Applied Statictics and Operation Research

Computer Modelling and New Technologies, 2010, Vol.14, No.4, 14–18

Transport and Telecommunication Institute, Lomonosova 1, LV-1019, Riga, Latvia

BETA-DISTRIBUTION MODELS IN STOCHASTIC

PROJECT MANAGEMENT

D. Greenberg1, A. Ben-Yair2
1
Department of Economics and Business Administration, Faculty of Social Science
Ariel University Center (AUC) of Samaria
P.O. Box 3, Ariel, 40700, Israel
E-mail: [email protected]
2
The Department of Industrial Engineering and Management, SCE-Shamoon College of Engineering
Beer-Sheva 84100, Israel
E-mail: [email protected]

A research is undertaken to justify the use of beta-distribution p.d.f. for man-machine type activities under random
disturbances. The case of using one processor, i.e., a single resource unit, is examined. It can be proven theoretically that under
certain realistic assumptions the random activity – time distribution satisfies the beta p.d.f. Changing more or less the implemented
assumptions, we may alter to a certain extent the structure of the p.d.f. At the same time, its essential features (e.g. asymmetry,
unimodality, etc.) remain unchanged. The outlined above research can be applied to semi-automated activities, where the presence
of man-machine influence under random disturbances is, indeed, very essential. Those activities are likely to be considered in
organization systems (e.g. in project management), but not in fully automated plants.

Keywords: random activity duration, time – activity beta-distribution, operating by means of a single processor, convergence
to a beta-distribution “family”

1. Introduction

In PERT analysis [1–24, etc.] the activity-time distribution is assumed to be a beta-distribution,

and the mean value and variance of the activity time are estimated on the basis of the “optimistic”, “most
likely” and “pessimistic” completion times, which are subjectively determined by an analyst. The creators
of PERT [3, 17] worked out the basic concepts of PERT analysis, and suggested the estimates of
the mean and variance values

μ =
1
( a + 4m + b ) , (1)
6

σ2 =
1
( b − a )2 , (2)
36
subject to the assumption that the probability density function (p.d.f.) of the activity time is

Γ ( α + β ) ( y − a )α −1 ( b − y )β −1
f y ( y) = , a < y< b , α, β > 0 . (3)
Γ (α ) Γ (β ) ( b − a )α +β −1
Here a is the optimistic time, b – the pessimistic time, and m stands for the most likely (modal)
time.
Since in PERT applications parameters a and b of p.d.f. (3) are either known or subjectively
determined, we can always transform the density function to a standard form,
Γ ( α + β ) α −1
f ( x) = x ( 1 − x )β −1 , 0 < x < 1 , α , β > 0 , (4)
Γ (α ) Γ (β )

y−a
where x = has the following parameters:
b−a

14
 Applied Statictics and Operation Research
μy − a σy my − a
μx = , σx = , mx = . (5)
b−a b−a b−a

Let α −1 = p , β −1 = q . Then p.d.f. (4) becomes

Γ ( p + q + 2)
f ( x) = x p ( 1 − x )q , 0 < x < 1 , p, q > − 1 , (6)
Γ ( p + 1) Γ ( q +1)
with the mean, variance and mode as follows:
p +1
μx = , (7)
p+q+2

σ x2 =
( p + 1)( q + 1) , (8)
p+q+2

p
mx = . (9)
p+q

From (6) and (9) it can be obtained

Γ ( p + q + 2)
f ( x) = x p ( 1 − x )p(1 m x −1) . (10)
Γ ( p + 1) Γ ( q + 1)

Thus, value m x , being obtained from the analyst’s subjective knowledge, indicates the density
function. On the basis of statistical analysis and some other intuitive arguments, the creators of PERT
assumed that p + q ≅ 4 . It is from that assertion that estimates (1) and (2) were finally obtained,
according to (6–9).
Although the basic concepts of PERT analysis have been worked out many years ago [3, 17], they
are open till now to considerable criticism. Numerous attempts have been made to improve the main
PERT assumptions for calculating the mean μ x and variance σ x2 of the activity-time on the basis of the
analyst’s subjective estimates. In recent years, a very sharp discussion [7, 10, 14, 21] has taken place in
order to raise the level of theoretical justifications for estimates (1) and (2).
Grubbs [12] pointed out the lack of theoretical justification and the unavoidable defects of the
PERT statements, since estimates (1) and (2) are, indeed, “rough” and cannot be obtained from (3) on the
basis of values a , m and b determined by the analyst. Moder [18–19] noted that there is a tendency to
choose the most likely activity – time m much closer to the optimistic value a than to the pessimistic
one, b , since the latter is usually difficult to determine and thus is taken conservatively large. Moreover,
it is shown [8] that value m , being subjectively determined, has approximately one and the same relative
location point in [ a , b] for different activities. This provides an opportunity to simplify the PERT
analysis at the expense of some additional assumptions. McCrimmon and Ryavec [16], Lukaszewicz [15]
and Welsh [22] examined various errors introduced by the PERT assumptions, and came to the
conclusion that these errors may be as great as 33%. Murray [20] and Donaldson [4] suggested some
modifications of the PERT analysis, but the main contradictions nevertheless remained. Farnum and
Stanton [6] presented an interesting improvement of estimates (1) and (2) for cases when the modal value
m is close to the upper or lower limits of the distribution. This modification, however, makes the
distribution law rather uncertain, and causes substantial difficulties to simulate the activity network.
In this paper, a research will be undertaken to develop some theoretical justifications for using
the beta-distribution p.d.f.

2. The Operation’s Description

We will consider a man-machine operation which is carried out by one processor, i.e., by one
resource unit. The processor may be a machine, a proving ground, a department in a design office, etc.
Assume that the operation starts to be processed at a pregiven moment T0 . The completion
moment F of the operation is a random value with distribution range [ T1 , T2 ] . Moment T1 is the operation’s

15
 Applied Statictics and Operation Research
completion moment on condition that the operation will be processed without breaks and without delays,
i.e., value T1 is a pregiven deterministic value. Assume, further, that the interval [ T0 , T1 ] is subdivided
into n equal elementary periods with length ( T1 − T0 ) n . If within the first elementary period
[ T0 , T0 + ( T1 − T0 ) n ] a break occurs, it causes a delay of length Δ = ( T2 − T1 ) n . The operation
stops to be processed within the period of delay in order to undertake necessary refinements, and
later on proceeds functioning with the finishing time of the first elementary period
T0 + ( T1 − T0 ) n + ( T2 − T1 ) n = T0 + ( T2 − T0 ) n .
It is assumed that there cannot be more than one break in each elementary period. The probability
of a break at the very beginning of the operation is set to be p . However, in the course of carrying out
the operation, the latter possesses certain features of self-adaptivity, as follows:
• the occurrence of a break within a certain elementary period results in increasing the probability of
a new break at the next period by value η , and
• on the contrary, the absence of a break within a certain period decreases the probability of a new
break within the next period, practically by the same value.

3. The Concept of Self-Adaptivity

The probabilistic self-adaptivity can be formalized as follows:

Denote A ik the event of occurrence of a break within the (i + 1) -th elementary period, on
condition, that within the i preceding elementary periods k breaks occurred, 1≤ k ≤ i ≤ n . It is assumed
that relation

( )
P A ik =
p + k ⋅η
1 + i ⋅η
(11)

holds. Note that (11) is, indeed, a realistic assumption.

( )
Relation (11) enables obtaining an important assertion. Let P A i0 be the probability of the
occurrence of a break within the (i + 1) -th period on condition, that there have been no breaks at all as
yet. Since

( )
P A i0 =
p
1 + i⋅p
, (12)

it can be well-recognized that relation

( ) ( )
P A ik +1 − P A ik
=
η
P(A ) 0
(13)
i p

holds. Thus, an assertion can be formulated as follows:

Assertion. Self-adaptivity (11) results in a probability law for delays with a constant ratio (13)
for a single delay.

4. Calculating the Activity-Time Distribution

Let us calculate the probability Pm ,n of obtaining m delays within n elementary periods, i.e.,
m
the probability of completing the operation at the moment F = T1 + m ⋅ Δ = T1 + ( T2 − T1 ) .
n
The number of sequences of n elements with m delays within the period [ T0 , F] is equal C m
n ,
while the probability of each such sequence equals

16
 Applied Statictics and Operation Research

⎡ m −1
( p + iη )⎤⎥ ⎡⎢ ∏ ( 1 − η + iη )⎤⎥
n − m −1
⎢ i∏
⎣ =0 ⎦ ⎣ i =0 ⎦ . (14)
n −1
∏ ( 1 + iη )
i =0

Relation (14) stems from the fact that if breaks occurred within h periods and did not occur
within k periods, the probability of the occurrence of the delay at the next period is equal
p + hη
, (15)
1 + ( k + h )η

while the probability of the delay’s non-appearance at the next period satisfies
1 − η + kη
. (16)
1 + ( k + h )η

Using (14–16), we finally obtain

⎡ m −1
( p + iη )⎤⎥ ⎡⎢ ∏ ( 1 − η + kη )⎤⎥
n − m −1
⎢ i∏
Pm, n = Cm ⎣ =0 ⎦ ⎣ i =0 ⎦ . (17)
n n −1
∏ ( 1 + iη )
i =0

Note that η = 0 , i.e., the absence of self-adaptivity, results in a regular binomial distribution.
Let us now obtain the limit value Pm ,n on condition that n → ∞ . From relation (17) we obtain

Pm +1, n n−m p + mη
= . (18)
Pm, n m + 1 1 − p + ( n − m − 1)η

p p⎛1 ⎞
Denoting =α, ⎜ −1⎟⎟ = β , we obtain
η η ⎜⎝ p ⎠
1− β
Pm +1, n − Pm, n (α −1) n + (2 − α − β ) m − β +1
(α −1) + (2 − α − β ) m +
= = n n .
Pm, n (m + 1)(β + n − m −1) n
m +1⎛
⎜1 −
m +1 β ⎞
+ ⎟
n ⎝ n n⎠
Denoting m n = x , (m + 1) n = x + Δx , Pm, n = y , Pm +1, n = y + Δy , via convergence n → ∞
or Δx → 0 and, later on, by means of integration, we finally obtain
y = C xα −1 (1 − x )β −1 . (19)

m
It can be well-recognized that the p.d.f. of random value ξ = lim satisfies
n →∞ n
pξ (x ) = xα −1 (1 − x )β −1 ,
1
(20)
B(α , β )

where B(α , β ) represents the Eiler’s function. Thus, relation (20) practically coincides with (10).
Thus, ξ is a random value with the beta-distribution activity – time p.d.f. By transforming
x = (y − a ) (b − a ) , we obtain the well-known p.d.f. (3).

Conclusions
The following conclusions can be drawn from the study:
• Under certain realistic assumptions we have proven theoretically that the activity-time distribution
satisfies the beta-distribution with p.d.f. (3) being used in PERT analysis.

17
 Applied Statictics and Operation Research
• Changing more or less the implemented assumptions, we may alter to a certain extent the structure of
the p.d.f. At the same time, its essential features (e.g. asymmetry, unimodality, etc.) remain
unchanged.
• The outlined above research can be applied to semi-automated activities, where the presence of man-
machine influence under random disturbances is, indeed, very essential. Those activities are likely to
be considered in organization systems (e.g. in project management), but not in fully automated plants.

References
1. Battersby, A. Network Analysis for Planning and Scheduling, 3rd edition. London: MacMillan, 1970.
2. Berny, J. A new distribution functions for risk analysis, J. Oper. Res. Soc., Vol. 40, No 12, 1989,
pp. 1121–1127.
3. Clark, C. E. The PERT model for the distribution of an activity, Opns. Res., Vol. 10, 1962, pp. 405–406.
4. Donaldson, W. A. The estimation of the mean and variance of PERT activity time, Opns. Res.,
Vol. 13, 1965, pp. 382–385.
5. Elmaghraby, S. E. Activity Networks: Project Planning and Control by Network Models. New-York:
Wiley, 1977.
6. Farnum, N. R. and L. W. Stanton. Some results concerning the estimation of beta distribution
parameters in PERT, J. Oper. Res. Soc., Vol. 38, 1987, pp. 287–290.
7. Gallagher, C. A note on PERT assumptions, Mgmt. Sci., Vol. 33, 1987, pp. 1360–1362.
8. Golenko-Ginzburg, D. Statistical Models in Network Planning and Control, Moscow: Nauka, 1966.
(In Russian)
9. Golenko-Ginzburg, D. On the distribution of activity time in PERT, J. Oper. Res. Soc., Vol. 39, No 8,
1988, pp. 767–771.
10. Golenko-Ginzburg, D. PERT assumptions revisited, Omega, Vol. 17, No 4, 1989, pp. 393–396.
11. Gonik, A. Planning and Controlling Multilevel Stochastic Projects: Ph.D. Thesis. Beer-Sheva:
Ben-Gurion University of the Negev, 1995.
12. Grubbs, F. E. Attempts to validate certain PERT statistics or “picking on PERT”, Opns. Res., Vol. 10,
1962, pp. 912–915.
13. Kelley, J. E., Jr. Critical path planning and scheduling: Mathematical basis, Opns. Res., Vol. 9, No 3,
1961, pp. 296–320.
14. Littlefield, T. K., Jr. and P. H. Randolph. Another note on PERT times, Mgmt. Sci., Vol. 33, 1987,
pp. 1357–1359.
15. Lukaszevicz, J. On the estimation of errors introduced by standard assumptions concerning the distribution
of activity duration PERT calculations, Opns. Res., Vol. 13, 1965, pp. 326–327.
16. MacCrimmon, K. R. and C. A. Ryaveck. An analytical study of the PERT assumptions, Opns. Res.,
Vol. 12, 1964, pp. 16–37.
17. Malcolm, D., Roseboom, J., Clark, C. and W. Fazar. Application of a technique for research and
development program evaluation, Opns. Res., Vol. 7, 1959, pp. 646–669.
18. Moder, J. J., Phillips, C. R., Davis, E. W. Project Management with CPM and PERT and Precedence
Diagramming. New-York: Van-Nostrand Reinhold Co., Inc., 1983.
19. Moder, J. J., Cecil, R. P. Project Management with CPM and PERT. New-York: Van-Nostrand
Reinhold Co., Inc., 1970.
20. Murray, J. E. Consideration of PERT Assumptions. Ann Arbour, Michigan: Conduction Corporation,
1962.
21. Sasieni, M. W. A note on PERT times, Mgmt. Sci., Vol. 32, 1986, pp. 1652–1653.
22. Welsh, D. Errors introduced by a PERT assumption, Opns. Res., Vol. 13, 1965, pp. 141–143.
23. Williams, T. M. Practical use of distributions in network analysis, J. Oper. Res. Soc., Vol. 43, No 3,
1992, pp. 265–270.
24. Williams, T. M. What are PERT estimates? J. Oper. Res. Soc., Vol. 46, No 12, 1995, pp. 1498–1504.

Received on the 21st of June, 2010

18