Received 12 February 2023, accepted 16 March 2023, date of publication 23 March 2023, date of current version 29 March 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3261139

A Two-Stage Algorithm Based on 12 Priority

Rules for the Stochastic Distributed
Resource-Constrained Multi-Project Scheduling
Problem With Multi-Skilled Staff
YINING YU, ZHE XU , AND SONG ZHAO
School of Economics and Management, Beihang University, Beijing 100191, China
Corresponding author: Zhe Xu ([email protected])
This work was supported by the National Natural Science Foundation of China under Grant 72271012.

ABSTRACT In practical multi-skilled resource-constrained multi-project management, the activity duration

is often affected by some factors (e.g., rework, increased workload), leading to uncertainty. Moreover,
multiple projects are often managed under a distributed decision-making environment. To deal with uncertain
activity durations in distributed multi-project management with multi-skilled staff, this paper studies
a stochastic distributed resource-constrained multi-project scheduling problem with multi-skilled staff
(MS-SDRCMPSP). In a distributed decision-making environment, a two-stage model with local scheduling
and global coordination stages is established to describe MS-SDRCMPSP. A two-stage algorithm with
12 priority rules (TSA-12PRs) is proposed, these 12 priority rules are composed of 4 activity priority rules
and 3 resource priority rules. In the local schedule stage, 4 activity priority rules (PRs) are applied to obtain
the local schedule plan. In the global coordination phase, we develop 3 resource PRs based on variable
neighborhood search (VNS), of which VNS is used to solve the execution order of conflicting projects,
and 3 resource PRs are developed to formulate multi-skilled resource assignment strategies. Based on the
multi-skilled instances adapted from benchmark instances, we evaluate the performance of the 12 PRs on
different instances. The experiment results show that two PRs among 12 PRs perform better than other
PRs in all-size instances. Comparing the two-stage algorithm with better two PRs with other approaches in
literatures, we find that our method performs better than other approaches, especially in large-size instances.
In addition, further experiments show that our method is more conducive to shortening the CPU runtime on
distributed problems than centralized methods.

INDEX TERMS Multi-project scheduling, multi-skilled staff, uncertain activity durations, stochastic
scheduling priority rules.

I. INTRODUCTION problem (MS-RCMPSP) [2]. A classic practical scenario

Resource-constrained multi-project scheduling problem about MS-RCMPSP can be found in software develop-
(RCMPSP) contains a series of projects, each of which ment [3]. There exists a multi-project management environ-
contains activities that satisfy the resource availability and ment in MS-RCMPSP. Multiple projects compete for limited
priority relationship constraints when obtaining a multi- multi-skilled staff that primary several skills. The assignment
project scheduling plan [1]. An extension to RCMPSP is the of multi-skilled resources involves the matching relationship
multi-skilled resource-constrained multi-project scheduling of ‘‘activity-skill-resource’’; i.e., implementing activities in
the project requires the completion of multi-skilled staff
The associate editor coordinating the review of this manuscript and who primary specific skills. The case of the ‘‘activity-skill-
approving it for publication was Zhenzhou Tang . resource’’ is shown in Fig. 1.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
29554 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

FIGURE 3. The case of MS-RCMPSP.

FIGURE 1. The relationship of the ‘‘activity-skill-resource’’. an online office platform and an e-commerce project. The
two project teams share the multi-skilled staff with coding,
testing, and other skills. Each project needs limited local
resources managed by the project manager and shared global
staff managed by the coordinating manager. This problem
needs to solve the independent scheduling of each project and
the allocation of global multi-skilled staff. Therefore, when
facing such distributed problems, it is necessary to study
the distributed resource-constrained multi-project schedul-
ing problem with multi-skilled staff (MS-DRCMPSP) [5].
In MS-DRCMPSP, each PA submits the global informa-
tion to CA, and then CA assigns the multi-skilled staff
by considering the characteristics of multi-skilled staff het-
erogeneity and the matching relationship of ‘‘activity-skill-
resource’’. Further, CA solves the global multi-skilled staff
conflicts by developing an effective coordination mecha-
FIGURE 2. The function of MAS in DRCMPSP.
nism, which includes determining the start time of the activ-
ity and assigning multi-skilled staff. When multi-skilled
The projects undertaken by enterprises become larger and staff primary several skills and the skill level is differ-
more complex, so enterprises often choose to undertake mul- ent (multi-skilled heterogeneity), the actual duration of the
tiple projects at the same time. At the same time, organiza- activity will be also changed. These further increase the
tional management is also gradually distributed (distributed difficulty of solving MS-DRCMPSP. As an extension of
decision-making environment). There are multiple project DRCMPSP, MS-DRCMPSP belongs to NP-hard. The case of
decision-makers, and all projects are independent. The only the MS-DRCMPSP is shown in Fig. 3.
connection among projects is to share limited resources. In the existing literature, scholars only study deterministic
When multiple projects compete for shared resources, problems where all parameters are deterministic. However,
shared resources are prone to conflict due to resource during the practical implementation of the project, there
limitations and project independence, which forms a dis- are often uncertain situations, the most common of which
tributed resource-constrained multi-project scheduling prob- are uncertain activity durations. When uncertainty occurs,
lem (DRCMPSP) [4]. The solution to DRCMPSP involves activities are interrupted, resulting in delays and losses in
the independent scheduling of multiple single-project exe- practical project management. For example, uncertain activ-
cuted simultaneously and the reasonable allocation of shared ity durations make the baseline scheduling plan and the
resources among projects. Generally, a multi-agent system resource assignment strategy impossible. Stochastic schedul-
(MAS) is used for solving the DRCMPSP, which includes ing is often used to solve the uncertainty in project schedul-
multiple project agents (PA) and one coordinating agent ing. In stochastic scheduling, a scheduling policy is obtained
(CA). Each PA has local resources to pursue its local interest instead of a specific scheduling plan [6]. In addition, stochas-
objective and does not disclose local information to each tic scheduling is a common method to solve the problem
other. CA coordinates shared resources (global resources) for of uncertain activity duration, in which the activity duration
each PA according to the global objective. The function of usually follows a known distribution. There is no litera-
MAS in DRCMPSP is shown in Fig. 2. ture that studies the multi-skilled resource-constrained multi-
In the distributed decision-making environment, it is more project scheduling problem with uncertain activity durations
and more common to share resources with multi-skilled by stochastic scheduling.
staff based on multiple projects. For instance, a software Based on the MS-DRCMPSP and uncertain activ-
development company has two project teams that undertake ity durations, we proposed the stochastic distributed

VOLUME 11, 2023 29555

 Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

resource-constrained multi-project scheduling problem with multi-objective multi-skilled resource-constrained project

multi-skilled staff (MS-SDRCMPSP). Each project is sched- scheduling problems [15], [16].
uled by its PA to optimize the expected project makespan There are few related types of research on the multi-skilled
(local objective) in this problem. CA assigns multi-skilled resource-constrained multi-project scheduling problem. The
staff for each project depending on minimizing the expected leading research is a centralized scheduling method, that is,
total tardiness costs (global objective). Further, heteroge- there is only one decision-maker. Heimerl and Kolisch [17]
neous characteristics of multi-skilled staff and uncertain studied the MS-RCMPSP in the centralized decision-making
activity durations significantly increase the difficulty of environment of IT enterprises. Considering multi-skilled
modeling and solving MS-SDRCMPSP. We attempt to outsourced resources, the meta-heuristic algorithm was used
develop a two-stage algorithm with 12 priority rules (TSA- to solve MS-RCMPSP [17]. Walter and Zimmermann [18]
12PRs) and variable neighborhood search algorithm (VNS) studied MS-RCMPSP from the perspective of the team size.
by multi-skilled staff heterogeneous characteristics for the Chen et al. [19] studied the learning effect of MS-RCMPSP,
MS-SDRCMPSP, especially designing some effective multi- established a multi-objective constrained model of
skilled staff assignment strategies. MS-RCMPSP, and used a non-dominated genetic algorithm
The main contributions of this article can be described as (NSGAII) to solve the problem; Some studies (Wu and
follows: Sun [20]; Gutjahr et al. [21]) also studied the centralized
Firstly, multi-skilled staff are considered shared resources MS-RCMPSP [20], [21], [22]. In summary, a centralized
in a distributed multi-project environment with uncertain approach with only one decision-maker is unsuitable for mul-
activity durations. tiple decision-makers. It lacks a certain degree of flexibility
Secondly, a two-stage algorithm with 12 priority rules and can not only partially satisfy all projects.
(TSA-12PRs) is proposed to solve this problem, among Conversely, the distributed approach is more suitable for
which 4 activity PRs for solving the local schedule stage and management environments with multiple decision-makers.
variable neighborhood search (VNS) with 3 resource PRs for So far, only scholar Yu et al. [5]studied the multi-skilled
solving the global coordination stage. distributed resource-constrained multi-project scheduling
Thirdly, the TSA-12PRs are evaluated on different size problem, established a two-stage decision-making model,
problems, and the two best PRs are selected on all and provided practical management suggestions. Therefore,
size instances. Further experiments compared a distributed there are more possibilities for the MS-DRCMPSP, especially
approach (SGNM) and a centralized method (BRKGA). under uncertain situations.
The rest of this article is structured as follows. A sum-
mary of existing publications in this field is presented in
Section II. Section III is devoted to the problem description
and mathematical formulations. In Section IV, the 12PRs and B. DISTRIBUTED RESOURCE CONSTRAINED
the proposed algorithms are used in this paper to solve the MULTI-PROJECT SCHEDULING PROBLEM
MS-SDRCMPSP. A comprehensive experimental analysis is Generally, the distributed scheduling problem is solved by
provided in Section V. Finally, Section VI summarizes the multi-agent systems (MAS). In the MAS, there are mul-
article. tiple Project Agents (PAs) as the project managers and
a Coordinating Agent (CA) as the coordinating manager
[23], [24], [25], [26]. Homberger solved DRCMPSP through
II. LITERATURE REVIEW
an electronic iterative negotiation mechanism based on the
In this section, a brief review of the existing literature rel-
mediation protocol [27]; Subsequently, Homberger contin-
evant to this article is conducted. It includes the distributed
ued to expand the DRCMPSP model, proposed the restart
resource-constrained multi-project scheduling problem, the
evolutionary algorithm, and solved the large-size example
multi-skilled resources-constrained multi-project scheduling
problem [28]. Zheng et al. [29]solved the DRCMPSP by a
problem, and the stochastic project scheduling problem.
critical-chain method. Wang et al. [30] researched a resource-
constrained project scheduling problem with a fractional
A. MULTI-SKILLED RESOURCE-CONSTRAINED shared resource by a column-generation-based algorithm.
MULTI-PROJECT SCHEDULING PROBLEM Rostami et al. [31] designed a lagrangian relaxation algorithm
Hegazy et al. were the first to propose the multi-skilled for DRCMPSP and verified the effectiveness of the method.
project scheduling problem [7]. The Bellenguez team consid- Recently, the distributed approach has also been applied in
ered the characterize of multi-skilled resources and defined different areas. For example, Zhao and Xu [32] discussed the
the MS-RCPSP in the current paper [8], [9]. Now the distributed multi-project scheduling problem with the transfer
researches about multi-skilled staff mainly concentrate on time. Kosztyan [33]studied the flexible multi-level project
a single-project environment. They are mainly divided scheduling problem by a matrix-based multi-level multi-
into multi-skilled resource-heterogeneous project schedul- mode project scheduling algorithm. Other scholars (Babaee
ing problems [10], [11], multi-mode multi-skilled resource- Tirkolaee et al. [34]; Tirkolaee et al. [35]; Mahdavi et al. [36])
constrained project scheduling problems [12], [13], [14], also studied the multi-trip scheduling problem in different

29556 VOLUME 11, 2023

Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

of activities noted as aij ∈ Vi (Vi = ai0 , . . . , aiJi ), where


fields, such as waste management, heat storage systems, and
two-echelon supply chain management. activity ai0 and activity aiJi represent two dummy activities at
the beginning and the end. Eij denotes the predecessor activity
C. STOCHASTIC PROJECT SCHEDULING PROBLEM set for aij , that is, aij can be started only after all activities
In stochastic scheduling, project activity duration is often in Eij are completed. A dummy activity has a duration of
known and assumed to follow a distribution; the objective is zero and no resource usage under any probability distribution;
to minimize the expected makespan of the project. Given the oppositely, the duration of a non-dummy activity is a random
randomness of the activity duration, the solution of SRCPSP variable and obeys a known probability distribution. Non-
is a scheduling policy rather than a deterministic schedule. dummy activities also require several types of local resources
For solving the SRCPSP, there are two main methods. One of and at most one type of global resource. The local resource
them is a meta-heuristic to solve the SRCPSP. For example, k(k = 1, 2, . . . , Ki ) stands for the ordinary staff who has only
genetic algorithms (Chen et al. [37]; Zaman et al. [38]), tabu one skill. The global resource g(g = 1, 2, . . . , G) refers to the
search (Servranckx and Vanhoucke [39]). The other is the multi-skilled staff who primary several skills.
priority rules-based heuristic [37], [38], [39]. When multiple PAs conflict with limited global resources,
In contrast to meta-heuristics, the priority rules-based some activities can not get global resources in time and will
heuristic relies on creating a priority rule instead of multiple cause their projects to be postponed. Each project has a unit
iterations. It includes two types of elementary policies, which delay cost tci and a completion time Bi (Bi = t · xiJi t ).
can be subdivided into static policies and dynamic policies. CA designs a coordination mechanism to assign global
The static policy means that the strategy is given before the staff to each PA. There is no connection among projects
project is executed. During the execution process, the strategy except for sharing limited global staff. During all project
will not change. Some static policies are often used in the execution, one staff can only use one skill to perform one
literature, such as resource-based policy class, activity-based activity simultaneously. At any moment, the skill demands
policy class, earliest-start policy class, preselective policy for all activities being performed in the multi-project cannot
class, preprocessor policy class, and generalized preprocessor exceed the total available global resources in providing skills
policy class [40], [41], [42]. The dynamic policy seeks the at current moment. Similarly, the local resources also have a
best policy step by step in a dynamic way. Its solutions must limited quantity Rik .
be constantly updated according to new information, such as Each PA performs local scheduling of one project to
the activity start time. Thus, dynamic policies usually need minimize the expected makespan of this project, which is
more CPU runtime. described as the following (1) and (2).
Scholars Thomas F, et al. studied uncertain problems
by stochastic scheduling under the centralized environ- min E(fp ) (1)
ment. The research expanded the deterministic model and T
X
solved the uncertain MS-RCMPSP through the Frank Wolfe fp = t · xiJi t . (2)
algorithm [6]. However, under the distributed decision- t=0
making environment, there is no literature that studies the
multi-skilled resource-constrained multi-project scheduling As a global decision-maker, the objective of CA is to find
problem with uncertain activity durations by stochastic a feasible schedule and global resources assignment plan
scheduling. with the goal of minimizing the expected total tardiness costs
Considering the complexity of the distributed environment (TTC) denoted by (1) and (4).
and multi-skilled heterogeneity, a dynamic strategy will sig-
nificantly increase the CPU runtime of the stochastic schedul- min E(TTC) (3)
ing. It is more realistic that a decision-maker needs to make M
X
decisions as quickly as possible with a static strategy. There- TTC = tci · (Bi − wi − CPDi ) (4)
fore, it is more important to study the stochastic distributed i=1
multi-skilled resource-constrained multi-project scheduling
problem by the static policy. This objective refers to the sum of the delay costs caused
by the actual makespan (Bi − wi ) of each project exceeding
III. PROBLEM MODELING the critical path duration (CPDi ).
In this section, stochastic distributed resource-constrained The following particular assumptions are considered in this
multi-project scheduling problem with multi-skilled staff is article:
introduced in detail, and a two-stage model is established. • An activity can only be executed after all the resources
required by the activity have arrived.
A. PROBLEM DESCRIPTION • All multi-skilled staff have the characteristics of het-
In this problem, M projects are scheduled simultaneously. erogeneity. In other words, each resource may have a
Each project t ∈ 1, 2, . . . , M has an arrival date wi ≥ 0 denot- different level for each skill s(s = 1, 2, . . . , S). The
ing its earliest possible start time. In project i, there are a set higher the skill level, the shorter the actual duration.

VOLUME 11, 2023 29557

 Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

B. NOTATION
1) PARAMETERS
lgs - the level of skill s primaried by the resource g.
rijs - the skill s demand of the activity aij .
vsij - a Boolean variable indicating when skill s is required
by activity aij that equals 1 and equals 0 otherwise.
d̄ij - the planned duration of activity aij , that is, the staff
with lgs = 1 performs the activity, d̄ij follows a random
variable with a known probability distribution.
dij - the actual duration of activity aij , the actual duration
is effected by the level of skill, see (5).

2) DECISION VARIABLE
xiJi t : 1, if the activity aij starts at time t; 0, otherwise;
gs
yijt : 1, if the resource g with the skill s perform activity aij
at time t; 0, otherwise;
 
gs
X
dij = rijs · d̄ij / lgs · yijt · vsij  (5)
g=1

IV. TWO-STAGE ALGORITHM BASED ON

12 PRIORITY RULES
In order to obtain a complete multi-project scheduling plan
in the distributed environment, a two-stage coordination pro-
cess is designed at each conflicting time. Here, conflicting
time refers to when more than two projects require global
resources, and these two projects are also called conflicting
projects. MAS comprises PA as a project manager and CA as
a coordinating manager.
In stage one, the local scheduling is a statistic scheduling
approach with the goal of the expected project makespan
by each PA. Each PA solves the local scheduling based on
activity priority rules (in Section IV-A1). In stage two, the FIGURE 4. An overview of the global coordination press.
global coordination process not only needs to determine the
execution order of conflicting projects (in Section IV-B) at
each conflicting time, but also needs to assign multi-skilled
A. PRIORITY RULE BASE HEURISTIC POLICY
resources (in Section IV-A2) reasonably. CA solves the global
The local scheduling is managed by each PA, which is to
coordination process to minimize the expected total delay
minimize the expected makespan of this project. In the local
costs of multiple projects. This paper combines the charac-
scheduling, each PA obtains the local scheduling by activity
teristics of multi-skilled resources and designes a heuristic
priority rules. At the conflicting time, CA assigns the global
strategy based on 12 priority rules.
resources for a selected project by resource priority rules.
Fig. 4 presents an overview of the global coordination
press.
Step1:From the time 0, after all PAs finish the local 1) ACTIVITY PRIORITY RULES
scheduling by activity PRs (in Section IV-A1), they submit This section exhibits the 4 activity priority rules for the local
the global resources demand information to the CA. SRCPSP. These rules perform better than other static priority
Step2:At this time, if there is only one project requir- rule heuristic policies for SRCPSP [43], [44], [45]. Then
ing global resources, CA assigns global resources for this we test the suitability of these 4 activity priority rules on
activity according to the resource PRs (in Section IV-A2); MS-SDRCMPSP.
Suppose there are more than two projects requiring This section exhibits the 4 activity priority rules
global resources, CA ranks conflicting projects by Vari- (in Table 1) for the local SRCPSP. These rules perform better
able Neighborhood Search (VNS, in Section IV-B) and than other static priority rule heuristic policies for SRCPSP.
coordinates global resource assignments by resource PRs Then we will test the suitability of these 4 activity priority
(in Section IV-B). rules on MS-SDRCMPSP.

29558 VOLUME 11, 2023

Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

TABLE 1. Four activity priority rules for the local SRCPSP. B. VARIABLE NEIGHBORHOOD SEARCH DESIGN
To minimize the global objective, CA determines an execu-
tion order of all conflicting projects at each conflicting time.
Firstly, an initial order L2 is selected from highest to lowest
unit cost; projects with higher unit costs are prioritized for
execution, which helps narrow the search to find a reason-
able solution. Then CA assigns the corresponding global
resources for projects in L2 according to the resource PRs
(in Section IV-A2), and calculates the total delay cost (an
2) RESOURCE PRIORITY RULES
initial solution Sinit ) of multi-projects in L2. Finally, starting
We choose three resource priority rules with good per- from position 1 (the first project at the current order ipro ),
formance from the existing literature (Snauwaert and exchanging the order with the following projects as a new
Vanhoucke [48]); that is, Highest Average Level (HAL), order. CA assigns the global resources based on the resource
Lowest Average Level (LAL), Lowest Number & Highest PRs (in Section IV-A2) and calculates the new solution Snew
Level (LN&HL). Then we design a new resource priority in the current order. If the new solution Snew is larger than
rule called Highest Level &Lowest Number (HL&LN). This the initial solution Sinit , the order will be updated again;
section explains that 4 resource priority rules are used at the otherwise, replace the initial solution and return to position 1.
conflicting time. Some of them are based on the skill-level or We continue to perform the variable neighborhood search
the skill-number. until the best solution Alldelay is obtained at the current
moment.
a: HIGHEST AVERAGE LEVEL (HAL)
This rule uses the skill level to assign multi-skilled staff.
Algorithm 1 The Postcode of Variable Neighborhood
Resources are ranked based on the average level of their mas-
Search (VNS)
tered skills, then activities with HAL are selected first, which
Require:
indicates that the most efficient resources will be prioritized.
LP: the set of all conflicting projects at the current con-
flicting time;
b: LOWEST AVERAGE LEVEL (LAL)
L1: the set of projects at least one activities meet the skill
This priority rule is the opposite of the previous rule.
availability for each project in LP.
Resources will be ranked from the lowest to the highest
1: if L1 ̸ = then
average depth, indicating that the least efficient resources will
2: Obtain initial order L2; %selected in order by highest
be prioritized. Similar to Highest Breadth First, this rule adds
unit deferred cost first
diversity to the set of priority rules.
3: Calculate Sinit ; %the initial solution.
4: Record ipro = 1; Count = 0.
c: HIGHEST LEVEL &LOWEST NUMBER (HL&LN)
5: if length(L1(1)) = 1 then
The rule considers both skill-level and skill-number of
6: L1delay = Sinit ;
resources. It gives priority to resources with the highest skill-
7: else
level and selects resources with the lowest skill-number as
8: %change the order of adjacent projects
tie-breakers. The objective of this rule is to minimize the
9: while ipro < allocatepro do
makespan of the activity by utilizing resources with the high-
10: Count = Count + 1; 0
est skill-level.
11: a = L2(1, ipro );
12: L2(1, ipro ) = L2(1, ipro + 1);
d: LOWEST NUMBER & HIGHEST LEVEL (LN&HL)
13: L2(1, ipro + 1) = a;
This rule also considers the skill-level and the skill-number. 14: Calculate Snew ; % the new solution
In this case, the lowest skill-number resources are prioritized, 15: if Snew ≥ Sinit then
and the highest skill-level resources are used as tie-breakers. 16: ipro = ipro + 1;
This rule aims to keep the most skill-number resources avail- 17: else
able while assigning the most efficient resources. 18: ipro = 1;
In the local scheduling, 4 activity priority rules are cho- 19: Sinit = Snew ;
sen. For each activity priority rule, there are 4 resource 20: end if
assignment rules. By Combining activity priority rules 21: L1delay = Sinit ;
and resource assignment rules, we obtain 16 priority 22: end while
rules based on heuristic strategies. They are LST-HAL, 23: end if
LST-LAL, LST-HL&LN, LST-LN&HL; LFT-HAL, LFT- 24: end if
LAL, LFT-HL&LN, LFT-LN&HL; SLST-HAL, SLST-LAL, Ensure:
SLST-HL&LN, SLST-LN&HL; SLFT-HAL, SLFT-LAL, Alldelay = L1delay .
SLFT-HL&LN, SLFT-LN&HL.

VOLUME 11, 2023 29559

 Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

TABLE 2. Five distributions for activity durations. Further multi-skilled information required to generate
instances are:
• Each problem instance is provided with at most one type
of global resource and three types of local resources.
• The types of skills with project numbers of 2, 5 are set
to 3, 5, respectively.
• The value of the parameters rijs are generated in the range
of [1] and [3] uniformly.
• Each staff primaries the types of skills are generated in
V. COMPUTATIONAL STUDY
the range of [2] and [3] uniformly.
A series of experiments are carried out to conduct the com-
• The level of skill is randomly generated in the range of
putational study. All designed test problems are solved in
0.6, 0.8, and 1, respectively. Such skill levels are valid
Matlab R2018b, with a core i7 CPU and 16 GB memory.
and guarantee the speed at which the resource executes
This section contains four parts: Section V-A introduce the
the activity [48], so staff with lower skill levels are not
problem instances; Section V-B analyzes the impact of sim-
considered.
ulated times on the priority rules; Section V-C verifies the
performance of the TSA-12 PRs on different distributions. TABLE 3. Problem instances of MS-SDRCMPSP.
Section V-D analysis the performance of the TSA-12 PRs
compared with other algorithms (including the distributed
method and centralized method).
In order to be able to compare the PR results with pre-
vious works on the SRCPSP, we first follow the existing
literature in line with their probability distribution types and
parameters [41], [46]. Assuming that the activity duration is B. THE INFLUENCE OF SIMULATION TIMES AND
a random variable, the mean value of the duration is equal SELECTION OF PRIORITY RULES
to the deterministic duration of the MPSPLIB data set. The
MP30_2 and MP90_2 problem sets are used for pre-
five distributions and their variances are shown in Table 2.
experiment when selecting simulated times. In order to test
Variances represent different degrees of uncertainty: the vari-
the significance of the results among the three simulated
ance of distribution U1 and B1 is the smallest, the variance
times, we first perform the paired samples Wilcoxon signed-
of U2 and B2 is the middle, and the variance of Exp is the
rank test for 12 PRs of each problem set. The significance
largest. Additionally, in each beta distribution, the shape of
level is set to 5%. Table 4 and 5 show the statistical test results
the distribution is determined by two main shape parameters,
on MP30_2 and MP90_2, respectively.
α and β. For subsequent experiments, we used the two most
commonly used parameters as reported in the literature, and TABLE 4. Wilcoxon signed rank test results for 12 PRs on MP30_2.
the specific values of the shape parameters are shown in
Table 2.

A. PROBLEM INSTANCES
Yu et al. [5] provided problem instances for the
MS-DRCMPSP under certainty. In this paper, we intro-
duce uncertainty in activity duration to the MS-DRCMPSP,
which increases the problem complexity in terms of solu-
tion times and solving difficulty for the MS-SDRCMPSP.
Wang et al. [47] demonstrated that the J30 dataset is already
representative in the multi-skilled project scheduling problem
under uncertainty. Therefore, it is reasonable that we selected
20 instances from the literature [5], including the J30 and J90
datasets generated from the MPSPLIB.
The problem instances are shown in Table 3. These The results indicate significant differences in the quality
instances are classified into 4 subsets. Each problem subset of solutions obtained for 12 PRs with 10 vs 30 simulated
is named as MPJi _m (MP subset), where the number of times and 10 vs 50 simulated trajectories, but no signifi-
activities Ji per project m is 30, 90, and the number of projects cant differences between 30 and 50 simulated trajectories.
is 2, 5, NOI denotes the number of instances primaried by Comparing the solutions, we find the results of 30 and
each problem subset. According to the problem size, MP90_5 50 simulations are statistically better than that of 10 simula-
is called a large-size instance, and the other three problem sets tions. To balance solution quality and computational runtime,
are called small-size instances. we set the number of simulated times to 30 in this paper.

29560 VOLUME 11, 2023

Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

TABLE 6. Average E (TTC ) of MP30_2 obtained by 12 priority rules.

TABLE 7. Average E (TTC ) of MP90_2 obtained by 12 PRs.

FIGURE 5. The deviation of the combination of the four worst heuristic

rules.

TABLE 5. Wilcoxon signed rank test results for MP90_2.

TABLE 8. Average E (TTC ) of MP30_5 obtained by 12 PRs.

To choose better combinations of priority rules, we test

the results of different priority rules with MP30_2 and TABLE 9. Average E (TTC ) of MP90_5 obtained by 12 PRs.
MP90_2 under the U1 distribution. We sort 16 rule combi-
nations and find that the last four combinations ranked had
poor results: LST-LAL, LFT-LAL, SLST-LAL, and SLFT-
LAL. After analyzing the deviation between the four poorly
ranked combinations and the 12th combination, we found that
the deviation was greater than 5%. As a result, we excluded
the four underperforming combinations and selected the top
12 combinations for further analysis. Fig. 5 shows the devia-
tion of the combination of the last four worst heuristic rules
on MP30_2 and MP90_2.

C. PERFORMANCE OF THE 12 PRIORITY RULES UNDER In Table 6, for MP30_2 problem set, LFT-HL&LN has the
DIFFERENT DISTRIBUTIONS best performance under B1, B2, and EXP, and SLFT-HL&LN
Table 6, 7, 8 and 9 show the average tardiness cost of 30 runs and LST-HL&LN perform best under U1 and U2, respec-
for MP30_2,MP90_2, MP30_5 and MP90_5 under different tively. In Table 7, for MP90_2 problem set, LFT-HL&LN has
distributions. The bold values in each table stand for the the best performance under U1, U2, and EXP, and SLFT-
minimum average E(TTC) under the same activity priority HL&LN performs best under B1 and B2. In Table 8, for
rule. The underlined values in each column are the best results MP30_5 problem set, SLST-HL&LN and SLFT-HL&LN are
obtained by all precedence rules under the same distribution. the best compared with other priority rules under U2, B2 and
According to the bold and underlined values, the combination U1, B1, respectively. However, LFT-HL&LN has the best
of several priority rules that perform better in each problem performance under EXP. While in Table 9, for MP90_5 prob-
set is shown in bold. lem set, LFT-HL&LN performs better than other priority

VOLUME 11, 2023 29561

 Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

rules under U2 and B2. However, SLFT-HL&LN has the best TABLE 10. Comparisons of E (TTC ) with the distributed approach of
MP30_2.
performance under U1, B1 and EXP.
Additionally, in all Tables, HL&LN is beneficial in reduc-
ing the objective under all distributions. Since resources with
the highest skill-level and the lowest skill-number can per-
form activities as soon as possible, effectively reducing actual
activity durations and the single project makespan. From the
global perspective, when the resource priority rule (HL&LN)
is determined, LFT-HL&LN and SLFT-HL&LN have the
best performance for all-size instances, especially for large-
size instances. However, LST-HL&LN and SLST-HL&LN
only perform well for small-size instances. Therefore, as a
global decision-maker, it is better to choose the LFT or SLFT
as the activity priority rules and the HL&LN as the resource TABLE 11. Comparisons of E (TTC ) with the distributed approach of
MP90_2.
priority rules, respectively.

D. PERFORMANCE ON THE TSA-12 PRs COMPARED

WITH OTHER ALGORITHMS
In order to verify the performance of TSA-12 PRs, our
comparative experiments mainly focus on two aspects:
(1) comparing the distributed methods on 12 heuristic priority
rules; (2) comparing the distributed and centralized methods
on different size instances.
The first comparative experiment aims to evaluate the
effectiveness of the VNS compared to other distributed meth-
ods for 12 heuristic priority rules, specifically the sequential
game-based negotiation mechanism (SGNM). TABLE 12. Comparisons of E (TTC ) with the distributed approach of
SGNM is a traversal search method where all conflict- MP30_5.

ing projects are arranged, and the best project sequence is

selected as the execution order at the conflicting time. Li and
Xu [49] introduced that SGNM had an excellent performance
in solving general distributed scheduling problems. There
is no literature on MS-SDRCMPSP with uncertain activity
duration, so we applied SGNM to MS-SDRCMPSP with
certain representativeness. The detailed SGNM is described
as follows:
SGNM: A sequential game-based negotiation mechanism
based on the distributed multi-agent system, introduced
in [49]. The sequential game is defined as a game consisting
of finite and at least two players where each player takes TABLE 13. Comparisons of E (TTC ) with the distributed approach of
actions at different times or in turn. CA as a coordinator orga- MP90_5.

nizes sequential games for PAs. After several sequential game

negotiations, CA determines the best subgame perfect Nash
equilibrium. Then multiple PAs resolve their local schedules
with the allocated global resources from CA independently.
Table 10, 11, 12, and 13 denote the comparison results
between our method and SGNM on 12 priority rules for
different instances. Since SGNM is also combined with a
two-stage distributed approach, its role as VNS is to decide
the project execution order in the global coordination press.
In order to test the impact of VNS and SGNM on results,
we ensure that each activity PR and resource PR as a com-
bination and 12 PRs are tested. The gap means the improved
percentage of VNS than SGNM in each priority rule. ‘‘+’’ Table 10 and 11 show no apparent difference between
denotes that VNS is better than SGNM. ‘‘−’’ means that the two projects for our method and SGNM. When there
SGNM is better than VNS. are only two projects, there are at most two sequences

29562 VOLUME 11, 2023

Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

TABLE 14. Friedman test under pm grouping.

TABLE 15. Comparisons of E (TTC ) for best priority rules among different
approaches.

FIGURE 6. The improved percentage of the solutions on different

instances.

at each moment, such as project 1-project 2 or project

2-project 1. If the sequence of random selection by SGNM
is the same as the sequence finally selected by VNS at the
decision time, the obtained objective refers to the same.
If the sequence of random selection by SGNM is opposite
to the VNS, the results of the two methods are different at
this moment. We also find that although the results from
VNS and SGNM are not the same, the deviation is controlled
within 5%, which is possible. Based on the comparison in
Fig. 6, it appears that the VNS method performs better than
SGNM, with an advantage of 8% on MP30_5and over 12%
on MP90_5. After analyzing the priority rules of VNS in
each table, it was found that LFT-HL&LN performed best in
MP30_2 and MP90_2, while SLFT-HL&LN performed best
in MP30_5 and MP90_5.
The second comparative experiment aims to compare
the performance of the distributed methods with a cen-
tralized method, which in this case is the BRKGA algo-
rithm. The BRKGA algorithm was originally proposed by
Almeida et al. [50] for the project scheduling problem with
flexible resources. Therefore, it is reasonable to test the
BRKGA algorithm on our instances. The detailed BRKGA
is as follows:
BRKGA: This centralized approach is a based random-
key genetic algorithm (BRKGA) for the project scheduling
problem with flexible resources [50]. In BRKGA, there are
three key parameters: the population size (pop), the crossover
rate(rc) and the mutation rate(rm). MP90_2 is a medium-
sized example in this paper, so we take it as an example
to set up a pre-experiment and select appropriate param-
eters for BRKGA. According to the pop = 60/100, rc =
0.7/0.8/0.9, rm = 0.05/0.1/0.15, different objetives are
FIGURE 7. Average CPU time obtained by three approaches.
obtained. Firstly, the Kolmogorov-Smirnov(K-S) test is used
to verify the sample data P = 0.000(< 0.05), and that does
not obey the normal distributionso this article uses a non- The bolded part of Table 14 indicates that there is a
parametric test. Then, the Wilcoxon signed rank-sum test is signifificant impact when pop = 60, rm = 0.05 and
applied since there are two kinds of population sizes. The pop = 100, rm = 0.1 are at the 5% signifificance level.
result is P = 0.000(<0.05). Therefore, at a significance level Therefore, this article compares the target values under the
of 5%, the value of population size has a significant impact. three crossover probabilities and selects the combination with
Finally, this article applies the Friedman test to multiple the smallest target value as pop = 100, rc = 0.8, and
related samples of mutation probability between the different rm = 0.1. Combined with the CPU runtime, the maximum
population sizes, and the results are shown in Table 14. number of iterations as Gen = 100.

VOLUME 11, 2023 29563

 Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

SGNM and BRKGA stand for the distributed method and with multi-skilled staff through rescheduling in an uncertain
the centralized algorithm, respectively. Table 15 shows the environment.
compared results among three algorithms on best priority
rules for different problem subsets. ACKNOWLEDGMENT
Table 15 and Fig 7 (a) shows that TSA-12PRs has The authors would like to thank their team for excellent tech-
greater difference than other two approaches on MP90_5. nical support and assisting in preparation of this manuscript.
And we also find that BRKGA performs well on MP30_2, REFERENCES
MP90_2 and MP30_5, but the advantages are not obvi- [1] A. Lova and P. Tormos, ‘‘Analysis of scheduling schemes and heuristic
ous. Therefore, TSA-12PRs is more suitable for large-size rules performance in resource-constrained multiproject scheduling,’’ Ann.
instances, while BRKGA is more suitable for small-size Oper. Res., vol. 102, no. 1, pp. 263–286, 2001.
[2] M. Hendriks, B. Voeten, and L. Kroep, ‘‘Human resource allocation in a
instances. Then SGNM performs well on MP30_2 and multi-project R&D environment: Resource capacity allocation and project
MP90_2, but SGNM has only a slight advantage over portfolio planning in practice,’’ Int. J. Project Manag., vol. 17, no. 3,
TSA-12PRs and only performs better on the instances with pp. 181–188, 1999.
[3] X. Shen, L. L. Minku, R. Bahsoon, and X. Yao, ‘‘Dynamic software project
2 projects. We can see that the number of projects is more scheduling through a proactive-rescheduling method,’’ IEEE Trans. Softw.
important to SGNM than the problem size. Fig.7(b) shows Eng., vol. 42, no. 7, pp. 658–686, Jul. 2016.
that BRKGA needs more CPU runtime, which is not appli- [4] G. Confessore, S. Giordani, and S. Rismondo, ‘‘A market-based multi-
agent system model for decentralized multi-project scheduling,’’ Ann.
cable to the actual situation. When there are more projects, Oper. Res., vol. 150, no. 1, pp. 115–135, Feb. 2007.
SGNM has no advantage, such as five projects. [5] Y. Yu, Z. Xu, D. Liu, and S. Zhao, ‘‘A two-stage approach with softmax
It appears that the distributed method, particularly TSA- scoring mechanism for a multi-project scheduling problem sharing multi-
skilled staff,’’ Exp. Syst. Appl., vol. 203, Oct. 2022, Art. no. 117385.
12PRs, outperforms the centralized method for large-size
[6] T. Felberbauer, W. J. Gutjahr, and K. F. Doerner, ‘‘Stochastic project
instances, while the centralized method is more suitable for management: Multiple projects with multi-skilled human resources,’’
small-size instances. Additionally, the proposed TSA-12PRs J. Scheduling, vol. 22, no. 3, pp. 271–288, Jun. 2019.
takes less time than the centralized method, making it a [7] T. Hegazy, A. K. Shabeeb, E. Elbeltagi, and T. Cheema, ‘‘Algorithm for
scheduling with multiskilled constrained resources,’’ J. Construct. Eng.
promising approach for addressing the MS-SDRCMPSP with Manag., vol. 126, no. 6, pp. 414–421, Dec. 2000.
uncertain activity duration. [8] O. Bellenguez and E. Neron, ‘‘Lower bounds for the multi-skill project
scheduling problem with hierarchical levels of skills,’’ in Proc. Int. Conf.
Pract. Theory Automated Timetabling. Cham, Switzerland: Springer, 2004,
pp. 229–243.
VI. CONCLUSION [9] O. Bellenguez-Morineau and E. Néron, ‘‘A branch-and-bound method
This article investigates the stochastic distributed resource- for solving multi-skill project scheduling problem,’’ RAIRO-Oper. Res.,
vol. 41, no. 2, pp. 155–170, Apr. 2007.
constrained multi-project scheduling problem with the multi- [10] M. Firat and C. A. J. Hurkens, ‘‘An improved MIP-based approach for a
skilled staff. A two-stage approach with 12 priority rules is multi-skill workforce scheduling problem,’’ J. Scheduling, vol. 15, no. 3,
developed for this problem. In the local scheduling stage, pp. 363–380, Jun. 2012.
[11] P. B. Myszkowski, M. E. Skowroński, Ł. P. Olech, and P. O. Oślizlo,
4 activity priority rules are applied to optimize the expected ‘‘Hybrid ant colony optimization in solving multi-skill resource-
project makespan; in the global decision stage, 3 resource constrained project scheduling problem,’’ Soft Comput., vol. 19, no. 12,
priority rules are designed to achieve the expected total pp. 3599–3619, 2015.
[12] Y. Kadrou and N. Najid, ‘‘Tabu search for the multi-mode resource con-
tardiness cost. In order to confirm the performance of our starined project scheduling problem whith resource flexibility,’’ in Proc.
approach, different size multi-skilled instances are solved. 19th Int. Conf. Prod. Res., 2007, pp. 1–7.
The experimental results show that the two best PRs, includ- [13] H. Kazemipoor, R. Tavakkoli-Moghaddam, and P. Shahnazari-Shahrezaei,
‘‘Solving a mixed-integer linear programming model for a multi-skilled
ing LFT-HL&LN and SLFT-HL&LN, perform better than project scheduling problem by simulated annealing,’’ Manag. Sci. Lett.,
other PRs on all-size instances. When the two best PRs with vol. 2, no. 2, pp. 681–688, Apr. 2012.
the two-stage algorithm are selected on all-size instances, our [14] H. Kazemipoor, P. Shahnazari-Shahrezaei, and R. Tavakkoli-Moghaddam,
‘‘Solving a novel multi-skilled project scheduling model by scatter
approach performs better than others(SGNM), especially for search,’’ South Afr. J. Ind. Eng., vol. 24, no. 1, pp. 121–135, 2013.
large-size instances. Additionally, further experiments show [15] L. L. Minku and X. Yao, ‘‘Software effort estimation as a multiobjective
that the centralized approach (BRKGA) is suitable for small- learning problem,’’ ACM Trans. Softw. Eng. Methodol., vol. 22, no. 4,
pp. 1–32, Oct. 2013.
size instances, but the CPU runtime of our method is within
[16] H. Maghsoudlou, B. Afshar-Nadjafi, and S. T. A. Niaki, ‘‘Multi-skilled
the controllable range. Therefore, managers can consider our project scheduling with level-dependent rework risk; Three multi-objective
method if there are more projects in practical project manage- mechanisms based on cuckoo search,’’ Appl. Soft Comput., vol. 54,
ment and the processing time is limited. pp. 46–61, May 2017.
[17] C. Heimerl and R. Kolisch, ‘‘Work assignment to and qualification of
Since our method is only applicable to large instances, multi-skilled human resources under knowledge depreciation and company
we plan to design a more effective distributed coordination skill level targets,’’ Int. J. Prod. Res., vol. 48, no. 13, pp. 3759–3781,
method that can be applied to instances of all sizes in the Jul. 2010.
[18] M. Walter and J. Zimmermann, ‘‘Minimizing average project team size
future. Additionally, since it is inevitable that multi-skilled given multi-skilled workers with heterogeneous skill levels,’’ Comput.
staff members will leave and return during a project, man- Oper. Res., vol. 70, pp. 163–179, Jun. 2016.
agers may need to reallocate these staff members to work [19] R. Chen, C. Liang, D. Gu, and J. Y.-T. Leung, ‘‘A multi-objective model
for multi-project scheduling and multi-skilled staff assignment for IT
on multiple projects. Therefore, future research aims to solve product development considering competency evolution,’’ Int. J. Prod.
the distributed resource-constrained multi-project scheduling Res., vol. 55, no. 21, pp. 6207–6234, Nov. 2017.

29564 VOLUME 11, 2023

Y. Yu et al.: Two-Stage Algorithm Based on 12 Priority Rules

[20] M.-C. Wu and S.-H. Sun, ‘‘A project scheduling and staff assignment [42] S. Rostami, S. Creemers, and R. Leus, ‘‘New strategies for stochastic
model considering learning effect,’’ Int. J. Adv. Manuf. Technol., vol. 28, resource-constrained project scheduling,’’ J. Scheduling, vol. 21, no. 3,
nos. 11–12, pp. 1190–1195, May 2006. pp. 349–365, Jun. 2018.
[21] W. J. Gutjahr, S. Katzensteiner, P. Reiter, C. Stummer, and M. Denk, [43] F. F. Boctor, ‘‘Some efficient multi-heuristic procedures for resource-
‘‘Competence-driven project portfolio selection, scheduling and staff constrained project scheduling,’’ Eur. J. Oper. Res., vol. 49, no. 1, pp. 3–13,
assignment,’’ Central Eur. J. Oper. Res., vol. 16, pp. 281–306, Apr. 2008. Nov. 1990.
[22] W. J. Gutjahr, S. Katzensteiner, P. Reiter, C. Stummer, and M. Denk, [44] R. Kolisch, ‘‘Efficient priority rules for the resource-constrained project
‘‘Multi-objective decision analysis for competence-oriented project port- scheduling problem,’’ J. Oper. Manag., vol. 14, no. 3, pp. 179–192,
folio selection,’’ Eur. J. Oper. Res., vol. 205, no. 3, pp. 670–679, Sep. 2010. Sep. 1996.
[23] Y. H. Lee, S. R. T. Kumara, and K. Chatterjee, ‘‘Multiagent based dynamic [45] Z. Chen, E. Demeulemeester, S. Bai, and Y. Guo, ‘‘Efficient priority rules
resource scheduling for distributed multiple projects using a market mech- for the stochastic resource-constrained project scheduling problem,’’ Eur.
anism,’’ J. Intell. Manuf., vol. 14, no. 5, pp. 471–484, Oct. 2003. J. Oper. Res., vol. 270, no. 3, pp. 957–967, Apr. 2018.
[24] T. Wauters, K. Verbeeck, P. De Causmaecker, and G. V. Berghe, [46] C. Fang, R. Kolisch, L. Wang, and C. Mu, ‘‘An estimation of distribu-
‘‘A learning-based optimization approach to multi-project scheduling,’’ tion algorithm and new computational results for the stochastic resource-
J. Scheduling, vol. 18, no. 1, pp. 61–74, Feb. 2015. constrained project scheduling problem,’’ Flexible Services Manuf. J.,
vol. 27, no. 4, pp. 585–605, Dec. 2015.
[25] J. S. K. Lau, G. Q. Huang, K. L. Mak, and L. Liang, ‘‘Distributed project
[47] M. Wang, G. Liu, and X. Lin, ‘‘Dynamic optimization of the multi-
scheduling with information sharing in supply chains: Part II—Theoretical
skilled resource-constrained project scheduling problem with uncertainty
analysis and computational study,’’ Int. J. Prod. Res., vol. 43, no. 23,
in resource availability,’’ Mathematics, vol. 10, no. 17, p. 3070, Aug. 2022.
pp. 4899–4927, Dec. 2005.
[48] J. Snauwaert and M. Vanhoucke, ‘‘A new algorithm for resource-
[26] J. S. K. Lau, G. Q. Huang, K. L. Mak, and L. Liang, ‘‘Distributed
constrained project scheduling with breadth and depth of skills,’’ Eur. J.
project scheduling with information sharing in supply chains: Part I—
Oper. Res., vol. 292, no. 1, pp. 43–59, Jul. 2021.
An agent-based negotiation model,’’ Int. J. Prod. Res., vol. 43, no. 22,
[49] F. Li and Z. Xu, ‘‘A multi-agent system for distributed multi-project
pp. 4813–4838, Nov. 2005.
scheduling with two-stage decomposition,’’ PLoS ONE, vol. 13, no. 10,
[27] J. Homberger, ‘‘A multi-agent system for the decentralized resource- Oct. 2018, Art. no. e0205445.
constrained multi-project scheduling problem,’’ Int. Trans. Oper. Res., [50] B. F. Almeida, I. Correia, and F. Saldanha-Da-Gama, ‘‘A biased random-
vol. 14, no. 6, pp. 565–589, Nov. 2007. key genetic algorithm for the project scheduling problem with flexible
[28] J. Homberger, ‘‘A (µ, λ)-coordination mechanism for agent-based multi- resources,’’ TOP, vol. 26, no. 2, pp. 283–308, Jul. 2018.
project scheduling,’’ OR Spectr., vol. 34, no. 1, pp. 107–132, Jan. 2012.
[29] Z. Zheng, Z. Guo, Y. Zhu, and X. Zhang, ‘‘A critical chains based dis-
tributed multi-project scheduling approach,’’ Neurocomputing, vol. 143,
pp. 282–293, Nov. 2014.
[30] Q. Wang, C. Liu, and L. Zheng, ‘‘A column-generation-based algorithm for
a resource-constrained project scheduling problem with a fractional shared YINING YU was born in Qingdao, Shandong,
resource,’’ Eng. Optim., vol. 52, no. 5, pp. 798–816, May 2020. China, in 1994. She received the B.S. degree
[31] M. Rostami and M. Bagherpour, ‘‘A Lagrangian relaxation algorithm in mathematics from Qufu Normal University,
for facility location of resource-constrained decentralized multi-project Shandong, in 2016, and the M.S. degree in
scheduling problems,’’ Oper. Res., vol. 20, no. 2, pp. 857–897, Jun. 2020. mathematics from Beihang University, Beijing,
[32] S. Zhao and Z. Xu, ‘‘New closed-loop approximate dynamic programming China, in 2018, where she is currently pursuing
for solving stochastic decentralized multi-project scheduling problem with the Ph.D. degree. Her research interests include
resource transfers,’’ Exp. Syst. Appl., vol. 185, Dec. 2021, Art. no. 115593.
project scheduling, optimization algorithm, and
[33] Z. T. Kosztyán, ‘‘An exact algorithm for the flexible multilevel risk management.
project scheduling problem,’’ Exp. Syst. Appl., vol. 158, Nov. 2020,
Art. no. 113485.
[34] E. B. Tirkolaee, I. Mahdavi, M. M. S. Esfahani, and G.-W. Weber,
‘‘A hybrid augmented ant colony optimization for the multi-trip capacitated
arc routing problem under fuzzy demands for urban solid waste manage-
ment,’’ Waste Manag. Res., vol. 38, no. 2, pp. 156–172, Feb. 2020.
[35] E. B. Tirkolaee, A. Mardani, Z. Dashtian, M. Soltani, and G.-W. Weber, ZHE XU received the Ph.D. degree from the
‘‘A novel hybrid method using fuzzy decision making and multi- Department of Economics and Management,
objective programming for sustainable-reliable supplier selection in two- Beihang University, Beijing, China, in 2006. Since
echelon supply chain design,’’ J. Cleaner Prod., vol. 250, Mar. 2020,
2009, she has been a Professor with the Depart-
Art. no. 119517.
ment of Economics and Management, Beihang
[36] A. Mahdavi, M. A. E. Moghaddam, and A. Mahmoudi, ‘‘Simultaneous
University. Her current research interests include
charging and discharging of multi-tube heat storage systems using copper
fins and Cu nanoparticles,’’ Case Stud. Thermal Eng., vol. 27, Oct. 2021, project management, quality management, and
Art. no. 101343. risk management.
[37] H. Chen, G. Ding, S. Qin, and J. Zhang, ‘‘A hyper-heuristic based
ensemble genetic programming approach for stochastic resource con-
strained project scheduling problem,’’ Exp. Syst. Appl., vol. 167, Apr. 2021,
Art. no. 114174.
[38] F. Zaman, S. Elsayed, R. Sarker, D. Essam, and C. A. C. Coello, ‘‘An evo-
lutionary approach for resource constrained project scheduling with uncer-
tain changes,’’ Comput. Oper. Res., vol. 125, Jan. 2021, Art. no. 105104. SONG ZHAO was born in Chengde, Hebei, China,
[39] T. Servranckx and M. Vanhoucke, ‘‘A tabu search procedure for the in 1991. He received the B.S. degree in mathe-
resource-constrained project scheduling problem with alternative sub- matics from the China University of Mining and
graphs,’’ Eur. J. Oper. Res., vol. 273, no. 3, pp. 841–860, Mar. 2019. Technology, Beijing, China, in 2014, and the M.S.
[40] B. Ashtiani, R. Leus, and M.-B. Aryanezhad, ‘‘New competitive results for degree in mathematics from Beihang University,
the stochastic resource-constrained project scheduling problem: Exploring Beijing, in 2017, where he is currently pursuing
the benefits of pre-processing,’’ J. Scheduling, vol. 14, no. 2, pp. 157–171, the Ph.D. degree. His research interests include
Apr. 2011. project scheduling, optimization algorithm, and
[41] F. Ballestín, ‘‘When it is worthwhile to work with the stochastic RCPSP?’’ risk management.
J. Scheduling, vol. 10, no. 3, pp. 153–166, Jun. 2007.

VOLUME 11, 2023 29565