the s-cut approach, when the fuzzy num- = [5, 6, 7, 9] and g = [4, 4.5, 5.25, 7.5]. The cor-
bers A and g are at possibility levels ct and ~', responding s-cuts of these two fuzzy numbers are
respectively, they become crisp lying in the range of [ l a ~ , ua(~)] = [5 + ~, 9 - 2~] and [/s(~),Us(~)] =

#T.

0.8

0.6

0.4

0.2

system length
0.25506 0.62750
Fig. 2. The membership function for E in Example 2 constructed from one hundred ct values.
 C. Kao et al. / Fuzzy Sets and Systems 107 (1999) 93-100 99

[(8 + ~)/2, (30 - 9~)/4]. Following (8), the para- From (20) it is not difficult to obtain the member-
metric programs to be solved are: ship function /~L by applying inverse transforma-
tion:
lL(~) = min. s/a
(18z -- 8)/(4z + 1), 4/9 ~< z ~< 9/14,
s.t. 5 + e ~< a ~< 9 - 2e,

(8 + e)/2 ~< s ~< (30 - 9e)/4,

UL(~) = max. s/a

(19a)

/tL(z) =
I1,
(30 -- 20z)/(9 + 4z),
l [o, lO/13),
9/14 ~ z ~< 21/24,
21/24~<z~< 1,
Z= OC.

(21)
s.t. 5 + ~ < a ~ < 9 - 2 ~ , (19b)
The approach of Negi and Lee [6] is to apply
(8 + e)/2 ~< s ~< (30 - 9~)/4. a single-value simulation proposed by Chanas and
Nowakowski [1]; therefore, only crisp solutions are
To minimize s/a, the smallest value of s and the derived.
largest value of a are desired; therefore, IL(~)=
[(8 + ~ ) / 2 ] / ( 9 - 2~). In maximizing s/a, on the
3.4. The F M / F M / 1 queue
contrary, the largest value of s and the smallest
value of a are desired, which results in
Li and Lee [5] have studied a one-server queue-
ULt~) = [(30 -- 9~)/4]/(5 + ~). However, referring to
ing system denoted by F M / F M / 1 , in which arrivals
Fig. 3, when ~ drops below 10/13, the largest service
and departures are both Poisson processes with
time is greater than the smallest interarrival time.
fuzzy rates ~ and ~. The membership functions for
According to (18), the average system length
E and lg~ they derive are:
approaches infinity. Combining these two parts
derives the a-cut of/~L as follows: #L(z) = sup
x, yeR+,x/y< 1
[(8 + ~)/(18 - 4~), (30 - 9~)/(20 + 4~)],
min. {/~.(x),/t~(y)l z = x/(y - x)}, (22a)
10/13<~<1,
L(~)= t p~(z) = sup
[(8 + ~)/(18 - 4~), ~), 0 <~ ~ < 10/13. x,y~R+,x/y< 1

(20) min. {pi(x), pa(y) l z = 1/(y - x)}, (22b)

w h e r e / ~ and/~i are the membership functions of

X and/~. Although the membership functions in (22)
are theoretically correct, they are not in the usual
'i form for practical use. By applying the concept of
a-cut, the F M / F M / 1 queue reduces to a family of

\ M / M / 1 queues. From the knowledge of queueing

theory, we have W = 1/(p - 2) and L = 2/(# - 2),
where 2 and p are the crisp arrival and service rates.
The procedure described in Section 2.2 can then be
applied to find the membership functions of the
performance measures.

Time Example 4. Consider an F M / F M / 1 queue, where

4 5 7.5 9 both the arrival rate and service rate are fuzzy
Fig. 3. B e l o w ~ = 10/13 the i n t e r a r r i v a l t i m e a n d service t i m e numbers represented by ~[ = [4,4.5,5.5,6.5] and
overlap. /2 = [8,9,11,12]. From these we can derive the
100 C. Kao et al. /Fuzzy Sets and Systems 107 (1999) 93-100

interval of confidence at possibility level ~ as reduce a fuzzy queue into a family of crisp queues
[(8 + ~)/2, (13 - 2~)/2] and [8 + z~, 12 - 7]. Fol- which can be described by a pair of parametric
lowing (8), the parametric p r o g r a m s to derive the p r o g r a m s to find the ~.-cuts of the membership
membership function for/~ are: functions of the performance measures. With the
7-cuts, the corresponding membership function is
IL~ = min. x / ( y - x)
derived consequently. Since the performance
s.t. (8 + ~)/2 ~< x ~< (13 - 2ct)/2, (23a) measures are expressed by membership functions
rather than possible values as derived in other
8+~<y~<12-~,
studies, they provide more information.
uL~,~ = max. x / ( y - x) In this paper, four typical fuzzy queues arc inves-
tigated. Obviously, the proposed procedure is not
s.t. (8 + ~)/2 ~< x ~< (13 - 2~)/2, (23b)
confined to those four queues. O t h e r fuzzy queues
8+~<),~< 12-~. such as M / F M / i , F M / M : I , finite capacity, or bulk
service are also applicable. Furthermore, although
W h e n x reaches its lower b o u n d and y reaches its
the procedure is devised for one and two fuzzy
upper bound, x / ( y - x) attains its minimum. C o n -
variables, systems with more than two fuzzy vari-
sequently, the optimal solution for (23a) is
ables, for instance, k servers each has a fuzzy service
It.~) = (8 + ~)/(16 - 3~). O n the contrary, to maxi-
time, the a p p r o a c h proposed in this paper can be
mize x / ( y - x) it is desired that x increases to its
extended by applying Zadeh's extension principle
upper b o u n d and y decreases to its lower bound. In
to derive the membership functions of interest.
this case (23b) has the optimal solution uLt,~ =
(13 - 2~t)/(3 + 4~). The inverse functions of
ILt,) and ULt,~ exist, which give the membership
function p£(z) as: References

[I] S. Chanas, M. Nowakowski, Single value simulation of

8(2z - 1 ) / ( 3 z + 1), 1/2 ~< z <~ 9/13,
fuzzy variable, Fuzzy Sets and Systems 21 (1988) 43-57.
/tr.(z) = 1, 9/13 ~< z ~< 11/7, I-2] T. Gal, Postoptimal Analyses, Parametric Programming,
and Related Topics, McGraw-Hill, New York, 1979.
(13 -- 3z)/(4z + 2), 11/7 ~< z ~< 13/3. [3] D. Gross, C.M. Harris, Fundamentals of Queueing
Theory, Wiley, New York, 1985.
(24) [4] A. Kaufmann, Introduction to the Theory of Fuzzy Sub-
The membership function for if" can be derived sets, vol. 1, Academic Press, New York, 1975.
1-5] R.J. Li, E.S. Lee, Analysis of fuzzy queues, Comput. Math.
similarly. Appl. 17 (1989} 1143-1147.
[6] D.S. Negi, E.S. Lee, Analysis and simulation of fuzzy
queues, Fuzzy Sets and Systems 46 (1992) 321-330.
4. Conclusion [7] 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. 147-153.
F u z z y sets theory has been applied to some clas- [8] R.E. Stanford, The set of limiting distributions for
sical queueing systems to provide wider applica- a Markov chain with fuzzy transition probabilities, Fuzzy
tions in some previous studies [5,6i. W h e n the Sets and Systems 7 (1982) 71-78.
interarrival time and service time are fuzzy vari- [9] R.R. Yager, A characterization of the extension principle,
ables, according to Z a d e h ' s extension principle, the Fuzzy Sets and Systems 18 [1986) 205-217.
[I 0] L.A.Zadeh, Fuzzy sets as a basis for a theory of possibility,
performance measures such as the average system Fuzzy Sets and Systems 1 (1978) 3-28.
length, the average waiting time, etc., will be fuzzy 1-11] H.J. Zimmermann, Fuzzy Set Theory and Its Applications,
as well. This paper applies the concept of ~-cut to 2nd ed., Kluwer-Nijhoff, Boston, 1991.