A Theoretical Discussion and Survey of Network Automation For IoT Challenges and Opportunity

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/353478255
A Theoretical Discussion and Survey of Network Automation for IoT: Challenges

and Opportunity
Article in IEEE Internet of Things Journal · August 2021

DOI: 10.1109/JIOT.2021.3075901
CITATIONS READS
47 2,663
6 authors, including:
Sisay Tadesse Arzo Claire Naiga Serugunda

University of New Mexico University of Bristol
28 PUBLICATIONS 275 CITATIONS 2 PUBLICATIONS 54 CITATIONS
SEE PROFILE SEE PROFILE
Fabrizio Granelli Riccardo Bassoli

Università degli Studi di Trento Technische Universität Dresden
375 PUBLICATIONS 5,012 CITATIONS 103 PUBLICATIONS 701 CITATIONS
SEE PROFILE SEE PROFILE
All content following this page was uploaded by Claire Naiga Serugunda on 30 September 2021.
The user has requested enhancement of the downloaded file.

IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 15, AUGUST 1, 2021 12021
A Theoretical Discussion and Survey of Network

Automation for IoT: Challenges and Opportunity
Sisay Tadesse Arzo , Graduate Student Member, IEEE, Claire Naiga , Student Member, IEEE,
Fabrizio Granelli , Senior Member, IEEE, Riccardo Bassoli , Member, IEEE,
Michael Devetsikiotis , Fellow, IEEE, and Frank H. P. Fitzek , Senior Member, IEEE
Abstract—The introduction of the Internet of Things (IoT) and presented possible future research directions and challenges
massive machine-type communications has implied an increase that need to be tackled to progress in achieving full network
in network size and complexity. In particular, there is already automation.
a huge number of IoT devices in the market in various sec-
tors, such as smart agriculture, smart city, smart home, smart Index Terms—Autonomic networking, machine learning,
transportation, etc. The IoT interconnectivity technologies are multiagent system (MAS), network function virtualization,
also increasing. Therefore, these are increasingly overwhelming network management system (NMS), network softwarization,
the efforts of network administrators as they try to design, software-defined networking (SDN).
reconfigure and manage such networks. Relying on humans
to manage such complex and dynamic networks is becom- I. I NTRODUCTION
ing unsustainable. Network automation promises to reduce the
NTERNET of Things (IoT) is expected to interconnect
cost of administration and maintenance of network infrastruc-
ture, by offering networks the capability to manage themselves.
Network automation is the ability of the network to manage
I almost everything. It has various areas of applications,
such as smart cities, smart homes, smart agriculture, and
itself. Various standardization organizations are taking the ini- smart transportation. It brings immense value into our lives,
tiative in introducing network automation, such as European
Telecommunication Standardization Institute (ETSI). ETSI is touching areas, such as health monitoring, green energy, and
leading the standardization activities for network automation. environment monitoring. IoT connects sensors, cameras, cars,
It has provided different versions of reference architecture etc., creating an automated environmental ecosystem. The IoT
called generic autonomic network architecture (GANA), which devices communicate with each other using IoT connectiv-
describes a four-level abstraction for network-management deci- ity technologies. Therefore, measurement values and other
sion elements (DEs), protocol level, function level, node level, and
network level. In this article, we review and survey the exist- information traverse between IoT devices so that the over-
ing works before and after the introduction of software-defined all processes can be automated, without the need for human
networking (SDN) and network-function-virtualization (NFV). intervention.
We relate the main trending paradigms being followed, such as There are various types of IoT connectivity technolo-
SDN, NFV, machine learning (ML), microservices, multiagent gies, such as Lora, wireless fidelity (WiFi), narrowband-IoT
system (MAS), containerization, and cloudification, as a piv-
otal enabler of full network automation. We also discuss the (NBIoT), etc. While IoT is helping automate our environ-
autonomic architectures proposed in the literature. Finally, we ment, it is necessary to manage the network communica-
tion and connectivity, which interconnect the IoT devices.
Manuscript received December 4, 2020; revised February 16, 2021 and Network automation provides the automated management of
March 26, 2021; accepted April 22, 2021. Date of publication April 26, 2021; IoT networks.
date of current version July 23, 2021. This work was supported in part by the
NATO Science for Peace and Security (SPS) Programme in the Framework The need for network automation is more pronounced when
of the Project “Dynamic Architecture based on UAVs Monitoring for Border we consider the requirements of a future network, such as IoT.
Security and Safety” under Grant SPS G5428, and in part by the German As indicated above, IoT is the interconnection of highly het-
Research Foundation (DFG, Deutsche Forschungsgemeinschaft) as part of
Germany’s Excellence Strategy—EXC 2050/1—Project ID 390696704— erogeneous networked entities and the networks that follow
Cluster of Excellence “Centre for Tactile Internet with Human-in-the- several different communication patterns, such as machine to
Loop” (CeTI) of Technische Universität Dresden. (Corresponding author: machine (M2M), human to human (H2H), human to thing
Sisay Tadesse Arzo.)
Sisay Tadesse Arzo, Claire Naiga, and Fabrizio Granelli are with (H2T), thing to thing (T2T), or thing to things (T2Ts). With
the Department of Information Engineering and Computer Science, newer wireless networks, superior sensors, self-healing soft-
University of Trento, 38122 Trento, Italy (e-mail: [email protected]; electronics, and revolutionary computing capabilities, the IoT
[email protected]; [email protected]).
Riccardo Bassoli and Frank H. P. Fitzek are with the Deutsche Telekom is the frontier in the race for technological advancement. IoT
Chair of Communication Networks, Institute of Communication Technology, also has a huge potential to revolutionize space exploration.
Faculty of Electrical and Computer Engineering, and the Centre for In all of these applications, network automation plays an
Tactile Internet with Human-in-the-Loop, Cluster of Excellence, Technische
Universität Dresden, 01062 Dresden, Germany (e-mail: riccardo.bassoli@ important role.
tu-dresden.de; [email protected]). In general, in the past decades, communication systems have
Michael Devetsikiotis is with the Department of Electrical and Computer experienced several technological advancements that have
Engineering, University of New Mexico, Albuquerque, NM 87131 USA
(e-mail: [email protected]). facilitated an increase in the use of the Internet and the number
Digital Object Identifier 10.1109/JIOT.2021.3075901 of devices [1]. As a result of this development, networks are
2327-4662 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Makerere University Library. Downloaded on August 25,2021 at 09:10:10 UTC from IEEE Xplore. Restrictions apply.
12022 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 15, AUGUST 1, 2021
evolving, showing an increase in network size and complexity.

In parallel, networks’ hardware and computational complexity
are also augmenting. However, this makes network manage-
ment very expensive and humanly infeasible. As networks
become bigger and bigger and more complex, there is a
need for more efficient network management methods. The
challenges of managing complex networks are numerous.
Autonomic networking is required to scale up the network
management capability to address the expected big growth
in the next-generation networks. Moreover, different strin-
gent performance indicators should concurrently be satisfied,
such as latency, reliability, and service continuity. This also
requires increasing expenditure and inefficiency as the network
size expands. In other words configuration, monitoring, con-
trol, and management of such networks become complex and
costly [2]. This means, from a business perspective, network
operators need to increase revenues, while reducing capital
(CapEx) and operational expenditures (OpEx).
Moreover, a dramatic increase in the amount of network
information will come from network softwarization and addi-
tional new verticals, such as ultrareliable massive machine-
type communications (mMTC). Due to this, a completely new Fig. 1. SDN architecture.
mechanism for network management is required. Therefore,
future networks [3] should mainly target alleviating the burden
through autonomic networking. deviation from the set of expectations and predefined-
Originally, network automation hails from a manifesto ideal standards.
by International Business Machines Corporation (IBM) in 3) Self-protecting is related to the security of the network
2001 [4]–[6], which was applying self-management on com- and it is a very crucial issue. Self-protection ensures that
puting. Through network automation, the network can adapt the network is able to shield itself against any potential
to any change in its environment, self-configure, and self- attacks, such as Denial-of-Service (DoS) attacks.
optimize without any human intervention [4], [7]. As defined 4) Self-healing is needed in case of failure of any network
by IBM, autonomic systems have self -“*” properties [1], [5], element. A network with self-healing capabilities is able
where the “∗” means configuration, optimization, healing, and to recover from such failures in the shortest time possi-
protection. ble. The network is able to discover and automatically
In particular, this can be through an autonomic control repair any failed elements to ensure service continuity.
loop (ACL) that could be implemented to collect detailed Current networking research mostly focuses on network
information about the network. Network data, along with data softwarization, relying on software-defined networking (SDN)
management, and users’ metadata, could be extracted and used and network-function-virtualization (NFV). Softwarization
to manage the network. Collecting such network data from maps hardware-based network functions into software. It is
big networks is very challenging but critically useful to pro- an important step in the automation process. SDN involves
vide dynamic and adaptive responses to the evolving service the separation of the control plane from the forwarding (data)
demands. Networks with these properties take their actions plane [8], [9]. The forwarding plane is made up of state-
based on a set of predefined policies. This can be overridden less forwarding tables, that are periodically populated by the
by the human network administration if needed. centralized controller, see Fig. 1 for the SDN architecture.
With a typical automation method, such as monitoring, ana- On the other hand, NFV is a software implementation
lyzing, planning, and executing methods, the network can of network function, which is traditionally implemented in
self-adapt to any changes in its environment. The follow- preparatory hardware [10]. The architecture of NFV contains
ing self-management properties [4] are described in IBM’s three main components, see Fig. 2 for the NFV architecture.
manifesto. 1) NFV infrastructure (NFVI) consists of the hardware and
1) Self-configuring is the capability of a network to config- software that host different virtual network functions
ure and reconfigure itself based on predefined policies (VNFs).
to achieve a given performance. This should happen 2) VNFs are softwarized network functions, such as fire-
seamlessly and with no human intervention. wall, network address translation (NAT), packet/serving-
2) Self-optimizing means to ensure that the network always gateway(P/S-G), and baseband unit (BBU), which could
uses the available resources to provide the best possible be deployed in an NFVI environment.
performance even in highly varying environments. The 3) NFV management and network orchestration (MANO)
network should always measure its current performance is the place where management and orchestration of
and set strategies to efficiently perform in case of any VNFs are implemented.
ARZO et al.: THEORETICAL DISCUSSION AND SURVEY OF NETWORK AUTOMATION FOR IoT: CHALLENGES AND OPPORTUNITY 12023
TABLE I
C OMPARISON OF O UR S URVEY W ITH THE E XISTING S URVEY ON
AUTONOMIC N ETWORKING
Fig. 2. High-level NFV framework [10].
surveys, and several past and present solutions in networking.

This article also indicates areas that are partially unexplored
but relevant automation issues, particularly from the commu-
nication perspective. Our survey addresses the research gap
by providing the current trends in network automation in the
era of network softwarization while also comparing them with
network automation trends before the arrival of SDN and NFV.
In the current state of the art, several earlier studies, such
as [12] and [13], (among others) surveyed issues related to
network automation but without considering SDN and NFV.
Fig. 3. Automation enabling technologies.
While other studies [14]–[16] have focused more on specific
scenarios, such as security and handover. Additionally, some
SDN and NFV enable the traditional static network to be publications [16], [17] have surveyed specific architecture
flexible, paving the way for network innovation. This opens without considering the current networking trends.
the door for rapid network evolution, transition to intelligent To the best of our knowledge, despite such previous works,
networking, and network automation. Today, the success of there is a lack of a survey combining SDN, NFV, and ML
the network automation industry is motivated by SDN, NFV, while showing the current trends, such as containerization
and machine learning (ML). Even if SDN and NFV enable and cloudification of network functions in a distributed cloud
the possibility of network programming, virtualization, and environment. Moreover, this survey article contributes to the
orchestration, they do not automate network management. relevant theoretical results. It presents a systematic review and
Therefore, a new framework should be developed to incorpo- organization of the research work in the network automation
rate SDN for network control and programmability, NFV for area in the last (almost) two decades.
virtualization and orchestration, and lastly, ML for knowledge Furthermore, our article provides an overview of the fol-
management and cognitive ability. Fig. 3 depicts a combination lowing topics in the context of network automation: network
of SDN, NFV, and ML for network automation. softwarization, SDN, NFV, cloudification, containerization,
ML has recently seen great advancements. It is expected microservices, multiagent-based system, fifth generation (5G),
to play a significant role in the automation process [11]. In and beyond network service requirements. We have classi-
fact, ML has been applied in various areas of networking, fied the literature as before and after the introduction of
such as traffic prediction, resource management, Quality of network softwarization. We have also surveyed some network
Service (QoS), and network security. ML provides cognition automation standardization efforts and automation architec-
and reasoning in automated decision making. The goal of ML ture. Moreover, we have identified and provided the research
in networks is to extract knowledge from network behavior, direction that should be followed in addressing the challenges
service, and users’ behavior, using historic data for training of full network automation. As a summary of our contribu-
and learning. tion, a comparative table with some representative surveys
In this article, we pioneer a discussion about network are presented in Table I. In general, this article presents the
automation, that will give the reader an understanding and following.
an overview of the different concepts in relation to network 1) Classification and comprehensive literature survey of
automation. We discuss the concerns that lead to the need for network automation into two categories namely; before
network automation. Even if the issue of network automation and after the introduction of SDN and NFV.
has been the subject of active and intense research, the con- 2) Discussion of challenges and opportunities of network
tribution of this article is a thorough comparison of previous automation.
Fig. 4. Organization of this article and possible navigation orders.
3) Discussion of conceptual and technological frameworks

for the autonomic network, such as SDN, NFV, network
function decomposition, network function atomization,
Fig. 5. SDN management architecture with respect to autonomic
microservice, multiagent system (MAS), virtualization, management [9].
containerization, and orchestration.
4) Identification of possible future research directions.
The remainder of this article is organized as follows. This In other words, the flexibility and dynamic programmability
article’s logical organization is depicted in Fig. 4. The figure of control enable network automation.
can also be used as a possible guild to the reader to go through Fig. 5 illustrates the SDN architecture. The data plane con-
this article, following the directions of arrows. The sections are sists of all network devices, such as switches, routers, and
organized as independently as possible considering the diverse firewalls. The control plane and the data plane communicate
interests of the target readers. Section II presents network over the south-bound interface (SBI) [18]. The concept behind
softwarization and 5G service requirements. Section III dis- SDN was to use flow tables within network devices and a com-
cusses an overview of network management systems (NMSs). mon interface for configuring, controlling, and manipulating
Various network automation standardization efforts are dis- the flow table. OpenFlow is the standard protocol employed
cussed in Section IV. Network automation architecture models at this interface. OpenFlow protocol is used by the centralized
are discussed in Section V. Section VI reviews existing liter- controller to manipulate entries in flow tables. Functionalities,
ature by classifying them before and after the introduction such as fault detection, the discovery of topology changes,
of SDN and NFV. A discussion of technological trends in etc., are performed by the SDN control plane [18]. The control
networking is presented in Section VII. The challenges and and the application planes communicate over the north-bound
opportunities that motivate the need for network automation interface (NBI). Over this interface, applications are able to
are presented in Section VIII. Finally, concluding remarks are exchange information with the control plane about the status
made in Section IX. of the network devices.
Li et al. [8] and Qi et al. [22] have proposed a number
II. OVERVIEW OF N ETWORK AUTOMATION E NABLERS of SDN-based architectures for autonomic management. For
As discussed in the introductory section, numerous research example, Qi et al. [22] proposed an SDN-based network virtu-
areas of networking are leading to network automation. These alization architecture with autonomic management (SDNVA)
are network softwarization through SDN and NFV, and ML. with two layers: 1) autonomic virtual physical virtual network
This section provides an overview of network automation management. The major objective of this approach is to
enabling technologies and 5G requirements. ensure isolation between the virtual network resources using
a virtualization module and also assert a hierarchical auto-
A. Software-Defined Networking nomic management approach among the physical and virtual
resources. Binsahaq et al. [18] discussed the current state-
Now let us see what SDN brings to network man- of-the-art literature on SDN about QoS management. The
agement. SDN is a promising solution in network literature discussed in their work is classified based on the
automation [8], [21], [22]. This is because, SDN involves the monitor, analyze, plan, execute, knowledge (MAPE-K) ref-
separation of the control plane and the data plane, which erence model [5] serve as basic functions, monitor, analyze,
significantly decreases the difficulties experienced in control- plan, execute, and knowledge.
ling networks. The control plane is implemented in a central
controller while the hardware functions, such as routing and
switching can be abstracted from hardware and implemented B. Network Function Virtualization
in software [23]. This allows for a range of considerably more NFV is another key component of network softwarization.
flexible and effective network management solutions, using Similar to SDN, NFV provides flexibility and programmabil-
network programming. This improves the role of network ity through dynamic orchestration of network functions. The
management in the event of any dynamic network changes. idea of NFV is decoupling network functions from hardware
This gives room for autonomic network management (ANM). and deploying them as software in virtual machines (VMs)
Fig. 7. Unified SDN-NFV architectural framework.
Fig. 6. ETSI-NFV architectural framework [10].

management and control. They are enablers of network
automation through the softwarization and orchestration of
network functions.
By rendering the network programmable, SDN enables
or containers [21]. NFV originated from a white paper [24]
accelerated innovation, contributing to greater responsive-
that was presented at a conference in Darmstadt, Germany, on
ness, stability, productivity, and cost effectiveness. Since NFV
SDN and OpenFlow in 2012. The architecture of NFV is illus-
replaces devices, such as load balancers, firewalls, and intru-
trated in Fig. 6. It represents the three main working domains
sion detection systems with software running on commodity
as 1) VNF; 2) NFVI; and 3) MANO [10].
hardware, it reduces the expense of installing the network. The
A VNF is software that implements a given network func-
forwarding, control, and management functions may also be
tion. VNFs run in one or more virtual environments (e.g.,
implemented in software as in NFV [2]. These consolidated
VM, containers, etc.) that run on general hardware. VNFs are
functions may then share the same resources, such as com-
deployed on-demand using an NFV architecture, removing the
puting, storage, and power. This reduces power consumption,
delivery delays associated with traditional network equipment,
maintenance costs and takes time to implement new services.
as well as the need for on-site specialist knowledge, when
Examples of functions that may be virtualized include switch-
deployed remotely [10]. NFVI includes physical and virtual
ing, routing, traffic load balancing, etc. Management software
network resources, hypervisors, VM, and virtual infrastruc-
that operates on converged SDN and cloud infrastructure dra-
ture managers (VIMs). It also includes the physical resources
matically reduce CapEx and OpEx, as well as the overall
for deploying and handling VNFs. The NFVI requires the
complexity of network operation [21].
virtualization layer above the hardware, which abstracts hard-
Moreover, automating the NMS can supplement SDN and
ware resources into virtual resources that support VNFs. The
NFV in harnessing the full benefits of the two technologies.
NFVI is also essential to the development of large, widely
For instance, some of the important benefits of network slic-
dispersed networks without the regional limitations associated
ing are network and service isolation. Enabling network slicing
with conventional network architectures [10].
requires dynamic network functions orchestration and dynamic
MANO is responsible for the management and orchestration
resource allocation depending on the service requirements.
of the virtualized network. NFV MANO consists of three mod-
An automated provisioning of resources and orchestration of
ular components: NFV Orchestrator manages VNF enrollment,
network functions would enable satisfying the dynamic ser-
lifecycle management, global view of resource management,
vice demands. A unified architecture for SDN-NFV, which is
and NFVI resource request validation and authorization. VNF
presented in [25] is depicted in Fig. 7.
Manager manages the instance management of the VNF life-
cycle, provides collaboration and adaptation function for the
configuration of NFVI and element/NMSs, and the recording D. 5G Service Requirements
of events. VIM controls and manages resources for the NFVI 5G and beyond networks are expected to have a hetero-
computing, storing, and networking. geneous environment made up of not just mobile phones
but different devices that support a wide number of appli-
cations. Some applications require a large number of devices
C. Network Softwarization or simultaneous connections at the same time while others
As we have introduced in the above section, SDN and require low latency in order to offer efficiency and reliability.
NFV are the two building blocks of network softwariza- 5G services are divided into three main categories [26], e.g.,
tion [21]. They are complementary technologies. Adopting enhanced mobile broadband (eMBB), ultrareliable low-latency
them together provides a myriad of innovative possibilities, communications (URLLC), and mMTC.
such as dynamic network slicing, dynamic network config- eMBB services require high data rates. Some of the use-
urations, network state measurement, and dynamic network cases include large-scale video streaming and virtual reality.
This type of service requires an enhancement of the existing

fourth-generation (4G) long-term evolution (LTE) broadband
data rate capability. URLLC services include autonomous driv-
ing, remote surgery, and industry control with robots. This
second category of 5G services targets very low-latency and
ultrareliability since it will support verticals, such as remote
surgery and automated driving. These services are also referred
to as mission-critical services because they will provide an
infrastructure for emergency-governmental communications.
mMTC services include billions of devices connected at the
same time. This third category of 5G services includes devices,
such as those in the IoT, smart cities, and smart buildings.
The main performance requirement for mMTC networks is
the need for highly dense connectivity (about 1 million devices
per km2 ).
Due to the heterogeneous nature of 5G and beyond
networks, stringent service requirements are targeted, e.g.,
high data rates, connectivity, and low latency. Thus, man-
agement and control of such networks are very difficult, the
CapEx increases with the need for connectivity, and more Fig. 8. Cyclic process of network management.
devices or sensors. Moreover, OpEx increases due to the con-
stant need for human intervention to manage the networks.
As previously mentioned, this motivates the need for network efficient utilization of network resources would be needed for
automation in 5G environments. the next phase of a multitenant network.
A. Network Management and Control as Cyclic Process

III. OVERVIEW OF N ETWORK M ANAGEMENT S YSTEMS Network management and control is a cyclic process that
AND N ETWORK AUTOMATION starts from observing the environment and ends the cycle to
Communication networks are very complex systems with acting and changing also influenced by the variations in the
numerous heterogeneous devices, services, and users involved. environment. Fig. 8 depicts a general network management
This is more pronounced in automating networks intercon- process.
necting IoT devices, as there will be a huge number of 1) Measuring and understanding the environment is
interconnected devices. For example, there are 20.6 bil- performed through observing and taking some mea-
lion connected devices in 2020 [27]. Thus, managing such surements to capture any behavioral change. Various
networks involves handling real-time events in a sophisti- techniques could be employed to observe the environ-
cated and heterogeneous environment. The problem is also ment. These techniques could be independent of the
complicated by the continued and huge amount of data gener- environment to observe and the parameter to be mea-
ated by IoT devices, users, services, and networks. On the sured. It could be sampling the instance of events or
other hand, even if it is cumbersome to manage this vast continuous measurement of network events and activi-
amount of data, data also contains valuable information about ties. Observing and measuring the environment provides
users, services, and network status, that could be extracted important information about the status of the network
for effective network management. Information could be used being managed. For instance, measuring the bandwidth
to facilitate the provisioning of services. This could happen across a given link or path would provide valuable
through systematic management and analysis of data to extract information that could be used in the routing decision of
valuable information to self-manage the network. new traffic and services. It could also be used to monitor
Legacy NMSs were mostly implemented in hardware but and guarantee the required QoS for a given service.
today, an increasing number of them have been implemented 2) Decision-making process is the process of extracting
in software. These systems are responsible for the good health and utilizing valuable information for the final decision.
of a network [28]. They monitor, maintain, and optimize the For instance, a decision could be made for routing of
network. NMSs provide a number of functionalities, such new traffic or rerouting of ongoing services by observ-
as fault detection, device management, performance analy- ing the current network state or amount of traffic in a
sis. Network management raises new problems, that need given link or path for optimal or maximum resource uti-
to be tackled to ensure a robust and secure networking lization depending on the policy stated by the network
system is completely achieved. The trend today is that future administrator. The decision-making process is the most
networks should be more independent, requiring almost no critical and complex part of the network management
human intervention for healing, protection, optimization, and process. Depending on the decision, it may involve per-
(re-)configuration. This results in ANM system [7]. It is gen- forming complex optimization algorithms including ML
erally agreed that a greater degree of service knowledge and techniques.
3) Planning action strategy is the procedural step needed networks. A dynamic inventory is also architected to be able
in order to execute the decisions, such as what kind to consume dynamically changing network data, providing a
of actions or configurations should be performed on real-time view to users and the wider system ecosystem. This
an element to have the required network behavior. For inventory architecture enables the processing of prodigious
example, once the decision is made to allocate resources, amounts of data about network state, performance, and health.
a configuration file needs to be prepared for execu- Recently, ETSI has taken a prominent role in auto-
tion in the network elements or devices. This allocates nomic architecture standardization and adaptation. An Industry
the required network resources, as agreed upon in the Specification Group (ISG) on autonomic network engineering
decision-making process. for the self-managing future Internet (AFI) has been estab-
4) Verification of planned action is the process of validat- lished under the ETSI [31]. Various versions of this model
ing the plan before executing it on the target network presented since its inception [32]–[34]. More recently, ETSI
elements or devices. These could be configuration files developed generic autonomic network architecture (GANA),
or steps, which have to be verified and checked for an autonomic network standard reference architecture [32].
accuracy and consistency. It is very important to ver- In [35], SDN enablers in the ETSI AFI GANA Reference
ify the action as the execution of the action may have Model for Autonomic Management & Control (emerging stan-
undesirable consequences on the target device and/or dard), and Virtualization impact. In [36], the implementation
the network behavior in general. Since the network is guide for the ETSI AFI GANA model is presented consider-
a dynamic environment and the process of settling for a ing a standardized reference model for autonomic networking,
given decision may take more time, the decision may be cognitive networking, and self-management. A Zero-touch
outdated or no longer useful, or suboptimal depending network and Service Management ISG (ZSM ISG) was estab-
on the scenarios considered. lished by ETSI. The ETSI ISG ZSM group was created
5) Executing the planned action is done through imple- to resolve the 5G E2E Network Slicing management issue.
menting the verified and accurate final action plan. It The visions are to enable full automation in terms of deliv-
could be a configuration path because of a requested ery, deployment, configuration, assurance, and optimization of
service, reservation of end-to-end network resources, network services. An example of an ongoing initiative recently
service isolation, or a part of QoS provisioning and launched in ETSI, [37] provides standardization of resilience
guaranteeing. & survivability, and autonomic fault-management, in evolving
6) Finally, monitoring the system behavior is analyzing the and future networks.
effect of the action on the overall system behavior. This Richard et al. [38] discussed a data and knowledge-driven
could be through loop-back control. intelligent network for future telecom transformation. The
In the decision organization of any network, control loops white paper presents the challenges and opportunities along
are a very essential part of a network administrator is with the autonomous network framework. In [39], the authors
to employ network automation. They provide information provided a specific full autonomic architecture for mobile
throughout different parts of the network in question thereby network realization relaying on NFV and SDN. In this work,
facilitating self-adaptation and later, self-management [29]. they extended the work in [39], providing a parallel auto-
They facilitate data collection from which decisions are made. nomic control plan for an SDN-NFV architecture. In the
However, loop-back control problems arise together with the proposed architecture, the control plane has a logical connec-
increase in the cost and complexity of the system. Another tion between different planes for data gathering and control
challenge exists, in the case of policy-based decisions where decisions exchange.
conflicts can arise from two policies with similar conditions GSMA [40] provided a high-level discussion for network
but different actions. automation. It is suggested that operators can use ML to
automate network operations, improve network planning and
strengthen security. The discussion is limited to the application
IV. S TANDARDIZATION OF N ETWORK AUTOMATION of either fixed rule automation or ML-based or a combina-
Significant standardization efforts are also taking place by tion of both. ML will enable cognitive functions that have
prominent standardization organizations, such as European not been possible before, supporting predictive maintenance,
Telecommunication Standardization Institute (ETSI), long-term network optimization, network planning, security,
International Telecommunication Union (ITU) Internet and deployment automation.
Engineering Task Force (IETF), 3rd Generation Partnership The first project launched to develop a network automa-
Project (3GPP), tmForum, and Global System for Mobile tion model is autonomic network architecture (ANA) [41]. The
Communications Association (GSMA). project is aimed at designing an autonomic architecture to mit-
3GPP [30] is also undertaking a standardization effort for igate network complexity enabling protocols and algorithms to
network automation. Two versions of specification by 3GPP operate in an autonomic manner.
has been put forward about the management and orchestra- An Internet Research Task Force group [42] is working on
tion which specifies the levels of the autonomous network. standards for automated network management which aims to
This standard specification, TS 28.100, specifies how software improve and make more efficient management of networks as
enables to dynamically create, deploy, and automate the end- they continue to increase in size and complexity. IETF is also
to-end delivery of services across both physical and virtual working on standardizing protocols for autonomic networking,
Fig. 10. Overview of the GANA reference model.
be important to self-* properties implementation. To iden-

tify and forecast any incidents of interest to the network, the
data collected from these measurements will then be analyzed.
Fig. 9. MAPE-K reference model.
Adaptation plans are then drawn based on the data analyzed.
Plans to configure the network and communicate the network
behavior to the autonomic manager are executed based on
which aims to enable networks to be self-managing, including
predefined policies. The knowledge base is used as storage
self-configuration, self-optimization, self-healing, and self-
for calculated data, plans, and predefined policies [18].
protection. Research on automated network management in
the Internet Research Task Force resulted in RFC7575 and
RFC7576 and led to the creation of the IETF autonomic
networking integrated model and approach (ANIMA) work- B. Generic Autonomic Network Architecture
ing group, which is in the process of defining protocols for Following IBM’s reference model, different standardization
autonomic network infrastructure. communities [6], [31] came up with proposals for different
The network 2030 architecture framework document (ITU- architecture for autonomic networks. As a leading group in the
T IMT-0040) [43] (ITU), discussed an automated arrangement, standardization of autonomic management & control (AMC),
governing, coordination of complex network systems and the AFI working group in ETSI’s network technologies tech-
functions including middleware for both physical and virtual nical (NTECH) [31] committee proposed GANA in Fig. 10 as
infrastructures. It is often discussed as having an inherent intel- a reference model for network automation [32].
ligence or even implicitly autonomic control. Orchestration The GANA reference model comprises of decision elements
results in automation with control network systems. (DEs) and managed elements (MEs) [31]. These together form
a hierarchical control loop (HCL) and the policies from the
superior elements are termed autonomic behaviors (ABs). The
V. A RCHITECTURES FOR N ETWORK AUTOMATION
DEs are further divided into four: 1) protocol level; 2) node
A. MAPE-K Reference Model level; 3) function level; and 4) network level [32].
Autonomy is achieved through the interaction of different 1) Network-Level DE: This is the highest level of the HCL.
network elements to met set goals without any human interven- It is responsible for the overall network functionality and
tion. Through IBM’s manifesto [1], an ACL exists 12 between behavior. To enable self-management in the network,
these network elements (managed components where monitor- the network-level DE (also known as the QoS manager)
ing of data is performed, the data is then analyzed, tasks are distributes ABs to the node-level DE.
then planned and the final decisions are executed. This process 2) Node-Level DE: Since this is the self-management func-
is the foundation of the MAPE-K reference model [5], [44] as tionality level, the DE interprets the ABs received from
illustrated in Fig. 9. the Network-level DE and passes them to the function-
The autonomic manager is located outside the ACL and per- level DE. The functionality and behavior of all nodes
forms the overseeing of the interactions between the network can be evaluated here.
elements (the managed components) to achieve the high-end 3) Function-Level DE: This then generates instructions and
objectives [18]. A managed component may be a hardware or passes them on to the protocol level DE. This level
software function, such as router, firewall, database, storage abstracts mechanisms and processes related to network
device, etc. functions, such as routing, forwarding, mobility, etc.
First, the monitoring component must take measurements 4) Protocol-Level DE: This is the lowest level of the HCL.
to capture any changes in the network environment that could This follows instructions from the function-level DE and
Fig. 11. ANIMA reference model and the structure of an AN.
implements the specified functions, e.g., scheduling. The

protocols here are associated with self-management.
C. Autonomic Networking Integrated Model and Approach

Reference Model
The GANA model is the inspiration behind most of the con-
cepts in the reference model proposed by ANIMA [6]. The
autonomic network shown in Fig. 11 comprises autonomic
nodes (ANs) and autonomic domains. Each AN asserts its
membership in an autonomic domain with a globally unique
domain certificate (a logical device identifier). Each AN
is comprised of autonomic networking infrastructure (ANI),
autonomic service agents (ASAs), and basic operating system
(OS) functions.
Autonomic functions (AFs) typically span several, possibly Fig. 12. Autonomic architecture with SDN-NFV [45].
all ANs in the network. AF refers to a function or feature that
can rely on self-knowledge, discovery, and intent to acquire
the information needed for operations without external con- elements as depicted in Fig. 12, which based is on uni-
figuration. The atomic entities of an AF are called the ASA, fied SDN-NFV model. This article attempted to combine
which is instantiated on nodes. Each AN provides a common SDN-NFV with an ML. Specifically, to provide unsuper-
set of capabilities across the network called ANI. The ANI vised and intelligent service toward optimal end-to-end QoS
provides functions like naming, addressing, negotiation, syn- provisioning, the autonomic mobile virtual network opera-
chronization, discovery, and messaging. The ANI is the ground tors (AMVNO) merges the SDN-NFV-based architecture with
for AFs and supports different ASAs. It comprises of: autonomics and resource management protocols to create a
1) bootstrapping remote secure key infrastructures unique system. Fig. 12 depicts the integration of the ACP
(BRSKI) protocol ensures automatic secure bootstrap in SDN-NFV architectures. The authors argue that consid-
using manufacturer-installed X.509 certificates; ering the effectiveness of unsupervised learning algorithms
2) autonomic control plane (ACP) is a self-managing, provides autonomics as a parallel control plane and not only
secure, and dedicated channel for management/control. as a block within existing architectures. This can be consid-
The interaction inside the ACP used Internet protocol ered as the extension of legacy SDN-NFV architectures as an
version 6 (IPv6) link-local addressing by default; additional 3-D. Consequently, representing interplane logical
3) generic autonomic networking protocol (GRASP) is a interactions, another type of logical interconnection is defined.
generic signaling protocol that runs in the ACP. It The logical structure of this plane strongly depends on
facilitates the interactions between ANs and ASAs. its unsupervised learning and data mining algorithms which
are indicated by the absence of specific blocks at the end
of pins at the ACP. This also significantly influences how
D. Autonomic Mobile Virtual Network Operators information from all the blocks in the foreground plane is gath-
Granelli and Bassoli [45] proposed customization and con- ered, processed, and used. The AMVNO, depicted in Fig. 12,
ceptual abstraction of autonomic network decision-making takes advantage of “cross-layer/cross-blocks” knowledge, as
QoS management in wired network [54]. This architecture

introduced the concept of autonomy and differentiated services
(ADiffServ) to solve issues of self-adaptation and flexibility.
Network automation deals with the decentralization of
network functionality. In order to achieve this, there is a need
to maintain policy states across all the network elements [56].
Fig. 13. Network automation development timeline. This led to the development of a cross-layer architecture,
autoXL [5] that adapts both a knowledge plane and a policy
plane. The knowledge plane is responsible for the provision of
it needs knowledge from both tenant SDN controllers (TC) information to equip all network elements with self-awareness
and infrastructure SDN controller (IC), from VNFs and ele- capabilities. The policy plan ensures that policy and policy
ment managements (EMs). Moreover, an AMVNO exploits states are updated so that network elements can meet their
the stateless and stateful information from both SDN switches predefined goals.
and VNFs. Finally, it requires strict interaction with both SDN In [55], an adaptive architecture for automating sensor-
and NFV, which spans from the Internet layer to the appli- actuator networks is introduced. Whereas, the work in [52]
cation layer. The ACP learns by performing a sequence of provides frameworks and protocols for pervasive networks.
operations called sensing, planning, deciding, proactive acting, Mathieu et al. [58] presented an autonomic method to cre-
and performance verification. ate, management. Liakopoulos et al. [53] applied distributed
monitoring techniques in autonomic networks. An architecture
VI. C LASSIFICATION OF E XISTING L ITERATURE for automated QoS management was presented in [54]. This
Network automation can be reviewed considering the state work provides a special emphasis on wired networks. A traffic-
of the art before and after the advent of network softwarization. aware network management automation was presented in [59]
In fact, SDN and NFV along with cloud computing created a and [60]. Finally, an organized survey of early attempts to
seismic change in the way networks will be designed, imple- provide network automation architecture was presented in [1]
mented, operated, and managed. Fig. 13 indicate the timeline and [13].
for chronological development and introduction of important 2) Literature Focusing on Network Type: The related
enabling technologies. research work discussed here covers network automation
network type areas, such as transport networks, overlay
networks, pervasive networks, wireless networks, wireless
A. Network Automation Before SDN and NFV sensor networks (WSNs), and sensor-actuator networks. An
Since 2004, after the introduction of autonomic comput- adaptive network traffic routing and scheduling are proposed
ing and networking [46], numerous research works had been in [61]. The solution introduces autonomy to traffic man-
performed in an attempt to have an autonomic network [4], agement. Autonomy had also been applied in optical burst
[47]–[50]. Let us further categorize the literature works on switched (OBS) networks. Praveen et al. [62] used learning
network automation into subcategories, depending on the main automata to realize the properties of an autonomic network,
focus of the articles. such as self-awareness, self-protection, and self-optimization
1) Architectural Models: Before SDN, most of the in a bid to solve the issues of routing and wavelength
networks included devices that each had a data plane and a assignment in OBS networks.
control plane as two separate entities. Different architectural Similarly, with a focus on a particular network type,
models had been proposed [51]–[56]. These architectural mod- few works attempted to incorporate autonomic in the
els made the management of such devices complex since the network operation and management [52], [55], [58], [63].
control planes were distributed among individual devices. A In [55], an adaptive architecture for automating sensor-
network administrator, in this case, would have to configure actuator networks is introduced. Whereas, the work in [52]
one device at a time. An early study of ANA is presented provides frameworks and protocols for pervasive networks.
in [57] where ANM is made easy through the collection of Mathieu et al. [58] presented an autonomic method to cre-
information using some monitoring devices. Using Darwin’s ate, configure, adapt, contextualize, and tear-down connections
theory as a basis, Tizghadam and Leon-Garcia [50] developed for context-aware overlay networks. A workload adaptation
autonomic network control and management system (AORTA) technique is presented in [63], which is based on queuing
with two major goals: 1) reducing network traffic congestion network modeling techniques in autonomic DBMSs. This arti-
and 2) increase the robustness of the network. These goals are cle addresses the need to manage workload and traffic in
achieved by a probabilistic routing algorithm that meets the DBMSs [63] where there is a need to assign resources and
traffic demands. assess performance based on the workload control schemes.
Bouabene et al. [41] proposed the ANA, which represents This article showed the big role of automation in offering
the network functionality as different functional blocks. This adaptation algorithms that can be used in DBMSs. Traffic
architectural proposal does not impose a one-fits-all structure management [60], monitoring [53], and scheduling [64] are
but rather a flexible framework that can be adapted to any very crucial parts of any transport network. The assessment of
protocol stack. The GANA model proposed by the ETSI [31] how critical a path is to topology and traffic demand changes
guides the development of a novel autonomic architecture for can be assessed using some kind of survival value using
Darwin’s theory [60]. Monitoring of traffic and traffic paths was discussed. It is of great importance to analyze any auto-
results in higher accuracy and reduced overhead as network nomic network for anomalous network traffic. Ma et al. [76]
nodes share monitoring data [53]. A scheduling algorithm recommended that the compromised nodes are identified and
is proposed for light-trail wavelength division multiplexing that all nodes are designed with self-recovering capabilities.
(WDM) networks [64] where the nodes in the network request Autonomic diagnosis of anomalous traffic is done using adap-
for bandwidth. The requests are made in advance and sent to tive anomaly detection components [73] that are composed of
the centralized controller. an entropy estimator to predict the onset of such traffic and a
In wireless networks, transmission and reception of sig- supervised naive Bayesian classifier. Detection and classifica-
nals are facilitated by radio waves moving through free tion of traffic are performed through collaboration among all
space. This kind of communication network is very complex network components. Autonomic devices are based on com-
given the ever-changing conditions of the channel through puters and they have weak points (vulnerabilities) that can
which the signal propagates [65]. Signals under phenom- easily be exploited by an intruder. Autonomic networks and
ena, such as multipath, interference, collisions, shadowing devices can be secured, traditionally, using firewalls and vir-
among others. Loureiro and Ruiz [65] provides a discussion tual private network (VPN) tunnels. This motivated the design
of network automation in wireless networks for environments. of a virtual autonomic network (VAN) firewall with logging of
Services run in most if not all wireless networks were orig- events to provide knowledge about the kind of security risks
inally designed to run in the wired networks. A testbed and instant messaging for alerts [71] which was implemented
implementation for mesh networks is presented in [66]. The in an intrusion detection system.
authors evaluated the performance of ANM. To alleviate the Rebahi et al. [72] analyzed the security concerns of auto-
problem of managing a femtocell, Rui et al. [67] presented an nomic networks at the architectural level concerning the
automated configuration using information modeling. GANA [6] reference model. The study in [57] focuses only on
The current technology design must meet the constraints of developing autonomic network monitoring for decision mak-
the wireless networks. Autonomic networks can be employed ing through information gathering using a random number
at different levels of the wireless network, such as in-building, of monitoring devices. An agent-based self-recovery (ASBR)
urban, suburban, and global levels. A typical deployment use Model [76] for WSN guarantees that the network is checked
case of wireless networks is the wireless sensors which are for compromised nodes and self-recovers in time. After the
very important in data collection. Usually, these sensors are development of the GANA, it was necessary to analyze the
placed in remote areas, making maintenance difficult for the security concerns around autonomic networks [72] by raising
technicians. This also makes their management and control points, such as security should be looked at as a requirement
difficult. Autonomy solves these issues by providing self- as well as analyze new threats that may arise from the self-
awareness and self-knowledge properties to these sensors. management aspect, communication between nodes and policy
Braga et al. [68] proposed the use of an AEs in the nodes management.
of WSN. They evaluated memory consumption and energy In general, the automation in legacy networks was per-
consumption. An AE is the smallest unit of an autonomic formed for nearly a decade, from 2004 to 2013. Even if
network. the attempts to performing network administration, without
administrators’ conscious knowledge, have been researched
3) QoS Monitoring, Fault and Traffic Management: In
since 2004, little has been done for real implementation and
addition to specific network type, there are also works focusing
adaptation. Therefore, the way we manage and design legacy
on automating specific functionalities, such as QoS monitor-
networks have been changed, requiring the new perspective
ing, fault management, and traffic management security [14],
provided by SDN and NFV.
[53], [54], [56], [59], [60], [69]–[74]. For instance, in [69],
a discussion on a model-based approach is presented. The
model-based method is to equip the fault management system B. Network Automation After SDN and NFV
with autonomic capabilities. Tcholtchev and Chaparadza [70]
SDN and NFV introduced a dramatic change in the
studied network operation automation for fault-Management
legacy static network, paving the way for network inno-
and resilience while Mokhtar and Eltoweissy [75] analyzed a
vations. The most notable contribution is the change from
memory-enabled autonomic resilient networking. The works
hardware-based to SDN. These have effectively replaced
in [56] investigated the consistency of policy states in decen-
specific network devices with general hardware. Network soft-
tralized ANM. Liakopoulos et al. [53] applied distributed
warization enhances the possibilities of innovation due to
monitoring techniques in autonomic networks. An architec-
flexibility, programmability, virtualization, and slicing. From
ture for automated QoS management was presented in [54].
an architecture point of view, SDN also provides centralized
This work provides a special emphasis on wired networks. A
network provisioning and management capability considering
traffic-aware network management automation was presented
a global view of the network. Now, we focus on navigat-
in [59] and [60].
ing through the most recent literature that has been done
4) Automaton for Network Security Application: Some since the widespread acceptance of SDN and NFV. Ironically,
works also focused on security application of automation [14], after the introduction of SDN and NFV, there are relatively
[71]–[73]. For instance, in [71], an automated intrusion detec- few organized efforts in the research community for network
tion system for devices within a virtual automation network automation [32], [77]–[79]. However, some researchers have
made efforts toward making improvements to the existing SDN layer—optical fiber, microwave links, Ethernet) and layer 3
and NFV concepts. (network layer—routers) of the open systems interconnect
1) Autonomic Network Architecture: Li et al. [8] presented (OSI) model. After this, the data is analyzed and decisions,
an autonomic management architecture for an SDN-based such as resource allocation are inferred from this analysis.
multiservice network. The first theoretical discussion for a van der Meer et al. [91] discussed the importance of network
software-based system is presented in [77]. The discussion automation in 5G networks in order to bring them to reality.
presents research and development with examples from differ- They also identify some of the areas where automation can
ent perspectives, such as biometric identification and tracking be applied in 5G networks. These include Radio access split,
systems, autonomic communications, and space exploration mobile edge computing (MEC), and C-RAN.
systems. Such works used an analogical approach to the bio- 3) Automating IoT Technologies: Here, are few works
logical autonomic nervous system. The study in [80] focuses focusing on automating IoT technologies. Asaamoning and
on two areas; data collection frameworks and a complete Mendes [85] analyzed a name-based coordination of auto-
refinement of the control plane as a functional split between nomic IoT functions. A more specific work, with a focus on
different control elements to enhance control flexibility, robust- enabling self-configuration, is dealt in [86]. Self-configuration
ness, and scalability. To handle real-time network changes, capability means providing a scalable and robust address for
Saadon et al. [81] also propose a dynamic architecture that future IoT networks. A control and management architecture
allows for reaction to such changes. They propose the intro- for optical transport networks is presented in [80]. The author
duction of a virtualization layer over the SDN controller with suggested an architecture to enable self-X capabilities using
its orchestrator. efficient network data collection and processing. They adopted
A specific work that focuses on wireless access networks the control plane functions in the ETSI NFV framework via
is presented in [82]. They developed a distributed generic dynamic instantiation. In [88], Abstraction and Control of
and autonomic management paradigm for decentralized the TE (Traffic Engineering) Networks refer to the set of
management of wireless access networks. More recently, virtual network operations needed to operate, control, and
Xilouris et al. [83] presented a demonstration of an open- manage large-scale multidomain, multilayer, and multivendor
source for MANO framework, an extension within the TE networks.
5GTANGO H2020 project scope. They discussed automated Very recently, an interesting application of artificial intelli-
service deployment considering policy-based service-level gence (AI) for IoT is introduced in [92]. The authors presented
agreement (SLA) provisioning and monitoring. frameworks of centralized and distributed AI-enabled IoT
Some works, network automation with a focus on specific networks. They argue that the AI-supported interconnected
scenarios, are also available in the literature [7], [39], [45], world will be the main part of future 6G wireless networks.
[84]–[88]. Granelli and Bassoli [45] provided a specific full This article further explored and discussed efficient and
autonomic architecture for mobile network realization, relay- overhead-friendly approaches using AI. They also identified
ing on NFV and SDN. In this work, they extended the work potential technical challenges for both centralized and dis-
in [39], providing a parallel autonomic control plan for an tributed. To address these challenges they suggested AI-based
SDN-NFV architecture. In the proposed architecture, the con- solutions.
trol plane has a logical connection between different planes for Industry 4.0 is the compilation of digital technology with
data gathering and control decisions exchange. Gifre et al. [84] a traditional production system. The Industrial IoT is adding
proposed a monitoring and data analytics (MDA) architec- abundant values in the production systems of companies in
ture. The architecture consists of centralized data storage addressing the current and future market competition and
along with data analytic capabilities and a generic node agent challenges [93]. IIoT is a subset of IoT which has shown
for monitoring. It provides support for disaggregation of a significant importance to businesses, safety, and lives. Unlike
multilayer network. An autonomic architecture for a con- other IoT areas of applications, the IIoT is focusing on con-
tent delivery network (CDN) is presented in [87]. Zhang and necting machines and devices in a diverse range of industries,
Wang [79] deliberated on an ANM scheme based on session such as manufacturing, agriculture, transportation, and health.
initiation protocol (SIP) in a collaborative environment. The Besides, combining IIoT with the recent and prevalent
work in [89] presents an interesting experimental demonstra- blockchain technology is one of the treads in intelligent man-
tion of an autonomic system in a cloud environment for a live ufacturing. The blockchain may act as a distributed ledger
streaming application. They discussed performance analysis for beneficially establishing a decentralized autonomous trad-
and challenges for a test-bed implementation of an autonomic ing platform for IIoT networks [94]. The authors study the
cloud network to manage live streaming applications. resource management and pricing problem between the cloud
2) Network Automation for Specific Network Type: Network provider and miners to alleviate the power and computation
automation research work after the introduction of SDN and constraints. More explicitly, the interaction between a cloud
NFV, discussed here, covers application areas, such as trans- provider and miners as a Stackelberg game is modeled. In the
port network, mobile cellular network, security, and C-RAN. model, the leader (cloud provider) makes the pricing first. The
As networks evolve, monitoring of the network becomes miners then act as the followers.
a key area of interest, especially in the transport network. 4) Automation of C-RAN: C-RAN, also referred to as
The research provided in [90] considers the use of autonomic Centralized-RAN, takes advantage of the current technolog-
agents for monitoring and data collection at layer 1 (physical ical advancements, such as common public radio interface
(CPRI) standard, low-cost Coarse or Dense WDM technolo- has been used in many areas [101] mainly focusing on works
gies to make the task of resource pooling very scalable, related to networking. We divide them by areas of application,
flexible, and efficient. With these technologies, resources are e.g., wireless, wireless mesh, security, mobility management,
offered within the shortest possible time, even over long dis- and NMS. A multiagent-based configuration in a ubiquitous
tances thereby providing a low cost and low latency network. wireless network is presented in [102]. The author showed how
van der Meer et al. [91] asserted that the C-RAN architec- agents are used in the RAN where the information is loaded
ture must allow for different automation approaches in order and fed back with the help of communication between various
to meet network requirements like low cost, low latency, agents. Q-learning technique is used to provide the agent with
reliability, spectral efficiency, etc. These approaches include cognitive abilities to make decisions on when to apply load
virtualization through NFV, centralization through SDN, man- balancing.
agement, and orchestration as well as coordination. A significant number of works exists utilizing agents for
5) QoS, Traffic Engineering, and Security: A specific func- WSN [103]–[110]. A multiagent-based system architecture
tion, such as QoS, Traffic engineering, and security are for WSN is proposed by [106] with four types of agents:
studied in [18], [23], and [95]–[97], together with SDN, auto- 1) interface; 2) regional; 3) cluster; and 4) query agents. In
nomic QoS management mechanism, and an autonomic and comparison to conventional client/server architecture, mobile
policy-based authorization framework for OpenFlow networks. agents have many advantages for WSN. In [103], a multiagent-
In [96], experimental validation of autonomic traffic engineer- based hybrid protocol is proposed, utilizing the advantages of
ing is studied. In [18] a survey on autonomic provisioning and combining value and decision fusion. The aggregation is per-
management of QoS in SDN networks is displayed. formed to avoid data redundancy before arriving at the sink.
As networks advance toward self-management, it is impor- Mobile agents perform the task of data aggregation at the
tant to consider the security aspects of these networks. The nodes rather than at the processing element resulting inef-
issue of security risks and vulnerabilities must be studied ficient bandwidth usage and improved network lifespan. A
with uttermost attention [14] and the duty of evaluating such source grouping along with a tree-based ordering mechanism
systems must move from being human dependent to the is proposed in [105]. Recently, Chen et al. [104] designed a
systems themselves. Systems must be empowered to know migration route planning algorithm based on mobile agents.
themselves and their missions, i.e., autonomic, to sense and Other related works in the area of WSN considered: a
decipher their situations, and to reply to occasions in an effort multiagent simulation [107] and performance comparison of
to protect themselves and to guarantee the coherence of mis- multiagent middle-ware platforms [110]. The mechanism for
sions they back [14]. Early expressions of this technique can Cooperation and consensus for multiagent networked systems
be seen in a few handled systems nowadays. is dealt in [108], [109], and [111]. Using the tuple space model,
Advanced network protection regularly highlights the capac- in [112], MAS architecture for intelligent home network is
ity to both evaluate the current state of network components presented.
and to reply promptly to distinguished concerns. DoS attacks Moreover, some other works focused on more specific appli-
are common in most networks. An attacker aims to make the cations, such as security, service discovery, and service migra-
network/service inaccessible for a given time. It is unreason- tion. Amudha et al. [113] proposed an approach for dynamic
able to undertake to dodge DoS attacks. Monitoring of network service discovery in service-oriented architecture based on
functionalities is very critical. This assists in revealing any a MAS using matchmaking technique. Patri et al. [114]
anomalies and plans for self-reconfiguration, self-healing, and developed a generic migration algorithm derived from a
self-protection. Scalable telemetry and network autonomics search-based rational agent decision process. Such an algo-
can be applied in abstraction and control [88]. rithm can deal with uncertainties to provide the migration
6) Recent Literature: Most recent literature focuses on the path. The migration is computed by using a maximized
utilization of network softwarization and machine learning tech- utility function. Several works also exist showing the use
nologies [17], [81], [98]–[100]. A demonstrating voice over an of intelligent autonomic agents for security applications,
autonomic network is elaborated in [98]. Stamou et al. [15] such as risk assessment [115], and network intrusion detec-
considered a survey of autonomic handover management for tion [116]. Zhang et al. [117] presented mobility management
heterogeneous networks in a future Internet context. Network over the satellite networks based on virtual agent clustering.
autoslicing supporting architecture is provided in [100]. Based Huang et al. [118] proposed the use of MAS methodology to
on software-defined infrastructure, adaptive load management facilitate the decentralized reconfiguration of power systems
of Web applications is presented in [99]. Saadon et al. [81] pro- to offer more flexibility and control as well as avoiding the
vided network virtualization and distributed orchestration-based problem of a single point of failure as witnessed in centralized
dynamic architecture for the management of the autonomic systems. Ma et al. [76] offered an ASBR model as a solution
network. A survey combining SDN, NFV, and ML is presented to DoS attacks in WSN. They describe a model in which a
in [17], for autonomic computing and communications in the network is able to monitor and identify nodes compromised
context of software-driven networks. by DoS attacks and self-recover in time.
None of the works considered an organized management
C. Application of Multiagents for Network Automation architecture, like in [119]–[121], for the overall network
There exist several research works emphasizing the appli- system management. Taoqing and Xiaoying [119] presented a
cation of the multiagent-based system in networking. MAS multiagent-based management system designed for distributed
network. They developed an intelligent management archi- networks. The GANA model in Fig. 10 illustrates the concept
tectural framework. The architecture organizes agents into of network function atomization by splitting the architecture
three types: 1) center-layer agent; 2) region-layer agent; and into different levels based on the different DEs implemented
3) access-layer agent. The functions of each agent and cooper- therein, i.e., protocol level, function level, node level, and
ation mechanisms in the framework are discussed in this arti- network level. These DEs implemented at different levels in
cle. Some interesting work about control systems is presented the system offer a distributed model of networking toward the
in [122] from the control research community. Several works management of resources and other devices.
existing related to steady states of the closed-loop system and Since the SDN architecture does not specify the internal
study-state analysis of multiagent-based system [123], [124]. design or implementation of an SDN controller, it could be
possible to design the SDN as a single monolithic process
or as a confederation of identical processes arranged to share
VII. T ECHNOLOGICAL T RENDS
the load or protect one another from failures or as a set of
The current technological trends in the research and indus- distinct functional components in a collaborative arrangement.
try are taking the direction of systems being distributed Controller components are free to be executed on arbitrary
and loosely coupled. This research journey can be sum- computing platforms, including computing resources local to
marized as the move from centralized and hardware-based a physical element. They may also execute on distributed and
toward distributed cloud-based and virtualization/softwarized possibly migratory resources, such as on VMs as a sequence
approaches. Before virtualization, applications were installed of VNFs in data centers [129].
as a complete system with its OS over some underlying
hardware. This, however, had several challenges, such as con-
strained mobility and difficulty running updates. As a result, B. Containerization and Cloudificaton
virtualization came along and offered the possibility of having Network automation could employ the concept of dis-
many applications, each in its virtual environment, but sharing tributed computing and centralized control. In a distributed
the same hardware infrastructure and resources. This greatly computing the solution to a problem is obtained through the
improved configuration and management of applications since division of tasks amongst a group of network management
VM are easy to clone and install in a different location. elements. This results in improved performance, flexibility as
Applications are designed to run in different isolated spaces tasks can be performed by different machines irrespective of
(e.g., containers, dockers, etc.) while utilizing the same their geographical location and reliability since a glitch in a
resources and OS. The concept of “containerization” [125] single element does not mean total DoS in the entire network.
comes with a number of benefits, such as efficient resource Distributed computing is employed in network automation
utilization, quick development and debugging of applications, through containerization, where different services of applica-
the possibility of fault isolation, and easier management. With tions are run in different containers while sharing the same
the network, automation arises the possibility for any network OS and other networking resources [125]. This happens in
to self-heal (fault isolation), self-configure in the shortest time an effort to implement virtualization of services, easy fault
possible, self-protect, and self-manage. isolation, and lower system costs.
Today, implementation is moving from on-premises applica- On the other hand, centralized network control is built
tions to cloud-based applications [126]. Applications can run around a single controller that manages all the major network
at the customer premises, i.e., in a server in some server room control. Network nodes or workstations connect to the cen-
in a building. This is becoming infeasible in terms of CapEx tralized controller and submit their requests to the central
and OpEx, scalability, and security (against theft and natural controller rather than performing them directly, depending
disasters like fires). Containerization alone does not imply on the required network control functions. The centralized
cloudification. Applications are hosted in the cloud and as long controller manages the flow by programming the devices to
as one has access (and an Internet connection), the services perform the required packet routing or management. The con-
are available. This has resulted in inefficient resource utiliza- troller directs traffic according to forwarding policies that a
tion, security against theft, and easy disaster recovery due to network operator puts in place. This minimizes manual con-
the availability of an off-site backup. The costs of migrat- figurations for individual network devices. The centralized
ing services and maintenance of equipment are significantly control will have a global information about the network.
lower than the potential OpEx [127] without virtualization and While maintaining a global view of the network and cen-
automation. tralized control principles, the controller could be designed
using a monolithic system or as a decoupled microservices or
A. Network Function Atomization MASs. The existing controller systems are designed as mono-
The concept of function atomization was introduced in [68], lithic systems, such as Ryu SDN controllers. However, very
where every node in the WSN is equipped with an auto- recently, a microserivce-based controller design is proposed by
nomic element (the smallest object in any autonomic network). the µONOS projectmicroONOS. It decomposes the controller
Besides, the idea in [128] was a considering the composi- functions into VNFs and deployed them as microservices in
tion of autonomic elements. This composition is intended to containers. The decomposition provides an independent imple-
include an autonomous overlay network management structure mentation of controller functions. Moreover, this approach
and a self-organizing composition against autonomous overlay provides flexibility in terms of dynamic scaling and backup
Fig. 14. Cloud computing versus fog computing versus edge computing.
deployment of functions. Such design provides the option of

distributed deployment of the centralized control. Fig. 15. Monolithic versus microservices-based architecture.
The network functions, that are either be designed as
microservice or multiagents, can be deployed as containers. A processing facility, that exists far away from the data source
container is a lightweight computing package that is abstracted (sensors). Accessing services here means accessing resources
away from the host OS. OS making it easy to migrate con- on the Internet, so experiencing considerably higher latency.
tainers from one device to another without the need to adapt Fog computing, on the other hand, is closer to the information
them to the new device. The applications deployed in each source and is characterized by its distributed nature, low-
container, each have their libraries and configuration files. latency. A number of standards and protocols are required in
The containerization technology today is divided into the order to access it. Edge computing (similar to fog computing)
following. happens in the edge and at the devices, where the sensors are
1) Container runtime, a software that executes container placed thereby, resulting in lower latency and faster system
images on a given node. These include Docker, response.
Docker Enterprise, CRI-O, rktlet, containerd, Microsoft
Containers, etc.
2) Cluster management and deployment technology, such C. Monolithic-Based System Versus Microservices-Based
as Kubernetes, Docker Swarm, etc., for workload man- System
agement and resource assignment. Two architectural styles are considered today by software
3) Storage containers like BlockBridge and EMC/libstorage developers, i.e., monolithic and microservices [132]. As a
for storage purposes. traditional approach toward software development, monolithic-
4) Container security like Twistlock, Aqua, Aporeto, etc., based systems saw the implementation of an application as
to provide for detection of any anomalies, such as a single unit [133]. Fig. 15 illustrates both the monolithic
intrusion and aid in processes like single sign-on, vul- and microservices-based architectures. In the monolithic-based
nerability scanning among others. architecture, all functions are handled and performed in one
The distributed nature of computing is employed either place, i.e., the business logic. The application has its internal
on-premises or in the cloud. The paradigm shift today is structure, consisting of a database, client-side interface, busi-
cloudification due to its advantages, such as faster disaster ness logic but remains an indivisible unit. For example, with
recovery, security, more resources, and easy access at any time online shopping, the business logic includes a shopping cart,
regardless of geographical location. Cloud computing can be product catalog, accounting, and customer care, all imple-
described as a model enabling the sharing of a variety of com- mented as one unit. This had its advantages, such as easy
puting resources in the form of services through remote access development, debugging/testing, and deployment. However,
over the Internet rather than a physical computer or storage when the size increases, the application becomes too complex
disk. This computer system is split into two parts, i.e., frontend for changes to be made in a very short time. Another major
(clients or user devices) and backend (servers). Terms, such as drawback of this approach is the need for the redeployment
edge computing and cloud computing have arisen because of of the whole application after each update has been made.
the overwhelming need for faster and better technologies. Edge On the other hand, the idea with microservices-based
and fog computing systems also bring data processing closer systems is to build an application as a set of small
to the source of data generated, i.e., the sensors [130]. The key interconnected services [133] instead of a single unit as is the
reason for this is to minimize the amount of data sent to the case of monolithic-based systems. Each service is equipped
cloud, thus reducing the latency. As a result, the response time with its logic and a database, in order to perform its func-
of the system particularly improves for applications requiring tions. This offers a simpler understanding of the application
low latency, such as the 5G URLLC services. and offers better scalability. Failure of one microservice does
Fig. 14 illustrates clearly the difference between cloud, fog not mean complete failure of the entire system. From Fig. 15,
and edge computing [131]. The cloud layer (cloud computing) it can be seen that the microservices-based architecture has its
can be described as a large and centralized data storage and business logic broken down into microservices, each of which
Fig. 16. General overview of a MAS.
Fig. 17. Comparison of microservice and MAS.

responsible for a particular business purpose, e.g., in an online
shopping example, the microservices include shopping cart,
product catalog, accounting, and customer care.
In most cases, the choice between microservices and any failures, tasks can always be reassigned to other agents.
monolithic-based architecture favors the former. The tendency Flexibility stems from the fact that each agent performs tasks
to concentrate only on the drawbacks of monolithic-based based on already predefined knowledge.
architecture and abandon it needs to be justified against some The networking industry is heading in the direction of
measurable metrics. The best approach is to build new prod- highly decentralized and loosely coupled softwarized systems,
ucts using the monolithic-based architecture, and then move such as NFV and MEC. Currently, large-scale software devel-
to the architecture of microservices when there is a justifi- opment is based on microservices. Microservice is a model,
able cause backed by proper observable quantities, such as in which systems are built using small and loosely coupled
performance monitoring. services. They are potential tools for building systems at scale,
such as NMSs. They adhere to the isolated state, distribution,
D. Multiagent System elasticity, automated management, and loose (IDEAL) cou-
MAS is part of the distributed ML paradigms used today to pling design principles. Microservice and multiagent are very
solve complex tasks [134]. Some of the areas where MAS has similar as they provide a mechanism in building systems using
been used include online trading, robotics, smart grids, com- smaller functional units. Such units could be used as build-
puter networks, etc. A MAS, therefore, is a collection of agents ing blocks of monolithic systems. Microservices have many
that interact with each other. “There are several definitions of similarities with MASs [135], [136]. Their core similarity
an agent but in simple terms, an agent is a physical or software is independence and loose-coupling. Collier et al. [136]
entity equipped with the capability to perform toward a given presented an interesting comparison of MASs and microser-
set of objectives [134].” In other words, an agent is capable vices.
of figuring out what to do on its own in order to achieve a From another perspective, VNFs are softwarized network
given objective. Fig. 16 depicts the overview architecture of functions that could potentially be implemented as microser-
a MAS with two agents. The agents interact with each other vices or multiagents. However, MASs have an advantage over
using agent communication language messages. Agents are microservices for dynamic and ANM system design. Recent
characterized by the following features. work presented in Collier et al. [136] showed the possibil-
1) Social Ability: Agents communicate with each ity of combining microservices and MASs. Fig. 17 presents a
other to exchange information and also request comparison between microservices and MASs.
information from other agents. Agents leave pheromone Multiagent deployment could be used to build systems, such
(information) for other nearby agents. This pheromone as cloud radio access networks (C-RAN) and traditional NMS.
may increase/decrease with time. This communication As indicated above even the SDN controller itself, could be
between agents is facilitated by agent communication decomposed into subfunction so that the subfunction could
language. be designed as agents. This provides a lot of benefits, such as
2) Autonomy: Agents can act independently without human efficient scaling, reusability, loos-coupling, distributed deploy-
intervention to perform specific tasks using some kind of ment. Future research directions could be explored in this
inference mechanism. These agents are also self-aware. direction.
3) Proactive/Reactive: Agents may react in response to Fig. 18 is an example illustrating an atomized NMS for
the environment (reactive) or they may not react in C-RAN where the concept of multiagents has been employed.
response to the environment and simply exhibit some Using the GANA reference model as inspiration [6], [137],
goal-oriented behavior (proactive). agents are employed in the four different levels of the system,
These features in turn are efficient and cheaper. Multiagents i.e., protocol level, function level, node level, and network
are flexible and reliable system [134]. Division of labor among level. At the highest level (node level), the agents here
the different agents results in efficiency and this distributed are responsible for QoS management, load balancing, and
nature of the system offers high reliability, i.e., in case of scheduling. These network-level agents enable the network
Devices had to be configured one at a time making the con-

figuration and management of the network complex and very
costly. As technology advancements happened, there was a
need to abstract control of devices.
The need for centralized control of network devices arose as
a solution to the complexity and cost incurred in configuring
devices individually and managing the network. This has led
to the birth of concepts like SDN where the control layer and
the data layer are separated. Currently, these are the networks
under development, and they are labeled SDN [140]. The oper-
ations, such as configuration, management, and update of flow
tables are done via two interfaces: 1) NBI and 2) SBI.
The current network transformation toward the future is
leading toward the employment of ML tools to help in the
decision making, control, and management of networks. As
networks evolve, they become more complex, and relying on
human support will become infeasible. Networks will be able
to learn from past data and experience to infer decisions in
the network [140].
Fig. 18. Atomized NMS for C-RAN. F. Intelligent Algorithms: Artificial Intelligence and Machine
Learning for Networking
to self-optimize and self-adapt to the changing network sit- Currently, there is a shift from MDNs to intelligent defined
uation. The node-level agents receive AB instructions from networks (IDNs) [11]. Operations, such as optimization and
the network-level agents and forward them to the function- configuration, in MDNs, are completely dependent on human
level agents. These in-turns forward the instructions to the intervention. Whereas the same operations in IDNs are not.
node-level agents. The function-level agents also implement The complexity introduced by network automation can be
BBU functions. At the protocol level, the agents receive the reduced by introducing ML technology into the network [140]
instructions and implement functions, such as error correction, thereby making it an intelligent network. Intelligent networks,
coding, decoding, and modulation. rather than following strictly defined policies/laws, can learn
from past data to guide new predictions or decisions.
On the one hand, through learning from new data, the
E. Intelligent and Smart Network of the Future IDN can dynamically adapt to a changing situation of the
Network development has continuously been evolving since network and develop its intelligence with it. There are var-
its first inception, in the late nineteenth century [138], [139]. ious methods to extract information for given data. ML
Networks have evolved from being fully human dependent tools, such as fuzzy logic, case-based reasoning, evolutionary
toward a scenario where no human intervention will be computing, artificial neural networks, and intelligent agents
required. At different stages, the network has faced different provide powerful aids in solving difficult applied problems,
and complex management challenges. The main characteris- that are often real time, involving large amounts of dis-
tics of this evolutionary trend for networking are depicted in tributed data, and benefiting from complex reasoning. Through
Fig. 19 [138]–[142]. The figure shows the continuous network sophisticated user interfaces and visualizations, IDN may
evolution stages: a mechanical defined network (which is typ- also strive to help human-based decisions through prepro-
ically characterized by mechanical switching and full human cessing data and providing insights to users and network
involvement), digital electronic defined network (this stage administrators. Most intelligent networks employ machine
is peculiar with the introduction of electronic devices, such learning paradigms to predict/classify network traffic pat-
as transistors for switching and memory), SDN (which is terns, application patterns and later make decisions based on
a shift from electronic hardware to virtualization and soft- these predictions/classifications. The ML paradigms include:
warization), intelligent-defined networking (which is typically supervised learning, unsupervised learning, and reinforcement
defined with the use of intelligent algorithms, such as ML learning [11]. These paradigms influence decisions how data
in the network management and operations), and quantum AI is collected, processed, and the inference of decisions from
defined networking (which will be the utilization of quantum the data.
computing and introduction of new communication paradigm, We continue to highlight some of the networking areas
such as teleportation). where machine learning paradigms have been applied. An
In the past, networks were termed as mechanical/manually interesting comprehensive survey paper on the application of
defined networks (MDNs) where all configurations or ML toward network automation is presented in [11]. ML
changes needed human intervention [140]. The operation and exploits information from the network and users’ data to
management of these networks were also human dependents. learn and perform the required changes to suit the demands.
Fig. 19. Continuous network evolution stages: mechanical defined network, digital electronic defined network, software-defined network, intelligent defined
network, and quantum AI defined network [138]–[142].
Recently, using SDN’s capability in collecting an enormous upon several approaches, e.g., AI and ML. In this case,
amount of packet and traffic data, lots of research has been network administrators and operators face the hurdle of data
done to explore the applicability of ML in networking for traf- accusation management and storage.
fic analysis [143], [144]. In [145], an autonomic VNF place-
ment is discussed considering a three data center hierarchy
A. Network Automation Challenge
for workload offloading between the data centers depending
on the traffic load. Most of the recent work on ML-based VNF 1) Neither Fixed-Rule NOR Zero-Touch Automation: The
placement has focused on exploring ML for dynamic traffic- fundamental questions in network automation are: to what
aware VNF placements [2], [146]. Sciancalepore et al. [2] extent should the network be autonomic or independent of
presented a method called z-TORCH orchestration mecha- the network administrator? Does it mean zero involvement? If
nism that uses unsupervised learning to monitor VNF key this is the case, then how does the system know what to do
performance indicator (KPI) and reinforcement learning to at each stage of the network, either it is in a transition state
find a tradeoff solution for reliability and complexity of the or steady state? Two extreme scenarios should be considered.
monitoring system. Subramanya and Riggio [147] discussed a The network elements are assumed to be capable of interacting
classification approach to dynamic resource scaling and VNF and learning their working domain to perform a given func-
placement. They used a specific version of neural network with tion and form the overall network. In this sense, they could
multilayer perceptron to predict the required number of VNFs start with a predetermined set of rules. On the other hand,
at a given time using a data set collected from a commercial they could start with no rules at all. To elaborate further on
mobile network. Time series prediction for VNF placement is the previous questions let us see two extreme scenarios.
presented in [148]. Tang et al. [148] first analyzed the traffic 1) Let us say DEs, such as devices, agents, or microser-
characteristics of the data center and devised a traffic fore- vices, can act according to preset rules or fixed policy or
casting technique. Based on traffic forecasting, they develop a common network management knowledge (preset with-
deterministic algorithm to determine the VNF resource scaling. out violating the rigid rules). They act with full authority
The application of ML in network automation raises a few in their routine function. Assuming zero perturbation,
fundamental questions. To what extent should the network be such as no device failure or misbehavior from the rou-
autonomic or independent of the network administrator? Does tine tasks, the overall network performs its functions
it mean zero involvement? If this is the case, then how does and behaves properly without digressing from the nor-
the system know what to do at each stage of the network, mal behavior all the time. This is advantageous in terms
either it is in transition state or steady state? These are two of system stability. However, such systems are limited
extreme scenarios that should be considered. The network ele- in their capability to adapt to changing and evolving
ments are assumed to be capable of interacting and learning environments. They are unable to capture the dynamic
their working domain to perform a given function and form the nature of networks, users, and services.
overall network. In this sense, they could start with a prede- 2) All DEs or agents could start with an empty set of rules
termined set of rules. On the other hand, they could start with or common knowledge, like aliens or people from dif-
no rules at all. In order to realize full network automation, a ferent places with different knowledge even different
comprehensive approach for the cumulative decisions consid- languages. They could start developing a mechanism
ering the overall decision latency, reliability, consistency, and to communicate, coordinate and accumulate common
computational efficiency is a necessity. knowledge that evolves in time, to work together and
accomplish a given task or perform the overall network
function. This approach to DEs definition is extremely
VIII. C HALLENGES AND O PPORTUNITIES flexible, with a full degree of freedom. It could cap-
Achieving complete network automation remains an enor- ture the dynamic and evolving nature of the network.
mous challenge as network administrators and operators are However, it is so flexible that the system could take a
struggling to create and simulate a proper network automation long time to converge in time. Moreover, we are not even
environment. The creation of this environment is dependent certain for the system to converge at all. Furthermore,
in such a system even if they converge in time due to amount of computational resources is required to execute ML
the evolving nature of the knowledge and agent’s behav- models. These resources are scarce and expensive to acquire.
ior, the system may exhibit instability at any point of the 1) Data Set Using Existing Data: Data is regarded as pri-
network. This could be a limitation to apply for a critical mary or secondary. Primary data is collected by the intended
part of the network, such as an NMS. users for a specific purpose while secondary data is collected
As this fundamental question would continue to persist, a focus by someone else and accessed by the intended user through
on defining a high-level involvement of the network admin- storage facilities, such as government databases and research
istrator is needed. This means that the administrator just sets institutions. Existing data sets are oftentimes referred to as
high-level network policy as a starting point or updates the secondary data. While using this data has proven to be very
policies or injects a new set of policy, whenever necessary to economical, it poses several challenges. The major challenge
tune the network to the business need of the operator. Still, an of using existing data is that it may be outdated or inaccurate
exact definition of high-level policy should be prescribed for and it may not address the researcher’s needs. The data sets
the network operations. These policies are to be used by the may also be presented in an undesirable format thereby mak-
automated management system to respond in managing the ing them inaccessible or difficult to process for the intended
evolving network and services demand. This mostly involves purposes.
a routing and cyclic process. 2) Data Set Generation Using Simulation: Most ML mod-
els require a lot of data which is usually expensive to
mine and may also be scarce due to privacy concerns, espe-
B. Challenges of Application of Machine Learning cially for medical records. Today, we see a shift toward a
Techniques for Network Automation generation of data sets through simulation—hence synthetic
To train an ML model, a good, clean, and sufficient amount data [149], [150]. Synthetic data, however, comes with a price.
of data set are very important to produce an accurate model. It introduces biases in the system model. For applications that
There are numerous challenges in acquiring data sets depend- require high accuracy, such as autonomous vehicles, it is most
ing on the kind of model that we would like to develop. This likely that a lot of real data will be needed in addition to syn-
could be at the stage of data collection, storage, processing, thetic data. For situations where the degree of accuracy can be
training a model, and even the availability of the required data comprised without serious consequences, synthetic data is very
set for research. For instance, there is a lack of 5G traffic to helpful. Synthetic data is helpful but the ML models, based
produce a traffic composition and workload amount prediction on real training data, are much more robust and precise. Even
model Or there is the problem of or difficulty of finding real- in situations that require synthetic data, it is good practice to
time 4G data set which could be due to operator/user security mix it with real training data to have an understanding of the
and privacy. The lack of such a type of traffic data set to train real-world in which the system will operate [151].
an ML model is a very challenging prospect in adopting an 3) Big Data and Big Data Analytics: In this era, we have
ML model in the network automation process. ML techniques very large volumes of complex data, i.e., big data [152]. Big
require training on large, unbiased data sets. Often these data data analytics involves the use of software to understand pat-
sets take a lot of time to produce them but they often need a terns and trends in large volumes of data. This requires a
huge storage space to preserve them [7]. comprehensive data management plan in terms of the right
1) Inability to extract the required data in real time. tools and skilled staff. Without these, the data is useless.
2) Data can sometimes be extracted in an undesirable for- Analysis of big data requires a lot of computing resources
mat which requires some manual clean-up before it can that are always not readily available or are very expensive to
be used for network automation. acquire. Security is another major concern when it comes to
3) The need for storage repositories for large amounts of big data [153]. A company database with large quantities of
data. confidential data is prone to malicious attacks.
4) Access to data, i.e., user-assigned permissions in order Device-level security considering the application of ML is
to manipulate data files. presented in [154]. They explored the attack model for IoT
5) The need to ascertain data integrity. systems and reviewed potential security solutions. Specifically,
Another critical challenge of ML is the delay in training they discussed ML-based IoT authentication, access control,
a model. ML algorithms require ample time to let the algo- secure offloading, and malware detection schemes focusing
rithms learn and fulfill their task with considerable precision. on protecting data privacy. They also identified potential
This also requires enormous resources to work, i.e., additional challenges that should be addressed while implementing the
computing power and time requirements. Another big obsta- proposed ML-based security schemes considering a realistic
cle is the ability to interpret the algorithm-generated results IoT system. However, most interconnected IoT systems gen-
correctly. The greatest obstacle of using ML in networking is erate a very large amount of data from the IoT devices and
the fact that networking and ML are very different areas and the network interconnecting them. This poses a tremendous
require expertise in both areas [7]. challenge that needs techniques from big data analytics.
The computational complexity of an ML model could be an An interesting survey on big data analytics is presented
issue depending on the network type. For instance, it could be in [155]. This article summarizes different aspects of big data
computationally expensive for some IoT networks with low- analytics for cyber–physical systems. The authors have also
powered devices and little computational capability. A huge discussed possible tools for data analytics. They also discussed
the security challenges of big data in a cyber physical system. (de)allocation, dynamic or variable coding/decoding, end-to-
Big data in a cyber–physical system requires protection against end QoS provisioning for services satisfying the dynamic
malicious attacks and unauthorized access. This is a tremen- demands. Each of these decisions affects the network state.
dous challenge, considering the enormous amount of data that Therefore, an organized approach has to be devised to have
are continuously generated in the network system. the best optimal behavior of the network.
In general, along with the data access, security is always an Existing approaches are automating individual decisions
added concern as we go from hardware-based environment to using different ML techniques, such as an ML algorithm for
software into the cloud and further to the intelligence domain. automated channel allocation. This would be suboptimal if it
There are always ranges of security challenges arising from does not have a global view of the network as well as the
the network perspective, either be the traditional network secu- global decision. Moreover, such approaches are limited if we
rity challenges, such as intrusion, DDoS, and the anomaly or consider automating the overall n network system. Automating
additional challenges due to centralized control, in a cloud the overall system need to make a decision at all level of the
environment with the added intelligence. It is always a chal- network. Finding such cumulative algorithms and architectures
lenge that someone with a better intelligent algorithm(s) could should be active research. An example could be automat-
be able to miss use or enable to miss behave the network. ing every part of the network, such as functions, devices,
Exploring the challenges and providing solutions should be and networks. And a coordinating mechanism for each level
an interesting but demanding area to work on. should be defined as well as the decision domain. Some of the
4) Challenges of Decision Generation: Automatic gener- proposed techniques are hierarchically organizing decisions
ation of various decisions by analyzing and reasoning from per level of decisions for abstraction. Defining and elaborat-
the collected data is very critical. Understanding, defining, and ing this architecture requires a huge work that should be an
parametric measurements of complex and intertwined network interesting challenge and an active research area.
events are required to manage using an automated decision. 7) Challenges of for Accuracy of Automated Decisions: The
ML tools provide powerful aids in solving difficult applied outcome of an autonomic decision should be assessed accord-
problems that are often real time, involving large amounts of ing to both the process to and outcome of, decision making.
distributed data, and benefiting from complex reasoning. ML The process of decision making is very critical. And it should
tools could be fuzzy logic, case-based reasoning, evolutionary be assessed since it has a big implication on the time it takes
computing, artificial neural networks, and intelligent agents. to decide and the additional computation and communication
Currently, machine learning is the most adopted means overhead. The quality of a final decision is also very impor-
of extracting valuable information about users, services, and tant that should be assessed based on a defined criterion, which
networks. Networks involve a colossal amount of data to learn could be specific for a given management demand. Assessing
from. The goal of machine learning is to extract the hidden a decision requires defining measurement matrices.
knowledge about the network and users through training, using We need to have a mechanism to validate the final decision
a network data set. Investigating specific decision generation for consistency, accuracy, and optimality. The final decision’s
mechanisms are necessary for various level of DEs, events, measurement and verification mechanisms should consider
and scenarios. consistency in collective system behavior. That is, because, in
5) Challenges of Network State Measurements a cumulative decision, individual decisions may not be the best
Understanding the network state is necessary to make or optimal one, since multiple factors should be taken into con-
decisions. Measuring the environment provides information siderations. It is important to model desired network behavior
about the network behavior in a reactive or proactive manner, pattern for proper evaluation of the performance of the automa-
based on continuous observation of the constituent network tion techniques. Typically, healthy network behavior initially
elements. Observation of the internal state can be a continuous is decided by a network service provider or network admin-
measurement of network state indicative metrics or service istrator, through a set of behavior constraints. The accuracy,
QoS metrics, such as latency, reliability, congestion, and time, and robustness of measurements have to be investigated
throughput. Network state measurements at each layer, such from network automation perspectives.
as the Physical, data link layer, and network layer have to 8) Challenges of Decision Execution: In network automa-
be performed. The challenges of accurate measurement of tion, action or decision execution is needed to fulfill network
network state are very necessary. The measured data have to management demands. An action could be network reconfig-
be timely available for processing and understanding to make uration, network state database updating, end-to-end network
automated decisions. Moreover, the size of the measurement path recomputing, and resource allocation, service authenti-
data also has a huge implication in the data accusation, cation, authorization, and network monitoring. The outcome
processing, and decision extraction process. There are various of decision generation needs to be applied to the system to
existing techniques of data measurement and accusation. change the part of an (overall) network system to fulfill ser-
However, contextualizing from the network automation vice or management demands. The network’s behavior change
perspective has to be investigated. comes by applying the final verified decision to change the
6) Organization of Automated Decisions: Several decisions network environment so that it behaves as per the network
have to be made to perform service admission, processing, management demands. The response to the change has to be
network configuration, and network maintenance. These dif- done appropriately and timely. This is very critical. A sim-
ferent decisions have to be made, such as dynamic channel ple example could be when we make a decision to route a
flow based on the currently available path. The path status or complexity of the network as a huge number of users, includ-
available resource changes very quickly, which may invalidate ing IoT devices, are joining the communication networks.
or outdated the decision. This is mostly because, in network Autonomics is the means of network management that is
automation, the aim is to develop a context-aware response aimed at reducing its overall complexity and its need for man-
to the demands. Decision execution as a part of the auto- ual/human intervention. In this work, we provided a global
mated process is also a vast and challenging area that needs view of research in network automation, to identify different
to be explored. A loopback control mechanism is the most challenges and opportunities surrounding the idea of automa-
applied method in most automation systems. An investigation tion. We identified the key enablers of network automation as
of the effective loop back controlling algorithms that would SDN, NFV, and ML. Moreover, we classified existing literature
be applied in network automation is active research. At the about network automation in terms of research done before
various level of a network system, measurements mechanism and after the introduction of SDN and NFV. We also recog-
are also an active research area to observe the effect of the nized the efforts of different stakeholders in the standardization
executed decisions. of different architecture for network automation that include:
9) Network Decision Element Communication: Various MAPE-K, GANA, and ANIMA reference models. This arti-
DEs may be involved to make a given decision. For instance, cle also discussed the possible research direction in network
if we consider the ETSI hierarchical organization of DEs, automation, e.g., the shift from on-premises services to virtu-
we will have various elements at each level. These elements alized services toward cloudification. This shift is driven by
could be implemented independently using software containers concepts, such as SDN, NFV, containerization, cloudification,
and deployed in a distributed environment. To have a com- ML, and MAS. Moreover, we highlighted the need for ML
plete system function communication between such devices is and its fundamental contribution to network automation.
needed. Effective communication mechanism has to be devices
for each type of communication. The common communication
between each element is remote procedural call (RPC), great R EFERENCES
Remote Procedural Programming (gRPC), representational [1] Z. Movahedi, M. Ayari, R. Langar, and G. Pujolle, “A survey of auto-
state transfer application programming interface (RESTful nomic network architectures and evaluation criteria,” IEEE Commun.
API), hypertext transfer protocol (HTTP), and ACL. Each Surveys Tuts., vol. 14, no. 2, pp. 464–490, 2nd Quart., 2012.
[2] V. Sciancalepore, F. Z. Yousaf, and X. Costa-Perez, “z-TORCH: An
has its own advantages and drawbacks. An investigation of automated NFV orchestration and monitoring solution,” IEEE Trans.
effective communication techniques for each DE should be Netw. Service Manag., vol. 15, no. 4, pp. 1292–1306, Dec. 2018.
investigated. [3] Consensus on 6G Is Gradually Forming. Accessed: Mar. 24, 2020.
[Online]. Available: https://telecoms.com/503142/consensus-on-6g-is-
10) Hardware Failure Management Automation: Existing gradually-forming/
electronic devices degrade over time due to fatigue, corrosion, [4] M. Behringer, A. Dutta, Y. Hertoghs, and S. Bjarnason, “Autonomic
or damage incurred during operation, leading to device fail- networking—From theory to practice,” in Proc. IEEE Netw. Oper.
Manag. Symp. (NOMS), Krakow, Poland, May 2014, pp. 1–17.
ure. The most challenging aspect of network automation is [5] W. Berrayana, H. Youssef, and G. Pujolle, “A generic cross-layer
when a network device or function failure occurs. For exam- architecture for autonomic network management with network wide
ple, a failure network switching could interrupt services and/or knowledge,” in Proc. 8th Int. Wireless Commun. Mobile Comput. Conf.
(IWCMC), Limassol, Cyprus, Aug. 2012, pp. 82–87.
reduce the available physical paths. The traditional mecha-
[6] X. Long et al., “Autonomic networking: Architecture design and
nism to tackle such kind of failure is mostly over-provisioning standardization,” IEEE Internet Comput., vol. 21, no. 5, pp. 48–53,
resources and switching devices to have alternative routes and Sep. 2017.
back nodes. However, these are not very expensive techniques [7] K. Tsagkaris, M. Logothetis, V. Foteinos, G. Poulios, M. Michaloliakos,
and P. Demestichas, “Customizable autonomic network management:
but also may not be effective solutions as the node could Integrating autonomic network management and software-defined
be recovered. Automatically recovering hardware failure is networking,” IEEE Veh. Technol. Mag., vol. 10, no. 1, pp. 61–68,
very difficult as it may need hardware component repair or Mar. 2015.
[8] H. Li, X. Que, Y. Hu, G. Xiangyang, and W. Wendong, “An auto-
replacements. nomic management architecture for SDN-based multi-service network,”
To alleviate such problems numerous recent research works in Proc. IEEE Globecom Workshops (GC Wkshps), Atlanta, GA, USA,
are being performed in material engineering. Using a novel Dec. 2013, pp. 830–835.
[9] Software-Defined Networking: The New Norm for Networks, Open
electronics device development technique called self-healing Netw. Found., Menlo Park, CA, USA, 2012.
soft-electronic materials that are mechanically robust and can [10] NFV Technology. Accessed: Oct. 30, 2019. [Online]. Available:
self-repair [156]. The investigation of such devices for network https://www.etsi.org/technologies/nfv
device building could also be an active research area. This [11] R. Boutaba et al., “A comprehensive survey on machine learning
for networking: Evolution, applications and research opportunities,”
could be interesting for network automation in that it could J. Internet Service Appl., vol. 9, no. 1, p. 16, Jun. 2018. [Online].
provide full network recovery from failure. It is also possi- Available: https://doi.org/10.1186/s13174-018-0087-2
ble to have automated self-healing mechanical maintenance [12] M. R. Nami and K. Bertels, “A survey of autonomic computing
systems,” in Proc. 3rd Int. Conf. Auton. Syst. (ICAS), Athens, Greece,
scheduling of network devices made of soft electronics. 2007, p. 26.
[13] N. Samaan and A. Karmouch, “Towards autonomic network manage-
ment: An analysis of current and future research directions,” IEEE
IX. C ONCLUSION Commun. Surveys Tuts., vol. 11, no. 3, pp. 22–36, 3rd Quart., 2009.
[14] M. Barrere, R. Badonnel, and O. Festor, “Vulnerability assessment in
The area of network automation has attracted significant autonomic networks and services: A survey,” IEEE Commun. Surveys
attention in research and industry. This is due to the growing Tuts., vol. 16, no. 2, pp. 988–1004, 2nd Quart., 2014.
[15] A. Stamou, N. Dimitriou, K. Kontovasilis, and S. Papavassiliou, [36] R. Chaparadza et al., “Implementation guide for the ETSI AFI GANA
“Autonomic handover management for heterogeneous networks in a model: A standardized reference model for autonomic networking,
future Internet context: A survey,” IEEE Commun. Surveys Tuts., cognitive networking and self-management,” in Proc. IEEE Globecom
vol. 21, no. 4, pp. 3274–3297, 4th Quart., 2019. Workshops (GC Wkshps), Atlanta, GA, USA, 2013, pp. 935–940.
[16] A. Sadasivarao, S. Bardhan, S. Syed, B. Lu, and L. Paraschis, [37] R. Chaparadza, M. Wodczak, T. Ben Meriem, P. De Lutiis,
“Optonomic: Architecture for secure autonomic optical transport N. Tcholtchev, and L. Ciavaglia, “Standardization of resilience & sur-
networks,” in Proc. IFIP/IEEE Symp. Integr. Netw. Service Manag. vivability, and autonomic fault-management, in evolving and future
(IM), Arlington, VA, USA, 2019, pp. 321–328. networks: An ongoing initiative recently launched in ETSI,” in
[17] Z. Zhao et al., “Autonomic communications in software-driven Proc. 9th Int. Conf. Design Rel. Commun. Netw. (DRCN), Budapest,
networks,” IEEE J. Sel. Areas Commun., vol. 35, no. 11, Hungary, 2013, pp. 331–341.
pp. 2431–2445, Nov. 2017. [38] A. Richard et al. (May 2019). Autonomous Networks: Empowering
[18] A. Binsahaq, T. R. Sheltami, and K. Salah, “A survey on autonomic Digital Transformation for Telecoms Industry. Accessed:
provisioning and management of QoS in SDN networks,” IEEE Access, Jan. 8, 2021. [Online]. Available: https://www.tmforum.org/wp-
vol. 7, pp. 73384–73435, 2019. content/uploads/2019/05/22553-Autonomous-Networkswhitepaper.pdf
[19] K. Dhanya, C. Jeyalakshmi, and A. Balakumar, “A secure autonomic [39] J. Matias, J. Garay, N. Toledo, J. Unzilla, and E. Jacob, “Toward an
mobile ad-hoc network based trusted routing proposal,” in Proc. Int. SDN-enabled NFV architecture,” IEEE Commun. Mag., vol. 53, no. 4,
Conf. Comput. Commun. Informat. (ICCCI), Coimbatore, India, 2019, pp. 187–193, Apr. 2015.
pp. 1–6. [40] The 5G Guide GSMA 2019. Accessed: Nov. 11, 2020. [Online].
[20] J. Oliveira et al., “Toward terabit autonomic optical networks based on Available: https://www.gsma.com/wp-content/uploads/2019/04/The-
a software defined adaptive/cognitive approach [invited],” IEEE/OSA 5G-Guide_GSMA_2019_04_29_compressed.pdf
J. Opt. Commun. Netw., vol. 7, no. 3, pp. A421–A431, Mar. 2015.
[41] G. Bouabene, C. Jelger, C. Tschudin, S. Schmid, A. Keller, and
[21] A. H. Shamsan and A. R. Faridi, “Network softwarization for IoT:
M. May, “The autonomic network architecture (ANA),” IEEE J. Sel.
A survey,” in Proc. 6th Int. Conf. Comput. Sustain. Global Develop.
Areas Commun., vol. 28, no. 1, pp. 4–14, Jan. 2010.
(INDIACom), New Delhi, India, Mar. 2019, pp. 1163–1168.
[22] Q. Qi, W. Wang, X. Gong, and X. Que, “A SDN-based network vir- [42] Automated Network Management. Accessed: Nov. 11, 2020. [Online].
tualization architecture with autonomie management,” in Proc. IEEE Available: https://www.ietf.org/topics/netmgmt/
Globecom Workshops (GC Wkshps), Austin, TX, USA, Dec. 2014, [43] “FG NET-2030 technical specification on network 2030 architec-
pp. 178–182. ture framework,” in FG-NET2030 Focus Group on Technologies for
[23] W. Wang, Q. Qi, X. Gong, Y. Hu, and X. Que, “Autonomic QoS man- Network 2030, Int. Telecommun. Union (ITU), Geneva, Switzerland,
agement mechanism in software defined network,” China Commun., 2020.
vol. 11, no. 7, pp. 13–23, Jul. 2014. [44] P. Arcaini, E. Riccobene, and P. Scandurra, “Modeling and analyz-
[24] M. B. Chiosi et al., “Network functions virtualisation,” Eur. ing MAPE-K feedback loops for self-adaptation,” in Proc. IEEE/ACM
Telecommun. Stand. Inst., Sophia Antipolis, France, Rep. DGS/NFV- 10th Int. Symp. Softw. Eng. Adapt. Self Manag. Syst., Florence, Italy,
REL004, 2012. [Online]. Available: http://portal.etsi.org/NFV/NFV_ May 2015, pp. 13–23.
White_Paper.pdf [45] F. Granelli and R. Bassoli, “Autonomic mobile virtual network oper-
[25] Á. L. V. Caraguay, P. J. Ludeña-González, R. V. T. Tandazo, and ators for future generation networks,” IEEE Netw., vol. 32, no. 5,
L. I. B. López, “SDN/NFV architecture for IoT networks,” in Proc. 14th pp. 76–84, Sep./Oct. 2018.
Int. Conf. Web Inf. Syst. Technol. Vol. 1 (ITSCO), 2018, pp. 425–429. [46] J. O. Kephart and D. M. Chess, “The vision of autonomic computing,”
[26] W. Lee, T. Na, and J. Kim, “How to create a network slice?—A 5G core Computer, vol. 36, no. 1, pp. 41–50, Jan. 2003.
network perspective,” in Proc. 21st Int. Conf. Adv. Commun. Technol. [47] J. Strassner, “Autonomic networking—Theory and practice,” in Proc.
(ICACT), PyeongChang, South Korea, Feb. 2019, pp. 616–619. IEEE/IFIP Netw. Oper. Manag. Symp., vol. 1. Seoul, South Korea,
[27] How Big is IoT? 20.6 Billion Connected Devices By 2020. Accessed: Apr. 2004, pp. 1–17.
Dec. 3, 2020. [Online]. Available: https://mitechnews.com/internet- [48] J. Strassner and J. O. Kephart, “Autonomic systems and networks:
of-things/how-big-is-iot-20-6-billion-connected-devices-by-2020/#: Theory and practice,” in Proc. IEEE/IFIP Netw. Oper. Manag. Symp.
∼:text=Bi%20i%20Io%3F-,20.%20Billio%20 Connecte%20Device% (NOMS), Vancouver, BC, Canada, 2006, p. 588.
20B%202020.,smar%20 citie%20an%20connecte%20healthcare [49] J. Strassner, “Autonomic networking,” in Proc. 5th IEEE Workshop
[28] C. Wannachakrit and T. Anwar, “Development of network management Eng. Auton. Auton. Syst. (EASE), 2008, p. 3.
system for network services provider,” in Proc. Malaysian Conf. Softw. [50] A. Tizghadam and A. Leon-Garcia, “AORTA: Autonomic network con-
Eng., Johor Bahru, Malaysia, Dec. 2011, pp. 439–444. trol and management system,” in Proc. IEEE INFOCOM Workshops,
[29] M. Behringer, J. Strassner, and J. Halpern, “The use of control Phoenix, AZ, USA, Apr. 2008, pp. 1–4.
loops in autonomic networking,” IETF, Internet-Draft. Accessed: [51] A. Mercian, P. Sharma, R. Aguiar, C. Chen, and D. Pinheiro, “UDAAN:
Aug. 31, 2020. [Online]. Available: https://tools.ietf.org/html/draft- Embedding user-defined analytics applications in network devices,”
strassner-anima-control-loops-01 in Proc. Workshop Netw. Meets AI ML, 2019, pp. 70–75. [Online].
[30] 3rd Generation Partnership Project, Technical Specification Group Available: https://doi.org/10.1145/3341216.3342216
Services and System Aspects, Management and Orchestration, Levels
[52] J. A. McCann and R. Sterritt, “Autonomic pervasive networks
of Autonomous Network. Accessed: Nov. 11, 2020. [Online]. Available:
(APNs)—Extended abstract,” in Proc. 7th IEEE Int. Conf. Workshops
https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDet
Eng. Auton. Auton. Syst., Oxford, U.K., 2010, pp. 145–148.
ails.aspx?specificationId=3756
[31] M. Wódczak et al., “Standardizing a reference model and autonomic [53] A. Liakopoulos, A. Zafeiropoulos, C. Marinos, M. Grammatikou,
network architectures for the self-managing future Internet,” IEEE N. Tcholtchev, and P. Gouvas, “Applying distributed monitoring tech-
Netw., vol. 25, no. 6, pp. 50–56, Nov. 2011. niques in autonomic networks,” in Proc. IEEE Globecom Workshops,
[32] T. B. Meriem, R. Chaparadza, B. Radier, S. Soulhi, Miami, FL, USA, Dec. 2010, pp. 498–502.
J.-A. Lozano-López, and A. Prakash, “GANA—Generic auto- [54] H. Zhang, M. Zhao, W. Wang, X. Gong, and X. Que, “A novel
nomic networking architecture,” ETSI, Sophia Antipolis, France, autonomic architecture for QoS management in wired network,” in
White Paper, 2016. Proc. IEEE Globecom Workshops, Miami, FL, USA, Dec. 2010,
[33] Autonomic Network Engineering for the Self-Managing Future Internet pp. 529–533.
(AFI): Scenarios, Use Cases and Requirements for Autonomic/Self- [55] M. ElGammal and M. Eltoweissy, “Towards aware, adaptive and auto-
Managing Future Internet, document ETSI GS AFI 001, ETSI, nomic sensor-actuator networks,” in Proc. IEEE 5th Int. Conf. Self
Sophia Antipolis, France, 2011. Adapt. Self Organizing Syst., Ann Arbor, MI, USA, 2011, pp. 210–211.
[34] “Network technologies (NTECH): Autonomic network engineering [56] J. C. Nobre and L. Z. Granville, “Towards consistency of policy
for the self-managing future Internet (AFI): Autonomicity and self- states in decentralized autonomic network management,” in Proc. IEEE
management in the broadband forum (BBF) architectures,” ETSI, Int. Symp. Policies Distrib. Syst. Netw., London, U.K., Jul. 2009,
Sophia Antipolis, France, ETSI Rep. TR 103 473, 2018. pp. 170–173.
[35] R. Chaparadza et al., “SDN enablers in the ETSI AFI GANA ref- [57] M. May, M. Siekkinen, V. Goebel, T. Plagemann, R. Chaparadza, and
erence model for autonomic management & control (emerging stan- L. Peluso, “Monitoring as first class citizen in an autonomic network
dard), and virtualization impact,” in Proc. IEEE Globecom Workshops universe,” in Proc. 2nd Bio Inspired Models Netw. Inf. Comput. Syst.,
(GC Wkshps), Atlanta, GA, USA, Dec. 2013, pp. 818–823. Budapest, Hungary, Dec. 2007, pp. 247–254.
[58] B. Mathieu et al., “Autonomic management of context-aware ambi- [80] R. Casellas, R. Vilalta, A. Mayoral, R. Martínez, R. Muñoz, and
ent overlay networks,” in Proc. 2nd Int. Conf. Commun. Netw. China, L. M. Contreras, “Control plane architectures enabling transport
Shanghai, China, 2007, pp. 727–733. network adaptive and autonomic operation,” in Proc. 19th Int. Conf.
[59] L. Noirie et al., “Semantic networking: Flow-based, traffic-aware, Transparent Opt. Netw. (ICTON), Girona, Spain, Jul. 2017, pp. 1–4.
and self-managed networking,” Bell Labs Techn. J., vol. 14, no. 2, [81] G. Saadon, Y. Haddad, and N. Simoni, “Dynamic architecture based
pp. 23–38, 2009. on network virtualization and distributed orchestration for management
[60] A. Tizghadam and A. Leon-Garcia, “Autonomic traffic engineering for of autonomic network,” in Proc. 15th Int. Conf. Netw. Service Manag.
network robustness,” IEEE J. Sel. Areas Commun., vol. 28, no. 1, (CNSM), Halifax, NS, Canada, Oct. 2019, pp. 1–5.
pp. 39–50, Jan. 2010. [82] S. Schuetz, K. Zimmermann, G. Nunzi, S. Schmid, and M. Brunner,
[61] S. Whiteson and P. Stone, “Towards autonomic computing: Adaptive “Autonomic and decentralized management of wireless access
network routing and scheduling,” in Proc. Int. Conf. Auton. Comput., networks,” IEEE Trans. Netw. Service Manag., vol. 4, no. 2,
New York, NY, USA, May 2004, pp. 286–287. pp. 96–106, Sep. 2007.
[62] J. Praveen, B. Praveen, and C. S. R. Murthy, “A first step towards [83] G. Xilouris et al., “Towards autonomic policy-based network service
autonomic optical burst switched networks,” in Proc. 2nd Int. Conf. deployment with SLA and monitoring,” in Proc. IEEE Conf. Netw.
Auton. Comput. (ICAC), Seattle, WA, USA, Jun. 2005, pp. 306–307. Funct. Virtualization Softw. Defined Netw. (NFV-SDN), Verona, Italy,
[63] Y. Qiang, Y. Li, and J. Chen, “The workload adaptation in autonomic Nov. 2018, pp. 1–2.
DBMSs based on layered queuing network model,” in Proc. 2nd Int. [84] L. Gifre, J.-L. Izquierdo-Zaragoza, M. Ruiz, and L. Velasco,
Workshop Knowl. Discovery Data Min., Moscow, Russia, Jan. 2009, “Autonomic disaggregated multilayer networking,” IEEE/OSA J. Opt.
pp. 781–785. Commun. Netw., vol. 10, no. 5, pp. 482–492, May 2018.
[64] A. Gumaste, T. Das, A. Mathew, and A. Somani, “An autonomic vir- [85] G. Asaamoning and P. Mendes, “A study for a name-based coordi-
tual topology design and two-stage scheduling algorithm for light-trail nation of autonomic IoT functions,” in Proc. 14th Int. Conf. Wireless
WDM networks,” IEEE/OSA J. Opt. Commun. Netw., vol. 3, no. 4, Mobile Comput. Netw. Commun. (WiMob), Limassol, Cyprus, 2018,
pp. 372–389, Apr. 2011. pp. 296–302.
[65] A. A. F. Loureiro and L. B. Ruiz, “Autonomic wireless networks in [86] M. Grajzer and M. Głabowski, “Neighbor discovery++—A scalable
smart environments,” in Proc. 5th Annu. Conf. Commun. Netw. Service and robust address auto-configuration for future Internet of Things
Res., Fredericton, NB, Canada, 2007, pp. 5–7. networks,” IEEE Access, vol. 7, pp. 61083–61108, 2019.
[66] A. Moursy, A. Aly, B. Xu, D. Perkins, and M. Bayoumi, “Testbed [87] L. Velasco, L. Gifre, and M. Ruiz, “Autonomic content delivery
implementation for autonomic network performance management network service,” in Proc. 21st Int. Conf. Transparent Opt. Netw.
of wireless mesh networks,” in Proc. IEEE Globecom Workshops, (ICTON), Angers, France, 2019, pp. 1–4.
Anaheim, CA, USA, 2012, pp. 903–907. [88] Y. Lee, R. Vilalta, R. Casellas, R. Martínez, and R. Muñoz, “Scalable
[67] L.-I. Rui, L.-M. Meng, and X.-S. Qiu, “Information modeling and auto- telemetry and network autonomics in ACTN SDN controller hierar-
nomic configuration management of femto network,” in Proc. 5th Int. chy,” in Proc. 19th Int. Conf. Transparent Opt. Netw. (ICTON), Girona,
Conf. Wireless Commun. Netw. Mobile Comput., Beijing, China, 2009, Spain, Jul. 2017, pp. 1–4.
pp. 1–5. [89] M. Rodrigues, P. Endo, J. Vitalino, G. Gonçalves, D. Sadok, and
[68] T. Braga, F. Silva, J. M. Nogueira, A. Loureiro, and L. B. Ruiz, “A tiny F. Wuhib, “Prototyping an autonomic cloud infrastructure to man-
and light-weight autonomic element for wireless sensor networks,” in age live streaming applications using a software defined network:
Proc. 4th Int. Conf. Auton. Comput. (ICAC), Jacksonville, FL, USA, Performance analysis and challenges,” in Proc. IEEE 6th Int. Conf.
Jun. 2007, p. 17. Cloud Comput. Technol. Sci., Singapore, 2014, pp. 591–598.
[69] Y. Liu, J. Zhang, M. Jiang, D. Raymer, and J. Strassner, “A model-
[90] D. Bogdanovic, “Autonomic agent for transport networks,” in Proc.
based approach to adding autonomic capabilities to network fault
Opt. Fiber Commun. Conf. Exhibit. (OFC), Los Angeles, CA, USA,
management system,” in Proc. IEEE Netw. Oper. Manag. Symp.
Mar. 2015, pp. 1–3.
(NOMS), Salvador, Brazil, 2008, pp. 859–862.
[91] S. van der Meer, J. Keeney, and L. Fallon, “5G networks must be
[70] N. Tcholtchev and R. Chaparadza, “Autonomic fault-management
autonomic!” in Proc. IEEE/IFIP Netw. Oper. Manag. Symp. (NOMS),
and resilience from the perspective of the network operation person-
Taipei, Taiwan, Apr. 2018, pp. 1–5.
nel,” in Proc. IEEE Globecom Workshops, Miami, FL, USA, 2010,
pp. 469–474. [92] H. Song, J. Bai, Y. Yi, J. Wu, and L. Liu, “Artificial intelligence enabled
[71] R. Kuchta, J. Kadlec, and R. Vrba, “Implementation of intrusion detec- Internet of Things: Network architecture and spectrum access,” IEEE
tion system for automation devices within virtual automation network,” Comput. Intell. Mag., vol. 15, no. 1, pp. 44–51, Feb. 2020.
in Proc. 4th Int. Conf. Syst., Gosier, France, Mar. 2009, pp. 243–246. [93] T. Mai, H. Yao, S. Guo, and Y. Liu, “In-network computing powered
[72] Y. Rebahi, N. Tcholtchev, R. Chaparadza, and V. N. Merekoulias, mobile edge: Toward high performance industrial IoT,” IEEE Netw.,
“Addressing security issues in the autonomic future Internet,” in Proc. vol. 35, no. 1, pp. 289–295, Jan./Feb. 2021.
IEEE Consum. Commun. Netw. Conf. (CCNC), Las Vegas, NV, USA, [94] H. Yao, T. Mai, J. Wang, Z. Ji, C. Jiang, and Y. Qian, “Resource trading
Jan. 2011, pp. 517–518. in blockchain-based industrial Internet of Things,” IEEE Trans. Ind.
[73] A. K. Marnerides, D. Hutchison, and D. P. Pezaros, “Autonomic Informat., vol. 15, no. 6, pp. 3602–3609, Jun. 2019.
diagnosis of anomalous network traffic,” in Proc. IEEE Int. Symp. [95] D. Rosendo, P. T. Endo, D. Sadok, and J. Kelner, “An autonomic
World Wireless Mobile Multimedia Netw. (WoWMoM), Montrreal, QC, and policy-based authorization framework for openflow networks,” in
Canada, Jun. 2010, pp. 1–6. Proc. 13th Int. Conf. Netw. Service Manag. (CNSM), Tokyo, Japan,
[74] M. B. Lehocine and M. Batouche, “Self-management in IP networks 2017, pp. 1–5.
using autonomic computing,” in Proc. Int. Conf. Workshop Netw. Future [96] V. Foteinos et al., “Experimental validation of autonomic traffic engi-
(NOF), Paris, France, 2014, pp. 1–3. neering,” in Proc. Eur. Conf. Netw. Commun. (EuCNC), Paris, France,
[75] B. Mokhtar and M. Eltoweissy, “Memory-enabled autonomic resilient 2015, pp. 533–537.
networking,” in Proc. 7th Int. Conf. Collab. Comput. Netw. [97] X. Yuan, H. Yao, J. Wang, T. Mai, and M. Guizani, “Artificial
Appl. Worksharing (CollaborateCom), Orlando, FL, USA, 2011, intelligence empowered QoS-oriented network association for next-
pp. 132–141. generation mobile networks,” IEEE Trans. Cogn. Commun. Netw., early
[76] C. Ma, X. Lin, H. Lv, and H. Wang, “ABSR: An agent based self- access, Mar. 11, 2021, doi: 10.1109/TCCN.2021.3065463.
recovery model for wireless sensor network,” in Proc. 8th IEEE Int. [98] L. Wang and E. Gelenbe, “Demonstrating voice over an autonomic
Conf. Depend. Auton. Secure Comput., Chengdu, China, Dec. 2009, network,” in Proc. IEEE Int. Conf. Auton. Comput., Grenoble, France,
pp. 400–404. 2015, pp. 139–140.
[77] R. Sterritt, “Autonomous and autonomic systems: Paradigm for engi- [99] N. Beigi-Mohammadi, M. Shtern, and M. Litoiu, “Adaptive load man-
neering effective software-based systems?” in Proc. 33rd Annu. IEEE agement of Web applications on software defined infrastructure,” IEEE
Softw. Eng. Workshop, Skovde, Sweden, Oct. 2009, p. 57. Trans. Netw. Service Manag., vol. 17, no. 1, pp. 488–502, Mar. 2020.
[78] A. Galis, S. Clayman, and L. Mamatas, “Towards autonomic manage- [100] L. Velasco et al., “An architecture to support autonomic slice
ment of software enabled networks,” in Proc. 8th Int. Symp. Adv. Topics networking,” J. Lightw. Technol., vol. 36, no. 1, pp. 135–141, Jan. 1,
Elect. Eng. (ATEE), Bucharest, Romania, May 2013, pp. 1–3. 2018.
[79] P. Zhang and J. Wang, “SIP based scheme for collaborative autonomic [101] G. Sakarkar and V. M. Thakar, “Autonomous software agent for local-
network management,” in Proc. 5th IEEE Int. Conf. Broadband Netw. ization,” in Proc. Int. Conf. Intell. Agent Multi Agent Syst., Chennai,
Multimedia Technol., Guilin, China, Nov. 2013, pp. 323–326. India, Jul. 2009, pp. 1–4.
[102] Q. Jiang, M. Xu, L. Tang, and Q.-B. Chen, “Research on load balanc- [124] M. Sharf and D. Zelazo, “Network feedback passivation of passivity-
ing based on multi-agent in ubiquitous networks,” in Proc. Int. Conf. short multi-agent systems,” IEEE Control Syst. Lett., vol. 3, no. 3,
Intell. Comput. Technol. Autom., vol. 3. Changsha, China, May 2010, pp. 607–612, Jul. 2019.
pp. 10–13. [125] J. Watada, A. Roy, R. Kadikar, H. Pham, and B. Xu, “Emerging
[103] P. Sethi, N. Chauhan, and D. Juneja, “A multi-agent hybrid protocol for trends, techniques and open issues of containerization: A review,” IEEE
data fusion and data aggregation in non-deterministic wireless sensor Access, vol. 7, pp. 152443–152472, 2019.
networks,” in Proc. Int. Conf. Inf. Syst. Comput. Netw., Mathura, India, [126] B. Loganayagi and S. Sujatha, “Creating virtual platform for cloud
Mar. 2013, pp. 211–214. computing,” in Proc. IEEE Int. Conf. Comput. Intell. Comput. Res.,
[104] J. Chen, X. Wang, Z. Cheng, and J. Qin, “On wireless sensor network Coimbatore, India, Dec. 2010, pp. 1–4.
mobile agent multi-objective optimization route planning algorithm,” [127] M. Ruffini and F. Slyne, “Moving the network to the cloud: The cloud
in Proc. IEEE Int. Conf. Agents (ICA), Beijing, China, Jul. 2017, central office revolution and its implications for the optical layer,” J.
pp. 101–103. Lightw. Technol., vol. 37, no. 7, pp. 1706–1716, Apr. 1, 2019.
[105] D. Lohani, P. Singh, and S. Varma, “Multi-agent data aggregation in [128] I. Al-Oqily and A. Karmouch, “A self-organizing composition towards
wireless sensor network using source grouping,” in Proc. Int. Conf. Adv. autonomic overlay networks,” in Proc. IEEE Netw. Oper. Manag. Symp.
Comput. Commun. Informat. (ICACCI), New Delhi, India, Sep. 2014, (NOMS), Salvador, Brazil, Apr. 2008, pp. 287–294.
pp. 2174–2179.
[129] D. Comer and A. Rastegarnia, “Toward disaggregating the SDN control
[106] S. Hussain, E. Shakshuki, and A. W. Matin, “Agent-based system archi-
plane,” IEEE Commun. Mag., vol. 57, no. 10, pp. 70–75, Oct. 2019.
tecture for wireless sensor networks,” in Proc. 20th Int. Conf. Adv. Inf.
Netw. Appl. Vol. 1 (AINA), vol. 2. Vienna, Austria, Apr. 2006, pp. 1–5. [130] K. Dolui and S. K. Datta, “Comparison of edge computing imple-
[107] C. Liao, K. Zhu, J. Tang, and S. Zhang, “Wireless sensor network mentations: Fog computing, cloudlet and mobile edge computing,” in
performance research for leach based on multi-agent simulation,” Proc. Global Internet Things Summit (GIoTS), Geneva, Switzerland,
in Proc. IEEE Int. Conf. Agents (ICA), Matsue, Japan, Sep. 2016, Jun. 2017, pp. 1–6.
pp. 98–99. [131] H. Tianfield, “Towards edge-cloud computing,” in Proc. IEEE Int. Conf.
[108] D. Zhang, C. Lu, and X. Jia, “Leader-following consensus of Big Data (Big Data), Seattle, WA, USA, Dec. 2018, pp. 4883–4885.
for multi-agent systems under event-triggering communication and [132] G. Mazlami, J. Cito, and P. Leitner, “Extraction of microservices
network-induced delays,” in Proc. Chin. Control Decis. Conf. (CCDC), from monolithic software architectures,” in Proc. IEEE Int. Conf. Web
Yinchuan, China, May 2016, pp. 2344–2349. Service (ICWS), Honolulu, HI, USA, Jun. 2017, pp. 524–531.
[109] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and coop- [133] L. De Lauretis, “From monolithic architecture to microservices archi-
eration in networked multi-agent systems,” Proc. IEEE, vol. 95, no. 1, tecture,” in Proc. IEEE Int. Symp. Softw. Rel. Eng. Workshops
pp. 215–233, Jan. 2007. (ISSREW), Berlin, Germany, Oct. 2019, pp. 93–96.
[110] M. A. B. Brasil, B. Bösch, F. R. Wagner, and E. P. de Freitas, [134] A. Dorri, S. S. Kanhere, and R. Jurdak, “Multi-agent systems: A
“Performance comparison of multi-agent middleware platforms for survey,” IEEE Access, vol. 6, pp. 28573–28593, 2018.
wireless sensor networks,” IEEE Sensors J., vol. 18, no. 7, [135] G. A. Oparin, V. G. Bogdanova, A. A. Pashinin, and S. A. Gorsky,
pp. 3039–3049, Apr. 2018. “Distributed solvers of applied problems based on microservices and
[111] Z. Cheng, T. Wang, and Y. Xin, “High-order distributed consensus in agent networks,” in Proc. 41st Int. Conv. Inf. Commun. Technol.
multi-agent networks,” in Proc. IEEE 7th Data Driven Control Learn. Electron. Microelectron. (MIPRO), Opatija, Croatia, May 2018,
Syst. Conf. (DDCLS), Enshi, China, May 2018, pp. 965–969. pp. 1415–1420.
[112] J. C. Moon and S. J. Kang, “A multi-agent architecture for intelligent [136] R. W. Collier, E. O’Neill, D. Lillis, and G. O’Hare, “MAMS:
home network service using tuple space model,” IEEE Trans. Consum. Multi-agent microservices*,” in Proc. Companion World Wide Web
Electron., vol. 46, no. 3, pp. 791–794, Aug. 2000. Conf., 2019, pp. 655–662. [Online]. Available: https://doi.org/10.
[113] T. Amudha, N. Saritha, and P. Usha, “A multi-agent based frame- 1145/3308560.3316509
work for dynamic service discovery and access using matchmaking [137] S. T. Arzo, R. Bassoli, F. Granelli, and F. H. P. Fitzek,
approach,” in Proc. Int. Conf. Intell. Agent Multi Agent Syst., Chennai, “Multi-agent based autonomic network management architecture,”
India, Jul. 2009, pp. 1–4. IEEE Trans. Netw. Service Manag., early access, Feb. 16, 2021,
[114] S. K. Patri, E. Grigoreva, W. Kellerer, and C. M. Machuca, “Rational doi: 10.1109/TNSM.2021.3059752.
agent-based decision algorithm for strategic converged network migra- [138] A. Odlyzko. (2000). The History of Communications and Its
tion planning,” IEEE/OSA J. Opt. Commun. Netw., vol. 11, no. 7, Implications for the Internet. Accessed: Mar. 20, 2020. [Online].
pp. 371–382, Jun. 2019. Available: https://ssrn.com/abstract=235284
[115] M. Mohammadian, “Network security risk assessment using intelligent [139] A. A. Huurderman, The Worldwide History of Telecommunications.
agents,” in Proc. Int. Symp. Agent Multi Agent Syst. Robot. (ISAMSR), New York, NY, USA: Wiley, 2003.
Putrajaya, Malaysia, Aug. 2018, pp. 1–6.
[140] “Next generation protocols (NGP); intelligence-defined network
[116] S. Ouiazzane, M. Addou, and F. Barramou, “A multi-agent model for
(IDN),” ETSI, Sophia Antipolis, France, ETSI Rep. GR NGP 006,
network intrusion detection,” in Proc. 1st Int. Conf. Smart Syst. Data
2018.
Sci. (ICSSD), Rabat, Morocco, Oct. 2019, pp. 1–5.
[117] X. Zhang, K. Shi, S. Zhang, D. Li, and R. Xia, “Virtual agent clustering [141] M. Katz, M. Matinmikko-Blue, and M. Latva-Aho, “6Genesis flagship
based mobility management over the satellite networks,” IEEE Access, program: Building the bridges towards 6G-enabled wireless smart soci-
vol. 7, pp. 89544–89555, 2019. ety and ecosystem,” in Proc. IEEE 10th Latin Amer. Conf. Commun.
[118] K. Huang, S. Srivastava, and D. Cartes, “Decentralized reconfiguration (LATINCOM), Guadalajara, Mexico, Nov. 2018, pp. 1–9.
for power systems using multi agent system,” in Proc. 1st Annu. IEEE [142] F. Cavaliere, E. Prati, L. Poti, I. Muhammad, and T. Catuogno,
Syst. Conf., Honolulu, HI, USA, Apr. 2007, pp. 1–6. “Secure quantum communication technologies and systems: From labs
[119] Z. Taoqing and H. Xiaoying, “Self-adaptive network service research to markets,” Quantum Rep., vol. 2, pp. 80–106, Jan. 2020.
based on multi-agents in hierarchy,” in Proc. 1st Int. Conf. Netw. [143] B. Raouf et al., “DeepNFV: A lightweight framework for intelligent
Distrib. Comput., Hangzhou, China, Oct. 2010, pp. 314–316. edge network functions virtualization,” J. Internet Service Appl., vol. 9,
[120] G.-Z. Liu and D. Juan, “Design of distributed network management no. 1, p. 16, Jun. 2018.
system based on multi-agent,” in Proc. 3rd Int. Symp. Inf. Process., [144] O. Mohammed and J. Kianfar, “A machine learning approach to short-
Qingdao, China, Oct. 2010, pp. 433–436. term traffic flow prediction: A case study of interstate 64 in Missouri,”
[121] H. Tianfield, “Multi-agent based autonomic architecture for network in Proc. IEEE Int. Smart Cities Conf. (ISC2), Kansas City, MO, USA,
management,” in Proc. IEEE Int. Conf. Ind. Informat. (INDIN), Banff, Sep. 2018, pp. 1–7.
AB, Canada, 2003, pp. 462–469. [145] F. Slim, F. Guillemin, A. Gravey, and Y. Hadjadj-Aoul, “Towards a
[122] N. A. M. Subha, G.-P. Liu, and N. H. M. Yusof, “External consen- dynamic adaptive placement of virtual network functions under ONAP,”
sus in networked multi-agent systems with random network delay,” in in Proc. IEEE Conf. Netw. Funct. Virtualization Softw. Defined Netw.
Proc. 57th Annu. Conf. Soc. Instrum. Control Eng. Jpn. (SICE), Nara, (NFV-SDN), Berlin, Germany, Nov. 2017, pp. 210–215.
Japan, Sep. 2018, pp. 427–432. [146] J. Pei, P. Hong, K. Xue, and D. Li, “Efficiently embedding service
[123] M. Sharf and D. Zelazo, “Analysis and synthesis of MIMO multi-agent function chains with dynamic virtual network function placement in
systems using network optimization,” IEEE Trans. Autom. Control, geo-distributed cloud system,” IEEE Trans. Parallel Distrib. Syst.,
vol. 64, no. 11, pp. 4512–4524, Nov. 2019. vol. 30, no. 10, pp. 2179–2192, Oct. 2019.
[147] T. Subramanya and R. Riggio, “Machine learning-driven scaling and Fabrizio Granelli (Senior Member, IEEE) received
placement of virtual network functions at the network edges,” in the M.Sc. and Ph.D. degrees from the University of
Proc. IEEE Conf. Netw. Softw. (NetSoft), Paris, France, Jun. 2019, Genoa, Genoa, Italy, in 1997 and 2001, respectively.
pp. 414–422. He is an Associate Professor with the Department
[148] H. Tang, D. Zhou, and D. Chen, “Dynamic network function instance of Information Engineering and Computer Science,
scaling based on traffic forecasting and VNF placement in opera- University of Trento, Trento, Italy. From 2012 to
tor data centers,” IEEE Trans. Parallel Distrib. Syst., vol. 30, no. 3, 2014, he was an Italian Master School Coordinator
pp. 530–543, Mar. 2019. in the framework of the European Institute of
[149] D. R. Jeske, P. J. Lin, C. Rendon, R. Xiao, and B. Samadi, “Synthetic Innovation and Technology ICT Labs Consortium.
data generation capabilties for testing data mining tools,” in Proc. IEEE He has authored or coauthored more than 250 papers
Mil. Commun. Conf. (MILCOM), Washington, DC, USA, Oct. 2006, published in international journals, books and con-
pp. 1–6. ferences on networking. He is a coordinator of the research and didactical
[150] J. W. Hu, I. T. Bowman, A. Nica, and A. Goel, “Distribution-driven, activities on computer networks.
embedded synthetic data generation system and tool for RDBMS,” in Mr. Granelli was an IEEE ComSoc Distinguished Lecturer for 2012 to
Proc. IEEE 35th Int. Conf. Data Eng. Workshops (ICDEW), Macao, 2015, the IEEE ComSoc Director for Online Content from 2016 to 2017, and
China, Apr. 2019, pp. 113–115. the IEEE ComSoc Director for Educational Services from 2018 to 2019.
[151] D. Libes, D. Lechevalier, and S. Jain, “Issues in synthetic data gener-
ation for advanced manufacturing,” in Proc. IEEE Int. Big Data (Big
Data), Boston, MA, USA, Dec. 2017, pp. 1746–1754.
[152] A. Londhe and P. P. Rao, “Platforms for big data analytics: Trend
towards hybrid era,” in Proc. Int. Conf. Energy, Commun. Data Anal.
Soft Comput. (ICECDS), Chennai, India, Aug. 2017, pp. 3235–3238. Riccardo Bassoli (Member, IEEE) received the
[153] C. A. Ardagna, P. Ceravolo, and E. Damiani, “Big data analytics as- B.Sc. and M.Sc. degrees in telecommunications
a-service: Issues and challenges,” in Proc. IEEE Int. Conf. Big Data engineering from the University of Modena
(Big Data), Washington, DC, USA, Dec. 2016, pp. 3638–3644. and Reggio Emilia, Modena, Italy, in 2008 and
[154] L. Xiao, X. Wan, X. Lu, Y. Zhang, and D. Wu, “IoT security techniques 2010, respectively, and the Ph.D. degree from
based on machine learning: How do IoT devices use ai to enhance 5G Innovation Centre, University of Surrey,
security?” IEEE Signal Process. Mag., vol. 35, no. 5, pp. 41–49, Guildford, U.K., in November 2016.
Sep. 2018. He is a Senior Researcher with the Deutsche
[155] R. Atat, L. Liu, J. Wu, G. Li, C. Ye, and Y. Yang, “Big data meet Telekom Chair of Communication Networks,
cyber-physical systems: A panoramic survey,” IEEE Access, vol. 6, Faculty of Electrical and Computer Engineering,
pp. 73603–73636, 2018. Technische Universität Dresden, Dresden, Germany.
[156] J. Kang, J. B.-H. Tok, and Z. Bao, “Self-healing soft electronics,” Nat. From 2011 to 2015, he was also a Marie Curie Early Stage Researcher
Electron., vol. 2, pp. 144–150, Apr. 2019. with the Instituto de Telecomunicações, Aveiro, Portugal, and a Visiting
Researcher with Airbus Defence and Space, Élancourt, France.
Michael Devetsikiotis (Fellow, IEEE) received the

Diploma degree in electrical engineering from the
Sisay Tadesse Arzo (Graduate Student Member, Aristotle University of Thessaloniki, Thessaloniki,
IEEE) received the B.Sc. degree from Hawassa Greece, in 1988, and the M.S. and Ph.D. degrees
University, Awassa, Ethiopia, in 2007, and the M.Sc. in electrical engineering from North Carolina State
degree in telecommunication engineering from the University, Raleigh, NC, USA, in 1990 and 1993,
University of Trento, Trento, Italy, in 2013, where respectively.
he is currently pursuing the Ph.D. degree. He joined the University of New Mexico,
He is a Former Visiting Research Fellow with Albuquerque, NM, USA, in July 2016, as a Professor
the University of Luxembourg, Luxembourg City, and the Chair of the ECE Department, School of
Luxembourg, in 2011 and 2012. He has more Engineering. His work has received well over 6700
than five years of industrial experience in Telecom citations. In 2017, he was inducted to the NC State ECE Alumni Hall of
from Huawei Technology, Shenzhen, China, and Fame. His research work has resulted in 44 refereed journal articles, 140
Ethio-Telecom, Addis Ababa, Ethiopia, and five years of academic expe- refereed conference papers, and 62 invited presentations, in the area of
rience working as a Lecturer with Addis Ababa University, Addis Ababa. design and performance evaluation of telecommunication networks, complex
His research interests include network softwarization, software-defined sociotechnical and cyber–physical systems, efficient simulation, and smart grid
networking, network function virtualization, Internet of Things, and network communications.
automation.
Frank H. P. Fitzek (Senior Member, IEEE)

received the diploma (Dipl.-Ing.) degree in electrical
engineering from the University of Technology—
Rheinisch-Westfälische Technische Hochschule—
Aachen, Aachen, Germany, in 1997, and the Ph.D.
Claire Naiga (Student Member, IEEE) received (Dr.-Ing.) degree in electrical engineering from
the B.Sc. degree in telecommunications engineering the Technical University Berlin, Berlin, Germany,
from Makerere University, Kampala, Uganda, and in 2002
the M.Sc. degree in information and communications In 2002, he was an Adjunct Professor with the
engineering from the University of Trento, Trento, University of Ferrara, Ferrara, Italy. He is a Professor
Italy, in 2018 and 2021, respectively. and the Head of the Deutsche Telekom Chair of
Her current research interests include network Communication Networks with Technische Universität Dresden, Dresden,
automation involving SDN and network function Germany, also coordinating the 5G Lab, Dresden. He is the spokesman of the
virtualization, and MASs. DFG Cluster of Excellence CeTI. In 2003, he joined the Aalborg University,
Aalborg, Denmark, as an Associate Professor and later became a Professor.
View publication stats

A Theoretical Discussion and Survey of Network Automation For IoT Challenges and Opportunity

Uploaded by

Copyright:

Available Formats

A Theoretical Discussion and Survey of Network Automation For IoT Challenges and Opportunity

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Theoretical Discussion and Survey of Network Automation For IoT Challenges and Opportunity

Uploaded by

Copyright:

Available Formats

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

A Theoretical Discussion and Survey of Network Automation for IoT: Challenges

Article in IEEE Internet of Things Journal · August 2021

Sisay Tadesse Arzo Claire Naiga Serugunda

SEE PROFILE SEE PROFILE

Fabrizio Granelli Riccardo Bassoli

SEE PROFILE SEE PROFILE

The user has requested enhancement of the downloaded file.

A Theoretical Discussion and Survey of Network

evolving, showing an increase in network size and complexity.

Fig. 2. High-level NFV framework [10].

surveys, and several past and present solutions in networking.

Fig. 4. Organization of this article and possible navigation orders.

3) Discussion of conceptual and technological frameworks

Fig. 7. Unified SDN-NFV architectural framework.

Fig. 6. ETSI-NFV architectural framework [10].

This type of service requires an enhancement of the existing

A. Network Management and Control as Cyclic Process

Fig. 10. Overview of the GANA reference model.

be important to self-* properties implementation. To iden-

Fig. 11. ANIMA reference model and the structure of an AN.

implements the specified functions, e.g., scheduling. The

C. Autonomic Networking Integrated Model and Approach

QoS management in wired network [54]. This architecture

deployment of functions. Such design provides the option of

Fig. 16. General overview of a MAS.

Fig. 17. Comparison of microservice and MAS.

Devices had to be configured one at a time making the con-

Michael Devetsikiotis (Fellow, IEEE) received the

Frank H. P. Fitzek (Senior Member, IEEE)

You might also like