Computing, Artificial Intelligence and Information Technology

Modeling uncertain activity duration by fuzzy number
and discrete-event simulation
Hong Zhang
a,
*
, C.M. Tam
b
, Heng Li
a
a
Department of Building and Real Estate, The Hong Kong Polytechnic University, Humg Hom, Kowloon, Hong Kong
b
Department of Building and Construction, City University of Hong Kong, Hong Kong
Received 20 March 2003; accepted 21 January 2004
Available online 24 March 2004
Abstract
This paper introduces how to incorporate fuzzy set theory and a fuzzy ranking measure with discrete-event simu-
lation in order to model uncertain activity duration in simulating a real-world system, especially when insufficient or no
sample data are available. Fuzzy numbers are used to describe uncertain activity durations, reflecting vagueness,
imprecision and subjectivity in the estimation of them. A fuzzy ranking measure is merged with an activity scanning
simulation algorithm for performing fuzzy simulation time advancement and event selection for simulation experi-
mentation. The uses of the fuzzy activity duration and the probability distribution-modeled duration are compared
through a series of simulation experiments. It is observed that the fuzzy simulation outputs are arrived at through only
one cycle of fuzzy discrete-event simulation, still they contain all the statistical information that are produced through
multiple cycles of simulation experiments when the probability distribution approach is adopted.
Ó 2004 Elsevier B.V. All rights reserved.
Keywords: Discrete-event simulation; Fuzzy sets theory; Fuzzy number; Fuzzy ranking; Uncertainty activity duration
1. Introduction
Discrete-event simulation is used to describe
and analyze the behavior of a system, ask what-if
questions about the real system, and aid analysis
and design of management policies for the con-
struction process. Developing a simulation model,
namely modeling, is key to discrete-event simula-
tion. Input modeling for defining the information
or parameters of a system being modeled strictly
determines the quality of the simulation results
(Maio et al., 2000). Inappropriate input leads to
misleading simulation outputs, and therefore
error-prone or sub-optimal planning and man-
agement decisions (Stephen, 1999). Difficulties in
input modeling, including the lack of confidence in
input information, have limited the use of discrete-
event simulation as a practical tool for construc-
tion (Fente et al., 2000).
Generally, modeling activity duration is ach-
ieved through importing observed data or proba-
bility distributions derived from sample data.
*
Corresponding author. Tel.: +852-27665795; fax: +852-
27645131.
E-mail addresses: bszhangh@polyu.edu.hk (H. Zhang),
bctam@cityu.edu.hk (C.M. Tam), bshengli@polyu.edu.hk
(H. Li).
0377-2217/$ - see front matter Ó 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2004.01.035
European Journal of Operational Research 164 (2005) 715–729
www.elsevier.com/locate/dsw
However, the use of real data in simulation is
time-consuming taken to collect large amounts of
field data for reproduction of the real operation
system (Law and Kelton, 1991). The definition of
probability distributions is considered to be
intractable; theoretically too complicated and
computationally too expensive (Gaul, 1981; Klein,
1986). The issues on how to define probability
distributions for activity duration through sample
data in the construction field have been widely
studied (AbouRizk and Halpin, 1992; Fente et al.,
2000; Maio et al., 2000; Wilson et al., 1982; Klein
and Baris, 1990; Touran, 1997; Weiler, 1965). It is
concluded that some real-life characteristics may
be left out if a small number of data are collected;
the goodness-of-fit between data and distributions
may be weak even when a large number of data are
available. Hence subjectivities in selecting distri-
butions and estimating related parameters are
unavoidable (Nasution, 1999; Fente et al., 2000;
Lu, 2002). The most difficult aspect of duration-
input modeling is gathering data of sufficient
quality, quantity, and variety (Stephen, 1998).
However, in some contexts it may be impossible or
infeasible to collect data due to, for example, the
uniqueness of construction activities and some-
times labor agreements that preclude collection of
performance data, while for other cases there may
be insufficient time or personnel to carry out
extensive data collection.
In respect to such situations, the estimation of
activity durations by experienced engineers or ex-
perts is often adopted. Therefore, more often than
not, there is subjectivity, vagueness or imprecision
in the estimated activity duration. Further, there
are some activity-duration uncertainties charac-
terizing the real-world construction process due to
factors such as changing weather conditions,
equipment properties, labor efficiency, supply
conditions of materials and coordination problems
among stakeholders, etc. Fuzzy set theory (Zadeh,
1965; Bellman and Zadeh, 1970) provides a good
mathematical methodology to describe and handle
the problem of vague and imprecise activity
durations.
Fuzzy set theory has been applied to network-
based planning techniques by some scholars such
as Duobis and Prade (1988), Nasution (1999),
Chanas and Kamburowski (1981), and Lootsma
(1989). In addition, fuzzy set theory has been ap-
plied to modeling uncertain production environ-
ments through continuous simulation (Dohnal,
1983; Fishwick, 1991; Negi and Les, 1992; Petrovic
et al., 1998; Southall and Wyatt, 1988). Fuzzy
logical control has been integrated with discrete-
event simulation to model flexible activity start
conditions and fuzzy-ruled computation of activity
duration varying with the amounts of involved
materials (Zhang et al., 2003). Examples of mod-
eling fuzzily estimated activity durations through
combining the fuzzy set theory with the discrete-
event simulation are scarce, especially for the
construction-oriented simulation.
The combination of fuzzy set theory with dis-
crete-event simulation to handle subjectivity,
vagueness or imprecision in estimating activity
duration so as to solve some problems in using
probability distributions is presented in this paper.
In addition, a fuzzy ranking measure and its
application to the control of fuzzy-time advance-
ment and event selection for the simulation exper-
iment of discrete-event simulation are described.
Illustrations on the fuzzy discrete-event simulation
and an example that compares the fuzzy discrete-
event simulation with the traditional probability
distribution-based simulation are also provided.
2. Description of activity duration with triangular
or trapezoidal fuzzy number
Although many types of fuzzy sets (Zadeh,
1965) have been used to describe uncertainties,
triangular and trapezoidal fuzzy sets (Fig. 1) are
very often used in the applications (e.g., fuzzy
controllers and managerial decision-making) be-
cause the parameters defining them can be easily
specified in linguistic terms (Yoon, 1996; Bojadziev
and Bojadziev, 1997). In addition, only triangular
or trapezoidal fuzzy sets suit the fuzzy ranking
measure (Tran and Duzkstein, 2002) that will be
introduced in Section 4. Therefore, triangular or
trapezoidal fuzzy sets are applied to describe
uncertain activity duration in the research. Fuzzy
numbers are the fuzzy sets that are normalized and
convex (Bojadziev and Bojadziev, 1997). Accord-
716 H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729
ingly triangular or trapezoidal fuzzy sets can be
called triangular or trapezoidal fuzzy numbers.
An uncertain activity duration can be described
by a fuzzy number defined by its membership
function l
D
ðxÞ, associated with numerical value
in the interval of [0,1], i.e. D ¼ fðx; l
D
ðxÞÞjx 2 D;
l
D
ðxÞ 2 ½0; 1Šg. When activity duration is described
using linguistic terms such as ‘‘most likely about
D
m
or between D
m1
and D
m2
, but definitely not less
than D
1
and not greater than Fig. 1 D
2
’’, the tri-
angular or trapezoidal fuzzy numbers (Fig. 1)
describing such an activity duration have the fol-
lowing membership functions:
l
D
ðxÞ ¼
xÀD
1
D
m1
ÀD
1
; D
1
6x 6D
m1
and D
1
6¼ D
m1
;
1; D
m1
6x 6D
m2
;
xÀD
2
D
m2
ÀD
2
; D
m2
6x 6D
2
and D
m2
6¼ D
2
;
0; otherwise:
8
>
>
>
<
>
>
>
:
ð1Þ
A trapezoidal fuzzy number can be reduced to a
triangular one when D
m1
¼ D
m2
. The triangular
and trapezoidal fuzzy numbers are defined by
three and four variables, respectively, as denoted
by (D
1
, D
m1
, D
2
) or (D
1
, D
m1
, D
m2
, D
2
). Therefore a
triangular fuzzy number (D
1
, D
m1
, D
2
) can be
written in the form of a trapezoidal fuzzy number
(D
1
, D
m1
, D
m1
, D
2
).
3. Simulation advancement problems with fuzzy
activity duration
Discrete-event simulation experimentation is
guided by a simulation strategy, which is generally
classified into three types: (a) activity scanning
(AS), (b) event scheduling (ES) and (c) process
interaction (PI) (Pidd, 1998). The AS is particu-
larly adept at modeling construction processes in
which activities are subject to complex activation
conditions (Martinez and Ioannou, 1999). The
three-phase strategy, an enhanced AS simulation
strategy, that requires classifying activities into
two groups, is generally used in construction-ori-
ented simulation systems such as CYCLONE
(Halpin, 1976). In our study, another improved AS
simulation strategy (Zhang et al., 2002) that does
not classify activities into two groups is merged
with the fuzzy number theory to develop the fuzzy
discrete-event simulation system.
This improved AS strategy advances simulation
experimentation through checking activities’ start
conditions (the availability of required resources
or logical dependencies, i.e. completion of the
preceding activities) and updating the simulation
time from now to next time of the end event that
will happen the earliest. Before advancing or
updating the simulation time, checking start con-
ditions and initiating start events are progressed
until no activities meet the conditions at the cur-
rent simulation time. The start time of an activity
should be the latest available time of the required
resources or logical dependencies. Fig. 2 illustrates
the time advancement and event management of
this AS strategy. As shown in Fig. 2, the simula-
tion time (m Now) will be successively updated to
the four end events that are due at times m Now1
(initialized as 0), m Now2, m Now3, and m Now4.
The numbers in the circles indicate the initiation
sequence of the start and end events.
D
µ
D
m1
D
1
D
2
X

D
µ
0

1

D

D
1
D
m1
0

1

D
m2
D
2
X

1-2 1-1

D

Fig. 1. Triangular and trapezoidal fuzzy numbers.
H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729 717
In this improved AS simulation strategy, there
are three parts in which operations on times are
needed.
• Initiation of a start event or activation of an
activity, where the start time of an activity is
determined; T
S
ðiÞ ¼ max
j¼1;...;J
fT
A
ði; jÞg, where
J is the total quantity of the entities (i.e. resources
and logical dependencies) required by activity i;
T
S
ðiÞ is the start time of activity i; T
A
ði; jÞ is the
available time of entity j at activity i.
• Determination of the due time of an end event,
which should equal the start time of the cur-
rent activity plus its activity duration; T
EE
ðkÞ ¼
T
S
ðiÞ þ DðiÞ, where T
EE
ðkÞ is the due time of
end event k, recorded in the end event list;
DðiÞ is the duration of activity i.
• Simulation advancement and end event selec-
tion, that is, the simulation time is updated
from the current time to the time of next one
or more end events that will happen the earliest
and the end events that are due at the updated
simulation time will be selected for initiation;
m Now ¼ min
k¼1;...;K
fT
EE
ðkÞg, where m Now is
the simulation time and K is the number of
end events recorded in the end event list.
When fuzzy numbers are used to represent
activity durations, all the times in the above
operations become fuzzy numbers. Hence the
above operations can be expressed in fuzzy arith-
metic operations as follows:
T
S
ðiÞ ¼ max
j¼1;...;J
fT
A
ði; jÞg; i ¼ 1; . . . ; I ð2Þ
T
EE
ðkÞ ¼ T
S
ðiÞ È DðiÞ; k ¼ 1; . . . ; K; i ¼ 1; . . . ; I;
ð3Þ
m Now ¼ min
k¼1;...;K
fT
EE
ðkÞg; ð4Þ
where I is the total number of activities. The
operation (3) is a fuzzy addition. If the fuzzy
duration of activity i is ðD
1
ðiÞ; D
m1
ðiÞ; D
m2
ðiÞ;
D
2
ðiÞÞ, the fuzzy occurring time of end event k is
ðT
EE
1
ðkÞ; T
EE
m1
ðkÞ; T
EE
m2
ðkÞ; T
EE
2
ðkÞÞ, and the fuzzy
start time of activity i is ðT
S
1
ðiÞ; T
S
m1
ðiÞ; T
S
m2
ðiÞ;
T
S
2
ðiÞÞ. Therefore, the fuzzy arithmetic operation
(3) has the following form:
T
EE
1
ðkÞ; T
EE
m1
ðkÞ; T
EE
m2
ðkÞ; T
EE
2
ðkÞ
À Á
¼ T
S
1
ðiÞ
À
þ D
1
ðiÞ; T
S
m1
ðiÞ þ D
m1
ðiÞ; T
S
m2
ðiÞ
þ D
m2
ðiÞ; T
S
2
ðiÞ þ D
2
ðiÞ
Á
;
i ¼ 1; . . . ; I; k ¼ 1; . . . ; K: ð5Þ
The fuzzy operations (2) and (4) require comparing
fuzzy numbers, i.e. the fuzzy number ranking
problem. The simplest ranking method is to con-
vert fuzzy numbers into a series of crisp numbers
and then compare them (Chen, 1985; Liou and
Wang, 1992). However, this crisp-ranking method
may lose much of the information that is purposely
kept throughout fuzzy operations (Bortolan and
Degani, 1985; Yufei, 1991). Of course, the fuzzy
ranking problems are more complicated, especially
when the ranked fuzzy numbers overlap or
intersect as shown in Fig. 3. For a fuzzy discrete-
event simulation that models complex construction
operations, the fuzzy simulation time, multiple
fuzzy times of end events and many fuzzy available
times of entities may encounter such overlapping
or intersection at a certain time interval.
4. Selection of a fuzzy ranking measure
Bortolan and Degani (1985) as well as Wang
and Kere (2001a,b) presented a comprehensive
survey of the available fuzzy number ranking
methods. Almost every method, however, has its
pitfalls in some respects, such as inconsistency with
human intuition, indiscrimination and difficulty of
interpretation, implying the non-existence of a
Start 1 Start 2 Start 3 Start 4
End 1 End 3 End 2
m_Now 1=0 m_Now 2 m_Now 3 m_Now 4
1 2 3 4 5 6 7
Simulation advancement (m_Now)
Fig. 2. Time advancement in the AS-based simulation strategy.
718 H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729
unique or best fuzzy number ranking method.
Suggested evaluation criteria (Bortolan and De-
gani, 1985; Chen and Hwang, 1992) to select a
reasonable ranking method include complexity,
robustness, flexibility, transitivity, and ease of
interpretation, which are detailed as follows:
• Complexity means the amount of computation
and the difficulties in implementing the ranking
method.
• Robustness refers to the ability of consistent
ranking for a diversity of cases, including the
ability to compare fuzzy and crisp numbers.
• Flexibility means the ability to compare differ-
ent shapes of fuzzy numbers (e.g. triangular
and trapezoidal) and allow participation of
decision makers in the process.
• Transitivity refers to the ability of giving a con-
sistent conclusion in the comparison of more
than two fuzzy numbers.
• Ease of interpretation is important to decision-
makers.
A fuzzy ranking measure (Tran and Duzkstein,
2002) possesses several advantages listed above
and can overcome problems inherent to the exist-
ing ranking methods, especially when used for
fuzzy discrete-event simulation. For instance, easy
implementation of this fuzzy ranking method en-
ables it to be integrated with the simulation algo-
rithm for discrete-event simulation. Its ability to
compare fuzzy numbers with crisp values is useful
because a discrete-event simulation generally sets
crisp initial simulation time (start point) or crisp
simulation time limit (end point). The flexibility of
ranking either triangular or trapezoidal fuzzy
numbers provides more choices to describe the
fuzzy activity duration. Keeping consistent results
among more than two fuzzy numbers is also
important because multiple fuzzy times may be
simultaneously ranked during simulation. Based
on the above descriptions, the fuzzy ranking
measure (Tran and Duzkstein, 2002) is chosen to
accomplish the fuzzy time ranking in simulation
advancement for fuzzy discrete-event simulation
if triangular or trapezoidal fuzzy numbers are
used to describe uncertain activity durations.
The fuzzy ranking measure (Tran and Duzk-
stein, 2002) is built based on comparison of from
fuzzy numbers (i.e., triangular and trapezoidal
fuzzy numbers) to some predetermined targets: the
crisp maximum (Max) and the crisp minimum
(Min). The idea is that a fuzzy number is ranked
first if its distance to the crisp maximum (D
max
) is
the smallest but its distance to the crisp minimum
(D
min
) is the greatest. Hence this fuzzy ranking
measure is also called fuzzy distance measure by
Tran and Duzkstein (2002). The Max and Min are
chosen as follows:
Min 6 inf
[
I
i¼1
sðAðiÞÞ
!
; ð6Þ
Max P sup
[
I
i¼1
sðAðiÞÞ
!
; ð7Þ
where sðAðiÞÞ is the support of fuzzy numbers AðiÞ
(i ¼ 1; . . . ; I) to be ranked. Based on the formulae
for the fuzzy ranking measure (Tran and Duzk-
stein, 2002), the formulae to compute D
max
and
D
min
for the trapezoidal fuzzy number A
(a
1
; a
2
; a
3
; a
4
) are as follows:
when the weighting function f ðaÞ ¼ a:
D
2
ðA; MÞ ¼
a
2
þ a
3
2

À M

2
þ
1
3
a
2
þ a
3
2

À M

 ½ða
4
À a
3
Þ À ða
2
À a
1
ފ þ
2
3
a
3
À a
2
2

2
þ
1
9
a
3
À a
2
2

½ða
4
À a
3
Þ þ ða
2
À a
1
ފ
þ
1
18
½ða
4
À a
3
Þ
2
þ ða
2
À a
1
Þ
2
Š
À
1
18
½ða
4
À a
3
Þða
2
À a
1
ފ; ð8Þ
Time 2 Time 3 Time 4 Time 5
0
Time
Time 1
1
Fig. 3. Overlapping or intersecting fuzzy numbers to be ranked.
H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729 719
when the weighting function f ðaÞ ¼ 1:
D
2
ðA; MÞ ¼
a
2
þ a
3
2

À M

2
þ
1
2
a
2
þ a
3
2

À M

 ½ða
4
À a
3
Þ À ða
2
À a
1
ފ þ
1
3
a
3
À a
2
2

2
þ
1
6
a
3
À a
2
2

½ða
4
À a
3
Þ þ ða
2
À a
1
ފ
þ
1
9
½ða
4
À a
3
Þ
2
þ ða
2
À a
1
Þ
2
Š
À
1
9
½ða
4
À a
3
Þða
2
À a
1
ފ; ð9Þ
where a
1
, a
2
, a
3
, and a
4
represent the four
parameters describing the trapezoidal fuzzy num-
ber A. M is either Max or Min. Hence, D
min
¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
D
2
ðA; MinÞ
p
and D
max
¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
D
2
ðA; MaxÞ
p
. f ðaÞ is a
continuous increasing function adopted in the
original formula (Tran and Duzkstein, 2002),
serving as a weighting function for computing
weighted sum of distances (Tran and Duzkstein,
2002). The f ðaÞ and a are all between 0 and 1. The
f ðaÞ is normally chosen to be equal to 1 (constant)
or a, depending on decision-makers or practitio-
ners’ preference. Generally, for risk-prone deci-
sion-makers, f ðaÞ ¼ 1 is used; for risk-averse or
risk-neutral decision-makers, f ðaÞ ¼ a is used
(Tran and Duzkstein, 2002). Because the trape-
zoidal fuzzy number Aða
1
; a
2
; a
3
; a
4
Þ can be chan-
ged to a triangular fuzzy number (i.e., Aða
1
; a
2
; a
4
Þ)
if a
2
¼ a
3
, the above formulae are applicable to
the triangular fuzzy number.
5. Fuzzy simulation advancement algorithm for
fuzzy discrete-event simulation
The selected fuzzy ranking measure is applied to
rank fuzzy times for fuzzy simulation advance-
ment and is developed into two fuzzy arithmetic
algorithms that are incorporated with the im-
proved AS simulation strategy of discrete-event
simulation.
5.1. Fuzzy-time-management algorithms based on
fuzzy ranking
No matter whether the times (e.g. start event
times, end event times, or entity available times) are
crisp or fuzzy, these times range from the simula-
tion start time (i.e. 0) to the simulation stop time
or simulation time limit (represented as m STL).
Therefore, the Min and Max in Eqs. (6) and (7)
are 0 and m STL, respectively. As a time-control
method for discrete-event simulation, it is unnec-
essary to consider decision-makers’ risk-proneness,
i.e. f ðaÞ ¼ 1. Hence, Eq. (8) for f ðaÞ ¼ a is used to
compute D
min
and D
max
for ranking the fuzzy times.
In order to comprehensively consider D
min
and
D
max
in ranking fuzzy times, two steps are adopted.
The first step is to compute D
min
ðMin ¼ 0Þ for the
fuzzy numbers and then decide that a fuzzy num-
ber with a smaller D
min
is smaller, or a fuzzy
number with a larger D
min
is larger. When this step
fails to rank the fuzzy numbers, that is, the D
min
of
the fuzzy times are equal, the second step is used.
The D
max
ðMax ¼ m STLÞ is computed in the sec-
ond step and then decide that a fuzzy number with
a smaller D
max
is larger, or a fuzzy number with a
larger D
max
is smaller. If the D
max
of the fuzzy times
are still found to be equal, these fuzzy times are
considered to be equal.
Therefore, the determination of the fuzzy start
time of an activity, i.e. operation (2), can be
accomplished through the following steps:
Step 1: For the activity i that is satisfied with
start conditions (i.e. logical feasibility and
resource requirements), initialize T
max
¼
ð0; 0; 0; 0Þ.
Step 2: For each entity’s fuzzy available time,
i.e. T
A
ði; jÞ ¼ ðT
A
1
ði; jÞ; T
A
m1
ði; jÞ; T
A
m2
ði; jÞ;
T
A
2
ði; jÞÞ, compute the D
min
ðT
A
ði; jÞ; 0Þ
through Eq. (8).
Step 3: Compute the D
min
ðT
max
; 0Þ through Eq.
(8).
Step 4: If D
min
ðT
A
ði; jÞ; 0Þ > D
min
ðT
max
; 0Þ then
T
max
¼ T
A
ði; jÞ
¼ T
A
1
ði; jÞ; T
A
m1
ði; jÞ; T
A
m2
ði; jÞ; T
A
2
ði; jÞ
À Á
:
Step 5: If D
min
ðT
A
ði; jÞ; 0Þ ¼ D
min
ðT
max
; 0Þ then
compute the D
max
ðT
A
ði; jÞ; m STLÞ for each
entity’s fuzzy available time T
A
ði; jÞ
through Eq. (8).
Step 6: Compute the D
max
ðT
max
; m STLÞ through
Eq. (8).
720 H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729
Step 7: If D
max
ðT
A
ði; jÞ; m STLÞ6D
max
ðT
max
; m STLÞ
then
T
max
¼ T
A
ði; jÞ
¼ T
A
1
ði; jÞ; T
A
m1
ði; jÞ; T
A
m2
ði; jÞ; T
A
2
ði; jÞ
À Á
:
Step 8: Repeat steps 2–7 until all entities required
by activity i are scanned.
Step 9: Get the start time of activity i, that is,
T
S
ðiÞ ¼ T
max
.
On the other hand, operation (4) that deter-
mines the next time that the current simula-
tion time will advance to contain the following
steps:
Step 1: Initialize T
min
¼ ðm STL; m STL; m STL;
m STLÞ.
Step 2: Scan end event list, compute the
D
min
ðT
EE
ðkÞ; 0Þ for each event-end’s due
time, i.e.
T
EE
ðkÞ ¼ ðT
EE
1
ðkÞ; T
EE
m1
ðkÞ; T
A
m2
ðkÞ; T
A
2
ðkÞÞ;
through Eq. (8).
Step 3: Compute the D
min
ðT
min
; 0Þ through Eq.
(8).
Step 4: If D
min
ðT
EE
ðkÞ; 0Þ < D
min
ðT
min
; 0Þ then
T
min
¼ T
EE
ðkÞ
¼ ðT
EE
1
ðkÞ; T
EE
m1
ðkÞ; T
EE
m2
ðkÞ; T
EE
2
ðkÞÞ:
Step 5: If D
min
ðT
EE
ðkÞ; 0Þ ¼ D
min
ðT
min
; 0Þ then
compute the D
max
ðT
EE
ðkÞ; m STLÞ for each
entity’s fuzzy available time T
EE
ðkÞ
through Eq. (8).
Step 6: Compute the D
max
ðT
min
; m STLÞ through
Eq. (8).
Step 7: If
D
max
ðT
EE
ðkÞ; m STLÞ PD
max
ðT
min
; m STLÞ
then
T
min
¼ T
EE
ðkÞ
¼ ðT
EE
1
ðkÞ; T
EE
m1
ðkÞ; T
EE
m2
ðkÞ; T
EE
2
ðkÞÞ:
Step 8: Repeat steps 2–7 until all end events are
scanned.
Step 9: Get the next time the current simulation
time will advance to, that is, m Now ¼
T
min
.
Because a triangular fuzzy number and a crisp
variable can be respectively represented by a
trapezoidal fuzzy number, that is, ðD
1
; D
m1
; D
2
Þ ¼
ðD
1
; D
m1
; D
m1
; D
2
Þ and D
1
¼ ðD
1
; D
1
; D
1
; D
1
Þ, all the
above algorithms are applicable to the triangular
fuzzy number-described activity duration or the
crisp activity duration.
5.2. Incorporation of fuzzy time management
with the improved as simulation strategy
The above fuzzy arithmetic algorithms for
fuzzy time management based on the fuzzy
ranking measure should be incorporated with the
improved AS simulation strategy so as to advance
simulation experimentation. The occurring times
of start and end events as well as the selection of
end events should be dynamically determined
through the two fuzzy algorithms. The fuzzy
advancement algorithm of the improved AS
simulation strategy is presented in the flowchart
shown in Fig. 4.
When a simulation experiment begins, the
simulation time is initialized at (0, 0, 0, 0) (i.e. a
crisp value 0) and the end event list is set to be
empty. For each fuzzy simulation time (m Now)
that has been advanced to, all end events that are
due at the current time will be activated, includ-
ing releasing involved entities (i.e. resources and
logical dependencies), which will be allocated
again. When all end events due at the current
time (m Now) are activated, the simulation time
limit (m STL) and all activities’ cycle limits will be
checked and stop the simulation experiment if
they are exceeded. After releasing and allocating
entities involved in the activities that have com-
pleted so far, all activities will be scanned to see
which are satisfied with start conditions. The
current simulation time will not be advanced
until no activities are satisfied with start condi-
tions. The simulation experiment will be ter-
minated when the simulation time limit (m STL)
is exceeded or when all activities’ cycle limits
H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729 721
are exceeded even if m STL is not exceeded.
Here an activity’s cycle limit is used to control a
repetitive activity’ operation required in a pro-
ject.
Start
Scan end event list to
select end event due
at m _ Now?

Remove the current
end event from the
event list
Advance the current
simulation time:
{ }
) (k T m _ Now = min
EE
k=1,...,K
Stop
Allocate these entities
by updating their
positions
Yes
No
Yes
No
Have end event
due at

m_Now?
No
Yes
Have activity
satisfied with start
conditions?
Initialize
simulation time,
end event list,
and others
Exceed m_STL or
all activities’ cycle
limits?
Scan Activities



Check logical dependencies
Check resource requirements
Release the involved
entities by updating
their available times:
m _ Now
j i T
A
) , ( =
Add the end event to
the end event list,
including its due time
Determine the end
event’s due time:
) ( ) ( ) ( i D i T k T
S EE
=
Update the involved
entities’ states from
‘idle’ to ‘busy’
Determine this
activity’s start time:
{ } ) , (
max
) ( j i T i T
A
j=1,...,J
S
=
Update the released
entities’ states from
‘busy’ to ‘idle’

Fig. 4. Fuzzy advancement algorithm of the improved AS simulation strategy.
722 H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729
6. Fuzzy discrete-event simulation system
The fuzzy advancement algorithm of the im-
proved AS simulation strategy is associated with
an activity-based graphical modeling interface
(Zhang et al., 2001, 2002) in order to develop a
fuzzy discrete-event simulation system for con-
struction. The activity-based graphical model is
developed on the basis of the activity-on-node
network of CPM and the activity cycle diagram
(ACD) (Hills, 1971), similar to the ‘‘Micro-Saint’’
graphical model (Catherine and Laughery, 1996).
Unlike other enhanced ACD graphical models
(e.g. CYCLONE) that require two types of activity
elements (i.e. conditional or unconditional activi-
ties) and queue element (for modeling entity’s
flow), the activity-based model requires only one
type of activity elements, reflecting the viewpoint
of the improved AS simulation strategy that fo-
cuses on activities.
Fig. 5 is the graphical modeling interface of the
fuzzy simulation system, where an activity-based
model and an activity-associated dialog box,
including its sub-dialog boxes, are shown. The
activity name and its label are marked at the top of
the activity element; required resources are marked
at the bottom of the activity element; the number
before the colon indicates the unit the activity re-
quires for each cycle and the number after the
colon indicates the unit initially allocated to this
activity. Logical dependencies and required units
(the completion number of the preceding activities)
are marked at the centre of the activity element.
The required unit smaller than 1 means that one
cycle of the preceding activity enables the current
activity to activate multiple times; replacing the
CON and GEN functions of CYCLONE. The
activity-associated dialog box, which can be
opened by double-clicking the activity elements, is
used for defining input information such as fuzzy
activity duration, required resources and logical
dependencies in modeling. The ‘‘Crisp resource
demand’’ is used and the fuzzy or flexible resource
demands are beyond the scope of this paper. This
Fig. 5. Graphical modeling interface and activity-associated dialog boxes.
H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729 723
fuzzy discrete-event simulation is implemented
with Visual C++ programming language, enabling
higher execution speed and easier subsequent en-
hancements.
7. Illustrative simulation experiments
In order to highlight the strengths of the fuzzy
discrete-event simulation system, a series of simu-
lation experiments using fuzzy activity durations
and probability distribution-modeled activity
durations are carried out and their outputs are
compared.
The illustrative simulation model is for a five-
story building construction (Fig. 6). The con-
struction starts with the erection of the structural
frame of each floor, then bricking exterior and
installing drywall successively. Bricking elevator
shaft follows erecting the structural frame and can
be operated simultaneously with bricking exterior
and installing drywall. After brickwork of all
floors is finished, the elevator is installed in one
operation. Three groups of workers are initially
allocated to ‘‘Erect floor frame’’, ‘‘Brick exterior’’
and ‘‘Install drywall’’; each contains 1, 1 and 2
units respectively. ‘‘Brick exterior’’ and ‘‘Brick
elevator shaft’’ share the same group of workers
and their priorities depend on which is first satis-
fied with the logical dependencies. ‘‘Start’’ and
‘‘Stop’’ are dummy activities with zero duration
and are stipulated to operate in turn, controlling
the simulation cycle. One cycle of ‘‘Start’’ enables
‘‘Erect floor frame’’ to activate 5 times. The
symbol ‘‘&’’ at ‘‘Install elevator’’ means logical
‘‘AND’’ relationships among related logical
dependencies, that is, 5 cycles of ‘‘Brick exterior’’
and ‘‘Brick elevator shaft’’ are both necessary to
activate ‘‘Install elevator’’ once.
For simplicity of illustration, all activities ex-
cept for ‘‘Brick exterior’’ are assumed to have crisp
duration. These activity durations are tabulated
in Table 1. Activity ‘‘Brick exterior’’, which is fo-
cused in the example, is assumed to have uncertain
duration that is estimated by the fuzzy number due
to the lack of any field sample data. Meanwhile,
probability distributions for the uncertain or sto-
chastic duration of ‘‘Brick exterior’’ are also pro-
vided. The comparisons of simulation outputs
from the fuzzy simulation and probability-based
simulation approaches are performed through two
simulation scenarios, each scenario consisting of
different fuzzy numbers and different probability
distribution groups.
In the first simulation scenario, the duration of
‘‘Brick exterior’’ is estimated to be lying around 8
Table 1
Duration of some activities for both simulation scenarios
Activity Start Erect floor frame Brick elevator shaft Install drywall Install elevator Stop
Duration 0 12 4 6 10 0
Start (0)

Erect floor frame (1)

0 (1/5)

Crew1 (1:1)
Brick exterior (2)

1 (1)

Crew2 (1:1)
Install drywall (4)

2 (1)

Crew3 (1:2)
Brick elevator shaft (3)

1 (1)

Crew2 (1)
Install elevator (5)

2 (5) & 3 (5)

Stop (6)

4 (5) & 5 (1)

Fig. 6. The activity-based graphical model for the construction of a five-story building.
724 H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729
and definitely between 6 and 10. Therefore, a tri-
angular fuzzy number is used. Other probability
distributions that are found to fit the estimated
activity duration include triangular, beta, normal
and lognormal distributions. The triangular fuzzy
number and the 4 probability distributions are
listed in Table 2. The related parameters of the
probability distributions are derived from the three
parameters defining the triangular fuzzy number.
In the second simulation scenario, the duration
of ‘‘Brick exterior’’ is estimated to be lying in the
interval of 7–8 and definitely between 6 and 11,
which is described by a trapezoidal fuzzy number.
Similarly, four probability distributions (i.e. one
triangular, one beta, and two lognormal distribu-
tions) that are considered to be able to reflect the
feature of the estimated activity duration are used.
These probability distributions and the related
parameters derived from the four parameters
defining the trapezoidal fuzzy number are listed in
Table 3.
Due to the internal mechanism of the fuzzy
arithmetic algorithms in the fuzzy advancement of
discrete-event simulation, only one cycle of the
simulation experiment for each scenario is suffi-
cient when using fuzzy activity durations. For the
probability distribution for activity durations,
however, multiple replications of the simulation
experiments are required to provide statistical
analyses of the simulation outputs and 30 cycles of
the simulation experiments for each of the total 8
(two scenarios) probability distributions are per-
formed.
When using the fuzzy activity durations, the
project duration of the building construction ob-
tained from one cycle of the simulation experiment
is also a fuzzy number. For the probability distri-
bution-based activity duration, the project dura-
tion should be presented in the statistical form of
frequency or cumulative frequency. For the sake
of comparing the simulation results from the
probability distributions with the fuzzy number,
the frequencies of the project durations are nor-
malized by dividing them with the maximum fre-
quency.
Figs. 7 and 8 show the outputs from simulation
scenarios one and two respectively, including fuzzy
number-described project durations and the nor-
malized frequency diagrams of the project dura-
tions for all 2 ·4 probability distributions. The
Table 2
Activity duration of ‘‘Brick exterior’’ for simulation scenario one
Duration type Triangular fuzzy
number
Triangular
distribution
Beta distribution Normal
distribution
Lognormal
distribution
Duration TriFuz. (6, 8, 10) Tri. (6, 8, 10) Beta (6, 10, 3, 3) Nor. (8, 2) Log. (8, 1)
Parameter explanation Minimum¼6 Minimum¼6 Minimum¼6 Mean l ¼ 8 Mean l ¼ 8
Mode ¼8 Mode ¼8 Maximum¼10 Standard variance
r ¼ 2
Standard variance
r ¼ 1
Maximum¼10 Maximum¼10 Shape a ¼ 3
Shape b ¼ 3
Table 3
Activity duration of ‘‘Brick exterior’’ for simulation scenario two
Duration type Trapezoidal fuzzy
number
Triangular
distribution
Lognormal
distribution 1
Lognormal
distribution 2
Beta distribution
Duration TraFuz. (6, 7, 8, 10) Tri. (6,7.5, 11) Log. (8, 1) Log. (7.5, 1) Beta (6, 11, 1.5, 4.5)
Parameter explanation Minimum¼6 Minimum¼6 Mean l ¼ 8 Mean l ¼ 7:5 Minimum¼6
Mode 1 ¼7 Mode ¼7.5 Standard variance
r ¼ 1
Standard variance
r ¼ 1
Maximum¼11
Mode 2 ¼8 Maximum¼11 Shape a ¼ 1:5
Maximum¼10 Shape b ¼ 4:5
H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729 725
fuzzy project duration from the first scenario is a
triangular fuzzy number, i.e. (70, 80, 90), and the
fuzzy project duration from the second scenario is
a trapezoidal fuzzy number, i.e. (70, 75, 80, 95).
The two figures show that the statistic outputs
from all the probability distribution-based activity
durations are generally contained in the fuzzy
number-described ones from the fuzzy activity
durations. It is also shown that the statistical
simulation outputs depend on the adopted prob-
ability distributions or related parameters, which
are actually subjectively selected or determined.
Besides, the statistical simulation outputs are
influenced by the simulation cycles. On the con-
trary, the fuzzy number-described simulation out-
puts produced from the fuzzy activity duration
and the fuzzy discrete-event simulation seem to be
able to explain the uncertainties modeled by
probability distributions even if they are appro-
priately selected.
Even though only one activity’s duration is
described with a fuzzy number in the above
experiments, considering more activities’ fuzzy
durations will not lead to much more work for
simulation experimentation. Other inputs such as
simulation time (for controlling simulation termi-
nation) remain unchanged and outputs requires no
further explanation because final project duration
0
0.2
0.4
0.6
0.8
1
70-72 72-74 74-76 76-78 78-80 80-82 82-84 84-86 86-88 88-90
Project duration
N
o
r
m
a
l
i
z
e
d

f
r
e
q
u
e
n
c
y

(
%
)
Tri. (6,8,10)
Beta (6,10,3,3)
Nor. (8,2)
Log. (8,1)
TriFuz. (6, 8, 10). (70, 80, 90)
Fig. 7. Comparison of simulation results for triangular fuzzy and distribution-based activity durations.
0
0.2
0.4
0.6
0.8
1
70-72.6 72.6-75.2 75.2-77.8 77.8-80.4 80.4-83 83-85.6 85.6-88.2 88.2-90.8 90.8-93.4 93.4-96
Project duration
N
o
r
m
a
l
i
z
e
d

f
r
e
q
u
e
n
c
y

(
%
)Tri. (6,7.5,11)
Log. (8,1)
Log. (7.5,1)
Beta (6,11,1.5,4.5)
TraFuz. (6, 7, 8, 11). (70, 80, 85, 95)
Fig. 8. Comparison of simulation results for trapezoidal fuzzy and distribution-based activity durations.
726 H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729
will remain to be a fuzzy number due to the
internal mechanism of the fuzzy arithmetic oper-
ations (e.g., fuzzy addition). Preparing parameters
(e.g., four parameters required for a trapezoidal
fuzzy number) for fuzzy numbers is easier than
defining probability-distributions, especially when
no or sufficient sample data are available.
8. Feature summary of the of fuzzy discrete-event
simulation
From the above descriptions and a series of
simulation experiments, the features of the fuzzy
discrete-event simulation can be summarized as
follows:
• The fuzzy discrete-event simulation uses the
uncertain input information (i.e. the fuzzy activ-
ity duration) that is described by fuzzy numbers
(including triangular and trapezoidal fuzzy
numbers) instead of probability distributions.
• The fuzzy activity durations are able to reflect
the subjectivity, vagueness or imprecision in
duration estimation and can be relatively easily
defined when insufficient or no sample data are
available to define probability distributions.
• The fuzzy discrete-event simulation uses fuzzy
arithmetic operations such as fuzzy addition
and fuzzy algorithms based on the fuzzy num-
ber ranking measure, controlling fuzzy simula-
tion advancement (e.g. time advancement and
event selection) for simulation experimentation.
• The fuzzy discrete-event simulation requires
only one cycle of the simulation experiment to
generate the simulation outputs, unlike the prob-
ability distribution-based simulation that needs
multiple cycles of the simulation experiments.
• The fuzzy discrete-event simulation produces
the fuzzy simulation outputs due to the internal
mechanism of the fuzzy arithmetic operations
in the simulation advancement algorithm.
• The fuzzy simulation outputs from the fuzzy
discrete-event simulation include the statistical
information from the probability-based simula-
tion; the latter is affected by distribution types,
related distribution parameters, and simulation
cycles.
It should be noticed that, however, the advan-
tages of the fuzzy discrete-event simulation can be
reflected only when probability-based simulation
cannot be adopted because no or sufficient sample
data are available to define probability distribu-
tions. In addition, because the fuzzy ranking
measure limits the fuzzy numbers to triangular or
trapezoidal fuzzy numbers, the fuzzy discrete-
event simulation based on the fuzzy ranking
measure cannot consider other kinds of fuzzy
numbers instead of triangular or trapezoidal fuzzy
numbers, which requires to be resolved in the
further research.
9. Conclusions
The applications of fuzzy set theory and a fuzzy
ranking measure to discrete-event simulation for
construction provide an alternative method to
model fuzzy activity durations through triangular
or trapezoidal fuzzy numbers. The fuzzy discrete-
event simulation can overcome the problems in
preparing probability distributions based on field
data and can take into account some vagueness,
imprecision and subjectivity in activity duration
estimation, in particular when a construction
project is proposed for the first time without any
former field data collected. The features of the
fuzzy discrete-event simulation are addressed
through a series of simulation experiments and
comparisons with the simulation when probability
distributions are used. The fuzzy discrete-event
simulation is a relatively speedy simulation tool
owing to its easiness of defining fuzzy numbers and
requirement of only one simulation cycle. Two
arithmetic algorithms based on the flexible and
easy-to-implement fuzzy number ranking measure
are developed and integrated with the improved
AS simulation algorithm, controlling fuzzy time
advancement and event selection during simula-
tion experimentation.
The use of fuzzy number for activity duration
and the combination of fuzzy arithmetic opera-
tions with the AS strategy for fuzzy discrete-event
simulation still needs to be studied further,
including applying other kinds of fuzzy num-
bers to describe uncertain activity duration. In
H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729 727
addition, the potentials of the fuzzy discrete-event
simulation and its internal connections or differ-
ence with/from the probability distribution-based
simulation approach need further exploration.
Acknowledgements
The work described in this paper was fully
supported by a grant from the Research Grants
Council of the Hong Kong Special Administration
Region, China (Project no. CityU 1015/02E).
References
AbouRizk, S.M., Halpin, D.W., 1992. Statistical properties of
construction duration data. Journal of Construction Engi-
neering and Management, ASCE 118 (3), 525–544.
Bellman, R.E., Zadeh, L.A., 1970. Decision-making in a fuzzy
environment. Management Science 17 (4), 141–164.
Bojadziev, G., Bojadziev, M., 1997. Fuzzy Logic for Business,
Finance, and Management. World Scientific, Singapore.
Bortolan, G., Degani, R., 1985. A review of some methods
for ranking fuzzy subsets. Fuzzy Sets and Systems 15,
1–19.
Catherine, E.D., Laughery, K.R., 1996. Advanced uses for
Micro Saint simulation software. In: Proceedings of 1996
Winter Simulation Conference IEEE, Piscataway, NJ,
pp. 510–516.
Chanas, S., Kamburowski, J., 1981. The use of fuzzy variables
in Pert. Fuzzy Sets and Systems 5, 11–19.
Chen, S.H., 1985. Ranking fuzzy numbers with maximizing set
and minimizing set. Fuzzy Sets and Systems 17, 113–129.
Chen, S.J., Hwang, C.L., 1992. Fuzzy Multiple Attribute
Decision Making. Springer, Berlin.
Dohnal, M., 1983. Fuzzy simulation of industrial problems.
Computers in Industry 4, 347–352.
Duobis, D., Prade, H., 1988. An approach to computerized
processing of uncertainty. In: Possibility Theory. Plenum,
New York. pp. 62–63.
Fente, J., Schexnayder, C., Knutson, K., 2000. Defining a
probability distribution function for construction simula-
tion. Journal of Construction Engineering and Manage-
ment, ASCE 126 (3), 234–241.
Fishwick, P.A., 1991. Fuzzy simulation: Specifying and iden-
tifying qualitative models. International Journal of General
Systems 19, 295–316.
Gaul, W., 1981. Bounds for the expected duration of a
stochastic project planning model. Journal of Information
and Optimization Sciences 2, 45–63.
Halpin, D.W., 1976. Cyclone––a method for modeling job site
processes. Journal of Construction Division, ASCE 103
(3), 489–499.
Hills, P.R., 1971. HOCUS, P.E. Group, Egham, Surrey,
Endland, UK.
Klein, H.W.K., 1986. Robustness against dependence in Pert:
An application of duality and distributions with known
marginals. Mathematical Programming Study 27, 153–182.
Klein, R.W., Baris, P.M., 1990. Selecting and generating
variates for modeling service times. In: Res. Memorandum
no. 90-2, Purdue University, West Lafayette, Ind.
Law, A.M., Kelton, W.D., 1991. Simulation Modeling and
Analysis. McGraw-Hill, New York.
Liou, T.S., Wang, M.J.J., 1992. Ranking fuzzy numbers with
integral value. Fuzzy Sets and Systems 50, 247–255.
Lootsma, F.A., 1989. Stochastic and fuzzy Pert. European
Journal of Operational Research 43, 174–183.
Lu, M., 2002. Simplified enhancing project evaluation and
review technique simulation through artificial neural net-
work-based input modeling. Journal of Construction Engi-
neering and Management, ASCE 128 (5), 438–445.
Maio, C., Schexnayder, C., Knutson, K., Weber, S., 2000.
Probability distribution function for construction simula-
tion. Journal of Construction Engineering and Manage-
ment, ASCE 126 (4), 285–292.
Martinez, J., Ioannou, P.G., 1999. General-purpose system for
effective construction simulation. Journal of Construc-
tion Engineering and Management, ASCE 125 (4), 265–
276.
Nasution, S.H., 1999. Techniques and applications of fuzzy
theory to critical path method. In: Fuzzy Theory Systems 4,
pp. 1561–1597.
Negi, D.S., Lee, E.S., 1992. Analysis and simulation of fuzzy
queues. Fuzzy Sets and Systems 46, 321–330.
Petrovic, D., Roy, R., Petrovic, R., 1998. Modeling and
simulation of a supply chain in an uncertain environment.
European Journal of Operational Research 109, 299–309.
Pidd, M., 1998. Computer Simulation in Management Science.
Wiley, Chichester, New York.
Southall J.T., Wyatt M.D., 1988. Investigation using simulation
models into manufacturing/distribution chain relationships.
BPICS Control 4/5, 29-34.
Stephen, E.C., 1998. Input data analysis. In: Banks, J. (Ed.),
Handbook of Simulation: Principles, Methodology Ad-
vances Applications and Practice. John Wiley & Sons Inc.,
New York. pp. 55–91.
Stephen, E.C., 1999. Steps to implement bayesian input
distribution selection. In: Proceedings of 1999 Winter
Simulation Conference, IEEE, Piscataway, NJ, pp. 317–
324.
Touran, A., 1997. Probabilistic model for tunneling project
using markov chain. Journal of Construction Engineering
and Management, ASCE 23 (1), 444–449.
Tran, L., Duzkstein, L., 2002. Comparison of fuzzy numbers
using a fuzzy distance measure. Fuzzy Sets and Systems 130,
331–341.
Wang, X., Kere, E.E., 2001a. Reasonable properties of fuzzy
quantities (I). Fuzzy Sets and Systems 118, 375–385.
Wang, X., Kere, E.E., 2001b. Reasonable properties of fuzzy
quantities II. Fuzzy Sets and Systems 118, 385–405.
728 H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729
Weiler, H., 1965. The use of the incomplete beta functions for
prior distributions in binomial sampling. Technometrics
7 (3), 335–347.
Wilson, J.R., Vaughan, D.K, Naylor, E., Voss, R.G., 1982.
Analysis of space shuttle ground operations. Simulation 38
(6), 187–203.
Yoon, K.P., 1996. A probabilistic approach to rank com-
plex fuzzy numbers. Fuzzy Sets and Systems 80, 167–
176.
Yufei, Y., 1991. Criteria for evaluating fuzzy ranking methods.
Fuzzy Sets and Systems 44, 139–157.
Zadeh, L.A., 1965. Fuzzy sets. Information and Control 8, 338–
353.
Zhang, H., Tam, C.M., Shi, J., 2001. Visual modeling and
simulation for construction operations. Automation in
Construction 11 (1), 47–57.
Zhang, H., Tam, C.M., Shi, J., 2002. Simulation-based meth-
odology for project scheduling. Construction Management
and Economics 20 (8), 667–678.
Zhang, H., Tam, C.M., Shi, J., 2003. Application of fuzzy logic
to simulation for construction operations. Journal of
Computing in Civil Engineering, ASCE 17 (1), 38–45.
H. Zhang et al. / European Journal of Operational Research 164 (2005) 715–729 729