A Survey on the Quality of Service

and Metaheuristic Based Resolution

Methods for Multi-cloud IoT Service
Selection

Ahmed Zebouchi(B) and Youcef Aklouf

RIIMA, USTHB, BP 32 El-Alia Bab-Ezzouar, Algiers, Algeria

{azebouchi,yaklouf}@usthb.dz

Abstract. The Internet-of-Things (IoT) generate increasingly signifi-

cant amount of data that needs to be stored and analysed. The use of
IoT devices as a service makes it more accessible and exploitable, this
could be achieved using of cloud computing. Multi-cloud service compo-
sition and selection are required to fulfill increasingly complicated user
requests for services. A service request is made from a cloud broker to
cloud providers (CP) to deliver the required Quality of Service (QoS).
Selecting services and optimizing service compositions to satisfy func-
tional and non-functional conflicting requirements across various cloud
service providers is an non-deterministic polynomial-time hardness prob-
lem (NP-hard). Multiobjective (MO) metaheuristics are known to be per-
formant to solve such a problem. This study examines how to select IoT
services to achieve the best performances on the eight selected QoS across
multiple CP. The experiment results reveal that among the 18 compared
algorithms, the parallel NSGAII provides the most efficient and optimal
outcomes.

Keywords: Internet of things · Multi-objective metaheuristics ·

Multi-cloud · Cloud brokering · Quality of service · Service
composition · Pareto front

1 Introduction

The IoT is a widely-adopted domain application. Nowadays, the internet of

things is becoming the trademark technology to provide specific services and
enable an intelligent world. The world will see a tripling of internet-connected
devices in the next decade to 30 billion by 2030 [9].
The IoT system should then exploit the advantage of cloud computing using
service oriented architecture [4]. Therefore, IoT can take advantage of the cloud
paradigm’s virtually unlimited capabilities and resources to compensate for its
technological constraints. Cloud service providers are increasing and offering
a wide range of services. Despite this, a given cloud provider likely won’t be
capable of meeting all of the consumer’s requirements [28]. The multi-cloud
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022
A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 412–424, 2022.
https://doi.org/10.1007/978-3-030-96299-9_40
 A Survey on the QoS and MO Metaheuristics for IoT 413

composition/selection is crucial to fulfilling a complicated user request, or when

the requested service is unavailable. The collaboration of several service providers
models the Broker-based service selection; this formulates what so-called Multi-
clouds IoT Environment.
What makes the request fulfillment more complex is that users generally focus
on some specific Quality of Service, like the cost of the service or response time,
with little or no consideration for the service location or energy efficiency. On the
other hand, the cloud broker focuses on cloud profit. These QoS are objectives
that the cloud broker goals to minimize or maximize. Some of these objectives
are conflicting, as in cost minimization and response time minimization. This
process is known for its complexity and is classified as an NP-hard problem.
Such problems are usually approached using metaheuristics, such as bio inspired
Pareto-based metaheuristics used for multiobjective problems.
Due to the worldwide popularity and success that metaheuristics have known
and the increasing publication counts of these studies. In this survey, we review
eighteen MO metaheuristic algorithms. We consider a need for a survey to review
and summarize the most appealing current studies on IoT service selection opti-
mization problems. Hence, as a first contribution, we have focused on studying dis-
tinguished metaheuristic algorithms that we term as the “new generation” meta-
heuristic algorithms. Also, through the studied papers, we noticed that most of the
researches on that issue has considered QoS optimization. Each study has only
focused on a few QoS, ignoring other important ones. In our research, we real-
ized the necessity of gathering all QoS considered in service selection to effectively
compare the existing solving methods and ensure the best service level agreement,
which constitutes the second contribution of this paper.
The rest of the paper is presented as follows. The next Section is dedicated
to raising the QoS that are considered in IoT service selection. In Sect. 3, we
expose a brief description of MO metaheuristic algorithms functioning. In Sect. 4,
we reveal the experimental results of applying the algorithms on IoT service
selection. Finally, Sect. 6 concludes the paper.

2 Quality of Services in IoT

Our work covers the QoS found in the most recent and relevant state-of-the-art
researches. We give a brief overview of each one below.

2.1 Energy Consumption

The required energy for IoT service computation is of high importance in the
QoS measurement. As mentioned in [11,24,26], it is considered a major issue in
cloud computing. Thus, service providers aim to complete the job with minimal
energy consumption efficiently. Assuming E(Sij ) is the energy consumption used
by service i to perform the job by cloud provider j and N is the number of
services included in the service composition. The service
N provider’s total energy
usage is computed within the equation: Eglobal = i=1 E(Sij )
414 A. Zebouchi and Y. Aklouf

2.2 Cost
With the rise of CP numbers and each one offers a tremendous number of ser-
vices, the cost of some services became a key element of the QoS. Thus, most
recent researches as in [24,32–34] focused on minimizing cloud prices. The cal-
culation ofthe global cost of a cloud service composition can be formulated as
N
Cglobal = i=1 C(Sij ). Where C is the cost of service i from provider j.

2.3 Profit
Profit is a monetary amount that cloud providers can be rewarded with. The
execution cost of the service is the fee that a tenant needs to pay for invoking
some operations. The broker is expected to profit from the task of finding the best
solution. The broker’s profit is, therefore, consideredN as a second goal. Profit is
considered in [3,10,31] and calculated by Pglobal = i=1 P (Sij ) − C(Sij ). Where
P (Sij ) is the consumer price of service i of the service provider j and C(Sij ) is
the service provider value of service i provided by cloud j.

2.4 Availability
Availability represents the uptime of cloud service during a specified time inter-
val. In [1,14,25], authors judged the Availability criteria as an impacting param-
eter that must be taken into account. The Availability A(Sij ) of service i in cloud
j is calculated as A = t/ts where t and ts represent the uptime and the total
time of service. As the value of A approaches Nto 1, availability increases. The
overall availability is calculated as Aglobal = i=1 A(Sij ).

2.5 Reliability
The Reliability calculates the level of assertion free of any cloud service software
or hardware fault. Several authors [1,14,23,34] took into account this require-
ment for its
relevance and its importance. The total reliability is calculated with
N
Rglobal = i=1 R(Sij ), where R(Sij ) function returns the reliability of service i
in cloud j, as the value of R approaches 1, the reliability increases.

2.6 Response Time

Response time is defined as the time required to send a request and receive the
response from the service. Numerous articles, for instance [1,14,25,32–34] have
been considering the customer query response time for their service selection
problem. Assuming L(Sij ) is the latency difference between customer request
time and service provider response time and ET (Sij ) is the execution time of
service i in cloud j. When a customer accepts a request from the service provider,
the service provider must spend ET (Sij ) time implementing the service request.
Thus, the goal is to minimize request response time (RT). The wording is given
N
in the next equation : RTglobal = i=1 L(Sij ) + ET (Sij ).
 A Survey on the QoS and MO Metaheuristics for IoT 415

2.7 Number of Clouds Involved in the Composition

The number of clouds included in the service composition (θ), including several
services from a large number of clouds, increases the inter-cloud communication,
which raises, in turn, the energy consumption and communication costs. As
mentioned in [2] is calculated following
|compositon|

θ := (α · Cloud(Si )) (α : boolean; 0 if Cloud already added, else 1).
i=1

2.8 Number of Clouds Involved Composition Plan

The number of cloud composition plans involved in the composition (γ),has been
considered as an effective QoS measure in [2]. Where each cloud has its own
predefined composition plan list that is already developed and optimized, incit-
ing composition to use a minimal number of cloud composition plans enhances
the response time and lowers the costs. The calculation of this QoS measure is
explained by
|compositon|

γ := (α · C plan(Si )) (α : boolean; 0 if C plan already added, else 1).
i=1

3 Multiobjective Metaheuristics

The multiobjective optimization with metaheuristics algorithms tested in this

project are the most used and suitable for the MO IoT service selection problem.
The algorithms are briefly described in the following sections. Multiobjective
optimization’s primary goal is to identify a set of solutions that are as close to
Pareto Front as possible. A solution here is a composition of IoT services from
multiple service providers which satisfies the client request. On the other hand,
the second goal is the solution set must be diverse and dispersed along the entire
optimal Pareto Front, in order to portray an accurate estimation.

3.1 Indicator-Based Evolutionary Algorithm

IBEA [35] is an indicator-based metaheuristic. This evolutionary algorithm’s fit-
ness assignment scheme is based on a pairwise comparison of solutions in a pop-
ulation using a binary quality indicator. A binary tournament among randomly
chosen individuals is used as a reproduction selection strategy. The replace-
ment technique involves deleting the worst individuals one by one and adjusting
the fitness values of the remaining solutions at each suppression. This process
is repeated until the required population size is met. In terms of convergence,
IBEA surpasses NSGAII. The computational Cost of the quality indicator value
is IBEA’s major weakness.
416 A. Zebouchi and Y. Aklouf

3.2 Parallel Speed-Constrained Multiobjective Particle Swarm

Optimization

pSMPSO [17] is a parallelized version of SMPSO which is the finest algorithm

in the context of the problems due to the velocity constriction mechanism. The
technique uses a standard dominance method that can be enhanced to generate
a population that converges more quickly in a multiobjective problem [15]. In
[27], the authors developed a parallel version of the algorithm using the master-
slave paradigm to parallelize the SMPSO evaluation operator. The algorithm
evaluates all of the solutions right after they have been created, and it does so
in parallel. The master process, which is implemented on the main execution
thread, assigns the solutions to be evaluated to each thread. The master process
then waits until all slave threads have completed their tasks and reported the
fitness values.

3.3 dual MO Particle Swarm Optimization Algorithm

dMOPSO is commonly used to solve MO problems. The usage of the NSGAII

crowding distance to filter out leader solutions is one of its primary characteris-
tics. The algorithm is the most effective method in terms of the quality of the
approximations to the Pareto front found. On the other hand, decomposition
approaches may worsen the algorithm’s performance, especially when working
with increasingly complicated situations [16].

3.4 Multiobjective Cellular Genetic Algorithm

MOCell is a multiobjective optimization technique based on cellular genetics.

The non-dominated individuals found throughout the search are saved in an
external archive by the algorithm. When it comes to problems with more than
two objectives, it’s a highly competitive method in terms of convergence and
hypervolume measures. The feedback of individuals from the archive to the pop-
ulation is the most notable characteristic of MOCell in comparison to other
existing cellular techniques [20]. However, it is possible to find a variant of the
algorithm that outperforms MOCell. Taking into account synchrony, archive
feedback, and replacement, Nebro et al. [19] offered several alternative versions:
Synchronous algorithms can be more effective in terms of hit rate [22].

– sMOCell1; It is the basic synchronous MOCell algorithm.

– sMOCell2; It consists of the original MOCell with archive feedback through
parent selection.
An asynchronous version can be obtained by updating the cells in sequential
order, using a unique population. Asynchronous algorithms can be more efficient
(faster) than synchronous ones.

– aMOCell1; Is the asynchronous version of MOCell

 A Survey on the QoS and MO Metaheuristics for IoT 417

– aMOCell2; It is a configuration of aMOCell1 where the worst neighbour of

each generation is replaced.
– aMOCell3; it is a configuration of asynchronous MOCell, archive feedback
through parent selection.

3.5 Non-dominated Sorting Genetic Algorithm II

NSGAII proposed in [5], is a fast and elitist multiobjective genetic algorithm
that can keep a wider variety of solutions and converge more quickly in the non-
dominated front. The present population is sorted by the number of solutions
that dominate each solution in NSGAII. As a result, a series of non-dominated
fronts with individuals of the same rank emerge. The algorithm then sorts each
of these fronts based on the distance between consecutive solutions, favouring
solutions in lightly populated regions of the search space before adding them into
the population of the following iteration. The algorithm uses selection, crossover,
and mutation to create a child population, then combined with the parent popu-
lation before the next iteration begins. It is noticeable that NSGAII has problems
with convergence closer to Pareto to the right front in some scenarios.

3.6 Parallel Non-dominated Sorting Genetic Algorithm II

Many approaches proposed to parallelize the NSGAII algorithm [6,12,23], the

most used and applicable, is the master-slave approach, which is the calcula-
tion of the objective formulate of each individual in the generation in parallel.
The master thread executes the NSGAII core code, and the agents evaluate the
objective functions of each new individual. Random selection of operators pro-
vides overall better results than the adaptive version in bi-objective problems,
while the latter outperforms the original NSGAII in three-Objective problems.

3.7 Non-dominated Sorting Genetic Algorithm II Steady State

A steady-state GA is an alternative to a generational GA, in which there is only

one population. In this manner, new individuals are instantly absorbed into
the evolutionary cycle, which allows the parents and offspring to dwell in the
same community. A steady-state variant of the NSGAII Durillo et al. proposed
in [7] converges better to the optimal Pareto front and spread, and it can be
simply implemented using a one-generation offspring population. The ranking
and crowding algorithms must be applied each time a new individual is created,
significantly increasing the algorithm’s computing complexity.

3.8 Non-dominated Sorting Genetic Algorithm II Random

NSGAIIr [21] is an extension of NSGAII that employs three different variation

operators: SBX, polynomial mutation, and DE’s variation operator. When a new
solution product is introduced, these operators are chosen at random. The key
418 A. Zebouchi and Y. Aklouf

difference from the original NSGAII is the selection method used by the parents,
and the way offspring are produced. One of the three variation operators is chosen
based on this value. The algorithm then operates like the original NSGAII once
the offspring has been formed.

3.9 Adaptative Non-dominated Sorting Genetic Algorithm II

Adaptative NSGAII (ANSGAII) [21] uses the variation operators in the same
manner that NSGAII Random does, but a more adaptive approach that considers
their contribution. In other words, each operator selection probability is altered
based on the operator’s success in the previous iteration. The adaptive strategy
for operator selection that is considered is based on the one used in AMAL-
GAM. Because it incorporates an adaptive mechanism, ANSGAII outperforms
the original algorithms in various experiments. ANSGAII can find solutions for
each occurrence that all fall within the first non-dominated front.

3.10 Multiobjective Evolutionary Algorithm

MOEAD [13] is an approach that combines Evolutionary Algorithms and the tra-
ditional mathematical Decomposition method. This hybridization makes apply-
ing the single objective optimizer to each sub-problem associated with a solution,
resulting in a more dispersed solution. The neighbourhood structure introduced
in MOEAD allows each sub-problem to be optimized using only the information
from its neighbours, greatly speeding up algorithm convergence. However, when
dealing with scaled problems, such as scaling challenges, its fundamental flaw is
the decrease of diversity and solution distribution.

3.11 Parallel Multiobjective Evolutionary Algorithm

The parallel version of MOEAD is proposed in [18], the authors chose the strat-
egy of distributing the population across multiple concurrent threads. This way,
each thread might work on a different segment of the population at the same
time. In most of the examined cases, the results show that there are no signif-
icant differences between the fronts calculated by MOEAD and pMOEAD in
terms of quality metrics.

3.12 Optimized Multi-Objective Particle Swarm Optimization

Optimized MOPSO [29] is a multiobjective particle swarm optimization tech-

nique. Its primary characteristics include using an external archive based on the
NSGAII crowding distance to filter out leader solutions. In a binary tournament,
the leaders are picked based on their crowding value, and only those with the
highest levels of overcrowding are maintained. According to the findings in [30],
OMOPSO outperforms the other MO algorithms. The idea of dominance is used
to limit the number of solutions saved in this archive.
 A Survey on the QoS and MO Metaheuristics for IoT 419

3.13 Multiobjective Random Search

Multiobjective random search (MORS) [8] random search is a single solution

search approach. The solutions are generated by randomly adding, removing, or
shifting one element of the solution to the preceding solution. For non-dominance
evaluation, the new solution is compared to the current Pareto optimal solutions,
and the Pareto optimal solution set is updated accordingly. The algorithm works
better for high dimensional, highly constrained nonlinear optimization problems;
it is challenging to outperform other MO algorithms when using it for many-
objective problems.

4 Experimental Evaluation

4.1 Dataset

As we could not find workloads fitting our purposes in the literature, we created
a generator of random multi-cloud environment’s configurations. We provide the
generator parameters, including the number of CP and the number of services
in the Multi-cloud environment. The script generates an XML file for each CP,
having randomly generated values, including QoS of each Service, Number of
composition plans, Number of available services in CP, Number of predefined
composition plans for every cloud, selected services in a CP, and selected services
from available services to be included in a composition plan of the CP.

4.2 Results

To effectively evaluate and compare the performance of each algorithm on our

dataset. All algorithms were run 50 times on the generated dataset of 100 cloud
providers and 1000 services for eight objective function evaluations, assuring
us that each algorithm performs the same amount of work. All experiments
are conducted on the same machine with Intel Core i7 2.60 GHz. Processor,
8 GB RAM, and Windows 10. The comparison of the algorithm performances is
made based on their quality metrics and execution time. The used metrics for
the evaluation of the convergence and diversity of the Pareto solution set are:
EPSILON, HV, and the Spread, such as the algorithms with a smaller value of
EPSILON, smaller value spread, and higher value of HV are the better-computed
fronts. Table 1 shows the normalized EPSILON, HV and Spread each have its
Mean Standard deviation and Median IQR of the obtained values from the runs
of the 18 Multiobjective algorithms considering the 8 QoS.
420 A. Zebouchi and Y. Aklouf

Table 1. Evaluation results.

NSGAII MOCell aMOCell1 aMOCell2 aMOCell3 sMOCell1 sMOCell2 pNSGAII ANSGAII
EPSILON.MeanAndStandardDeviation 4,37E+03 6,45E+03 6,45E+03 5,81E+03 6,22E+03 7,27E+03 6,19E+03 2,87E+03 6,33E+03
EPSILON.MedianAndIQR 4,37E+03 6,45E+03 6,45E+03 5,81E+03 6,22E+03 7,27E+03 6,19E+03 2,87E+03 6,33E+03
HV.MeanAndStandardDeviation 2,98E+05 3,50E+06 4,64E+05 2,39E+05 1,11E+05 4,34E+06 2,07E+05 3,00E+05 8,21E+05
HV.MedianAndIQR 2,98E+05 3,50E+06 4,64E+05 2,39E+05 1,11E+05 4,34E+06 2,07E+05 3,00E+05 8,21E+05
SPREAD.MeanAndStandardDeviation 3,72E+03 4,68E+03 4,90E+03 4,17E+03 4,04E+03 4,09E+03 3,48E+03 3,51E+03 3,92E+03
SPREAD.MedianAndIQR 3,72E+03 4,68E+03 4,90E+03 4,17E+03 4,04E+03 4,09E+03 3,48E+03 3,51E+03 3,92E+03
NSGAIIr NSGAIIss dMOPSO pSMPSO OMOPSO MOEAD pMOEAD IBEA MORS
EPSILON.MeanAndStandardDeviation 4,82E+03 2,65E+03 4,36E+03 4,54E+03 5,76E+03 8,80E+03 6,95E+03 5,17E+03 7,73E+03
EPSILON.MedianAndIQR 4,82E+03 2,65E+03 4,36E+03 4,54E+03 5,76E+03 8,80E+03 6,95E+03 5,17E+03 7,73E+03
HV.MeanAndStandardDeviation 7,72E+05 5,49E+05 6,29E+05 3,79E+05 3,08E+05 5,47E+05 4,91E+05 6,03E+05 1,68E+05
HV.MedianAndIQR 7,72E+05 5,49E+05 6,29E+05 3,79E+05 3,08E+05 5,47E+05 4,91E+05 6,03E+05 1,68E+05
SPREAD.MeanAndStandardDeviation 3,81E+03 3,91E+03 1,37E+02 4,60E+03 3,97E+03 6,96E+03 1,24E+02 5,23E+03 3,92E+03
SPREAD.MedianAndIQR 3,81E+03 3,91E+03 1,37E+02 4,60E+03 3,97E+03 6,96E+03 1,24E+02 5,23E+03 3,92E+03

Fig. 1. Multiobjective metaheuristics Fig. 2. Multiobjective metaheuristics

ranks comparison. execution time comparison.

For each metric, the mean is calculated from both the Mean Std and Median
IQR of that metric, which have very similar values, the algorithms are then
sorted according to each metric from the worst to the best algorithm, the higher
the metric, the better the algorithm. The sum slope represents the sum of the
ranks of each metric rank. The higher mean value of HV and the lower mean
value of SPREAD show a good approximation to the algorithm’s Pareto Front.
As shown in Fig. 1 and Table 1 pNSGAII has a higher level of 36 which suggests
that pNSGAII produce the best computed non-dominated front. Figure 2, shows
the means execution time for each algorithm; it shows that OMOPSO performed
the worst execution time, while pNSGAII overpassed pMOEAD by slightly two
seconds which makes pSNGAII having the best execution time among all the
tested algorithms.

5 Discussion
We can deduce several facts from the previous results. The outcomes of MOCell
and its derivatives were inconsistent, and it did not outperform many other
algorithms. sMOCell only brings a minor improvement to the solution quality
in comparison with other synchronous algorithms. The asynchronous versions
 A Survey on the QoS and MO Metaheuristics for IoT 421

aMOCell 1 and 2, performed better in terms of execution time. In contrast to

aMOCell3, which offered a better Pareto front and increased execution time a
bit compared to the original algorithm. We remark that OMOPSO algorithm
had the slowest execution time, four times that of the second MO algorithm.
Still, it showed potential in producing high-quality distributions. MOEAD and
pMOEAD had the best execution time; dMOPSO, IBEA, PSMPSO, and MORS
all produce non-satisfying results in terms of execution time, and did not perform
better solution quality.
The pNSGAII variants outperformed all other MO algorithms in terms of
metrics and execution time, the algorithm provided the best Pareto front qual-
ity and improved execution time over the original algorithm. The number of
objectives made it more challenging for other algorithms to meet the pNSGAII
quality metrics in a reasonable amount of time thanks to pNSGAII low compu-
tational needs. Mutation operators and appropriate population distribution were
two important reasons for this superiority since they prevented the algorithm
from becoming trapped in local optimum. The crowding distance operator for
pNSGAII played great during the selection and the population diversity between
generations was well conserved. Therefore, the pNSGAII is the most suitable for
improving multiobjective multi-cloud service selection with additional clouds
and objectives.

6 Conclusion
This paper has collected and presented the most targeted QoS for multi-cloud
Computing applied to IoT. Then, we proposed a survey on the most impor-
tant Multiobjective optimization methods that are based on metaheuristics. We
also implemented and compared their performances based on four well-known
metrics. Thus, we found out that the pNSGAII algorithm outperforms other
approaches and that it is the most adequate for the multi-cloud IoT service
selection issue. We hope that the findings in this paper will spark some thoughts
for new research works that offer more efficient methods in this field. In future,
we will extend this study to include other MO Algorithms such as games theory
and Deep Reinforcement Learning based approaches. Additionally, we plan to
work on the amelioration of the best Metaheuristic found, namely pNSGAII.

References
1. Asghari, P., Rahmani, A.M., Javadi, H.H.S.: Privacy-aware cloud service compo-
sition based on QoS optimization in internet of things. J. Amb. Intell. Human.
Comput. (2020). https://doi.org/10.1007/s12652-020-01723-7
2. Baker, T., Asim, M., Tawfik, H., Aldawsari, B., Buyya, R.: An energy-aware service
composition algorithm for multiple cloud-based IoT applications. J. Netw. Comput.
App. 89, 96–108 (2017). https://doi.org/10.1016/j.jnca.2017.03.008
422 A. Zebouchi and Y. Aklouf

3. Chauhan, S.S., Pilli, E.S., Joshi, R., Singh, G., Govil, M.: Brokering in intercon-
nected cloud computing environments: a survey. J. Parallel Distrib. Comput. 133,
193–209 (2019). https://doi.org/10.1016/j.jpdc.2018.08.001
4. Choudhary, G., Jain, A.: Internet of things: a survey on architecture, technologies,
protocols and challenges. In: 2016 International Conference on Recent Advances
and Innovations in Engineering (ICRAIE). IEEE (2016). https://doi.org/10.1109/
icraie.2016.7939537
5. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multi-objective
genetic algorithm: NSGA-II. IEEE Trans. Evol. Comput. 6(2), 182–197 (2002).
https://doi.org/10.1109/4235.996017
6. Durillo, J.J., Nebro, A.J., Luna, F., Alba, E.: A study of master-slave approaches
to parallelize NSGA-II. In: 2008 IEEE International Symposium on Parallel and
Distributed Processing. IEEE (2008). https://doi.org/10.1109/ipdps.2008.4536375
7. Durillo, J.J., Nebro, A.J., Luna, F., Alba, E.: On the effect of the steady-state
selection scheme in multi-objective genetic algorithms. In: Ehrgott, M., Fonseca,
C.M., Gandibleux, X., Hao, J.K., Sevaux, M. (eds.) Evolutionary Multi-criterion
Optimization. EMO 2009. Lecture Notes in Computer Science. LNCS, vol. 5467,
pp. 183–197. Springer, Berlin, Heidelberg (2009). https://doi.org/10.1007/978-3-
642-01020-0 18
8. Feng, J., Shen, W.Z., Xu, C.: Multi-objective random search algorithm for simul-
taneously optimizing wind farm layout and number of turbines. J. Phys. Conf. Ser.
753, 032011 (2016). https://doi.org/10.1088/1742-6596/753/3/032011
9. Hatton, M.: The IoT in 2030: 24 billion connected things generating $1.5 trillion
(2020). https://alhena.io/the-iot-in-2030-24-billion-connected-things-generating-
1-5-trillion/. Accessed 23 Sep 2021
10. Kumrai, T., Ota, K., Dong, M., Kishigami, J., Sung, D.K.: Multi-objective opti-
mization in cloud brokering systems for connected internet of things. IEEE Internet
Things J. 4(2), 404–413 (2017). https://doi.org/10.1109/jiot.2016.2565562
11. Lakhdari, A., Bouguettaya, A., Mistry, S., Neiat, A.G.G.: Composing energy ser-
vices in a crowdsourced IoT environment. IEEE Trans. Serv. Comput. 99, 1 (2020).
https://doi.org/10.1109/tsc.2020.2980258
12. Lancinskas, A., Zilinskas, J.: Approaches to parallelize pareto ranking in NSGA-
II algorithm. In: Wyrzykowski, R., Dongarra, J., Karczewski, K., Wasniewski, J.
(eds.) Parallel Processing and Applied Mathematics. PPAM 2011. Lecture Notes
in Computer Science, vol. 7204, pp. 371–380. Springer, Heidelberg (2012). https://
doi.org/10.1007/978-3-642-31500-8 38
13. Li, H., Zhang, Q.: Multi-objective optimization problems with complicated pareto
sets, MOEA/d and NSGA-II. IEEE Trans. Evol. Comput. 13(2), 284–302 (2009).
https://doi.org/10.1109/tevc.2008.925798
14. Liu, J., et al.: A cooperative evolution for QoS-driven IoT service composition.
Automatika 54(4), 438–447 (2013). https://doi.org/10.7305/automatika.54-4.417
15. Maltese, J., Ombuki-Berman, B.M., Engelbrecht, A.P.: A scalability study of
many-objective optimization algorithms. IEEE Trans. Evol. Comput. 22(1), 79–96
(2018). https://doi.org/10.1109/tevc.2016.2639360
16. Martı́nez, S.Z., Coello, C.A.C.: A multi-objective particle swarm optimizer based
on decomposition. In: Proceedings of the 13th Annual Conference on Genetic and
Evolutionary Computation - GECCO 2011. ACM Press (2011). https://doi.org/
10.1145/2001576.2001587
17. Nebro, A.J., Durillo, J., Garcı́a-Nieto, J., Coello, C., Luna, F., Alba, E.: SMPSO:
a new PSO metaheuristic for multi-objective optimization (2009)
 A Survey on the QoS and MO Metaheuristics for IoT 423

18. Nebro, A.J., Durillo, J.J.: A study of the parallelization of the multi-objective meta-
heuristic MOEA/D. In: Blum, C., Battiti, R. (eds.) Learning and Intelligent Opti-
mization. LION 2010. Lecture Notes in Computer Science. LNCS, vol. 6073, pp.
303–317. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-13800-
3 32
19. Nebro, A.J., Durillo, J.J., Luna, F., Dorronsoro, B., Alba, E.: Design issues in a
multi-objective cellular genetic algorithm. In: Obayashi, S., Deb, K., Poloni, C.,
Hiroyasu, T., Murata, T. (eds.) Evolutionary Multi-Criterion Optimization. EMO
2007. LNCS, vol. 4403, pp. 126–140. Springer, Heidelberg (2007). https://doi.org/
10.1007/978-3-540-70928-2 13
20. Nebro, A.J., Durillo, J.J., Luna, F., Dorronsoro, B., Alba, E.: MOCell: a cellular
genetic algorithm for multi-objective optimization. Int. J. Intell. Syst. 24(7), 726–
746 (2009). https://doi.org/10.1002/int.20358
21. Nebro, A.J., Durillo, J.J., Machin, M., Coello Coello, C.A., Dorronsoro, B.: A study
of the combination of variation operators in the NSGA-II Algorithm. In: Advances
in Artificial Intelligence. CAEPIA 2013. LNCS, vol. 8109, pp. 269–278. Springer,
Heidelberg (2013). https://doi.org/10.1007/978-3-642-40643-0 28
22. Olariu, S., Zomaya, A.Y. (eds.): Handbook of Bioinspired Algorithms and Appli-
cations. Chapman and Hall/CRC, Boco Raton (2005). https://doi.org/10.1201/
9781420035063
23. de Oliveira, L.B., Marcelino, C.G., Milanes, A., Almeida, P.E.M., Carvalho, L.M.:
A successful parallel implementation of NSGA-II on GPU for the energy dispatch
problem on hydroelectric power plants. In: 2016 IEEE Congress on Evolutionary
Computation (CEC). IEEE, July 2016. https://doi.org/10.1109/cec.2016.7744337
24. Pang, B., Hao, F., Park, D.S., Maio, C.D.: A multi-criteria multi-cloud service
composition in mobile edge computing. Sustainability 12(18), 7661 (2020). https://
doi.org/10.3390/su12187661
25. Singh, M., Baranwal, G., Tripathi, A.K.: QoS-aware selection of IoT-based service.
Arabian J. Sci. Eng. 45(12), 10033–10050 (2020). https://doi.org/10.1007/s13369-
020-04601-8
26. Sun, M., Zhou, Z., Wang, J., Du, C., Gaaloul, W.: Energy-efficient IoT service
composition for concurrent timed applications. Future Gener. Comput. Syst. 100,
1017–1030 (2019). https://doi.org/10.1016/j.future.2019.05.070
27. Toutouh, J., Alba, E.: Parallel multi-objective metaheuristics for smart communi-
cations in vehicular networks. Soft Comput. 21(8), 1949–1961 (2015). https://doi.
org/10.1007/s00500-015-1891-2
28. Vakili, M., Jahangiri, N., Sharifi, M.: Cloud service selection using cloud service
brokers: approaches and challenges. Front. Comput. Sci. 13(3), 599–617 (2018).
https://doi.org/10.1007/s11704-017-6124-7
29. Wang, H., Qian, F.: Improved PSO-based multi-objective optimization using iner-
tia weight and acceleration coefficients dynamic changing, crowding and mutation.
In: 2008 7th World Congress on Intelligent Control and Automation. IEEE (2008).
https://doi.org/10.1109/wcica.2008.4593644
30. Wang, H., Qian, F.: Improved PSO-based multi-objective optimization using iner-
tia weight and acceleration coefficients dynamic changing, crowding and mutation.
In: 2008 7th World Congress on Intelligent Control and Automation, pp. 4479–4484
(2008). https://doi.org/10.1109/WCICA.2008.4593644
31. Wang, W., Niu, D., Li, B., Liang, B.: Dynamic cloud resource reservation via cloud
brokerage. In: 2013 IEEE 33rd International Conference on Distributed Computing
Systems. IEEE, July 2013. https://doi.org/10.1109/icdcs.2013.20
424 A. Zebouchi and Y. Aklouf

32. Yang, C., Peng, T., Lan, S., Shen, W., Wang, L.: Towards IoT-enabled dynamic
service optimal selection in multiple manufacturing clouds. J. Manuf. Syst. 56,
213–226 (2020). https://doi.org/10.1016/j.jmsy.2020.06.004
33. Zhang, M., Liu, L., Liu, S.: Genetic algorithm based QoS-aware service compo-
sition in multi-cloud. In: 2015 IEEE Conference on Collaboration and Internet
Computing (CIC). IEEE, October 2015. https://doi.org/10.1109/cic.2015.23
34. Zhang, X., Geng, J., Ma, J., Liu, H., Niu, S.: A QoS-driven service selection
optimization algorithm for internet of things, September 2020. https://doi.org/
10.21203/rs.3.rs-69961/v1
35. Zitzler, E., Künzli, S.: Indicator-based selection in multi-objective search. In: Yao,
X., et al. (eds.) Parallel Problem Solving from Nature - PPSN VIII. PPSN 2004.
LNCS, vol. 3242, pp. 832–842. Springer, Heidelberg (2004). https://doi.org/10.
1007/978-3-540-30217-9 84