You are on page 1of 14

Activity Uncrashing Heuristic with Noncritical

Activity Rescheduling Method for the


Discrete Time-Cost Trade-Off Problem
Rifat Sonmez 1; Saman Aminbakhsh 2; and Tankut Atan 3
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

Abstract: Despite intensive research efforts that have been devoted to discrete time-cost optimization of construction projects, the current
methods have very limited capabilities for solving the problem for real-life–sized projects. This study presents a new activity uncrashing
heuristic with noncritical activity rescheduling method to narrow the gap between the research and practice for time-cost optimization. The
uncrashing heuristic searches for new solutions by uncrashing the critical activities with the highest cost-slope. This novel feature of
the proposed heuristic enables identification and elimination of the dominated solutions during the search procedure. Hence, the heuristic
can determine new high-quality solutions based on the nondominated solutions. Furthermore, the proposed noncritical activity rescheduling
method of the heuristic decreases the amount of scheduling calculations, and high-quality solutions are achieved within a short CPU time.
Results of the computational experiments reveal that the new heuristic outperforms state-of-the-art methods significantly for large-scale
single-objective cost minimization and Pareto front optimization problems. Hence, the primary contribution of the paper is a new heuristic
method that can successfully achieve high-quality solutions for large-scale discrete time-cost optimization problems. DOI: 10.1061/(ASCE)
CO.1943-7862.0001870. © 2020 American Society of Civil Engineers.
Author keywords: Scheduling; Optimization; Algorithms; Multiple objective analysis; Project management.

Introduction relations have been the main subject of recent research due to its
practical relevance (Vanhoucke 2005). In construction management
Critical path method (CPM) is commonly used for scheduling of research, discrete time-cost tradeoff problem (DTCTP) has been the
construction projects. In CPM, the durations of activities are deter- main focus, as the majority of the techniques used to crash durations
mined according to the manpower and machinery resources allo- (such as increasing crew size, increasing capacity of the construction
cated, and construction methods selected. The normal duration machines, using overtime, applying different construction methods,
of an activity corresponds to the resource level and construction etc.) lead to discrete time-cost alternatives.
method for which the duration cannot be shortened without increas- Two types of DTCTP have mainly been studied in the construc-
ing the direct cost of the activity. In many construction projects, the tion management literature. The first problem aims to minimize the
contractors may need to shorten (crash) durations of some activities project cost considering the direct and indirect costs, penalties and
below their normal levels for several reasons such as avoiding the bonuses for a given project deadline. The second problem involves
delay penalty, decreasing the indirect expenses, or to mobilizing determination of the Pareto front, which is the complete and non-
the resources to another project. Expediting activities below their dominated time-cost profile over the set of feasible project dura-
normal duration comes at a cost, and this trade-off between time tions (Vanhoucke and Debels 2007). The Pareto front provides
and cost is known as the time-cost trade-off problem. The objective decision makers multiple nondominated project time-cost alterna-
of the general time-cost trade-off problem is to identify the set of tives that enable consideration of different strategies at a broader
activity duration-cost alternatives (execution modes) that will min- perspective for portfolio management. The nondominated solutions
imize the total cost of the project. can also be used to form a basis for negotiation and bargaining
Time-cost trade-off problem has been the subject of research (Yang et al. 2013).
under various assumptions since the development of CPM. While Numerous heuristic and metaheuristic methods have been pro-
early research assumed linear and continuous relations between posed for the single-objective cost minimization DTCTP (Hegazy
time and cost (Siemens 1971; Goyal 1975), the discrete time-cost 1999; Elbeltagi et al. 2005, 2007; Vanhoucke and Debels 2007;
Sonmez and Bettemir 2012; Aminbakhsh and Sonmez 2016;
1
Professor, Dept. of Civil Engineering, Middle East Technical Univ., Agdas et al. 2018), and also the Pareto front problem (Feng
Cankaya, Ankara 06531, Turkey (corresponding author). Email: rsonmez@ et al. 1997; Zheng et al. 2005; Kandil and El-Rayes 2006; Yang
metu.edu.tr 2007; Ng and Zhang 2008; Afshar et al. 2009; Zhang and Li
2
Assistant Professor, Dept. of Civil Engineering, Atılım Univ., Incek, 2010; Aminbakhsh and Sonmez 2017). The main limitation of
Ankara 06830, Turkey. ORCID: https://orcid.org/0000-0002-4389-1910 the majority of the current methods is that they were tested with
3
Associate Professor, Dept. of Industrial Engineering, Bahçeşehir Univ., problem instances including only up to 18 activities; hence they
Beşiktas, Istanbul 34353, Turkey. ORCID: https://orcid.org/0000-0002
have very limited capabilities for time cost optimization of con-
-3241-4617
Note. This manuscript was submitted on July 19, 2019; approved on
struction projects, which typically include more than 300 activities
February 3, 2020; published online on May 18, 2020. Discussion period (Liberatore et al. 2001). Few recent methods for the single-objective
open until October 18, 2020; separate discussions must be submitted for DTCTP (Sonmez and Bettemir 2012; Aminbakhsh and Sonmez
individual papers. This paper is part of the Journal of Construction En- 2016; Agdas et al. 2018) and the Pareto front problem (Kandil
gineering and Management, © ASCE, ISSN 0733-9364. and El-Rayes 2006; Aminbakhsh and Sonmez 2017) were validated

© ASCE 04020084-1 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


with larger size test instances. However, these methods also have problems in Sonmez and Bettemir (2012). The method of Agdas
several limitations. One of the main limitations of the majority of et al. (2018) required 16.8 h with eight processors to achieve a
current large-scale DTCTP methods is that they require large com- 7% deviation from the optimal for the 6,300 activity problem.
putation time to achieve high-quality solutions. These methods also Metaheuristic attempts to solve the Pareto front problem, in-
have a limitation with respect to the instances used in the computa- clude genetic algorithms (Feng et al. 1997; Zheng et al. 2005), par-
tional analysis as they have used few problems that were created by ticle swarm optimization methods (Yang 2007; Zhang and Li
copying small projects serially, which do not reflect the complexity 2010), ant colony optimization algorithms (Ng and Zhang 2008;
of large-scale construction projects adequately. Afshar et al. 2009), and an integrated multiobjective particle swarm
This paper aims to contribute to both researchers and professio- optimization and support vector regression method (Yang et al.
nals by narrowing the gap between the literature and real-life project 2013). In all of these metaheuristic studies the largest problem used
management with multiple objectives. The primary objective of the to evaluate the performance of the methods only contained 18 ac-
paper is to achieve an advancement in time-cost optimization of proj- tivities. Kandil and El-Rayes (2006) presented a parallel multiob-
ects by presenting a novel heuristic that can obtain high-quality so- jective genetic algorithm framework to enable optimal planning of
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

lutions within a short CPU time for large-scale single-objective cost large-scale construction projects. In computational experiments
minimization and Pareto front DTCTP. The paper also aims to enable Kandil and El-Rayes (2006) used three Pareto front instances with
improved performance evaluation of state-of-the-art methods for 180, 360, and 720 activities that were created by copying the 18
real-life–sized projects by introducing realistic discrete time-cost activity problem serially. The global parallel genetic algorithm
trade-off problem instances. of Kandil and El-Rayes (2006) required 15.4 h over a supercom-
puting cluster of 50 processors to obtain the Pareto front solutions
for the 720 activity problem. The same 180, 360, and 720 activity
Literature Review instances were also used in the performance evaluation of the recent
Pareto front optimization method of Aminbakhsh and Sonmez
A variety of methods including exact methods, constraint program- (2017) based on Siemens (1971) heuristic and particle swarm
ming approaches, heuristics, and metaheuristics have been pro- optimization.
posed for the different versions of the discrete time-cost trade-off The existing research on the DTCTP mainly developed meta-
optimization problem. De et al. (1995) formulated a mixed-integer heuristic methods that usually rely on random search algorithms.
programming (MIP) model that minimized the total direct costs for There are scarce studies that focus on heuristic search procedures.
a given deadline. Vanhoucke (2005) proposed a branch and bound Among the few research that report performance of the methods for
algorithm for the deadline problem with time-switch constraints. large-scale projects, the majority require long CPU times, and may
Klanšek (2015) proposed a general mixed-integer nonlinear model not achieve high-quality solutions for complex problems even with
for the discrete time-cost trade-off problems, including generalized long CPU times. The instances that are used to evaluate the perfor-
precedence constraints, project deadlines, and incremental and mances of the methods is another limitation of these studies,
total budget limitations. Al Haj and El-Sayegh (2015) formulated because they were created by copying small problems that do
a nonlinear-integer programming model that considers the impact not fully represent the complexity of the large-scale projects. The
of total float loss for time-cost optimization. Abuwarda and Hegazy primary objective of this paper is to fill these gaps in the single-
(2019) proposed a comprehensive schedule optimization frame- objective and Pareto front literature by proposing a novel heuristic
work based on constraint programming that enabled simultaneous method and new set of large-scale DTCTP instances for contribut-
use of variety of constraints related to deadline, milestones, and ing optimal planning and scheduling of the real-life–sized construc-
multiple resource limits for schedule compression. tion projects.
Numerous studies have focused on heuristic and metaheuristic
procedures for achieving high-quality solutions within a short
amount of CPU time for the DTCTP. In an early study Hegazy Problem Formulation
(1999) developed a genetic algorithm (GA) for the single-objective
problem and used an instance including 18 activities for perfor- The objective of general DTCTP is to determine the optimal set of
mance evaluation. The same problem was used to compare the time-cost alternatives (modes) that will minimize the sum of direct
results of five evolutionary-based algorithms in a later research and indirect costs. The general single-objective cost minimization
(Elbeltagi et al. 2007). Vanhoucke and Debels (2007) discussed DTCTP can be formulated by modifying the formulation of De et al.
a new metaheuristic approach for the DTCTP and demonstrated (1995) to include the indirect costs and delay penalty as follows:
the results of the algorithm with instances including 50 activities.
Sonmez and Bettemir (2012) presented a hybrid metaheuristic strat- X
S X
mðjÞ

egy for the single-objective cost minimization DTCTP and evalu- minimize ðdcjk xjk Þ þ D × ic þ Delay × dp ð1Þ
j¼1 k¼1
ated the performance of the new algorithm for different instances,
the largest of which included 630 activities and was created by
copying a base problem with 63 activities serially. Aminbakhsh X
mðjÞ
and Sonmez (2016) presented a discrete optimization method based xjk ¼ 1; ∀ j ¼ f1; : : : ; Sg ð2Þ
on Siemens (1971) heuristic and particle swarm optimization k¼1
(PSO), and used instances including up to 630 activities for perfor-
mance evaluation. Siemens (1971) and Goyal (1975) were the early
X
mðjÞ
heuristic methods, which assumed linear and continuous relations djk xjk þ Stj ≤ Stl ; ∀ l ∈ Scj and ∀ j ¼ f1; : : : ; Sg
between time and cost and were applied on an example including k¼1
eight activities. In a recent study, Agdas et al. (2018) used graph ð3Þ
theory, parallel computing, and genetic algorithms to solve the cost
minimization DTCTP for large-scale instances up to 6,300 activ-
ities, which were also created by serially copying the 63 activity D ≥ StSþ1 ð4Þ

© ASCE 04020084-2 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


Noncritical Activity Rescheduling Method
Delay ≥ D − Deadline ð5Þ
The proposed heuristic creates new solutions by uncrashing the
critical activities. Meanwhile, to improve the solutions obtained by
St0 ¼ 0 ð6Þ uncrashing critical activities, the heuristic may also uncrash the
noncritical activities. Traditional CPM calculations would require
xjk ∈ f0; 1g; ∀ j ¼ f1; : : : ; Sg and ∀ k ¼ f1; : : : mðjÞg a full forward and backward pass calculations for all of the activ-
ities in the project network to reschedule their early start times
ð7Þ (ES), early finish times (EF), late start times (LS), late finish times
(LF), and floats, even if the duration of a single noncritical activity
is changed. A new scheduling method, hereafter named noncritical
Stj ≥ 0; ∀ j ¼ f1; : : : ; Sg ð8Þ activity rescheduling method (NARM) is presented to decrease the
amount of scheduling calculations to achieve high-quality solutions
D; Delay ≥ 0 ð9Þ for time-cost optimization of large-scale projects within a short
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

CPU time.
where S = number of activities not including the dummy activities; NARM determines the noncritical activities that are impacted by
0 and S þ 1 represent the start and end of the project; dcjk = direct a change in the duration of a noncritical activity, and only resched-
ules those activities that are impacted by the change. The proposed
cost of mode k for activity j; xjk = mode selection variable; D =
method only updates the required start and finish times by perform-
project duration; ic = daily indirect cost; Delay = project’s delay
ing the noncritical partial forward pass and the noncritical partial
amount; dp = daily delay penalty; djk = duration of mode k for
backward pass scheduling procedures. When the duration of a non-
activity j; mðjÞ = number of modes for activity j; Deadline = critical activity is changed, NARM first determines the updated EF
project deadline; Stj = start time for activity j; Scj = set of imme- of the activity. If the updated EF of the activity is not later than the
diate successors for j; and St0 and Stsþ1 = start times of start and LF of the activity NARM is suitable for rescheduling. Along with
finish dummy activities. The minimum cost Ca solution for a par- the EF, the method determines the updated LS and calculates the
ticular project deadline Da can be obtained by setting a constraint to updated total float (TF). Because the duration change does not
the project duration D as follows: impact the ES and the LF of the activity only EF, LS, and TF
are updated for this activity. The updated noncritical activity is in-
D ¼ Da ð10Þ cluded in the critical path if the updated TF is equal to zero.
In NARM, the noncritical successors of the updated activity
The set of solutions forming the optimal Pareto front can be with an ES earlier than the updated EF of the rescheduled activity
obtained by determining the minimum cost solutions for all feasible are added to the activities to be rescheduled (ATRS) list. In the par-
project durations with the Eqs. (1–10) and then removing the do- tial forward pass scheduling procedure of the NARM, the ES, EF,
minated solutions (Aminbakhsh and Sonmez 2017). and the TFs of all succeeding activities that are included in the
ATRS list are updated. The ES of the activities in ATRS list are
rescheduled so that these activities start immediately after the EF
Activity Uncrashing Heuristic with Noncritical times of their predecessors and the EF times and TFs of the all
Activity Rescheduling Method activities in the ATRS list are updated. The activities with an up-
dated zero float are included in the critical path.
The heuristic methods used in the existing time-cost trade-off stud- The ATRS list is updated so that it will include the successors of
ies (Siemens 1971; Goyal 1975; Aminbakhsh and Sonmez 2016; the previous ATRS list with an ES earlier than the updated EF of
Aminbakhsh and Sonmez 2017) are based on a procedure that starts the preceding activity; afterward, the updated activities are removed
with the longest project duration solution and determines new so- from the list. The ES, EF times, and TFs of all of the required
lutions by crashing the critical activities with the least cost-slope. activities are updated by partial forward pass procedure, and the
These heuristics usually require a large amount of computational procedure is repeated until all the succeeding activities that are im-
time for large-scale networks especially when there are several par- pacted by the duration change are updated as shown in Fig. 1. In the
allel critical paths. Due to their crashing schemes, existing crashing partial backward pass scheduling procedure of this method, the LF,
heuristics can determine the dominated solutions only after per- LS, and the TFs of all preceding activities that are impacted by the
forming and completing the whole crashing procedure. However, duration change are updated. The predecessors with an LF later
the proposed heuristic procedure in this work starts with the short- than the updated LS of the rescheduled activity are impacted by
est project duration solution (the all-crashed schedule) and searches the duration change, hence these activities are rescheduled. NARM
for new solutions by uncrashing the critical activities with the high- determines the activities that are impacted by the duration change
est cost-slope. This novel feature of the proposed heuristic presents and performs partial backward pass scheduling until all the im-
two major advantages. By uncrashing only one activity on the criti- pacted preceding activities are updated similar to the forward pass
cal path, the proposed heuristic can achieve a new solution with a as shown in Fig. 1. In the final stage of the partial backward pass,
few calculations in a very short time compared to the heuristic or the activities with an updated zero float are included in the criti-
metaheuristic methods that achieve a new solution by crashing cal path.
the critical path. The proposed uncrashing heuristic also enables
determination and elimination of the dominated solutions during
Case Example for Noncritical Activity Rescheduling
the search procedure; hence, new solutions can be built based on
Method
the nondominated or high-quality solutions. These features of the
new heuristic enable achieving high-quality nondominated Pareto A CPM network given in Fig. 2 will be used as a case example to
front solutions in a short amount of CPU time. Furthermore, a novel describe the NARM. The ES, EF, LS, LF, dates, and the TFs of
scheduling method is also developed in the new uncrashing heu- activities in Fig. 2 are calculated using traditional forward and
ristic for rescheduling the noncritical activities effectively. backward pass calculations. In the case example the duration of

© ASCE 04020084-3 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


Update the LS and TF of the non critical
activity for which the duration has
changed

Include the activity in Is the updated TF of the


Yes
the critical path activity equal to zero?

No
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

Is there a successor with


ES earlier than updated EF
of the activity?

Yes

Add the successors with an ES earlier


than the updated EF of the activity to the
activities to be re-scheduled (ATRS) list

Update ES, EF and TF of all activities in


the ATRS list

Is the updated TF of the


Include the activities
Yes any activity in the ATRS list
in the critical path
equal to zero?
No

No
For the activities
in the ATRS list is there a
Update the ATRS list For the activities predessor with LS later than
so that it will only include in the ATRS list updated LF of the
the successors with an ES Yes is there a successor with ES activity?
earlier than the updated earlier than updated EF of
EF of the activity the activity?

Yes Update LS, LF and TF of all


No activities in the ATRS list

For the activity


Update the ATRS list
which the duration has
so that it will only include
changed is there a
Yes the predecessors with an
predecessor with LS later
LS later than the updated
than updated LF of the
LF of the activity
activity?
No

No

Report the results

Fig. 1. Flow chart of noncritical activity rescheduling method.

© ASCE 04020084-4 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


5 18 23 23 22 45
2 3
5 0 23 23 0 45

0 5 5 5 3 8 8 10 18 18 11 29 29 6 35 35 2 37 45 5 50
1 4 5 6 7 8 12
0 0 5 13 8 16 16 8 26 26 8 37 37 8 43 43 8 45 45 0 50

5 7 12 12 8 20 20 16 36
9 10 11
14 9 21 21 9 29 29 9 45
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

ES EF
Critical Path Act. No
LS TF LF

Fig. 2. Case example for noncritical activity rescheduling method.

5 18 23 23 22 45
2 3
5 0 23 23 0 45

0 5 5 5 3 8 8 10 18 18 11 29 29 6 35 35 2 37 45 5 50
1 4 5 6 7 8 12
0 0 5 13 8 16 16 8 26 26 8 37 37 8 43 43 8 45 45 0 50
a b
5 7 12 12 11 23 23 16 39
9 10 11
14 9 21 18 6 29 29 6 45

ES EF
Critical Path Act. No
LS TF LF

Fig. 3. Partial forward pass procedure of noncritical activity rescheduling method.

Activity-10 is increased to 11 days from its original duration of In the partial backward pass, Activity-9 is added to the ATRS
8 days. NARM first determines the updated EF of the activity list, because it has an LF date of 21 that is later than the updated LS
as 23 by adding the revised duration to its ES time of 12, in the date of Activity-10 (18). The LF of Activity-9 is updated as 18, so
partial forward pass procedure as shown by “a” in Fig. 3. Because that it does not finish later than the updated LS time of Activity-10,
the updated EF of the Activity-10 (23) is not later than its LF (29), as shown by “c” in Fig. 4. The LS of the Activity-9 is updated by
NARM can be used to update the schedule. The method then up- subtracting its duration of seven from its updated LF. The updated
dates the LS of the activity as 18 by subtracting the revised duration TF of Activity-9 is calculated as six, by subtracting its EF of 12
from its LF time of 29. The updated TF of Activity-10 is calculated from its updated LF of 18. Because Activity-1, which is the only
as six, by subtracting its updated EF of 23 from its LF of 29. predecessor of Activity-9, has an LF date of five that is not later
Activity-11, which is the successor of Activity-10 is added to than the updated LS date of Activity-9 (11), partial backward pass
the ATRS list, because it has an ES date of 20 that is earlier than is completed.
the updated EF date of Activity-10 (23). The partial forward pass is The case example illustrates that the proposed method performs
continued, by updating the ES of Activity-11 as 23, so that it starts rescheduling of noncritical activities by only performing partial
immediately after the updated EF time of Activity-10, as shown by scheduling of three activities (Activities 10, 11, and 9) when the du-
the “b” in Fig. 3. The EF of Activity-11 is updated by adding its ration of Activity-10 is changed to 11 days; whereas traditional criti-
duration of 16 to its updated ES. The updated TF of Activity-11 is cal path method will require full scheduling of all of the 12 activities.
calculated as six, by subtracting its updated EF of 39 from its LF of Hence, the example illustrates the effectiveness of the NARM for
45. Because Activity-12, which is the only successor of Activity-11, rescheduling of noncritical activities, which is crucial for reducing
has an ES date of 45 that is not earlier than the updated EF date of the CPU time of time-cost or resource optimization methods with
Activity-11 (39), partial forward pass is completed after removal of iterative and repetitive nature requiring rescheduling of noncritical
Activity-11 from the ATRS list. activities, such as the proposed activity uncrashing heuristic.

© ASCE 04020084-5 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


5 18 23 23 22 45
2 3
5 0 23 23 0 45

0 5 5 5 3 8 8 10 18 18 11 29 29 6 35 35 2 37 45 5 50
1 4 5 6 7 8 12
0 0 5 13 8 16 16 8 26 26 8 37 37 8 43 43 8 45 45 0 50
c a b
5 7 12 12 11 23 23 16 39
9 10 11
11 6 18 18 6 29 29 6 45
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

ES EF
Critical Path Act. No
LS TF LF

Fig. 4. Partial backward pass procedure of noncritical activity rescheduling method.

Activity Uncrashing Heuristic alternative is selected (this index orders the time-cost alternatives
according to the decreasing cost-slope).
Unlike earlier heuristic approaches, in lieu of using the all-normal
Because the uncrashing heuristic starts from the all-crashed
schedule (CPM max ) initially, the activity uncrashing heuristic ini-
tiates from the all-crashed schedule (CPM min ), by selecting the schedule, an uncrashing scheme is designed to optimize the solu-
shortest/costliest time-cost alternative for each activity. The duration tions. The uncrashing scheme incorporates two different phases,
of this schedule is determined using the CPM technique and the total namely uncrash noncritical activities that can be uncrashed free
cost is calculated by including the direct and indirect costs and the of cost without increasing the project duration, or uncrash free-float
delay penalty. In the next step of the uncrashing heuristic, duration phase, and UCS phase that uncrashes critical activities based on
differences, cost differences, and cost-slopes of discrete time-cost their cost-slopes. The UFF phase is performed to improve the
modes for all activities are calculated according to Eqs. (11–13) solutions obtained by the UCS phase by decreasing the direct cost
of a given schedule. Through UFF phase the activities with total
DDiffjn ¼ djðnþ1Þ − djðnÞ ; ∀ j ¼ f1; : : : ; Sg; floats (TF) greater than or equal to their effective (immediate avail-
able) crash amounts, DDiffjn ≤ TFj , are uncrashed to their next
∀ n ¼ f1; : : : ; mðjÞ − 1g ð11Þ
available option, by uncrashing one activity at a time. After per-
forming each UFF, the noncritical activities are rescheduled using
CDiffjn ¼ dcjðnþ1Þ − dcjðnÞ ; ∀ j ¼ f1; : : : ; Sg; the NARM. The UFF process is continued until either all of the
∀ n ¼ f1; : : : ; mðjÞ − 1g ð12Þ noncritical activities are fully uncrashed, or their total floats be-
come less than their effective crash amounts. While each solution,
Sol, is recorded in an archive called Solutions, the nondominated
CDiff jn
CSjn ¼ ; ∀ j ¼ f1; : : : ; Sg; solutions are recorded in an external repository called Pareto.
DDiff jn After the first nondominated solution is generated and stored in
∀ n ¼ f1; : : : ; mðjÞ − 1g ð13Þ Pareto with a status of Open, the uncrashing heuristic performs first
uncrashing cycle on this Sol. In an uncrashing cycle, the activities
where DDiff jn = difference between durations of mode n þ 1 that can be uncrashed are uncrashed to their next available option,
and n for activity j; CDiff jn = difference between direct costs by executing UCS process one activity at a time. After performing
of mode n þ 1 and n for activity j; and CSjn = cost-slopes from each UCS, uncrashing heuristic implements the regular CPM
mode n to n þ 1 for activity j. calculations to calculate the duration and TFs. For this new Sol
Time-cost alternatives are then sorted in descending order, first uncrashing heuristic executes UFF process and searches for the
according to CDiff values then according to CS values; an index activities that can be uncrashed without increasing the project du-
named uncrash free-float index (UFF − index) is assigned to each ration. The heuristic then stores the current solution in Solutions.
alternative. Next, time-cost alternatives are sorted in descending During the cycle all nondominated solutions will be recorded in
order first according to CS values then in ascending order accord- Pareto with a status Open. Once the uncrashing cycle is completed
ing to DDiff values, and an index named Uncrash Cost-Slope all of the solutions stored in Solutions are deleted, and the status
ðUCSÞ − index is assigned to each alternative. According to the of Sol ∈ Pareto, which was selected at the start of the cycle is
UFF − index, an alternative providing a greater cost saving is changed to Closed. New schedules are generated through UCS
uncrashed first. In case of a tie, the alternative with a greater and UFF processes based on the nondominated solutions which
cost-slope rate is selected; if the tie is still not broken, the first alter- are already stored in Pareto, as long as there exists at least one
native is selected (this index orders the time-cost alternatives Sol in Pareto with Open status, or, until all activities are uncrashed
according to the decreasing cost difference). Conversely, according to their modes with the longest durations. The heuristic terminates
to UCS − index, an alternative with a greater cost-slope rate is un- when the status of the all of the nondominated Sols stored in Pareto
crashed first. In case of a tie, the alternative providing a smaller are Closed. The procedure of the uncrashing heuristic is summa-
duration reduction is selected; if the tie is still not broken, the first rized in Fig. 5.

© ASCE 04020084-6 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


Start Uncrash Free-Float

Retrieve No Yes
Calculate Deadline
Project Info. j S

Yes
n m
No
No Yes Calculate CS, DDiff, Index by CDiff, then
j S CDiff by CS for UFF Increment j
DDiff Total-Float
No
Yes
n m
Index by CS, then Increment n Yes
Increment j No Increment n
by DDiff for UCS
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

Uncrash based on
Calculate Cost Perform NARM
UFF Index

Fully Crash all Uncrash Free-


Calculate CPM Calculate Dur/Cost
Activities Float
Return

No
Yes Uncrash based on
Store Sol in Pareto Sol in not Closed Calculate CPM
UCS Index

Uncrash Free- Store Sol in


Calculate Dur/Cost Close Sol
Float Solutions

Remove dominated Remove dominated


Increment Sol index
Sol’s from Solutions Sol’s from Pareto

No max Sol index Yes Copy Solutions to Output


reached Pareto Pareto

End

Fig. 5. Flow chart of uncrashing heuristic.

Case Example for Activity Uncrashing Heuristic penalty is $2,000=day. In the first step of the heuristic, the duration
The proposed uncrashing heuristic is explained using a small case of all-crashed schedule is determined by CPM. The initial solution
(Schedule-1) has a project duration of 31 days, and a total cost of
problem which is presented in Fig. 6. The indirect cost for the case
$270,600 as shown in Table 1. In the Schedule-1 which consists of
problem is $1,000=day, the deadline is 45 days, and the delay
all-crashed modes, Activities 1, 3, and 6 are on the critical path.
According to UFF, among the noncritical activities that have a to-
3 tal float (TF) greater than or equal to its effective (immediate avail-
1. (16, $52,000) able) crash amounts, Activity-4 with a CDiff of $18,000 provides
2. (20, $45,000)
1 3. (28, $33,000) the greatest cost saving as shown in Table 2. Thus, Activity-4 is
1. (12, $21,600) uncrashed first, to its second mode (M2) and the noncritical activ-
2. (14, $17,000)
3. (26, $14,000) 4 6 ities are rescheduled using the NARM. The same procedure is ap-
Start 1. (5, $52,000) 1. (3, $36,000) Finish plied to Activity-5, which is uncrashed to its second mode (M2);
2
2. (8, $34,000) 2. (5, $24,000) similarly, Activity-2 is then uncrashed to its second mode (M2).
1. (7, $38,000) In Schedule-1.iv and Schedule-1.v, Activity-2, and Activity-5
2. (10, $26,000) are uncrashed to their third option (M3), respectively. The resulting
3. (12, $20,000) 5 ID
1. (10, $40,000)
Schedule-1.v has a duration of 31 days and a total cost of $216,600
1. (Dur., Cost)
2. (15, $24,000) n. (Dur., Cost) (Table 1) and is recorded in Pareto with a status Open as the first
3. (16, $22,000) cycle terminates.
The effective (immediately available) uncrashing alternatives
Fig. 6. Case example for uncrashing heuristic.
for the only Sol in Pareto with a status of Open (Schedule-1.v)

© ASCE 04020084-7 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


Table 1. Candidate solutions for the case example
Schedule number Cycle number Continue from Activity Uncrash Duration (days) Direct cost ($) Indirect cost ($) Total cost ($)
1 1 — — All Crashed 31 239,600 31,000 270,600
1.i 1 1 4 M1 to M2 31 221,600 31,000 252,600
1.ii 1 1.i 5 M1 to M2 31 205,600 31,000 236,600
1.iii 1 1.ii 2 M1 to M2 31 193,600 31,000 224,600
1.iv 1 1.iii 2 M2 to M3 31 187,600 31,000 218,600
1.v 1 1.iv 5 M2 to M3 31 185,600 31,000 216,600
2 2 1.v 6 M1 to M2 33 173,600 33,000 206,600
— 2 1.v 1 M1 to M2 33 — — —
3 2 1.v 3 M1 to M2 35 178,600 35,000 213,600
4 3 2 1 M1 to M2 35 169,000 35,000 204,000
5 3 2 3 M1 to M2 37 166,600 37,000 203,600
6 4 4 3 M1 to M2 39 162,000 39,000 201,000
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

7 4 4 1 M2 to M3 47 166,000 47,000 217,000


8 5 5 1 M1 to M2 39 162,000 39,000 201,000
9 5 5 3 M2 to M3 45 154,600 45,000 199,600
10 6 6 3 M2 to M3 47 150,000 47,000 201,000
11 6 6 1 M2 to M3 51 159,000 51,000 222,000
12 7 9 1 M1 to M2 47 150,000 47,000 201,000
Note: Boldface distinguishes the nondominated solutions.

Table 2. Cost-slopes, DDiff, CDiff, and index values of uncrashing Single-Objective Activity Uncrashing Heuristic
alternatives
The proposed uncrashing heuristic can be used for solving both
Crash Cost-slope DDiff CDiff
single-objective cost minimization and Pareto front DTCTPs.
Activity mode ($/day) (days) ($) UCS-index UFF-index
The Pareto front uncrashing heuristic described in the previous sec-
1 M1 to M2 2,300 2 4,600 6 8 tion is slightly modified for the cost minimization problem. In the
1 M2 to M3 250 12 3,000 10 9 UCS phase of the cost minimization problem, for a candidate sol-
2 M1 to M2 4,000 3 12,000 3 4 ution within the same cycle if a solution with same duration already
2 M2 to M3 3,000 2 6,000 5 7
exists in the Solutions that solution is discarded to accelerate the
3 M1 to M2 1,750 4 7,000 8 6
3 M2 to M3 1,500 8 12,000 9 5 search process. Cost minimization uncrashing heuristic also uses
4 M1 to M2 6,000 3 18,000 2 1 Result to store only the least-cost Sol among Solutions rather than
5 M1 to M2 3,200 5 16,000 4 2 recording the whole Pareto.
5 M2 to M3 2,000 1 2,000 7 10
6 M1 to M2 6,000 2 12,000 1 3
Computational Experiments

Computational experiments are conducted to evaluate the perfor-


are Activity-1: M1 to M2, Activity-3: M1 to M2, and Activity-6: mance of the new uncrashing heuristics for single-objective cost
M1 to M2. Activity-6 with the largest cost-slope of $6,000=day minimization and Pareto front DTCTPs. The proposed heuristics
is uncrashed to its second mode (M2) in the course of first UCS is coded in C# and compiled within Visual Studio 2013 on a
phase of the second cycle and the result is stored as Schedule-2 in 64-bit platform. All of the experiments are carried out on a personal
Solutions. The next least-cost-slope activity is Activity-1; however, computer with a 16 GB 667 MHz DDR3 random-access memory
uncrashing it to its second mode (M2) results in a solution with (RAM), Intel Core i7-3.40 GHz CPU, and 64-bit Windows 10 op-
a duration of 33 days, which is the duration of the Schedule-2 eration system. In the tests, the new heuristic is executed using a
in Solutions, and none of the activities can be uncrashed free single processor and overclocking was not performed.
of cost without increasing the project duration, hence the heuristic
does not calculate the cost for the current solution and the solu-
tion is discarded. The final uncrashable activity in this cycle is New Problem Instances
Activity-3, which is uncrashed to its second mode (M2) and New large-scale instances are generated to enable improved perfor-
Schedule-3 is obtained. The UFF process is not executed for mance evaluation of state-of-the-art methods for real-life–sized
Schedule-2 and Schedule-3, as none of the noncritical activities problems. RanGen2 random network generator (Vanhoucke et al.
in these schedules can be uncrashed free of cost without increasing 2008) is used to generate the networks for the problems. Five rep-
the project duration. At the end of the second uncrashing cycle the lications of networks with serial/parallel factor (I 2 ) 0.8 are used to
only nondominated Sol in Solutions, i.e., Schedule-2 (Schedule-3 generate 20 networks with 100, 200, 500, and 990 activities. Six
is dominated by Schedule-2), is copied to Pareto and the status of time-cost alternatives are generated by means of nonincreasing con-
Schedule-1.v is changed to Closed, and all the Sol in Solutions are vex time-cost functions using the procedure of Akkan et al. (2005).
deleted. In the next uncrashing cycle the nondominated Sols, The details of instance generation procedure are described in
Schedule-4 and Schedule-5, are obtained by continuing uncrashing Aminbakhsh (2018). The deadline for each instance is set as the
from Schedule-2. The uncrashing procedure is repeated to obtain average of CPM max and CPM min . An indirect cost of $1,000=day
the final Pareto including the nondominated Sols Schedule-1.v, and a delay penalty of $2,000=day are used in all of the problems.
Schedule-2, Schedule-4, Schedule-5, Schedule-6, and Schedule-9 All of the new instances were solved to optimality for the cost
as shown in Table 1. minimization problem with the mixed-integer programming model

© ASCE 04020084-8 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


Table 3. Performance of uncrashing heuristic for cost minimization Table 4. Comparison of the results for the 630-activity problems
problem
630a 630b
Instance Optimal ($) Heuristic ($) PD (%) CPU time (s)
CPU CPU
T100_101 8,027,232 8,059,341 0.40 0.11 Method APD (%) time (s) APD (%) time (s)
T100_102 8,272,300 8,347,578 0.91 0.07
HA (Sonmez and 2.41 4,380 2.47 4,380
T100_103 8,155,670 8,205,420 0.61 0.06
Bettemir 2012)
T100_104 8,709,207 8,810,234 1.16 0.11
DPSO (Aminbakhsh 0.33 15 0.33 15
T100_105 8,194,660 8,246,286 0.63 0.06
and Sonmez 2016)
T200_101 15,729,352 15,861,479 0.84 0.29
GA (Agdas et al. 2018) 0.79 NA 1.12 NA
T200_102 16,995,377 17,180,627 1.09 0.28
Uncrashing heuristic 0.05 5 0.1 5
T200_103 16,311,611 16,492,670 1.11 0.29
(this study)
T200_104 16,670,816 16,822,520 0.91 0.28
T200_105 16,663,623 16,791,933 0.77 0.28
T500_101 40,872,814 41,297,891 1.04 2.31
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

T500_102 42,070,915 42,558,938 1.16 2.37


T500_103 41,819,190 42,362,839 1.30 2.29
T500_104 41,219,686 41,656,615 1.06 1.93 Table 5. Comparison of uncrashing heuristic with GA
T500_105 41,444,839 41,933,888 1.18 2.42 Method GA (Agdas et al. 2018) Uncrashing heuristic
T990_101 82,463,696 83,296,579 1.01 12.33
T990_102 80,748,487 81,499,448 0.93 10.15 Instance APD (%) CPU time (s) PD (%) CPU time (s)
T990_103 80,695,576 81,454,114 0.94 7.94 1,800 7.05 20,952 0.00 41
T990_104 81,906,165 82,667,892 0.93 10.19 1,800 14.72 21,024 0.00 41
T990_105 81,188,735 81,935,671 0.92 10.18 3,150 6.50 32,940 0.04 548
— — Average 0.94 3.19 3,150 4.73 33,876 0.10 529
6,300 7.66 59,112 0.04 4,485
6,300 6.96 60,336 0.10 4,293
Average 7.94 38,040 0.05 1,656
presented in Eqs. (1–9) and using GUROBI solver version 5.6.3
within a time limit of 1 h. The optimal Pareto front was determined
by solving the mixed-integer programming model presented in
Eqs. (1–10) several times for all the durations between CPM min
and CPM max , and discarding the dominated solutions. For the outperforms all the three state-of-the art methods for the two large-
Pareto problem the instances with 100 and 200 activities were scale cost minimization problems by achieving results with a 0.1%
solved to optimality within the 1 h limit with GUROBI solver; how- or less deviation from the optimal within 5 s.
ever, the global optimal solution for any of the problems with 500, Agdas et al. (2018) created six instances including 1,800, 3,150,
and 990 activities was not achieved. The instances and their optimal and 6,300 activities by copying the 18 activity and 63 activity prob-
solutions are available on the websites provided in the Data Avail- lems serially. These six instances are used to provide a comparison
ability Statement. of the proposed heuristic with the recent GA of Agdas et al. (2018)
as shown in Table 5. The results of GA are the average of five trials
executed with eight CPUs of a high performance computing
Single-objective Cost Minimization Problem facility. GA (Agdas et al. 2018) was able to obtain solutions with
Performance Comparison an average percent deviation (APD) of 7.05% and 14.72% for two
The performance of the single-objective uncrashing heuristic is first instances with 1,800 activities within 5.8 h, whereas the uncrashing
tested for the cost minimization problem using the new instances. heuristic determined the optimal solutions for both problems within
The results of the proposed heuristic for the instances are summa- 41 s. For the problems including 3,150 and 6,300 activities GA
rized in Table 3. The proposed heuristic was able to determine was able to achieve solutions with an average percent deviation
solutions within an average percent deviation (PD) of 1% from the of 4.73%–7.66% within a CPU time between 9.2 and 16.8 h. The
optimal in an average CPU time of 3.19 s. The largest deviation uncrashing heuristic achieved solutions with a percent deviation of
from the optimal was 1.30% for a problem with 500 activities, 0.04%–0.10% within a CPU time between 9.1 min and 1.3 h. The
and the longest CPU time was 12.33 s for a problem with 990 ac- computational experiments reveal that the new uncrashing heuristic
tivities. Hence the new heuristic was able to achieve high-quality outperforms state-of-the-art methods significantly for the cost
solutions within few seconds for the large-scale cost minimization minimization problem by achieving high-quality solutions in less
problems including up to 990 activities. CPU time.
The large-scale instances that were created by copying small
problems are used to compare the performance of the uncrashing
Pareto Front Problem—Hypervolume
heuristic with the state-of-the-art methods for the cost minimization
problem. The results of uncrashing heuristics for two large-scale The results obtained from multiobjective optimizations comprise a
instances included in Sonmez and Bettemir (2012) with 630 activ- set of solutions rather than a single optimal solution. Of the multi-
ities are compared with the results of hybrid genetic algorithm (HA) objective optimization problems, Pareto front DTCTP yields a
(Sonmez and Bettemir 2012), discrete particle swarm optimization sequence of solutions for the problems with objectives of time and
(DPSO) method (Aminbakhsh and Sonmez 2016), and genetic cost. Consequently, the definition of quality of solutions is compli-
algorithm (Agdas et al. 2018) in Table 4. The results of the HA, cated within this context and for Pareto front DTCTP. Hence,
DPSO, and GA are average results of several trials due to the sto- numerous performance indices have been proposed for evaluation
chastic nature of the methods, whereas the results of the uncrashing of solution quality of multiobjective optimization methods, which
heuristic are based on a single trial, as the proposed method is not a mainly engage two aspects of the generated solutions (Zitzler et al.
stochastic method. The results reveal that the new heuristic method 2000): (1) the distance of achieved nondominated Pareto solutions

© ASCE 04020084-9 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


Table 6. Example Pareto solution HVPF
HR ¼ ð14Þ
TP PFa HVTP
Duration (days) Cost ($) Duration (days) Cost ($) where HVPF = hypervolume of the Pareto front being evaluated;
44 648,000 44 648,000 and HVTP = hypervolume of the optimal/true Pareto front (TP).
53 621,000 53 642,000 In case the optimal Pareto front is not available, the HVTP is taken
64 620,000 62 641,000 as the hypervolume of the unified Pareto front (UPF), that includes
— — 64 637,000 the nondominated solutions obtained by combining nondominated
— — 69 620,000
solutions of different optimization methods for the same problem.
The hypervolume of a Pareto front for the DTCTP can be
calculated using Eq. (15)
to the optimal Pareto front should be minimized, and (2) a good
X
NS
distribution and spread of the solutions (representing the distribu- HV ¼ ðCost ðRÞ − Cost ðpÞÞ
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

tion of the optimal Pareto) is desired. The amount of CPU time p¼1
required to achieve the solutions is also crucial along with the sol-
ution quality in performance evaluation of different time-cost opti- × ðDuration ðp þ 1Þ − Duration ðpÞÞ ð15Þ
mization methods.
where Cost ðRÞ = cost value of the reference point; Duration
In previous studies (Kandil and El-Rayes 2006; Afshar et al.
ðp þ 1Þ = duration value of the ðp þ 1Þth Sol ∈ Pareto; Cost ðpÞ
2009; Aminbakhsh and Sonmez 2017) the number of nondomi- and Duration ðpÞ = cost and duration values of pth Sol ∈ Pareto;
nated Pareto solutions is commonly used to evaluate the quality of and NS = total number of Sol ∈ Pareto.
solutions. The number of nondominated Pareto solutions provides Hypervolume calculation requires selection of a reference point
an indication for the distribution of the solutions; however, this R. In general the reference point is selected such that it will give a
metric is not sufficient to evaluate the quality solutions if used space that is slightly larger than the actual objective space (Bader
solely, because the metric may give misleading results under certain 2010). Beume et al. (2007) recommended incrementing the ex-
circumstances. For example, in a problem with an optimal Pareto tremal objective values by one, and Knowles (2006) suggested in-
front (true Pareto front—TP) consisting of three nondominated cluding a corner box which is 1% larger than the objective space. In
solutions, a solution PFa including five nondominated solutions this research, reference point was determined based on suggestions
as shown in Table 6, would be considered as a better solution than of Beume et al. (2007) and Knowles (2006) according to the mag-
the optimal Pareto front, if only the number of nondominated nitude of the objective values. The cost value of the reference point
Pareto solutions is used to evaluate the solution quality. was taken as 0.5% larger than the maximum cost value of the ob-
Hypervolume (HV), or hyperarea, indicator was proposed in jective space, and the duration value of the reference point was
recent years to evaluate the quality solutions of multiobjective op- taken as 0.5% larger than the maximum duration value of the ob-
timization methods (Zitzler and Thiele 1998; Veldhuizen 1999; jective space.
Knowles 2006; Beume et al. 2007). Hypervolume indicates the The example of Table 6 will be used to illustrate the hyperarea
partition of the solution space bounded by the nondominated fron- ratio calculations performed to evaluate the performance of the new
tier and a reference point (R) (Zitzler and Thiele 1998). For each uncrashing heuristic for the Pareto front problem. The cost value of
Sol ∈ Pareto, an area is constructed using a reference point R as the reference point was taken as $651,240 which is 0.5% larger
the diagonal corner of the area. than the maximum cost value ($648,000) of the objective space.
A normalized, HV-based approach called hyperarea ratio (HR) Merely in order to provide an improved visual representation of
(Veldhuizen 1999) is adopted in this paper to evaluate the quality of the hyperarea given in Fig. 7, duration value for the reference point
the solutions.Unary HR is calculated according to Eq. (14) was set by increasing the maximum duration by 5%, rather than the

655,000

650,000

645,000
Project Cost ($)

640,000

635,000

630,000

625,000

620,000

615,000
40 45 50 55 60 65 70 75
Project Duration (Days)

Fig. 7. Hyperarea ratio calculation for HVPFA.

© ASCE 04020084-10 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


Table 7. Performance of uncrashing heuristic for Pareto front problem
Number of solutions Hyperarea ratio CPU time (s)
Problem Optimal PFPSO Heuristic PFPSO Heuristic Optimal PFPSO Heuristic
T100_101 149 14 128 0.61 0.99 92.7 0.5 0.5
T100_102 141 16 135 0.62 1.00 75.6 0.5 0.5
T100_103 183 23 153 0.64 0.98 119.7 0.6 0.5
T100_104 247 15 207 0.50 0.99 96.3 0.8 0.8
T100_105 168 10 125 0.55 0.99 83.1 0.5 0.5
T200_101 398 14 320 0.45 0.99 256.0 6.8 6.8
T200_102 424 17 355 0.42 0.98 260.7 8.1 8.0
T200_103 292 8 261 0.35 1.00 223.7 5.8 5.7
T200_104 247 9 196 0.45 0.99 226.2 4.5 4.5
T200_105 300 16 247 0.43 0.99 272.3 5.5 5.4
— —
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

T500_101 21 852 0.43 0.99 147.2 146.3


T500_102 — 42 1,071 0.53 0.99 — 191.4 188.3
T500_103 — 41 1,079 0.43 1.00 — 191.3 189.1
T500_104 — 34 1,021 0.46 1.00 — 184.6 183.1
T500_105 — 32 914 0.43 0.99 — 160.7 159.2
T990_101 — 72 1,983 0.53 1.00 — 1,466.9 1,457.5
T990_102 — 46 1,650 0.47 1.00 — 1,168.5 1,148.0
T990_103 — 38 1,668 0.46 1.00 — 1,211.0 1,193.4
T990_104 — 35 1,817 0.48 1.00 — 1,232.8 1,217.0
T990_105 — 56 2,106 0.49 1.00 — 1,533.7 1,516.6
— Average 30.69 979.06 0.46 0.99 — 469.96 464.34

0.5% used to obtain the results (Table 7) of this study. Hence, in this PFPSO for the new instances are provided in the Data Availability
example the duration value of the reference point was taken as Statement.
72.45, which is 5% larger than the maximum duration value (69) Three DTCTP instances that were created by Kandil and
of the objective space. The hypervolume of the TP (HVTP ) is cal- El-Rayes (2006) including 180, 360 and 720 activities are used to
culated as $625,778 according to Eq. (15) as shown in Fig. 7. The compare the performance of the new Pareto optimization uncrash-
HVPFa is the hypervolume of PFa , and is calculated as $415,778 ing heuristic with the performance of the global parallel and coarse-
similarly. Finally, the hyperarea ratio of HVPFa is calculated as grained genetic algorithms of Kandil and El-Rayes (2006) and
0.664 using Eq. (14). According to the hyperarea ratio indicator PFPSO (Aminbakhsh and Sonmez 2017). The genetic algorithms
the optimal Pareto front TP is a much better solution than the were implemented using 50 processors on the tungsten supercom-
HVPFa , because it has a perfect hyperarea ratio of 1, which is puting cluster, which was composed of 640 Dell PowerEdge 1750
significantly larger than the hyperarea ratio of PFa . servers, each with two Intel Xeon 3.2 GHz processors, 1.5 MB of
cache memory, and a total of 3 GB of SDRAM (Kandil and
El-Rayes 2006). The supercomputing cluster had a peak perfor-
Pareto Front Problem Performance Comparison mance of 6.4 Gflops.
Because only the number of nondominated Pareto solutions and
The performance of the proposed Pareto optimization uncrashing
CPU time were reported in Kandil and El-Rayes (2006) and Kandil
heuristic is compared with two the state-of-the-art optimization
(2005) these two performance indicators are used in comparisons.
methods (Kandil and El-Rayes 2006; Aminbakhsh and Sonmez
Table 8 reveals that for the 180-activity problem the Pareto op-
2017) that are capable of solving the large-scale Pareto front prob-
timization uncrashing heuristic with 586 solutions in 3 s, was able
lem. The results of the proposed uncrashing heuristic is first com- to achieve significantly more nondominated Pareto solutions in a
pared with the results of the Pareto front particle swarm optimizer shorter time, compared to the coarse-grained GA (68 solutions
(PFPSO) (Aminbakhsh and Sonmez 2017) using the new problem in 684 s), the global parallel GA (267 solutions in 2,556 s), and the
instances and considering two performance indicators, namely the PFPSO (304 solutions in 21 s). The uncrashing heuristic (1,176
number of nondominated Pareto solutions and hyperarea ratio. In solutions in 31 s) also outperformed the coarse grained GA
computational experiments, for each problem PFPSO was termi- (94 solutions in 1,836 s), the global parallel GA (232 solutions
nated when the CPU time of the uncrashing heuristic is reached. in 10,404 s), and the PFPSO (536 solutions in 43 s) for the 360
The results that are provided in Table 7 show that the proposed activity problem as shown in Table 9. The coarse grained GA
heuristic is significantly better than the PFPSO both in terms of and the global parallel GA over a cluster of 50 processors required
number of nondominated solutions and the hyperarea ratio. The 7,091 and 55,296 s to obtain 132 and 303 nondominated solutions
new uncrashing heuristic was able achieve an average of 979 so- for the 720 activity problem as summarized in Table 10. The un-
lutions with an average HR of 0.99, whereas PFPSO obtained an crashing heuristic with a single processor achieved 2,356 solutions
average of 31 solutions with an average HR of 0.46. The HR values for the same problem in 292 s, which was also significantly better
of the proposed heuristic for the problems with optimal solutions than the number of solutions obtained by the PFPSO (1,023 solu-
indicate that the new heuristic was able to obtain high-quality tions in 292 s). The computational experiments reveal that the
solutions that are very close to the optimal Pareto front. The results new uncrashing heuristic presents an advancement in both single-
for the first problem T100_101 are shown in Fig. 8 to illustrate the objective and Pareto front optimization of large-sized projects as it
quality of the solutions of the uncrashing heuristic compared to outperforms all of the state-of-the-art methods that were used for
the PFPSO. All of the results of the uncrashing heuristic and solving the large-scale instances for both problems.

© ASCE 04020084-11 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


8,140,000
Optimal
8,120,000 Heuristic
PFPSO

8,100,000
Project Cost ($)
8,080,000

8,060,000

8,040,000
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

8,020,000
350 400 450 500 550 600 650 700
Project Duration (Days)

Fig. 8. Performance comparison of new heuristic and PFPSO for problem T100_101.

Table 8. Performance of uncrashing heuristic for the 180-activity problem Add-in Software
Number of CPU Number of
Algorithm processors time (s) solutions An add-in was developed for Microsoft Project 2019 to enhance the
practical use of the new uncrashing heuristic. The add-in can solve
Global parallel GA (Kandil 50 2,556 267a both single-objective and Pareto front discrete time-cost optimiza-
and El-Rayes 2006)
tion problems for medium and large-scale projects. The add-in in-
Coarse grained GA (Kandil 50 684 68a
and El-Rayes 2006)
tegrates the proposed heuristics to Microsoft Project to enable easy
PFPSO (Aminbakhsh and 1 21 304 use of the uncrashing heuristic for single-objective cost minimiza-
Sonmez 2017) tion and Pareto front optimization. Details regarding availability of
Uncrashing heuristic 1 3 586 the add-in are given in the Data Availability Statement.
(this study)
a
Kandil (2005).
Conclusions
In this paper, an uncrashing heuristic is presented for both single-
Table 9. Performance of uncrashing heuristic for the 360-activity problem objective cost minimization and Pareto front time-cost trade-off
problems. The contributions of the paper are threefold: (1) the
Number of Number of
Algorithm processors CPU time (s) solutions new uncrashing heuristic provides an advancement in time-cost
optimization of large-scale projects as it is outperforms state-
Global parallel GA (Kandil 50 10,404 232 of-the-art methods consistently for the single objective and Pareto
and El-Rayes 2006)
front problems and is capable of achieving successful solutions
Coarse grained GA (Kandil 50 1,836 94
and El-Rayes 2006)
within a reasonable CPU time for both problems; (2) new discrete
PFPSO (Aminbakhsh and 1 43 536 time-cost trade-off problems are created and majority of the prob-
Sonmez 2017) lems were solved to optimality to enable improved performance
Uncrashing heuristic 1 31 1,176 evaluation of state-of-the-art methods for both large-scale cost
(this study) minimization and Pareto front problems; and (3) a novel sched-
uling method is presented for effective rescheduling of the non-
critical activities to achieve solutions within short CPU time for
time-cost optimization of large-scale projects. The new scheduling
method can also be used for decreasing scheduling calculations of
Table 10. Performance of uncrashing heuristic for the 720-activity other heuristic and metaheuristic optimization methods requiring
problem
repetitive rescheduling of noncritical activities, such as resource
Number of CPU Number of Pareto leveling or cash flow optimization heuristics or metaheuristics.
Algorithm processors time (s) front solutions Hence the paper has multiple contributions to the literature and
Global parallel GA (Kandil 50 55,296 303 practice in particular for time-cost optimization of large-scale
and El-Rayes 2006) projects.
Coarse grained GA (Kandil 50 7,092 132 Despite the new uncrashing heuristic outperforms the state-of-
and El-Rayes 2006) the-art time-cost optimization methods, it requires long CPU times
PFPSO (Aminbakhsh and 1 292 1,023 for large size projects, particularly for the Pareto front problem.
Sonmez 2017) Future research efforts may focus on reducing the CPU time of
Uncrashing heuristic 1 292 2,356 heuristic and metaheuristic methods for the Pareto front problem,
(this study)
as achieving high-quality solutions for this problem requires

© ASCE 04020084-12 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


significantly more time than the cost minimization problem. De, P., E. James Dunne, J. B. Ghosh, and C. E. Wells. 1995. “The discrete
Network reduction techniques that are not included in this heuristic time-cost tradeoff problem revisited.” Eur. J. Oper. Res. 81 (2):
could be included in the future, such as exact, heuristic, and meta- 225–238. https://doi.org/10.1016/0377-2217(94)00187-H.
heuristic time-cost optimization methods to reduce the computa- Elbeltagi, E., T. Hegazy, and D. Grierson. 2005. “Comparison among five
tional time requirements for solving the real-life–sized projects. evolutionary-based optimization algorithms.” Adv. Eng. Inf. 19 (1):
43–53. https://doi.org/10.1016/j.aei.2005.01.004.
The mixed-integer programming method presented in this research
Elbeltagi, E., T. Hegazy, and D. Grierson. 2007. “A modified shuffled
was able to achieve optimal solutions for the Pareto problem instan-
frog-leaping optimization algorithm: Applications to project manage-
ces with up to 200 activities. Future exact methods that enable ment.” Struct. Infrastruct. Eng. 3 (1): 53–60. https://doi.org/10.1080
determination of global optimal solutions for real-life–sized proj- /15732470500254535.
ects would facilitate improved performance evaluation of time-cost Feng, C. W., L. Liu, and S. A. Burns. 1997. “Using genetic algorithms to
optimization capabilities of state-of-the-art methods. solve construction time-cost trade-off problems.” J. Comput. Civ. Eng.
11 (3): 184–189. https://doi.org/10.1061/(ASCE)0887-3801(1997)11:3
(184).
Goyal, S. K. 1975. “A note on a simple CPM time-cost tradeoff algorithm.”
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

Data Availability Statement


Manage. Sci. 21 (6): 718–722. https://doi.org/10.1287/mnsc.21.6.718.
Data generated by the authors or analyzed during the study are Hegazy, T. 1999. “Optimization of construction time-cost trade-off analysis
available at https://blog.metu.edu.tr/rsonmez/research/dtctp, and using genetic algorithms.” Can. J. Civ. Eng. 26 (6): 685–697. https://doi
.org/10.1139/l99-031.
http://etd.lib.metu.edu.tr/upload/12622120/index.pdf.
Kandil, A. 2005. “Multi-objective optimization for large-scale highway
construction projects.” Ph.D. thesis, Dept. of Civil Engineering, Univ.
of Illinois at Urbana Champaign.
Acknowledgments Kandil, A., and K. El-Rayes. 2006. “Parallel genetic algorithms for
optimizing resource utilization in large-scale construction projects.”
The authors gratefully acknowledge the financial support provided J. Constr. Eng. Manage. 132 (5): 491–498. https://doi.org/10.1061
by the Scientific and Technological Research Council of Turkey /(ASCE)0733-9364(2006)132:5(491).
(TÜBİTAK), under Grant No. 213M253. Klanšek, U. 2015. “Mixed-integer nonlinear programming model for non-
linear discrete optimization of project schedules under restricted costs.”
J. Constr. Eng. Manage. 142 (3): 04015088. https://doi.org/10.1061
References /(ASCE)CO.1943-7862.0001074.
Knowles, J. 2006. “ParEGO: A hybrid algorithm with on-line landscape
Abuwarda, Z., and T. Hegazy. 2019. “Multi-dimensional optimization approximation for expensive multiobjective optimization problems.”
model for schedule fast-tracking without over-stressing construction IEEE Trans. Evol. Comput. 10 (1): 50–66. https://doi.org/10.1109
workers.” Can. J. Civ. Eng. 46 (12): 1160–1173. https://doi.org/10 /TEVC.2005.851274.
.1139/cjce-2018-0544. Liberatore, M., B. Pollack-Johnson, and C. Smith. 2001. “Project manage-
Afshar, A., A. Ziaraty, A. Kaveh, and F. Sharifi. 2009. “Nondominated ment in construction: Software use and research directions.” J. Constr.
archiving multicolony ant algorithm in time–cost trade-off optimiza- Eng. Manage. 127 (2): 101–107. https://doi.org/10.1061/(ASCE)0733
tion.” J. Constr. Eng. Manage. 135 (7): 668–674. https://doi.org/10 -9364(2001)127:2(101).
.1061/(ASCE)0733-9364(2009)135:7(668). Ng, S., and Y. Zhang. 2008. “Optimizing construction time and cost
Agdas, D., D. J. Warne, J. Osio-Norgaard, and F. J. Masters. 2018. “Utility using ant colony optimization approach.” J. Constr. Eng. Manage.
134 (9): 721–728. https://doi.org/10.1061/(ASCE)0733-9364(2008)
of genetic algorithms for solving large-scale construction time-cost
134:9(721).
trade-off problems.” J. Comput. Civ. Eng. 32 (1): 04017072. https://doi
Siemens, N. 1971. “A simple CPM time-cost tradeoff algorithm.”
.org/10.1061/(ASCE)CP.1943-5487.0000718.
Manage. Sci. 17 (6): B-354–B-363. https://doi.org/10.1287/mnsc.17
Akkan, C., A. Drexl, and A. Kimms. 2005. “Network decomposition-based
.6.B354.
benchmark results for the discrete time–cost tradeoff problem.” Eur. J.
Sonmez, R., and Ö. H. Bettemir. 2012. “A hybrid genetic algorithm for the
Oper. Res. 165 (2): 339–358. https://doi.org/10.1016/j.ejor.2004.04
discrete time–cost trade-off problem.” Expert Syst. Appl. 39 (13):
.006.
11428–11434. https://doi.org/10.1016/j.eswa.2012.04.019.
Al Haj, R. A., and S. M. El-Sayegh. 2015. “Time–cost optimization model
Vanhoucke, M. 2005. “New computational results for the discrete time/cost
considering float-consumption impact.” J. Constr. Eng. Manage.
trade-off problem with time-switch constraints.” Eur. J. Oper. Res.
141 (5): 04015001. https://doi.org/10.1061/(ASCE)CO.1943-7862 165 (2): 359–374. https://doi.org/10.1016/j.ejor.2004.04.007.
.0000966. Vanhoucke, M., J. Coelho, D. Debels, B. Maenhout, and L. V. Tavares.
Aminbakhsh, S. 2018. “Heuristic and exact methods for the large-scale dis- 2008. “An evaluation of the adequacy of project network generators
crete time-cost trade-off problems.” Ph.D. thesis, Dept. of Civil Engi- with systematically sampled networks.” Eur. J. Oper. Res. 187 (2):
neering, Middle East Technical Univ. 511–524. https://doi.org/10.1016/j.ejor.2007.03.032.
Aminbakhsh, S., and R. Sonmez. 2016. “Discrete particle swarm optimi- Vanhoucke, M., and D. Debels. 2007. “The discrete time/cost trade-
zation method for the large-scale discrete time–cost trade-off problem.” off problem: Extensions and heuristic procedures.” J. Scheduling
Expert Syst. Appl. 51 (Jun): 177–185. https://doi.org/10.1016/j.eswa 10 (4–5): 311–326. https://doi.org/10.1007/s10951-007-0031-y.
.2015.12.041. Van Veldhuizen D. A. 1999. “Multiobjective evolutionary algorithms:
Aminbakhsh, S., and R. Sonmez. 2017. “Pareto front particle swarm op- Classifications, analyses, and new innovations.” Ph.D. thesis,
timizer for discrete time-cost trade-off problem.” J. Comput. Civ. Eng. Graduate School of Engineering, Air Force Institute of Technology,
31 (1): 04016040. https://doi.org/10.1061/(ASCE)CP.1943-5487 Air Univ.
.0000606. Yang, I. 2007. “Using elitist particle swarm optimization to facilitate bicri-
Bader, J. M. 2010. Hypervolume-based search for multiobjective optimi- terion time-cost trade-off analysis.” J. Constr. Eng. Manage. 133 (7):
zation: Theory and methods (No. 112). Zurich, Switzerland: Johannes 498–505. https://doi.org/10.1061/(ASCE)0733-9364(2007)133:7(498).
Bader. Yang, I. T., Y. C. Lin, and H. Y. Lee. 2013. “Use of support vector regres-
Beume, N., B. Naujoks, and M. Emmerich. 2007. “SMS-EMOA: Multiob- sion to improve computational efficiency of stochastic time-cost trade-
jective selection based on dominated hypervolume.” Eur. J. Oper. Res. off.” J. Constr. Eng. Manage. 140 (1): 04013036. https://doi.org/10
181 (3): 1653–1669. https://doi.org/10.1016/j.ejor.2006.08.008. .1061/(ASCE)CO.1943-7862.0000784.

© ASCE 04020084-13 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084


Zhang, H., and H. Li. 2010. “Multi-objective particle swarm optimization Zitzler, E., K. Deb, and L. Thiele. 2000. “Comparison of multiobjective
for construction time-cost tradeoff problems.” Constr. Manage. Econ. evolutionary algorithms: Empirical results.” Evol. Comput. 8 (2):
28 (1): 75–88. https://doi.org/10.1080/01446190903406170. 173–195.
Zheng, D., S. Ng, and M. Kumaraswamy. 2005. “Applying Pareto ranking Zitzler, E., and L. Thiele. 1998. “Multiobjective optimization using
and Niche formation to genetic algorithm-based multiobjective time– evolutionary algorithms—A comparative case study.” In Proc., Int.
cost optimization.” J. Constr. Eng. Manage. 131 (1): 81–91. https://doi Conf. on Parallel Problem Solving from Nature, 292–301. Berlin:
.org/10.1061/(ASCE)0733-9364(2005)131:1(81). Springer.
Downloaded from ascelibrary.org by East Carolina University on 05/25/20. Copyright ASCE. For personal use only; all rights reserved.

© ASCE 04020084-14 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(8): 04020084

You might also like