Advances in Engineering Software 39 (2008) 1023–1028

Contents lists available at ScienceDirect

Advances in Engineering Software

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

Genetic algorithm-based multi-objective model for scheduling of linear

construction projects
Ahmed Senouci *, Hassan R. Al-Derham
Department of Civil Engineering, University of Qatar, P.O. Box 2713, Doha, Qatar

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

Article history: This paper presents a genetic algorithm-based multi-objective optimization model for the scheduling of
Received 31 January 2006 linear construction projects. The model allows construction planners to generate and evaluate optimal/
Received in revised form 16 July 2007 near-optimal construction scheduling plans that minimize both project time and cost. The computations
Accepted 15 August 2007
in the present model are organized in three major modules. A scheduling module that develops practical
Available online 21 April 2008
schedules for linear construction projects. A cost module that computes the project’s costs. A multi-
objective module that searches for and identifies optimal/near-optimal tradeoffs between project time
Keywords:
and cost. An application example is analyzed to illustrate the use of the model and to demonstrate its
Genetic algorithms
Multi-objective optimization
capabilities in optimizing the scheduling of linear construction projects.
Linear construction projects Ó 2008 Elsevier Ltd. All rights reserved.
Scheduling
Cost control
Information technology (IT)

1. Introduction malized Selinger’s approach into a two-state variable, N-stage dy-

Linear construction projects are characterized by repetitive
activities. Examples of such construction projects include high-
ways, tunnels, railways, pipeline networks, high-rise buildings,
and housing development projects. In linear construction projects,
construction crews repeat the same work in various sections of the
project, moving from one section to another.
It is well documented that network scheduling methods such as
CPM and PERT are not suitable for the scheduling of linear con-
struction projects. Graphical methods, such as the line of balance
[1,2] have been developed for the scheduling of linear construction
projects. These methods provide tools suitable for the overall plan-
ning as well as summary level scheduling of such projects. How-
ever, they are not efficient in scheduling large linear construction
projects.
In the last three decades, a number of computerized methods
for the scheduling of linear construction projects were developed.
Selinger [11] presented the framework of the first dynamic pro-
gramming solution of linear construction projects. This formula-
tion, however, did not incorporate activity costs as decision
variables in the optimization process. Russell and Caselton [10] for-
* Corresponding author.
E-mail addresses: a.senouci@qu.edu.qa (A. Senouci), alderham@qu.edu.qa (H.R.
Al-Derham).
namic programming solution that determines the minimum
project time. This formulation also did not incorporate activity
costs as decision variables in the optimization process. Reda [9]
took another scheduling approach and developed a linear pro-
gramming formulation to minimize project costs by maintaining
constant production rates. The method can only be used for the
scheduling of linear projects with identical activity durations with-
in each section. Eldin and Senouci [6] developed a two-state vari-
able, N-stage dynamic programming formulation that included
both activity durations and costs in the optimization formulation
whose objective formulation was to determine the minimum pro-
ject cost. Hegazy and Wassef [8] presented a genetic algorithm
model for the scheduling of non-serial linear construction projects.
The objective of the model was to minimize project costs.
All these models are capable of generating a single optimal solu-
tion that either minimizes the project time or cost of linear con-
struction projects. There is a need for advanced models that can
help construction planners in generating and evaluating all the fea-
sible trade-off between project times and costs in order to select an
optimal schedule that satisfies the specific requirements with re-
spect to time and cost of the linear construction project being
considered.
The objective of this paper is to present the development of a
multi-objective model for the scheduling of linear construction
projects. The model provides planners and decision makers in

0965-9978/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.advengsoft.2007.08.002
1024 A. Senouci, H.R. Al-Derham / Advances in Engineering Software 39 (2008) 1023–1028
linear construction projects with an optimization model that is
capable of: (1) generating optimal/near-optimal resource utiliza-
tion plans that optimize construction time and cost and (2) visual-
izing the trade-offs among project time and cost in order to
support decision makers in evaluating the impact of various re-
source utilization plans on project performance.
2. Model formulation
The primary purpose of this development stage is to formulate a
robust optimization model that supports a two-dimensional time-
cost trade-off analysis. To this end, the present model is formulated
in two major steps: (1) determining the major decision variables in
this resource utilization problem and (2) formulating the two ma-
jor objectives of optimizing construction time and cost in a robust
optimization model.
2.1. Decision variables
For each construction activity in the project, the present model
is designed to consider all relevant decision variables that may
have an impact on project time and cost. This include: (1) the con-
struction method, which indicates the availability of different
types of materials and/or methods that can be utilized; (2) the
crew formation, which represents feasible sizes and configurations
for construction crews; and (3) the crew overtime policy, which
represents available overtime hours and nighttime shifts. In order
to control the complexity of the optimization model, the present
model combines these three major decision variables into a single
variable called a resource utilization option.
For example, the feasible resource utilization options for per-
forming the concrete paving activity in a highway project may in-
clude the utilization of: (1) construction material of concrete that
provides a strength of 30 or 35 MPa; (2) a crew formation A that
consists of one paving machine, one grader, and a labor force of
one foreman, three laborers, two equipment operators, and one ce-
ment finisher, or crew B that has the same composition except that
it has a larger and more powerful concrete paving machine; and (3)
and overtime policy of 0 or 4 h of overtime. Various combinations
of these available options can be represented by different resource
utilization options, as shown in Table 1. Each of these feasible re-
source utilization options has an expected daily production rate
and cost rate; and accordingly it leads to a time and cost for this
activity, as shown in Table 1.
Similarly, each of the remaining activities in the project can
have a similar set of feasible resource utilization options that can
be used to construct the activity. The major challenge confronting
construction planners in this problem is to select an optimal re-
source utilization option, from the available set of feasible alterna-
tives (n = 1 to N), for each activity (i = 1 to l) in the project. The
possible combinations of these alternatives create a large search
Table 1
Feasible resource utilization options for concrete paving
Resource utilization Resource composition Performance
option (n)
Material Crew Overtime Cost
(MPa) (h) ($/m2)
1 30 A 0 42
2 30 B 0 47
3 35 A 0 43
4 35 B 0 48
5 30 A 4 48
6 30 B 4 54
7 35 A 4 43
8 35 B 4 56
space, where each solution in this space represents a possible re-
source utilization option for delivering the project. For example,
a small-size project that includes 20 activities and 5 possible re-
source utilization for each activity creates a search space of
approximately 95 trillion (i.e. 520) possible solutions. The present
model is designed to help planners in the challenging task of
searching this large solution space in order to identify optimal re-
source utilization plans that achieve multiple project objectives.
2.2. Optimization objectives
The present optimization model is formulated in order to pro-
vide the capability of minimizing construction time and cost. The
model is also designed to quantify and measure the impact of var-
ious resource utilization decisions on performance in each of the
identified two project objectives. It incorporates two major objec-
tives functions as shown in the following two equations to enable
the evaluation of project performance in construction time and
cost, respectively
X
l
Minimize project time ¼ T ni
i¼1
where T ni is the duration of activity (i) on the critical path using re-
source utilization (n). In this model, the project time is computed
using previously developed algorithms for scheduling linear con-
struction projects [6]
X
l
Minimize project cost ¼ ½ðM ni þ Dni  Rni Þ þ ðBni Þ
i¼1
where Mni is the material cost of activity (i) using resource utiliza-
tion (n); Dni is the duration of activity (i) using resource utilization
(n); Rni is the daily cost rate in $/day in activity (i); and Bni is the sub-
contractor lump sum cost for resource utilization (n) in activity i, if
any.
3. Model implementation
In order to support decision makers in their search for optimal
trade-offs between time and cost, the present model is imple-
mented using a multi-objective genetic algorithm. Genetic algo-
rithms are search and optimization tools that assist decision
makers in identifying optimal or near-optimal solutions for prob-
lems with large search space. They are inspired by the mechanics
of evolution and they adopt the survival of the fittest and the struc-
tured exchange of genetic materials among population members
over successive generations as a basic mechanism for the search
process [7]. As such, the present model is implemented in three
major phases: (1) initialization phase that generates an initial set
of (S) possible solutions for this resource utilization problem; (2)
fitness evaluation phase that calculates the cost and time of each
generated solution; and (3) population generation phase that seeks
to improve the fitness of solutions over successive generations. The
detailed computation procedure in these three phases is explained
in the following sections and is shown in Fig. 1.
3.1. Phase 1: initialization
Productivity
(m2/day)
The main purpose of this phase is to initialize the optimization
2100 procedure in the present model, using the following two major
2500
steps.
2100
2500
2600 (1) Read project and genetic algorithm parameters needed to
3100 initialize the search process. The project parameters include:
2600
(1) number of project activities and sections; (2) number of
3100
resource utilization options for each activity; (3) activity
 A. Senouci, H.R. Al-Derham / Advances in Engineering Software 39 (2008) 1023–1028 1025

Start

Phase 1: Initialization
Read project and genetic algorithm parameters

Generate random solutions (s = 1 to S) for parent population P1 of

first generation (g = 1)

Phase 2: Fitness Function Evaluation

Phase 3: Population Generation
Solution (s=1)
Yes

Child population Cs
Compute project time for solution s in
generation g
No

Calculate Pareto optimal rank and crowding

Compute project cost for solution s in distance for each solution (s=1 to S) of the
generation g parent population Pg

Create a child population Cg using selection,

Yes crossover, and mutation
Last solution S
Combine child population Cg and parent
population Pg to form a new combined
population Ng of size 2S
No

Calculate Pareto optimal rank and crowding

Next solution ( s + 1)
distance for each Solution (s=1 to 2S) of the
combined population Ng

Sort new combined population Ng using

Niched comparison

Keep top S solutions to form next generation's

parent population Pg+1

Yes
Last generation g

No

Next generation g = g + 1
End

Fig. 1. Model computational flowchart.

duration and direct cost for each resource utilization option
and for each section; and (4) lag/lead time between succes-
sive activities and their precedence relationships. The
required genetic algorithm parameters for this initialization
phase include: (1) string size; (2) number of generations; (3)
population size; (4) mutation rate; and (5) crossover rate.
The string size is determined by the model, considering
the total number of construction activities included in the
analyzed project. The number of generations (G) and popula-
tion size (S) are identified based on the selected string size in
order to improve the quality of the solution. The population
size typically contains between 30 and 500 individuals/
strings [7]. The crossover rate represents the probability that
two strings will swap their bits. The crossover operation cre-
ates variations in the solution population by producing new
solution strings that consist of parts taken from selected par-
1026 A. Senouci, H.R. Al-Derham / Advances in Engineering Software 39 (2008) 1023–1028
ent solution strings. Its value varies between 0.7 and 0.9. A
value of 0.8 is commonly used. The mutation rate represents
the probability that a bit within a string will be flipped (0
becomes 1, 1 becomes 0). The mutation operation introduces
random changes in the solution population. In genetic algo-
rithms, the mutation operation can be beneficial in reinforc-
ing diversity in a population. The mutation rate has usually a
very low value for binary encoded genes, say 0.005.
(2) Generate random solutions (s = 1 to S) for the initial popula-
tion P1 in the first generation (g = 1). These solutions repre-
sent an initial set of activity resource utilization options.
This set of possible solutions is then evolved in the following
two phases in order to generate a set of activity optimal
resource utilization options that establishes an optimal
trade-off between project time and cost.
3.2. Phase 2: fitness functions evaluation
The main purpose of this phase is to evaluate the time and cost
for each possible solution (s) in generation (g) in order to deter-
mine the fitness of the solution. This fitness determines the likeli-
hood of survival and reproduction of each solution in following
generations. As such, this phase evaluates the two identified fitness
functions for each solution using the following two steps:
(1) Calculate the project time (D(s, g) for solution (s) in genera-
tion (g)).
(2) Calculate the project cost (Cost(s, g) for solution (s) in gener-
ation (g)).
3.3. Phase 3: population generation
The purpose of this phase is to create three types of population in
each of the considered generations: (1) parent; (2) child; and (3)
combined. For each generation (g), a parent population (Pg) is used
to generate a child population (Cg) in a manner similar to that used
in traditional genetic algorithms [7]. The purpose of generating this
child population is to introduce a new set of solutions by rearrang-
ing and randomly changing parts of the solutions of the parent pop-
ulation. This child population can then be combined with the parent
population to create an expanded set of possible solutions that
forms the combined population (Ng) for generation (g). This com-
bined population (Ng) is used to facilitate the comparison among
the initial solutions in the parent population and those generated
in the child population. The best solutions in this combined popula-
tion regardless of their origin are retained and passed to the follow-
ing generation as a parent population [3–5,12]. The computational
procedure in this phase is implemented in the following steps.
(1) Calculate Pareto optimal rank and crowding distance for
each solution (s = 1 to S) in the parent population (Pg). First,
this is done by ranking the solutions in the population
according to their Pareto optimal domination of other solu-
tions, where a solution is identified as dominant if it is better
than all other solutions in all of the optimization objectives
considered simultaneously. For example, solution A in Fig. 2
can be considered a Pareto optimal solution that dominates
(i.e. provides lower cost and time than) solution E. Second,
this step calculates the crowding distance of each solution,
which represents the closeness of neighboring solutions to
the solution considered. The crowding distance values help
the algorithm spread the obtained solutions over a wider
Pareto front instead of converging to points that cover only
a small part of the tradeoff surface [4].
Fig. 2. Pareto optimal solutions.
(2) Create a new child population (Cg) using the genetic algo-
rithm operations of selection, crossover, and mutation. The
selection operation chooses the individuals that will go
through the reproduction process, by favoring those with
higher optimal ranks and wider crowding distances. The
crossover operation, on the other hand, crosses each pair
of the selected individuals at a randomly determined point
and swaps the variables coded in the springs at this point,
resulting in two new individuals. The mutation operation
randomly changes the value of one of the variables in the
string to induce innovation and to prevent premature con-
vergence to a local optima [7]. The fitness of the generated
child population is then analyzed using the earlier described
steps of phase 2 in order to obtain the values of project time
and cost for each solution as shown in Fig. 1.
(3) Combine child population (Cg) and parent population (Pg) to
form a new combined population (Ng) of size (2S). This com-
bined population acts as a vehicle for the elitism, where
good solutions of the initial parent population are passed
on to the following generation to avoid the loss of good solu-
tions of the initial parent population once they are found [4].
(4) Calculate Pareto optimal rank and crowding distance for
each solution (s = 1 to 2S) of the newly created combined
population (Ng). This step performs the same operations as
step 1 of this phase on the new combined population (Ng).
(5) Sort the new combined population (Ng) using the niched
comparison rule. This sorting rule selects solutions with
higher Pareto optimal ranks and breaks ties between solu-
tions with the same rank by favoring solutions with higher
crowding distances.
(6) Keep the top (S) solutions from the combined population
(Ng) to form the parent population (Pg+1) of the next genera-
tion. This parent population is then returned to step 1 of this
phase for generating a new child population as shown in
Fig. 1. This iterative execution of the second and third phases
of the model continues until the specified number of gener-
ations is completed.
4. Illustrative example
A project example is analyzed in order to illustrate the use of
the present optimization model and demonstrate its capabilities.
 A. Senouci, H.R. Al-Derham / Advances in Engineering Software 39 (2008) 1023–1028 1027

This example was originally presented by Reda [9] and has been Table 3
modified to allow for multiple resource utilization options. The Optimal solutions and their impact on project performance

example project is the relocation of 18 km of natural gas pipeline. Project Project Resource utilization option
The repetitive activities in the pipeline relocation have been iden- time cost $
Activity Activity Activity Activity Activity Activity
tified as locate and clear, excavate, string pipe, lay pipe, test, and (work
1 2 3 4 5 6
days)
backfill, as shown in Fig. 3. The project is divided into five sections.
Each activity is performed by a single crew progressing from the 14 4000 1 4 2 5 1 1
19 3250 1 4 2 4 1 1
first to the fifth section. The precedence relationships among suc- 20 2750 1 3 1 4 1 1
ceeding activities are finish to start with no lag time. Table 2 sum- 24 2500 1 4 2 3 1 1
marizes the input values for the example project. 25 2000 1 3 1 3 1 1
The present optimization model was used to search the space of 26 1750 1 2 1 3 1 1
30 1250 1 3 1 2 1 1
possible solutions using the developed multi-objective optimiza-
31 1000 1 2 1 2 1 1
tion model. The rate of crossover and mutation were set equal to 32 750 1 1 1 2 1 1
their most commonly used values (i.e. 0.8 and 0.005, respectively). 35 500 1 3 1 1 1 1
After a number of trial-and-error adjustments, a population size 36 250 1 2 1 1 1 1
equal to 30 individuals and a number of generations equal to 150 37 0 1 1 1 1 1 1

were found to meet the accuracy requirements of the example.

The model was able to significantly reduce the search space by
precluding dominated solutions in the successive generations of 4800
the genetic algorithm, using the Pareto optimality principles 4400
shown in Fig. 2. This led to the selection of 12 Pareto optimal 4000
(i.e. non-dominated) solutions for this example. Each of these solu- 3600

Project Costs ($)

tions identifies an optimal trade-off among project time and cost. 3200
Table 3 summarizes these optimal solutions and their impact on 2800
project performance. Fig. 4 shows the time-cost trade-off curve 2400
of the project. In Fig. 4, the horizontal axis represents project times 2000
and the vertical axis representing project additional costs as a re- 1600
sult of reducing project times. As can be seen from Fig. 4, the pro- 1200
ject time varies from 37 to 14 days. The values of the different
800
points on the curve are obtained from Table 3. Figs. 5 and 6 show
400
the 37-days and 14-days project schedules, respectively. In order
0
to validate the results provided by the present model, they are 14 16 18 20 22 24 26 28 30 32 34 36 38
compared to those reported in the literature for the same applica- Project Duration (Work Day)
tion example [9]. The comparison confirms that the present model
Fig. 4. Project time-cost trade-off curve.

Excavate
Locate &Clear Excavate String Pipe Lay Pipe Test Backfill

40

Locate and String Lay

Test Backfill 35
Clear Pipe Pipe
Project Time (Work Day)

30
Fig. 3. Project network diagram.

25

Table 2
Activity cost and duration data for proposed project example 20

Activity Activity Resource Activity Activity 15

description number utilization duration additional cost ($/
option number (d) d)
10
Locate and 1 1 1 0
clear
Excavate 2 1 4 0 5
2 3 50
3 2 100 0
4 1 150 0 1 2 3 4 5
String pipe 3 1 2 0
Pipeline Section
2 1 50
Fig. 5. Project schedule for project duration of 37 days.
Lay pipe 4 1 5 0
2 4 150
3 3 300 is capable of generating the same set of optimal solutions as those
4 2 450
reported by Reda [9]. Furthermore, the present model has the
5 1 600
added capability of allowing multiple resource utilization options,
Test 5 1 1 0
multiple precedence-relationships, and a general time-cost trade-
Backfill 6 1 2 0
off analysis procedure.
1028 A. Senouci, H.R. Al-Derham / Advances in Engineering Software 39 (2008) 1023–1028

Locate & Clear String Pipe Excavate Lay Pipe Test Backfill ti-objective optimization module that searches for and identifies
16 near-optimal construction plans. An application example was ana-
lyzed to illustrate the capabilities of the developed model in gener-
14 ating all optimal trade-off solution between project time and cost
in a single run, where each provides the least project time that
Project Time (Work Day)

12
10
8
6
4 Manage, ASCE 1990;116(4):689–704.
2 Constr Div, ASCE 1986;112(3):411–24.
0 NY: John Wiley & Sons, Ltd.; 2001.
0 1 2 3
Pipeline Section
Fig. 6. Project schedule for project duration of 14 days.
5. Conclusions
A robust multi-objective optimization model was developed to
support the scheduling of linear construction projects. The model
enables construction planners to generate and evaluate optimal
construction plans that establish optimal trade-offs between pro-
ject duration and total cost. Each of these plans identifies, from a
set of feasible alternatives, an optimal resource utilization option
for each project activity. To accomplish this, the model incorpo-
rates: (1) a scheduling module that calculate the project time;
(2) a cost module that computes the project’s costs; and (3) a mul-
can be achieved for a given project cost. This new capability should
prove useful to construction planners and is expected to advance
existing scheduling practices for linear construction projects.
References
[1] Al Serraj ZM. Formal development of line-of-balance technique. J Constr Eng
[2] Arditi D, Albulak ZM. Line-of-balance scheduling in pavement construction. J
[3] Deb K. Multi-objective optimization using evolutionary algorithm. New York,
[4] Deb K, Agrawal S, Pratap A, Meyarivan T. A fast elitist non-dominated sorting
4 5
genetic algorithm for multi-objective optimization. KANGAL Report No.
200001. Kanpur, India: Genetic Algorithm Laboratory, Indian Institute of
Technology; 2001.
[5] Deb K, Pratap A, Agrawal S, Meyarivan T. A fast elitist multi-objective NSGA-II.
IEEE Trans Evolut Comput 2002;6(2):182–97.
[6] Eldin NN, Senouci AB. Scheduling and control of linear projects. Canad J Civil
Eng 1994;21(2):219–30.
[7] Goldberg DE. Genetic algorithms in search optimization and machine
learning. Reading, Mass: Addison-Wesley Publishing Company, Inc.; 1989.
[8] Hegazy T, Wassef N. Cost optimization in projects with repetitive nonserial
activities. J Constr Eng Manage, ASCE 2001;127(3):183–91.
[9] Reda RM. RPM: repetitive project modeling. J Constr Eng Manage, ASCE
1990;116(2):316–30.
[10] Russell AD, Caselton WF. Extensions to linear scheduling optimization. Phys
Rev Lett 1988;114(1):36–52.
[11] Selinger S. Construction planning for linear projects. J Constr Div, ASCE
1980;106(2):195–205.
[12] Zitzler E, Thiele L. Multiobjective evolutionary algorithms: a comparative case
study and the strength Pareto approach. IEEE Trans Evolut Comput
1999;3(4):257–71.