Automation in Construction 75 (2017) 1–9

Contents lists available at ScienceDirect

Automation in Construction

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

An effective approach with feasible space decomposition to solve

resource-constrained project scheduling problems
Zhenyuan Liu a,b,⁎, Liu Yang a, Raoyi Deng a, Jing Tian a
a
College of Automation, Huazhong University of Science and Technology, Wuhan 430074, China
b
Key Laboratory of Education Ministry for Image Processing and Intelligent Control, Wuhan 430074, China

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

Article history: Resource-constrained project scheduling problem is a classic problem in construction project. Aimed at solving
Received 8 May 2016 this problem, an effective approach with decomposition on time windows is proposed in this paper. This ap-
Received in revised form 14 November 2016 proach is to select one activity to do decomposition and to partition the feasible space of the original problem
Accepted 24 November 2016
into some feasible subspaces, in which solutions are generated by using an extended serial scheduling scheme.
Available online 8 December 2016
Double justification is also performed in the process of searching in subspace. Four strategies for selecting activity
Keywords:
to do decomposition, three strategies for decomposition and a strategy on sampling size in various subspaces are
Project scheduling designed. The results of experiments on two real construction projects show that the strategy based on degree for
Decomposition-based approach selecting activity and the strategy based on initial schedule for decomposition can obtain the best results. When
Extended serial scheduling scheme compared with some other exiting algorithms, it is proven that the decomposition-based approach is effective
Double justification and competitive.
Construction management © 2016 Elsevier B.V. All rights reserved.

1. Introduction During the past 30 years, kinds of meta-heuristic methods have been
put forward to find more accurate solutions within shorter computing
The resource-constrained project scheduling problem (RCPSP) is a time. In their procedures, solutions are usually encoded firstly. Then
classic NP-hard optimization problem [1] which can be found frequent- the codes are visited by using some meta-heuristic strategy and finally
ly in construction project, production development etc. Various proce- decoded to generate schedules. Random-key and activity-list are two
dures for solving the RCPSP including exact and heuristic ones have typical encoding methods [8]. Also choosing a decoding process accord-
been developed in the last several decades years [2,3]. Among them, ing to characteristics of the problem reflects the flexibility of the algo-
branch-and-bound method is a typical exact procedure, which is able rithm [9]. The widely used meta-heuristic algorithms include Genetic
to find optimal solutions in feasible time [4,5]. However, exact solution Algorithms (GAs) [10–12], particle swarm optimization [13,14], simu-
procedures are restricted to small or medium-scale RCPSPs. In order to lated annealing algorithm [15] and ant system algorithm [16], etc. For
solve large-scale problems, heuristic methods have been proposed example, Zamani presented a Genetic Algorithm with a magnet-based
which could generate optimal or suboptimal solutions within accept- crossover operator [10], Kadam and Mane combined GA with local
able time. Serial Scheduling Scheme (SSS) and Parallel Scheduling search algorithm [11], and Chen and Weng presented a two-phase GA
Scheme are two popular schedule generation schemes (SGSs) in this model for RCPSP [12]. Georgios Koulinas et al. proposed a particle
area, which can generate feasible schedules with priority rules-based swarm optimization based hyper-heuristic algorithm by using random
selection of activities stage by stage [6]. In recent decades, these two keys as the solution representation [13]. Similarly, Anantathanvit and
schemes have been widely combined with lots of meta-heuristic algo- Munlin extended the particle swarm optimization by regrouping the
rithms to solve the large-scale RCPSPs. Besides, priority rules were pro- agent particle within the appropriate radius of the circle to solve the
posed to construct schedules based on the SGS. The most popular problem [14]. Anagnostopoulos and Koulinas proposed simulated an-
priority rules include latest finish time (LFT), shortest process time, nealing hyper-heuristics [15] and Li and Zhang put forward an ant colo-
minimum slack etc. The priority rules can make different influences ny optimization-based multi-mode scheduling algorithm [16].
on the performance of heuristics [7]. As for the new approaches in recent years, such as bee algorithm
[17], analogous immune algorithm [18], shuffled frog-leaping algorithm
[19], and neurogenic algorithms [20] have also been applied to the
⁎ Corresponding author at: College of Automation, Huazhong University of Science and
Technology, Wuhan 430074, China. RCPSP. For other heuristic algorithms, Wang and Fang developed a hy-
E-mail address: [email protected] (Z. Liu). brid estimation of distribution algorithm [21] and Liu et al. designed

http://dx.doi.org/10.1016/j.autcon.2016.11.012
0926-5805/© 2016 Elsevier B.V. All rights reserved.
2 Z. Liu et al. / Automation in Construction 75 (2017) 1–9

an activity-list-based nested partitions algorithm recently [22]. Cheng et Table 1

al. proposed a novel fuzzy clustering chaotic-based differential evolu- Earliest finish times and latest finish times.

tion with serial method [23]. Finish times 1 2 3 4 5 6

Evaluation on the performance of these heuristic algorithms shows EF 3 4 5 6 6 10
that the meta-heuristic often costs far more time than the heuristic. In LF 17 14 19 16 20 20
order to saving computing time, some researchers reduce solution
space by decomposing the RCPSP into sub-problem. Benders decompo-
sition is an effective method to solve mixed-integer liner programming of activity j. Another variable is At, the set of activities being in progress
problem [24]. Recently, a hybrid bender decomposition algorithm out- in period t.
performs pure constrain programming in solution quality and speed The model of RCPSP can be formulated as the following:
for solving project scheduling problem with multi-purpose resources,
through decomposing the mixed-integer linear programming into a re- minx J ð1Þ
laxed master assignment problem and a feasibility scheduling sub-
problem [25]. These two sub-problems are connected by benders cuts s:t:xi ≤x j −d j ; ∀i∈P j ; ∀j ¼ 1; …; J ð2Þ
to exclude infeasible solutions.
When searching feasible schedules in solution space, we often need ∑ r jk ≤ Rk ; ∀k ¼ 1; …; K; t ¼ 1; …; D ð3Þ
j∈At
to rank activities by checking their earliest finish (start) time and latest
finish (start) time which can be called finish (start) time windows of ac- x j ≥ 0; ∀j ¼ 1; …; J ð4Þ
tivities to satisfy precedence relation constraints and resource con-
straints. These time windows are combined to form the whole feasible The formula (1) shows the objective function which minimizes the
space without consideration of resource constraints. RCPSP has been completion time of the project. (2), (3), (4) are constraints. Constraint
solved with various algorithms in previous work, but there is still no re- (2) considers immediate predecessor and successor relationship of all
search trying to decompose these time windows into sub-windows activities and ensures that start time of every activity is not earlier
which may reduce searching space. In one of our work, a rough ap- than the finish times of its immediate predecessors. (3) indicates that
proach with the idea of decomposition has been developed and some the total resource usage per period is in the range of availability. (4) en-
rules for decomposition have been designed [26]. However, decomposi- sures that finish time of an activity is not negative.
tion on feasible space has not been defined clearly. The SSS embedded in
the procedure just operates directly on the original sub-windows with- 3. Decomposition on feasible space
out considering that the feasible start times of activities will change dur-
ing the scheduling procedure. There are only some simple rules for 3.1. A notation of feasible space
decomposition and sampling. The numerical studies are not done on
real cases and the results are not satisfying. Actually, the feasible space of RCPSP can be derived from the con-
This work is an extension of our previous work in [26] where the straints (1) to (4). But it is not obvious because of the resource con-
procedure will be defined in a systematic way. The remainder of the straints. If this type of constraints is not considered, the feasible space
paper is organized as follows. Section 2 provides the description of the can be denoted as a J-dimension-space Ω which is actually the combina-
RCPSP. And then the decomposition methods on time windows are pro- tion of finish time windows of activities.
posed in Section 3. Section 4 introduces the decomposition method-


based approaches and Section 5 shows several case studies about exper- Ω ¼ ½EF 1 ; LF 1   ½EF 2 ; LF 2   ⋯  EF J ; LF J ð5Þ
imental design and result comparisons on the PSPLIB. Finally, some
conclusions are presented in Section 6. For example, Fig. 1 is the activity on node network of a project exam-
ple. The deadline is set as 20 and the finish times obtained with CPM are
2. Problem description shown in Table 1.
Feasible space of this example can be defined as:
In RCPSP, it is assumed that a project consists of J activities. Without
loss of generalization, we can assume that activity 1 is the only start activ- Ω ¼ ½3; 17  ½4; 14  ½5; 19  ½6; 16  ½6; 20  ½10; 20
ity and activity J is the only end activity. The duration of activity j is dj
periods. In this case, solving the RCPSP can be regarded as a procedure of
In the project, an activity j ∈ {2, ⋯ ,J} may not be started before any of finding a schedule in Ω to satisfy resource constraints and precedence
its immediate predecessors i ∈ Pj (Pj is the set of immediate predecessors relation constraints with minimizing the makespan of the project.
of activity j) has been finished. There are K-type of renewable resources.
Rk indicates the capacity of resource k. rjk(∀k) units of resource k are re- 3.2. Decomposition operator
quired to process activity j during every period of its duration.
The time parameters in the problem are all integer valued. We use D With the above notation of feasible space, we can employ a decom-
to denote the deadline of the project. The earliest finish time EFj and the position operator to partition Ω into several subspaces, as Fig. 2 shows.
latest finish time LFj of activity j can be obtained with this D. We use a set The finish time window of activity i is [EFi, LFi]. It is assumed that the
of integer decision variables xj ∈ [EFj,LFj] (j = 1, … , J) to be the finish time time window will be partitioned into M feasible sub-windows: [EF1i , LF1i ],

3 2 1 di
1 3 5
i
2 4 3 ri
start end
4 2 4 Deadline=20
2 4 6 Rmax=4
3 4 2

Fig. 1. Activity on node network of project example. Fig. 2. Decomposition Operation on Feasible Space.
 Z. Liu et al. / Automation in Construction 75 (2017) 1–9 3

Start windows at time point 10, then the subspaces after decomposition are
as the following:

Calculate the finish times of activities Ω1 : ½3; 17  ½4; 14  ½5; 19  ½6; 10  ½6; 20  ½10; 20

Ω2 : ½3; 17  ½4; 14  ½5; 19  ½11; 16  ½6; 20  ½10; 20
Generate an Initial Schedule

3.3. Redefinition of feasible subspaces

Define the feasible space
Definition 1. Transitive predecessors and transitive successors [27]

Select an decomposed activity Assuming the precedence relationship between activities couple
(i, j) is denoted as Pre(i, j). If Pre(i, j) = 1, then activity i is a predecessor
of activity j, otherwisePre(i, j) = 0. For an activities couple (i, j) in an ac-
Do decomposition on feasible space tivity network G = (V,A) where V is the activities set and A is the prece-
dence relationship set, ifPre(i0, i1) × ⋯ × Pre(is − 1, is) =
1 , ∃(i0,i1) , ⋯ , (is−1,is) ∈ A , i0 , i1 , ⋯ , is−1 ,is ∈ V, i0 = i , is =j, then activ-
Generate schedules in subspaces ity i is called a transitive predecessor of activity j.
TPj and TSj are respectively the set of all the transitive predecessors
of j and the set of all transitive successors of j. The union of all transitive
Output the optimal solution
successors and all transitive predecessors of activity j is denoted as
Qj , Qj =TPj ∪ TSj.
Stop After partitioning on time window of some activities, it is probably
that time windows of the related activities will be changed because of
the precedence relationship. If the latest finish time (the earliest finish
Fig. 3. The flow chart of OPDA. time) of some activity was changed, we need to redefine the latest finish
times (the earliest finish time) of its transitive predecessors (transitive
successors). Critical Path Method is employed to redefine the changed
[EF2i , LF2i ], …, [EFM M 1 M l l−1
i , LFi ], where EFi = EFi, LFi = LFi, EFi = LFi + 1, time windows of activities accordingly so that the feasible subspace will
l l
2 ≤ l ≤ M. Then, the lth sub-window [EFi, LFi] of activity i will be utilized be changed.
to substitute its original finish time window [EFi, LFi ] in Ω to form the In the example in Fig. 1, subspace 1 and 2 are redefined as the follow-
lth feasible subspace: ing, where the finish time window of activity 2 is changed in Ω1 and the
finish time window of activity 6 is changed in Ω2.
h i

Ωl : ½EF 1 ; LF 1   …  EF li ; LF li  …  EF J ; LF J ð6Þ
Ω1 : ½3; 17  ½4; 8  ½5; 19  ½6; 10  ½6; 20  ½10; 20

In the example in Fig. 1, if we choose activity 4 for decomposition

and its feasible finish time window is divided into two feasible sub- Ω2 : ½3; 17  ½4; 14  ½5; 19  ½11; 16  ½6; 20  ½15; 20

Fig. 4. The base procedure of the ESSS.

4 Z. Liu et al. / Automation in Construction 75 (2017) 1–9

a) Case 1

4 5 6 di
6 3 6 10 5 i
1 2 9 6 16 5 ri
2 6 1 2 4 19
5 11 14
3 2 2
3 4 3 end
start 2 18
3
3 2 2 3 5 20
8 12 15
6 8 5
5 8 6
4 2 2 4 17 Rmax=9
5 7 9 13 1
3 3 5

b) Case 2

15 10 10
7 9 16 di
5 8 10 5
11 12 21 24 27 i
2 5 4 10 32
2 14 11 10 7 4 5 5 ri
23 3 2 7 14
2 18 5 13 17 10
7 end
start 4 4 4 14 30 34 36
4 5 19
10 14 7 3 10 1 1
25 6 9 26 9 1 7
12 12 14 16 4 37
1 6 4 28 33 35
9 4 5 3 1
2 21 3
7 9 10 14 6 5 2 1
3 5 12
15 18 20 22 29
4 7 31 Rmax=12
4 6 6 3 2
25 4
5

Fig. 5. Network of projects [23].

It has been proved that any feasible solution in Ω must be in some Without loss of generalization, it is assumed that the scheduling
feasible subspace after decomposition [28]. procedure will be called in some subspace: Ω′ = [EF1′, LF1′] ×
[EF2′, LF2′] × … × [EF′,J LF′].
J

4. One-Point Decomposition-Based Approach for RCPSP Similar to the classic SSS, the whole base procedure of ESSS consists of
J iterations shown as Fig. 4. In each iteration, an activity is selected by
4.1. The framework of OPDA using a priority rule. Its start time and finish time will be decided accord-
ing to the finish time constraints in the subspace and the constraints on
Here, we only select one activity to do decomposition, so we call the resource availability and precedence relationships. The procedure will
approach One-Point Decomposition-Based Approach (OPDA). Fig. 3 is be called iteratively until all activities are scheduled. If resource con-
the flow chart of the approach. The detailed procedure is described as straints cannot be satisfied during the procedure, it will stop.
the following. In Fig. 4, En is an activity set in which activities could be scheduled
with precedence constraints in iteration n. Cn is an activity set in
Step 1: Calculate the earliest and latest finish times of activities. which activities have already been scheduled in iteration n. Vj is the pri-
ority value of activity j with some rule for selection.
Step 2: Generate an initial schedule using the SSS. Priority rules for selecting activities in each iteration are usually ESj,
Step 3: Set the makespan of the initial schedule as the deadline of the LS j ; LF j ; r j ; FT j 0 ; SLK which means earliest start time, latest start time,
project and calculate the finish times of activities again. Then, the latest finish time, resource consumed in unit time, finish time of the op-
feasible space Ω can be defined. timal result in initial scheduling of the activity and time lag between
Step 4: Select an activity for decomposition with some selection earliest start time and latest start time [7]. In this paper we use the pri-
strategy. ority rule of LFj in ESSS for activity selection. The mode used to select ac-
Step 5: Do decomposition on Ω with some decomposition strategy to tivities to do scheduling randomly is Regret-based Biased Sampling [6],
form feasible subspaces.
Step 6: Generate schedules in each feasible subspace respectively
using an extended SSS (ESSS) with a strategy on sampling size. Table 2
Step 7: Find the optimal solution from all the schedules generated in Design of Experiments for the OPDA.

Step 6 and output it. Parameters Case 1 Case 2

Combination of strategies S1-R1/S1-R2/S1-R3 S1-R1/S1-R2/S1-R3

4.2. The extended SSS S2-R1/S2-R2/S2-R3 S2-R1/S2-R2/S2-R3
S3-R1/S3-R2/S3-R3 S3-R1/S3-R2/S3-R3
Here the classic SSS will be extended for generating schedules in fea- S4-R1/S4-R2/S4-R3 S4-R1/S4-R2/S4-R3
Sampling size 1000/3000/5000 1000/5000/50000
sible subspaces. We call it the extended SSS (ESSS).
 Z. Liu et al. / Automation in Construction 75 (2017) 1–9 5

a) Case 1

9 17
7 3
10 15 13 18
5 4 2 6 20
12
3 8
7 5 11
9 14 16
1 1 19
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 T

b) Case 2

12
2 5 16 22
10 8 7 14 25 24 27 30
29
8 10 18 20
15 21
6 10 12 6 23
4
4 9 11 13 28 33
2 3 17 19 26 31 37
1 32 34 35 36

8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128 136 144 152 160 168 176 184 190 T

Fig. 6. Optimal solutions for two case.

in which regret value ρ(j) on selection of activity j will be set according Definition 2. Degree is defined as the number of other nodes connected
to the priority value with the worst sequence among the feasible activ- with a node in a graph. Let de1 , de2 , …,deJ be the degree of activities 1,
ities in En, 2 , … , J respectively. Degree of activity j is represented as dej = -
count(Pj)+count(Sj), where function count(.) is to record the number of
ρð jÞ ¼ maxi∈En φðiÞ−φð jÞ ð7Þ elements in a set and Sj is the set of immediate successors of activity j.

Wherein it assumed that a “minimal” priority rule is employed. Definition 3. Central degree is defined as the absolute value of the differ-
Then, the probability for activity j to be selected is: ence between the number of activities in the set of all the transitive pre-
decessors and the number of activities in the set of all the transitive
ðρð jÞ þ 1Þα successors. Let ctj =|count(TPj)−count(TSj)| be the central degree of ac-
Φð jÞ ¼ X ð8Þ tivity j.
i∈E
ðρðiÞ þ 1Þα
n
According to the above definitions, four strategies are developed for
In ESSS, double justification is additionally performed to improve the selecting the decomposed activity. They are denoted as S1 , S2 , S3 , S4.
results obtained in the base procedure, which has testified to be an effi- n o
cient method to improve a feasible solution [29]. S1: min jj max pr j ð9Þ
j∈V

4.3. Strategy for selecting activity to do decomposition

In (9), the index of transitive relationship for activity j is represented
Different strategy for selecting an activity to do decomposition as prj =count(Qj). Activity j that has larger prj is preferred to make decom-
would diversify the decomposing operation. We give two definitions position more effective because of precedence constraints among activi-
firstly. ties. The ties will be broken to choose the activity with a smaller

Table 3
Results comparison between different strategies in sampling size 1000.

Rules Case 1 Case 2

Success rate Optimal value Time(s) Success rate Optimal value Time(s)

Best Avg. Std. Worst. Avg. Best Avg. Std. Worst Avg.

R1 S1 80% 42 42.2 0.407 43 0.013 3.33% 190 191.733 1.2299 195 0.0598
S2 80% 42 42.2 0.407 43 0.0125 3.33% 190 191.733 1.2299 195 0.0593
S3 83.3% 42 42.2 0.379 43. 0.0198 6.67% 190 191.533 1.2794 196 0.0796
S4 80% 42 42.2 0.407 43 0.0125 3.33% 190 191.733 1.2298 195 0.0582
R2 S1 93.33% 42 42.07 0.25 43 0.0146 6.67% 190 191.13 0.6814 194 0.084
S2 93.33% 42 42.07 0.25 43 0.0166 6.67% 190 191.13 0.6814 194 0.0852
S3 93.33% 42 42.07 0.25 43 0.0156 16.67% 190 191.2 0.9613 194 0.0822
S4 83.33% 42 42.17 0.38 43 0.0177 6.67% 190 191.13 0.6814 194 0.0848
R3 S1 53.33% 42 42.5 0.57 44 0.0109 0 191 191.93 1.285 195 0.0567
S2 53.33% 42 42.5 0.57 44 0.0119 0 191 191.93 1.285 195 0.0535
S3 73.33% 42 42.27 0.45 43 0.0156 3.33% 190 192.33 1.539 196 0.0614
S4 53.33% 42 42.5 0.57 44 0.0114 0 191 191.93 1.2847 195 0.0557
6 Z. Liu et al. / Automation in Construction 75 (2017) 1–9

Table 4 4.4. Strategy for decomposition

Results comparison between different strategies in sampling size 3000 for Case 1.

Rules Case 1 In this approach, the finish time window of the selected activity j will
Success rate Optimal value Time(s)
be partitioned into some sub-windows. Here, we take two sub-win-
dows for example. The two sub-windows will be left: [EFj , Δj] and
Best Avg. Std. Worst. Avg.
right: [Δj + 1 , LFj]. In this case, there are two rules set to get Δj, which
R1 S1 100.00% 42 0 0 42 0.0354 have been proposed in our previous work [26].
S2 100.00% 42 0 0 42 0.0338
R1. Average: Decompose the original time window into two equal
S3 100.00% 42 0 0 42 0.0557
S4 100.00% 42 0 0 42 0.03537
sub-windows. ⌈.⌉ is the minimum integer not less than “.”.
R2 S1 100.00% 42 42 0 42 0.0551  
S2 100.00% 42 42 0 42 0.0523 Δ j ¼ ⌈ L F j −E F j =2þE F j ⌉ ð13Þ
S3 100.00% 42 42 0 42 0.0536
S4 100.00% 42 42 0 42 0.0556
R2. Biased: Select the best finish time of activity j generated in the
R3 S1 100.00% 42 42 0 42 0.0322
S2 96.67% 42 42.033 0.1826 43 0.0458 initial scheduling as the decomposition time point.
S3 100.00% 42 42 0 42 0.0307
S4 100.00% 42 42 0 42 0.0343 Δ j ¼ FT j 0 ð14Þ

For three-sub-windows: the sub-windows will be set as [EFj , Δj1],

sequence number when several activities have the same index of transi- [Δj1 + 1 , Δj2] and [Δj2 + 1 , LFj]. We will use an average rule R3 in this
tive relationship. case.
R3. Decompose the original time window into three equal sub-win-
n o dows.
S2: min jj max pr j ; min ct j ð10Þ
j∈V  
Δ j1 ¼ ⌈ LF j −E F j =3þE F j ⌉ ð15Þ
 
Activity j with smaller ctj shows the uniform distribution of its im- Δ j2 ¼ ⌈2 L F j −E F j =3þE F j ⌉ ð16Þ
mediate predecessors and immediate successors, which may reduce
the solution space. Hence we select activity j with minimum ctj from
the activity set in which the activities have the maximum prj, and we
4.5. Strategy on sampling size
choose the activity with smaller sequence number when several activi-
ties have the same index shown as (10).
In numerical experiments for solving RCPSP with various heuristics,
the sampling sizes are usually set as 1000, 5000 and 50,000 [8]. One
 
S3: min jj max de j ð11Þ sampling is the procedure for searching one solution. The sampling
j∈V
sizes of feasible subspaces means the number of searching solutions.
We use the following strategy to set the sampling proportion of each
Degree is a measure of the impact on the decomposed time windows subspace. If the total sampling sizes is N and the time window of the
in the project network. The larger dej is, the more activity j impacts. We decomposed activity is partitioned into M sub-windows with the length
select activity j that has the maximum dej, and we choose the activity n1 , n2 , … , nm , … , nM, then the sampling size of the mth feasible sub-
with smaller sequence number when several activities have the same space is:
index in S3. X
SSm ¼ N  nm = nm0 ð17Þ
( !) m0 ¼1⋯M

S4: min jj max de j þ dei ∑ ð12Þ

j∈V i∈Q j
5. Case studies

This strategy considers the degree of transitive successors and tran- In this section we present the results of numerical experiments. In
sitive predecessors of activity j on the basis of S3. the analysis of experimental results, we firstly discuss the results by

Table 5
Results comparison between different strategies in sampling size 5000.

Rules Case 1 Case 2

Success rate Optimal value Time(s) Success rate Optimal value Time(s)

Best Avg. Std. Worst. Avg. Best Avg. Std. Worst Avg.

R1 S1 100% 42 42 0 42 0.0629 13.33% 190 191.033 0.7184 194 0.3026

S2 100% 42 42 0 42 0.06343 13.33% 190 191.033 0.7184 194 0.3047
S3 100% 42 42 0 42 0.0868 53.30% 190 190.567 0.8172 194 0.3853
S4 100% 42 42 0 42 0.06343 13.33% 190 191.033 0.7183 194 0.3026
R2 S1 100% 42 42 0 42 0.0852 33.30% 190 190.667 0.4795 191 0.404
S2 100% 42 42 0 42 0.0868 33.30% 190 190.667 0.6814 191 0.3963
S3 100% 42 42 0 42 0.0842 66.70% 190 190.667 0.4795 191 0.0962
S4 100% 42 42 0 42 0.0874 23.33% 190 191 1.05 195 0.3805
R3 S1 100% 42 42 0 42 0.049 10.00% 190 191.067 0.5833 193 0.2767
S2 100% 42 42 0 42 0.05 10.00% 190 191.067 0.5833 193 0.2714
S3 100% 42 42 0 42 0.0722 36.67% 190 190.767 0.8172 194 0.3614
S4 100% 42 42 0 42 0.051 10.00% 190 191.067 0.5833 193 0.2689
 Z. Liu et al. / Automation in Construction 75 (2017) 1–9 7

Table 6 5.2.1. Results comparison between different strategies

Results comparison between different strategies in sampling size 50,000 for Case 2. First of all, let us discuss the performance of the approach with dif-
Rules Case 2 ferent strategies that we choose (see Tables 3–6). Table 3 shows that
Success rate Optimal value Time(s)
success rates of experiments in which the feasible time window is
partitioned into two sub-windows (R1 and R2) are significantly higher
Best Avg. Std. Worst. Avg.
than that of R3 when sampling size is small. The best strategy we pro-
R1 S1 93.33% 190 190.0667 0.2537 191 3.0067 posed is R2. When sampling size and decomposition strategy of activity
S2 93.33% 190 190.0667 0.2537 191 2.9822
are the same, selection strategy for activity –S3 is the best one, which
S3 100% 190 190 0 190 3.395
S4 93.33% 190 190.0667 0.2537 191 3.0217
has the highest success rate and the lowest average result. For Case 1,
R2 S1 100.00% 190 190 0 190 4.39 Table 4 shows that success rate of different experiments almost reach
S2 100.00% 190 190 0 190 4.3821 100% when sampling size is 3000, in addition to the combination of
S3 100.00% 190 190 0 190 3.739 strategies S2-R3. When sampling size is 5000, the success rate is up to
S4 100.00% 190 190 0 190 4.494
100% and optimal values are all the same for all experiments as Table 5
R3 S1 76.67% 190 190.233 0.4302 191 2.764
S2 76.67% 190 190.233 0.4302 191 2.717 shows. For Case 2, when sampling size is 1000 and 5000, experimental
S3 100.00% 190 190 0 190 3.547 results are not satisfactory because the number of activities is more
S4 76.67% 190 190.233 0.4302 191 2.698 than that of Case 1 and the activity network is more complex. Table 6
shows that some of success rates are 100% when sampling size is 50,000.
Computational times are additional characteristics used to evaluate
using the OPDA with different strategies. Then we compare results of the performance of the OPDA. Tables 3–6 show that there is no signifi-
the approach with those of some other algorithms for the same cases. cant differences in computational times when the decomposition strat-
In this experiment, we use Visual Studio 2008 for coding and the Win- egy is the same, while the computational times of the decomposition
dows 7 (CPU is 2.50 GHz and RAM is 4.00GB) as the operating strategy R3 are significantly less than that of R2 in most cases and are
environment. near to those of R1.

5.1. Experimental design

5.2.2. Results comparison between different algorithms
The experiments have been performed on two real construction case To verify the effectiveness of the OPDA, we compare its performance
studies adapted from Jun and El-Rayes [30] and Chen and Weng [12], with those of other four different algorithms. These four algorithms in-
the same with those in [23]. The two cases fully meet the demand of clude the original DE [31], Genetic Algorithm (GA) [12], CDE and FCDE
the data the approach needs. Fig. 5 shows the precedence relationships [23]. The results of optimal value and computational times of other
of the network projects. Cases 1 and 2 consists of 20 activities and 37 ac- models from [23] are listed in Table 7. We choose the result of the
tivities respectively. Two cases both have one renewable resource, least computational time to compare with other models when success
where Case 1 has a maximum availability limit of 9 per day and Case 2 rate is up to 100%. For Case 1, we choose the result of strategy S3-R3
has 12 per day. The duration of each activity is noted above the corre- with sampling size 3000. For Case 2, we choose the result of strategy
sponding circle node. The amount of required resources is indicated S3-R3 with sampling size 50,000. Table 7 shows that the performance
below the circle node. The precedence constraints among activities are of OPDA is competitive. In terms of success rate and average results,
described by using arrow lines [23]. comparing with other models, the performances of OPDA for RCPSP
The purpose we apply the OPDA for RCPSP is minimizing project du- are both the best for Cases 1 and 2. The OPDA for RCPSP found optimal
ration while satisfying all constraints. The basic framework and search solutions with highest rate of success: 100% and 100%, lowest fitness so-
algorithm of the OPDA combine with the strategies for selecting activity lution: 42 and 190 and standard deviation: 0 and 0 in all executions of
S1–S4, the strategies for decomposition R1–R3 and strategy on sam- Cases 1 and 2, respectively. Although the results of the algorithms are
pling size proposed in chapter 4.3 to 4.5. We have carried on the exper- not running on the same computer, the speed of OPDA for RCPSP is
iments for several different combination strategies. For example, S1-R1 also competitive. It can be concluded that the OPDA for RCPSP we pro-
denotes strategy which choose activity with the maximal index of tran- posed is a competitive approach from the above comparison.
sitive relationship to decompose and the original time window is
partitioned into two equal sub-windows. The design of the experiments
is shown in Table 2 in order to obtain comprehensive experimental re- Table 7
sults of the OPDA. The total sampling sizes are usually 1000, 5000 and Results comparison between OPDA for RCPSP and benchmarked algorithms.
50,000. However, we find that 5000 of sampling size has made success
Performance measurement GA DE CDE FCDE OPDA for
rate to 100% no matter which strategy is used in Case 1. So we set the
RCPSP
total sampling size in Case 1 as 1000, 3000 and 5000.
Case 1 Success rate 70% 73.33% 93.33% 100% 100.00%
Optimal value Best. 42 42 42 42 42
5.2. Result comparisons Avg. 42.51 42.5 42.07 42 42
Std. 0.85 0.86 0.25 0 0
Fig. 6 shows the corresponding optimal solution schedule by using Worst. 44 44 43 42 42
the OPDA for both cases. It is worthy pointing out that Fig. 6 shows Comp. time(s) Best. 5.67 2.34 6.67 6.77 0.015
Avg. 8.91 5.57 12.45 10.91 0.0307
just one of optimal solutions for each case. The optimal objectives of Std. 2.20 3.78 3.46 2.38 0.0096
two cases are 42 and 190 which we can use in the analysis of experi- Worst. 13.23 12.91 16.08 14.01 0.047
ments results. In order to evaluate the effectiveness and stability of Case 2 Success rate 60% 70% 80% 93.33% 100.00%
the approach proposed, the performance of the OPDA for RCPSP is Optimal value Best. 190 190 190 190 190
Avg. 191.33 191.27 190.47 190.13 190
expressed in terms of success rate, best result found (best), average re-
Std. 2.07 2.02 1.01 0.51 0
sult (avg.), standard deviation (std.), and worst result (worst) after 30 Worst. 198 196 194 1.92 190
runs. The success rate denotes the proportion of optimal solution ob- Comp. time(s) Best. 45.67 8.87 31.5 71.89 2.277
tained. The best results and worst results demonstrate the capacity of Avg. 92.78 62.21 200.45 147.16 3.547
the approach to find the optimal solution. Average and standard devia- Std. 16.90 34.86 79.75 40.91 0.972
Worst. 136.67 102.6 275.02 250.65 5.148
tion are two additional characteristics used to describe solutions quality.
8 Z. Liu et al. / Automation in Construction 75 (2017) 1–9

6. Conclusion [3] C. Schwindt, J. Zimmermann, Handbook on Project Management and Scheduling,

vol. 1, Springer International Publishing, 2015, ISBN 978-3-319-05442-1 http://dx.
doi.org/10.1007/978-3-319-05443-8.
RCPSP is a common hard problem to solve in construction project [4] P. Brucker, S. Knust, A. Schoo, O. Thiele, A branch and bound algorithm for the re-
management. An effective approach with feasible space decomposition source-constrained project scheduling problem, Eur. J. Oper. Res. 107 (2) (1998)
272–288, http://dx.doi.org/10.1016/S0377-2217(97)00335-4.
is proposed for it where a decomposition operator is employed to parti- [5] L. Bianco, M. Caramia, An exact algorithm to minimize the makespan in project
tion the feasible space of the original problem into some subspaces and scheduling with scarce resources and generalized precedence relations, Eur. J.
then these subspaces are searched with an extended serial scheduling Oper. Res. 219 (1) (2012) 73–85, http://dx.doi.org/10.1016/j.ejor.2011.12.019.
[6] R. Kolisch, Serial and parallel resource-constrained project scheduling methods
scheme, which is presented in a more systematic way than that in our revisited: theory and computation, Eur. J. Oper. Res. 90 (2) (1996) 320–333,
previous work [26]. Based on the general framework of the approach, http://dx.doi.org/10.1016/0377-2217(95)00357-6.
we have designed 4 strategies for selecting activity to do decomposition, [7] R. Kolisch, Efficient priority rules for the resource-constrained project scheduling
problem, J. Oper. Manag. 14 (3) (1996) 179–192, http://dx.doi.org/10.1016/0272-
3 strategies for decomposition and one strategy on sampling size in var-
6963(95)00032-1.
ious subspaces where these three kinds of strategies are not clearly de- [8] R. Kolisch, S. Hartmann, Experimental investigation of heuristics for resource-
fined in the procedure in [26]. constrained project scheduling: an update, Eur. J. Oper. Res. 174 (1) (2006)
We used two real construction projects to validate the effectiveness 23–37, http://dx.doi.org/10.1016/j.ejor.2005.01.065.
[9] R. Kolisch, A. Drexl, Adaptive search for solving hard project scheduling problems,
of the OPDA for RCPSP. The experiments verify the approach to diminish Nav. Res. Logist. 43 (1) (1996) 23–40 (http://dx.doi.org/10.1002/(SICI)1520-
the searching space and reduce the amount of calculation in the process 6750(199602)43:1%3C23::AID-NAV2%3E3.0.CO;2-P).
of search to improve the efficiency of solving the RCPSP. According to [10] R. Zamani, A competitive magnet-based genetic algorithm for solving the resource-
constrained project scheduling problem, Eur. J. Oper. Res. 229 (2) (2013) 552–559,
the experiments, the sampling size has an influence on the success http://dx.doi.org/10.1016/j.ejor.2013.03.005.
rate of the solutions and solving efficiency. In terms of success rate, [11] S.U. Kadam, S.U. Mane, A Genetic-Local Search Algorithm Approach for Resource
the strategy of decomposing the time window of activity based on the Constrained Project Scheduling Problem, The 2015 International Conference on
Computing Communication Control and Automation, Pune, India, 26–27, Feb.
initial schedule is better than that of decomposing the original time 2015 841–846, http://dx.doi.org/10.1109/ICCUBEA.2015.168.
window into three equal sub-windows, while it is opposed to compare [12] P. Chen, H. Weng, A two-phase GA model for resource-constrained project schedul-
the strategy for computational times. The strategy based on degree for ing, Autom. Constr. 18 (4) (2009) 485–498, http://dx.doi.org/10.1016/j.autcon.2008.
11.003.
selecting activity to do decomposition is the best one. It is proven that [13] G. Koulinas, L. Kotsikas, K. Anagnostopoulos, A particle swarm optimization based
the approach proposed in this paper is fast and efficient. hyper-heuristic algorithm for the classic resource constrained project scheduling
In addition to the strategy for selecting activity, strategy for decom- problem, Inf. Sci. 277 (2014) 680–693, http://dx.doi.org/10.1016/j.ins.2014.02.155.
[14] M. Anantathanvit, M. Munlin, Radius Particle Swarm Optimization for Resource
position and strategy on sampling size proposed in this paper, some
Constrained Project Scheduling, The 16th International Conference on Computer
other strategies can be designed and more activities can be selected to and Information Technology(ICCIT), Khulna, Bangladesh, 8–10, March 2014
do decomposition to improve the efficiency of the approach in the fu- 24–29, http://dx.doi.org/10.1109/ICCITechn.2014.6997330.
ture research. The decomposition-based approach is a framework for [15] K.P. Anagnostopoulos, G.K. Koulinas, A simulated annealing hyper heuristic for con-
struction resource levelling, Constr. Manag. Econ. 28 (2) (2010) 163–175, http://dx.
solving the RCPSP, which can be combined with various heuristics or doi.org/10.1080/01446190903369907.
metaheuristics in feasible subspace to solve the problem and computa- [16] H. Li, H. Zhang, Ant colony optimization-based multi-mode scheduling under re-
tional experiments are also worthy of further attention. Moreover, an- newable and nonrenewable resource constraints, Autom. Constr. 35 (2013)
431–438, http://dx.doi.org/10.1016/j.autcon.2013.05.030.
other extension of this work is that distributed computing would be [17] K. Ziarati, R. Akbaria, V. Zeighamib, On the performance of bee algorithms for re-
induced into the framework based on the approach by using different source-constrained project scheduling problem, Appl. Soft Comput. 11 (4) (2011)
processors to do searching in various feasible subspaces. In this work, 3720–3733, http://dx.doi.org/10.1016/j.asoc.2011.02.002.
[18] M. Mobinia, Z. Mobinib, M. Rabbania, An Artificial Immune Algorithm for the project
experiments are conducted on two real construction cases which are scheduling problem under resource constraints, Appl. Soft Comput. 11 (2) (2011)
still small size and it will be meaningful to do more experiments on var- 1975–1982, http://dx.doi.org/10.1016/j.asoc.2010.06.013.
ious cases with small, medium and large size (just like the instances in [19] C. Fang, L. Wang, An effective shuffled frog-leaping algorithm for resource-
constrained project scheduling problem, Comput. Oper. Res. 39 (5) (2012)
PSPLIB [8] or cases collected from construction projects of various sizes). 890–901, http://dx.doi.org/10.1016/j.cor.2011.07.010.
The problem discussed in this work is a classic resource-constrained [20] A. Agarwal, S. Colak, S. Erenguc, A neurogenetic approach for the resource-
project scheduling problem where there is only Finish-Start (FS) rela- constrained project scheduling problem, Comput. Oper. Res. 38 (1) (2011) 44–50,
http://dx.doi.org/10.1016/j.cor.2010.01.007.
tionship among activities and the resource are available all through
[21] L. Wang, C. Fang, A hybrid estimation of distribution algorithm for solving the re-
the duration of project. However, there are more complex types of rela- source-constrained project scheduling problem, Expert Syst. Appl. 39 (3) (2012)
tionships among activities in practice, e.g. Start-Start (SS), Start-Finish 2451–2460, http://dx.doi.org/10.1016/j.eswa.2011.08.095.
(SF), Finish-Finish (FF) and FS relations with leads/lags, maximal prece- [22] Z. Liu, L. Xiao, J. Tian, An activity-list-based nested partitions algorithm for resource-
constrained project scheduling, Int. J. Prod. Res. 54 (16) (2016) 4744–4758, http://
dence relations. There are usually time windows or calendars con- dx.doi.org/10.1080/00207543.2015.1065353.
straints on activities or resources in the project. Besides minimizing [23] M. Cheng, D. Tran, Y. Wu, Using a fuzzy clustering chaotic-based differential evolu-
completion time, managers may take other objectives into consider- tion with serial method to solve resource-constrained project scheduling problems,
Autom. Constr. 37 (2014) 88–97, http://dx.doi.org/10.1016/j.autcon.2013.10.002.
ation such as maximizing net present value, minimizing cost of the pro- [24] J.F. Benders, Partition procedures for solving mixed variables programming prob-
ject. Many more cases of project scheduling problems can be found in lems, Numer. Math. 4 (1962) 238–252, http://dx.doi.org/10.1007/s10287-004-
[32,33]. The OPDA proposed in this paper cannot be utilized directly to 0020-y.
[25] H. Li, K. Womer, Optimizing the supply chain configuration for make-to-order
solve these various project scheduling problems, but it provides a way manufacturing, Eur. J. Oper. Res. 221 (1) (2012) 118–128, http://dx.doi.org/10.
to do feasible space decomposition and to embed other approaches 1016/j.ejor.2012.03.025.
into the procedure for solving these problems. So, the approaches [26] J. Tian, Z. Liu, W. Yu, An Approach with Decomposition on Time Windows for Re-
source-constrained Project Scheduling, The 26th Chinese Control and Decision Con-
based on feasible space decomposition for these practical problems
ference, Changsha China, May 31, 2014–June 2, 2014 4897–4903, http://dx.doi.org/
are also worth studying. 10.1109/CCDC.2014.6853050.
[27] Z. Liu, H. Wang, Heuristic algorithm for RCPSP with the objective of minimizing ac-
tivities' cost, J. Syst. Eng. Electron. 17 (1) (2006) 96–102, http://dx.doi.org/10.1016/
References S1004-4132(06)60018-2.
[28] W. Yu, An Approach with Nested Partition for Two Kinds of Resource-Constrained
[1] J. Blazewicz, J.K. Lenstra, A.H.G. Rinnooy Kan, Scheduling subject to resource con- Project Scheduling Problem, Huazhong University of Science and Technology,
straints: classification and complexity, Discret. Appl. Math. 5 (1) (1983) 11–24, Wuhan China, 2013 (Master Thesis, In Chinese).
http://dx.doi.org/10.1016/0166-218X(83)90012-4. [29] V. Valls, F. Ballestin, S. Quintanilla, Justification and RCPSP: a technique that pays, Eur.
[2] H. Xu, X. Wang (Eds.), Optimizations in Project Scheduling: A State-of-Art Survey, in J. Oper. Res. 165 (2) (2005) 375–386, http://dx.doi.org/10.1016/j.ejor.2004.04.008.
Optimization and Control Methods in Industrial Engineering and Construction, [30] D.H. Jun, K. El-Rayes, Multi-objective optimization of resource scheduling in con-
Springer, Netherlands, ISBN: 978-94-017-8043-8 2014, pp. 161–177, http://dx.doi. struction projects, Construction Research Congress 2009, Seattle, United States
org/10.1007/978-94-017-8044-5_10. 2009, pp. 806–815, http://dx.doi.org/10.1061/41020(339)82.
 Z. Liu et al. / Automation in Construction 75 (2017) 1–9 9

[31] P. Kenneth, R.M. Storn, J.A. Lampinen, Differential Evolution: a Practical Approach to [33] W. Herroelen, E. Demeulemeester, B. De Reyck, A classification scheme for project
Global Optimization, Springer-Verlag, Berlin Heidelberg, 2005, ISBN 978-3-540- scheduling, Project Scheduling: Recent Models, Algorithms and Applications,
31306-9http://dx.doi.org/10.1007/3-540-31306-0. Springer, US, ISBN: 978-1-4613-7529-6 1999, pp. 1–26, http://dx.doi.org/10.1007/
[32] M. Hajdu, Point-to-point versus traditional precedence relations for modeling activ- 978-1-4615-5533-9_1.
ity overlapping, Procedia Eng. 123 (2015) 208–215, http://dx.doi.org/10.1016/j.
proeng.2015.10.079.