See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/341472611

Security-Aware Dynamic Scheduling for Real-Time Optimization in Cloud-

Based Industrial Applications

Article  in  IEEE Transactions on Industrial Informatics · May 2020

DOI: 10.1109/TII.2020.2995348

CITATIONS READS
9 137

7 authors, including:

Shunmei Meng Qianmu Li

Nanjing University of Science and Technology Nanjing University of Science and Technology
42 PUBLICATIONS   568 CITATIONS    241 PUBLICATIONS   1,474 CITATIONS   

SEE PROFILE SEE PROFILE

Shaohua Wan
Zhongnan University of Economics and Law
144 PUBLICATIONS   2,763 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Fuzzy Functional Observer Based Adaptive Controller Design and Fault Detection for Nonlinear Systems under Time-varying Delays View project

Routing in UWSNs View project

All content following this page was uploaded by Mohammad Khosravi on 19 August 2020.

The user has requested enhancement of the downloaded file.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2995348, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (ACCEPTED) 1

Security-aware Dynamic Scheduling for Real-time

Optimization in Cloud-based Industrial Applications
Shunmei Meng, Weijia Huang, Xiaochun Yin, Mohammad R. Khosravi, Qianmu Li, Shaohua Wan,
Lianyong Qi*


Abstract—Nowadays, large number of cloud-based techniques I. INTRODUCTION
have been used in Industrial Control Systems (ICS), which also
brings many security threats. The emergence of security-aware
industrial control has paved the way of security-aware scheduling
in cloud-based industrial applications. Actually, most cloud-based
I N recent years, Industrial Control Systems (ICS) are
gradually migrated from monolithic to distributed
cloud-based architectures. Cloud computing diagram opens up
industrial applications are time-sensitive, which need real-time new potentials in the industrial sector, which provide scalable
processing. Edge cloud computing paradigm extends the resources with high availability to ICS. However, it also makes
computing ability of traditional cloud model with low-latency ICS more complex and vulnerable, and brings up security
local resources. Thus heterogeneous clouds that consist of both
centralized resources and edge resources may be a promising problem [1-2]. The emergence of security-aware industrial
resource model to provide both scalable and low-latency resources control has paved the way of security-aware scheduling,
for cloud-based industrial applications. In view of these challenges, real-time optimization, and data analysis in cloud-based
in this paper, we propose a security-aware dynamic scheduling industrial applications [3].
method for real-time resource allocation in ICS. Firstly, a Because of the explosive growth of mobile devices and other
three-level security model is designed for both tasks and cloud smart infrastructures, the amount of mobile industrial
resources in ICS, and a two-tier heterogeneous cloud architecture
is introduced. Accordingly, a security-aware scheduling method applications has been growing rapidly [4-5]. Most cloud-based
based on distributed PSO (Particle Swarm Optimization) is mobile industrial applications consisting of hundreds of
presented for resource allocation with security concerns. To deal individual tasks could generate big data problems. These “big”
with the dynamics of edge resources and the mobility of mobile mobile industrial applications are usually computing-intensive
industrial applications, a dynamic scheduling mechanism based and security-sensitive, which need scalable and real-time
on dynamic workflow model is proposed for real-time solutions. Benefiting from the cloud computing diagram, most
optimization. Experimental results validate that the scheduling
control policy proposed in this paper can achieve a good balance computing workloads of mobile industrial applications could
between scheduling performance and security performance. be migrated to centralized cloud resources for scalable
solutions [6-7].
Index Terms—Security, Cloud-based industrial applications, However, connecting with remote clouds and offloading
Dynamic scheduling, Real-time optimization, PSO computing workload to remote cloud instances will yield high
latency, since most mobile industrial applications are time
This paper is partially supported by the National Natural Science sensitive [8]. Recently, edge computing has been envisioned as
Foundation of China under Grant No. 61702264, No. 61872219, the
Fundamental Research Funds for the Central Universities, No. 30918014108,
a prospective resource model, which could provide low-latency
the Natural Science Foundation of Shandong Province, No. ZR2019MF001. resources for mobile industrial applications [9-11]. Since edge
Shunmei Meng is with Department of Computer Science and Engineering, clouds could provide low-latency resources and remote
Nanjing University of Science and Technology, Nanjing, China (e-mail:
[email protected]).
distributed clouds could provide scalable resources, a
Weijia Huang is with Department of Computer Science and Engineering, heterogeneous cloud model with both of these two positive
Nanjing University of Science and Technology, Nanjing, China (e-mail: features can be designed to provide scalable and low-latency
[email protected]). resources.
Xiaochun Yin is with Facility Horticulture Laboratory of Universities in
Shandong, WeiFang University of Science & Technology, Shouguang, China Resource allocation in scheduling problem aims to do
(e-mail: [email protected]). optimized resource deployment to each task of applications so
Mohammad R. Khosravi is now with Department of Computer Engineering, as to satisfy users’ diverse requirements [12-13]. However,
Persian Gulf University, Bushehr, Iran, and has been with Department of
Electrical and Electronic Engineering, Shiraz University of Technology, Shiraz, most existing scheduling researches mainly focus on schedule
Iran ( e-mail: [email protected], [email protected]). planning, and seldom consider dynamic scheduling during the
Qianmu Li is with Department of Computer Science and Engineering, execution [14-15]. Actually, in mobile environment, local edge
Nanjing University of Science and Technology, Nanjing, China (e-mail:
[email protected]). resources are changing with the movement of mobile users,
Shaohua Wan is with School of Information and Safety Engineering, thus reactive dynamic scheduling for mobile industrial
Zhongnan University of Economics and Law, Wuhan, China (e-mail: applications should be conducted for dynamic resource
[email protected]).
Lianyong Qi is with School of Information Science and Engineering, Qufu
deployment to deal with the dynamics [16]. Moreover, the
Normal University, Shandong, China (Corresponding author: e-mail: complex and insecure network environment also brings many
[email protected]). security threats [17], which will put users in risk, thus

1551-3203 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Skovde. Downloaded on May 19,2020 at 07:00:59 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2995348, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (ACCEPTED) 2

security-aware scheduling is also needed. multi-heuristic resource allocation algorithm is used to get
Based on the above observations, in this work, a solutions in polynomial time. The authors in [24] propose an
security-aware dynamic scheduling method based distributed uncertainty-aware dynamic resource allocation method to deal
PSO (Particle Swarm Optimization) is proposed for real-time with the scheduling optimization problem. Singh et al. [26]
resource allocation, which aims to get the optimized cost with design a cost-aware scheduling approach based on the K-means
the deadline and security constraints. The main contributions of clustering algorithm, which partitions problem for dynamic
our work are listed as follows: scheduling and aims to satisfy the deadline. The researches
 Firstly, a two-tier heterogeneous cloud architecture is [27-28] focus on edge-based scheduling methods, which aim to
introduced to offer both centralized and low-latency respond to users’ real-time requirement. However, most of
resources for real-time ICS, and a three-level security existing dynamic or real-time scheduling researches seldom
model is designed for both tasks and resources. address the adaptive scheduling across heterogeneous clouds
 Then a security-aware dynamic scheduling method based and rarely consider the security problem in scheduling.
on distributed PSO is presented, where both baseline
C. Security-aware Scheduling
scheduling and dynamic scheduling schemes are designed
for real-time optimization. Concerning the privacy and security problem in scheduling
 Finally, experiments are designed and conducted to problems, there have been some studies on security-aware
validate the effectiveness of our method. scheduling algorithms [29-35]. Sharif et al. [29] design a
The remainder of this work is organized as follows: Section II security-aware scheduling method in hybrid clouds to optimize
reviews related researches of our proposal. In Section III, cost while satisfying privacy and deadline constraints. Wen et
system overview is presented. In Section IV, a security-aware al. [30] proposes a multi-objective scheduling method with
dynamic scheduling based on distributed PSO is proposed. privacy protection constraints based on the genetic algorithm.
Section V empirically studies the empirical performance and Zhang et al. [31] integrate the security constraint into task
effectiveness of our proposal. Finally, Section VI concludes scheduling by constructing a pairwise Markov chain, which
this paper and gives an outlook on our future work. aims to achieve image feature extraction on edge with privacy
preservation. In the scheduling mechanism of [34],
II. RELATED WORK intermediate data encrypting is conducted by utilizing tasks'
laxity time for security-aware applications. The reference [35]
Scheduling and resource allocation problem in clouds and
focuses on security-aware recommendation algorithms for
industrial applications has been in wide research. Related
efficient service scheduling. Overall, there are few researches
researches are reviewed from the following three categories.
considering both dynamic scheduling for real-time
A. Scheduling in Cloud-based Industrial Systems optimization and security concerns. In this paper, we design a
In this section, we mainly discuss the scheduling problem in security-aware dynamic scheduling algorithm where both
cloud-based industrial systems [18-22]. Abrishami et al. [18] adaptive scheduling across heterogeneous clouds and the
propose a deadline-constrained scheduling method based on the security concerns are under consideration.
partial critical path strategy for IaaS-cloud-based industrial
applications. The reference [19] provides an energy-aware VM III. SYSTEM OVERVIEW
scheduling approach on cloud-based cyber systems. Recently, Fig. 1 presents the framework of our proposal. Firstly, a
with rapid development of mobile edge computing, scheduling security model and a two-tier heterogeneous cloud architecture
problem in edge clouds also attracts the attention of many are provided. Then based on the above models, a
researchers. Boveiri et al. [21] introduce a high-performance security-aware scheduling method based on the distributed
scheduling technique for cloud-based industrial systems based PSO scheme is proposed. Moreover, to handle the resource
on a modified ant colony optimization algorithm. Zhang et al. dynamics, a dynamic scheduling approach based on mobile
[22] present an energy-aware scheduling approach based on a dynamic workflow is presented for real-time optimization.
double deep Q-learning model. There have been many More details will be discussed in the following.
researches focusing on scheduling algorithms in cloud-based
industrial systems. However, most of them didn’t consider
dynamic scheduling with reactive actions during the execution
of cloud-based applications, and few researches focus on the
scheduling problem in heterogeneous-cloud-based systems.
B. Dynamic and Real-time Scheduling
To deal with the mobility of mobile users and resource
dynamics in mobile environment, there also appear some
researches focusing on dynamic scheduling during execution of
applications [23-28]. Juarez et al. [23] present an energy-aware
runtime scheduling method for parallel task-based application
with the consideration of green computing, where a
Fig.1. Security-aware dynamic scheduling framework on heterogeneous clouds

1551-3203 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Skovde. Downloaded on May 19,2020 at 07:00:59 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2995348, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (ACCEPTED) 3

A. Security Model for Scheduling of Industrial Applications For a task ti, its execution time contains both transmission time
In this research, a security model is designed for and execution time. The execution time on local edge cloud
configuration security and resource management in resources can be calculated based on equation (1). Here,
cloud-based industrial applications. Firstly, the definition of the transmission time between the local edge cloud resources
security degree is defined in the following. and tasks is ignored.
Definition 1. (Security Degree) In this security model, each (1)
task of an industrial application and each cloud resource are
tied with a security degree indicating the security performance where is the workload of task ti, and is the CPU
of the task or resource. Here, we consider three levels of capacity based on MIPS of the local edge cloud resource lr.
security degrees in our security model: public, semi-public, and The execution time of task ti offloaded to the remote
private, which are respectively noted as lpu, lspu, and lpr . The centralized cloud resources not only consists of the execution
security degree relationship is . Higher degree time but also the data transmission time between the task and
means that the task or resource has higher security concerns. resource, which is defined in equation (2):
According to the definition of security degree, there are
three kinds of tasks in mobile industrial applications, i.e., public (2)
tasks, semi-public tasks, and private tasks. Similarly, there are
where is the CPU capacity based on MIPS of the cloud
three kinds of cloud resources, i.e., public instances,
resource pr and is the output data size transferred from
semi-public instances, and private instances. Table I gives the
task ti to the resource pr, and is the input data size
mapping relations between tasks and resources with different
transferred from pr to task ti, and
security degrees. In Table I, “√” represents that the task can be
respectively represent the data transmission rate sending or
deployed in the resource instance of the corresponding row, and
receiving data from resource pr to task ti.
“×” represents that the task cannot be deployed in the
Some other basic scheduling notions are described in the
corresponding resource instance. Specifically, tasks with public
following. ES(ti) indicates the earliest start time of task ti. EF
tags can be deployed on all kinds of resource instances.
(ti) is the earliest finish time of task ti, LF (ti) is the latest finish
Semi-public tasks can be deployed on semi-public instances
time of task ti .
and private instances, and cannot be deployed on public
instances. Private tasks can only be deployed on private ( ) ( )( ( ) ( ) ( ))
resource instances. { ( ) ( ) ( ) (3)
TABLE I ( ) ( ){ ( ) ( ) ( )}
MAPPING RELATIONS BETWEEN TASKS AND RESOURCE INSTANCE
where P(ti) indicates the predecessor tasks of ti, S(ti) indicates
Public task Semi-public task Private task
the predecessor tasks of ti , ME(tj) is the minimum execution
Public instance √ × ×
time of task tj on the fastest candidate resource, TT(ti ,tj) is the
Semi-public instance √ √ ×
data transmission time between ti and tj.
Private instance √ √ √

B. Two-tier Heterogeneous Cloud Architecture D. PSO Mechanism

The two-tier heterogeneous cloud resource architecture The PSO algorithm is a population-based global search
proposed in our work could meet diverse requirements of swarm intelligent algorithm [36]. The main scheme of the PSO
algorithm is the updating of velocity and position, which is
mobile users, which can not only provide scalable centralized
described as follows:
cloud resources but also low-latency cloud resources. On the
first tier (Tier-1), the edge clouds could provide local edge ( ) ( )
cloud resource instances, which could supply low-latency { ( )
services with the fine grain location for real-time cloud-based
industrial systems. The second tier (Tier-2) could provide where is denoted as the position vector,
scalable centralized resource instances with high performance, is denoted as the velocity vector,
which have the ability of scalability and high availability. Users , D is the dimensional of the search space which
can offload their industrial applications to local edge clouds or corresponds to the number of tasks in the scheduling problem,
distributed scalable clouds by the wireless access network. This denotes the d-th dimension of . and are the d-th
heterogeneous cloud architecture integrated with the above velocities of particle i at the k-th iteration and the (k+1)-th
security model could address the requirements of users in iteration, respectively. is the inertia weight, and are
scalability, low-latency and security concerns. acceleration coefficients, and are random numbers
C. Application Model and Task Computing Model ranging between the interval [0, 1], is the personal best
position of particle i, is the global best position of the
Each industrial application can be modeled as a DAG (direct entire swarm. The main procedure of the PSO algorithm
acyclic graph) graph G= (T, E). includes initialization, fitness function, velocity and position
represents the task set. Each task updating. More details of PSO could be referred to [36].
ti is tied with a security tag .
E={ei,j|ei,j=(ti, tj)} represents the dependencies between tasks.

1551-3203 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Skovde. Downloaded on May 19,2020 at 07:00:59 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2995348, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (ACCEPTED) 4

IV. SECURITY-AWARE DYNAMIC SCHEDULING METHOD ON the number of tasks in the critical path is far less than the
DISTRIBUTED PSO number of tasks of the workflow. It will consume less time to
Inspired by the advantages (such as high-precision solutions filter bad initial particles based on the critical path strategy.
and fast convergence) of PSO, PSO is applied in our work for Algorithm 2: CP & Priavcy-based initialization
dynamic scheduling. Specifically, a baseline scheduling Input: critical path CP of workflow G, deadline UD, particle number N,
method based on distributed PSO algorithm is proposed for security constraint of tasks and resources
schedule planning. Then to deal with the mobility and resource Output: N Initial particles
1: for i = 1 to N do
dynamics, a dynamic scheduling method is presented based on 2: for j =1 to |T | do
the mobile dynamic workflow for real-time optimization. 3: () ()
4: if (PC(Res(ti)) PC(ti)) do
A. Security-aware Scheduling based on Distributed PSO 5: j=j-1
In this paper, we improve the search capability of PSO-based 6: end if
scheduling by the improved parameter setting strategies and the 7: end for
8: while (FT(CP) > UD) do
distribution strategy. 9: for each do
(1) Improved settings in PSO-based scheduling 10: [ ] () [ ] ()
1) Initialization based on critical path and security constraint 11: end for
In our PSO-based scheduling method, to improve quality of 12: end while
the initial particles, we use the critical path (CP) strategy and 13: end for
the security constraint to filter the poor initial population 2) Parameters settings in adaptive PSO
solutions, which are described as follows: The setting of parameters, such as inertia weight and
Determine the CP of the workflow: To meet the deadline acceleration parameters , , has great influence on the search
constraint, the exit task texit should be finished before deadline efficiency. Thus in our work, we apply adaptive parameters to
UD, i.e., LF(texit) UD. Algorithm 1 gives the procedure to update the velocity and position of particles.
determine the CP. Inertia weight: For adaptive learning of PSO, an adaptive
Algorithm 1: Fine the CP of a Workflow G={T, E} inertia weight is adopted for velocity updating [37]. The inertia
Input: Workflow G weight in the k-th iteration can be determined by equation (5).
Output: the CP of workflow G
( ) ( ( ))
1: t = texit , (5)
2:
3: while ( ( ) ) do ( ) ( ∑ ( ( ) ̅) ) (6)
4: for each ( ) do
where are the maximum and minimum inertia
5: MT =0
6: if EF(tp)+TT(tp, t) > MT then weight values, is a diversity function decided by equation
7: MT =EF(tp)+TT(tp, t) (6), ( ) is the fitness value of particle in k-th iteration, ̅
8: is the average fitness of the swarm in k-th iteration. ( ( ))
9: end if
is an adjustment coefficient where equals to K/3 and K is total
10: end for
11: number of iterations, N is the number of particles in the swarm.
12: Acceleration parameters: The acceleration parameters
13: end while and are defined by equation (7):
14: return CP.
( ) ( ) (7)
In a critical path, each task t has a critical parent tcp. And tcp is
the task that has the latest finish time to task t, i.e., EF (tp)+TT(tp, where and are the initial values of and , and
t) is maximal (Line 6-9). After determining the critical parent tcp are constants which are empirically determined.
of task t in CP, tcp will be added to the critical path CP (Line 11).
Repeatedly, the overall critical path for the workflow will be 3) Fitness Function
determined. In this paper, our scheduling goal is to minimize the cost
CP&Priavcy-based initialization: Algorithm 2 presents the while satisfying the deadline and security constraints. Thus the
process to decide an initial swarm. Firstly, initial particles are fitness function can be formulated as follows:
generated randomly (Line 3). If the security degree of the ∑ (8)
resource assigned to task ti is lower than the security degree of
task ti., i.e., PC(Res(ti)) PC (ti), then the random solution ∑ ( ( ) ( )
should be filtered as it cannot satisfy the security constraint where eci and eti respectively denote the execution cost and the
(Line 4-6). Here, Res(ti) denotes the resource assigned to task ti, execution time of task ti, and eti contains not only the execution
( ( ) represents the security constraint of the resource time but also the output/input data transmission time between
assigned to task ti, and ( ) is the security constraint of task ti. task and the assigned resource.
Moreover, if the finish time of the critical path is bigger than (2) Scheduling based on distributed PSO
deadline UD, then the solution is infeasible and will be seen as To deal with big data problem in scheduling, we apply a
a bad initial particle which will also be filtered (Line 8-12). As distributed PSO strategy to find the best scheduling solution,

1551-3203 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Skovde. Downloaded on May 19,2020 at 07:00:59 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2995348, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (ACCEPTED) 5

where the original swam are divided into multiple sub-swarms, the movement of mobile users, as the local edge resources are
noted as , where M is the number of continually changing with the movement of mobile industrial
sub-swarms. Algorithm 3 presents the overall procedure of the applications. Thus it is necessary to update the scheduling for
distributed PSO-based scheduling strategy. the uncompleted tasks dynamically at each waypoint. To
Algorithm 3: Distributed PSO-based scheduling address this issue, mobile dynamic workflow model is designed
Input: the total number of iterations K for rescheduling.
Output: gbest Definition 2. (Mobile Dynamic Workflow, MDW) The
1: Initialize the particle swarm randomly based on Algorithm 2. mobile dynamic workflow at a waypoint is a sub workflow
2: Split original swarm into M sub-swarms
cut-off from the original application workflow, which only
3: for each do
4: gbesti=0; contains executing tasks (tasks which are being executed) and
5: while (k < K) do unexecuted tasks of the original workflow. And each waypoint
6: for each particle j in do has a different mobile dynamic workflow.
7: Calculate the fitness values of particle j (denoted as F(j)) in Fig. 2 presents the evolutionary process of the mobile
8: if F(j) > pbestj do
dynamic workflow at each waypoint, where MDWi denotes as
9: pbestj = F(j)
10: end if the dynamic partial workflow at waypoint pi. As shown in Fig. 2,
11: if F(j) > gbesti do different waypoint has different local edge cloud resources, thus
12: gbesti = F(j) re-scheduling should be conducted for the remaining tasks for
13: end if real-time optimization. At a waypoint pi, the application
14: Update the velocity and position of the particle based on the
workflow has three kinds of tasks: completed tasks, executing
updating equations (equation (4)).
15: end for tasks and unexecuted tasks. The dynamic workflow at each
16: k++; waypoint only contains executing tasks and unexecuted tasks.
17: end while
18: end for Heterogeneous Clouds
19: gbest max{ gbest1 ,…, gbestM} Centralized Clouds
In Algorithm 3, Firstly, initialization is conducted (Line 1).
In each sub-swarm SWi, the particles are updated iteratively
Edge Edge Edge Edge
until the stop criterion is satisfied (the number of iterations is no Cloud Cloud Cloud Cloud
less than the total number of iterations K). In each iteration,
pbest of each particle and gbesti of each sub-swarm should be
Waypoint Waypoint Waypoint Waypoint
updated accordingly (Line 7-Line 13). Finally, merge gbesti of p0 p1 ... ph-1 ph
each sub-swarms, the best solution among all gbesti will be
determined as the global gbest (Line 19). Then the global best
scheduling solution will be determined. ...

B. Dynamic Scheduling Based on Mobile Dynamic Workflow MDW0 MDW1 MDWh-1 MDWh

Model
Fig. 2. Evolutionary process of mobile dynamic workflow.
In this section, a dynamic scheduling method is proposed to
handle the influence of the mobility of users and resource Fig. 3 gives an example of the specific generation process of a
dynamics on scheduling for real-time optimization. mobile dynamic workflow. When the mobile user researches a
(1) Mobility model of mobile industrial applications new waypoint, the original application workflow in Fig. 3 (a)
can be pruned into the real-time partial workflow shown in Fig.
In this work, we use a well-known mobility model, i.e.,
3 (b). In Fig. 3 (a), red tasks, green tasks and black tasks are
random waypoint (RWP) model [38], to simulate the moving of
denoted as completed tasks, executing tasks and unexecuted
mobile users.
tasks, respectively. As shown in Fig. 3 (b), the completed tasks
In RWP model, the moving trajectory of a mobile user is
in Fig. 3(a) is not contained in the new waypoint, where the
modeled as the moving trajectory between different waypoints
executing tasks and unexecuted tasks will be rescheduled in the
where mobile users move at preset speeds and stop at each
current waypoint, which will be described in detail in the
waypoint for a preset time period. It is assumed that the signal
following. Specifically, only the uncompleted workload of
strength keeps stable in each square of the two-dimensional
executing tasks is contained in the DPW workflow. For
grid. Specifically, user starts by choosing its first waypoint p1
example, task t5’ in Fig. 3 (b) is the uncompleted workload of
randomly and stop at p1 for a random time period which is
task t5 in Fig. 3 (a).
between [Tmin, Tmax]. Then the user selects another waypoint,
i.e., the second waypoint p2, and moves to p2 at a random speed t1 t5 t9 t14 t5' t9 t14
between [vmin, vmax]. User will also pause at p2 for another t2 t6 t11 t15 t11' t15
random time period. This procedure will be repeated for the ten ten
tex tex
t3 t7 t12 t7' t12
whole execution process of the mobile industrial application.
t4 t8 t16
(2) Generating mobile workflows based on RWP model t13 t 8' t13 t16

During the movement of cloud-based mobile applications, (a) The workflow of previous waypoint (b) The mobile dynamic workflow
of current waypoint
the optimized resources to each individual task may vary with
Fig. 3. An example of generation of a mobile dynamic workflow at a waypoint.

1551-3203 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Skovde. Downloaded on May 19,2020 at 07:00:59 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2995348, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (ACCEPTED) 6

(3) Rescheduling for real-time optimization 2) Dataset (Application workflow model)

As the local edge resources are continually changing with the Application Workflow Dataset 1: Realistic workflows [18]
movement of mobile users, there may be better solution for the with different structures and computational features are used to
unexecuted tasks during the movement of users. For executing model the task structures of mobile industrial applications in
tasks, we will not consider the re-scheduling if there is no our experiment. There are five realistic workflows: Montage,
potential better resource. If there exists any potential better CyberShake, Epigenomics, LIGO and SIPHT. Because of
resource for the executing tasks, then the uncompleted limited space, only the experimental results of Epigenomics
workload of executing tasks is also needed to be rescheduled. and CyberShake are presented and analyzed.
Here, it is said to exist potential better resource if there exists Application Workflow Dataset 2: Some artificial workflows
a new local edge resource that dominates the current resource are generated randomly with sequential tasks to analyze the
allocated to the executing task. Pareto dominance [39] is a scalability of the distributed PSO-based scheduling approach.
well-known method used to compare solutions, which is 3) Comparative methods
appropriate for conflicting objectives. For , x is IC-PCPD2: The IC-PCPD2 method [18] is a cost-aware
considered to dominate y if and only if ( ) deadline-constrained scheduling approach on IaaS cloud,
( ) ( ) ( ). Then a solution x* is said to be which is based on the partial critical path scheme.
pareto optimal if it is not dominated by any other solution. In PSOS [40]: This method is a heuristic scheduling based the
scheduling problems, for a task, a resource scheduling r* PSO mechanism, which did not consider dynamic scheduling.
dominates resource r’ if both the execution time and the cost of PDS-Loss: It is a variation of our proposal, which did not
scheduling r* for the task is no more than that of scheduling r’, consider distributed computing and dynamic scheduling.
and at least one of them is less.
If there exists any potential better resource for some B. Experimental Result
executing tasks, the uncompleted workload of theses executing In this experiment, our approach is compared with other three
tasks will be treated as unexecuted tasks in MDW workflow, methods in cost and makespan. Besides, security performance
which will be rescheduled together with other unexecuted tasks. is also analyzed.
Other executing tasks that have no potential better resource will 1) Effect of the change of deadline factor
be executed in the original resources. In each waypoint, based In our experiment, we use the deadline factor to set
on the updated candidate resources, re-scheduling will be different deadlines for the application workflows. The
conducted for the MDW workflow according to the distributed deadlines are set between the lower bound (executed on the
PSO-based scheduling method until all tasks of the application fastest instances, noted as lowestmakespan) and the upper
has been completed. bound (executed on the slowest instances, noted as
highestmakespan) of the execution time for the workflow. The
V. EXPERIMENT difference between the lowest bound and the highest bound of
In this section, simulation experiments are designed to the execution time is divided into four partitions, and each
evaluate the effectiveness of our method (named as PDS). partition is denoted as Diff . Then means that the
Experimental settings are described firstly in the following. deadline set as .
( )
A. Experimental Settings
1) Experimental setup Fig. 4 presents the mean cost of the four methods with the
In our experiment, to simulate the heterogeneous cloud increase of deadline factor where the number of initial
resource architecture, Amazon T2 instances are used to simulate particles and iterations is fixed. We can see that PDS is the
the scalable centralized clouds in Tier-2. Amazon T2 instances lowest-expensive policy compared with other three methods as
are public clouds which have scalable and elastic cloud services the line for PDS is at the bottom of the four lines. Moreover, it
with high availability. The details of Amazon T2 instances are can be observed that the execution cost of other three methods
presented in Table II. A group of device nodes are used to (ICPCPD2, PSOS and PDS-Loss) decreases with increase of
simulate the edge resources in Tier-1. It is assumed that the when , and the performance of execution cost almost
prices of edge resources are most proportional to the computing remain stable when . It is s due to the fact that when
capacity. RWP model is utilized to model the moving trajectory is small, the deadline is hard to satisfy, then high-performance
of mobile industrial applications. The parameters in the resources with higher prices should be assigned to tasks to meet
adaptive PSO mechanism are set as follows: the deadline. Thus The execution cost will decrease with the
, which increase of (The deadline becomes looser) when .
are all empirical values. When the deadline limitation is much looser ( ), and the
workflow can be finished before deadline in the scheduling
TABLE II AMAZON T2 INSTANCES
VCPU RAM (GiB) CPU (points/hour) Price/Hour solution where the execution cost is approximately minimized,
t2.nano 1 0.5 3 0.0058USD then the cost will become stable. While the execution cost of
t2.micro 1 1.0 6 0.0116USD PDS continually decreases with the increase of deadline factor,
t2.small 1 2.0 12 0.023USD which is more obvious in CyberShake workflow. It is due to
t2.medium 2 4.0 24 0.0464USD that fact that the edge resources are dynamic in mobile
t2.large 2 8.0 36 0.0928USD

1551-3203 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Skovde. Downloaded on May 19,2020 at 07:00:59 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2995348, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (ACCEPTED) 7

environment. There may exist better resources with high both Epigenomics and CyberShake workflows. It indicates that
performance and less cost during the movement of mobile users. adopting distribution computing in PSO mechanism could get a
Thus the cost of PDS will be continually optimized with the global best scheduling solution stably, even when the number
increase of deadline factor. of initial particles and iterations is a bit small. Moreover, from
As shown in Fig. 5, the makespan of our method may higher Fig. 6 and Fig. 7, it can be observed that our method always
than other methods. This is because the goal of our method is to consumes less cost than PSOS and PDS-Loss with the increase
minimize the cost while satisfying the deadline and security of the number of both initial particles and iterations which also
constraints. So if there exist less expensive solutions within the indicates that PDS is the lowest-expensive policy compared
deadline, the resource allocation for uncompleted tasks will be with other two PSO-based scheduling methods.
updated, which will yield less cost while higher makespan.
Moreover, it can be observed that the results of Epigenomics
and CyberShake are very different. It is due to the fact that
Epigenomics and CyberShake have different structures and
computational features, and edge clous resources are dynamic.
Then the evaluation results of Epigenomics and CyberShake
may not have consistent patterns.

(a) Mean cost for Epigenomics (b) Mean cost for CyberShake

Fig. 6. Mean cost with the increase of number of initial particles (dealine
factor=2.0, number of iterations = 80, number of distribution partitions =3).

(a) Mean cost for Epigenomics (b) Mean cost for CyberShake
Fig. 4. Mean cost under different deadlines (number of initial particles=200,
number of iterations=60, number of distribution partitions=3).

(a) Mean cost for Epigeonmics (b) Mean cost for CyberShake

Fig. 7. Mean cost with the increase of the number of iterations (deadline factor =
2.0, initial number of particles = 100, number of distribution partitions =3).

3) Runtime of PDS
Fig. 8 provides the runtime of PDS with the increase of the
number of distribution partitions under different number of
initial particles. This simulation is run on the artificial
(a) Mean makespan for Epigenomics (b) Mean makespan for CyberShake workflow dataset which has much larger tasks than
Epigenomics and CyberShake. In Fig. 8, PDS-200, PDS-400,
Fig. 5. Mean makespan under different deadlines (number of initial particles=200,
number of iterations=60, number of distribution partitions =3). PDS-800, and PDS-1600 denote that the number of initial
particles in PDS is set as 200, 400, 800 and 1600 respectively.
2) Effect of the number of iterations and initial particles The horizontal ordinate represents the number of distribution
Fig. 6 and Fig. 7 respectively show the performance of the partitions (denoted as NDP). “NDP=1” represents that the
initial particles are not divided, which is denoted as PDS-WD.
three PSO-based scheduling methods, i.e., PSOS, PDS-Loss
From Fig. 8, it can be found that distributed PSO-based
and PDS, with the increase of the number of initial particles and
scheduling mechanism ( ) has higher search efficiency
iterations. In this case, the deadline factor is set as 2.0 and the
than PDS-WD (NDP=1). It also can be observed that the
number of distribution partitions is set as 3. As shown in the runtime of PDS becomes lower with the increase of the number
simulation results of Fig. 6 and Fig. 7, with the increase of of partitions, while with the increase of partitions, runtime of
number of initial particles and iterations, the cost of PSOS and PDS reduces slower. It is due to the fact that the generation time
PDS-Loss for Epigenomics workflow is not stable, while the and transmission time of the partitions will also increase with
cost of PSOS and PDS-Loss for CyberShake workflow is the increase of the number of partitions. Moreover, PDS
stable. The mean execution cost of our method remains stable (NDP ) finishes the scheduling faster than PDS-WD
with the increase of the number of particles and iterations for ( ), and the effect is more evident when the number of

1551-3203 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Skovde. Downloaded on May 19,2020 at 07:00:59 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2995348, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (ACCEPTED) 8

initial particles is larger. That is to say that the data computing VI. CONCLUSION
workload is larger, PDS is more effective. It indicates that In this paper, a security-aware dynamic scheduling approach
distributed PSO strategy could improve the search efficiency, is proposed for real-time provisioning in cloud-based industrial
especially in big data environment. applications. It aims at getting an optimized execution cost
dynamically with the consideration of user-defined deadline
and security configuration constraints. Firstly, a three-level
security model is introduced to deal with the configuration
security problem in mobile cloud environment, and a two-tier
heterogeneous cloud architecture is presented. Then a
security-aware distributed PSO-based scheduling approach is
proposed for resource allocation of cloud-based industrial
applications. Moreover, to deal with the dynamics in mobile
clouds, a dynamic scheduling mechanism based on mobile
dynamic workflow scheme is proposed for real-time
optimization. Experimental results show that our proposal
could achieve a good balance between scheduling performance
and security performance, which enhance the configuration
Fig. 8. Runtime of PDS with different distribution partitions (deadline
factor=2.0, number of iterations = 100). security of cloud-based industrial applications. In our future
work, we will do further research in multi-objective real-time
4) Security analysis scheduling problem for industrial applications with AI
To study configuration security performance of our method techniques.
for cloud-based industrial applications, the execution cost and
makespan in three security degrees are analyzed, which are REFERENCES
denoted as PDS-P1, PDS-P2 and PDS-P3. In PDS-P1, there is
[1] S. Zhang, X. Li, Z. Tan, et al. A caching and spatial K-anonymity driven
no security constraint. In PDS-P3, there is a strict security privacy enhancement scheme in continuous location-based services.
constraint. In PDS-P2, there exists weak security constraint, Future Generation Computer Systems, vol. 94, pp. 40-50, 2019.
where both semi-public tasks and public tasks can be executed [2] Hua Q S, Shi Y, Yu D, et al. Faster Parallel Core Maintenance Algorithms
on resources with all security degrees, and private tasks could in Dynamic Graphs. IEEE Transactions on Parallel and Distributed
Systems, DOI: 10.1109/TPDS.2019.2960226, 2020.
only be executed on privacy resources and semi-public [3] Jolfaei A, Wu X W, Muthukkumarasamy V. On the security of
resources. Table III and Table IV respectively give the permutation-only image encryption schemes. IEEE transactions on
execution cost and makespan under different deadlines for information forensics and security, 2015, 11(2): 235-246.
PDS-P1, PDS-P2 and PDS-P3. From Table III and Table IV, [4] W. Gong, L. Qi, Y. Xu. Privacy-aware multidimensional mobile service
quality prediction and recommendation in distributed fog environment.
we can see that the PDS approach with stricter security Wireless Communications and Mobile Computing, vol. 2018, Article ID
constraint has more expensive cost than the PDS approach with 3075849, 2018.
looser security constraint. Because the strict security constraint [5] S. Zhang, G. Wang, M. Z. A. Bhuiyan, et al. A dual privacy preserving
makes the number of optional candidate instances with lower scheme in continuous location-based services. IEEE Internet of Things
Journal, vol. 5, no. 5, pp. 4191-4200, 2018.
prices become less. In our work, we aim to get the optimized [6] X. Xu, X. Zhang, H. Gao, et al. BeCome: Blockchain-enabled
execution cost with the deadline security constraints, so we will computation offloading for IoT in mobile edge computing. IEEE
choose the PDS-P3. In multi-objective scheduling considering Transactions on Industrial Informatics, DOI: 10.1109/TII.2019.2936869,
optimizing both cost and security constraints, then we would to be published, 2019.
[7] Wang S, Hu L, Wang Y, et al. Modeling multi-purpose sessions for
make a trade-off between execution cost and security concern, nextitem recommendations via mixture-channel purpose routing
such as PDS-P2. Based on the above analysis, it can be networks. Proceedings of the 28th International Joint Conference on
observed that the security model designed in our proposal cloud Artificial Intelligence. AAAI Press, 2019: 1-7.
achieve a good balance between scheduling performance and [8] B. Zhao, Y. Wang, Y. Li, et al. Task allocation model based on worker
friend relationship for mobile crowdsourcing. Sensors, vol. 19, no. 4, pp.
security performance, which could not only meet the 921, 2019.
scheduling requirements but also enhance the configuration [9] S. Meng, Q. Li, T. Wu, et al. A fault-tolerant dynamic scheduling method
security of cloud-based industrial applications. on hierarchical mobile edge cloud computing. Computational Intelligence,
vol. 35, no. 3, pp. 577-598, 2019.
TABLE III EXECUTION COST ON SECURITY CONSTRAINT [10] Y. Wang, Z. Cai, Z. H. Zhan, et al. An optimization and auction-based
Deadline factor 1.5 2 2.5 3 3.5 4 incentive mechanism to maximize social welfare for mobile
PDS-P1 16.89 14.86 13.68 12.17 11.68 10.36 crowdsourcing. IEEE Transactions on Computational Social Systems, vol.
PDS-P2 17.27 15.39 13.99 12.62 11.87 10.38 6, no. 3, pp. 414-429, 2019.
PDS-P3 21.92 20.77 18.73 17.84 18.84 14.75 [11] Yu D, Zou Y, Yu J, et al. Implementing Abstract MAC Layer in Dynamic
Networks. IEEE Transactions on Mobile Computing, DOI:
TABLE IV MAKESPAN ON SECURITY CONSTRAINT 10.1109/TMC.2020.2971599, 2020.
Deadline factor 1.5 2 2.5 3 3.5 4 [12] S. Zhang, K. K. R. Choo, Liu Q, et al. Enhancing privacy through uniform
PDS-P1 30.40 36.02 40.13 41.88 64.91 71.57 grid and caching in location-based services. Future Generation Computer
PDS-P2 30.86 34.01 39.40 46.50 63.60 71.51 Systems, vol. 86, pp. 881-892, 2018.
PDS-P3 30.05 31.95 39.24 42.98 52.06 66.19 [13] H. Liu, H. Kou, C. Yan, et al. Link prediction in paper citation network to
construct paper correlation graph. EURASIP Journal on Wireless
Communications and Networking, vol. 2019, no. 1, pp. 1-12, 2019.

1551-3203 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Skovde. Downloaded on May 19,2020 at 07:00:59 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2995348, IEEE
Transactions on Industrial Informatics
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS (ACCEPTED) 9

[14] Y. Wang, Z. Cai, X. Tong, et al. Truthful incentive mechanism with [37] A. Al-Maamari, F. A. Omara. Task scheduling using PSO algorithm in
location privacy-preserving for mobile crowdsourcing systems. cloud computing environments. International Journal of Grid and
Computer Networks, vol. 135, pp. 32-43, 2018. Distributed Computing, vol. 8, no. 5, pp. 245-256, 2015.
[15] Jolfaei A, Kant K. A lightweight integrity protection scheme for low [38] N. Kumar, S. Zeadally, N. Chilamkurti, et al. Performance analysis of
latency smart grid applications[J]. Computers & Security, 2019, 86: Bayesian coalition game-based energy-aware virtual machine migration
471-483. in vehicular mobile cloud. IEEE Network, vol. 29, no. 2, pp. 62-69, 2015.
[16] L. Qi, W. Dou, W. Wang, et al. Dynamic mobile crowdsourcing selection [39] S. Ganguly, N. C. Sahoo, D. Das. Multi-objective particle swarm
for electricity load forecasting. IEEE Access, vol. 6, pp. 46926-46937, optimization based on fuzzy-Pareto-dominance for possibilistic planning
2018. of electrical distribution systems incorporating distributed generation.
[17] Jolfaei A, Wu X W, Muthukkumarasamy V. Comments on the security of Fuzzy Sets and Systems, vol. 213, pp. 47-73, 2013.
“Diffusion–substitution based gray image encryption” scheme. Digital [40] S. Pandey, L. Wu, S. M. Guru, et al. A particle swarm optimization-based
signal processing, 2014, 32: 34-36. heuristic for scheduling workflow applications in cloud computing
[18] S. Abrishami, M. Naghibzadeh, D. H. J. Epema. Deadline-constrained environments. 2010 24th IEEE international conference on advanced
workflow scheduling algorithms for infrastructure as a service clouds. information networking and applications, pp. 400-407, April, 2010.
Future Generation Computer Systems, vol. 29, no. 1, pp. 158-169, 2013.
[19] X. Xu, X. Zhang, M. Khan, et al. A balanced virtual machine scheduling
method for energy-performance trade-offs in cyber-physical cloud
systems. Future Generation Computer Systems, vol. 105, pp. 789-799,
2017.
[20] Li F, Yu D, Yang H, et al. Multi-Armed-Bandit-Based Spectrum
Scheduling Algorithms in Wireless Networks: A Survey[J]. IEEE
Wireless Communications, 2020, 27(1): 24-30.
[21] H. R. Boveiri, R. Khayami, M. Elhoseny, et al. An efficient
Swarm-Intelligence approach for task scheduling in cloud-based internet
of things applications. Journal of Ambient Intelligence and Humanized
Computing, vol. 10, no. 9, pp. 3469-3479, 2019.
[22] Q. Zhang, M. Lin, L. T. Yang, et al. A double deep Q-learning model for
energy-efficient edge scheduling. IEEE Transactions on Services
Computing, vol. 12, no. 5, pp. 739-749, 2018.
[23] F. Juarez, J. Ejarque, R. M. Badia. Dynamic energy-aware scheduling for
parallel task-based application in cloud computing. Future Generation
Computer Systems, vol. 78, pp. 257-271, 2018.
[24] X. Xu, H. Cao, Q. Geng, X. Liu, F. Dai, and C. Wang. Dynamic Resource
Provisioning for Workflow Scheduling under Uncertainty in Edge
Computing Environment. Concurrency and Computation-Practice &
Experience, DOI: 10.1002/cpe.5674, to be published, 2020,.
[25] Wang S, Hu L, Cao L, et al. Attention-based transactional context
embedding for next-item recommendation. Thirty-Second AAAI
Conference on Artificial Intelligence. 2018.
[26] V. Singh, I. Gupta, P. K. Jana. A novel cost-efficient approach for
deadline-constrained workflow scheduling by dynamic provisioning of
resources. Future Generation Computer Systems, vol. 79, pp. 95-110,
2018.
[27] C. A. Aslam, Y. L. Guan, K. Cai. Edge-based dynamic scheduling for
belief-propagation decoding of LDPC and RS codes. IEEE Transactions
on Communications, vol. 65, no. 2, pp. 525-535, 2016.
[28] X. Li, J. Wan, H. N. Dai, et al. A hybrid computing solution and resource
scheduling strategy for edge computing in smart manufacturing. IEEE
Transactions on Industrial Informatics, vol. 15, no.7, pp. 4225-4234,
2019.
[29] S. Sharif, P. Watson, J. Taheri, et al. Privacy-aware scheduling SaaS in
high performance computing environments. IEEE Transactions on
Parallel and Distributed Systems, vol. 28, no. 4, pp. 1176-1188, 2016.
[30] Y. Wen, J. Liu, W. Dou, et al. Scheduling workflows with privacy
protection constraints for big data applications on cloud. Future
Generation Computer Systems, 2018.
[31] H. Zhang, K. Zeng. Pairwise Markov Chain: A Task Scheduling Strategy
for Privacy-Preserving SIFT on Edge. IEEE INFOCOM 2019-IEEE
Conference on Computer Communications, pp. 1432-1440, April 2019.
[32] Sheng H, Zheng Y, Ke W, et al. Mining Hard Samples Globally and
Efficiently for Person Re-identification. IEEE Internet of Things Journal,
2020.
[33] Xiao Y, Jia Y, Liu C, et al. Edge computing security: State of the art and
challenges. Proceedings of the IEEE, 2019, 107(8): 1608-1631.
[34] H. Chen, X. Zhu, D. Qiu, et al. Scheduling for workflows with
security-sensitive intermediate data by selective tasks duplication in
clouds. IEEE Transactions on Parallel and distributed systems, vol. 28, no.
9, pp. 2674-2688, 2017.
[35] X. Chi, C. Yan, H. Wang, W. Rafique, and L. Qi. Amplified LSH-based
Recommender Systems with Privacy Protection. Concurrency and
Computation: Practice and Experience. DOI: 10.1002/CPE.5681, to be
published, 2020
[36] T. Zeugmann, P. Poupart, J. Kennedy, et al. Particle swarm optimization.
Encyclopedia of machine learning, pp. 760-766, 2010.

1551-3203 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Skovde. Downloaded on May 19,2020 at 07:00:59 UTC from IEEE Xplore. Restrictions apply.
View publication stats