You are on page 1of 10

174 European Journal of OperationalResearch43 (1989) 174-183

North-Holland

Theory and Methodology

Stochastic and fuzzy PERT


F.A. L O O T S M A
Faculty of Mathematics and Informatics, Delft University of Technology, P. 0. Box 356, 2600 AJ Delft,
Netherlands

Abstract: Network-planning techniques are considered based on stochastic and fuzzy models of the activity
durations. The stochastic versions of PERT are generally intractable, and they cannot be used to draw up
tight plans for action. Fuzzy models are closer to reality, simpler to use, but theoretically not well
established. The impact of a stochastic and a fuzzy version of PERT is illustrated via a numerical example.
Finally, the representation of uncertainty in network planning is discussed, when the activity durations are
estimated by human experts.

Keywords: Network planning, minimum project duration, gamma distributions, maxima of stochastic
variables, fuzzy numbers, triangular membership functions, maxima of fuzzy numbers, normative modell-
ing, descriptive modelling, randomness, vagueness

1. Introduction materials at the right moment, and to control the


execution of the project.
Ever since the first publication (the PERT Sum- The PERT team also tried to model uncertainty.
mary Report, 1958; see also Malcolm et al., 1959), They proposed to ask three estimates for each
the Project Evaluation and Review Technique PERT activity duration (the optimistic, the most-likely,
has been a successful tool of Operations Research. and the pessimistic estimate), they modelled each
With the activities of a project as the arcs and the activity duration as a stochastic variable with an
milestones or events as the nodes, the network appropriate beta distribution, and they proposed a
model visualizes the order in which the activities simple method to calculate the expectation and the
must be carried out. It is typically a directed, variance of the event times approximately.
acyclic network with exactly one initial node, the There is a systematic error in this simple version
start, and one terminal node, the completion of of PI~RT, however. At every milestone where several
the project. Moreover, if there is enough certainty activities join together, the earliest event time is
to express each activity duration by exactly one the maximum of a number of generally dependent
estimate, the minimum project duration is unam- stochastic variables. By the simple approximate
biguously determined by the longest path from the maximization in the original method, the expected
initial to the terminal node. The earliest and the event times, and henceforth the expected mini-
latest event times, also calculated by the longest- mum project duration, are underestimated. Ini-
path algorithm, provide a framework for plan- tially, this prompted several authors (Clark, 1961;
ning: they enable the user to contract labour and Fulkerson, 1962; MacGrimmon and Ryavec, 1964;
Martin, 1965; Hartley and Wortham, 1966;
Lootsma, 1966; Elmaghraby, 1967, 1977; Burt and
Received May 1988; revisedJuly 1988 Garman, 1971) to improve the method by a more
0377-2217/89/$3.50 1989, ElsevierSciencePublishers B.V.(North-Holland)
F.A, Lootsma / Stochastic and fuzzy PERT 175

sophisticated, but nevertheless approximate maxi- resent the length of the corresponding arc in the
mization, and also by simulation. Later authors project network. It is a common experience that
(Kleindorfer, 1971; Robillard and Traham, 1977; project leaders and their subcontractors generally
Shogan, 1977; Meilijson and N~tdas, 1979; Nhdas, expect the planned minimum project duration to
1979; Gaul, 1981; Cleef and Gaul, 1983; Klein be equal to the length of a longest path from the
Haneveld, 1986) admitted that the problem of initial to the terminal node. Otherwise, there would
finding the distribution or the moments of the be positive floats along every path from start to
minimum project duration is intractable: theoreti- finish: this would create serious conflicts during
cally too complicated, even if the activity dura- the execution of the project. Next, we summarize
tions are supposed to be independent, and compu- our extended stochastic version of PERT (1966) as
tationally too expensive. Instead, they turned to a well as a fuzzy version based on fuzzy activity
worst-case analysis. They attempted to derive an durations with triangular membership functions.
upper bound for the expected violation of a pre- Thereafter, we calculate the earliest event times
determined project delivery moment by a maximi- and the minimum duration of a particular project
zation over the class of joint distributions compat- via these two methods, and also via simulation, in
ible with the marginal activity durations. order to show the discrepancies between the re-
Since the pioneering work of Zadeh (1965) in sults; to our knowledge, such a comparison has
fuzzy-set theory, it is questionable whether the never been made. With this example as a vehicle
above, stochastic models yield a proper represen- for discussion, we finally turn to the representa-
tation of the uncertainty in a project. The activity tion of uncertainty in projects where the activity
durations are estimated by human experts, some- durations are estimated by human experts. In
times under unique circumstances, so that we are doing so, we also refute the biting criticism of
confronted with the vagueness of their judgemen- French (1987) on the modelling of vague phenom-
tal statements. Modelling with fuzzy concepts was ena.
considerably improved by Dubois and Prade
(1978a) when they established arithmetic oper-
ations on fuzzy numbers. They applied the ideas 2. Tight planning
immediately to the PERT problem of finding the
event times and the minimum project duration We consider a directed, acyclic network with
(1978b). In addition, Chanas and Kamburowski node set (1, 2 . . . . . n }. The unique initial node has
(1981) suggested to use fuzzy numbers with trian- number 1, the unique terminal node number n.
gular membership functions based on the optimis- The network of project NABLAin Figure 1, which
tic, the most-likely, and the pessimistic estimates will be used to clarify the results of the stochastic
of the respective activity durations. Fuzzy num- and the fuzzy versions of PERT, is a typical exam-
bers of this type, thoroughly studied by Van ple of such a 'project network'.
Laarhoven and Pedrycz (1983), have successfully
been used in multi-criteria analysis (see Lootsma,
1987). We are convinced that they will also be
workable tools for project planning.
This leads us to the purposes of the present
paper. First, we discuss the important issue of
tight planning. Even if the activity durations are
uncertain, the project leader and the subcontrac-
tors cannot work without commitments, both be-
tween themselves and with the outside world. They
have to make certain promises, they have to make
agreements about the delivery of labour, raw
materials, tools, machinery, etc. Normally, each
activity duration, uncertain or not, is replaced by Figure 1. Network visualizing the order relations of the activi-
a planned activity duration (in practice always the ties (the arcs) in project NABLA. Activity ( j , k) cannot start
expected activity duration), which is taken to rep- before all activities (i, j ) , i ~ Bj, have been completed
176 F.A. Lootsma / Stochastic and fuzzy P~Rr

In the deterministic case, alternatively referred whereas for many distributions the strict in-
to as the crisp or non-fuzzy case, each arc (i, j ) equality sign holds. This has significant implica-
has a non-negative number dij associated with it: tions. Taking the expected activity durations to
the activity duration, estimated by the subcontrac- stand for the lengths of the corresponding arcs, we
tor who is responsible for the activity. The earliest find that the expected minimum project duration
event times are defined by the forward recursion is generally greater than the length of a longest
path from the initial to the terminal node. So a
ta = 0 , tj= max(ti+dij ), (1)
plan, drawn up with the expected event times to
i~Bj
schedule the supply of labour and materials and to
where Bj stands for the set of nodes 'immediately control the project, cannot be tight. There will be
before' j. Obviously, t n represents not only the positive floats, even along critical paths. The ex-
earliest event time of the terminal node, but also ample of Figure 1 will show that the expected
the minimum project duration. minimum project duration may easily exceed the
The latest event times are similarly defined by length of the longest path by more than 10%.
the backward recursion Calculating the distribution or the moments of
r, = t,, ~ = rain (% - djk ), (2) the minimum project duration _tn has appeared to
k~Aj be an intractable problem. There are not only
theoretical complications. One also has to solve
where Aj denotes the set of nodes 'immediately
the computational question of how to process and
after' j. It must be true, of course, that Zl = 0.
to store the stochastic dependence between the
There is at least one longest path, a so-called
quantities _ti + _dij in formula (4). These are heavily
critical path, from 1 to n, It is easy to show that
dependent if they have a long 'common history' in
node j is on a critical path if, and only if, tj -- ~).
the project. Dependence is due to the network
Moreover, activity (i, j ) is on a critical path if,
structure, and it occurs even when the activity
and only if, its total float z j - (t i + d~j) is zero.
durations are independent.
With these properties, the event times constitute
an excellent framework for a tight plan. By execut-
The original, simple version of PERT. The PERT
team originally introduced approximate expecta-
ing each activity within the respective time limits,
tions Ep(t_;) calculated according to the forward
the project leader and the subcontractors will
recursion
complete the project as soon as possible. The
completion time (delivery time) will indeed be Ep (t~) = 0,
minimized because the plan contains at least one
path without floats (a critical path). E p ( , j ) = max
, Bj [Ep(ti) + E(_d,j)]. (5)
In a stochastic version, each activity (i, j ) has
associated with it a stochastic activity duration Obviously, these quantities constitute a framework
_dij. Usually, the probability distribution of _d~j for a tight plan, when the expected activity dura-
has a non-negative support; this is at least the case tions are taken to be the arc lengths in the project
with the beta distribution originally proposed by network. That is the reason why (5) is currently
the PERa- team (1958), and with the gamma distri- used in network planning (see Smith, 1982). More-
bution introduced by Lootsma (1966, 1969). The over, uncertainty was, and still is, incorporated as
earliest event times are clearly defined by the follows. Let Pk stand for the longest path from 1
forward recursion to k; then the approximation Vp(_tk) to the vari-
ance V(t k) is given by
~1 = 0 , tj = max(t_i+d_ij). (3)
i~Bj Vp(-tk) = E V(_d,j). (6)
(i,j)~Pk
The expected earliest event times satisfy the
inequalities If Pk is not unique, then Ve(t_k) is the maximum
total variance over all longest paths from 1 to k.
E(t_j)=E[max(t_i+d_ij)] The earliest event times are further assumed to be
t i~j normally distributed.
>/max [ E ( t i ) + E ( d i j ) ] , (4) We did not realize how important the require-
i~Bj ment of tight planning is until we came up with a
F.A. Lootsma / Stochastic and fuzzy PERT 177

proposal to implement an extended version of for it, at least at first sight. Now, let us consider
PERT in a network-programming code, a version an activity duration _d with such a distribution,
with improved approximations to E(t_j) and V(t_j). and let us take Pk to denote the probability, sup-
With the expected activity durations as arc lengths, plied by the subcontractor, that the activity will be
the resulting plans would naturally have positive completed in k units of time. To be even more
floats along any path from 1 to n. For the project specific, let us suppose that he has chosen P8 =
leaders and their subcontractors, this was totally 0.10, P9 = 0.20, Pl0 = 0.60, Pll = 0.10. Now,
unacceptable! A tight plan constitutes a trans- imagine that the subcontractor comes under pres-
parent basis for discussions and negotiations. A sure and that he is asked to reconsider the chance
non-tight plan, however, creates conflicts about of completing the activity in 9 time units. He still
the floats. The project leaders fear that they can- has at least one degree of freedom: he may argue
not keep the subcontractors under control and that the activity should be completed to such an
that the floats will prematurely be squandered. extent that succeeding activities can start. So, a
The subcontractors will indeed use up any availa- slight shift in the concept of 'completion' is suffi-
ble slack in the plan, thus jeopardizing the planned cient for him to come up with a new 'probability
delivery of the project. estimate' for k = 9, almost as high as Pl0. We may
To our knowledge, the issue of tight planning is go even further. The vague definition of 'comple-
overlooked in the literature on PERT. Hence, one tion' enables him to assign to any integer k a new
also ignores the phenomenon that crude ap- number c k, 0 <~c k <~1, expressing the credibility
proximations like the Ep(t_j) are acceptable for (at least in his eyes) that the activity will be
planning purposes whereas the true values E(_tj) 'completed' in k time units. The slightly varying
are not! Other frameworks for planning (possibly notions of 'completion' may be overlapping. It
the medians of the event times, with the medians may also happen that they do not cover all possi-
or the modes of the activity durations as arc ble states of 'completion'. Anyway, the credibili-
lengths) have not been considered either. In fact, ties do not necessarily add up to 1. By these steps
we found in many textbooks for Operations Re- we have remodelled the activity duration as a
search a n d / o r Project Planning that the authors fuzzy number J with a discrete support.
ignored the opportunity to discuss the representa- The transition from discrete to continuous
tion of uncertainty via the example of PERT-plan- probability distributions or from discretely to con-
ning (see, for instance, Moder and Phillips, 1970; tinuously supported fuzzy numbers is frequently
Hillier and Liebermann, 1974; Wagner, 1975; made when the resulting distributions or member-
Eiselt and yon Frajer, 1977; Elmaghraby, 1977; ship functions can be described by a few parame-
Smith, 1982; Eppen, Gould and Schmidt, 1987; ters. We shall also do it here with the activity
Ravindran, Phillips and Solberg, 1987). Un- duration. There are three estimates at our dis-
certainty is a major issue in Operations Research, posal: the optimistic estimate a which would al-
and network planning with uncertain activity most always (in roughly 99% of all cases) be
durations is typically an area where stochastic exceeded, the most-likely estimate m, and the
concepts may be unworkable, as we will see in the pessimistic estimate b which would almost never
sections to follow. (in roughly 1% of all cases) be exceeded. With
these data we modelled (see Lootsma, 1966, 1969)
the stochastic activity duration d_ as a gamma dis-
3. Stochastic and fuzzy activity durations tribution with the density function

Activity durations are mostly expressed in days,


weeks, or months, and not seldom in integer num- ~F--~a)(d-a) 5-1 exp(-X(.- a))
bers of these time units. Even if the activity ends f(d) =
in the middle of a day, for instance, the project | if d>~a,
\0 if d < a.
organization is often so rigid that the succeeding
activities cannot start before the next morning. So,
if the activity duration is uncertain, a discrete The shape parameter a and the scale parameter X
probability distribution seems to be a proper model must be positive. If a > 1, the duration d has a
178 FA. Lootsma / Stochastic and fuzzy PERT

unique mode, which is set to m so that For the fuzzy activity duration d we adopt the
simplest possible model: a triangular membership
a-1
m=--+a. function with the basis [a, b] and the top value of
X 1 at m. In what follows, we shall refer to a, m and
Moreover, we have b as the lower value L ( d ) , the modal value M ( d ) ,
and the upper value U ( d ) respectively.
Ot
E(_d) = ~- + a.
Note. One of the referees suggested us to use a
It is usual in PERT-planning tO employ the triangular density function for a stochastic activity
expected activity duration in order to draw up a duration, so that the stochastic and the fuzzy
time schedule (see Section 2). The expectation model would be as similar as possible. Although it
should not therefore be below the most-likely is a valuable idea, we finally decided not to follow
estimate, because the subcontractor responsible the suggestion. First, the resulting stochastic ver-
for the activity and for the estimate must be taken sion of PERT would also lead to a non-tight plan.
seriously. Of course, there is no obvious reason Second, such a version would yield a higher
why the expectation should be above the most- expected project duration than the version based
likely estimate, but mostly we have b - m >> m - on a gamma distribution: the triangular distribu-
a, and then some extra time seems to be accepta- tion has the expectation ~(a + m + b), which
ble (see also Section 5; the fuzzy model will not exceeds (9) in the usual case that b - m >> m - a.
give extra time for planning). Hence, we have Our gamma distribution is rather similar to the
E(_d) > m, and we obtain traditional, widely-accepted beta distribution but
it has, in its turn, a somewhat higher expectation.
)t = 1 > 0, (7)
E(d)-m
e(d)-a 4. Maxima of stochastic variables and fuzzy num-
= - > 1. (8) bers
E(d ) -m

The original PERT formula for the expected In stochastic PERT, whether we have beta or
activity duration was based on a beta distribution gamma distributions for the activity durations, we
for _d. In an attempt to find an almost similar shall follow the original assumption that the
formula for E ( d ) , such that E(_d) > m, we started earliest event times _ti, as well as the quantities
off by setting _ti + d o are reasonably taken to be normally dis-
tributed. We are accordingly in the position to
E(d_)= ~(b+ 5m). (9) calculate the moments of
With the results (7)-(9), the tall probability P(_d > tj = m a x (t, + 4 , j )
b) appears to be dependent on the ratio r = (b - iEBj
m)/(m-a) only. Moreover, for a reasonable
range of values of r (r = 1, 2 . . . . ,5) the tail prob- by using the formulas derived by Clark (1961).
For ease of exposition, we consider two normally
ability is roughly 1%, so that we accepted formula
distributed variables _x and y, and we set _z=
(9). Finally, the variance of _d is simply expressed
max(x, y). Then
by
Ot E ( z ) = E ( x ) ~ ( 7 / ) + E ( y ) # ( - ~ l ) + sea(n),
V ( d ) = -~ = (E(d_) - a)(E(d_) - m). (10)
(11)
Obviously, the gamma distribution is not only e(z 2) = + E(_/) ( - . )
more natural than the original beta distribution, it
is also simpler: there is a straightforward relation- + + (12)
ship between the estimates a, m, and b on the one
hand and the parameters a and X on the other. where @ and ea denote the distribution function
This is important for simulation experiments. and the probability density function of the normal
F.A. Lootsma / Stochastic andfuzzy PERT 179

distribution with zero mean and variance 1. The In what follows we shall take Ec(t_j ) and Vc(_tj)
quantities s > 0 and ~ are defined by to denote the resulting approximations to the
expectation and the variance of _tj. Simulation
s 2 = V ( x ) + V ( y ) - 2 cov(x, y ) , experiments with small networks have shown that
these approximations are workable (see also the
= {E(x)- E(_yl)/s next section with a hitherto unpublished example).
The dependence between z, x and y is processed We always have
via the expressions
Ec Ep(#),
cov(z, x ) = V ( x ) ~ ( ~ / ) + cov(x, y ) O ( - 7 / ) ,
which implies that the improved approximations
(13) E c ( t j ) do not constitute a framework for a tight
plan when the expected activity durations are taken
cov(z, y) = cov(_ , +
to stand for the arc lengths.
(14) (b) A fuzzy version of PERT. In fuzzy PERT the
calculations are much simpler, at least as far as
Finally, if x, y and w_ stand for normally distrib- the lower, modal and upper values of the earliest
uted variables with known covariances, then we event times are concerned. The membership func-
have tions, however, do not keep the simple triangular
cov(_, w)= cov(x, coy(y, form associated with the activity durations. In the
discussion of the results, we shall nevertheless
(15) assume that deviations from that form may be
ignored. Considering the relation
In the rest of the paper we concern ourselves
with two versions of PERT: a stochastic version
with a gamma distribution for the activity dura-
max(
i~Bj
,+ 4),
tions and Clark's formulas to carry out the maxi-
and using the results of Dubois and Prade (1978a)
mization at the milestones (in contrast to the
for the lower, modal and upper values of maxima,
simple version of PERT we call it extended), as well
we have
as a fuzzy version with triangular membership
functions for the activity durations and the maxi- L(Tj) = m a x ( L ( 7 , ) + L ( a ~ j ) ) , (16)
mization formulas of Dubois and Prade (1978a). i~B: ~
(a) An extended, stochastic version of PERT. The
essential feature of an extended version is an
M(Tj) = max{ M(-t, ) + M(c~j)), (17)
i~Bj "
improved approximation to the moments of
U(?,) = m a x ( U ( 7 , ) + U ( Z j ) ). (18)
_tj = m a x ( L + d_ij). i~B/
i~Bj
Hence, the lower, modal, and upper values can be
Suppose that Bj = { p, q, r }, and set calculated separately via a longest-path algorithm.
x = _tp + _dpj, _y = _tq + dqj, w = _tr + G : , Moreover, with the modal values of the activity
durations as arc lengths, the modal values of the
where p, q and r are chosen such that E ( x ) > event times provide a framework for tight plan-
E ( y ) >t E(w) (numerical experience indicates that ping. It is worth noting that 'dependence' between
this order may be disrupted without serious effects fuzzy numbers is ignored in fuzzy logic.
on the quantities to be calculated). Compute the
desired moments of z=max(_x, y) and _tj=
max(y, w) according to Clark's formulas (11)- 5. Numerical e x a m p l e
(15). It will be obvious how the calculations pro-
ceed if Bj has more or less than three elements. Let us now consider the example of project
Finally, compute the covariances of _tj and any NABLA. The network is shown in Figure 1, and the
other event time in the network which has already estimates of the activity durations, as well as the
been 'attained' in the computational process. expectations and the variances, in Table 1. First,
180 F.A. Lootsma / Stochastic andfuzzy PERT

Table 1 extended version with the formulas (11)-(15) of


The three estimates (optimistic, most likely, and pessimistic), Clark (1961). Finally, Table 2 exhibits the ap-
the expectation Exp, and the variance Var of the activity
durations in project NABLA;Exp = ~(b + 5m), and Var = (Exp
proximate expectations and variances
- a)(Exp- m)
Es(tj ) and Vs(_tj),
Activity a rn b Exp Var
1-2 1 3 9 4 3
which are obtained by simulation (10000 realiza-
1-3 3 4 10 5 2 tions of the project, by sampling from the gamma
2-3 1 2 8 3 2 distributions of the activity durations). These
2-4 3 9 27 12 27 quantities are helpful yardsticks to decide how
3-4 2 8 20 10 16 well Ec(_tj) and Vc(tj) approximate the expecta-
1-7 22 30 54 34 48
3-7 18 24 42 27 27
tion and the variance of the earliest event times.
4-7 8 14 26 16 16 Let us now turn to the fuzzy version of PERT.
1-5 2 6 18 8 12 Table 3 shows the lower, modal, and upper values
5-7 16 22 40 25 27 of the fuzzy earliest event times of project NABLA.
1-6 11 14 20 15 4 Because a c o m p a r i s o n with the results i n T a b l e 2
6-7 15 19 25 20 5
c a n n o t readily be made, we have also recorded the
5-8 16 20 32 22 12
6-8 12 14 26 16 8 1% u n c e r t a i n t y ranges of the earliest event times
3-9 22 25 31 26 4 (expectation _+2 - s t a n d a r d deviation) o b t a i n e d via
4-9 14 16 22 17 3 the simple PERT version a n d via simulation.
7-10 9 13 25 15 12 F u z z y PERT a n d the original, simple version of
8-10 12 15 33 18 18
PERT e n a b l e us to consider critical paths from the
9-10 8 12 36 16 32
initial to the t e r m i n a l node. I n the fuzzy version,
with the modal values of the activity d u r a t i o n s as
arc lengths, the p a t h (1, 6, 7, 10) appears to be
we summarize the results of stochastic PERT. T a b l e
the u n i q u e critical path. I n the original, simple
2 displays the quantities
version, with the expected activity d u r a t i o n as arc
Ev(t_j ) and Vp(_tj), length, there are two critical paths, (1, 6, 7, 10}
a n d (1, 2, 3, 4, 9, 10}.
calculated according to the simple version origi- F o r a m o r e elaborate example, a project called
n a l l y proposed b y the PERT team. T a b l e 2 also DELTA with a n e t w o r k of 27 n o d e s a n d 44 arcs,
shows the a p p r o x i m a t i o n s the reader is referred to a n earlier p a p e r of the
a u t h o r (1966). The expectation a n d the variance of
Ec(_tj) and Vc(_tj)
the m i n i m u m project d u r a t i o n were according to
of E(tj) and V(ts) respectively, calculated via the the simple version of PERT 156.0 a n d 53, accord-

Table 2
Approximations to the expectation and the variance of the earliest event times in project NABLA
Milestone Expectation Variance
Original Extended Simulation Original Extended Simulation
PERT by Clark's 10000 PERT by Clark's 10000
(1958) formulas runs (1958) formulas runs
1 0 0 0 0 0 0
2 4.0 4.0 4.1 3 3 3
3 7.0 7.4 7.4 5 4 4
4 17.0 19.5 19.4 21 16 20
5 8.0 8.0 7.9 12 12 12
6 15.0 15.0 15.0 4 4 3
7 35.0 40.9 41.5 9 13 26
8 31.0 32.9 32.7 24 12 16
9 34.0 37.1 37.2 24 13 18
10 50.0 58.5 59.1 42 19 34
F.A. Lootsma / Stochastic and fuzzy PERT 181

Table 3
Lower, modal and upper values of the fuzzy earliest event times in project NABLA.For a comparative analysis we also show the 1%
uncertainty ranges (expectation + 2. standard deviation) produced by stochastic PERT
Milestone Fuzzy PERT Original PERT Simulation
Crit. Lower Modal Upper Crit. Exp - 2 o Exp Exp + 2 o Exp - 2 o Exp Exp + 2 o
path value value value paths
l * 0 0 0 */* * 0 0 0 0 0 0
2 1 3 9 * * 0.5 4.0 7.5 0.6 4.1 7.6
3 3 5 17 * * 2.5 7.0 11.5 3.4 7.4 11.4
4 5 13 37 * * 7.8 17.0 26.2 10.5 19.4 28.3
5 2 6 18 1.1 8.0 14.9 1.0 7.9 14.8
6 * 11 14 20 * 11.0 15.0 19.0 11.5 15.0 18.5
7 * 26 33 63 * 29.0 35.0 41,0 31.3 41.5 51.7
8 23 28 50 21.2 31.0 40.8 24.7 32.7 40.7
9 25 30 59 * * 24.2 34.0 43,8 28.7 37.2 45.7
10 * 35 46 95 */ * * 37.0 50.0 63.0 47.4 59.1 70.8

ing to the extended version 161.5 and 51, via would have been e m p l o y e d for small and medium-
simulation 161.5 and 66, whereas the fuzzy version size networks.
yields the lower, m o d a l and upper values 114, 140,
and 240 respectively.
W h e n the modal values of the earliest event 6. Conclusions
times are taken as a framework for planning,
fuzzy PERT gives a shorter project duration and a In the never-ending debate between ' p r o b -
larger range of uncertainty than the original, sim- abilists' and 'fuzzicists' about the representation
ple version of PERT where the plan is based on the of uncertainty, F r e n c h (1984, 1986, 1987) is cer-
expected activity durations. The shorter project tainly one of the most critical commentators. Basi-
duration is easy to explain. In fuzzy PERT, we take cally, he holds the viewpoint that the modelling of
the most-likely estimates of the activity duration uncertainty in a normative analysis is best accom-
to draw up a plan; in simple PERT, the plan is plished by the use of subjective probabilities. To
based on the expected activity durations, which avoid any misunderstanding, he clearly states that
always exceed the most-likely estimates when we he is concerned with h o w people ought to organize
use a g a m m a distribution, and practically always their j u d g e m e n t s (normative analysis), not how
(if b - m > m - a ) for a beta distribution. Inci- they actually do (descriptive analysis). This leads
dentally, it remains obscure w h y this extra time us straightaway to the heart of the matter, because
should be given into the bargain, as all stochastic the conflict between normative and descriptive
PERT versions do, the simple version as well as the modelling clearly emerges in PERT.
extended versions or simulation! W h y should the H u m a n j u d g e m e n t plays a d o m i n a n t role in
project leader be so generous to the subcontrac- PERT at two different points: (a) in the estimation
tors? Possibly because others m a d e h i m aware of of an activity duration, and (b) in the requirement
the law of large n u m b e r s in probability theory, that the resulting plan must be tight. Let us con-
without telling him that the law is invalid for sider b o t h situations in m o r e details.
unique or practically unique projects! (a) Estimation of an activity duration. As we
Extended versions and simulation have never have seen in Section 3, subcontrators m a y employ
been used for network planning, at least to our slightly varying definitions of the concept of ' c o m -
knowledge. This m a y be due to the c o m p u t a t i o n a l pletion' of their activities. This is what they actu-
b u r d e n (both Clark's formulas and simulation are ally do. T h e question of whether they ought to use
expensive for large networks), but we suspect that one single, proper definition is irrelevant. T h e y
the requirement of tight planning is the d o m i n a n t have the freedom, or they take the freedom, to
factor ruling out these versions. Otherwise, they complete their activities to a certain extent only,
182 F..A. Lootsma / Stochastic and fuzzy PERT

whereafter they signalize that succeeding activities be different (identical), and it is hardly possible to
may start. They finally complete their own activi- decide which is which.
ties in parallel, without jeopardizing the 'comple- Vagueness is not properly modelled by prob-
tion' of the project. A similar phenomenon may be ability theory. In many respects, the fuzzy PERT
observed at the last milestone, the delivery of the model is therefore closer to reality and more
project. Those who ever attended the opening of a workable than a stochastic model, notwithstand-
new building will understand this. They must have ing the weak foundations of fuzzy-set theory. The
noticed the loose ends behind the curtains or algebraic operations on fuzzy numbers as pro-
under the carpet. Otherwise, they should go and posed by Dubois and Prade (1978a), for instance,
visit the same building on the day after the open- rely on the notorious minimum and maximum
ing ceremony. Work is then still in progress, de- operators for the intersection and the union of
spite the 'official' delivery, and the observer may fuzzy sets. These operators are not the only ones
wonder why the ceremony took place at all. The to define intersection and union respectively, but
vagueness of the definition of 'completion' is one the resulting algebraic operations are simple and
of the reasons why we reject probabilistic PERT transparant. The most confusing deficiency of the
models. operations is that dependence is ignored. Ex-
(b) The requirement of tight planning. In Sec- tended, stochastic PERT models do take into
tion 2, we argued that the project leader and the account the dependence due to the underlying
subcontractors need a tight plan for action, al- network. In fuzzy models, however, this is left out
though they ought to (!?) accept a non-tight plan of consideration.
on the basis of a probabilistic model. This leads to In summary, we reject stochastic models in
the peculiar phenomenon that the simple version PERT planning when the activity durations are
of P~RT, with crude approximations to the mo- estimated by human experts. We hesitate to be-
ments of the earliest event times, is acceptable for lieve, however, that the fuzzy arithmetic, in its
planning whereas extended versions with improved present form, should be sufficiently well estab-
approximations are not. This is the second argu- lished to model the vagueness of human judge-
ment against stochastic PERT. ment.
In fact, we are confronted with two types of
uncertainty, which are due to randomness and to
vagueness respectively. Probability theory is con- Acknowledgement
ceptually based on experiments which are fre-
quently repeated, under identical circumstances,
It is a pleasure to acknowledge my students L.
and without mutual dependence. Thus, we throw a
van Bedaf, J. van Dorp, and D. den Hertog for
die and we 'ask' it to show one of its faces, or we
their excellent support and their critical com-
draw a ball from an urn, and we 'ask' it to show
ments.
its colour. The experiment may be simple or com-
plicated, the result may be due to a small or a
large number of interactions, but as soon as the
experiment terminates, the outcome is clear: the References
die shows one of the faces 1, 2 . . . . . 6, the ball
shows the colour red or white. So far, we have Burt, J.M., and Garman, M.B. (1971), "Conditional Monte
seen randomness only. Vagueness enters into the Carlo, a simulation technique for stochastic network analy-
sis", Management Science 18, 207-217.
experiment when the outcome cannot properly be
Chanas, S., and Kamburowski, J. (1981), "The use of fuzzy
observed. This happens, for instance, when we use variables in PERT", Fuzzy Sets and Systems 5, 11-19.
a die with coloured faces in the evening twilight. It Clark, C.E. (1961), "The greatest of a finite set of random
also happens when we ask a human expert to variables", Operations Research 9, 145-162.
estimate the duration of an activity. Even his Cleef, H.J., and Gaul, W. (1982), "Project scheduling via
answer, the outcome of the experiment, will leave stochastic programming", Mathematische Operations-
forschung und Statistik, Series Optimization 13, 449-469.
us in doubt about the state of completion which Dubois, D., and Prade, H. (1978a), "Operations on fuzzy
he has in mind. If two human referees make numbers", International Journal for Systems Sciences 9,
identical (different) statements, their opinions may 613-626.
F.A. Lootsma / Stochastic and fuzzy PERT 183

Dubois, D., and Prade, H. (1978b), "Algorithmes de plus Lootsma, F.A. (1966), "Network planning with stochastic ac-
courts chemins pour traiter des donnres floues", R A I R O / tivity durations, an evaluation of PERT", Statistica Neer-
Recherche Op&ationelle 2, 213-227. landica 20, 43-69.
Eiselt, H.A., and Frajer, H. von (1977), Operations Research Lootsma, F.A. (1969), "A gamma distribution as a model of an
Handbook, Walter de Gruyter, Berlin. activity duration", in: M~thodes d Chemin Critique, Actes
Elmaghraby, S.E. (1967), "On the expected duration of PERT du Congrbs International, Vienne 1967, Dunod, Paris,
type networks", Management Science 5, 299-306. 219-225.
Elmaghraby, S.E. (1977), Activity Networks, Wiley, New York. Lootsma, F.A. (1987), "Modrlisation du jugement humain
Eppen, G.D., Gould, F.J., and Schmidt, C.P. (1987), Introduc- dans l'analyse multicrit+re au moyen de comparaisons par
toO, Management Science, Prentice-Hail, Englewood Cliffs, paires", RAIRO/Recherche Op&ationelle 21,241-257.
NJ. Malcolm, D.G., Roseboom, J.H., Clark, C.E., and Fazar, W.
French, S. (1984), "Fuzzy decision analysis, some criticisms", (1959), "Application of a technique for research and devel-
in: H.J. Zimmermann, L.A. Zadeh and B.R. Gaines (eds.), opment program evaluation", Operations Research 7, 646-
Fuzz)' Sets and Decision Analysis, North-Holland, Amster- 669.
dam. MacGrimmon, K.R., and Ryavec, C.A. (1964), "An analytical
French, S. (1986), Decision Theory, an Introduction. to the study of the PERT assumptions", Operations Research 12,
Mathematics of Rationality, Ellis Horwood, Chichester. 16-37.
French, S. (1987), "Fuzzy sets: The unanswered questions", Martin, J.J. (1965), "Distribution of the time through a directed,
Research report 112/SF/2, Department of Mathematics, acyclic network", Operations Research 13, 46-66.
University of Manchester. Moder, J.D., and Phillips, C.R. (1970), Project Management
Fulkerson, D.R. (1962), "Expected critical path length in PERT with CPM and PI~RT, Van Nostrand-Reinhold, New York.
type networks", Operations Research 10, 808-817. Meilijson, I., and N~tdas, A. (1979), "Convex majorization with
Gaul, W. (1981), "Bounds for the expected duration of a an application to the length of critical paths", Journal of
stochastic project planning model", Journal of Information Applied Probability 16, 671-677.
and Optimization Sciences 2, 45-63. N~das, A. (1979), "Probabilistic PERT", IBM Journal of Re-
Hartley, H.O., and Wortham, A.W. (1966), "A statistical the- search and Development 23, 239-258.
ory for PERT critical path analysis", Management Science PERT Summary Report (1958), Bureau of Ordnance, Depart-
12, 469-481. ment of the Navy, Washington, DC.
Hillier, F.S., and Liebermann, G. (1974), Operations Research, Ravindran, A., Phillips, D.T., and Solberg, J.J. (1987), Oper-
Holden-Day, San Francisco, CA. ations Research, Principles and Practice, Wiley, New York.
Kleindorfer, G.B. (1971), "Bounding distributions for a sto- RobiUard, P., and Traham, M. (1977), "'The completion time
chastic, acyclic network", Operations Research 19, 1586- of PERT networks", Operations Research 25, 15-29.
1601. Shogan, A.W. (1977), "Bounding distributions for a stochastic
Klein Haneveld, W.K. (1986), "Robustness against depen- PERT network", Networks 7, 359-381.
dence in PERT: An application of duality and distributions Smith, D.K. (1982), Network Optimization Practice, Ellis
with known marginals", Mathematical Programming Study Horwood, Chichester.
27, 153-182. Wagner, H.M. (1975), Principles of Operations Research, Pren-
Laarhoven, P. van, and Pedrycz, W. (1983), "A fuzzy extension tice-Hall, Englewood Cliffs, NJ.
of Saaty's priority theory", Fuzzy Sets and Systems 11, Zadeh, L.A. (1965), "Fuzzy sets", Information and Control 8,
229-241. 338-353.

You might also like