Expert Systems With Applications 51 (2016) 177–185

Contents lists available at ScienceDirect

Expert Systems With Applications

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

Discrete particle swarm optimization method for the large-scale

discrete time–cost trade-off problem
Saman Aminbakhsh, Rifat Sonmez∗
Department of Civil Engineering, Middle East Technical University, Ankara 06800, Turkey

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

Keywords: Despite many research studies have concentrated on designing heuristic and meta-heuristic methods for
Project management the discrete time–cost trade-off problem (DTCTP), very little success has been achieved in solving large-
Particle swarm optimization
scale instances. This paper presents a discrete particle swarm optimization (DPSO) to achieve an effective
Discrete time–cost trade-off problem
method for the large-scale DTCTP. The proposed DPSO is based on the novel principles for representation,
Construction projects
initialization and position-updating of the particles, and brings several benefits for solving the DTCTP,
such as an adequate representation of the discrete search space, and enhanced optimization capabilities
due to improved quality of the initial swarm. The computational experiment results reveal that the new
method outperforms the state-of-the-art methods, both in terms of the solution quality and computation
time, especially for medium and large-scale problems. High quality solutions with minor deviations from
the global optima are achieved within seconds, for the first time for instances including up to 630 ac-
tivities. The main contribution of the proposed particle swarm optimization method is that it provides
high quality solutions for the time–cost optimization of large size projects within seconds, and enables
optimal planning of real-life-size projects.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction problem and involves determination of the complete and non-

In project scheduling, the project duration can be shortened by
allocating additional resources to the activities. However, short-
ening the project durations below their normal levels comes at
an additional cost. The objective of time–cost trade-off problem
(TCTP) is to identify the set of time–cost alternatives that will
provide the optimal schedule under certain conditions. Since in
practice many resources (e.g., crews, equipment) are available in
discrete units, numerous researches have focused on the discrete
version of this problem, called the discrete time–cost trade-off
problem (DTCTP).
In the literature, three types of the DTCTP have been included
commonly; the cost optimization problem, the duration optimiza-
tion problem and the Pareto front problem. The objective of the
cost optimization problem is to determine the set of time–cost al-
ternatives that will minimize the total cost under certain condi-
tions such as: a given strict project deadline, or a delay penalty.
The budget problem aims to identify the time–cost alternatives
to minimize the project duration without exceeding a given bud-
get. The Pareto front problem is a multi-objective optimization
∗
Corresponding author. Tel.: +90 312 210242.
E-mail addresses: saman.aminbakhsh@metu.edu.tr (S. Aminbakhsh), rsonmez@
metu.edu.tr, rifats@hotmail.com (R. Sonmez).
dominated time–cost profile over the set of feasible project dura-
tions (Vanhoucke & Debels, 2007) which is called the Pareto front.
This paper focuses on the cost optimization problem.
The methods proposed for the DTCTP could be categorized into
three areas: exact methods, heuristics, and meta-heuristics. All
three types of DTCTP are NP-hard in the strong sense (De, Dunne,
Ghosh, & Wells, 1997). Hence, few studies presented exact meth-
ods for the DTCTP to solve small and medium scale instances.
De, Dunne, Ghosh, and Wells (1995) presented network decom-
position/reduction approaches for solving the Pareto front prob-
lem. Demeulemeester, Reyck, Foubert, Herroelen, and Vanhoucke
(1998) presented a branch and bound algorithm for the Pareto
front problem, and solved instances up to 50 activities. Vanhoucke
(2005) proposed a branch and bound algorithm for the cost op-
timization problem considering strict deadlines and time-switch
constraints and achieved optimal solutions for instances up to 30
activities. Akkan, Drexl, and Kimms (2005) provided lower and up-
per bounds for the cost optimization problem with strict deadlines.
Hazır, Haouari, and Erel (2010) presented an exact method based
on Benders Decomposition for the duration optimization problem,
and was able to solve instances including up to 136 activities and
10 modes within 90 min. Szmerekovsky and Venkateshan (2012)
studied four integer programming formulations for irregular time–
cost trade-offs and achieved optimal solutions for instances with
up to 90 activities.

http://dx.doi.org/10.1016/j.eswa.2015.12.041
0957-4174/© 2016 Elsevier Ltd. All rights reserved.
178 S. Aminbakhsh, R. Sonmez / Expert Systems With Applications 51 (2016) 177–185
In an early attempt to achieve an efficient method for solv-
ing the TCTP, Siemens (1971) presented a heuristic called the
Siemens approximation method (SAM) for the cost optimization
problem with strict deadlines, and implemented it on an exam-
ple including eight activities. Goyal (1975) proposed a modified
version of the Siemens approximation method and used the same
example with eight activities to demonstrate the modified heuris-
tic. Numerous meta-heuristic solution procedures have been pre-
sented in the literature for the DTCTP. Genetic algorithms (GAs) are
among the most commonly used meta-heuristics for the DTCTP.
Feng, Liu, and Burns (1997) proposed a GA for Pareto front op-
timization. Hegazy (1999) developed a GA for the cost optimiza-
tion problem. Zheng, Ng, & Kumaraswamy, 2005 proposed a GA
based multiobjective model for the Pareto problem. Kandil and El-
Rayes (2006) explored the performance of supercomputing clusters
through a GA for Pareto front optimization. Eshtehardian, Afshar,
and Abbasnia (2008) presented a GA for Pareto Front optimiza-
tion of stochastic DTCTP. Fallah-Mehdipour, Bozorg Haddad, Reza-
pour Tabari, and Mariño (2012) considered a nondominated sort-
ing genetic algorithm along with a multi-objective particle swarm
optimization method for Pareto front optimization of DTCTP and
time–cost-quality trade-off (TCQTO) problems. Sonmez and Bet-
temir (2012) developed a hybrid strategy based on GAs, simu-
lated annealing, and quantum simulated annealing techniques for
the cost optimization problem. Zheng (2015) presented a GA for
the discrete time–cost–environment trade-off problem. Zhang, Zou,
and Qi (2015) proposed a GA for the DTCTP in repetitive projects.
Ant colony optimization, shuffled frog leaping, tabu search,
Electimize, ε -constraint based evolutionary algorithm, and particle
swarm optimization (PSO), are among the meta-heuristic methods
proposed for the DTCTP, other than GAs. Afshar, Ziaraty, Kaveh, and
Sharifi (2009), Ng and Zhang (2008) and Xiong and Kuang (2008)
proposed ant colony optimization algorithms for the Pareto front
problem. Elbeltagi, Hegazy, and Grierson (2007) presented a shuf-
fled frog-leaping optimization algorithm for the cost optimization
problem. Ashuri and Tavakolan (2015) considered the Pareto front
optimization of resources along with the time and cost and pre-
sented a shuffled frog leaping algorithm. Vanhoucke and Debels
(2007) developed a meta-heuristic approach involving tabu-search
and truncated dynamic programming for the three extensions of
the cost optimization problem with strict deadlines. Abdel-Raheem
and Khalafallah (2011) also focused on the cost optimization prob-
lem, and proposed an evolutionary algorithm which simulates the
behavior of electrons moving through electric circuit branches.
In a recent study, Tavana, Abtahi, and Khalili-Damghani (2014)
presented two multi-objective procedures based on ε -constraint
method and dynamic self-adaptive evolutionary algorithm for solv-
ing the discrete time–cost-quality trade-off problem.
Elbeltagi, Hegazy, and Grierson (2005) and Bettemir (2009) ex-
plored the potential of PSO for the cost optimization problem. Yang
(2007) and Zhang and Xing (2010) proposed multi-objective PSO
algorithms for the Pareto front problem. In a comparison of five
evolutionary based algorithms for the cost optimization problem,
PSO had the best performance (Elbeltagi et al., 2005). Among eight
meta-heuristic methods, including a sole genetic algorithm, four
hybrid genetic algorithms, PSO, ant colony optimization, and elec-
tromagnetic scatter search, PSO was one of the top performing al-
gorithms along with the hybrid genetic algorithm with quantum
simulated annealing for the large-scale cost optimization problem
(Bettemir, 2009). Although DTCTP is a discrete optimization prob-
lem, to our best knowledge, all of the previous PSO research on
DTCTP are restricted in real number space.
The majority of the previous DTCTP research (Abdel-Raheem &
Khalafallah, 2011; Afshar et al., 2009; Elbeltagi et al., 2007; Es-
htehardian et al., 2008; Fallah-Mehdipour et al., 2012; Feng et al.,
1997; Hegazy, 1999; Ng & Zhang, 2008; Xiong & Kuang, 2008;
Yang, 2007; Zhang & Xing, 2010; Zheng et al., 2005) used few
problem instances including up to eighteen activities to evaluate
the performances of the proposed meta-heuristics and did not in-
clude the optimal results in comparisons. Even majority of the re-
cent methods was tested with problems including up to twenty ac-
tivities (Ashuri & Tavakolan, 2015; Zhang et al. 2015; Zheng 2015).
Vanhoucke and Debels (2007) included instances up to fifty ac-
tivities in computational experiments. Tavana et al. (2014) gener-
ated instances up to 100 activities for solving the discrete time–
cost-quality trade-off problem. Very few studies have focused on
optimization of large-scale discrete time–cost trade-off problems.
Kandil and El-Rayes (2006) used instances up to 720 activities;
however the proposed genetic algorithm required 21 h with a sin-
gle processor to obtain a Pareto front for a problem including
360 activities. For instances including 630 activities, the best of
the eight meta-heuristics tested by Bettemir (2009) was able to
achieve solutions with a 2 percent deviation from the optimum in
73 min.
Despite a large amount of the research on the DTCTP has con-
centrated on designing heuristics and meta-heuristics, very few
of the proposed methods can be applied to real-life construc-
tion projects which typically include more than 300 activities
(Liberatore, Pollack-Johnson, & Smith, 2001). Besides, a few meth-
ods that are capable of solving large-scale discrete time–cost trade-
off problems usually require a significant amount of computation
time to achieve high quality solutions. Hence, for the time–cost
trade-off problem there is a significant gap between the literature
and real-life project management (Vanhoucke, 2005).
The main objective of this paper is to develop a discrete particle
swarm optimization method that is capable of providing high qual-
ity solutions for the large-scale discrete time–cost trade-off cost
optimization problems within short computational time. The pro-
posed method attempts to improve existing methods for DTCTP by
designing a PSO, which can adequately represent the discrete so-
lution space of the DTCTP. A modified version of the Siemens ap-
proximation method is integrated to the new discrete PSO (DPSO)
to improve the quality of the initial swarm for accelerating the op-
timization. The paper aims to fill the gap in the literature by pre-
senting a method that can handle the cost optimization problem
for real-life-size projects. The remainder of the paper is organized
as follows: in Section 2, a mixed-integer formulation is presented
for the DTCTP. Section 3 is devoted to the novel discrete particle
swarm optimization method. The results of the computational ex-
periments are presented in Section 4, and concluding remarks are
made in Section 5.
2. Discrete time–cost trade-off problem
The general discrete time–cost trade-off cost optimization prob-
lem in which the objective is to minimize the sum of direct and
indirect costs can be formulated as follows (De et al., 1995):
 
S m ( j)
minimize (dc jk x jk ) + D × ic (1)
j=1 k=1
subject to:
m( j )

x jk = 1, ∀ j = {1, ..., S} (2)
k=1
m( j )

d jk x jk + St j ≤ Stl , ∀ l ∈ Sc j and ∀ j = {1, ..., S} (3)
k=1
D ≥ StS+1 (4)
where dcjk is the direct cost of mode k for activity j; xjk is a 0-1
variable which is 1 if mode k is selected for executing activity
 S. Aminbakhsh, R. Sonmez / Expert Systems With Applications 51 (2016) 177–185 179

j, and 0 otherwise; ic is the daily indirect cost; D is the project
duration; djk is the duration of mode k for activity j; Stj is the
start time for activity j; and Scj is the set of immediate successors
for j.
3. Discrete particle swarm optimization method for DTCTP
In this section following a background of particle swarm
optimization and Siemens approximation method, the proposed
discrete particle swarm optimization method for the DTCTP is
described.
(DPSO) method, a semi-random initialization scheme is imple-
mented to accelerate the swarm optimization.
The DPSO creates a certain percentage (pct) of the initial pop-
ulation by using the modified SAM method, and generates the re-
maining particles randomly. The solutions obtained by using SAM
are represented by yijk , while, each xijk denotes a particle generated
by PSO phase. In DPSO, each yijk and xijk is represented by a binary
value that holds solution i’s position for the jth activity, which can
only have a single k value equal to one, and all the remaining posi-
m( j )
tions for the jth activity are set to zero ( k=1 x(jkt ) = 1). Precedence
constraints are satisfied according to Eq. (6).

ES(jt ) + d (jt ) − ESl( ) ≤ 0, ∀ l ∈ Sc j

t
(6)
3.1. Particle swarm optimization
where t is the generation number; dj is activity j’s duration; ESj is
The first paradigm of the PSO was developed by Eberhart & the early start time of jth activity; and Scj are the immediate suc-
Kennedy (1995) and Kennedy & Eberhart (1995) based upon the cessors of the jth activity. In the first iteration, bound by the fea-
principles of the swarm intelligence who later introduced a bi- sible region [−vmax , vmax ], random velocity vectors, vi(jk
1)
, are gener-
nary version of this algorithm for problems with discrete search ated for all the initial seeds (i.e., deterministic and random parti-
spaces (Kennedy & Eberhart, 1997). PSO is rooted upon imitating cles). Generation of the initial particles is completed with the de-
the choreography of bird flocks that communicate together as they t t
termination of the pbest, Pi 0 and the gbest, Pg0 positions. Each par-
fly; therefore, the population is called “swarm”, while, the poten- ticle acquires the “best” positions, using Eq. (7) as follows:
tial solutions are named as “particles”. Particles iteratively fly over 
the search space in explicit directions, and are attracted to self- Pitjk
0
= Pgt0jk = yi jk , i = [ 1, pct.N]
(7)
= Pgt0jk = xi(jk) ,
1
attained historical best position (personal best; pbest), as well as Pitjk
0
i = ( pct.N, N]
the best position among the entire swarm (global best; gbest). Each
particle records the coordinates of the best location it has visited where N is the number of birds. Each particle i’s fitness is evalu-
so far. At each time step, particles evaluate their own positions ated with respect to Eqs. (8) and (9), which minimize the sum of
with respect to fitness criteria, then by comparing the fitness val- direct and indirect costs.
ues they communicate to identify the particle located at the best  

S 
m
position. Each particle moves towards the best position using a ve- d (jk) x(jk)
t t
Di = max (8)
locity that incorporates coordination of the personal best location j=1 k=1
as well. Each particle, then evaluates the domain from its new lo-
cation, and the process reiterates until either the swarm reaches to 
S 
m
a predefined target, or a computational limit. Ci = dc(jkt ) x(jkt ) + Di × ic (9)
j=1 k=1

3.2. Siemens approximation method

The Siemens approximation method – SAM (Siemens, 1971),
is one of the first heuristics developed for TCTP. SAM was orig-
inally developed for the continuous time–cost trade-off problem.
We have modified SAM for the discrete version of the time–cost
trade-off problem to improve the quality of the initial swarm.
This modified procedure is initiated with the construction of the
project network using the normal (uncrashed) modes. The activity
with minimum cost-slope is identified and crashed according to
Eq. (5):
CS j = (C jk − C j (k−1) )(D jk − D j (k−1) )−1
∀ j = {1, ..., S}, ∀ k = {1, ..., m( j )}
where S is the number of activities, m(j) is the number of available
time–cost options for activity j, Cjk is the direct cost of kth time–
cost option of activityj, Djk is the duration of thekth time–cost op-
tion of activity j. The cost-slopes of the first network are evaluated
by setting k = m( j ) initially; an then, decreasing the value of op-
tion k, one at a time, as the jth activity is crashed.
3.3. Discrete particle swarm optimization method
Few research has presented PSO algorithms for the DTCTP
(Bettemir, 2009; Fallah-Mehdipour et al., 2012; Yang, 2007; Zhang
& Xing, 2010) which have operated in continuous space. In this
paper, a discrete particle swarm optimization algorithm is de-
signed to achieve an improved particle swarm representation for
the DTCTP. Previous research also generated the initial particles
randomly. In the proposed discrete particle swarm optimization
∀ j = {1, . . . , S}, ∀ k = {1, . . . , m}
where djk and dcjk represent duration and direct cost of the kth
options of the jth activity, respectively; and ic denotes the daily
indirect cost.
The quality of the solutions are compared with each other ac-
cording to Eq. (10):
u>v if Cu < Cv (10)
which makes the discrimination in favor of decision vector u in
case the total cost of that particle is less than decision vector v,
while, in the case of equality (Cu = Cv ) discrimination is made in
(5) favor of the particle with smaller duration by Eq. (11).

Cu = Cv
u>v if (11)
Du < Dv
For the occasion in which both particles u and v have the same
total costs and durations, discrimination is made randomly. Pi ’s and
Pg ’s are updated after the better fitted individuals are identified.
Particles are flown to their new positions using the velocity vec-
tor given in Eq. (12), which incorporates Kennedy and Eberhart’s
(1997) linearly decreasing time-varying inertia weight (w).
   
vi(jkt+1) = w(t ) vi(jkt ) + c1 r1 Pi(jkt ) − xi(jkt ) + c2 r2 Pg(tjk) − xi(jkt ) (12)
where r1 and r2 are random vectors with uniformly distributed
components within the range [0, 1], and the constants c1 and c2
are the cognitive and social parameters, respectively.
The velocity vectors are transformed into probabilities
(Aminbakhsh, 2013) and are normalized to the range [0, 1]
180 S. Aminbakhsh, R. Sonmez / Expert Systems With Applications 51 (2016) 177–185

Fig. 1. Case Example.

Modes
Activities k1 k2 k3 k4 k5
j1 0.50 0.11 0.71 0.03 0.73
j2 0.94 0.94 0.97 0.62 -
j3 0.82 0.12 0.27 - -
j4 0.75 0.96 0.31 0.91 0.02

Fig. 2. Probability matrix. Fig. 3. Position matrix.

using a logistic transformation function given in Eq. (13). Each par- representation was used instead, a dominant discrimination of the
ticle is then migrated to a new position subject to the probabilistic selected modes would not be always possible, as in many cases
condition according to Eq. (14). the position of the selected mode could be close to the position of
  1
an undesired mode.
sig vi(jkt ) = (13) The optimization process is reiterated until the pre-set number
1 + exp (−vi(jk) )
t
of iterations is reached. DPSO will return the final gbest particle as
 the solution for the DTCT problem when the optimization process
1 if sig (vi(jk ) ) = max sig (vi(jk ) )
t+1 t+1
(t+1 ) is terminated. The flowchart and the pseudo-code of DPSO algo-
xi jk = (14)
0 otherwise rithm are illustrated in Figs. 4 and 5, respectively.

Eq. (14) differs from the position update equation of the binary
PSO proposed by Kennedy and Eberhart (1997), in such a way that,
it involves determination of the alternative(s) associated with the
maximum amount of probability for every activity. This condition
indicates that in each row of position matrix, a single alternative
with the largest probability will have a value of one. If the value of
max{sig(vi(jk
t+1 )
)} is same for more than one alternative, then, dis-
crimination is made randomly.
DPSO’s binary representation for the particles’ positions is il-
lustrated for a case example given in Fig. 1. The probability matrix
for the 4th particle in the 3rd iteration for the case example is
illustrated in Fig. 2. The velocity matrix (v4(3 ) ), which is calculated
by Eq. (12), is transformed to the probability matrix (sig(v4(3 ) ))
using the sigmoid function given in Eq. (13). The 4th particle’s
new position matrix (x4(4 ) ) in the 4th iteration is determined by
using Eq. (14) as shown in Fig. 3. The discrete position matrix in
Fig. 3, assigns a value of “one” to the mode selected, and a value
of “zero” to the remaining modes, to enable a dominant discrimi-
nation of the selected modes. For example, for the first activity the
fifth mode is selected, hence the position x15 is assigned a value
of “one”, and the remaining positions of the first activity (x11 ,
x12 , x13 , x14 ) are assigned a value of “zero”. If continuous position
4. Computational experiments
Computational experiments are conducted to evaluate the per-
formance of the proposed DPSO method for the DTCTP using
benchmark instances. The proposed algorithm is coded in C++ and
compiled within Visual Studio 2013. All of the tests were carried
out on a computer with an Intel Core i7-3.40 GHz CPU. Pilot exper-
iments were conducted to determine an adequate set of parameter
values for the DPSO. The pilot experiments revealed that the set
of parameters that are summarized in Table 1 provided an ade-
quate combination for the DPSO. 50,000 schedules (objective func-
tion evaluations) are used as the stopping criteria in all of the ex-
periments (Kolisch & Hartmann, 2006; Sonmez & Bettemir, 2012).
Since PSO is a stochastic search algorithm, performance of DPSO
is evaluated ten times for existing benchmark instances, and av-
erage percent deviation (APD) from the global optimal solution is
reported.
4.1. Small-scale instances
The performance of the novel DPSO method was first tested
using the small-scale DTCTP test instances which are commonly
 S. Aminbakhsh, R. Sonmez / Expert Systems With Applications 51 (2016) 177–185 181

Start

Randomly
Feed initial
Generate Precedence
solutions
Remaining Constraint Valid
to PSO
Construct Particles
Project Network YES
NO

Calculate Determine
Determine Total Cost Duration
Duration

Compare
pbest has change
Calculate Fitness
Total Cost
YES
NO

YES
Update pbest gbest has changed Update gbest
Store Solution

NO

YES
All critical Transform Velocity
Evaluate Velocity
activities to Probability
crashed

NO
NO Update Inertia
Update Position Stop Iteration
Weight
Evaluate
Cost Slopes

YES

Crash least Output gbest

Cost Slope
in Critical path

End

Fig. 4. Flowchart of the proposed discrete PSO algorithm.

used in the literature (Afshar et al., 2009; Elbeltagi et al., 2005;
Elbeltagi et al., 2007; El-Rayes & Kandil, 2005; Fallah-Mehdipour
et al., 2012; Kandil & El-Rayes, 2006; Monghasemi, Nikoo, Khak-
sar Fasaee, & Adamowski, 2015; Ng & Zhang, 2008; Sonmez & Bet-
temir, 2012; Xiong & Kuang, 2008; Zhang & Ng, 2012; Zheng et al.,
2005) for performance evaluation. The small-scale test instances
consisted of 18 activities (Feng et al., 1997) including the time–
cost alternatives defined in Hegazy (1999). The first project which
included 5.90 × 109 different time–cost alternatives is examined
under three different conditions. In problem 18a, the indirect costs
are $200/day, the deadline is 110 days, the liquidated damages are
$20,000/day, and the incentives are $1000/day. In problem 18b, the
indirect costs are $1500/day, and in problem 18c, the indirect costs
are $500/day. The optimal solutions for problems 18a and 18b are
$128,270 and $271,270 (Sonmez & Bettemir, 2012). Exact solution
for 18c is determined by applying the mixed integer programming
formulation given in Eqs. (1)–(4) using GUROBI solver version 5.6.3.
The optimal solution for 18c is calculated as $161,270. The aver-
age deviation of DPSO from the global optimal for problems 18a,
18b, and 18c is summarized in Tables 2–4, respectively. In all of
the ten trials for the three problems, DPSO was able to obtain the
global optimal results within 50,000 schedules (objective function
182 S. Aminbakhsh, R. Sonmez / Expert Systems With Applications 51 (2016) 177–185

Table 3
Average deviations from the optimal for problem 18b.

Algorithm No. of runs APD (%)

MAWA-GA (Zheng et al., 2005) 1 0.903

ACS-TCO (Ng & Zhang, 2008) 1 0.018
NA-ACO (Afshar et al., 2009) 1 0.00
ACS-SGPU (Zhang & Ng, 2012) 1 0.698
ACS (Zhang & Ng, 2012) 1 0.018
GA (Sonmez & Bettemir, 2012) 10 1.29
HA (Sonmez & Bettemir, 2012) 10 0.00
DPSO (This Study) 10 0.00

Table 4
Average deviations from the optimal for problem 18c.

Algorithm No. of runs APD (%)

GA (Elbeltagi et al., 2005) 20 2.171

MA (Elbeltagi et al., 2005) 20 0.759
PSO (Elbeltagi et al., 2005) 20 0.415
ACO (Elbeltagi et al., 2005) 20 3.351
SFL (Elbeltagi et al., 2005) 20 2.960
MSFL (Elbeltagi et al., 2007) 20 0.00
ACS-TCO (Ng & Zhang, 2008) 1 0.00
Electimize (Abdel-Raheem & Khalafallah, 2011) 20 0.00
ACS-SGPU (Zhang & Ng, 2012) 1 0.00
ACS (Zhang & Ng, 2012) 1 0.00
DPSO (This Study) 10 0.00

the genetic algorithms (Sonmez & Bettemir, 2012; Zheng et al.,

2005), ant colony system time–cost optimization algorithm (ACS-
TCO) of Ng and Zhang (2008), and ant colony system (ACS) and
ant colony system with global updating strategy (ACS-SGPU) algo-
rithms of Zhang and Ng (2012) for obtaining the optimal solution.
The proposed DPSO also outperformed the genetic, memetic, PSO,
ant colony optimization (ACO), and shuffled frog-leaping optimiza-
tion (SFL) algorithms of Elbeltagi et al. (2005) for the problem 18c.
Fig. 5. Pseudo-code of the proposed discrete PSO algorithm.
The modified shuffled frog-leaping optimization algorithm
(MSFL) (Elbeltagi et al., 2007), the nondominated archiving ACO
(NA-ACO) (Afshar et al., 2009), the Electimize algorithm (Abdel-
Table 1
Raheem & Khalafallah, 2011), and the hybrid genetic algorithm
Parameter settings of the DPSO.
with simulated annealing (HA) (Sonmez & Bettemir, 2012) are
Parameter Description Value among the state-of-the-art methods that are competitive with the
i No. of Birds 250
proposed DPSO for obtaining high quality solutions for the small
pct % of deterministic swarm 0.8 instances up to 18 activities. Though, the elapsed CPU time for
c1 Cognitive parameter 5 MSFL was 8 s compared to 0.4 s of the DPSO. Besides, except for
c2 Social parameter 1 GA and HA (Sonmez & Bettemir, 2012), none of the existing state-
wmax Max. inertia weight 1.2
of-the-art methods were tested for larger instances.
wmin Min. inertia weight 0.0
vmax Max. velocity 3.7

4.2. Medium-scale instances

Table 2
Average deviations from the optimal for problem 18a. The medium-scale instances consisted of two instances, includ-
Algorithm No. of runs APD (%)
ing 63 activities with up to five modes (Sonmez & Bettemir, 2012).
The activity-on-node diagram for the instances with normal mode
GA (Hegazy, 1999) 1 8.139 durations is shown in Fig. 6. The time–cost alternatives for the in-
GA (Sonmez & Bettemir, 2012) 10 2.17
stances are given in Table 5. The medium-scale instances which
HA (Sonmez & Bettemir, 2012) 10 0.00
DPSO (This Study) 10 0.00
included 1.37 × 1042 different time–cost alternatives are examined
with two different indirect expenses. In the first problem (63a) the
indirect costs are $2300/day, and in the second problem (63b) the
indirect costs are $3500/day. The optimal solutions for the prob-
evaluations). The CPU time of DPSO for the small-scale test in- lems are $5,421,120 and $6,176,170, respectively.
stances was 0.4 s. The proposed DPSO algorithm was very successful for obtaining
The comparison of DPSO with the state-of-the-art methods re- high quality solutions for the medium-scale test instances and out-
veal that proposed DPSO is among the top performing algorithms performed the sole genetic algorithm and hybrid genetic algorithm
for the small-scale DTCTP. The DPSO has outperformed the genetic (Sonmez & Bettemir, 2012) as shown in Table 6. DPSO achieved
algorithms (Hegazy, 1999; Sonmez & Bettemir, 2012) for instance APD values of 0.02% and 0.05% for instances 63a and 63b within
18a (Table 2). Similarly, for instance 18b DPSO was better than 50,000 schedules. However, HA was able to achieve APD values of
 S. Aminbakhsh, R. Sonmez / Expert Systems With Applications 51 (2016) 177–185 183

Table 5
Data for the 63 activity TCT problem.

Act. No. Pred. Mode 1 Mode 2 Mode 3 Mode 4 Mode 5

Dur. Cost Dur. Cost Dur. Cost Dur. Cost Dur. Cost
(days) $ (days) $ (days) $ (days) $ (days) $

1 - 14 3750 12 4250 10 5400 9 6250 - -

2 - 21 11,250 18 14,800 17 16,200 15 19,650 - -
3 - 24 22,450 22 24,900 19 27,950 17 31,650 - -
4 - 19 17,800 17 19,400 15 21,600 - - - -
5 - 28 31,180 26 34,200 23 38,250 21 41,400 - -
6 1 44 54,260 42 58,450 38 63,225 35 68,150 - -
7 1 39 47,600 36 50,750 33 54,800 30 59,750 - -
8 2 52 62,140 47 69,700 44 72,600 39 81,750 - -
9 3 63 72,750 59 79,450 55 86,250 51 91,500 49 99,500
10 4 57 66,500 53 70,250 50 75,800 46 80,750 41 86,450
11 5 63 83,100 59 89,450 55 97,800 50 104,250 45 112,400
12 6 68 75,500 62 82,000 58 87,500 53 91,800 49 96,550
13 7 40 34,250 37 38,500 33 43,950 31 48,750 - -
14 8 33 52,750 30 58,450 27 63,400 25 66,250 - -
15 9 47 38,140 40 41,500 35 47,650 32 54,100 - -
16 9, 10 75 94,600 70 101,250 66 112,750 61 124,500 57 132,850
17 10 60 78,450 55 84,500 49 91,250 47 94,640 - -
18 10, 11 81 127,150 73 143,250 66 154,600 61 161,900 - -
19 11 36 82,500 34 94,800 30 101,700 - - - -
20 12 41 48,350 37 53,250 34 59,450 32 66,800 - -
21 13 64 85,250 60 92,600 57 99,800 53 107,500 49 113,750
22 14 58 74,250 53 79,100 50 86,700 47 91,500 42 97,400
23 15 43 66,450 41 69,800 37 75,800 33 81,400 30 88,450
24 16 66 72,500 62 78,500 58 83,700 53 89,350 49 96,400
25 17 54 66,650 50 70,100 47 74,800 43 79,500 40 86,800
26 18 84 93,500 79 102,500 73 111,250 68 119,750 62 128,500
27 20 67 78,500 60 86,450 57 89,100 56 91,500 53 94,750
28 21 66 85,000 63 89,750 60 92,500 58 96,800 54 100,500
29 22 76 92,700 71 98,500 67 104,600 64 109,900 60 115,600
30 23 34 27,500 32 29,800 29 31,750 27 33,800 26 36,200
31 19, 25 96 145,000 89 154,800 83 168,650 77 179,500 72 189,100
32 26 43 43,150 40 48,300 37 51,450 35 54,600 33 61,450
33 26 52 61,250 49 64,350 44 68,750 41 74,500 38 79,500
34 28, 30 74 89,250 71 93,800 66 99,750 62 105,100 57 114,250
35 24, 27, 29 138 183,000 126 201,500 115 238,000 103 283,750 98 297,500
36 24 54 47,500 49 50,750 42 56,800 38 62,750 33 68,250
37 31 34 22,500 32 24,100 29 26,750 27 29,800 24 31,600
38 32 51 61,250 47 65,800 44 71,250 41 76,500 38 80,400
39 33 67 81,150 61 87,600 57 92,100 52 97,450 49 102,800
40 34 41 45,250 39 48,400 36 51,200 33 54,700 31 58,200
41 35 37 17,500 31 21,200 27 26,850 23 32,300 - -
42 36 44 36,400 41 39,750 38 42,800 32 48,300 30 50,250
43 36 75 66,800 69 71,200 63 76,400 59 81,300 54 86,200
44 37 82 102,750 76 109,500 70 127,000 66 136,800 63 146,000
45 39 59 84,750 55 91,400 51 101,300 47 126,500 43 142,750
46 39 66 94,250 63 99,500 59 108,250 55 118,500 50 136,000
47 40 54 73,500 51 78,500 47 83,600 44 88,700 41 93,400
48 42 41 36,750 39 39,800 37 43,800 34 48,500 31 53,950
49 38, 41, 44 173 267,500 159 289,700 147 312,000 138 352,500 121 397,750
50 45 101 47,800 74 61,300 63 76,800 49 91,500 - -
51 46 83 84,600 77 93,650 72 98,500 65 104,600 61 113,200
52 47 31 23,150 28 27,600 26 29,800 24 32,750 21 35,200
53 43, 48 39 31,500 36 34,250 33 37,800 29 41,250 26 44,600
54 49 23 16,500 22 17,800 21 19,750 20 21,200 18 24,300
55 52, 53 29 23,400 27 25,250 26 26,900 24 29,400 22 32,500
56 50, 53 38 41,250 35 44,650 33 47,800 31 51,400 29 55,450
57 51, 54 41 37,800 38 41,250 35 45,600 32 49,750 30 53,400
58 52 24 12,500 22 13,600 20 15,250 18 16,800 16 19,450
59 55 27 34,600 24 37,500 22 41,250 19 46,750 17 50,750
60 56 31 28,500 29 30,500 27 33,250 25 38,000 21 43,800
61 56, 57 29 22,500 27 24,750 25 27,250 22 29,800 20 33,500
62 60 25 38,750 23 41,200 21 44,750 19 49,800 17 51,100
63 61 27 9500 26 9700 25 10,100 24 10,800 22 12,700

2.61% and 2.50% within the same number of schedules. By search-
ing only 50,000 solutions out of 1.37 × 1042 potential solutions,
DPSO was able to determine very high quality solutions that are
either optimal or very close to the optimal. The CPU time of DPSO
for the medium-size instances was 1.3 s.
4.3. Large-scale instances
The large-scale instances consisted of two instances with 630
activities including up to five modes, and representing 2.38 × 10421
possible time–cost alternatives (Sonmez & Bettemir, 2012). The
184 S. Aminbakhsh, R. Sonmez / Expert Systems With Applications 51 (2016) 177–185

Table 6 Table 7
Average deviations from the optimal for problems 63a and 63b. Average deviations from the optimal for problems 630a and 630b.

Algorithm 63a 63b Algorithm 630a 630b

No. of runs APD (%) No. of runs APD (%) No. of runs APD (%) No. of runs APD (%)

GA (Sonmez & Bettemir, 2012) 10 5.86 10 5.16 GA (Sonmez & Bettemir, 2012) 10 8.83 10 7.50
HA (Sonmez & Bettemir, 2012) 10 2.61 10 2.50 HA (Sonmez & Bettemir, 2012) 10 2.41 10 2.47
DPSO (This Study) 10 0.02 10 0.05 DPSO (This Study) 10 0.33 10 0.34

instances represent the size of a real-world construction projects.

The indirect costs for the large-scale instances (630a and 630b) are
the same as the indirect costs of the medium-scale instances (63a
and 63b).
The performance of the DPSO for large-scale instances is sum-
marized in Table 7. DPSO achieved very successful results and out-
performed the hybrid genetic algorithm (HA) (Sonmez & Bettemir,
2012) for large-scale instances. The APD values for instances 630a
and 630b were 0.33% and 0.34%, respectively, and the processing
time of DPSO was 14.6 s for the large-scale instances. By search-
ing only 50,000 solutions out of 2.38 × 10421 potential solutions,
DPSO was able to achieve high quality solutions for the large-
scale instances. HA was able to achieve APD values of 2.41% and
2.47% within 50,000 schedules. The tests of HA with 1,000,000
schedules revealed that HA was able to achieve an APD value of
2% in 73 min (Bettemir, 2009). DPSO was able to achieve sig-
nificantly better solutions than the HA in a much shorter CPU
time.
4.4. New instances
Two large-scale benchmark problems included in the literature
were generated by copying small problems several times and have
certain limitations for representing the complexity of some large-
scale construction projects. 80 new large-scale instances were gen-
erated to evaluate performance of the proposed DPSO. ProGen/max
network generator (Schwindt, 1995) is used to generate the net-
works for the new instances. 10 replications of networks with The-
sen restrictiveness coefficient of 0.2, 0.4, 0.6, and 0.8 are used to
generate 40 networks with 200 activities, and 40 networks with
500 activities. Two to five time–cost alternatives are generated for
each network by means of concave cost functions, using the pro-
cedure described in Akkan et al. (2005). The indirect expenses are
set as $500/day.
The new instances were solved to optimal with the mixed
integer programming model presented in Eqs. (1)–(4) and using
GUROBI solver version 5.6.3. Optimal solutions are achieved for all
the 40 instances with 200 activities, and for 36 of the instances
with 500 activities. Since the problem is NP-hard, a time limit of
five hours is used for solving each instance. The results of the pro-
posed DPSO at the end of the 50,000 schedules are compared with
the optimal solutions for the 76 instances that are solved to opti-
mal by the mixed integer programming method. DPSO achieved an
average 0.19% deviation from the optimal within an average CPU
time of 4.6 s for 40 instances with 200 activities and succeeded an
average 0.21% deviation from the optimal within an average CPU
time of 17.2 s for 36 instances with 500 activities. The instances
and the solutions are available at http://www2.ce.metu.edu.tr/∼
sonmez/research/dpso.
The majority of the existing methods for the cost optimization
problem was tested with few instances, including up to 18 activi-
ties and cannot handle real-life-size projects. The hybrid genetic al-
gorithm (Sonmez & Bettemir, 2012) which was proposed for large-
scale cost optimization problem requires 73 min to achieve a 2%
deviation from the optimal solution. To our best knowledge, pro-
posed DPSO is the first method that can achieve high quality solu-
tions for the large-scale DTCTP within seconds. Computational re-
sults reveal the effectiveness of the DPSO.

6 12 20 27
44 68 41 67

1 34 40 47 52 58
14 7 13 21 28 74 41 54 31 24

39 40 64 66

2 8 14 22
55 59
21 52 33 58
29 35 41 29 27
Start
76 138 37
Finish
42 48
3 9 15 23 30
44 41
24 63 47 43 34
36
53
54 62
43 39
16 24 25
75 60
75 66
31

4 10 17 25 31 37 44 49 54
19 57 60 54 96 34 82 173 23

63
18 26 32 38 56
27
81 84 43 51 38
5 11
45 50 61
28 63
59 101 29
19 33 39 46 51
36 52 67 66 83 57 Activity Critical Path
41 Duration

Fig. 6. Activity on node (AoN) representation of the 63 activity network.

 S. Aminbakhsh, R. Sonmez / Expert Systems With Applications 51 (2016) 177–185
5. Conclusions
In this paper, we present a new discrete particle swarm op-
timization method for the discrete time–cost trade-off problem.
The results of the computational experiments reveal that DPSO can
achieve high quality solutions for small, medium, and large-scale
DTCTP and outperforms the state-of-the-art methods with respect
to solution quality and computation time especially for medium
and large-scale problems. For instances including 630 activities,
DPSO achieved high quality solutions within seconds for the first
time. The main contribution of the proposed DPSO is that it en-
ables time–cost optimization of real-life-size projects.
New benchmark DTCTP instances, including up to 500 activities
are generated to evaluate the performance of the proposed DPSO
method. The majority of the instances are solved to optimal for
the cost optimization problem using mixed integer programming
method. The new instances with optimal solutions enables perfor-
mance evaluation of heuristic and meta-heuristic methods for the
large scale discrete time–cost trade-off cost optimization problem.
The results indicate that DPSO provides an effective and robust
alternative for solving real-world time–cost optimization prob-
lems. However, the large-scale benchmark problems used in this
study included up to 630 activities and up to five time–cost al-
ternatives and may have certain limitations for representing the
complexity of some large-scale construction projects. Research
focusing on generation of large-scale complex DTCTP instances that
includes more activities and modes, would enable a better under-
standing of the performance of heuristic and meta-heuristic meth-
ods for real world projects. The proposed DPSO focused on the cost
optimization problem. Development of efficient discrete optimiza-
tion methods for Pareto-front optimization of large-scale DTCTP in-
stances appears to be another promising area for future research.
Acknowledgments
This research is funded by The Scientific and Technological Re-
search Council of Turkey (TÜBİTAK), under grant #213M253. We
would also like to acknowledge the other research team member
Mert Bilir.
References
Abdel-Raheem, M., & Khalafallah, A. (2011). Using electimize to solve the time-
cost-tradeoff problem in construction engineering. Computing in civil engineering
(2011) (pp. 250–257). American Society of Civil Engineers.
Afshar, A., Ziaraty, A., Kaveh, A., & Sharifi, F. (2009). Nondominated archiving multi-
colony ant algorithm in time–cost trade-off optimization. Journal of Construction
Engineering and Management, 135(7), 668–674.
Akkan, C., Drexl, A., & Kimms, A. (2005). Network decomposition-based benchmark
results for the discrete time–cost tradeoff problem. European Journal of Opera-
tional Research, 165(2), 339–358.
Aminbakhsh, S. (2013). Hybrid particle swarm optimization algorithm for obtaining
Pareto front of discrete time–cost trade-off problem [M.Sc. Thesis]. Ankara, Turkey:
Middle East Technical University.
Ashuri, B., & Tavakolan, M. (2015). Shuffled frog-leaping model for solving time-
cost-resource optimization problems in construction project planning. Journal of
Computing in Civil Engineering, 29(1).
Bettemir, Ö. H. (2009). Optimization of time–cost–resource trade-off problems in
project scheduling using meta-heuristic algorithms [Doctoral dissertation, Ph.D.
Thesis]. Ankara, Turkey: Middle East Technical University.
De, P., Dunne, E. J., Ghosh, J. B., & Wells, C. E. (1997). Complexity of the discrete
time-cost tradeoff problem for project networks. Operations Research, 45(2),
302–306.
De, P., Dunne, James, E, Ghosh, B, J., & E, C. (1995). The discrete time-cost tradeoff
problem revisited. European Journal of Operational Research, 81(2), 225–238.
Demeulemeester, E., Reyck, B. De, Foubert, B., Herroelen, W., & Vanhoucke, M.
(1998). New computational results on the discrete time/cost trade-off problem
in project networks. The Journal of the Operational Research Society, 49(11), 1153–
1163.
Eberhart, R., & Kennedy, J. (1995). A new optimizer using particle swarm theory.
In Proceedings of the sixth international symposium on micro machine and human
science, 1995. MHS.
Elbeltagi, E., Hegazy, T., & Grierson, D. (2005). Comparison among five evolutionary-
based optimization algorithms. Advanced Engineering Informatics, 19(1), 43–53.
185
Elbeltagi, E., Hegazy, T., & Grierson, D. (2007). A modified shuffled frog-leaping op-
timization algorithm: Applications to project management. Structure and Infras-
tructure Engineering, 3(1), 53–60.
El-Rayes, K., & Kandil, A. (2005). Time-cost-quality trade-off analysis for highway
construction. Journal of Construction Engineering and Management, 131(4), 477–
486.
Eshtehardian, E., Afshar, A., & Abbasnia, R. (2008). Time–cost optimization: Using
GA and fuzzy sets theory for uncertainties in cost. Construction Management and
Economics, 26(7), 679–691.
Fallah-Mehdipour, E., Bozorg Haddad, O., Rezapour Tabari, M. M., & Mariño, M. A.
(2012). Extraction of decision alternatives in construction management projects:
Application and adaptation of NSGA-II and MOPSO. Expert Systems with Applica-
tions, 39(3), 2794–2803.
Feng, C., Liu, L., & Burns, S. (1997). Using genetic algorithms to solve construction
time-cost trade-off problems. Journal of Computing in Civil Engineering, 11(3),
184–189.
Goyal, S. K. (1975). Note—A note on “a simple cpm time-cost tradeoff algorithm”.
Management Science, 21(6), 718–722.
Hazır, Ö., Haouari, M., & Erel, E. (2010). Discrete time/cost trade-off problem: A
decomposition-based solution algorithm for the budget version. Computers &
Operations Research, 37(4), 649–655.
Hegazy, T. (1999). Optimization of construction time-cost trade-off analysis using
genetic algorithms. Canadian Journal of Civil Engineering, 26(6), 685–697.
Kandil, A., & El-Rayes, K. (2006). Parallel genetic algorithms for optimizing resource
utilization in large-scale construction projects. Journal of Construction Engineer-
ing and Management, 132(5), 491–498.
Kennedy, J., & Eberhart, R. C. (1995). Particle swarm optimization. In Proceedings of
IEEE international conference on neural network: 4 (pp. 1942–1948).
Kennedy, J., & Eberhart, R. C. (1997). A discrete binary version of the particle swarm
algorithm. In Proceedings of the IEEE international conference on systems, man,
and cybernetics: 5 (pp. 4104–4108).
Kolisch, R., & Hartmann, S. (2006). Experimental investigation of heuristics for
resource-constrained project scheduling: An update. European Journal of Oper-
ational Research, 174(1), 23–37.
Liberatore, M., Pollack-Johnson, B., & Smith, C. (2001). Project management in con-
struction: Software use and research directions. Journal of Construction Engineer-
ing and Management, 127(2), 101–107.
Monghasemi, S., Nikoo, M. R., Khaksar Fasaee, M. A., & Adamowski, J. (2015). A novel
multi criteria decision making model for optimizing time–cost–quality trade-off
problems in construction projects. Expert Systems with Applications, 42(6), 3089–
3104.
Ng, S., & Zhang, Y. (2008). Optimizing construction time and cost using ant colony
optimization approach. Journal of Construction Engineering and Management,
134(9), 721–728.
Schwindt, Ch. (1995). ProGen/max: A new problem generator for different resource-
constrained project scheduling problems with minimal and maximal time lags. Uni-
versity of Karlsruhe Technical Report WIOR 449.
Siemens, N. (1971). A simple CPM time-cost tradeoff algorithm. Management Science,
17(6), B-354–B-363.
Sonmez, R., & Bettemir, Ö. H. (2012). A hybrid genetic algorithm for the discrete
time–cost trade-off problem. Expert Systems with Applications, 39(13), 11428–
11434.
Szmerekovsky, J. G., & Venkateshan, P. (2012). An integer programming formulation
for the project scheduling problem with irregular time–cost tradeoffs. Comput-
ers & Operations Research, 39(7), 1402–1410.
Tavana, M., Abtahi, A.-R., & Khalili-Damghani, K. (2014). A new multi-objective
multi-mode model for solving preemptive time–cost–quality trade-off project
scheduling problems. Expert Systems with Applications, 41(4), 1830–1846.
Vanhoucke, M. (2005). New computational results for the discrete time/cost trade-
off problem with time-switch constraints. European Journal of Operational Re-
search, 165(2), 359–374.
Vanhoucke, M., & Debels, D. (2007). The discrete time/cost trade-off problem: Ex-
tensions and heuristic procedures. Journal of Scheduling, 10(4-5), 311–326.
Xiong, Y., & Kuang, Y. (2008). Applying an ant colony optimization algorithm-based
multiobjective approach for time–cost trade-off. Journal of Construction Engineer-
ing and Management, 134(2), 153–156.
Yang, I. (2007). Using elitist particle swarm optimization to facilitate bicriterion
time-cost trade-off analysis. Journal of Construction Engineering and Management,
133(7), 498–505.
Zhang, L., Zou, X., & Qi, J. (2015). A trade-off between time and cost in scheduling
repetitive construction projects. Journal of Industrial and Management Optimiza-
tion, 11(4), 1423–1434.
Zhang, H., & Xing, F. (2010). Fuzzy-multi-objective particle swarm optimization for
time–cost–quality tradeoff in construction. Automation in Construction, 19(8),
1067–1075.
Zhang, Y., & Thomas Ng, S. (2012). An ant colony system based decision support
system for construction time-cost optimization. Journal of Civil Engineering and
Management, 18(4), 580–589.
Zheng, D., Ng, S., & Kumaraswamy, M. (2005). Applying Pareto ranking and niche
formation to genetic algorithm-based multiobjective time–cost optimization.
Journal of Construction Engineering and Management, 131(1), 81–91.
Zheng, H. (2015). Multi-mode discrete time-cost-environment trade-off problem of
construction systems for large-scale hydroelectric projects. In Proceedings of the
ninth international conference on management science and engineering manage-
ment (pp. 337–346).