Expert Systems with Applications 36 (2009) 10004–10015

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

Using Ant Colony Optimization algorithm for solving project management problems
Hazem Abdallah *, Hassan M. Emara, Hassan T. Dorrah, Ahmed Bahgat
Cairo University, Automatic Control, 29 Eshreeny Street, Faculty of Engineering, P.O. Box 12613, Cairo, Makour Faysl, Giza 12111, Egypt

a r t i c l e i n f o a b s t r a c t

Keywords: Network analysis provides an effective practical system for planning and controlling large projects in con-
Ant Colony Optimization struction and many other fields. Ant Colony System is a recent approach used for solving path minimiza-
PERT tion problems. This paper presents the use of Ant Colony Optimization (ACO) system for solving and
Project management calculating both deterministic and probabilistic CPM/PERT networks. The proposed method is investi-
CPM
gated for a selected case study in construction management. The results demonstrate that – compared
to conventional methods – ACO can produce good optimal and suboptimal solutions.
Ó 2008 Elsevier Ltd. All rights reserved.

1. Introduction 2. Network analysis

Effective project management techniques are important to en-
sure successful project performance; a poor strategy can easily
turn expected profit into loss. With the availability of computer
facilities the design, calculation, modeling, managing and checking
processes and projects can be done in a more efficient and effective
manner. The management of construction project involves plan-
ning of tasks from large numbers of disciplines which require dif-
ferent pieces of information at various times. This results in the
production of a huge quantity of complex information, which must
be managed efficiently. Network analysis provides a comprehen-
sive practical system for planning and controlling large projects
in construction and many other fields. One of the most needed
tasks is to accomplish a forecast of optimal and suboptimal paths
of the network for construction management due to the complex-
ity of the project and the possibility of crash or delay occurrence
which is not so easy with conventional methods. The integration
of optimization algorithms based on metaheuristic opens new per-
spectives of applications in real life. Ant Colony System has been
introduced in the early 1990s. It mimics the performance of natural
ants while searching food and finding the shortest path between
the nest and the food source thanks to local message exchange
(Bonabeau, Dorigo, & Theraulaz, 1999).
This paper proposes the use of Ant Colony System to analyze
PERT network problems to solve decision making problem in pro-
ject management.
Network is a graphical representation of a project. Network
analysis provides a practical way to monitor the progress of the
project till its accomplishment in the minimum time; it can also
be used to assist in allocating resources and to minimize total cost.
The solution of network models is accomplished through a variety
of network optimization algorithms.
2.1. Critical Path Method (CPM)
CPM (Critical Path Method) models are extremely useful for the
purpose of planning, analyzing, controlling the progress and the
completion of large and complex projects (Paul Loomba, 1978).
We must use some definition to complete the computation:
A typical network is shown in Fig. 1 to show its components.
The purpose of the Critical Path Method (CPM) is to identify
critical activities on the critical path so that resources may be con-
centrated on these activities in order to reduce project length time.
Besides, CPM has proved very valuable in evaluating project perfor-
mance and identifying bottlenecks. Thus, CPM is a vital tool for the
planning and control of complex projects (Yao & Lin, 2000). To
identify the critical path, three parameters for each of its activities
are determined: (1) earliest event time, (2) latest event time and
(3) slack time.
Paths other than the critical path offer flexibility in scheduling
because thy take less time to complete less than the critical path.

2.2. Stochastic activity network

One of the most important theoretical problems in project man-

* Corresponding author. Tel.: +20 102593278; fax: +20 237225566.
E-mail addresses: hazemsad2005@yahoo.com (H. Abdallah), hmrashad@ieee.org agement is to obtain the probability distribution of the total com-
(H.M. Emara), dorrahht@aol.com (H.T. Dorrah), agbahgat@hotmail.com (A. Bahgat). pletion time in PERT networks (Paul Loomba, 1978).

0957-4174/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2008.12.064
 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015 10005

In project management, the duration of tasks is seldom pre-
cisely known in advance, at the time when the plan of the project
is designed. Detailed specification of the methods and resources in-
volved for the realization of activities are often not available when
the tentative plan is made up. In stochastic network, the activity
duration is a random variable with a known probability density
function. Authors that introduced the PERT approach have noticed
this difficulty very early. They proposed to model the duration of
tasks by probability distributions and tried to evaluate the mean
value and standard deviation of earliest starting times of activities.
Since then, there is an extensive literature on probabilistic PERT.
Even if the task durations are independent random variables, it is
admitted that the problem of finding the distribution of the ending
time of a project is intractable, due to the dependencies induced by
the topology of the network. Another difficulty is the possible lack
of statistical data validating the choice of activity duration distri-
butions. Even if statistical data are available, they may be partially
inadequate because each project takes place in a specific environ-
ment and is not the exact replica of past projects. Standard PERT
assumes three point estimates for probabilistic duration times in
order to approximate project completion and the relative probabil-
ity at each milestone using the normal distribution function (Paul
Loomba, 1978). PERT and CPM are similar, however they differ in
two terms:
 1-PERT activity estimates are probabilistic while in CPM activity
time is deterministic,
 2-PERT activity costs are not provided while in CPM they are
explicitly provided.
Unlike CPM, in stochastic activity networks the duration time of
individual activities varies and so activities are critical for some
combinations of duration times and may not be critical for other
combinations. Therefore, activities have a given probability of
being critical. PERT assumption of a normally distributed project
completion time typically leads project managers into optimistic
planning based on less than actual project completion estimates
due to a failure in considering the absolute bounds to project com-
pletion. These bounds arise from the fact that the project comple-
tion time is the maximum sum of the duration of each and every
path, which in turn is the result of adding the bounded durations
of its activities (Copertari & Archer, 2001). Project completion can-
not be an unbounded random variable because the sum of
bounded (beta distributed) activity duration times yields bounded
path (and project) completion times. Thus, the normal distribution,
which is unbounded, should not be used to portray completion
times.
For a normalized beta density function the mean value is given
by
l ¼ as =as þ bs : ð1Þ
The corresponding variance value is given by
r2 ¼ as bs =ðas þ bs Þ2 ðas þ bs þ 1Þ; ð2Þ
where (as, bs) are shape (skewing) parameters.
The beta function has various shapes according to different
combinations of shape parameters (a, b). It is a common practice
to portray activity duration times using bell-shaped beta distribu-
tion (Malcolm, Roseboom, Clark, & Fazar, 1959). To ensure a bell-
shaped beta distribution the sum of as + bs P 4 as shown in Fig. 2.
The PERT textbook formula to calculate expected (mean) activ-
ity duration times-which are supposed to follow beta density func-
tions – is given by
l ¼ ða þ 4m þ bÞ=6; ð3Þ
where a, minimum time; m, most likely; and b, maximum time.
The PERT textbook formula considers three parameters when in
fact the beta distribution has four parameters (two range parame-
ters and two shape parameters). It turns out that the PERT formula
used to calculate the mean as a function of the minimum, most
likely, and maximum activity duration time estimates, ignores
how the biases to the right or left affect the shape of the beta
distribution.
In fact, the PERT formula assumes a fixed value for the sum of
the shape parameters (as + bs = 4) to calculate the mean, and it cal-
culates the variance as an approximation to that assumption. Fur-
thermore, PERT does not consider the variance when determining
which path is the longest, since the variance of the project comple-
tion time is assumed to be the same as the variance of the path
with the longest sum of mean duration times.
These assumptions typically lead to optimistic planning due to
less than actual project completion times (Copertari & Archer,
2001).
2.2.1. Forward calculation
Traditional forward pass calculations such as those performed
in CPM are employed. Each event has duration according to
equation
Dnew ¼ D þ DD: ð4Þ
The mean value of the beta function
l ¼ DD  ðas =as þ bÞ: ð5Þ
The mean value of each event
l event ¼ D þ l: ð6Þ
The variance value of each event

Te=6 3 te=5 5
te=9
r2 ¼ DD2  as b=ðas þ bÞ2 ðas þ b þ 1Þ: ð7Þ
Tl=6 Te=11 Te=19
S=0 te=5 Tl=11 Tl=19
S=0 S=0
1 2 7
te=6 te=3 Bell shaped beta distribution
Te=12 Te=14
Te=0 Tl=16 Tl=22
Te=28 αs + βs 4
te=6 te=6 Tl=28
Tl=0 S=4 S=8
S=0 S=0

te=2 6
4

Event
Activity
αs > 1 β s = αs αs > 2
Activity on βs> 2 βs > 1
critical path
Left Bias Symmetrical Right Bias
Fig. 1. A typical PERT network showing its components, earliest time, latest time
and slack time. Fig. 2. Shapes of the beta distribution that ensures a bell-shaped beta distribution.
10006 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015
The critical path is calculated the same way as deterministic CPM
with adding the mean value and ±2r.
2.2.2. Backward calculation
Backward calculation produces unrealistically large uncertain-
ties due to the accumulation while the backward pass. On comple-
tion of the backward pass calculation some of the earlier activities
could end up with negative late time, which is unrealistic. So the
right spread of the late times is at least as uncertain as the early
times. For a linear-type project network the early and late start
times must be identical (Malcolm et al., 1959).
2.3. Fuzzy PERT
The problems of computing the intervals of possible values of
the latest starting times and floats of activities with imprecise dura-
tions represented by fuzzy or interval numbers have attracted
intensively attentions and many solution methods have been pro-
posed, in which it is suggested to use fuzzy numbers to model the
activity times. Most of them are straightforward extensions of
deterministic CPM. They are mainly based on the CPM with formu-
las for the forward and the backward recursions, in which the deter-
ministic activity times are replaced with the fuzzy activity times.
The classic formula used in the CPM as well as some dependencies
true for a network with deterministic activity times are used where
the common operations are replaced with the operations general-
ized on the fuzzy numbers (Chanas, Dubois, & Zieliński, 2002).
With great progress in fuzzy set theory there have been also
several efforts in using this theory in network planning methods
(PERT and CPM) to make more realistic usage of them and better
dealing with uncertainties that is intrinsic in estimating duration
of activities (Lorterapong & Moselhi, 1996).
The application of fuzzy reasoning allows the quantitative eval-
uation of the effect to be produced on the project duration by eval-
uation factors of intuitive and ambiguous judgment such as
‘‘Good”, ‘‘Middle” or ‘‘Bad”. Membership functions are a tool to de-
scribe the degree of belonging of an element to a fuzzy set. In net-
work planning applications, fuzzy sets comprise time units, with
which the duration of each activity is estimated, considering vari-
ous factors such as expert’s opinion, availability of tools or materi-
als, personnel capabilities. Resorting to fuzzy set and possibility
theory may help building a tradeoff between the expressive power
and the computational difficulties of stochastic scheduling tech-
niques while tackling uncertainty and accounting for local specifi-
cations of preferences. When fuzzy intervals represent ill-known
processing times does not lie in introduction of fuzzy sets. It is al-
ready present when only usual intervals are involved. Solving the
interval valued case is the main difficulty. The fuzzy case can then
be rather easily solved, via the use of level-cuts. Modeling ill-
known processing times with fuzzy numbers is rather simple. It
is justified by the limited expressive power of intervals: if too
small, there is little confidence in them, too large; they are not
informative enough to be exploited. Rommelfanger (1990) sug-
gests the use of three confidence intervals: the core, containing
typical values, the 0.5 level cut, containing unsurprising values,
and the support, outside which values are physically unattainable
(Dubois, Fargier, & Galvagnon, 2003). A simpler model is the trian-
gular fuzzy number using an interval and a plausible value in it.
Consider a set of activities (or tasks) related by precedence con-
straints expressing that some activities cannot start before the end
of others. Each activity i is assigned a duration pi. This partially or-
dered set models a project (or a job in the production engineering
context). Usually, two fictitious tasks v and w of null duration are
added to this set, standing for the starting point and the ending
point of the project, respectively: v precedes all activities, and w
takes place after all activities. Let Succ(i) denote the set of activities
immediately following activity i, while. SUCC(i) denote set of all
activities taking place after activity i. And let Pred(i) denote the
set of activities immediately preceding activity i, while PRED(i) de-
note set of all activities taking place before activity i. Given two
activities i and j in the graph, let C(i, j) denotes the set of all paths
from i to j, T  þ
i and T i are earliest and latest starting times of activ-
ity j. An activity is said to be critical whenever its float is zero. The
fuzzy propagation approach enables a fuzzy set of more or less fea-
sible starting times to be computed rather than just the optimal
ones. The starting time Ti of activity i is constrained by the earliest
ending times of activities in PRED(i) and the latest starting times of
activities in SUCC(i). Earliest ending times in PRED(i) and latest
starting times in SUCC(i) are independent quantities here. These
parameters are modeled by fuzzy earliest starting times T  i and lat-
est starting times T þ i . They can be computed by a straightforward
extension of the standard forward and backward recursive
schemes, performed independently, using fuzzy additions and sub-
tractions and the fuzzy extensions of the minimum and the maxi-
mum operations

T i ¼ max T j  Pj for i > v and T v ¼ R; ð8Þ
j2PredðiÞ

T þi ¼ max T j HPj for i < w and T w ¼ D: ð9Þ
j2SuccðiÞ
Assume that the project starts at time 0. T v ¼ 0. Let C(v, i) be the
set of paths (activity sequences) from the initial task to activity i.
When processing times are precisely known the earliest starting
time of activity i is the maximal length L(v, i) of paths in C(v, i) sup-
posing that arcs (i, j) are assigned length Pi. The minimal comple-
tion time is the maximal length L (v, w) of path C in C(v, w).
When processing times are fuzzy, the length of path C is easily de-
fined by adding the fuzzy numbers representing the processing
times of activities in C, and applying the extended maximum

LðCÞ ¼ i2C P i and Lðv ; wÞ ¼ max LðCÞ: ð10Þ
c2Cðv ;wÞ
Fuzzy earliest starting times Ti of activities can be computed by
applying the recursion Eq. (8) until the last activity is reached,
using fuzzy number arithmetic. However, a difficulty arises when
it comes to checking for critical activities, computing latest starting
times and floats.
For criticality analysis, it is assumed that the project has mini-
mal duration. Then floats can be computed and are never negative.
Floats of activities represent the minimal temporal shift of starting
times which do not alter the completion time of the project. Criti-
cal activities form one or several paths in C(v, w). When durations
Pi are ill-known, it is tempting to compute the fuzzy latest starting
times T þi using the backward recursion method Eq. (9), initializing
the process as T þ 
W ¼ T W and using fuzzy subtraction. However, as
pointed out by several authors (Dubois, Fargier, & Fortemps,
2003; Dubois et al., 2003; Hapke et al., 1994; Nasution, 1993;
Rommelfanger, 1994), this method does not work for reasoning
under uncertainty because of assuming that the equality
Tþ 
W ¼ T W does constrain the earliest starting time of the last task
to be equal to its latest starting time. Also fuzzy subtraction on fuz-
zy numbers, that are actually interactive, is considered as a mis-
take. So the computed values of latest starting times will be too
imprecise, and the determination of critical activities cannot be
made by comparing T  þ
i and T i . MacCahon (1993) proposes to go
back to standard critical path methods via defuzzification of the
fuzzy processing times. Kaufmann and Gupta (1988), Hapke et al.
(1994) and Rommelfanger (1994) suggest substitutes to the fuzzy
subtraction, so as to improve the situation, but these techniques
remain poor. The computation of distributions of latest starting
times of activities cannot be achieved using elementary tech-
niques of fuzzy arithmetic, not even of interval arithmetic in the
 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015
non-fuzzy case. It is clear that the difficulty stems for the presence
of intervals, be they fuzzy or not. Nasution (1993) resorts to
symbolic computations on the variable processing times, however
this technique is unwieldy and highly combinatorial (Dubois et al.,
2003).
3. Ant Colony Optimization System
Ant Colony Optimization (ACO) is a metaheuristic technique for
hard discrete optimization problems that was first proposed in the
early 1990s. One of the problems studied by ethnologists was to
understand how almost blind insects like ants could manage to
establish shortest route paths from their colony to feeding sources
and back. When searching for food, ants initially explore the area
surrounding their nest in a random manner. As soon as an ant finds
a food source, it evaluates quantity and quality of the food and car-
ries some of the found food to the nest. During the return trip, the
ant deposits a chemical pheromone trail on the ground. The quan-
tity of pheromone deposited, which may depend on the quantity
and quality of the food, will guide other ants to the food source
(Blum & Dorigo, 2004; Gambardella, Taillard, Dorigo, 1997). While
an isolated ant moves essentially at random, an ant encountering a
previously laid trail can detect it and decide with high probability
to follow it, thus reinforcing the trail with its own pheromone. The
collective behavior that emerges is a form of autocatalytic behavior’
where the more the ants following a trail, the more attractive that
trail becomes for being followed. The process is characterized by a
positive feedback loop, where the probability with which an ant
chooses a path increases with the number of ants that previously
chose the same path. Also pheromone concentration vanishes with
time, and then the less used path will be much lower in pheromone
concentration. The indirect communication between the ants via
the pheromone trails allows them to find shortest paths between
their nest and food sources (Bonabeau et al., 1999). Fig. 3A shows
the real ants following a path between nest and food source.
Fig. 3B shows an obstacle appears on the path. Fig. 3C and D
shows that the pheromone is deposited faster on the shorter path,
and all ants have chosen the shorter path.
In the field of Ant Colony Optimization (ACO), models of collective
intelligence of ants are transformed into useful optimization tech-
niques that find applications. In this paper, the problem-solving
paradigm of ACO is explicated. As we are not interested in simula-
tion of ant colonies, but in the use of artificial ant colonies as an
optimization tool, ACO will have some major differences with a
real (natural) one:
(1) artificial ants will have some memory,
(2) they will not be completely blind,
(3) they will live in an environment where time is discrete.
Fig. 3. Real ants follow a path between nest and food source.
10007
The idea is that if at a given point an ant has to choose among
different paths, they do this based on two main components:
(1) Pheromone trails that were heavily chosen by preceding
ants (that is, those with a high trail level) are chosen with
higher probability. Furthermore, high trail levels (phero-
mone values) are synonymous with short paths.
(2) Heuristic, which is visibility, it is defined as the inverse of
the distance, and so closer nodes have a higher chance of
being next.
Two parameters (a and b) are introduced to control the relative
weight of pheromone trail intensity and visibility. At run-time,
ACO algorithms try to update the pheromone values in such a
way that the probability to generate high-quality solutions in-
creases over time. The pheromone values are updated using previ-
ously generated solutions. The reinforcement of solution
components depending on the solution quality is an important
ingredient of ACO algorithms. It implicitly assumes that good solu-
tions consist of good solution components. To learn which compo-
nents contribute to good solutions can help to assemble them into
better solutions (Dorigo, Maniezzo, & Colorni, 1996).
3.1. ACO algorithm description
ACO has been applied in many combinatorial optimization
problems such as the asymmetric traveling salesman problem
(Fogel, 1993; Gambardella & Dorigo, 1996), graph coloring problem
and vehicle routing problem, etc.
The basic ACO algorithm involves the following: ants choose the
trail with largest concentration of pheromone s with a certain
probability p as shown in Fig. 4.
The shorter path will have more pheromone concentration. For
a whole colony following this behavior the shortest path will be
much higher in pheromone concentration than the other paths.
Then the probability of choosing any other paths will be very small.
Also the less used path will be much lower in pheromone concen-
tration due to pheromone volatility with time.
Let an ant k resides in a node i at time t. Assume that one of the
possible trails for its next step is (i, j) connecting node i to node j.
The ant selects this trail with probability of
pðk; i; jÞ ¼ f ðsði; jÞÞ; ð11Þ
where s(i, j) is the pheromone concentration on the path connecting
i to j in the way of food source.
Then the probability of choosing node j as the next node is given
by
8 sði;jÞa gði;jÞb
<P a b
sði;jÞ gði;jÞ
f ðsði; j; tÞÞ ¼ ; ð12Þ
: r2C
0 rRC
where the pheromone trails update is given by
sði; jÞðt þ 1Þ ¼ v  sði; jÞðtÞ þ dðk; i; jÞ; ð13Þ
v 2 (0, 1) is the evaporation coefficient. d(k, i, j) is the pheromone
released by the ant k on the trail(i, j)
τ3
p(τ1)
p(τ3) p(τ2)
τ2 τ1
Pheromone trail depositing
Probabilistic rule to choose the path
Fig. 4. Probability of real ant to choose the path.
10008 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015

0 if ant’s selected trail – j
dðk; i; jÞ ¼ ; ð14Þ
Q =L if ant’s selected trail ¼ j
where Q is a positive constant and L is the total path time. U is a
tabu list, which contains all the trails that the ant has already
passed and must not be chosen again, this act as the memory of
the ant. g (i, j) is the visibility function which is usually selected
as the inverse of the distance for i to j, i.e.
gði; jÞ ¼ c=dði; jÞ
where c is the visibility parameter, a and b express the relative
weight between the importance of s and g. The steps of the ACO
algorithm are illustrated as follows (Sim & Sun, 2003):
Optimization Algorithm of standard ACO
Step 1: initialize the pheromone table.
Step 2: randomly allocate ants to every node.
Step 3: every ant walk to next node, depending on the prob-
ability distribution (local search).
Step 4: compute the length of the path traveled by each ant,
and allocate a quantity amount of pheromone and visibility
to the path, according to the length of its path.
Step 5: perform a local update.
Step 6: compute whether a better solution is obtained in
this time step than the last; if so, then perform a global
update on the solution and empty the Tabu value.
Step 7: repeat Steps 2 to 6.
4. Proposed ACO technique for solving CPM
The basic ACO has been applied to solve the traveling salesman
problem (Bonabeau et al., 1999). Compared to this benchmark
problem, PERT networks are more imitating the natural ants. This
is due to the fact that the PERT network has a start node that is
equivalent to the ants’ nest and an end node equivalent to the food
location. Two types of PERT networks will be considered, namely:
deterministic PERT network and Uncertain PERT network.
4.1. Deterministic PERT network
ACO technique is used primarily for path minimization prob-
lem. The proposed algorithm modifies the basic ACO algorithm in
order to look for the longest path (CPM).
The visibility function g(i, j) which is usually selected as the in-
verse of the distance for i to j, it is proposed here to be proportional
to the time of the activity
g ¼ ð1 þ dði; jÞ=cÞ:
Also for d(k)(i, j) is assumed proportional to the path length
 be described by modifying Step 4 in Algorithm 1.
0 if ant’s selected trail – j
dðk; i; jÞ ¼ : ð17Þ
Q  L if ant’s selected trail ¼ j
For the first step we put random amount of pheromone trails
randomly on different paths, this step acts as the initialization of
the pheromone level. The purpose of this step is to give a random
guide to the first to follow. Then, we put the first ant at the first ini-
tial state. The ant chooses one of the available paths guided by the
pheromone trail and visibility level, and goes to the next node,
putting its pheromone trace on this path. The node reached by
the first ant is now considered as the initial state. A tabu list, which
contains all the trails that the ant has already passed is made, (e.g.
node 1 for the first ant after the first step) also the nodes with no
oriented link (activity) with the initial node, (e.g. node 6–3–8–4 for
initial node 1); and must not be chosen are included in the tabu
list. The tabu list acts as the memory of the ant. The pheromone le-
vel is also updated. The ant now goes to the next node guided by
the pheromone trail and visibility level, and so on until it reaches
the final node. The decay factor of the pheromone density due to
evaporation is taken into consideration. Then another ant is al-
lowed to choose a path till it reaches the last node. The pheromone
ð15Þ and visibility level on this path is updated. The second ant chooses
randomly at node 1 but at node 2 it is guided with the trail of the
first ant and chooses node 8 and continues to node 4. The critical
path, which is the longest path, can be now determined by choos-
ing the path with lowest pheromone level. Fig. 5 shows the path
chosen randomly by first ant, which deposits pheromone on the
path: 1–2–8–4. The next ant chooses another path: 1–7–8–4 and
also deposits pheromone on the new path while the pheromone le-
vel on the first path (chosen by first ant) decays due to evaporation
as shown in Fig. 6.
The Optimization Algorithm for the developed ACO to solve the
deterministic critical path method can be briefly described as
follow:
Optimization Algorithm of the proposed ACO
Step 1: initialize the pheromone table.
Step 2: allocate first ant to initial node.
Step 3: every ant must walk to next node, depending on the
probability distribution (local search) until it reaches the
last node.
Step 4: compute the length of the path traveled by each ant,
and determine the amount of pheromone to be deposit on
every activity arc on this path based on the total path
length.
Step 5: perform a local update.
Step 6: compute whether a better solution is obtained in
this time step than the last; if so, then perform a global
update on the solution and empty the Tabu value.
Step 7: repeat Steps 2 to 6.
4.2. Uncertain PERT network
The same developed ACO is used to solve probabilistic PERT net-
work with a change that every ant has to go from a node to another
taking into consideration the uncertainty of the duration of each
event and dealing with the event time of each event as a random
variable and then accumulate the uncertainty while the ant is trav-
eling to the end node. The path chosen by the ant the same way
ð16Þ mentioned before with adding the mean value.
The Optimization Algorithm for the developed Ant Colony Sys-
tem to solve the critical path method with uncertainty can briefly
Step 4: compute the length of the path traveled by each ant
including the variation of each event as a random variable, and
allocate a quantity amount of pheromone to the path, according
to the length of its path.
5. Case study results
5.1. Introduction
Four case studies are presented: the first is a typical PERT net-
work solved by ACO in order to test all parameters which influence
the solution, in the second case the developed model is applied to
 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015 10009

5 6

1
2 3 4

7 8

Fig. 5. The path randomly chosen by the first ant.

5 6

1
2 3 4

7 8

Fig. 6. The path randomly chosen by the 2nd ant and the pheromone level update.

solve crashing in a PERT network then in Case study 3 we apply the 12

developed ACO to a PERT network solved by fuzzy PERT and a com- 5 E 29 H 35
8
parison is made between the two results, in the fourth one the 7
25 4 35
developed ACO is applied to a medium scale project. 29
6
B 12 G
M
5.2. Case study 1: House construction activities
10 J 2
5
The duration of construction activities of a house are usually 0 A D 17 37 L 43
3 7 4 10 11
estimated by experts in order to identify areas that need more
1 0 I
management attention. Table 1 lists the activities of our case study 7 17 37 43
6
10
and the stages involved in the construction process. No new activ- 5 12
ity can begin until the event logically preceding it has occurred. C
2
The corresponding network is shown in Fig. 7. The critical path is 7 K
calculated through the traditional procedure of the CPM method.
7 17 Earliest
The critical path is: 1–2–3–4–6–7–8–10–11 and the overall time 6
N 19
Event time
2
of the critical path is 43 days. To investigate the applicability of 9
17 2 41
the proposed ACO technique, it has been applied to solve the 7

CPM in the case study network, the influence of the following x

y

parameters are considered: z

Event

(1) number of ants (n) Latest

Critical path Event time
(2) the decay rate (v)
(3) visibility parameter (c) Fig. 7. Case Study 1 Network.
(4) numerical parameter (Q).

The influence of number of ants on reaching the optimal solu-

tion for 100 tests is described in Table 2. During these tests c; v
and q are kept constants {6; 0.02; 0.5  L} respectively. The number
Table 1
Activities and stages in project construction of Case study 1. of ants with is increased from 15 to 1000. The initial pheromone
level is assumed randomly on all trails. In the initialization phase,
Activity reference Activity nature Linking effect Estimated days
the ants walk randomly according to the randomly existing phero-
A Obtain bricks 1&3 5 mone level from the initialization step. While they are moving they
B Obtain roof tiles 1&5 12
leave their own pheromones. Then, they follow the usual phero-
C Prepare foundations 1&2 7
D Erect shell 3&4 10 mone and visibility level. It is clear that using starting from 250
E Roof 5&7 4 ants leads to a satisfying results and converge to the optimal solu-
F Lay drains 2&6 7 tion (43 days–74%) and give an indication on a suboptimal results
G Wiring 4&7 10 (41 days–26%). For 750 ants results converge to the optimal solu-
H Plastering 7&8 6
I Plumping 6&7 12
tion (43 days–98%) and suboptimal results (41 days–2%). These re-
J Flooring 7 & 10 5 sults show the ability of ACO to produce a good suboptimal
K Landscaping 9 & 11 2 solution in a short time.
L Painting 10 &11 6 The influence of decay rates upon the first case study solution is
M Doors & fitting 8 & 10 2
described in Table 3. In this case, the number of ants, g and Q are
N Lay pathways 6&9 2
kept constant at {50; 0.05; 0.5  L} respectively. Using high decay
10010 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015

Table 2
The influence of number of ants on the probability distribution of the maximum time found by ACO.

Ants No. Time

27 30 37 Sub 39 (%) Optimal 40 (%) Solutions 41 (%) Critical path 43 (%)
15 0 0 0 12 26 55 7
25 0 0 0 4 21 57 17
50 0 0 0 0 13 61 26
75 0 0 0 0 3 64 33
100 0 0 0 0 0 39 61
150 0 0 0 0 0 50 15
250 0 0 0 0 0 26 74
500 0 0 0 0 0 13 87
750 0 0 0 0 0 2 98
1000 0 0 0 0 0 0 100

rate leads the ants to make all paths randomly and the ant does not time. This is clear from Eq. (9). In this case, the ants are not heavily
influence the successor ones. Low decay rate leads to fast conver- attracted to the pheromones of preceding ants. This may lead to
gence to the preceding path and may lead to immature conver- inaccurate identification of the critical path. Using extremely low
gence due to the lack of exploration of other paths. For example, c encourages the ants to follow the preceding ants’ pheromones.
when the decay rate is set to 0.02, in 13% of the simulations, the By considering number of ants, decay rate, c constants {50; 0.02;
Ant colony converges immaturely to a path with length 40 days. 6} respectively and increasing numerical parameter. The influence
The influence of visibility rate (c) on the results of the Case of numerical parameter Q upon the results is described in Table 5.
study 1 is described in Table 4, the number of ants, v and q are kept It should be noted that adequate value of the parameter Q depends
constants {50; 0.02; 0.5  L} respectively. Using high c encourages upon the total network time. As we can see that the factor that most
the ants to choose the next node as the having the largest trail influences the results is the number of ants which yields to a fast
convergence to the solution Ant Colony System has a remarkable
advantage over traditional optimization algorithms is the ability
Table 3
to produce a good suboptimal solution in a very quick time. With
The influence of decay rate on the probability distribution of the maximum time
found by ACO. the increasing of the computational power ACO has proved effi-
ciency to solve the networks in a very short time.
Decay rate Time
The proposed ACO has been applied to solve PERT network case
27 30 37 39 40 (%) 41 (%) 43 (%) study under uncertainty by dealing with the event time of each
0.01 0 0 0 0 4 67 29 event as a random variable. During these tests number of ants, c,
0.02 0 0 0 0 13 61 26 v and q are kept constants {50; 6; 0.02; 0.5  L} respectively. The
0.05 0 0 0 0 6 64 30 activity A and B have a variation (Dt) of 10% of there most likely
0.1 0 0 0 0 3 62 35
0.2 0 0 0 0 4 53 43
time due to their purchasing nature (limited risk), all other activi-
0.3 0 0 0 0 2 44 54 ties have a variation of 20%. The network is tested while activities
0.6 0 0 0 0 3 51 46 have a symmetrical, left biased, right biased beta distribution func-
0.9 0 0 0 0 1 61 38 tion, respectively.
The results are solved for each path in order to compare with
the results obtained by ACO.
The time for each path taking a = b = 4 the critical path is 43 for
Table 4
The influence of visibility rate on the probability distribution of the maximum time
optimistic solution, 47.3 for expected solution, 48.58 with the
found by ACO.

c Time
Pert Analysis
27 30 37 39 (%) 40 (%) 41 (%) 43 (%) α =4 β =4
60
1 0 0 0 1 42 51 6
3 0 0 0 1 15 73 11
6 0 0 0 0 13 61 26
15 0 0 0 0 4 50 46
50 0 0 0 0 18 47 51
100 0 0 0 1 1 50 48 40
500 0 0 0 0 2 41 57
Time

Table 5
20
The influence of parameter Q on the probability distribution of the maximum time
found by ACO.
Optimistic
Q Time Expected
Expected+2Seg
27 30 37 (%) 39 (%) 40 (%) 41 (%) 43 (%) Expected-2Seg
0
0.5 0 0 2 8 25 57 8
bejl

adghml

adgjl

adihml

adijl

afihml

afijl

adnk

afnk

cdghml

cdgjl

cdihml

cdijl

cdnk

cfihml

cfijl

cfnk
behml

0.3 0 0 0 2 16 64 18
0.1 0 0 0 1 8 65 26
0.08 0 0 0 0 8 60 32 Path
0.05 0 0 0 0 13 61 26
Fig. 8. PERT network analysis-time required for all paths with a = b = 4.
 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015 10011

adaptation of the standard deviation +2r, 46.01 with the adapta- the adaptation of the standard deviation +2r, 44.04 with the adap-
tion of the standard deviation 2r. This result is shown in Fig. 8. tation of the standard deviation 2r. This result is shown in Fig. 9.
The time for each path taking a = 2 b = 6 the critical path is 43 The time for each path taking a = 6 b = 2 the critical path is 43
for optimistic solution, 45.15 for expected solution, 46.25 with for optimistic solution, 49.45 for expected solution, 50.55 with

Pert Analysis 3
15
α =2 β=6 10
15 5
60
40
10 5
40
19
0 5 70 10
2 5 80
40
1 0 15 7 70 8
5 80
10 20
Time

10
5
12
20 30
6 60
4
Optimistic 30 19 60
Expected
Expected+2Seg
Expected-2Seg Fig. 12. Case study 2 Network.
0
bejl

adghml

adgjl

adihml

adijl

afihml

afijl

adnk

afnk

cdghml

cdgjl

cdihml

cdijl

cdnk

cfihml

cfijl

cfnk
behml

Table 6
Path
Activities crashing in an electrical installation project of Case study 2.

Fig. 9. PERT network analysis-time required for all paths with a = 2 b = 6. Job Normal duration Crash duration Normal cost Crash cost
1.2 5 150 150 5
2.3 10 200 200 10
Pert Analysis 2.4 5 250 310 2
α =6 β =2 3.4 15 900 900 15
60 3.5 5 750 1150 1
3.8 10 1000 1250 8
4.5 10 300 540 7
4.6 19 400 960 11
5.6 20 500 600 18
5.7 19 600 900 15
40
6.7 10 700 1210 4
6.8 12 600 800 10
Time

7.8 10 250 300 9

20
Table 7
Optimistic Results of crashing of Case study 2 PERT network.
Expected
Expected+2Seg Project Activity Normal Crash Number Slope Project Project
Expected-2Seg duration duration duration of days cost cost (after
0 reduced (before crashing)
bejl

adghml

adgjl

adihml

adijl

afihml

afijl

adnk

afnk

cdghml

cdgjl

cdihml

cdijl

cdnk

cfihml

cfijl

cfnk
behml

crashing)
80 6600
Path
79 7–8 10 9 1 50 6600 6650
77 5–6 20 18 2 50 6650 6750
Fig. 10. PERT network analysis-time required for all paths with a = 6 b = 2.
74 4–5 10 7 3 80 6750 6990
68 6–7 10 4 6 85 6990 7500

70
UNCERTAINTY RESULTS
60
RIGHT BIAS
SYMMETRICAL 50
2 ~ 5
OCCURANCE

LEFT BIAS
40 ~ ~ ~
30

20
1 4
~ 7
9

10
~ ~ ~
0

~ ~
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52

3 6 8
TIME

Fig. 11. The influence of a and b on reaching the critical path. Fig. 13. Case study 3 Network.
10012 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015

Table 8
The fuzzy activity times Tij, left shape function Lij(x), right shape function Rij(x), the a-cuts of fuzzy activity times.

Tij Lij(x) Rij(x) ðT ij Þiac ðT 12 ÞUac

L
pffiffiffiffiffiffiffiffiffiffiffiffi
T12 = (1; 1.5; 1; 1)L12R12 L12(x) = max(1  x ,0)2
R12(x) = max(0,1  x) ðT 12 Þac ¼ 1  1  ac ðT 12 ÞUa ¼ 2:5  ac
T13 = (2; 3; 0; 2)L13R13 L13(x) = ex R13(x) = max(0,1  x) ðT 13 ÞLac ¼ 2 ðT 13 ÞUac ¼ 5  2a
T24 = (0;0;0;0)L24R24 L24(x) = max(0,1  x) R24(x) = max(0,1  x) ðT 24 ÞLac ¼ 0 ðT 24 ÞUac ¼ 0
p ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
T25 = (2; 3; 1; 2)L25R25 L25(x) = max(0,1  x4) R25(x) = ex ðT 25 ÞLac ¼ 2  4
1  ac ðT 25 ÞLac ¼ 3  2 ln ac
T34 = (0;0;0;0)L24R34 L34(x) = max(0,1  x) R34(x) = max(1  x2,0) ðT 34 ÞLac ¼ 0 ðT 34 ÞUac ¼ 0
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðT 36 ÞLac ¼ 6 ðT 36 ÞUac ¼ 7 þ 2 1  ac
2
T36 = (6; 7; 0; 2)L36R36 L36(x) = ex R36(x) = max(1  x2,0)
4 L U
pffiffiffiffiffiffiffiffiffiffiffiffi
T46 = (5; 5; 1; 1)L46R46 L46(x) = max(0,1  x) R46(x) = max(0,1  x ) ðT 46 Þac ¼ 4 þ ac ðT 46 Þac ¼ 5 þ 2 4 1  ac
pffiffiffiffiffiffiffiffiffiffiffiffi
T47 = (9; 9; 1; 1)L47R47 L47(x) = max(0,1  x4) R47(x) = ex ðT 47 ÞLac ¼ 9  4 1  ac ðT 47 ÞUac ¼ 9  ln ac
L
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
T59 = (8; 9; 2; 4)L59R59 L59(x) = max(0,1  x4) R59(x) = max(1  x2,0) ðT 59 Þac ¼ 8  2 4 1  ac ðT 59 ÞUac ¼ 9 þ 4 1  ac
p ffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
T68 = (4; 4; 2; 2)L68R68 L68(x) = max(1- x2,0) R68(x) = max(0,1  x4) ðT 68 ÞLac ¼ 4  2 1  ac ðT 68 ÞUac ¼ 4 þ 2 4 1  ac
T78 = (3; 4; 2; 0)L78R78 L78(x) = max(0,1  x) R78(x) = max(0,1  x4) ðT 78 ÞLac ¼ 1 þ 2ac ðT 78 ÞUac ¼ 4
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffi
T89 = (6; 9; 2; 3)L89R89 L89(x) = max(1  x2,0) R89(x) = ex2 ðT 89 ÞLac ¼ 6  2 1  ac ðT 89 ÞUac ¼ 9 þ 3  ln ac

D = max T12X12 + T13X13 + T24X24 + T34X34 + T25X25 + T36X36 + T46X46 + T47X47 + T59X59 + T68X68 + T78X78 + T89X89.
Such that:
X12 + X13 = 1; X12 = X24 + X25; X13 = X34 + X36; X24 + X34 = X47 + X46; X25 = X59; X46 + X36 = X68; X47 = X78; X78 + X68 = X89; X59 X89 = 1;
X12; X13; X24; X34; X25; X36; X46; X47; X59; X68; X78; X89 P 0.

the adaptation of the standard deviation +2r, 48.34 with the adap-
tation of the standard deviation 2r. This result is shown in
Fig. 10.
Fig. 11 shows the applicability of the developed ACO to solve
stochastic PERT network indicating the influence of parameters a
and b on the network.
Comparing the results obtained by ACO with all paths solution
we see that ACO converges to right solution under any assump-
tions of the project.
Table 9
The a-cuts of the total duration and the corresponding critical paths at 12 a values for
Case study 3.
5.3. Case study 2: Crashing activities
The events of a project is given in table, the corresponding net-
work is shown in Fig. 12 (see Table 6).
The proposed ACO is then applied to the crashing procedure and
the critical path is calculated through the traditional procedure of
the CPM method.
During this test number of ants, c, v and q are kept constants
{500; 6; 0.02; 0.5  L} respectively. The critical path is: 1–3–4–5–
6–7–8 and the overall time of the critical path is 80 days with cost
equal 6600 pounds.
The developed ACO reaches the same results reached through
the traditional crashing algorithm. Results are shown in Table 7

a-cut Lower bound Upper bound

Table 11
DLa Critical path DUa Critical path The network solved by the developed ACO for several Dt.
1 20.0000 1–3–4–7–8–9 25.0000 1–3–4–7–8–9
Dt Critical path Time of critical path
0.9 18.6052 1–3–4–7–8–9 26.2791 1–3–4–7–8–9
0.8 18.0368 1–3–4–7–8–9 27.0491 1–3–6–8–9 0.2  Dt 1–3–4–7–8–9 17.3
0.7 17.5645 1–3–4–7–8–9 27.9673 1–3–6–8–9 0.5  Dt 1–3–4–7–8–9 19.25
0.6 17.1398 1–3–4–7–8–9 28.7996 1–3–6–8–9 0.75  Dt 1–3–4–7–8–9 20.87
0.5 16.7449 1–3–4–7–8–9 29.5937 1–3–6–8–9 Dt 1–3–4–7–8–9 22.5
0.4 16.3707 1–3–4–7–8–9 30.3811 1–3–6–8–9 1.25  Dt 1–3–4–7–8–9 24.1
0.3 16.0120 1–3–4–7–8–9 31.1945 1–3–6–8–9 1.5  Dt 1–3–4–7–8–9 24.75
0.2 15.6654 1–3–4–7–8–9 32.0863 1–3–6–8–9 2  Dt 1–3–4–7–8–9 29
0.1 15.3286 1–3–4–7–8–9 33.6549 1–3–4–7–8–9 2.5  Dt 1–3–4–7–8–9 31.5
0.01 15.0325 1–3–4–7–8–9 38.0231 1–3–4–7–8–9 4  Dt 1–3–6–8–9 42
0.001 15.0033 1–3–4–7–8–9 41.7905 1–3–4–7–8–9 5  Dt 1–3–6–8–9 49

ΔT α - cut
Table 10
Activities of Case study 3 as random variables. 4.5 Total duration time by ACO =22.5
4 Fuzzy total duration time =23.26
Tij t Dt as bs 1
3.5
T12 0 2.5 3 3
3 0.8
T13 2 3 6 2
2.5
T24 0 0 4 4 0.6
T25 1 5 2 4 2
T34 0 0 4 4 1.5 0.4
T36 6 3 6 2 1
T46 4 2 3 3 0.2
0.5
T47 8 3 2 4
T59 6 7 2 4
T68 2 4 3 3 5 10 15 20 25 30 35 40
T78 1 3 4 2
T89 4 5 3 4 Fig. 14. The membership function of the fuzzy total duration time and the total
duration time by ACO for Case study 3.
 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015 10013

5.4. Case study 3: Uncertain PERT network where L and R are continuous, non-increasing functions that define
the left and right shapes of lT(t), respectively; and L(0) = R(0) = 1.
The third case study is presented in order to apply the devel- The a-cuts of all fuzzy activity times can be summarized in Ta-
oped ACO to another PERT network solved by fuzzy PERT method ble 8.
and compare the results.Consider a project whose corresponding According to Chen (2007), there is a way to determine the total
network is given in Fig. 13. The activities times are fuzzy numbers duration and find a critical path is by using LP technique. The idea
of Lij_Rij type, (i, j) .The concept of L–R fuzzy numbers is explained is based on the concept that a CPM problem can be thought of as
in Chen (2007); a fuzzy set T = (tla¼1 ; t la¼1 , ls, rs)LR is called a fuzzy the opposite of the shortest path problem, to determine a critical
number of L–R type with the subset ½tlac¼1  consisting of the real path in the project network it is sufficient to find the longest path
numbers with the highest chance of realization, and ls and rs being from start to finish. Then the length of this longest path is the total
the left and right spreads that are non-negative real numbers, duration time of the project network. The LP formulation assumes
respectively, if its membership function l T ðtÞ has the form as that a unit flow enters the project network at the start node and
follows: leaves at the finish node.
8 tl t A mathematical programming software package Lingo is used to
>
>
> L ac¼1
ls
for t 6 t la¼1 solve the above linear programs. Table 9 lists the a-cuts of the total
<
lT ðtÞ ¼ duration time and the fuzzy critical paths at twelve distinct a val-
l for tla¼1 6 t 6 tUa¼1 ; ð18Þ
>
>   ues: 1, 0.9. . .0.1, 0.01, and 0.001. The approximate membership
> l
: R ttac¼1
rs
for t P t ua¼1 function of the fuzzy total duration time is shown in Table 9.

Table 12
Activities duration of a robotics automation system.

No. Description Time No. Description Time

1 Define System Requirements 10 51 Review Bids for Temp Control Equipment 2
2 System Design 20 52 Award Contract for System2 Controller 1
3 Prepare Drawings for Temp Control Equipment 10 53 Fabricate & Deliver Temp Control Equipment 20
4 Approve System Design 10 54 Site 2 Preparation 10
5 Prepare Drawings for System Controller 10 55 Install Robot Base 18
6 Review and Approve Temp Control Equipment 5 56 Run Sealant, Air, and Water Piping 13
7 Prepare and Solicit Bids for Temp Control Equip 5 57 Installation Begins 0
8 Review Bids for Temp Control Equipment 2 58 Install Electrical Power 24
9 Site Preparation 10 59 Dummy Link 0
10 Award Contract for Temp Control Equipment 1 60 Dummy Link 0
11 Install Robot Base 18 61 Dummy Link 0
12 Run Sealant, Air, and Water Piping 13 62 Dummy Link 0
13 Installation Begins 0 63 Deliver system2 50
14 Install Electrical Power 24 64 Workshop Drawings 20
15 Fabricate & Deliver Temp Control Equipment 20 65 Review of Workshop Drawings 5
16 Review and Approve System Controller 10 66 Prepare Drawings for Temp Control Equipment 15
64 Workshop Drawings 20 67 Review and Approve System Controller 12
17 Prepare and Solicit Bids for System Controller 5 68 Prepare and Solicit Bids for System Controller 8
18 Review Bids for System Controller 4 69 Review Bids for System Controller 5
19 Award Contract for System Controller 1 70 Award Contract for System Controller 2
20 Fabricate and Deliver System Controller 60 71 Workshop Drawings 22
21 Dummy Link 0 72 Review of Workshop Drawings 6
22 Dummy Link 0 73 Deliver System 2 75
23 Dummy Link 0 74 Design System 3 10
24 Dummy Link 0 75 Review and Approve System3 Controller 10
25 Install Temperature Control Equipment 20 76 Prepare and Solicit Bids for System3 Controller 5
26 Set & Connect Robots 18 77 Prepare and Solicit Bids for Temp Control Equip 4
27 Install System & Misc. Components 30 78 Review Bids for Temp Control Equipment 1
28 Install System Controller 16 79 Workshop Drawings 20
29 Test & Debug Line B 24 80 Review of Workshop Drawings 5
30 Test & Debug Line A 24 81 Deliver System 3 60
31 Install Bar Rails 20 82 Prepare Drawings for System4 Controller 13
32 Trim Robot Paths Line B 24 83 Review and Approve System4 Controller 8
33 Trim Robot Paths Line A 15 84 Prepare and Solicit Bids for System4 Controller 5
34 Start-Up Line B 1 85 Prepare and Solicit Bids for Temp Control Equip 4
35 Pilot Start Line A 1 86 Review Bids for Temp Control Equipment 2
36 Path Refinement and Shakedown-Line B 10 87 Workshop Drawings 20
37 Path Refinement and Shakedown-Line A 10 88 Review of Workshop Drawings 4
38 Switch System2 0 89 Deliver System 4 60
39 Define System2 Requirements 10 90 Production Start 0
40 System2 Design 20 91 Install Processor/Software/Data Tapes 2
41 Prepare Drawings for System Controller 41 92 Calibrate Robot Controller & Power Up 1
42 Prepare Drawings for Temp Control Equipment 42 93 Load System Software 20
43 Approve System2 Design 10 94 Manual Operation Ready 0
44 Review and Approve System2 Controller 10 95 Program 15
45 Prepare and Solicit Bids for System2 Controller 5 96 Test Software 18
46 Prepare and Solicit Bids for Temp Control Equip 4 97 Debug Software 7
47 Review Bids for Temp Control Equipment 1 98 Automatic Operation Ready 0
48 Fabricate & Deliver Temp Control Equipment 60 99 Startup Procedure Development 3
49 Review and Approve Temp Control Equipment 5 100 Training Manuals/System Operation 9
50 Prepare and Solicit Bids for System2 Controller 5 101 System Buyoff 0
10014 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015

The Critical Path is : 1 2 3 7 9 19 20 24 25 26 27 28 29 30 31 51 53 55 58 75 76 77 78 79 80 81 82 82

214
50
N7 1
N1 2
214 124
74 194 291
50 306
N2 8 114
N7124
6 N3 7 N1 7 N5 1 N7 5
119 306 207 63 194 291
114 148 204
N3 3 74 N5 3 N4 1 N5 0
119 142 N2 4 207 343 40 126
148 114 124 182
N4 4 52 74 N8 1 N7N2 7 64N3 5
N3 9
142 N8 166
10 50 343 40 114 N1126
4
N2 142
52 N4 8 181
124 331 N6 59
148
10 N4
2379
50
235 N3 4 N7 9 166 N4 2
246N6 3 124 181
N6 4 134 114
209 331 148
N5 9 N6 234
5 334 237
128 N3 1 N3 2 40
209 247 227 166 N8 0 N3 6 134 114
60 N1 1
291 N5 5 N4 5 50 94 104
334 N2 2 127 231N2 1 40
N7 4 227 166 N9 45 N2129
6 N6 1 54 124
0 237 94
50 274 104N2 9 60N1 N3
231 0 8 343
N5 112 220
N1 N7 0 129 53 124 N8 2
148 22145 N2 3 N5 6
0 133 216214 271 343
N4N3
0 N6
N70 2 112 220
2170
324 137 30 N5 2
221
216
133
N7 3 331
N7 7 N3214 262
217 84 N7 8
324 30 N6 8
74 N2 5 331
257 134 262 190
N1 9 84270
74 N6
215 6 74N1 5 74 N4 6
N5257
4 N6 9 190
40 N165
6 266 229 N2 0
N4 266 215 68 N6 2 74
40 N6 7 74
229 164
266 N1 229
8
N4 3
50 N5 7 231
224 148 N5 8
55 231
N1 3 204
The Time of critical path is: 343
55 N4 7
The Total Cost is: 101 191

Fig. 15. Results of Case study 4 solved by developed ACO.

The same example is solved by the proposed ACO algorithm: same critical path is reached from 1ac until 0.8ac then the critical
each fuzzy activity time is transformed into random variable with path changes to 1–3–6–8–9 from 0.1ac until 0.001ac. While for
four parameters; t; the most optimistic time for each activity, Dt; ACO the critical path is 1–3–4–7–8 from 0.2 Dt until 2.5Dt then
the variation of time for each activity, as, bs the skewing parame- it changes for bigger Dt to 1–3–6–8–9. These differences are due
ters as shown in Table 10. The critical path and its time duration to the variation of handling random variables and fuzzy numbers,
are solved for several Dt and the results are shown in Table 11. solving by fuzzy PERT the support of exponential equation is finite
Comparing the results of the two methods of solving the PERT when considered finite in the second method, also due to the
network (see Fig. 14). approximation of the defuzzification methods.
Solving by fuzzy PERT the critical path is 23.26 while solving by
the developed ACO yields to 22.5. Also for the lower bound of fuzzy 5.5. Case study 4: Robotics automation system
PERT the critical path is 1–3–4–7–8–9, for the upper bound the
The fourth case study is presented in order to apply the devel-
oped ACO to a medium scale network. Table 12 shows a robotics
automation system with 101 activities, 82 nodes.
Table 13
Robotics automation system durations with uncertainty.
The case study is solved in deterministic case. The results are
shown in Fig. 15. The critical path is determined and the total time
Act. Time DT Act. Time DT Act. Time DT Act. Time DT of the project is 343 days.
1 10 0.1 76 22 2.2 26 0 0 51 0 0 Also uncertainty is taken into consideration of some activities as
2 20 7 77 8 0.8 27 5 0.5 52 0 0 shown in Table 13. The system is capable to analyze this uncer-
3 10 1 78 6 2 28 18 1.8 53 0 0
tainty as shown in Fig. 16. The critical path is determined and
4 10 1 79 5 0.5 29 30 3 54 50 5
5 10 2 80 2 0.2 30 10 1 55 24 2.4 the total time of the project has increased to 373.96 days.
6 5 0.5 81 10 1 31 20 2 56 24 2.4 The system is capable of dealing with all kinds of projects either
7 5 0.5 82 75 7.5 32 10 1 57 16 1.6 deterministic or probabilistic, the developed ACO and its parame-
8 2 0.6 83 10 2 33 10 1 58 24 2.4
ters are tested through small, medium scale projects dealing with
9 10 1 84 20 2 34 10 1 59 1 0.1
10 1 0.1 85 5 0.5 35 10 1 60 10 1
all natures of activities and their degrees of uncertainty. Also the
11 18 1.8 86 5 0.5 36 5 0.5 61 15 1.5 developed ACO is applied to a crashing problem. The results are
12 13 1.3 87 4 1 37 4 0.4 62 1 0.1 compared with other ways of solving PERT to test its power and
13 0 0 88 1 0.1 38 1 0.1 63 10 2 applicability to analyze all kinds of PERT problems.
14 24 2.4 89 60 20 39 60 20 64 20 2
15 20 2 90 2 0.2 40 5 1.25 65 13 1.3
16 10 1 91 1 0.1 41 5 0.5 66 15 1.5 6. Conclusions
64 20 7 92 20 2 42 2 0.2 67 0 0
17 5 0.5 93 0 0 43 1 0.1 68 8 0.8 Several researchers have investigated the critical path analysis
18 4 0.4 94 15 1.5 44 20 7 69 20 2
19 1 0.1 95 18 1.8 45 10 0.1 70 5 0.5
in the project network with fuzzy activity times, random time. This
20 60 6 96 7 0.7 46 18 6 71 4 0.4 paper develops an approach to use Ant Colony Systems for solving
21 0 0 97 0 0 47 13 1.3 72 4 0.4 both deterministic and probabilistic CPM/PERT networks. The
22 0 0 98 3 0.3 48 0 0 73 2 0.2 applicability of the proposed algorithm is demonstrated for a typ-
23 0 0 99 9 0.9 49 24 8 74 60 6
ical construction project and another project with uncertain data.
24 0 0 100 0 0 50 0 0 75 12 1.2
Investigation of the results reveals that the main advantage of
 H. Abdallah et al. / Expert Systems with Applications 36 (2009) 10004–10015 10015

The Critical Path is: 1 2 3 4 5 6 8 10 20 24 25 26 27 28 29 30 31 51 59 61 64 65 67 70 80 81 82

289.25
181.25 159.8192.1 201.55
N67
N43 235.85N41 N49
79.75
240.05 289.25 234.8 N46
60.3 142.75 159.8
N20N56 N72 131.8192.1 201.55
N13 235.85 364.5N54
N31
79.75 251.6
60.3 240.05
142.75 241.1
N58
N80 234.8 256.5
79.75
111.25 N60 364.5 293.45
247.4 251.6 N63
N24
N26 241.1 N69 254.75
N55
294.5 131.8 361.35 79.75
111.25 10.05 289.85
233.75
N70 247.4 N38 N79 44.05
N52 68.05
131.8N2 54.55
159.8
294.5 141.7 136.3 N9 356.6 N7
233.75 N21 10.05
127.375 44.05
349.25
N30N40 N36
N3368.05 100.75 54.55
233.75 145.45 135.775 49.3 N77
141.7 127.375 N22
159.8 N5 349.25
N71
132.625 N42 212.75 149.75
100.75 330.35
49.3
233.75 N51 N15N76
N34
176.35 159.8
314.6 73.3
132.625
N45 228.5 330.35
212.75
N74
176.35 N59 222.75
0 257.9 373.95
228.5 N50 79.75
N1 132.25226.4 N82
0 193.15 N16
N28 N53 373.95
121.75 90.25 121.7573.45
N25 373.95 132.25226.4
N27 79.75
249.5
N81 N32
121.75 90.25 44.55 N17 121.75
N11 N62 373.9554.55
236.9 68.2
44.05
249.5280.1 79.75
56.8544.55 N4 249.5 N6
N73
356.6
N57 N66 236.9 N18N8 54.55
N78 54.55 79.75 44.05 137.5
244.25280.1 72.25 56.85
356.6 N19 N29
58.75 N14 137.5
268.25
64.5 79.75 159.8
N10 251.6 N39 N65
119.65 57.9 N61 152.8 269.1
N23
251.6
119.65 55.05
223.25 151.15
176.35
121.75 33.55 N12
N47 134.725N37 N48 N44 314.6 N3 55.05
202.6
The Time of critical path is: 373.95 N75
151.15
176.35 33.55 285.35 258.6
N35 121.75
314.6 N68 N64
The Total Cost is: 101 134.725
285.35258.6

Fig. 16. Results of Case study 4 with uncertainty solved by ACO.

the proposed method over traditional optimization algorithms is
the ability of the Ant system to produce good optimal and subop-
timal solutions. Other advantages of the proposed method can be
marked. The proposed method can be easily applied to more com-
plicated application in network analysis; like networks with uncer-
tainty (random or fuzzy data), crash analysis in CPM network, etc.
Finally the proposed method has introduced the power of ACO
algorithm in computing and analyzing PERT networks.
References
Blum, C., & Dorigo, M. (2004). The hyper-cube framework for Ant Colony
Optimization. Transactions on Systems, Man and Cybernetics – Part B:
Cybernetics, 34(2).
Bonabeau, E., Dorigo, M., & Theraulaz, G. (1999). Swarm intelligence from natural to
artificial systems. Oxford University Press.
Chanas, S., Dubois, D., & Zieliński, P. (2002). On the sure criticality of tasks in activity
networks with imprecise durations. IEEE Transactions on Systems, Man, and
Cybernetics – Part B, 32.
Chen, S.-P. (2007). Analysis of critical paths in a project network with fuzzy activity
times. European Journal of Operational Research, 183, 442–459.
Copertari, Luis F., & Archer, Norman P. (2001). Calculating the theoretical project
completion time of large networks in polynomial processing time. International
Conference on Management of Engineering and Technology, 1, 405.
Dorigo, M., Maniezzo, V., & Colorni, A. (1996). Ant system: Optimization by a colony
of cooperating agents. IEEE Transactions on Systems, Man, and Cybernetics – Part
B, 29–41.
Dubois, D., Fargier, H., & Fortemps, P. (2003). Fuzzy scheduling: Modelling flexible
constraints vs. coping with incomplete knowledge. European Journal of
Operational Research, 147, 231–252.
Dubois, D., Fargier, H., & Galvagnon, V. (2003). On latest starting times and floats in
activity networks with ill-known durations. European Journal of Operational
Research, 147(2), 266–280.
Fogel, D. (1993). Applying evolutionary programming to selected traveling
salesman problems. Cybernetics and Systems: An International Journal, 24,
27–36.
Gambardella, L. M., Dorigo, M., 1996. Solving symmetric and asymmetric tsps by ant
colonies. In IEEE conference on evolutionary computation (ICEC’96).
Gambardella, L. M., Taillard, E., Dorigo, M. (1997). Ant Colonies for QAP. Technical
Report Number: IDSIA 97-4, IDSIA, Lugano, Switzerland.
Lorterapong, P., & Moselhi, O. (1996). Project-Network Analysis Using Fuzzy Sets
Theory. Journal of Construction Engineering and Management, 122(4). Paper No.
8797.
Malcolm, D. G., Roseboom, J. h., Clark, C. E., & Fazar, W. (1959). Application of a
technique for research and development program evaluation. Operations
Research, 7, 646–669.
Paul Loomba, N. (1978). Management – A quantitative perspective. International
Editions: Collier Macmillan.
Sim, K. M., & Sun, W. H. (2003). Ant Colony Optimization for routing and load-
balancing: Survey and new directions. IEEE Transactions on Systems, Man, and
Cybernetics – Part A: Systems and Humans, 33(5) (November).
Yao, J.-S., & Lin, F.-T. (2000). Fuzzy critical path method based on signed distance
ranking of fuzzy numbers. IEEE Transactions on Systems, Man, and Cybernetics –
Part A: Systems and Humans, 30(1).