Computers & Industrial Engineering 42 (2002) 521±531

www.elsevier.com/locate/dsw

Ef®cient Monte Carlo simulation method of GERT-type network

for project management
Kenzo Kurihara*, Nobuyuki Nishiuchi
Department of Industrial Engineering and Management, Kanagawa University, 3-27-1 RokkakubashI, Kanagawa-ku,
Yokohama 221-8686, Japan

Abstract
This paper proposes a new method to estimate network characteristics such as project time, project cost, etc.
There are many kinds of projects such as new product development and plant construction. Since these projects
consist of various uncertain jobs, we can model those processes as stochastic networks such as GERT network.
Only simulation technique can analyze the networks' characteristics of practical project model. However, the
method is extremely time-consuming. Our method can estimate the network characteristics rapidly and accurately
by analyzing one set of Monte Carlo simulation results based on the theory of probability. q 2002 Elsevier
Science Ltd. All rights reserved.

Keywords: Monte Carlo simulation; Project management; GERT network; Sensitivity analysis

1. Introduction

Projects such as new product development and plant construction have the similar uncertainty about
the job sequence/duration of the project. Because the job sequence is usually represented by a semi-order
relationship, it is dif®cult to determine the sequence in advance. These jobs are often performed in a
complex order that is determined as the job proceeds. For example, jobs may be done repeatedly, when
mistakes are found. In many cases, the duration of each job should be treated as stochastic variables. To
manage the projects effectively, we should ®nd the main problems of the projects. To ®nd and solve the
problems, it is important to analyze the in¯uences on network characteristics caused by the change in
network parameter values such as branching probability and duration of job execution. In the case of
general stochastic network, analytical treatment is dif®cult because of the network complexity. Simula-
tion technique is practical in realistic networks (Moore & Clayton, 1976; Pritsker, 1966).

* Corresponding author.
E-mail address: [email protected] (K. Kurihara).

0360-8352/02/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved.
PII: S0 3 6 0 - 8 3 5 2 ( 0 2 ) 0 0 05 0 - 5
522 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531

This paper proposes a GERT-type network analyzing method. This method can estimate ef®ciently
the characteristic changes caused by branching probability changes, and also estimates the sensitivity to
the network characteristics. By processing one set of Monte Carlo simulation results, we do not need to
repeat the Monte Carlo simulation. Speci®cally speaking, we can estimate the in¯uence of branching
probabilities on the network characteristics. In our research, we observed the branching probability that
should be improved to satisfy both the required time and budget of the project.

2. Project management problem

2.1. Research history

We usually represent project processes as stochastic networks, such as GERT, when we want to
manage and control projects. In the case of restricted simple GERT networks, in which all nodes are
of the `Exclusive-OR' type, the problem can be solved analytically (Pritsker, 1966). In realistic projects,
we must analyze generalized complex GERT networks including `AND' and `OR' nodes. In this case,
analytical treatment is dif®cult because of the network complexity. Simulation is a practical technique
for complex GERT networks.
By Monte Carlo simulation, we can easily calculate network characteristics under the given parameter
values. However, sensitivity analysis through the simulation is dif®cult, because it is time-consuming;
®rst we have to repeat simulations under various parameter values, then we analyze the parameter
sensitivity based on the difference of the results. This sensitivity analysis is essential to manage the
projects effectively. Based on the sensitivity, we can adjust network parameters such as branching
probability and job duration, and optimize the network characteristics such as project duration and
budget requirement.
In our former research, we presented an analytical approach combined with simulation technique to
overcome these problems (Kurihara, Seki, & Akashi, 1984). We picked up job duration as control
variables, and estimated network characteristics and parameter sensitivity under various control values.
This method was applied to the problem of optimally investing a ®xed sum of money to achieve a
maximal reduction in project time.

2.2. Project process analyzing problem related to branching probability

Another main control parameter for project management is branching probability, which is deter-
mined by the result of each job. Our research proposes a new estimation method of network character-
istics with regard to this branching probability. The method can estimate the network characteristics
rapidly and accurately by analyzing one set of Monte Carlo simulation results.
Now, we think about a new product development process. In this kind of project, we apply up-to-date
yet still experimental engineering ingenuity. Failure ratio of the project is moderate, and it results in
re-designing, re-producing and re-testing. When we analyze project processes as stochastic network
models, we sometimes encounter problems on how to get branching probabilities for all arrows in the
network model. We can estimate probabilities based on the experiences and actual results. For example,
for a new logic semiconductor LSI developing project, we can obtain actual results of many LSIs.
As branching probabilities are decided subjectively, we cannot rely on the absolute numerical values
 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531 523

Fig. 1. GERT network

of the results themselves. However, for the purpose of planning and managing projects, it proves
informative to analyze the relationships between branching probabilities and network characteristics
such as the project time and the probability successfully ®nishing. Sensitivity analysis on branching
probabilities is also essential for the improvement of network characteristics.
Though it is almost impossible to change a certain branching probability from exactly 0.2 to 0.28, we
recognize how effective each change is for the project. We can make an appropriate management plan on
manpower requirements and engineering skill. To a certain extent, the branching probabilities can be
controlled by the plan. For example, if there are many parallel projects, and if some of the signi®cant
projects must be ®nished successfully, experienced engineers should be allocated properly to those
projects. Project managers understand the key points and can decide the strategies for project manage-
ment. Our method can give the managers adequate operational policies mentioned earlier.

3. Project process model

3.1. Stochastic network

The GERT network example is shown in Fig. 1. The arrow number and the branching probability are
described beside each arrow. The dotted arrows represent `modi®cation function'. This function can
modify the structure of the network based on the processing situation by substituting a node for another.
There are two types of nodes, the deterministic node with round tail and the stochastic node with
triangular tail. When a deterministic node is attained, all arrows starting from the node will begin.
On the contrary, in the case of a stochastic node, only one of the arrows will be triggered.
As shown in Fig. 2, there are two types of nodes in the GERT network, and they have three

Fig. 2. Deterministic node, stochastic node.

524 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531

parameters. The right-hand number in each node represents the arrow number. The left-hand and upper
side number is the initial number of jobs required to release this node, while the lower side number is the
subsequent number. When the left-hand and upper side number is 1, the node is an OR-node. If the
number is more than 2, the node is an AND-node.

3.2. A new product development project

The following is a standard model of a new product development project. It consists of the following
three groups of jobs.

(1) Design jobs. Two jobs (arrows 2, 3) can be performed simultaneously. After the jobs are ®nished,
proceed to trial production jobs. In some situations, the development may be terminated.
± Proceed to trial production (to node 5); probability ˆ 0.9.
± Terminate the development (to node 12); probability ˆ 0.1.

Design jobs may be performed again, if more than one of the other group jobs fail. At this time, the
duration of each job (arrows 10 and 11) is shorter than the initial one. This situation is modeled by
`modi®cation (Pritsker, 1966)' that substitutes node 3 for node 10 in the case of re-designing.

(2) Trial production jobs (arrows 5, 6). Two jobs can be performed simultaneously. When one job is
®nished, the other job stops. Then, proceed to evaluation jobs. In some cases, design failures may be
found and design jobs are performed again.
± Proceed to evaluation (to node 7); probability ˆ 0.6.
± Proceed to re-design (to node 10); probability ˆ 0.4.

(3) Evaluation job (arrow 8). Trial production result is evaluated. When it is good, the development
will be ®nished successfully. Otherwise, one of the following decisions is selected.
± Finish the development in success (to node 9); probability ˆ 0.5.
± Quit the development in failure (to node 12); probability ˆ 0.1.
± Proceed to re-evaluation (to node 7); probability ˆ 0.1.
± Proceed to re-trial production (to node 5); probability ˆ 0.1.
± Proceed to re-design (to node 10); probability ˆ 0.2.

The problem considered here is to look for the most effective branch for project management by
clarifying `the relationship between the branching probabilities and the project characteristics, such as
project time and probability of success'.

4. Estimation method of network characteristics

Symbols

p0k : a branching probability of arrow k as the simulation condition;

P0 ˆ p01 ; p02 ; ¼; p0k ; ¼; p0K † : a set of branching probability p0k of each arrow as the simulation condition;
 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531 525

pk : a branching probability of arrow k as the estimation condition;

P ˆ p1 ; p2 ; ¼; pk ; ¼; pk † : a set of branching probability pk of each arrow as the estimation condition;
Ha0k : executed number of arrow k at the a th trial.
H0a ˆ Ha1 0 0
; Ha2 ; ¼; Ha0k ; ¼; Ha0K † : a set of the arrow branching history;
Xa : value of output variable X at the a th trial. Project characteristics, such as project duration time and
cost, can be de®ned as the output variable X.

4.1. Expectation X and its sensitivity analysis

First, we perform Monte Carlo simulation, whose trial number is N, for this network under the given
condition of branching probabilities P0 : As a result, we can obtain the output values Xa a ˆ
1; 2; 3; ¼; N†: Then, we estimate the result under another condition that the branching probabilities
are changed from P0 to arbitrary P. In this case, the occurring probability of history Ha0k is calculated
by the following equation:
   
Prob Ha0k P† ˆ Prob Ha0k P0 †´f P; P0 ; H0a 1†

where
!H 0
  YK
pk ak

f P; P0 ; H0a ˆ 2†
kˆ1 p0k

The ®rst term, Prob Ha0k uP0 †; in the right-hand side of Eq. (1) shows `the probability of getting arrow k`s
execution number, Ha0k ; at the a th trial under the condition of branching probability set P 0'. On the other
hand, the left-hand side, Prob Ha0 k uP0 †; shows `the probability of getting the same execution number,
Ha0k ; under the condition of another branching probability set P'. The second term, f P; P0 ; H0a † in the
right-hand side represents the ratio of the above two probabilities. Therefore, if we can calculate the ratio
and if we have the practical simulation results under the original branching probability set P0 ; we can
estimate the simulation results under an arbitrary branching probability set P That is because we can
consider that the history Ha0k will occur f P; P0 ; H0a † times under the new condition.
The ratio f P; P0 ; H0a † can be calculated by Eq. (2), because the denominator of the right hand shows
the probability of getting the history under the original condition P0 ; and the numerator shows the
probability of getting the same history under the new condition P.
Eq. (1) can be understood consequently that each trial result of the original simulation will occur
f P; P0 ; H0a † times under the new condition P 0. Although the same trial results might appear several
times in the original simulation, we can treat them independently. When all f P; P0 ; H0a † are summarized
for all trial results, we can estimate a set of simulation results under an arbitrary condition without
executing new simulation again.
Using this relationship, we can estimate the cumulative distribution function of X, F X; P†; under
arbitrary condition of P by the following equation:

1 X N  
0 0
F X; P† ù f P; P ; H a S X; Xa † 3†
N p aˆ1
526 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531

where
(
1 X $ Xa
S X; Xa † ˆ 4†
0 X , Xa

X
N  
Np ˆ f P; P0 ; H0a 5†
aˆ1

As shown in Eq. (4), we introduce a new function, S X; Xa †: According to the de®nition of the
cumulative distribution function, we can estimate the function F X; P† by summarizing every
f P; P0 ; H0a † that is multiplied by S X; Xa †:
In the same way, if S X; Xa † is replaced by g X†; we can get the expectation of g X† by the following
equation:

1 X N  
0 0
Ex g X†; P† ˆ f P; P ; H a g Xa † 6†
N p aˆ1

In the case of g X† ˆ X and when X is taken as the project time, the expected project time under arbitrary
condition of P is given by the following equation:

1 X N  
0 0
Ex X; P† ˆ f P; P ; H a Xa 6a†
N p aˆ1

On the other hand, parameter sensitivity for expectation g X† can be estimated by the partial differential
of Eqs. (3) and (6) with respect to pk : We can examine the in¯uence to the network characteristics caused
by the branching probability change. If there are more than three arrows from one node, the changed
portion of the arrow ks branching probability will be distributed to the remaining arrows in proportion to
the original probabilities.
Let express the start node of arrow k in terms of subscript `n', and let de®ne bn as a set of arrows,
whose start node is n, except for k. The total executed number Ha0k of arrows in bn can be calculated by
the following equation:
X 0
Ha0k ˆ Hab 7†
b[bn

Using this equation, Eq. (2) is transformed into the following form:
!H 0 !H 0
  p0k ak
1 2 p0k ak
0
f P; P ; H0a ˆ 2a†
pk 1 2 pk

By substituting Eq. (2a) in Eqs. (3) and (6), we can estimate parameter sensitivity for the network
characteristics, 2F X; P†=2pk and 2Ex g X†; P†=2pk ; by the following equations:
( )
2 2 1 X N  
F X; P† ˆ f P; P0 ; H0a S X; Xa † 8†
2pk pk ˆp0k 2pk N p aˆ1
 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531 527
( )
2 2 1 X N  
0 0
Ex g X†; P† ˆ f P; P ; Ha g Xa † 9†
2pk 0
pˆpk 2p k N p aˆ1

Here, we introduce f 0 de®ned by the following equation:

2   H0 Ha0k
f0 ù f P; P0 ; H0a ˆ a0k 2 10†
2pk pk ˆp0k pk 1 2 p0k

In the case of g X† ˆ X; sensitivity for expectation X can be estimated by the following equation. To
simplify this equation, we apply the relation f P; P0 ; H0a †upk ˆp0k ˆ 1 :

2 1 XN
1 X
N XN
Ex X; P† ˆ f 0 Xa † 2 2 f 0 †´ Xa 9a†
2pk pk ˆp0k N aˆ1 N aˆ1 aˆ1

By processing one set of Monte Carlo simulation results by the above equations, we can analyze network
characteristics ef®ciently. Therefore, it is not necessary to repeat Monte Carlo simulation. That is to say
that, by substituting simulation results {X1 ; X2 ; X3 ; ¼; Xa ; ¼; XN } and branching probabilities before and
after into Eqs. (3) and (6), we can obtain network characteristics under changed branching probabilities.
We can treat various values as the network characteristics, such as the project time, the project cost, the
number of each arrow's performance and the project probability of success.
The merit of our method is not only to decrease simulation trial numbers, but also to make it possible
to analyze the sensitivities. We can estimate the in¯uences of branching probability changes on the
network characteristics by substituting simulation results in Eqs. (8) and (9a).
In the proposed method, the variance of expectation X and its sensitivity are given by substituting the
following equation in Eqs. (6) and (9), respectively:
!2
1 X
N  
g Xa † ˆ Xa 2 p f P; P0 ; H0a Xa 11†
N aˆ1

The variance is given by

8 !2 9
XN < X
N     =
1 1
V X; P† ˆ p Xa 2 p f P; P0 ; H0a Xa f P; P0 ; H0a 6b†
N aˆ1 : N aˆ1 ;

In the same manner, the sensitivity for variance is represented by the following equation:
2 8 ! 93
2 2 4 1 <X N     1 N  
X   2=
5
V X; P†† ˆ p: f P; P0 ; H0a Xa2 2 p f P; P0 ; H0a Xa
2pk pk ˆp0 2p k N aˆ1
N aˆ1
;
k
pk ˆp0k

9b†
528 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531

Table 1
Input data for arrows

Arrow no. Mean Min. Max. Std dev.

2 30.0 15.0 40.0 5.0

3 25.0 10.0 50.0 10.0
5 5.0 2.0 6.0 1.0
6 5.0 2.0 6.0 1.0
8 3.0 2.0 6.0 2.0
10 5.0 2.0 10.0 2.0
11 4.0 1.0 8.0 2.0

5. Numerical results

5.1. Example model

A new product development project (Fig. 1) is assumed to evaluate the proposed method. The input
data for arrow parameters are shown in Table 1. In this example, it is assumed that the each probability
density function shows a normal distribution.

5.2. Estimation of network characteristics

Monte Carlo simulation experiment was carried out with 50,000 trials under the parameters of Table
1. Considering the simulation results, we estimate the network characteristics on condition that the
branching probability of arrow 16 has been changed from 0.20 to 0.28.
All new product development projects do not always result in success. It is useful to know the
probability of failure ratio. We also want to investigate which branching probability is the most vital
to decrease the probability of failure. Fig. 3 shows the probability chart for complete or incomplete
project. The ®gure shows three results; the original simulation results, the estimation result by our
method and the re-simulation results under the new branching probability. Estimated result shows
that the succeeding probability will drop 4% by enlargement of the branching probability of arrow
16. Fig. 4 shows the expected project times for both cases of success and failure. Estimated result shows
that the expected project time will extend by 1.8 days because of the change in branching probability.

Fig. 3. Probability of success/failure.

 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531 529

Fig. 4. Project time.

These estimation results are consistent with the simulation results under the changed branching
probability.

5.3. Estimation of sensitivity for network characteristics

We evaluated the sensitivity for expected project time T and for expected project success probability S
(Figs. 5 and 6). Since the real sensitivity cannot be obtained in this evaluation, we compare the results of
our method with another estimated sensitivity, dT=dpk and dS=dpk : dT and dS represent the difference
between the simulation results, T and S, under different parameter values, and dpk is the difference in the
parameter values for the simulations. In the ®gures, the vertical axis represents the sensitivities,
estimated by the proposed method Eq. (9a), and the horizontal axis represents dT=dpk and dS=dpk :

Fig. 5. Sensitivity (project time).

Fig. 6. Sensitivity (success probability).

530 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531

Fig. 7. Convergence of the solutions (proposed method).

As shown in the ®gure, all experimental data points are approximately on a straight line inclined at
458, and errors from the line are within 10% in the entire measured results. Unfortunately, we cannot
logically explain whether the 10% difference is suf®cient or not. We can only show that the magnitude
order of the 8 data is almost the same in two axes. (There is only one inversion as shown in Fig. 5. As the
both are close, no signi®cant problems exist from viewpoint of realistic usage.)
In this experiment, we investigated the sensitivities for eight arrows. The proposed method required
only one set of Monte Carlo simulation under original branching probabilities, though the simulation
repetition method needed nine sets.

5.4. Convergence of the solutions

During the course of these numerical evaluations, we examined the convergence of the project time.
In general, the convergence performance depends
p on the network complexity. The accuracy of Monte
Carlo simulation results is proportional to 1= N ; where N represents the trial number of the simulation.
In the above-mentioned example, after 50,000 trials, we can obtain converged results of Fig. 4. Figs. 7
and 8 show the convergence toward the result in Fig. 4.
In this example, the Monte Carlo simulations are performed for ®ve series of random variables. In
Fig. 7, white circle plots represent the original simulation results and black circle plots represent
estimated results that are calculated by the proposed method Eq. (9a) under the changed condition of
branching probabilities. Fig. 8 shows the results of the Monte Carlo simulation.
As shown in these ®gures, an accuracy of the numerical result is good enough by 10,000 simulation
trials. At this time, we can get the following results.

² Original simulation results: Average Ex(X) ˆ 50.94, standard deviation ˆ 0.18.

² Proposed method results: Average Ex(X) ˆ 52.71, standard deviation ˆ 0.25.
² Re-simulation results: Average Ex(X) ˆ 52.94, standard deviation ˆ 0.25.

According to this numerical example, the convergence property of the proposed method is as good as
the original Monte Carlo simulation.

6. Conclusion

We proposed a new analysis method for project process management. By processing one set of Monte
Carlo simulation results by our method, we can analyze network characteristics ef®ciently. Therefore, it
is not necessary to repeat the Monte Carlo simulation. Judging from the numerical evaluations, the
 K. Kurihara, N. Nishiuchi / Computers & Industrial Engineering 42 (2002) 521±531 531

Fig. 8. Convergence of the solutions (re-simulations).

method has suf®cient accuracy and calculation speed comparable to the original Monte Carlo simulation
performance. The convergence of the solution is also the same as the simulation result.

Acknowledgements

We express our sincere thanks to Mr Seki, former chief researcher at Hitachi Company, System
Development Laboratory, who gave us invaluable time and advice for this study.

References

Kurihara, K., Seki, S., & Akashi, K. (1984). An optimal method for investment in a project expressed by GERT network. The
Transaction of The Institute of Electrical Engineers of Japan, 104 (3), 9±16 in Japanese.
Moore, L. J. & Clayton, E. R. (1976). GERT modeling and simulationÐFundamentals and applications: PETROCELL
CHAPTER.
Pritsker, A. A. B. (1966). GERT graphical evaluation and review technique, Memorandom RM-4973-NASA, Rand
Corporation.