5g1 s2.0 S0140366418300380 Main

Computer Communications 129 (2018) 248–268
Contents lists available at ScienceDirect
Computer Communications
journal homepage: www.elsevier.com/locate/comcom
Review
From 4G to 5G: Self-organized network management meets machine T

learning
⁎
Jessica Moysen ,a, Lorenza Giupponib
a
Department of Signal and Theory Communications, Universitat Politècnica de Catalunya (UPC), C/ Esteve Terradas, 7, Castelldefels (Barcelona) 08860, Spain
b
Communications Network Division, Centre Tecnològic de Telecomunicacions de Catalunya (CTTC), Av. Carl Friedrich Gauss 7, Castelldefels (Barcelona) 08860, Spain
A R T I C LE I N FO A B S T R A C T
Keywords: Self-organization as applied to cellular networks is usually referred to Selforganizing Networks (SONs), and it is a
Network management key driver for improving Operations,Administration, and Management (OAM) activities. SON aims at reducing
Self-organizing networks the cost of installation and management of 4G and future 5G networks, by simplifying operational tasks through
Mobile networks the capability to configure, optimize and heal itself. To satisfy 5G network management requirements, this
Machine learning
autonomous management vision has to be extended to the end to end network. In literature and also in some
Big data
instances of products available in the market, Machine Learning (ML) has been identified as the key tool to
implement autonomous adaptability and take advantage of experience when making decisions. In this paper, we
survey how 5G network management, with an end-to-end perspective of the network, can significantly benefit
from ML solutions. We review and provide the basic concepts and taxonomy for SON, network management and
ML. We analyze the available state of the art in the literature, standardization, and in the market. We pay special
attention to 3rd Generation Partnership Project (3GPP) evolution in the area of network management and to the
data that can be extracted from 3GPP networks, in order to gain knowledge and experience in how the network is
working, and improve network performance in a proactive way. Finally, we go through the main challenges
associated with this line of research, in both 4G and in what 5G is getting designed, while identifying new
directions for research.
1. Introduction work is inspired by a set of requirements defined by the operators’

Alliance Next Generation Mobile Networks (NGMN). The main objec-
Traditionally, and up to 4G, the evolution from one generation of tive of SON can be roughly classified into three main points: 1) to bring
mobile networks to another, has been driven by hardware technology intelligence and autonomous adaptability into cellular networks, 2) to
advancements. The revolution of 5G is different, and novel advance- reduce capital and operational expenditures (CAPEX/OPEX), 3) to en-
ments of software technology will be critical, especially in the way the hance network performances in terms of network capacity, coverage,
network will be managed. With the advent of these software advance- spectral efficiency, offered service/experience, etc. In particular, SON is
ments, and unprecedented levels of computational capacity, the vision a driving technology providing capabilities for self-configuration to au-
of autonomous Network Management (NM) can be put into practice tomatize the network deployment phase, self-optimization to optimize
taking advantage of cross-disciplinary knowledge advancements in the the network performance while it is operating, by properly adjusting
area of ML. This vision aligns with the concepts of self-awareness, self- network parameters, and self-healing to detect faults and to correct/
configuration, self-optimization, and self-healing, which have already prevent network failures. The overall complex SON problem has been
been defined in the area of NM, in the access segment of Long Term decomposed in a set of useful use cases, which have been identified by
Evolution (LTE), through the concept of Self-organizing Network 3GPP, NGMN, 5G Infrastructure Public Private Partnership (5GPPP)
(SON). and different EU projects [1–6]. The academic literature has dedicated
SON is a common term used to refer to mobile network automation significant effort to SON algorithms in the context of the above men-
and minimization of human intervention in the cellular/wireless net- tioned 4G use cases, providing smart solutions to optimize network
work management. This concept has been introduced by 3GPP in operator performance, expenses and users’ experience. Many of these
Release 8 and it has been expanding across subsequent releases. 3GPP works are already reviewed here [7]. The market also offers already
⁎
Corresponding author.
E-mail addresses: jessica.moysen@tsc.upc.edu (J. Moysen), lorenza.giupponi@cttc.es (L. Giupponi).
https://doi.org/10.1016/j.comcom.2018.07.015
Received 13 January 2018; Received in revised form 4 July 2018; Accepted 6 July 2018
Available online 27 July 2018
0140-3664/ © 2018 Elsevier B.V. All rights reserved.
J. Moysen, L. Giupponi Computer Communications 129 (2018) 248–268
complete sets of SON solutions, (e.g. [8–12], among others) many self-aware, by exploiting the information already available in the net-
products have been advertised and presented in Mobile World Congress work to gain experience in the network management, 2) more self-
(MWC) 2016–2018 [13–16]. For example, AirHop’s eSON from Jio & adaptive, by exploiting intelligent control decisions procedures which
AirHop communications [11], which employs a multi-vendor, multi- allow to automatize the decision processes based on the experience. In
technology, real-time SON solution based on scalable and virtualized this paper, we focus on the application of ML to 5G network manage-
software platform has recently been awarded for the 2016 Small Cell ment. Differently from other surveys on SON proposed before [7,25,26]
Forum Heterogenous Network (Het-Net) management software and or from other surveys related with 5G network management [27], we
service award [17]. focus here on the study and analysis of the available literature on SON
Up to date SON solutions have been designed as an add-on feature and network management considering ML as the tool to implement
to LTE, associated to a specific set of use cases affecting the Radio automation and self-organization, from a 5G perspective. Another work
Access Network (RAN). As a result, as of today, SON represents mainly [26] has focused on the application of ML to SON, with emphasis on 4G
a RAN solution, rather than an integral feature of the end-to-end net- LTE and on the specific set of SON use cases defined for the LTE RAN.
work, capable of fulfilling the requirements of different kinds of vertical However, differently from this work, in this paper we focus on NG SON
service instances [18]. In addition, several open challenges are still for 5G networks, and on how ML will be the driving tool for autono-
unsolved, like the problem of coordination of SON functions [19,20], or mous end-to-end 5G network management. In particular, we review
the proper solution of the trade-off between centralized and distributed and provide the basic concepts and taxonomy of traditional SON and 5G
SON implementations [21]. To achieve this vision, the networking network management in Section 2. We pay special attention to the
world is exploring new directions. On the one hand, Software Defined evolution of 3GPP in the area, following its nomenclature. We then
Network (SDN) is expected to bring the benefits of programmability provide, in Section 3, guidelines to select the most appropriate ML al-
and flexibility to the management of the network, while Network gorithm and approach, based on the network management issue to
Functions Virtualisation (NFV) is expected to extend the economy of address. In Section 4, we review the main sources of information re-
scale of the Information Technology industry to the Telecom industry, levant for a knowledge based network management, data that is actu-
by simplifying the deployment of new features and functionalities ally already generated by the network, and we survey the literature on
through implementation in Commercial off-the-shelf (COTS) without ML-based network management. We then highlight challenges for fu-
the need of specialized hardware. These virtualized functions can be ture works in Section 5. Finally, Section 6 concludes the survey.
centralized in the cloud (e.g, cloud-RAN) or can be migrated to the edge
through Mobile Edge Computing (MEC), where new processing cap- 2. Self organizing network (SON) and network management
abilities are introduced in the base station for new applications [22]. On
the other hand, as we observed in [23] a huge amount of data is cur- SON is a key driver to maximize total performance in cellular net-
rently already generated in 4G networks during normal operations by works. The main idea is to bring into them intelligence and autonomous
control and management functions, and more is expected to come in 5G adaptability by diminishing human involvement, while enhancing
networks due to the densification process [24], heterogeneity in layers network performance, in terms of network capacity, coverage and
and technologies, the additional control and management complexity in service quality.
NFV and SDN architectures, the advent of Machine to Machine (M2M) The main motivation behind the increasing interest in the in-
and Internet of Things (IoT) paradigms, the increasing variety of ap- troduction of SON from operators, standardization bodies and projects
plication and services, each with distinct traffic patterns and Quality of is twofold. On the one hand, from the market perspective, the ever
Service (QoS)/Quality of Experience (QoE) requirements, etc. As a re- increasing demand for a diversity of offered services, and the need to
sult, 5G architectures should support Big Data platforms to handle reduce the time to market, further add to the pressure to remain
measurements, traces, logs and configuration data from the entire competitive by effectuating cost reductions. On the other hand, from
network, providing the basis for automatic decision making. the technical perspective, the complexity and large scale of future radio
5G network management is expected to provide a whole new set of access technologies imposes significant operational challenges due to
challenges due to 1) the need to manage future network complexity, the multitude of tunable parameters and the intricate dependencies
associated to ultra dense deployments, heterogeneous nodes, networks, among them. In addition, the advent of heterogeneous networks is ex-
applications, different RANs coexisting in the same setting, 2) the need pected to tremendously increase the number of nodes in this new
to manage very dynamic networks, where operators may not have ecosystem, so classic manual and field trial design approaches are just
complete control on the deployment of some nodes (e.g. femto-cells impractical.
deployed by end users), energy saving policies are in place generating a Similarly, manual optimization processes or fault diagnosis and
fluctuating number of nodes, active antennas are a reality, etc. 3) the cure, performed by experts are no longer efficient and need to be au-
need to support 1000x traffic, and 10x users, and improve energy ef- tomatized, as this causes time intensive experiments with limited op-
ficiency, 4) the need to improve the experience of the users by enabling erational scope, or delayed, manual and poor handling of cell/sites
Gbps speeds, and highly reduced latency, 5) the need to manage new failures. Key operational tasks, such as radio network planning and
virtualized architectures, 6) the need to handle heterogeneous spectrum optimization are largely separated nowadays and this causes intrinsic
access privileges through novel LTE-Unlicensed (LTE-U), Licensed shortcomings like the abstraction of access technologies for network
Assisted Access (LAA), MuLTEFire, New Radio (NR), NR-U paradigms planning purposes, or the consideration of performance indicators that
and the availability of both traditional sub- 6 GHz bands, and above are of limited relevance to the end user’s service perception. These
6 GHz mmWave bands. In this challenging context, we believe that the problems have been approached through SON by European projects
use of next generation (NG) SON and of smart network management such as SOCRATES [2], and Gandalf [1]. Also FP7 and 5GPPP EU
policies is crucial and inevitable for operators running multi-RAT, projects have been dealing with SON. In particular FP7 SEMAFOUR [3],
multi-vendor, multi-layer networks, where an overwhelming number of which develops a unified self-management to operate complex HetNets.
parameters need to be configured and optimized. Among 5GPPP projects, we highlight SESAME [4], which proposes the
We believe that Machine Learning (ML) can be effectively used to Cloud-Enabled Small Cell (CESC) concept, i.e., a new multi-operator
allow the network to learn from experience, while improving perfor- enabled Small Cell by deploying Network Functions Virtualisation
mance. In particular, big data analytics, through the analysis of data (NFV), supporting powerful self-management inside the access network
already generated by the network, can pursue the self-awareness by infrastructure. In terms of self-organized network management in SDN
driving the network management from reactive to predictive. The use of and NFV, the SELFNET project aims at enabling the use of ML to
NG SON based on ML allows operators to make the networks 1) more achieve real time autonomous 5G network management [5]. In
249
Fig. 1. Self-organizing networks.
particular, the project explores a smart integration of state-of-the-art functions. Finally, concepts of MDT and Energy Saving (ES) are also
technologies in SDN, NFV), SON, Cloud computing, Artificial in- introduced and then enrich in Release 11. Release 11 SON functions are
telligence. The COGNET project [6] has similar objectives and aims at related to the automated management of heterogeneous networks. It
developing several operators’ use cases by applying ML algorithms. includes mobility robustness optimization enhancements and inter-
SON has been introduced by 3GPP as a key component of LTE radio access technology Handover (HO) optimization. Release 12 in-
network starting from the first release of this technology in Release 8, troduces optimization and enhancements for small cells including de-
and expanding to subsequent releases. In SOCRATES [2] and in 3GPP ployments in dense areas. In Release 13 novel concepts of unlicensed
[28], meaningful SON use cases have been defined, which can be LTE have been introduced. Besides that, Release 13 studied the in-
classified according to the phases of the life cycle of a cellular systems tensification of OAM, with respect to centralized and distributed ar-
(planning, deployment, maintenance and optimization) into: self-con- chitecture. In particular, focuses on distributed MLB, as well as on en-
figuration, self-healing and self-optimization, as depicted in Fig. 1. In hanced NM or centralized CCO. Release 14 focuses on meeting the 5G
this section, first we give an overview of the evolution of SON in 3GPP. requirements in terms of latency reduction, use of unlicensed spectrum
We go through self-configuration, optimization and healing function- in a fair manner, support for carrier aggregation, energy efficiency at
alities, introducing the use cases that have been defined for each one of OAM level, SON for active antennas, etc. Finally, Release 15 has studied
them. We discuss about the self-coordination problem, to handle the the pre-configuration of the eNB with alternative coverage configura-
potential conflicts that may exist between the parallel execution of tions to select the coverage change during Cell Shaping and set the
multiple SON functions. We then present a standardized database used proper handover margin with neighbour cells. Self-Configuration and
for different 3GPP use cases known as the Minimization of Drive Tests Self-Optimization are not finalized yet, and are targeted for completion
(MDT) functionality. Finally, we focus on and end-to-end vision by in June 2018 [82]. Other documents of interest also include the pro-
extending SON principles to the core, and we discuss the role of vir- tocol neutral SON policy Network Resource Model (NRM) Integration
tualized and software defined architectures in the context of 5G Net- Reference Point (IRP), with the Information Service (IS) [37,47] and
work Management. Notice that here we do not focus on the academic Solution Sets (SS) [48,49]. Table 1 summarizes the evolution of SON in
literature, as it has already been reviewed in other interesting works 3GPP.
[7,26]. Differently from previous work, we focus on the taxonomy de-
fined by 3GPP, on the related roadmap, and we pay attention to the 2.2. SON implementations
market penetration.
In this section, we discuss about the different architectures that have
2.1. SON evolution in 3GPP been considered for SON, ranging from Centralized SON (C-SON)
functionalities, where the self-organizing algorithms reside in the net-
3GPP Release 8 started defining LTE and already sets the basis for work management system, in the OAM or in the Network Management
concepts and requirements, and for SON functionalities regarding self- Systems (NMS), to a Distributed SON (D-SON) solution, where the SON
configuration, initial equipment installation and integration. The functions are distributed, in the control plane, across the edges of the
Automatic Neighbour Relation (ANR) functionality is introduced here network, typically in the Enhanced Node Base stations (eNBs). On the
to reduce manual work when configuring the neighbouring list in newly one hand, C-SON can take into consideration data from all nodes in the
deployed eNBs. Concepts of self-optimization are defined in the context network to identify and address network-wide issues. It can potentially
of Release 9. It includes optimization of coverage, capacity, handover orchestrate the behavior of radio network equipments across an entire
and interference. The functions which are introduced (and that will be network of multi-vendor and multi-technology environments. C-SON
detailed in the following sections) are Mobility Load Balancing (MLB), solutions can be more robust against network instabilities caused by the
Mobility Robustness/Handover Optimisation (MRO), Inter-Cell concurrent operation of SON functions with conflicting objectives be-
Interference Coordination (ICIC) and Random Access Channel (RACH) cause control is centralized. However, centralized systems may respond
optimization. Release 10 focuses on improvements to already defined too slowly in the emerging world of small cells that experience very
SON functions to enhance interoperability between small cells and transitory traffic loads. The SON implementations are depicted in Fig. 2.
macro-cells and includes NGMNs recommendations, i.e., new func- Market implementations of C-SON are offered by vendors like
tionalities such as Coverage and Capacity Optimization (CCO), en- Celcite (acquired by AMDOCS), Ingenia Telecom and Intucell (acquired
hanced ICIC, and it defines all the concepts related to self-healing, so by Cisco), while D-SON solutions have traditionally been more chal-
Cell Outage Detection (COD) and Cell Outage Compensation (COC) lenging to implement and vendor specific, not allowing for easy
250
Table 1
Evolution of SON in 3GPP.
Release WI Feature TS or TR
Rel.8 SA5-SON concepts and requirements SON concepts and requirements [29]
Rel.8 SA5-Self establishment of eNBs Self configuration [30–35]
Rel.8 SA5-SON Automatic Neighbour Relation (ANR) list management ANR, PCI [36–39]
Rel.9 SA5: Study of SON related OAM Interfaces for HeNBs SON related OAM Interfaces for HeNBs [40]
Rel.9 SA5: Study of self-healing of SON Self-healing management [41]
Rel.9 SA5:SON OAM aspects: Automatic radio network configuration data preparation Automatic radio network configuration data preparation [30–32]
Rel.9 SA5:SON OAM aspects self-organization management Self-optimization (MRO, MLB, ICIC) [42]
Rel.9 RAN3: Self-organizing networks CCO, MRO, MLB, RACH opt. [43–46]
Rel.10 SA5: SON self-optimization management continuation Self-coordination, self-optimization (MRO, MLB, ICIC, [37,42,47–49]
RACH opt.)
Rel.10 SA5: Self-healing management CCO, COC [50]
Rel.10 SA5: OAM aspects of ES in radio networks ES [37–39,42,51–53]
Rel.10 RAN2-3: LTE SON enhancements CCO, ES, MLB, MRO enhancements [44–46,54]
Rel.11 SA5: ULTRAN SON management SON management [29,47,48,55–58]
Rel.11 SA5: LTE SON coordination management SON coordination [59] [29,37,42,47,48,57]
Rel.11 SA5: Inter RAT ES management OAM aspects of ES management [47,48,53,55,58,60]
Rel.11 RAN3: Further SON enhancements MRO, MDT enhancements [43–46,54,61–64]
Rel.12 SA5: Enhanced NM centralized CCO Enhanced NM centralized CCO [42,65–70]
Rel.12 SA5: Multi-vendor plug and play eNB connection to the network Multi-vendor plug and play eNB connection to the [30,71,72]
network
Rel.12 SA5: Enhancements on OAM aspects of distributed MLB OAM aspects of distributed MLB [73]
Rel.12 SA5: Energy efficiency related performance measurements Energy efficiency related performance measurements [42]
Rel.12 SA5: Het-Nets management/OAM aspects of network sharing Het-Nets/network sharing [74,75]
Rel.12 RAN2-3: Next generation SON for ULTRAN/EUTRAN SON per UE type, active antennas, small cells [76]
Rel.12 RAN2-3: ES enhancements for EUTRAN ES [77]
Rel.13 RAN2-3: Enhanced Network Management centralized CCO CCO [67]
Rel.13 SA5: Study on Enhancements of OAM aspects of Distributed Mobility Load MLB [78]
Balancing SON function
Rel.14 RAN: OAM (SON for Active Antenna Systems (AAS)-based deployments) Energy efficiency [79,80]
Rel.15 RAN: OAM (Study on OAM aspects of SON for AAS-based deployments) Coverage configuration [81]
interaction of products from different vendors, so that a supervisory 2.3. Self configuration
layer is commonly still needed to coordinate the different instances of
D-SON across a much broader scope and scale. Only recently, vendors Self-configuration is the process of bringing a new network element
like Qualcomm or Airhop have started proposing D-SON as a SON into service with minimal human operator intervention [30]. This
mainstream, as small cells and Het-Net require the millisecond response covers the cellular system life cycle phase related to planning and de-
times of D-SON. ployment. Self-configuring algorithms take care of all configuration
aspects of the eNB. When the eNB is powered on, it detects the transport
Fig. 2. SON implementations.
251
link and establishes a connection with the core network elements. After supported by signalling among cells over X2 interface. To support
this, the eNB is ready to establish OAM, S1 and X2 links and finally sets proactive coordination among cells the High Interference Indicator
itself in operational mode. After the eNB is configured, it performs a (HII) and the Relative Narrowband Transmit Power (RNTP) in-
self-test to deliver a status report to the network management node. dicators have been defined, while to support reactive coordination,
Since Release 8 ANR and Automated Configuration of Physical Cell the Overload Indicator (OI) has been introduced [46].
Identity (PCI) use cases have been considered [83,84]. The ANR func- 4. RACH. RACH optimization aims at optimizing the random access
tion resides in the eNB and manages the conceptual Neighbour Relation channels in the cells based on UE feedback and knowledge of its
Table (NRT). Located within ANR, the Neighbour Detection Function neighbouring eNBs RACH configuration. RACH optimization can be
finds new neighbours and adds them to the NRT. ANR also contains the done by adjusting the Power control (Pc) parameter or change the
Neighbour Removal Function which removes outdated NRs. The preamble format to reach the set target access delay [90].
Neighbour Detection Function and the Neighbour Removal Function
are implementation specific [85]. The PCI is a physical layer signature In Release 10, 3GPP defined new use cases.
to distinguish signals from different eNBs. It is based on synchroniza-
tion signals. The total number of PCIs is LTE is 504, so that reuse is 1. Coverage and Capacity Optimization (CCO) is a SON function that
inevitable, especially in dense deployments. The Automatic PCI as- aims to design self-optimizing algorithms that achieve optimal
signment aims at an automatic conflict and confusion free identification trade-offs between coverage and capacity. Different mechanisms can
of cells [86]. Recommended practices for both use cases can be found in be considered to dynamically improve coverage and capacity, such
[87]. as ICIC, scheduling, and the combination of such mechanisms. The
targets that can be optimized may be vendor dependent and include
2.4. Self optimization coverage, cell throughout, edge cell throughput, or a weighted
combination of the above.
Self-optimization embraces all the set of mechanisms which opti- 2. ES aims at providing the quality of experience to end users with
mize the network parameters during operation, based on measurements minimal impact on the environment. The objective is to optimize the
received from the network. In the following we provide a brief over- energy consumption, by designing Network Elements (NEs) with
view of the main self-optimization function that have been introduced lower power consumption and temporarily shutting down unused
across the different recent releases [86]. From Release 9, we highlight capacity or nodes when not needed [51]. In particular, many works
work on: in literature have been focusing on switching ON/OFF eNBs or small
cells, in an efficient way, in order to guarantee a target level of
1. MLB. The MLB is the SON function in charge of managing cells’ Quality of service/experience, while minimizing the dissipated en-
congestion through load transfer to other cells. The main objective is ergy.
to improve the end-user experience and achieve higher system ca-
pacity by distributing user traffic across the system radio resources. Release 11 provides enhancements to MLB optimization, HO opti-
The implementation of this function is generally distributed and mization, CCO, and ES. Release 12 has focused on a study on en-
supported by the load estimation and resource status exchange hancements of OAM aspects for distributed MLB [78].
procedure. The messages containing useful information for this SON
function (resource status request, response, failure and update) are 2.5. Self-healing
transmitted over the X2 interface [46]. MLB can be implemented by
tuning the Cell Individual Offset (CIO) parameter. The CIO contains Self-healing [50] focuses on the maintenance phase of a cellular
the offsets of the serving and the neighbour cells that all UEs in this network. Wireless cellular systems are prone to faults and failures, and
cell must apply in order to satisfy the A3 handover condition [88]. the most critical domain for fault management is the RAN. Every eNB is
2. MRO. The MRO is a SON function designed to guarantee proper responsible for serving an area, with little or none redundancy. If a NE
mobility, i.e. proper handover in connected mode and cell re-se- is not able to fulfill its responsibilities, it results in a period of de-
lection in idle mode. Among the specific goals of this function we gradation of performances, during which users are not receiving a
have the minimization of call drops, the reduction of Radio Link proper service. This results in severe revenue loss for the operator. Self-
Failures (RLFs), the minimization of unnecessary handovers, ping- healing was initially studied in Release 9 [41], but it is in Release 10,
pongs, due to poor handover parameters settings, the minimization when the main work has been carried out and features for detection,
of idle problems. Its implementation is commonly distributed. The and adjustment of parameters have been specified [91]. These specifi-
messages containing useful information are: the S1AP handover cations have been further updated in Release 11 [50]. The main defined
request or X2AP handover request, the handover report, the RLF use cases are the following.
indication/report. Release 11 focused on different improvements of
the handover optimization [89]. MRO operates over connected 1. Self-recovery of NE Software. If the NE software failed due to load
mode and idle mode parameters. In connected mode, it tunes earlier software version and/or configuration, the most important
meaningful handover trigger parameters, such as the event A3 offset thing is to ensure that the NE runs normally by removing the fault
(when referring to intra-RAT, intra-carrier hand-overs), the Time to software, and restoring the configuration.
Trigger (TTT), or the Layer 1 and Layer 3 filter coefficients. In idle 2. Self Healing of board Faults. This use case aims to solve hardware
mode, it tunes the offset values, such as the Qoff-set for the intra- failures in the NE [92].
RAT, intra-carrier case. 3. Cell Outage Management. This use case is split in two main func-
3. Inter-Cell Interference Coordination (ICIC). ICIC aims to minimize tions: 1) Cell Outage Detection. The main objective here is to detect
interference among cells using the same spectrum. It involves the a cell outage through the monitor performance indicators, which are
coordination of physical resources between neighbouring cells to compared against thresholds and profiles, and 2) Cell Outage
reduce interference from one cell to another. ICIC can be done in Compensation. This use case aims at alleviating the outage caused
both uplink and downlink for the data channels Physical Downlink by the loss of a cell from service [50]. It refers to the automatic
Shared Channel (PDSCH), and Physical Uplink Shared Channel mitigation of the degradation effect of the outage by appropriately
(PUSCH), or uplink control channel Physical Downlink Control adjusting suitable radio parameters, such as the pilot power and the
Channel (PDCCH). ICIC can be static, semi-static or dynamic. antenna parameters of the surrounding cells.
Dynamic ICIC relies on frequent adjustments of parameters,
252
2.6. Self coordination An example of this can be observed in Fig. 3, where we provide an
analysis of how the iterations among several SON functions im-
SON functionalities are often designed as stand-alone functional- plemented in centralized and distributed manner can generate con-
ities, by means of control loops. When they are executed concurrently flicts in the network. In particular, this figure focuses on the SON
in the same or different network elements, the impact of their inter- output parameter conflict, i.e., when two or more SON functions aim
actions is not easy to be predicted, and unwanted effects may even at optimizing the same output parameter with different actions re-
occur among instances of the same SON function, when implemented in quest, and where at least three possible conflicts can arise: 1) the re-
neighbouring cells. The risk of unacceptable oscillations of configura- source conflict between MRO and MLB, which is generated by the
tion parameters or undesirable performance results increase with the concurrent execution of actions 3 and 4; 2) the one among CCO and
number of SON functions. ICIC, which is generated by the concurrent execution of actions 1 and
As we stated earlier, 3GPP has proposed different architectures for 6, and/or 3) the one among COC and ICIC, which is generated by the
SON implementation, ranging from centralized C-SON to distributed D- concurrent execution of actions 2, 5 and 7. We can identify output
SON. The choice of the architecture has a strong impact on the effi- parameters, which are affected by two opposite decisions of two dif-
ciency of the self-coordination framework. If C-SON is used, SON ferent functions, trying to achieve their own targets. As a result, to
functions are implemented in the OAM or in the NMS, as part of the define and implement a self-coordination framework is considered a
Operation and Support System (OSS). This implementation benefits necessity [2,93,94].
from global information about metrics and Key Performance Indicator
(KPI)s, as well as computational capacity to run powerful optimization
2.7. Minimization of drive tests
algorithms involving multiple variables or cells. However, it suffers
from long time scales. In order to avoid oscillations of decision para-
MDT enables operators to collect User Equipments (UEs) measure-
meters, 3GPP requires [59] that each SON function asks for permission
ments together with location information, if available, with the purpose
before changing any configuration parameter. This means that a request
of optimizing network management while reducing operational effects
must be sent from the SON function to the SON coordinator and a re-
and maintenance costs. This feature has been studied by 3GPP since
sponse has to be returned. In C-SON all these requests must pass
Release 9 [95], among the targets there are the standardization of so-
through the Interface-N, which is not suitable for real-time commu-
lutions for coverage optimization, mobility, capacity optimization,
nication, so that there is no possibility to give priority to SON co-
parametrization of common channels, and QoS verification [92]. Since
ordination messages over other OAM messages. If in turn, distributed
operators are also interested in estimating QoS performance, in Release
coordination is used, the interaction between the SON function and the
11, MDT functionality has been enhanced through QoS performance to
local SON coordinator will be over internal vendor-specific interfaces,
properly dimension and plan the network by collecting measurements
with much lower latency characteristics. This makes the D-SON archi-
indicating throughput and connectivity issues [96]. These MDT func-
tecture much more flexible and adequate for small cell networks, which
tions have been further elaborated in Release 11, while Release 12 has
experience very transitory traffic loads, thus requiring high reactivity to
included specific enhancements in terms of correlation of information,
propagation and traffic conditions.
which can be found in the study on enhanced network management
Fig. 3. High-level example of how the iterations of multiple SON functions may interfere.
253
centralized CCO. These improvements and extensions of SON en- of scale of the IT industry, by moving traditional network functions
hancements introduced until Release 13 can be found in [29]. away from specialized hardware to general purpose computation, sto-
rage and memory pools, distributed throughout the network and in data
2.8. Core networks centers. NFV virtualizes the functional elements of the network, in-
stantiating the corresponding functions as programs that run on COTS,
The core network operations can be managed through self-orga- and less expensive hardware. This concept, combined with a SDN ar-
nizing functionalities. The benefits also in this case come from the re- chitecture, in charge of exposing the network resources, through
duced human intervention and from reduced operational costs. self- Application Programming Interfaces (APIs), to external applications, is
organization in the core network allows to self-adapt traffic loads and introduced to make mobile network deployments more cost-effective
prevent bottlenecks. In addition, self-organization for Core enables the [27,98].
core network to handle signalling more efficiently. In this regard, Nokia On the one hand, SDN is promoted by the Open Networking
[97] already automates core networks operations based on SON tech- Foundation (ONF) and defines an architecture of three layers, the Data
nology. The objective is to automatically and rapidly allocate core layer, the control layer and the application layer. The data layer ex-
network resources to meet unpredictable behaviours and demands in poses the hardware resources towards the control layer, which manages
terms of broadband. Notice that SON use cases for core networks are the network devices according to policies or high level applications.
not limited to LTE networks, but many of them can be taken into ac- The application layer allows the development of high level applications
count also for other kinds of networks, like 2/3G. to customize and program the network behaviour using well defined
interfaces. Communication between the layers is handled by the
Northbound and Southbound interfaces, as shown in Fig. 4(a).
2.9. SON in virtualized and softwarized 5G architecture On the other hand, the virtualization refers to the abstraction of
logical from physical resources. In this context, NFV proposes to
The wireless industry is currently working towards being prepared transfer network functions, such as routing, firewall, gateway, deep
for a 1000x data traffic growth. It is unlikely, though, that users will packet inspection, etc., towards virtual software-based applications,
want to pay more for the service than they are paying today, which set a executed on Information Technology (IT) platforms (servers, switches,
serious challenge for both mobile operators and vendors, i.e. how to storage). This allows improved flexibility and scalability, which facil-
improve the infrastructure 1000 times, without increasing the CAPEX itates the implementation of new network functions, while decreasing
and OPEX. Besides SON, another trend in this direction, initiated by an costs. For this, the network functions have to migrate from being im-
European Telecommunications Standards Institute (ETSI) industrial plemented in specialized hardware to run on top of COTS hardware.
study group in 2012, is the NFV, which allows to exploit the economies
Fig. 4. SDN and NFV architectures.
254
The ETSI has proposed a NFV architecture based on three main mod- implementations proposed for SON, by proposing a SON over NFV ar-
ules, the NFV Infrastructure (NFVI), NFV and the NFV Management and chitecture, where SON functions, aimed at tackling the main radio ac-
Orchestration, as it is shown in 4. NFVI represents the hardware and cess and backhauling challenges of extremely dense deployments, are
software resources where NFV is applied. The virtualization layer al- virtualized and run over generic purpose hardware. These virtualized
lows the abstraction of the different resources. The Virtual Network SON functions can be deployed in decentralized manner (D-SON), close
Function (VNF) represents the virtual instance that runs over the NFVI. to the edge of the network, following the MEC paradigm, or in cen-
Finally, the orchestration and management of VNF and NFVI ensures tralized nodes (C-SON), in order to fulfill tradeoffs between latency and
the optimal operation of VNFs over the available infrastructure. The optimal operations, and while responding online to the specific situa-
orchestrator, the VNF manager and the Virtualized Infrastructure tion of the network in terms of load, perceived QoE/QoS, etc. Moreover,
Manager, in charge of the resource inventory to check on resource SON operating over a SDN/NFV architecture would be compatible also
availability, ensures the proper intelligence of the network. This layer is with non-3GPP systems. The market offers start-ups which already
the brain with the broadest view of the network, it coordinates the advertise the concept of virtual-SON in the cloud. Cellwize [13] is one
allocation of functions across the different segments of the dense, het- of them.
erogeneous network. When the RAN is provided with IT and cloud However, this research line is extremely novel. A joint standardized
computing capabilities by deploying COTS at the access side, we talk architecture including fundamental concepts like NFV and MEC, SDN,
about MEC. MEC aims to place the service platform at the edge of the and SON is still under discussion and cannot be easily defined here. We
network, to provide nearby service computing and data caching for want to highlight the work that is under development in the context of
mobile users. It utilizes virtualization and cloud computing technolo- the SELFNET project [5], who has designed an architecture to imple-
gies to run and manage applications at the edge of the mobile network, ment this SON over NFV/SDN view, by integrating novel SDN, NFV,
to provide low latency and high bandwidth experience. Both MEC and SON, Cloud computing artificial intelligence technologies [99]. In
NFV architectures require a network infrastructure virtualization plat- Fig. 5, we have represented a SON over SDN/NFV architecture, inspired
form, and an orchestrator making decisions to meet the requirements of by the work carried out by the SELFNET project [5]. The self-organizing
the operational system and service applications. At the methodological capabilities are provided by two layers, the SON control layer and the
level, the orchestrator can take advantage of the huge amount of in- SON autonomic layer. The SON control layer contains the applications
formation travelling through the network, in terms of measurements, that will enable the collection of data from sensors deployed throughout
signaling information, QoS and QoE indicators, etc., by means of ma- the network, as well as the applications that will enforce decisions to-
chine learning based approaches. wards elements of the network. The SON autonomic layer offers the
Even if SDN and NFV have already a value when implemented in- mechanisms to provide network intelligence. It performs monitoring,
dependently, they have an added value when they are combined. The collects information about the network behaviour, diagnose problems
NFV technology allows to easily create and program new network and makes decisions about what functions to activate, where to locate
functions and to place them in the most suitable location in the net- them, if at the edge or not, etc. As a result, automation is implemented
work, coping with capacity, coverage and quality issues, allowing at two levels. First inside the specific instantiated VNF, and then in the
seamless integration of new network elements, all while using the most definition of the network itself, about where and how to locate the
appropriate amount of resources. This requires though the SDN to be specifics VNFs.
able to adjust the network accordingly, enabling network program- In the market there already exist start-ups which advertise the
mability. concept of virtual-SON in the cloud. Cellwize [13] is one of them. They
In this context, SON can be useful to achieve real time autonomous are promising a technology with deployment in the cloud, capable of
network management. In this novel softwarized visions, we can benefit working seamlessly across different vendors, spectrum and technolo-
of all the opportunities offered by centralized, distributed and local gies.
Fig. 5. SON over NFV-SDN architecture.
255
3. How to address SON and NM through ML amount of data. For an appropriate processing and to extract useful
information, it is convenient to eliminate the noise present in the
In this section we classify at high level the different network man- data base, by reducing the dimensionality of data. Solutions to this
agement classes of problems that one may need to deal with when problem are to be found in the UL literature, and specifically among
aiming at managing the network in a self-organized manner. For each the dimensionality reduction solutions.
class of problem, we identify the machine learning tools that can be • Pattern identification, grouping: The tasks belonging to this class aim
used. The objective of ML is to improve performance of a particular sets at identifying patterns, group of nodes with similar characteristics,
of tasks by creating a model that helps find patterns through learning according to some kind of criteria. An objective may be to apply to
algorithms. ML taxonomy is traditionally organized onto: 1. Supervised them similar optimization approaches. Self-configuration use cases
Learning (SL), 2. Unsupervised Learning (UL), 3. Reinforcement are intuitive application for these issues. Solutions to these problems
Learning (RL), 4. New trends: Deep Learning (see Table 2). Recently, can be translated into learning the set of classes the data belongs to.
new trends in the area of ML are taking momentum, thanks to the UL literature offers solutions in the area of clustering.
progress of software engineering, computational capabilities and • Sequential decision problems for online parameter adjustment: This class
memory availability. Deep learning has been proven feasible and ex- of problems is extremely common in the area of autonomous man-
tremely effective in different applications, like language, video, speech agement, where we face control decision problems to online adjust
recognition, object and audio detection, among others. The most ex- network parameters, with the objective to meet certain performance
emplary one is the win of AlphaGo, beating the world champion at the targets. This kind of decision problems, where we learn the most
Chinese board game Go. The victory of AlphaGo was due to the im- appropriate decision online, based on the reaction of the environ-
plementation of a deep reinforcement learning algorithm capable of ment to the actions the network is taking, can be addressed through
self-learning. RL solutions. All self-optimization use cases can be addressed
Keeping in mind the SON and NM functions introduced in the pre- through these solutions, as well as COC problems.
vious section, the classes of problems that need to be addressed when
managing the network autonomously are: In the rest of the section, we relate each class of NM problem to the
possible ML literature to solve it. The review of ML literature provided
• Variable estimation or classification: The tasks belonging to this class in the following, is far from being exhaustive. Many methods and
of problem aim at e.g. estimating the QoS or the QoE of the network, techniques will not be described, because the purpose is here to provide
at predicting performances or behaviours of the network, by a useful taxonomy to address NM and SON problems and to analyze and
learning from the analysis of data obtained from past behaviours of understand the related literature using ML solutions. For a deeper un-
the network. NM and SON functions where these tasks are useful are derstanding of ML solutions, the reader is referred to more specific
QoS estimation and other MDT use cases, the prediction of beha- literature.
viours to optimize network parameters, etc. Solutions to these pro-
blems can be translated into finding the relationship between one 3.1. Supervised learning (SL)
variable and some others, or Identifying which class of a set of pre-
defined classes the data belongs to. Solutions are then to be found in This ML technique could be extremely useful when the NM function
the SL literature, with both regression and classification tasks. to address requires estimation, prediction, classification of variables. SL
• Diagnosis of network faults or misbehaviours: The tasks belonging to is a ML technique which takes training data (organized into an input
this class of problems aim at detecting issues ongoing in the net- vector (x ) and a desired output value (y)) to develop a predictive
work, which may be associated to faults and anomalous setting of model, by inferring a function f (x), returning the predicted output y .̂
network parameters. This kind of problems relates to self-healing For that, the construction of a dataset is needed. The dataset contains
issues and solutions can be found in UL literature, and in particular training samples (rows), and features (columns), and is usually divided
in the anomaly detection solution. into 2 sets. The training set, used to train the model, and the test set,
• Dimensionality reduction: The network generates continuously a huge used to make sure that the predictions are correct. The goal of the
training model is to minimize the error between the predictions and the
Table 2 actual values. Hence, by applying ML, we aim to estimate how well a
Summary of selected ML techniques. learning algorithm generalizes beyond the samples in the training set.
The input space is represented by a n-dimensional input vector
Supervised learning
x = (x (1), …, x (n) )T ∈ Rn . Each dimension is an input variable. In addition
Classification K-Nearest Neighbours a training set involves m training samples ((x1, y1), …, (x m, ym )) . Each
Generalized Linear Model sample consists of an input vector x i, and a corresponding output yi.
Support Vector Machines
Hence x i(j) is the value of the input variable x(j) in training sample i, and
Naive Bayes
Neural Networks the error is usually computed via  yi − yi . The SL technique has two
Regression K-Nearest Neighbours main applications, classification and regression. On the one hand,
Generalized Linear Regression classification is applied when y, the output value we try to predict is
Support Vector Regression discrete, e.g., we want to predict if a cancer is benign or malign, based
Neural Networks
Decision Trees
on a dataset constructed based on medical records, and collecting many
Unsupervised learning features, e.g. tumour size, age, uniformity of cell size, uniformity of cell
Clustering Non-overlapping clustering shape. On the other hand, a regression problem is applied when y is a
Hierarchical clustering real number.
Overlapping clustering
A huge amount of SL algorithms for classification can be found in
Dimensionality Reduction Feature Extraction
Feature Selection the literature, and a study to evaluate the performance of some of them
Anomaly Detection Pruning techniques can be found in [100]. In the following we briefly introduce the most
Rule-based systems common algorithms.
Reinforcement learning
Model-based Dynamic Programming
1. k-Nearest Neighbors (k-NN) can be used for classification and re-
Monte Carlo
Model-free Temporal Difference gression. k-NN is a non-linear method where the input consists of the
256
k closest training samples in the input space. The predicted output is leaves in the tree [108]. Also, they do not require any prior
the average of the values of its k nearest neighbours. A commonly knowledge of the data, are robust (i.e., do not suffer the curse of
used distance metric for continuous variables is the Euclidean dis- dimensionality as they focus on the salient attributes) and work well
tance. The k-NN method has the advantage of being easy to inter- on noisy data. However, DTs are dependent on the coverage of the
pret, fast in training, and the amount of parameter tuning is training data as with many classifiers. Moreover, they are also sus-
minimal. However, the accuracy of the prediction is generally lim- ceptible to over-fitting.
ited.
2. Generalized Linear Models (GLM). The linear model describes a Methodologies have also been proposed to take the best out of the
linear relationship between the output and one or more input available data, to boost the prediction performance. Some of these
variables, and where the approximation function maps from xi to yi methodologies are classified among the so called Ensemble methods.
as follows, Ensemble methods combine the predictions of multiple learning algo-
rithms to produce a final prediction. This technique has been in-
yi = θ0 + θ1 x i(1) + …+θn x i(n)
 (1) vestigated in a huge variety of works [109,110]. A general method is
where θi are the unknown parameters. The idea is to choose θi so sub-sampling the training examples, where the most useful techniques
that yi minimizes the loss function. Typically, we make the as- are referred to as bagging and boosting [111]. Bagging manipulates the
sumption that the samples in each dataset are independent from training examples to generate multiple hypothesis. It runs the learning
each other, and that the training set and testing set are identically algorithm several times, each one with different subset of training
distributed. Note that if the relation is not linear, the model should samples. On the other hand, AdaBoost maintains a set of weights over
be generalized, in an attempt to capture this relationship [101]. the original training set, and adjusts these weights by increasing the
3. Naive Bayesian. The method is used for classification and is based on weight of samples that are misclassified, and decrease the weight of
Bayes theorem, i.e., calculating probabilities based on the prior samples that are correctly classified [112,113].
probability. The main task is to classify new data points as they
arrive. A NB classifier assumes that all attributes are conditionally 3.2. Unsupervised learning (UL)
independent, and is recommended when the dimensionality of the
input is high [102]. Since NB assumes independent variables, it only This kind of learning can be extremely useful when the NM function
requires a small amount of training data to estimate the means and requires identifying anomalous behaviours, recognizing patterns or
variances of the variables. reducing the dimensionality of the data. UL is a ML technique, which
4. Support Vector Machines (SVMs) can be used for classification and receives unlabelled input patterns with the objective to find a pattern in
regression. SVMs are inspired by statistical learning theory, which is it. In this case, we let the computer learn by itself, without providing
a powerful tool for estimating multidimensional functions the correct answer to the problem we want to solve. The goal is to
[103,104]. This method can be formulated as a mathematical opti- construct representation of inputs that can be used for predicting future
mization problem, which can be solved by known techniques. For inputs without giving the algorithm the right answer, as in turn we do
this problem, given m training samples ((x1, y1), …, (x m, ym )), the in case of supervised learning [114]. The three most important families
goal is to learn the parameters of a function which best fit the data. of algorithms are clustering, dimensionality reduction and anomaly
It samples hyperplanes. Thus, the hyperplane with the main detection techniques. There are many examples of UL applications in
minimum distance from the sample points is maintained. The our daily life, e.g., news.google.com, understanding genomics, organize
sample points that form margin are called support vectors and es- computer clusters, social network analysis, astronomical data analysis,
tablish the final model. This method in general shows high accuracy market segmentation, etc. In the context of SON, UL algorithms are
in the prediction, and it can also behave very well with non-linear applied mainly on self-optimization and self-healing use cases.
problems when using appropriate kernel methods. Also, when we
cannot find a good linear separator, kernel techniques are used to 1. Clustering. This technique aims at identifying groups of data to build
project data points into a higher dimensional space where they can representation of the input. The most common methods to create
become linearly separable. Hence the correct choice of kernel clusters by grouping the data are: non-overlapping, hierarchical and
parameters is crucial for obtaining good results. In practice, this overlapping clustering methods. K-means [115] and Self-organizing
means that an exhaustive search must be conducted on the para- Maps (SOMs) [116] methods belong to non-overlapping clustering
meter space, thus complicating the task [105]. techniques. When the clusters at one level are joined as clusters at
5. Artificial Neural Network (ANN) is a statistical learning model in- the next level (cluster-tree), this is referred in literature as a hier-
spired by the structure of a human brain, where the interconnected archical clustering method [117]. In case that an observation can
nodes represent the neurons producing appropriate responses. ANN exist in more than one cluster simultaneously, this is known as
supports both classification and regression algorithms. The basic overlapping or fuzzy clustering. Fuzzy C-means and Gaussian mix-
idea is to efficiently train and validate a neural network. Then, the ture models belong to this kind of technique [115,118]. Also Hidden
trained network is used to make a prediction on the test set. In this Markov Model (HMM) can be used for clustering. This kind of al-
method the weights are the parameters in charge of manipulating gorithms have been proposed in a wide range of fields, such as,
the data in the calculations. Here, the interconnection pattern be- robotics, wireless systems, and routing algorithms for mobile ad-hoc
tween the different layers of neurons, the learning process for up- networks, among others.
dating the weights of the interconnections, and the activation 2. Dimensionality Reduction. High-dimensional datasets present many
function that converts a neuron’s weighted input to its output acti- challenges. One of the problems is that, in many cases, not all the
vation are the most important parameters to be trained [106]. ANNs measured variables are necessary to understand the problem of in-
methods require parameters or distribution models derived from the terest. In the state of the art we can find a huge amount of algo-
data set, and in general they are also susceptible to over-fitting. rithms to predict models with good performance from high-dimen-
6. Decision Trees (DT) is a flow-chart model in which each internal sional data. However it is of interest for many problems to reduce
node represents a test on an attribute. Each leaf node represents a the dimension of the original data. For example, in [119,120], the
response, and the branch represents the outcome of the test [107]. authors face the problem of the huge amount of potential features
DTs can be used for classification and regression, and they have the system has as input, and they suggest that the regression analysis
nuisance parameters, such as the desired depth and number of has a better performance in a reduced space. In this context, the
most common methods are: Feature Extraction (FE) and Feature
257
Selection (FS) [121]. Both methods seek to reduce the number of estimating a so called value function, which is a function of the states
features in the dataset. FE methods do so by creating new combi- estimating how good it is for an agent to be in a given state. For MDPs
nations of features (e.g. Principal component analysis (PCA)), which the state-value function, denoted as Vπ(s), is the expected return when
project the data onto a lower dimensional subspace by identifying starting in state s and following policy π thereafter. For more in-
correlated features in the data distribution. They retain the Principal formation the reader is refereed to [130].
Components (PCs) with the greatest variance and discard all others RL literature offers two approaches to solve MDPs. These two ap-
to preserve maximum information and retain minimal redundancy proaches are: model-based and model-free.
[122]. Correlation based FS methods include and exclude features
present in the data without changing them. For example, Sparse 1. Model-based. Dynamic Programming (DP) and Monte Carlo (MC)
Principal Component Analysis (SPCA) extends the classic method of methods fall into the category of model-based approach.
PCA for the reduction of dimensionality of data by adding sparsity (a) DP is able to solve MDPs by relying on the knowledge of the
constraint on the input features. state transition probability between two states after executing a
3. Anomaly Detection. Anomaly detection identifies events that do not certain action. DP is an algorithmic paradigm that solves a given
correspond to an expected pattern. By modeling the most common complex problem by breaking it into sub-problems and stores
behaviors, the machine selects the set of unusual events [123]. Self the results of sub-problems to avoid computing the same results
healing is one of the main functionality in which this kind of tech- again. DP algorithms are based on update rules derived from the
niques are applied, some examples are [124,125]. The two most Bellman equation. The first key component is known as the
common techniques are: Rule based systems, which are very similar policy evaluation process, according to which a policy provides
to DTs, but they are more flexible than DTs as new rules may be information about how much reward is going to be received in
added, without creating a conflict with the existing ones [123] and the MDP. This solution is used to build the first overall solution
Pruning techniques, which aim at identifying outliers, where there by finding the optimal policy known as the policy iteration pro-
are errors in any combination of variables. cess. Finally the value iteration makes the value function better
and better by applying Bellman’s equation intuitively. DP is used
3.3. Reinforcement learning (RL) to solve problems such as, scheduling, graph algorithms, bioin-
formatics, among others.
The ML approaches under this category can be used to address NM (b) MC method only requires experience, i.e., sample sequences of
functions which require network parameter control. Differently from states, actions and rewards. The estimations are only updated
the case of SL, RL aims to learn from interactions how to achieve a after the episodes conclude. Although their application on
certain goal. In many real applications and in particular, in sequential practical cases is limited, they provide foundation for other RL
decision and control problems, it is not possible to provide an explicit methods.
supervision to the training (i.e. the right answer to the problem). In 2. Model-free. Temporal Difference (TD) methods are model free ap-
these cases, we can only provide a reward/cost function, which in- proaches to solve RL problems. TD learning is a combination of MC
dicates to the algorithm when it is doing well and when it is doing and DP ideas. It uses the current estimate Vtπ of the value function
poorly. RL has already been proven effective in many real world ap- instead of the exact Vπ, as it happens in DP. If is known, we can
plications, such as autonomous helicopters, network routing, robot solve the MDP through DP, otherwise we need to rely on TD
legged automation, etc.[126–128]. methods.
The learner or decision maker is called agent, and it interacts con- Some common examples of TD methods are: Q-learning, Sarsa and
tinuously with the so-called environment. The agent selects actions and Actor Critic (AC) [130]. TD methods can be found in each SON
the environment responds to those actions and evolves into new situa- functionality.
tions. In particular, the environment responds to the actions through (a) Q-learning and Sarsa are based on the estimation of the state-
rewards, i.e., numerical values that the agent tries to maximize over time. action value function, Q(s, a). Learning is performed by itera-
The agent has to exploit what it already knows in order to obtain a tively updating the Q-values, which represent the expert
positive reward, but it also has to explore in order to take better actions knowledge of the agent, and have to be stored in a representa-
in the future. Learning can be centralized in a single agent or dis- tion mechanism. The most intuitive and common representation
tributed across a multiple agents. In single agent systems, ML ap- mechanism is the lookup table, i.e., the TD methods represent
proaches are capable of finding optimal decision policies in dynamic their Q-values in a Q-table, whose dimension depends on the
scenarios with only one decision maker. In multi-agent systems, the size of the state and action sets. The difference between them is
distributed decisions are made by multiple intelligent decision makers, that, Q-learning is an off-policy learner. This means that, the
and the optimal solutions or equilibria are not always guaranteed agent will use the policy corresponding to the best action in the
[129]. next state, given the current agent experience, whereas Sarsa is
The problem is then defined by means of a Markov Decision Process an on-policy learner. On-policy learners evaluate the policy π, to
(MDP) { , , , , γ }, where , is the set of possible states of the en- perform the decisions. This means that, the policy followed by
vironment = {s1, s2, …, sn}, , is the set of possible actions the agent to select its behaviour in a given state is the same used
= {a1, a2 , …, aq} that each decision maker may choose, (s′ s, a), is to select the action based on which it evaluates the followed
the transition function denoting the probability of getting s′ giving an behaviour.
action a in state s, (s, a) is a reward function, which specifies the (b) AC methods have a separate memory structure to represent the
expected immediate return obtained by executing action a in state s, policy independently of the value function. The policy structure
and 0 ≤ γ ≤ 1 is a discount factor, which gives more importance to is known as the actor, since it is used to select the actions, while
immediate rewards compared to rewards obtained in the future [130]. the estimated value function is known as the critic. The critic
The MDP represents the theoretical basis for the RL framework learns and critiques whatever policy is currently being followed
[130]. At each time step, the agent implements a mapping from states by the actor and takes the form of a TD error δ, which is used to
to probabilities of selecting each possible action. This mapping is the determine if at was a good action or not. δ is a scalar signal,
agent’s policy. which is the output of the critic and drives the learning proce-
The objective of each learning process is to find an optimal policy dure. After each action selection, the critic evaluates the new
π * (s ) ∈ for each s, to maximize some cumulative measure of the state to determine whether things have gone better or worse
reward r received over time. Almost all RL algorithms are based on than expected.
258
3.4. New trends: deep learning the various bearers (both on the radio part and in the back-haul),
idle and connected mode mobility.
Deep learning is a new trend in ML that allows computer systems to 3. Minimization of Drive Tests (MDT). The data extracted from this
improve with experience and data. It achieves great power and flex- source refers to the radio measurements of both idle and connected
ibility to operate in complicated real-world environments, by learning mode mobility, coverage items, such as, power measurements and
to represent the world through a nested hierarchy of concepts. The ML radio link failure events, and can be associated with position in-
algorithms that we have reviewed in this paper have a strong de- formation of the UE performing the measurement. More information
pendency on the features on which the algorithms are applied. Based on on these data has already been provided also in Section 2.7.
that, much effort has been devoted to design ML algorithms that yield 4. E-UTRA Control plane protocols and interfaces, such as Radio Resource
to useful representations. This is known as representation learning, and Control (RRC), S1-AP, X2-AP protocols, are another huge source of
deep learning is one way of learning representations [131,132]. The information, especially concerning aspects, such as cell coverage, user
main representation of deep learning is through a Multilayer Percep- connectivity, mobility in idle and connected mode, inter-cell inter-
tron, which is a multilayer neural network function mapping some sets ference, resource management, load balancing, among others.
of input values to output values. Each layer of this representation learns 5. Data plane traffic flow statistics, also are a huge source of informa-
a hierarchy of the output. Deep Learning has the ability to do successful tion, which can be gathered at various points of the network, like the
training from the bottom layers to the higher ones. This is done by eNB, or the PDN Gateway (PGW) and Serving Gateway (SGW). The
applying computational models that are composed of multiple levels of Internet Protocol Flow Information Export (IPFIX) is an example of
representation and abstraction that help make sense of data. standardized format to exchange this kind of statistics [136].
4. Machine learning enabled network management All these data are available to the network operators, but in most cases
this is not made available to the academic community due to privacy
As we have mentioned in the introduction of this work, mobile issues and network operators’ interests. There are some exceptions, like
networks constitute a huge source of data which could be analyzed with the Data for Development (D4D) initiative from the Orange group [137],
proper tools, with the primary goal to make more informed decisions which made available anonymous data extracted from the Senegal’s
when it comes to efficiently manage the overall 4G or 5G network. In network to research laboratories. However these data are in general ag-
this context, ML is a great opportunity due to its capability of providing gregated and do not allow deep insight into the operator’s network.
insightful information from the analysis of data already available to This lack of data represents a great limitation for the advancement
operators, which can be used to make improvements or changes. of the ML based network management research. It is not easy to find
In this section we focus on how ML can specifically be applied to contributions analyzing real network management data. Some work can
SON and novel network management concepts. First, we present all the be found in the context of 3G networks, but for 4G networks it is very
relevant sources of information that could be extracted from a mobile hard to find works considering real data [138,139]. These works
network. All these data are available to operators, and may happen to though do not take into account the data analyzing them through ML
be sensitive data for the users’ privacy. However, some interesting data techniques, to extract experience from them.
can be derived from open databases or sniffed from unencrypted control However, some network data can be derived in other ways. Some
channels like the PDCCH. We will then discuss on these options. databases are available, providing a huge insight in mobile network
Second, we will go through again the main SON and network man- operators. Some examples are listed in the following, together with
agement functions and provide a classification of the main inputs and information that can be extracted from them.
outputs that we would need available in the form of data, when de-
signing an appropriate ML algorithm to target the specific use case, and • opencellid: It contains information on specific cells, such as: net-
the KPI indicators that we would need to monitor. Finally, we provide work type (GSM, UMTS, LTE), Mobile Country Code (MCC), Mobile
an overview of SON and network management’s related work, where Network Code (MNC), Location Area Code (LAC) for GSM and
ML techniques have been adopted, classifying this work as a function of UMTS, Tracking Area Code (TAC) for LTE, Cell ID (CID) for GSM and
the targeted use case, the specific high level problem to solve and the LTE networks, Primary Scrambling Code (PSC) for UMTS networks,
ML technique that the authors have picked to address the problem. PCI for LTE networks, longitude and latitude in degrees, estimates of
range in meters, total number of measurements collected from the
4.1. Data generated by mobile cellular networks tower, defines if the coordinates of the cell tower are exact or esti-
mated, information of the date when the cell tower was first added
As we observed in [23] a huge amount of data is currently already to the data base and updated, average signal strength from all
generated in mobile networks during normal operations by control and measurements received from the cell in dBm, or as defined in [140].
management functions. This kind of data can be exploited to find pat- This data base also receives funding from important vendors like
terns and extract useful information from them. This allows to take Qualcomm [141] and offers some formula of free access to portion
more informed decisions to effectively manage network performance. of data for academic purposes.
Some examples of the different sources information generated by mo- • opensignal: It offers information on achievable data rates, latency
bile networks, together with the kind of usage currently provided by and availability, per operators, but not information per cell
operators, and related references of interest, is detailed in Table 3. tower [142].
• antenasgsm: It offers information on maps and positions of cells,
1. Charging Data Records (CDR). They are defined in [133] and pro- with added information on the operator and the assigned band-
vide a comprehensive set of statistics at the service, bearer and IMS width [143].
levels. These records are typically stored for offline processing by • Google geolocations API: It allows queries based on the cell ID to
the operator. The granularity of this information in the time domain get cell related information and WiFi Access Points (APs), such as
is however quite coarse, as records are generated in correspondence latitude and longitude [144].
with high-level service events (e.g., start of a call).
2. Performance management functionality. This data source [42,134] The information provided by these databases is precious, but still
provides data regarding the network performance and it covers, does not give sufficient insight on the behaviour of the network, and
among others, aspects of the performance of the radio access net- mainly offers an overview of the coverage provided by the single op-
work, such as, radio resource control and utilization, performance of erators.
259
Table 3
Information elements relevant for ML enabled SONs.
Source Data Usage TS
Charging Data Records Includes statistics at the service, bearer and IP These records are typically stored, but only used by TS 32.298 [133]
(CDR) Multimedia System (IMS) levels. customer service. The network operation departments
typically do not leverage this information and do not
have access to it, as much as customer service does not
leverage network management data.
Performance management It covers long-term network operation functionalities, The data is currently mostly used for fault identification, TS32.401 [134],
(data on network such as Fault, Configuration, Accounting, Performance e.g., triggering alarms when some performance TS32.425 [42]
performance) and Security management (FCAPS), as well as customer indicator passes some threshold, so that an engineer can
and terminal management. An example is that defined investigate and fix the problem. Typically, the only
for OAM, which consists of aggregated statistics on automatic use of this info is threshold-based triggering,
network performance, such as number of active users, which can be done with very low computational
active bearers, successful/failed handover events, etc. complexity.
per BS, as well as information gathered by means of
active probing.
Minimization of Drive Tests Radio measurements for coverage, capacity, mobility This data is used for identified use cases such as TS37.320 [135]
(MDT) optimization, QoS optimization/verification coverage, mobility and capacity optimization, and QoS
verification
E-UTRA Control plane Control information related to regular short-term A This information is normally discarded after network TS36.331 [54],
protocols and interfaces network operation, covering functionalities such as call/ operation purposes have been fulfilled. Some data can TS36.413 [45],
session set-up, release and maintenance, security, QoS, be gathered via tracing functionality or used by SON TS36.423 [46]
idle and connected mode mobility, and radio resource algorithms which normally discards the information
control. after usage
To get more information, still we can do something more and access Management (RRM) and SON algorithms. Consequently it could be a
directly to the unencrypted PDCCH and extract information exchanged very useful tool to build realistic scenarios based on information
between the users and the associated eNB. In particular, it is possible to available in public databases, generate data to analyze, build algo-
build a sniffer, as the one described in [145], from which to collect raw rithms based on this analysis and close the loop on the simulator to test
communication traces exchanged by the users and the associated eN- the designed algorithms. The lack of data to analyze when designing
odeB. This allows to have access not only to aggregate base station future networks will always be a challenge also for vendors, in the
statistics, but also to more valuable information derived from the radio phase that the networks are still not running, as a result, relying on high
protocols, such as the resource block allocation and the link adaptation fidelity open source simulators, like those provided by ns-3 consortium,
mechanism of the system. In particular, the OWL sniffer [145] is an available to anyone and allowing collaborative development, and re-
online decoder of the LTE control channel, which uses a Software De- sults reproducibility, is a fundamental aspect of research in this line.
fined Radio (SDR) to send the raw LTE signal to a PC running the de- Finally, let us review the main SON use cases in Table 4, by ana-
coding software. This open-source software is capable of reliably log- lyzing the main input information that their design would require, in
ging the LTE downlink control information (DCI) broadcasted by base terms of data, together with the main identified output actions and
stations. In fact, LTE uses an unencrypted control channel to assign meaningful associated KPIs.
network resources to users for both downlink and uplink communica-
tions. Resources are assigned to devices through their radio network 4.2. Overview of ML based network management’s relevant literature
temporary identifiers (RNTIs), every millisecond, specifying the
number of resource blocks (RBs) and the modulation and coding This section reviews SON and Network management’s recent work
scheme (MCS) to be used. There are works in literature using this sniffer in the area of ML. We will go through each main function and use case
to collect and analyze traces from different European cities [146]. and review significant literature and ML approach that has been used to
Another option is to use a high-fidelity network simulator like ns-3 approach the problem. Table 5 summarizes the main works in this area
LTE/LTEEPC Network Simulator (LENA) module, to generate realistic and classifies them per 3GPP use case, technique and specific algorithm
data [147]. This simulator has been built around industrial Application adopted by the authors.
Programming Interface (API) defined by the small cell forum and offers
high-fidelity models from Media Access Control (MAC) to application 1. Use case: Indicates the 3GPP targeted use case.
layers. It has also been designed with the requirement to simulate tens 2. Reference: Indicates the reference of the related work.
of eNBs and hundreds of UEs, and to specifically test Radio Resource 3. Technique: Indicates the applied ML method (Supervised Learning,
Table 4
SON inputs, outputs and KPIs.
SON function Inputs Output actions KPIs
MLB X2 resource status and load estimation information. Tuning the CIO, i.e. offsets of serving and neighbour cells to Improved QoS and capacity
satisfy handover conditions.
MRO S1AP and X2AP handover requests, handover reports, A3 offsets, TTT, L1 and L3 filter coefficients, in connected Minimized call drops, RLFs and ping
RLF reports and indications. mode, and Qoffset in Idle mode. pong effects.
CCO UE measurements Transmission power, pilot power, antenna parameters, Maximized coverage and cell and edge
coordinated Almost Blank Subframes (ABS) throughput
ICIC HII, RNTP, OI, UE Measurements. Transmission power, pilot power, antenna parameters, Minimized Intercell interference.
coordinated ABS
COC UE Measurements. Transmission power, antenna parameters of neighbouring Minimized outage.
cells
ES Resource status, UE Measurements. Switch ON and OFF policies Minimized energy consumption.
260
Unsupervised Learning, Reinforcement Learning). adjacent cells. Approaches incorporating fuzzy logic with RL cap-
4. Problem: Indicates the general problem to solve. abilities have the advantage to capture the uncertainty existing in real
5. Algorithms: Indicates the specific ML algorithm applied to the data world complex scenarios, while schemes considering only learning
(see Table 5). approaches may be limited by the fixed variable definition. When
combining fuzzy logic with RL, also the subjectivity with which the
4.2.1. Mobility load balancing fuzzy variable may be defined is overcome by the adjusting capabilities
The literature offers some examples of application of ML techniques of the learning. Alternatively, a centralized solution is approached in
to the MLB use case. The majority of applications fall in the area of RL, [152], where a central server in the cellular network determines all HO
as the main problem to solve is a sequential decision problem about margins among cells by means of a dynamic programming approach.
how to set configuration parameters, which optimize network perfor- Besides RL, also clustering schemes have been proposed in this area, to
mance and user experience. An example of a RL application for MLB use group cells with similar characteristics and provide for them similar
case can be found in [149]. Here the authors present a distributed Q- configuration parameters [153]. Considering clustering in large rea-
learning approach that learns for each load state the best MLB action to listic scenarios is an added value to reduce computational complexity
take, while also minimizing the degradation in HO metrics. Another and take advantage of what is learnt in other regions of the network
option to take advantage also of fuzzy logic capabilities of dealing with where we observe similar environment characteristics.
heterogeneous sources of information is provided in [150], where fuzzy
logic is combined with Q-learning in order to target the load balancing 4.2.2. Mobility robustness optimization
problem. For similar reasons, fuzzy logic is also proposed in [151] to Also for the case of MRO, we find in literature different solutions
enhance the network performance by tuning HO parameters at the based on RL to solve a control decision problem. In [155,156], the
Table 5
Related work.
Reference ML technique Problem Algorithm
Self-configuration
PCI [148] UL Planning Clustering
Self-optimization
MLB [149] RL Control optimization Q-learning
[150] RL Control optimization Q-learning
[151] RL Control optimization Fuzzy Q-learning
[152] RL Control optimization Dynamic Programming
[153] UL Grouping K-means clustering
[154] SL Prediction Multivariate polynomial regression
MRO [155] RL Control optimization Q-learning
[157] RL Control optimization Fuzzy control
[158,159] UL Pattern identification SOM
[160] UL Prediction Semi-Markov model
[161–164] SL Prediction ANN
CCO [165] RL Control optimization Fuzzy Q-learning
[167] RL Control optimization Fuzzy Q-Learning
[168] UL, RL Control optimization Fuzzy ANN/Q-learning
ICIC [169] RL Control optimization Q-learning
[171,172] RL Control optimization Q-learning
ES [173] RL Control optimization Q-learning
[174] UL Decision making Fuzzy logic
[175,176] UL Grouping, pattern identification Clustering
Self-healing
COC [177] RL Control optimization Actor Critic
[178] RL Control optimization Actor-Critic
[179] SL Control optimization Fuzzy logic
COD [180] UL Anomaly detection Diffusion Maps
[181] SL Anomaly detection Fuzzy logic
[182] SL/UL Diagnosis Naive Bayesian
[183] SL Anomaly detection SVM, Ensemble methods
[178] SL/UL Anomaly detection k-NN, local-outlier-factor
[184] UL Grouping, pattern identification Hidden Markov Model
[185–187] SL Fault Detection k-NN
[188,189] SL Diagnosis Naive Bayesian
Self-coordination
[190] SL Classification Decision Trees
[191] RL Control optimization Actor Critic
[193] RL Control optimization Actor Critic
[194] SL/UL Prediciton Bagged-SVM
MDT
[195,196] SL Verification/estimation Linear correlation
[119] SL Prediction Regression models
[197] SL/UL Prediction/curse of dimensionality Regression models/DR
[120,198] SL Prediction Bagged-SVM/DR
Core Networks
[199] SL Prediction Adaboost, SVM
261
authors focus on the optimization of the users’ experience and of the HO efficiencies. Finally, in [172] also a decentralized Q-learning approach
performance. In [155] the authors take advantage of the Q-learning for interference management is presented. The goal is to improve the
approach to effectively reduce the call drop rates, whereas in [156], systems performance of a macro-cellular network overlaid by femto-
unlike other solutions that assume a general constant mobility, the cells. In order to improve the time of convergence, a mitigation ap-
authors adjust the HO settings in response to the mobility changes in proach has been introduced, allowing them to have significant gains in
the network by means of a distributive cooperative Q-learning. Differ- terms of throughput for both, macro and femto-users. Interesting trade-
ently from [155,156], in [157], the authors take advantage also of fuzzy offs can be studied to compare centralized vs. distributed solutions. In
logic capabilities. These solutions are based on control optimization of the novel context of small cells distributed solutions to interference
HO parameters through RL, so they propose similar solutions to those management are to be preferred over more complex centralized solu-
found in the literature of MLB. In this case we can do the same con- tions, but convergence and instability approaches may appear to affect
siderations about the advantages of considering fuzzy logic in order to the TD learning schemes, compromising system performances [169].
gain in flexibility in the uncertain and complex real network context.
Different approaches in turn, address the problem by identifying suc- 4.2.5. Energy savings
cessful HO events, through solutions based on unsupervised learning. In Energy saving schemes for wireless cellular systems have been
particular, the works of [158] and [159] propose an approach to HO proposed in the past, enabling cells to go into a sleep mode, in which
management based on UL and SOM analysis. The idea is to exploit the they consume a reduced amount of energy. In order to reduce the en-
experience gained from the analysis of data of the network based on the ergy consumption of the eNBs, we can found several works related to
angle of arrival and the received signal strength of the user, to learn ML techniques. An example can be found in [173], where the authors
specific locations where HOs have occurred and decide whether to take advantage of RL to propose a decentralized Q-learning approach to
allow or forbid certain handovers to enhance the network performance. allow energy savings by learning a policy by the iterations with the
The solutions enable self-tuning of HO parameters to learn optimal environment taking into account different aspects over time, such as the
parameters’ adaptation policies. Similarly, in [160] the authors exploit daily solar irradiation. Also, in [174], the authors switch off some un-
the huge amount of information generated in the network to predict derutilized cells during off peak hours. The proposed approach opti-
user traffic distribution. In particular, they take advantage of semi- mizes the number of base stations in dense LTE pico-cell deployments in
Markov model for spatiotemporal mobility prediction in cellular net- order to maximize the energy saving. For the purpose, they use a
works. Finally, the works in [161–164], propose schemes to make combination of Fuzzy Logic, Grey Relational Analysis and Analytic
predictions about UE’s mobility, which allows to anticipate smart HO Hierarchy Process tools to trigger the switch off actions, and jointly
decisions. consider multiple decision inputs for each cell. This last work uses
smart decision theory approaches, which though are not able to take
4.2.3. Coverage and capacity optimization advantage of the previous decisions made in the same environment, as
In case of CCO, different approaches in literature focus on RL so- in turn does the work proposed in [173], as a result of the TD learning
lutions based on continuous interactions with the environment, or- approach. This allows that the work in [173] offers a more solid solu-
iented to online adjusting antenna tilts and transmission power levels tion, considering also past information in the decision. Also for HetNets,
through TD learning approaches. In [165] and [166] a fuzzy Q-learning we find several works, such as, [175,176], where the authors take ad-
approach to optimize the complex wireless network, by learning the vantage of KPIss available in the network for the construction of dif-
optimal antenna tilt control policy has been proposed, and a similar ferent kind of databases to analyze the potential gains that can be
approach is followed also in [167] and [168]. In addition, they also achieved in clustered small cell deployments.
propose to combine fuzzy logic with Q-learning, in order to deal with
continuous input and output variables[167]. also proposes a central 4.2.6. Cell outage compensation
control mechanism, which is responsible to initiate and terminate the The literature already offers different works targeting the problem
learning optimization process of every learning agent deployed in each of COC. For this use case RL has been proven as a valid solution since it
eNB. Finally, [168] innovates with respect to other approaches since in is a continuous decision making/control problem. In this context a
order to adjust the antenna tilt and transmission power parameters, it contribution in the area of self-healing has been presented in [177,178],
considers the load distribution of the different cells involved in the where the authors present a complete solution for the automatic miti-
optimization process, and introduces novel mechanisms to facilitate gation of the degradation effect of the outage by appropriately ad-
cooperative learning among the different SON entities. justing suitable radio parameters of the surrounding cells. The solution
consists of optimizing the coverage and capacity of the identified
4.2.4. Inter-cell interference coordination outage zone, by adjusting the gain of the antenna due to the electrical
Similarly to the CCO case, ML has been proposed in the literature of tilt and the downlink transmission power of the surrounding eNBs. To
ICIC use case as a valid solution, where RL is the principle used tool, implement this approach, the authors propose a RL based on actor-critic
with special emphasis to TD methods, in order to target the optimiza- theory to take advantage of its capability of making online decisions at
tion of control parameters. Several works target the problem to mini- each eNB, and of providing decisions adapting to the evolution of the
mize the interference among cells by using the most common TD scenario in terms of mobility of users, shadowing, etc., and of the de-
learning method, Q-learning [169–172]. The work in [169] is related to cisions made by the surrounding nodes to solve the same problem. A
control inter-cell interference in a heterogeneous femto-macro network. COC contribution also based on ML is targeted in [179], where fuzzy
The work combines information handled by the multi-user scheduling logic is proposed as the driving techniques to fill a coverage gap. The
with decisions taken by a learning agent based on Q-learning, which authors show performance gains by using different parameters, such as,
tries to control the cross-tier interference per resource block. [170] the power transmission, the antenna tilt, and a combination of the two
proposes a distributed solution for ICIC in OFDMA networks based on a schemes. These two works are compared in [177] and the work in [177]
Fuzzy Q-learning implementation. The proposed solution achieves joint is proven superior thanks to the ability to learn from the past experience
improvement for all users, i.e., the improvements of users with bad introduced by the RL actor-critic approach.
quality do not come at the expense of users with good quality. More-
over, a decentralized Q-learning framework for interference manage- 4.2.7. Cell outage detection
ment in small cells is proposed in [171]. The authors focus on a use case As we already mentioned, COD aims to autonomously detect cells
in which the small cell networks aim to mitigate the interference caused that are not operating properly due to possible failures. For this kind of
to the macro-cell network, while maximizing their own spectral problem, anomaly detection algorithms offer an interesting solution
262
that allows to identify outliers measurements, which can be high- 4.2.9. Minimization of drive tests
lighting a hidden problem in the network. Proposals of solutions for this The great majority of literature using the MDT functionality to
problem can be found in [180] and [181]. In particular, [180] presents target MDT use cases, takes advantage of supervised and unsupervised
a solution based on diffusion maps, by means of clustering schemes, learning techniques to provide different solutions for the different use
capable of detecting anomalous behaviours generated by a sleeping cases. An example of that can be observed in [195,196], where the
cell. [181] presents a solution based on fuzzy logic for the automatic authors address the QoS estimation by selecting different KPIs and
diagnosis of a troubleshooting system. In order to determine if there is a correlating them with common nodes measurements, to establish
failure, the authors propose a controller, which receives as an inputs a whether a UE is satisfied with the received QoS. A similar objective is
set of representative KPIs. A similar approach is presented by [182], targeted in [119], however, differently from the previous works, here
where the authors present an automated diagnosis model for Universal the authors focus on multi-layer heterogeneous networks, so in a more
Mobile Telecommunications System (UMTS) networks based on Naive complex and realistic scenario than the traditional macro-cell one. In
Bayesian classifier, and where the model uses both network simulator particular, they present an approach, based on regression models,
and real UMTS network measurements. In the context of this king of which allows to predict QoS in heterogeneous networks for UEs, in-
classifiers , the works in [188,189], also take advantage of NB for au- dependently of the physical location of the UE. This work is extended in
tomated diagnosis based on different inputs network performances. The [197] by taking into account the most promising regression models, but
work in [183] addresses both the case of outage and the one where in also analyzing dimensional reduction techniques. By doing PCA/SPCA
turn the cell can provide a certain level of service, which though does on the input features, and promoting solutions in which only a small
not allow to fulfill the expected UEs requirements. The approach relies number of input features capture most of the variance, the number of
on ensemble methods to train KPIs extracted by human operators to random variables under consideration is reduced. Based on previous
make informed decisions. In [200], the authors consider large data sets results, in [120,198] the same authors define a methodology to build a
to identify anomaly behaving base station. They proposed an algorithm tool for smart and efficient network planning, based on QoS prediction
consisting of preprocessing, detection and analysis phases. The results derived by proper data analysis of UE measurements in the network.
show that by using dimensionality reduction and anomaly detection Moreover, the work in [202] presents a system based on a fuzzy
techniques irregularly behaving base stations can be detected in a self- logic controller to improve network performances by adjusting antenna
organized manner. In [178] data gathered through MDT reports is used tilts values in a LTE system. Differently from previous works, the au-
for anomaly detection purposes. Furthermore, the works of [185–187] thors consider the use of call traces to identify the level of coverage,
take advantage of k-NN algorithm to propose a self-healing solution, in overshooting and overlapping problems, which are the inputs to the
particular to tackle the fault detection domain. Finally, in [184], the algorithm. Also, in [203], the same authors take advantage of con-
authors consider a HetNet and they take advantage of HMM to auto- nection traces (signal strength, traffic, and resource utilization mea-
matically capture the dynamic’s of four different states and probabil- surements) to improve the network infrastructure in terms of spectral
istically estimate if there exist a possible failure. efficiency. The proposed method is designed to be integrated in com-
mercial network planning tool. Finally, in [204] the authors take ad-
vantage of the MDT measurements to build a Radio Environment Map
4.2.8. SON conflicts coordination (REM) by applying spatial interpolation techniques (Bayesian kriging).
As the deployment of stand-alone SON functions is increasing, the The REM (Radio Environmental Map) is then used to detect coverage
number of conflicts and dependencies between them also increases. holes and predict the shape of those areas.
Hence, an entity has been proposed for the coordination of this kind of
conflicts. In this context, current literature includes several works based 4.2.10. Core networks
on ML. In [190] the authors focus on the classification of potential SON As we already mentioned in Section 2, the operational aspects of
conflicts and on discussing the valid tools and procedures to implement core networks elements can be enhanced through, for example, the
a solid self-coordination framework. Q-learning, as a RL method, has automatic configuration of the neighbour cell relations function. In this
been proposed in [191] to take advantage of experience gained in past regard, the idea of applying ML to this function is not new. In [199] the
decisions, in order to reduce the uncertainty associated with the impact authors study the benefits of using ML to root-cause analysis of session
of the SON coordinator decisions when picking an action over another drops, as well as drop prediction for individual sessions. They present
to resolve conflicts. In [192], the authors use Q-learning to deal with an offline Adaboost and SVM method to create a predictor, which is in
the conflict resolution between two SON instances. Decision trees have charge of eliminating/mitigating the session drops by using real LTE
been proposed in [201] to properly adjust Remote Electrical Tilt (RET) data.
and transmission power. Additionally, in [193] the authors provide a
functional architecture that can be used to deal with the conflicts 4.2.11. Virtualized and softwarized 5G architectures
generated by the concurrent execution of multiple SON functions. They Currently 5G-PPP projects like SELFNET and key vendors [18] in
show that the proposed approach is general enough to model all the the industry are working towards the vision of a virtualized and soft-
SON functions and their derived conflicts. First they introduce these werized architecture bringing together the advantages proposed by
SON functions in the context of the general SON architecture, together SDN, NFV and MEC paradigms to introduce programmability and vir-
with high-level examples of how they may interfere. Second, they de- tualization not only in the cloud, but also at the edge of the network.
fine the state and action spaces of the global MDP that models the self- These novel architectures are expected to be able to support Big Data
optimization procedure of the overall RAN segment. They show that the platforms to collect, manage and analyze all the possible data that can
global self-optimization problem can be decomposed onto as many be extracted by the network, to facilitate autonomous decision making.
Markov decision sub-process (subMDP) as SON functions. Finally, the When we go beyond the RAN and we focus on the end-to-end network,
authors in [194] propose a scheme that provides a solution to deal with ML concepts have already been proposed in different works to build
the potential conflicts between two of the most important SON func- cognitive based techniques to operate the network. An example of these
tions in the context of mobility, namely MLB and MRO, which require proposals is well summarized by Clark et al. [205]. In this work, a
the updating of the same set of handover parameters. The proposed Knowledge Plane is advocated, which would bring many advantages to
scheme shifts the complexity to the creation of a prediction model that the networks in terms how the network is operated, automated, opti-
uses historical data and that allows to anticipate the network perfor- mized and troubleshooted. Conceptually this vision is aligned with
mance. Therefore, fast performance evaluations are allowed when the different others proposals in other areas, such as the black-box opti-
optimization is running. mization [206], the autonomic self-x architectures [207], or the work
263
presented in [208]. In this context, the work in [209] analyzes the context of big data analytics, where two of the most important families
reasons why the vision proposed in [205] has still not been brought to are: K-anonymity and differential privacy. A review in more detail of
reality, and the main reason that they find is in the challenges that the aforementioned methods can be found in [212]. Moreover, in order
appear when it comes to autonomously manage a network in a dis- to protect personal data in databases (data storage phase), encryption is
tributed fashion. In particular, the work argues that the emerging trend a fundamental security technique, which transforms data in a way that
of centralization in control brought by the novel SDN vision, will sig- only authorized parties can read it. For more information, the reader is
nificantly reduce this complexity and favour the realization of the ML referred to [212,213].
vision in the network. As a result, in [209] some initial experimental
results based on the vision defined in [205] are brought into reality in 5.2. Implementation of multi-agent systems for network management
the context of a SDN based architecture. Further work in this area is
carried put in the context of different European H2020 projects [6]. The As we have seen in the state of the art sections, reinforcement
work in [210] presents a novel cognitive management architecture that learning solutions and in particular temporal difference learning, offer
manages multiple use cases, like the Service Level Agreement (SLA) and an interesting framework for taking online control decisions based on
the Mobility Quality Predictor. Both use cases are tackled using ma- learned experience, while optimizing long term objectives. This is then
chine learning approaches, the Long Short Term Memory, and a per a useful tool to apply in the context of SON solutions, for the optimi-
user bandwidth predictor. The work in [211] implements Service Level zation of network functions. When the scenario is limited to a decision
Agreement (SLA) through ML approaches. It uses an ANN for evaluation maker, or agent, single agent reinforcement learning can be applied,
of cognitive SLA enforcement of networking services involving Virtua- and optimal solutions can be found and demonstrated. However, when
lized Network Functions and SDN controllers. in the environments multiple agents are coexisting and interacting
among each others, the current approach is to use single agent algo-
5. Challenges for future works rithms and extend them to multi-agent settings. However, this kind of
algorithms require a considerable amount of time before finding a so-
In this section, we focus on some open challenges that still need to lution, which also increases with the state and action spaces. So, re-
be addressed when it comes to making ML based network management inforcement learning approaches dealing with these issues have to be
a reality. investigated. Moreover, no proof of convergence is available demon-
strating that this approach actually reaches meaningful conclusions.
5.1. Big data and privacy Game theoretic effects also appear, since the policy found by an agent
can be affected by the executed policy of the neighbour agent. Even
As already mentioned, the ability to handle data from the network, though ML literature offers different algorithms that can find inter-
in order to continuously optimize it, will be a fundamental feature of 5G esting solutions (e.g. NashQ [214]), the space of possible solutions is so
network management, NG SON and future network architectures. In big that this kind of approaches is not feasible to be used in a realistic
particular, advanced architectures will need to support data platforms network where the time constraints of RRM/SON problem have to be
sourced from the Big Data concept in order to take full advantage of met. So, more research in the area of multi-agent systems, which are
measurement data, trace data, logs, configuration data, etc., from the also compatible with real network requirements need to be in-
entire network. This huge amount of data will need to be handled in vestigated. But also, research involving cognitive tool kits (knowledge
order to provide optimal and automatic decision making solutions. representation, spatial reasoning, acquisition heuristics, consciousness/
Verticals like industrial IoT, smart cities, autonomous cars will generate theory mind among others) and multi-agent systems that rely on re-
huge amount of data to be handled through 5G’s connectivity with low inforcement learning and evolutionary strategies.
latency and high data throughout. At the same time, 5G will need to
support fast data transfers into the cloud, analytics at the cloud and the 5.3. Next generation SON for 5G NR
edge and in the real time to support mission critical scenarios, public
safety, emergency care, etc. For this, traditional ML solutions can be Currently SON only addresses the list of standardized use cases that
used, or novel approaches based on deep learning, which have been we have been describing in the survey, and is defined as an optional
already proven successful when dealing with big amount of data. element. However, to fulfill the objective of a self-organized 5G net-
This research line, additionally, faces important privacy and con- work, SON has to be designed as an integral part of 5G networks. Telus
fidentiality issues. Big Data raises multiple security concerns. As men- and Huawei propose in [18] a joint view on the evolution of the NG-
tioned in Section 4, data come from different sources of the network, SON and propose that in order to provide sustainable network growth
but can also be offered by third parties, e.g., data generated by the user, and to enable 5G success, NG-SON should be considered as an integral
open data, sensor data, among others. Therefore, to come up with a part of the RAN and not just as an add-on feature with a limit set of use
unified privacy policy is extremely challenging at security and privacy cases. NG-SON has to be designed together with NR, the new radio
levels, due to the variety and the granularity of the data. If we add to access technology for 5G, which is currently under definition in 3GPP.
this the speed at which data are created and need to be analysed, the However, this is not an easy task. It is to expect that the user experience
security challenges are huge. In this context, big data is changing the requirements will generate requirements for improved spectrum effi-
security analytics, where robust and scalable privacy preserving mining ciency and higher data rates, and this will come at the cost of increased
algorithms are critical to ensure that the most sensitive private data is network complexity. 5G networks will have to handle an increased
secure. As a result, privacy-preserving data mining is a challenging number of carriers and aggregated bandwidth, larger frequency range,
research line that has to be investigated. In particular, in order to much higher Multiple-input Multiple-output (MIMO) order, increased
guarantee privacy protection, it is important to define the privacy re- number of sites, increased number of RAN features and parameters.
quirements taking into account the lifecycle of the analytics. For ex- This will require increased coordination between the sites and in-
ample, in the data collection phase, it is important to identify the creased number of technologies (in particular LTE/5G). But also, the
personal data needed for processing. The idea is to extract only the network architecture must be capable of supporting an hybrid archi-
needed data for the specific purpose. Aggregated information can also tecture (C-SON and D-SON), a platform to take full advantage of all the
be used instead of personal data. In this context, one of the most re- data generated in the network, and operators should be able to visualize
levant techniques is anonymization, which is the process of modifying and monitor the SON adjustment configurations for optimal decision
personal data in such a way that no identification is possible. Regarding making. Thus, this is an important research line to be considered.
the data analysis phase, different privacy models are available in the In addition, 3GPP networks will also have to interface other kind of
264
networks. The autonomous management of multi-technology networks, a crucial and inevitable tool in order to address automation, self-
where heterogeneous networks including different Radio Access awareness and self-organization in current and future mobile networks.
Technologies (RAT)s, or different layers of the network are coexisting, SON capabilities have been included LTE as add on features of LTE.
e.g., Wi-Fi, mmWave, mobile network layer, transport layer, among However, the need of automation will be further enhanced with the
others, is still immature. These scenarios will tend to emerge with the expected complexity that future 5G network management will have to
advent of the unlicensed spectrum paradigm and with technologies handle, so that NG-SON needs to be redefined as an embedded feature
such as LAA or NR. NR, in particular, will be defined to work over a of 5G networks. On top of that, we have shown that current cellular
wide range of spectrum opportunities, ranging from sub 6 GHz and up networks already generate a huge amount of data that if properly stored
to mmWave spectrum, and under multiple spectrum paradigms, such as and managed could bring new insights in how the networks work and
licensed, unlicensed and shared. The opportunities of autonomous offer new challenges for improving network management taking into
network management in this area are huge. ML has still not been account the experience that can be gained from these data. We have
exploited to handle these networks with intelligence and self-aware- reviewed the main taxonomy of machine learning and the novel trends
ness. In particular, the management of densified and heterogeneous, in that could make this exploitation of data to gain insight of the network
both technologies and layers, architectures, requires the evolution of a reality. Also, we have discussed open data options, as much as al-
NG-SON concepts, which have traditionally been designed and stan- ternatives to get data from the networks, which otherwise are not made
dardized for LTE based networks. available to the academic community. With these motivations in mind,
we have started by reviewing the main concepts and taxonomy of SON,
5.4. Network management of novel softwarized and virtualized network management and ML, and we have reviewed significant aca-
architectures demic literature in the area of network management focusing only on
solutions based on ML. The work has reviewed the status of this exciting
The 5G network architecture must be carefully designed to support research line, while at the same time highlighting open challenges that
the NG-SON introduced in previous subsection. First of all, the archi- we need to deal with in order to make future autonomous network
tecture should be hybrid, in the sense that it has to be able to support management a reality.
both D-SON and C-SON, in order to take full advantage of both para-
digms. D-SON shortens the response cycles and C-SON more easily 7. Competing interests
manage large-sized networks. The architecture should also support
platforms to collect, and analyze data derived from the network, in order The authors declare that there is no conflict of interest regarding the
to take full advantage of measurement data, trace data, logs and con- publication of this paper.
figurations of the entire network. A platform sourced from Big Data
concept should be able to collect as much data as possible, to provide Acknowledgment
automatic decision making. The architecture should also be program-
mable in line with the novel SDN paradigm, in order to allow for easy fix The research leading to these results has received funding from the
and reconfigurability. The platform should also be virtualized in order to Spanish Ministry of Economy and Competitiveness under grant
allow for cost effective introduction of new features and functions, while TEC2017-88373-R (Project 5G-REFINE) and by Generalitat de
meeting the stringent new service verticals’ requirements. This infra- Catalunya grant 2017 SGR 1195. This work also was supported by the
structure is to be managed by an orchestrator entity (in coordination Spanish National Science Council and ERFD funds under Project
with the corresponding virtual network function and virtual infra- TEC2014-60258-C2-2-R.
structure managers), as proposed in ETSI architecture. The network
should not only react to failures, but adapt to the demand, predict it, Supplementary material
based on data analytics collected through the network in order to facil-
itate the network management task. Before reaching this vision, multiple Supplementary material associated with this article can be found, in
challenges need to be addressed, e.g. the self-coordination problem and the online version, at doi:10.1016/j.comcom.2018.07.015.
the solution of conflicts among SON functions executed in different
nodes, or networks, which put the network at risk of instability, or the References
most appropriate location of SON functions and algorithms, to solve
properly the D-SON vs. C-SON implementation issue. Many aspects have [1] Gandalf, Monitoring and self-tuning of RRM parameters in a multi-system net-
to be considered when locating and designing a SON function, e.g. re- work, [Online]. Available:http://www.celtic-initiative.org/.
[2] SOCRATES, Self-Optimisation and self-ConfiguRATion in wirelEss networkS,
sponse time, complexity, size of databases, computational capability of [Online]. Available:http://www.fp7-socrates.org/.
nodes, etc. Centralized (i.e. a large number of cells is involved), dis- [3] SEMAFOUR, Self-Management for Unified Heterogeneous Radio Access Networks,
tributed (approx. 2 cells are involved, coordinating through X2) and local [Online]. Available:http://fp7-semafour.eu/.
[4] SESAME, Small Cells Coordination for Multi-tenancy and Edge services, [Online].
(only one cell is involved) implementations of SON functions have been Available:https://5g-ppp.eu/sesame/.
proposed. No architecture can be claimed superior to the other. The [5] SELFNET, Framework for self-organised network management in virtualised and
virtualized and softwarized paradigm offers the opportunity to build an software defined networks,[Online]. Available:https://5g-ppp.eu/selfnet/.
[6] COGNET, Cognitive Networks,[Online]. Available:https://5g-ppp.eu/cognet/.
elastic SON vision, dynamically moving from one implementation to the [7] O. Aliu, A. Imran, M. Imran, B. Evans, A survey of self organization in future
other. Research on deep reinforcement learning implementations of the cellular networks, IEEE Commun. Surv. Tut. (99) (2012) 1–26.
orchestrator will allow it to self-learn after every decision. Automation [8] Sistelbanda, SN4G SON,[Online]. Available:http://sistelbanda.es/.
[9] Qualcomm, Self managing and enabling seamless roaming,[Online].
will require also all the advancements of the Information Technology
Available:https://www.qualcomm.com/videos/qualcomm-wi-fi-son.
sector, with increased computational capacity, more CPUs and memory [10] Huawei, Huawei’s innovative single SON solution,[Online]. Available:http://
space. However, future orchestrators will need to handle a huge amount www1.huawei.com/en/solutions/broader-smarter/.
of data and learn from them through novel deep learning approaches, the [11] A. communications, Powering 4G networks,[Online]. Available:http://www.
airhopcomm.com.
smart network management decisions. This research line is still highly [12] Elisa, Elisa automate, mobile word congress 2018,[Online]. Available:https://
immature and requires a lot of efforts. elisaautomate.com/.
[13] Cellwize, Driving value through SON,[Online]. Available:http://www.cellwize.
com.
6. Conclusions [14] AVIAT, Wireless products for small cell applications,[Online]. Available:https://
startupgenome.co/aviat-networks.
In this work we have motivated the need for ML to be considered as [15] INTRACOM TELECOM, SON automation in Small cell backhaul, INTRACOM
265
TELECOM, Mobile Word Congress 2018,[Online]. Available:http://www. [48] 3GPP, Self-Organizing Networks (SON): Policy Network Resource Model (NRM)
intracom-telecom.com/en/products/wireless_network_systems/4G_smallcell_son_ Integration Reference Point (IRP); Information Service(IS); Solution Set (SS) de-
backhaul/son_Automation_solutionPaper.htm. finitions, (Release 10), TS, 3GPP, 2011.
[16] C.C.S. (CCS), Ccs metnet self-organising mmwave access and backhaul,[Online]. [49] 3GPP, Evolved Universal Terrestrial Radio Access Network (E-UTRAN); Network
Available:https://www.ccsl.com/. Resource Model (NRM) Integration Reference Point (IRP); Solution Set(SS) defi-
[17] S.C. Forum, Small cell industry award,[Online]. Available:http://www. nitions, (Release 10), TS, 3GPP, 2011.
smallcellforum.org/events/awards. [50] 3GPP, Telecommunication management; Self-Organizing Networks (SON); Self-
[18] TELUS, Huawei, Next generation son for 5g, white paper, 2016,[Online]. healing concepts and requirements (Release 12), TS, 3GPP, 2014.
Available:http://www.huawei.com/en/industry-insights/outlook/mbb-2020/ [51] 3GPP, Technical Specification Group Services and System Aspects;
trends-insights/next-generation-son-for-5g. Telecommunication Management; Study on Energy Savings Management (ESM)
[19] 3GPP, Technical Specification Group Services and System Aspects; (Release 10), TR, 3GPP, 2010.
Telecommunication Management; Self-Organizing Networks (SON); Concepts and [52] 3GPP, Technical Specification Group Services and System Aspects; Study on
requirements (Release 13), TS, 3GPP, 2016. Operations, Administration and Maintenance (OAM) aspects of inter-Radio-
[20] A. Imran, A. Zoha, A. Abu-Dayya, Challenges in 5G: how to empower SON with big Access-Technology (RAT) energy saving (Release 11), TR, 3GPP, 2012.
data for enabling 5G, IEEE Netw. 28 (6) (2014) 27–33. [53] 3GPP, Universal Mobile Telecommunications System (UMTS); LTE;
[21] NOKIA, Intelligent self-organizing networks: increased automation for higher Telecommunication management; Energy Saving Management (ESM); Concepts
network performance with lower costs, White Paper (2013). and requirements (Release 10), TS, 3GPP, 2011.
[22] Ovium, Why fronthaul matters: a key foundation for centralized and cloud ran, [54] 3GPP, Technical Specification Group Radio Access Network; Evolved Universal
white paper, 2015, [Online]. Available:http://e-blink.com/wp-content/uploads/ Terrestrial Radio Access (E-UTRA); Radio Resource Control (RRC); Protocol
Why_Fronthaul_Matters_EBLINK_Ovum_whitepaper.pdf. Specification (Release 10), TS, 3GPP, 2010.
[23] N. Baldo, L. Giupponi, J. Mangues-Bafalluy, Big data empowered self organized [55] 3GPP, Performance Management (PM); Performance Measurements UTRAN
networks, 20th IEEE European Wireless, (2014). (Release 9), TS, 3GPP, 2009.
[24] B. Romanous, N. Bitar, A. Imran, H. Refai, Network densification: challenges and [56] 3GPP, Automatic Neighbor Relation (ANR) management, Concepts and
opportunities in enabling 5G, 20th IEEE International Workshop on Computer Requirements, (Release 11), TR, 3GPP, 2011.
Aided Modelling and Design of Communication Links and Networks (CAMAD, [57] 3GPP, Technical Specification Group Services and System Aspects;
Special Session: Dense and Elastic Big Data Empowered SON, (2015), pp. Telecomunications Management; Self-Organizing Networks (SON) Policy Network
129–134. Resource Model (NRM) Integration Reference Point (IRP)(Release 10), TR, 3GPP,
[25] D. Falko, A study of self-organization mechanisms in ad hoc and sensor networks, 2009.
Els. Comput. Commun. (2008) 3018–3019. [58] 3GPP, Configuration Management (CM); UTRAN network resources Integration
[26] PauloValente. Klaine, MuhammadAli. Imran, Oluwakayode Onireti, Reference Point (IRP); Network Resource Model (NRM)(Release 9), TS, 3GPP,
RichardDemo Souza, A survey on ML techniques applied to self-organizing cellular 2009.
networks, IEEE Commun. Surv. Tut. (2017). [59] 3GPP, TSG SA WG5 (Telecom Management) Meeting 85; Study of implementation
[27] D. Kreutz, F.M.V. Ramos, P.E. Verissimo, C.E. Rothenberg, S. Azodolmolky, alternative for SON coordination, Technical Report, 3GPP, 2012.
S. Uhlig, Software-defined networking: a comprehensive survey, Proc. IEEE 103 [60] 3GPP, Configuration Management (CM); UTRAN network resources Integration
(1) (2015) 14–76, https://doi.org/10.1109/JPROC.2014.2371999. Reference Point (IRP); Solution Set (Release 10), Technical Report, 3GPP, 2011.
[28] 3GPP, Evolved Universal Terrestrial Radio Access Network (EUTRAN); Self con- [61] 3GPP, Radio Resource Control (RRC); Protocol specification (Release 10), TS,
figuring and self optimizing network uses case and solutions (Release 9), 3GPP, TR 3GPP, 2011.
T36.902, v9.2.0, 2010. [62] 3GPP, UTRAN overall description (Release 10), TS, 3GPP, 2011.
[29] 3GPP Telecommunication management; Self-Organizing Networks (SON); [63] 3GPP, UTRAN Iu Interface: General Aspects and Principles (Release 10), TS, 3GPP,
Concepts and requirements (Release 11), 3GPP, TS 32.500, v11.1.0, 2011. 2011.
[30] 3GPP, Self Configuration of Network Elements; Concepts and Requirements, [64] 3GPP, UTRAN Iur interface Radio Network Subsystem Application Part (RNSAP)
Release 9, 3GPP, TS 32.501, 2010. signalling (Release 9), TS, 3GPP, 2010.
[31] 3GPP, Self-configuration of network elements Integration Reference Point (IRP); [65] 3GPP, Study on Network Management (NM) centralized Coverage and Capacity
Information Service (IS) (Release 10), 3GPP, TS 32.502, v10.1.0, 2010. Optimization (CCO) Self-Organizing Networks (SON) function, TS, 3GPP, 2013.
[32] 3GPP, Self-Configuration of Network Elements Integration Reference Point (IRP); [66] 3GPP, Integration Reference Point (IRP) overview and usage guide (Release 10),
Common Object Request Broker Architecture (CORBA) Solution Set (SS)(Release TS, 3GPP, 2011.
8), TR, 3GPP, 2008. [67] 3GPP, Self-Organizing Networks (SON) Policy Network Resource Model (NRM)
[33] Software Management (SWM); Software and Integration Reference Point (IRP) Integration Reference Point (IRP); Requirements (Release 11), TS, 3GPP, 2012.
Requirements (Release 10), 3GPP, TR 32.531, 2010. [68] 3GPP, Self-Organizing Networks (SON) Policy Network Resource Model (NRM)
[34] 3GPP, Software management Integration Reference Point (IRP); Information Integration Reference Point (IRP); Information Service (IS) (Release 11), TS, 3GPP,
Service (IS) (Release 8), TR, 3GPP, 2009. 2014.
[35] 3GPP, Software Management (SWM) Integration Reference Point (IRP): Common [69] 3GPP, Evolved Universal Terrestrial Radio Access Network (E-UTRAN) Network
Object Request Broker Architecture (CORBA) Solution Set (SS) (Release 9), Resource Model (NRM) Integration Reference Point (IRP); Information Service (IS)
Technical Report, 3GPP, 2009. (Release 11), TS, 3GPP, 2012.
[36] 3GPP, Technical Specification Group Services and System Aspects; [70] 3GPP, Evolved Universal Terrestrial Radio Access Network (E-UTRAN) Network
Telecommunications management; Evolved Universal Terrestrial Radio Access Resource Model (NRM) Integration Reference Point (IRP); Solution Set (SS) defi-
Network (E-UTRAN) Network Resource Model (NRM) Integration Reference Point nitions (Release 11), TS, 3GPP, 2012.
(IRP); Requirements (Release 9), TS, 3GPP, 2009. [71] 3GPP, Gap analysis between 3GPP SA5 specifications and NGMN Top Operational
[37] 3GPP, Evolved Universal Terrestrial Radio Access Network (E-UTRAN) Network Efficiency (OPE) Recommendations (Release 12), TS, 3GPP, 2013.
Resource Model (NRM) Integration Reference Point (IRP); Information Service (IS) [72] 3GPP, Data formats for multi-vendor plug and play eNodeB connection to the
(Release 11), TS, 3GPP, 2014. network (Release 12), TS, 3GPP, 2013.
[38] 3GPP, Evolved Universal Terrestrial Radio Access Network (E-UTRAN) Network [73] 3GPP, Gap analysis between 3GPP SA5 specifications and NGMN Top Operational
Resource Model (NRM) Integration Reference Point (IRP): CORBA solution set Efficiency (OPE) Recommendations (Release 12), TR, 3GPP, 2013.
(Release 9), TR, 3GPP, 2009. [74] 3GPP, Study of heterogeneous networks management (Release 12), TR, 3GPP,
[39] 3GPP, Evolved Universal Terrestrial Radio Access Netwrok (E-UTRAN) Network 2014.
Resource Model (NRM) Integration Reference Point (IRP): eXtensible Markup [75] 3GPP, Study on Operations, Administration and Maintenance (OAM) aspects of
Language (XML) definitions (Release 9), TR, 3GPP, 2004. Network Sharing (Release 12), TR, 3GPP, 2013.
[40] 3GPP, Study of Self-Organizing Networks (SON) related Operations, [76] 3GPP, Study on Next Generation Self-Optimizing Network (SON) for UTRAN and
Administration and Maintenance (OAM) for Home Node B (HNB) (Release 9), TR, E-UTRAN (Release 12), TR, 3GPP, 2013.
3GPP, 2004. [77] 3GPP, Study on Energy Saving Enhancement for E-UTRAN (Release 12), TR, 3GPP,
[41] 3GPP, Telecommunication management; Self-Organizing Networks (SON); Study 2013.
on self-healing (Release 9), TR, 3GPP, 2010. [78] 3GPP, Study on Enhancements of OAM aspects of distributed Self-Organizing
[42] 3GPP, Performance measurements Evolved Universal Terrestrial Radio Access Networks (SON) functions, Release 12, TR, 3GPP, 2014.
Network (E-UTRAN)(Release 10), TS, 3GPP, 2012. [79] 3GPP, Overview of 3gpp release 14.
[43] 3GPP, Universal Mobile Telecommunications System (UMTS); UTRAN Iu Interface [80] 3GPP, RAN evolution of LTE in release 14, 2016,[Online]. Available:http://www.
RANAP Signalling, TS, 3GPP, 1999. 3gpp.org/news-events/3gpp-news/1768-ran_rel14.
[44] 3GPP, Radio Access (E-UTRA) and Evolved Universal Terrestrial Radio Access [81] 3GPP, Study on OAM aspects of SON for AAS-based deployments, TR, 3GPP, 2017.
Network (E-UTRAN); Overall description; Stage 2 (Release 12), TS, 3GPP, 2015. [82] 3GPP, 3rd Generation Partnership Project; Technical Specification Group Radio
[45] 3GPP, LTE; Evolved Universal Terrestrial Radio Access Network (E-UTRAN); S1 Access Network; NR; NR and NG-RAN Overall Description; Stage 2 (Release 15),
Application Protocol (S1AP) (Release 9), TS, 3GPP, 2010. TS, 3GPP, 2018.
[46] 3GPP, Technical Specification Group Radio Access Network; Evolved Universal [83] 3GPP, Technical Specification Group Services and System Aspects;
Terrestrial Radio Access Network (EUTRAN); X2 Application Protocol (X2AP) Telecommunication Management; Project scheduling and open issues for SA5,
(Release 11), TS, 3GPP, 2013. (Release 8), Technical Report, 3GPP, 2010.
[47] 3GPP, Self-Organizing Networks (SON) Policy Network Resource Model (NRM) [84] M. Amirijoo, P. Frenger, F. Gunnarsson, Neighbor cell relation list and physical
Integration Reference Point (IRP); Information Service(IS), 3rd Generation cell identity self-organization in LTE, ICC Workshops-IEEE International
Partnership Project (3GPP), (Release 10), TS, 3GPP, 2011. Conference on Communications Workshops, (2008), pp. 31–41.
266
[85] F. Parodi, M. Kylvaja, G. Alford, J. Li, J. Pradas, An automatic procedure for [121] S.T. Roweis, K.S. Lawrence, Nonlinear dimensionality reduction by locally linear
neighbor cell list definition in cellular networks, IEEE International Symposium on embedding, American Association for the Advancement of Science, (2000), pp.
a World of Wireless, Mobile and Multimedia Networks (WoWMoM), (2007), 2323–2326.
pp. 1–6. [122] S.T. Roweis, EM algorithms for PCA and SPCA, Advances in Neural Information
[86] 3GPP, Technical Specification Group Radio Access Network, Evolved Universal Processing Systems, The MIT Press, 1998.
Terrestrial Radio Access Network (E-UTRAN), Self configuring and Self optimizing [123] V.J. Hodge, J. Austin, A survey of outlier detection methodologies, Artif. Intell.
Network (SON) Uses Case and Solutions (Release 9), TR, 3GPP, 2011. Rev. 22 (2004).
[87] F. Parodi, M. Kylvaja, G. Alford, J. Li, J. Pradas, Recommended practices for multi- [124] I. de la Bandera, R. Barco, P. Muñoz, I. Serrano, Cell outage detection based on
vendor SON deployment, D2- NGMN Aliance V1.0, (2014). handover statistics, IEEE Commun. Lett. 19 (7) (2015).
[88] 3GPP, Evolved Universal Terrestrial Radio Access (E-UTRA); Physical layer pro- [125] P. Muñoz, R. Barco, A. Gómez Andrades, Correlation based time-series analysis for
cedures, TS, 3GPP, 2011. cell degradation, IEEE Commun. Lett. (2016).
[89] NGMN, Recommended practices for multi-vendor SON deployment, NGMN P- [126] S. Schaal, C. Atkeson, Robot juggling: an implementation of memory-based
SmallCell Work Stream 2 (WS2) first deliverable, (2014). learning, Control Syst. Mag. (1994).
[90] 3GPP, Evolved Universal Terrestrial Radio Access (E-UTRA) and Evolved [127] S. Thrun, Learning to play the game of chess, Advances in Neural Information
Universal Terrestrial Radio AccessE-UTRAN; Overall description; Stage 2, TS, Processing Systems 7, 1995, The MIT Press., Cambridge, MA, 1995.
3GPP, 2015. [128] M.L. Littman, Markov games as a framework for multi-agent reinforcement
[91] 3GPP, SON Self-healing management, Technical Report, 3GPP, 2011. learning, 11th International Conference on Machine Learning, (1994), pp.
[92] Seppo Hämäläinen, Henning Sanneck, Cinzia Sartori, LTE Self-Organising 157–163.
Networks (SON): Network Management Automation for Operational Efficiency, [129] P. Liviu, L. Sean, Cooperative multi-agent learning: the state of the art, Auton.
John Wiley and Sons, 2012. Agent Multi Agent Syst. 11 (3) (2005) 387–434.
[93] L.C. Schmelz, M. Amirijoo, A. Eisenblaetter, R. Litjens, M. Neuland, J. Turk, A [130] R.S. Sutton, A.G. Barto, Reinforcement Learning: An Introduction, MIT Press,
coordination framework for self-organisation in LTE networks, IEEE International 1998.
Symposium on Integrated Network Management, (2011). [131] Y. Bengio, A. Courville, P. Vincent, Representation learning: a review and new
[94] Z. Altman, M. Amirijoo, F. Gunnarson, H. Hoffmann, I.Z. Kovács, D. Laselva, perspectives, IEEE Trans. Pattern Anal. Mach. Intell. 35 (8) (2013) 1798–1828.
K. Spaey, A. Tall, H. Van Den Berg, K. Zetterberg, On design principles for self- [132] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, 2016. http://
organizing network functions, 11th International Symposium on Wireless www.deeplearningbook.org .
Communications Systems (ISWCS), (2014), pp. 454–459. [133] 3GPP, Technical Specification Group Services and Systems Aspects; Charging
[95] 3GPP, Study on Minimization of drive-tests in Next Generation Networks, TR, management; Charging Data Record (CDR) parameter description (Release 13),
3GPP, 2009. TS, 3GPP, 2016.
[96] J. Johansson, W.A. Hapsari, S. Kelley, G. Bodog, Minimization of drive tests in [134] 3GPP, Telecommunication management; Performance Management (PM);
3GPP (release 11), IEEE Commun. Mag. (2012). Concept and requirements (Release 10), TS, 3GPP, 2010.
[97] Nokia Siemens Networks, Automating core network management, mobile world [135] 3GPP, Technical Specification Group TSG RAN; Radio measurement collection for
congress, 2012,[Online]. Available:https://networks.nokia.com/solutions/cloud- Minimization of Drive Tests (MDT), Release 10, TS, 3GPP, 2010.
native-core-network. [136] Internet Engineering Task Force (IETF), Information Model for IP Flow
[98] A. Blenk, A. Basta, M. Reisslein, W. Kellerer, Survey on network virtualization Information Export (IPFIX), Technical Report, 3GPP, 2013.
hypervisors for software defined networking, IEEE Commun. Surv. Tut. 18 (1) [137] Orange, The D4D challenge,[Online]. Available:http://www.d4d.orange.com/en/
(2016) 655–685. Accueil.
[99] A Valdivieso Caraguay L. Garcia Vilalba, Monitoring and discovery for self-orga- [138] F. Ricciato, P. Widhalm, M. Craglia, F. Pantisano, Estimating population density
nized network management in virtualized and software defined networks, Sensors distribution from network-based mobile phone data, JRC and AIT, (2015).
17 (2017) 700–731. [139] M. Laner, P. Svoboda, P. Romirer, N. Nikaein, F. Ricciato, A Comparison Between
[100] R. Caruana, A. Niculescu-Mizil, An empirical comparison of supervised learning One-Way Delays in Operating HSPA and LTE Networks, 8th Int’l workshop on
algorithms, 23rd International Conference on Machine Learning, (2006). Wireless Network Measurements (WINMEE’12), Pedeborn, Germany, (2012).
[101] P. McCullagh, J. Nelder, Generalized Linear Models, Second Edition, Chapman and [140] 3GPP, Technical Specification Group Core Network and Terminals; AT command
Hall, Boca Raton, 1989. set for User Equipment (UE) (Release 12), TS, 3GPP, 2013.
[102] H. Zhang, The optimality of Naive Bayes, 17th International Florida Artificial [141] OpenCellID, Largest open database of cell towers and geolocation, [Online].
Intelligence Research Society Conference, (2004). Available:https://opencellid.org/.
[103] V. Vapnik, An overview of statistical learning theory, IEEE Trans. Neural Netw. 10 [142] OpenSignal, 3G and 4G LTE Cell Coverage Map,[Online]. Available:https://
(5) (1999). opensignal.com/.
[104] A.J. Smola, B. Scholkopf, A tutorial on support vector regression, Stat. Comput. [143] Antenas GSM, [Online]. Available:http://antenasgsm.com/.
(2004) 199–222. [144] Google geolocation api, [Online]. Available:https://developers.google.com/
[105] A. Müller, Kernel approximations for efficient SVMs (and other feature extraction maps/.
methods), Data Sci. Mach. Learn. Resour. (2012). [145] N. Bui, J. Widmer, Owl: a reliable online watcher for LTE control channel mea-
[106] M.C. Bishop, Pattern Recognition and Machine Learning, Busines Dia, Llc. Springer surements, arXiv:1606.00202l (2016).
Science, 2006. [146] H.D. Trinh, N. Bui, J. Widmer, L. Giupponi, P. Dini, Analysis and modeling of
[107] J.R. Quinlan, Induction of Decision Trees. Machine Learning, Kluwer Academic mobile traffic using real traces, submitted to IEEE Personal Indoor and Mobile
Publishers, 1986, pp. 81–106. Radio Communications (PIMRC) (2017).
[108] L. Rokach, O. Maimon, Data mining with decision trees: theory and applications, [147] Centre Tecnològic de Telecomunicacions de Catalunya (CTTC), The LENA-EPC
Series in Machine Perception and Artificial Intelligence, (2008). network simulator, [Online]. Available:http://networks.cttc.es/mobile-networks/
[109] D. Opitz, R. Maclin, Popular ensemble methods: an empirical study, J. Artif. Intell. software-tools/lena/.
Res. 11 (1999) 169–198. [148] M. Peng, D. Liang, Y. Wei, J. Li, H.H. Chen, Self-configuration and self-optimi-
[110] L. Rokach, Ensemble-based classifiers, Artif. Intell. (2010) 1–39. zation in LTE-advanced heterogeneous networks, IEEE Commun. Mag. 51 (5)
[111] T.G. Dietterich, An experimental comparison of three methods for constructing (2013) 36–45.
ensembles of decision trees: bagging, boosting and randomization, Mach. Learn. [149] S.S. Mwanje, A. Mitschele-Thiel, Minimizing handover performance degradation
40 (2000) 139–157. due to LTE self organized mobility load balancing, 77th IEEE Conference on
[112] T.G. Dietterich, Machine-learning research: four current directions, Am. Assoc. Transactions on Vehicular Technology (VTC Spring), (2013), pp. 1–5.
Artif. Intell. (1997). [150] P. Muñoz, R. Barco, I. de la Bandera, Optimization of load balancing using fuzzy Q-
[113] Y. Freund, R.E. Schapire, A decision-theoretic generalization of on-line learning Learning for next generation wireless networks, Expert Syst. Appl., Els. 40 (2013)
and application to boosting, 2nd European Conference on Machine Learning, 984–994.
(1995), pp. 148–156. [151] P. Muñoz, R. Barto, I. de la Bandera, Load balancing and handover joint optimi-
[114] G. Zoubin, Unsupervised learning, Gatsby Computational Neuroscience Unit zation in LTE networks using fuzzy logic reinforcement learning, Comput. Netw.
University College London, UK, (2004). Els. (2014) 113–125.
[115] W.H. Press, S.A. Teukolsky, W.T. Vetterling, B.P. Flannery, Section 16.1. Gaussian [152] J. Suga, Y. Kojima, M. Okuda, Centralized mobility load balancing scheme in LTE
Mixture Models and k-Means Clustering, The Art of Scientific Computing, Third, systems, 8th International Symposium on Wireless Communication Systems,
York: Cambridge University, 2007. Aachen, (2011).
[116] A. Ultsch, H.P. Siemon, Kohonen’s self organizing feature maps for exploratory [153] E. Bergner, Unsupervised learning of traffic patterns in self-Optimizing 4th gen-
data analysis, International Neural Network Conference, (1990). eration mobile networks, Master of Science Thesis, KTH Computer Science and
[117] T. Hastie, R. Tibshirani, J. Friedman, The Elements of Statistical Learning (2nd Communications, Stockholm, Sweden (2012).
ed.), New York: Springer, 2009. [154] C.A.S. Franco, J.R.B. de Marca, Load balancing in self-organized heterogeneous
[118] R. Nock, F. Nielsen, On weighting clustering, IEEE Trans. Pattern Anal. Mach. LTE networks: a statistical learning approach, IEEE Latin-American Conference on
Intell. 28 (2006) 1–13. Communications (LATINCOM), (2015), pp. 1–5.
[119] J. Moysen, N. Baldo, L. Giupponi, J. Mangues-Bafalluy, Predicting QoS in LTE [155] W. Qin, Y. Teng, M. Song, Y. Zhang, X. Wang, A Q-learning approach for mobility
hetnets based on location-independent UE measurement, 20th IEEE International robustness optimization in LTE-SON, 15th IEEE International Conference on,
Workshop on Computer Aided Modelling and Design of Communication Links and Communication Technology (ICCT), (2013), pp. 818–822.
Networks, Guildford, UK, (2015). [156] S.S. Mwanje, A. Mitschele-Thiel, Distributed cooperative Q-learning for mobility-
[120] J. Moysen, L. Giupponi, J. Mangues-Bafalluy, A mobile network planning tool sensitive handover optimization in LTE SON, IEEE Symposium the Computers and
based on data analytics, Mobile Inf. Syst., Hindawi (2017). Communication (ISCC), (2014), pp. 1–6.
267
[157] P. Muñoz, R. Barco, I. de la Bandera, On the potential of handover parameter IEEE 25th Annual International Symposium on Personal, Indoor, and Mobile Radio
optimization for self-Organizing networks, IEEE Trans. Veh. Technol. (2013). Communication (PIMRC), (2014), pp. 1626–1630.
[158] N. Sinclair, D. Harle, I.A. Glover, J. Irvine, R.C. Atkinson, Parameter optimization [187] S. Chernov, F. Chernogorov, D. Petrov, T. Ristaniemi, Data mining framework for
for LTE handover using an advanced SOM algorithm, IEEE Conference on random access failure detection in LTE networks, 25th IEEE Annual International
Transactions on Vehicular Technology (VTC), (2013). Symposium on Personal, Indoor, and Mobile Radio Communication (PIMRC),
[159] N. Sinclair, D. Harle, I. Glover, J. Irvine, R. Atkinson, An advanced SOM algorithm (2014), pp. 1321–1326.
applied to handover management within LTE, IEEE Trans. Veh. Technol. (2013) [188] R. Barco, V. Wille, L. Díez, System for automated diagnosis in cellular networks
183–1894. based on performance indicators, Eur. Trans. Telecommun. (2005) 399–409.
[160] H. Farooq, A. Imran, Spatio-temporal mobility prediction in proactive self-orga- [189] R.M. Khanafer, B. Solana, J. Triola, R. Barco, L. Moltsen, Z. Altman, P. Lazaro,
nizing cellular networks, IEEE Commun. Lett. 21 (2) (2017) 370–373. Automated diagnosis for UMTS networks using Bayesian network approach, IEEE
[161] Z. Ali, N. Baldo, J. Mangues-Bafalluy, L. Giupponi, Machine learning based Trans. Veh. Technol. 57 (2008) 2451–2461.
handover management for improved qoe in LTE, IEEE/IFIP Network Operations [190] H.Y. Lateef, A. Imran, A. Abu-dayya, A framework for classification of self-orga-
and Management Symposium, (2016), pp. 794–798. nising network conflicts and coordination algorithms, IEEE Personal Indoor and
[162] E. Ostlin, H.J. Zepernick, H. Suzuki, Macrocell radio wave propagation prediction Mobile Radio Communications PIMRC, (2013).
using an artificial neural network, IEEE Vehicular Technology Conference (VTC [191] O. Iacoboaeia, B. Sayrac, S. Jemaa, P. Bianchi, SON coordination for parameter
Fall), (2004), p. 57. conflict resolution: a reinforcement learning framework, IEEE Wireless
[163] A. Quintero, O. García, A profile-based strategy for managing user mobility in Communications and Networking Conference 2014 (IEEE WCNC 2014), (2014).
third-generation mobile systems, IEEE Commun. Mag. (2004) 134–139. [192] P. Muñoz, R. Barco, S. Flores, Conflict resolution between load balancing and
[164] K. Majumdar, N. Das, Mobile user tracking using a hybrid neural network, handover optimization in LTE networks, IEEE Commun. Lett. 18 (2014)
Wireless Networks, (2005), pp. 275–284. 1795–1798.
[165] R. Razavi, S. Klein, H. Claussen, A fuzzy reinforcement learning approach for self- [193] J. Moysen, L. Giupponi, Self-coordination of parameter conflicts in D-SON archi-
optimization of coverage in LTE networks, Bell Labs Tech. J. 15 (2010) 153–175. tectures: a markov decision process framework, EURASIP J. Wirel. Commun.
[166] M. Naseer ul Islam, A. Mitschele-Thiel, A cooperative fuzzy Q-learning for self- Netw. (2015).
organized coverage and capacity optimization, IEEE 23rd International [194] J. Moysen, M. García Lozano, L. Giupponi, S. Ruiz, Conflict resolution in mobile
Symposium on Personal Indoor and Mobile Radio Communications (PIMRC), networks: a self-coordination framework based on non-dominated solutions and
(2012). machine learning for data analytics, IEEE Comput. Intell. Mag. 13 (2018) 52–64.
[167] J. Li, J. Zeng, X. Su, W. Luo, J. Wang, Self-Optimization of coverage and capacity [195] F. Chernogorov, T. Nihtilä, QoS verification for minimization of drive tests in LTE
in LTE networks based on central control and decentralized fuzzy Q-learning, Int. networks, 75th IEEE Vehicular Technology Conference (VTC Spring), (2012),
J. Distrib. Sensor Netw., Hindawi (2012). pp. 6–9.
[168] S. Fan, H. Tian, C. Sengul, Self-optimization of coverage and capacity based on a [196] F. Chernogorov, J. Puttonen, User satisfaction classification for minimization of
fuzzy neural network with cooperative reinforcement learning, J. Wireless drive tests QoS verification, 24th IEEE Personal Indoor and Mobile Radio
Commun. Netw. (2014). Communications (PIMRC), London, United Kingdom, (2013), pp. 2165–2169.
[169] A. Galindo-Serrano, L. Giupponi, G. Auer, Distributed learning in multiuser [197] J. Moysen, L. Giupponi, J. Mangues-Bafalluy, On the potential of ensemble re-
OFDMA femtocell networks, 73rd IEEE Vehicular Technology Conference (VTC gression techniques for future mobile network planning, 21th IEEE Symposium on
Spring), (2011), pp. 1–6. Computers and Communications (ISCC), Messina, Italy, (2016).
[170] M. Dirani, Z. Altman, A cooperative reinforcement learning approach for inter-cell [198] J. Moysen, L. Giupponi, J. Mangues-Bafalluy, A machine learning enabled network
interference coordination in OFDMA cellular networks, 8th Int Modeling and planning tool, 27th IEEE Personal Indoor and Mobile Radio Communications
Optimization in Mobile, Ad Hoc and Wireless Networks (WiOpt), (2010), pp. (PIMRC), Valencia, Spain, (2016).
170–176. [199] B. Daróczy, A. Benczúr, P. Vaderna, Machine learning based session drop pre-
[171] M. Bennis, S.M. Perlaza, P. Blasco, Self-Organization in small cell networks: a diction in LTE networks and its SON aspects, 5th International Workshop on Self-
reinforcement learning approach, IEEE Trans. Wireless Commun. 12 (7) (2013). Organizing Networks (IWSON), Glasgow, (2015).
[172] M. Simsek, A. Czylwik, A. Galindo-Serrano, L. Giupponi, Improved decentralized [200] J. Turkka, D. Gil, Anomaly detection framework for tracing problems in radio
Q-learning algorithm for interference reduction in LTE-femtocells, 5th networks, 10th International Conference on Networks (ICN), (2011).
International Workshop on Self-Organizing Networks IWSON, Glasgow, (2015). [201] R. Romeikat, B. Bauer, T. Bandh, G. Carle, H. Sanneck, L.C. Schmelz, Policy-driven
[173] M. Miozzo, L. Giupponi, M. Rossi, P. Dini, Distributed Q-learning for energy workflows for mobile network management automation, 6th International
harvesting heterogeneous networks, IEEE ICC 2015 Workshop on Green Wireless Communications and Mobile Computing Conference (IWCMC), (2010),
Communications and Networks with Energy Harvesting, Smart Grids, and pp. 1111–1115.
Renewable Energies, (2015). [202] V. Buenestado, M. Toril, S. Luna, J.M. Ruiz, A. Mendo, Self-tuning of remote
[174] A. Dudnikova, P. Dini, L. Giupponi, D. Panno, Multiple criteria plus fuzzy logic electrical tilts based on call traces for coverage and capacity optimization in LTE,
switch off method for dense heterogeneous networks, 20th IEEE International IEEE Trans. Veh. Technol. (2017).
Workshop on Computer Aided Modelling and Design of Communication Links and [203] M. Toril, R. Acedo-Hernàndez, S. Luna-Ramírez, A. Sánchez, C. Úbeda, Automatic
Networks, Guildford (UK), (2015). clustering algorithms for indoor site selection in LTE, Hindawi Mobile Inf. Syst.
[175] E. Ternon, P. Agyapong, L. Hu, A. Dekorsy, Energy savings in heterogeneous (2017).
networks with clustered small cell deployments, 11th IEEE International [204] A. Galindo-Serrano, B. Sayrac, S. Ben Jemaa, J. Riihijärvi, P. Mähönen, Harvesting
Symposium on Wireless Communications Systems (ISWCS), (2014). MDT data: radio environment maps for coverage analysis in cellular networks,
[176] E. Ternon, P. Agyapong, L. Hu, A. Dekorsy, Database-aided energy savings in next International Conference on Cognitive Radio Oriented Wireless Networks
generation dual connectivity heterogeneous networks, IEEE WCNC’14 Track 3: (CROWNCOM), (2013).
Mobile and Wireless Networks, (2014). [205] D. Clark, et al., A knowledge plane for the internet, Proc. of the 2003 Conference
[177] J. Moysen, L. Giupponi, A reinforcement learning based solution for self-healing in on Applications, Technologies, Architectures, and Protocols for Computer
LTE networks, 80th IEEE Vehicular Technology Conference (VTC Fall), Vancouver, Communications, (2003).
Canada, (2014). [206] L.M. Rios, N.V. Sahinidis, B. Sayrac, Derivative free optimization: a review of al-
[178] O. Onireti, A. Zoha, J. Moysen, A. Imran, L. Giupponi, M. Ali Imran, A. Abu-Dayya, gorithms and comparison of software implementations, J. Global Optim. (2013)
A cell outage management framework for dense heterogeneous networks, IEEE 1247–1293.
Trans. Veh. Technol. 64 (2015) 2097–2113. [207] H. Derbel, N. Agoulmine, M. Salan, Anema: autonomic network architecture to
[179] A. Saeed, O.G. Aliu, M.A. Imran, Controlling self healing cellular networks using support self-configuration and self-optimization in ip networks, Comput. Networks
fuzzy logic, IEE Wireless Communications and Networking Conference (WCNC), (2009) 418–430.
Paris, France, (2012). [208] e.a.M. Zorzi, COBANETS: a new paradigm for cognitive communications systems,
[180] F. Chernogorov, J. Turkka, T. Ristaniemi, A. Averbuch, Detection of sleeping cells International Conference on Computing, Networking and Communications ICNC,
in LTE networks using diffusion maps, IEEE 73rd Vehicular Technology (2016).
Conference (VTC Spring), (2011), pp. 1–5. [209] A. Mestres, et al., Knowledge defined networking (2016) arXiv:1606.06222.v2.
[181] E.J. Khatib, R. Barco, A. Gomez-Andrades, I. Serrano, Diagnosis based on genetic [210] I. Ben Yahia, et al., Cognitive 5G networks: comprehensive operator use cases with
fuzzy algorithms for LTE self-Healing, IEEE Trans. Veh. Technol. (2015). machine learning for management operations, 20th Conference on Innovations in
[182] R.M. Khanafer, B. Solana, J. Triola, R. Barco, L. Moltsen, Z. Altman, P. Lazaro, Clouds, Internet and Networks (ICIN), (2017).
Automated diagnosis for UMTS networks using bayesian network approach, IEEE [211] J. Bendriss, et al., AI for SLA management in programmable networks,
Trans. Veh. Technol. 57 (2008). International Conference on Design of Reliable Communication Networks, (2017).
[183] G.F. Ciocarlie, S.N. Sanneckaczki, Detecting anomalies in cellular networks using [212] European Union Agency for Network and Information Security (enisa), Privacy by
an ensemble method, 9th CNSM and Workshops, (2013). design in big data, An overview of privacy enhancing technologies in the era of big
[184] A. Multazamah, S. Navrati, R. Abhishek, Efficient cell outage detection in 5G het- data analytics (2015).
nets using hidden Markov model, IEEE Commun. Lett. 20 (3) (2016). [213] K.A. Sallehab, L. Janczewski, Technological, organizational and environmental
[185] W. Xue, M. Peng, Y. Ma, H. Zhang, Classification-based approach for cell outage security and privacy issues of big data: a literature review, Procedia Computer
detection in self-healing heterogeneous networks, IEEE Wireless Communications Science, Elsevier, 2016, pp. 19–28.
and Networking Conference (WCNC), (2014), pp. 2822–2826. [214] J. Hu, M.P. Wellman, Nash Q-learning for general-sum stochastic games, J. Mach.
[186] A. Zoha, A. Saeed, A. Imran, M.A. Imran, A. Abu-Dayya, A SON solution for Learn. Res. 4 (2003) 1039–1069.
sleeping cell detection using low-dimensional embedding of mdt measurements,
268

5g1 s2.0 S0140366418300380 Main

Uploaded by

Copyright:

Available Formats

5g1 s2.0 S0140366418300380 Main

Uploaded by

Document Information

Original Description:

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

5g1 s2.0 S0140366418300380 Main

Uploaded by

Copyright:

Available Formats

Computer Communications 129 (2018) 248–268

Contents lists available at ScienceDirect

From 4G to 5G: Self-organized network management meets machine T

1. Introduction work is inspired by a set of requirements deﬁned by the operators’

Fig. 1. Self-organizing networks.

Fig. 2. SON implementations.

Fig. 4. SDN and NFV architectures.

Fig. 5. SON over NFV-SDN architecture.

You might also like