electronics

Article
A Machine Learning Approach for 5G
SINR Prediction
Ruzat Ullah 1 , Safdar Nawaz Khan Marwat 1, * , Arbab Masood Ahmad 1 ,
Salman Ahmed 1 , Abdul Hafeez 2, * , Tariq Kamal 1 and Muhammad Tufail 3
1 Department of Computer Systems Engineering, University of Engineering and Technology Peshawar,
Peshawar 25120, Pakistan; [email protected] (R.U.); [email protected] (A.M.A.);
[email protected] (S.A.); [email protected] (T.K.)
2 Department of Computer Science and Information Technology, Jalozai Campus, University of Engineering
and Technology Peshawar, Nowshera 24240, Pakistan
3 Department of Mechatronics Engineering, University of Engineering and Technology Peshawar,
Peshawar 25100, Pakistan; [email protected]
* Correspondence: [email protected] (S.N.K.M.); [email protected] (A.H.);
Tel.: +92-91-922-2233 (S.N.K.M.)




Received: 4 September 2020; Accepted: 6 October 2020; Published: 12 October 2020


Abstract: Artificial Intelligence (AI) and Machine Learning (ML) are envisaged to play key roles in 5G
networks. Efficient radio resource management is of paramount importance for network operators. With
the advent of newer technologies, infrastructure, and plans, spending significant radio resources on
estimating channel conditions in mobile networks poses a challenge. Automating the process of predicting
channel conditions can efficiently utilize resources. To this point, we propose an ML-based technique,
i.e., an Artificial Neural Network (ANN) for predicting SINR (Signal-to-Interference-and-Noise-Ratio) in
order to mitigate the radio resource usage in mobile networks. Radio resource scheduling is generally
achieved on the basis of estimated channel conditions, i.e., SINR with the help of Sounding Reference
Signals (SRS). The proposed Non-Linear Auto Regressive External/Exogenous (NARX)-based ANN
aims to minimize the rate of sending SRS and achieves an accuracy of R = 0.87. This can lead to vacating
up to 4% of the spectrum, improving bandwidth efficiency and decreasing uplink power consumption.

Keywords: 5G; machine learning; artificial intelligence; sounding reference signals;

Signal-to-Interference-and-Noise-Ratio

1. Introduction
The use of technology in general and portable devices in particular has skyrocketed in the
past two decades. As the technology continues to evolve, various requirements of newly developed
standards have emerged, and these requisites have been fulfilled significantly through research and
innovation. Wireless and cellular communication technologies have played a vital role in the evolution
of technology. Mobile communication marks its birth in the 1970s with the advent of 1G mobile
communication systems, which incorporated circuit switching technologies such as Frequency Division
Multiple Access (FDMA) along with analog communication techniques. Eventually, the 2G standards
evolved with the help of a mechanism where FDMA and Time Division Multiple Access (TDMA) were
combined. The first digital communication system designed was 2G [1]. However, the 3G used Code
Division Multiple Access (CDMA) and High Speed Packet Access (HSPA) technologies to provide
IP- based services in an attempt to meet the demand of higher data rates. In a long Term Evolution
(LTE) also termed as 3.9G and LTE-Advanced (LTE-A) also referred to as 4G, incorporate Orthogonal
Frequency Division Multiple Access (OFDMA) and Single Carrier Frequency Division Multiple Access

Electronics 2020, 9, 1660; doi:10.3390/electronics9101660 www.mdpi.com/journal/electronics

 Electronics
Electronics 2020,
2020, 9, x FOR PEER REVIEW
9, 1660 2 of219
of 19

to as 4G, incorporate Orthogonal Frequency Division Multiple Access (OFDMA) and Single Carrier
(SC-FDMA)
Frequencyin order toMultiple
Division achieveAccess
much higher rates in
(SC-FDMA) compared
order to to 3G. Figure
achieve much 1higher
showsrates
the evolution
compared from
to
1G3G.
toElectronics
5G in terms
Figure 9,of
1 shows
2020, services
thePEER
x FOR offered.
evolution
REVIEW from 1G to 5G in terms of services offered. 2 of 19

to as 4G, incorporate Orthogonal Frequency Division Multiple Access (OFDMA) and Single Carrier
Frequency Division Multiple Access (SC-FDMA) in order to achieve much higher rates compared to
3G. Figure 1 shows the evolution from 1G to 5G in terms of services 4G
offered. 5G
1G 3G • HD video
2G 2G • video
• analog • internet • self driving
• ccc
digital sms • digital sms • gaming
• calls • roaming • VR
• Mobile TV
• D2D
5G
4G
1G 3G • HD video
2G 2G • video
• analog • internet • self driving
• ccc
digital sms • digital sms • gaming
• calls Figure 1. • roaming
Evolution from 1G to 5G. • VR
• Mobile TV
Figure 1. Evolution from 1G to 5G. • D2D
The technological innovation is prospering at an unprecedented rate in the form of Internet-of-
The (IoT),
Things technological innovationis is
where connectivity prospering
required at an and
ubiquitously unprecedented rate in from
pervasively. Ranging the form
fan to of
Internet-of-Things (IoT),all where Figure 1. Evolution
connectivity from 1G to 5G.
vehicle, CEO to peon, are required to be isinrequired
the loop ubiquitously
and stay connected and pervasively. Ranging from
to remain up-to-date in
fanthe
to vehicle,
ever CEO
Thechanging to
technological peon,
global all are required
technological
innovation to be
scenario.
is prospering in the
at anSmart loop and stay
homes, cities,
unprecedented connected
rate in cars to remain
andofindustries
the form up-to-date
Internet-of-are
theThings
ever (IoT),
in envisaged changingwhereglobal
to further connectivity
soar and is required
technological
toughen ubiquitously
thescenario.
spectrum, Smart and
speed, pervasively.
homes,
QoS cities, Ranging
cars andfrom
(Quality-of-Service) fan to
industries
and QoEare
vehicle,toCEO
envisaged to peon,
(Quality-of-Experience)
further soarall are required
requirements
and toughen tothe
for be in the loop
networks.
spectrum, Asandper stay
speed, connected
Gartner
QoS stats to
of remain up-to-date
2016, 20.4
(Quality-of-Service) billion
andinIoT
QoE
the
devices ever
are changing
envisaged global
to be technological
connected by scenario.
2020 [2]. Smart
The total homes,
number cities,
of
(Quality-of-Experience) requirements for networks. As per Gartner stats of 2016, 20.4 billion IoT devices cars
connected and industries
devices are
including
envisaged to further soar and will
toughen the spectrum, 50speed, QoSby (Quality-of-Service) andscientific
QoE
areIoT devices to
envisaged andbe mobile
connected devices
by 2020 reach
[2]. Themore
total than
number billion
of connected 2021devices
[3]. IoTincluding
devices, IoT devices
(Quality-of-Experience)
research requirements for networks. As per Gartner stats of 2016, 20.4 billion IoT
and mobile and
devicessocial
willmedia
reach have
more alsothanresulted
50 billioninbythe 2021proliferation of internet
[3]. IoT devices, traffic
scientific leading
research andto social
an
devices
estimate are
of 2.5envisaged
exabytes/day to be connected
and 40 by 2020
petabytes/year[2]. Theby total number
CERN’s of
Large connected
Hadron devices
Collider including
(LHC) [4].
mediaIoT have also resulted
devices and mobile in the proliferation
devices will reachofmoreinternet
thantraffic leading
50 billion to an[3].
by 2021 estimate of 2.5 scientific
IoT devices, exabytes/day
andConsidering
40 internet by streaming data, approximately 400 h’/min videos are uploadedinternet
to YouTube and
research and social media have also resulted in the proliferation of internet traffic leading streaming
petabytes/year CERN’s Large Hadron Collider (LHC) [4]. Considering to an
2.5 million photos and videos are posted using the platform of Instagram. Consequently, a total of
data,estimate
approximately 400 h’/min and
of 2.5 exabytes/day videos are uploadedby
40 petabytes/year to CERN’s
YouTube and Hadron
Large 2.5 million photos
Collider (LHC)and[4].
videos
2.5 quintillion bytes of data are uploaded every day [5]. To cope with all these challenges, the research
are posted usinginternet
Considering the platform
streaming of data,
Instagram. Consequently,
approximately 400 h’/min avideos
total are
of uploaded
2.5 quintillion
to YouTubebytesand of data
community is gradually heading towards the next generation 5G network.
2.5 millionevery
are uploaded photos dayand[5].videos
To copeare posted
with all using thechallenges,
these platform of the Instagram.
research Consequently,
communitya istotal of
gradually
2.5Figure 2 shows
quintillion bytes the
of journey
data are towards 5G
uploaded everydeployment
day [5]. To ever
cope since
with LTE
all waschallenges,
these introduced. the Substantial
research
heading towards the next generation 5G network.
research is going
community on in the area
is gradually of mobile communication, which 5Gisnetwork.
evident from the number of papers
Figure 2 shows the heading
journeytowards towards the next
5G generation
deployment ever since LTE was introduced.
being published in this area. From January 2014
Figure 2 shows the journey towards 5G deployment ever since LTE to January 2018, a total
was of 389 papers
introduced. have been
Substantial
Substantial
published research
in IEEE is going
Xplore, on
whilein the
588 area
in of mobile
ScienceDirect communication,
[6]. which is evident from the number
research is going on in the area of mobile communication, which is evident from the number of papers
of papers being published in this area. From January 2014 to January 2018, a total of 389 papers have
being published in this area. From January 2014 to January 2018, a total of 389 papers have been
been published
publishedininIEEE IEEE Xplore,
Xplore, while
while 588 588 in ScienceDirect
in ScienceDirect [6]. [6].

Figure 2. Plan execution timeline [7].

Figure2.2.Plan
Figure Plan execution
execution timeline
timeline[7].
[7].
Electronics 2020, 9, 1660 3 of 19

The 5G is still undergoing standardization and New Radio (NR) was standardized in 3GPP
Release 15 [8]. Its requirements include up to 10 Gbps speed and different operating bands (from below
1 GHz up to 100 GHz) of frequency [1,9]. The 5G will inherit several features of legacy 4G system like
sub-carrier spacing (with few additional options), Carrier Aggregation (CA) and most of the Reference
Signals (RS) [10].
Nevertheless, it is anticipated that 5G will augment some novel and versatile dynamics to networks
such as Big Data (BD), Artificial Intelligence (AI) and Machine Learning (ML). AI and ML have evolved
as irrefutably important disciplines in the last decade. In the case of 5G, the quality of a network of an
operator can be assessed by the quality of deployment of AI techniques [11]. It is making appliances
automatic and hence decreasing human intervention, which is also a demand of 5G [12]. In a nutshell,
it can be inferred that introduction of AI and its sub-categories in 5G is inevitable [13].
“Spectrum is our invisible infrastructure; it’s the oxygen that sustains our mobile communication”—stated
by US Federal Communications Commission Chairman, Julius Genachoeski. It is obvious that spectrum
is one of the most precious resources in mobile communication. Therefore, handling spectrum with due
attendance is indispensable [14]. In cellular communication, a considerable amount of the spectrum
is used for the reference signals. The part of the spectrum that is consumed by the reference signals
aggregates up to 25%. Availability of additional spectrum helps in smoothening of the communication.
Transmission of these signals adds up overhead as well as consumes a considerable amount of uplink
power. In this research, an attempt has been made to save a fraction of the above-mentioned spectrum
i.e., 25% mainly utilized by the reference signals. This will help in mitigating the motioned problems.
The recent research advances in Mobile Networks (MN) demonstrates several attempts for saving
the spectrum resources as well as power constraints. Some of the researchers are focusing on ML
and AI techniques to achieve the said goal. The individual proposed solutions in the literature are
orchestrated to address a particular/single problem at a time. As discussed in the literature survey
section in detail, the proposed solutions are not multifarious and do not present a single solution for
the stated problems. Hence, an efficient technique is required, which addresses the spectrum as well as
power usage problems simultaneously.
In this paper, we present a novel method for predicting Signal-to-Interference-and-Noise-Ratio
(SINR) based on the location of a Cyber Physical System (CPS). An Artificial Neural Network
(ANN)-based model predicts the channel condition, i.e., SINR on the basis of current location of CPS.
The Base Station (BS) sends the location in the form of different signals including Position Reference
Signal (PRS) and Observed Time Difference of Arrival (OTDoA). In terms of ML, the model has shown
good enough (R = 0.87 and MSE = 12.88) accuracy. Furthermore, employing such an approach into the
cellular network has contributed in:

1. Increased throughput of up to 1.4 Mbps in the case of 16 QAM (Quadrature Amplitude Modulation)
and 2.1 Mbps in the case of 64 QAM.
2. Saved power of 6.2 dBm at the CPS end.
3. Increased bandwidth (BW) efficiency i.e., approximately equal to 4% by saving the fraction of
spectrum that is used for the frequent transmission of Sounding Reference Signals (SRS).

The rest of the paper is organized as follows: Section 2 covers the background, Section 3 discusses
the literature survey on the use of AI in wireless and cellular networks, and Section 4 is about data
collection. In Section 5, simulation setup for data acquisition has been discussed, Section 6 discusses
training and testing process for ANN, Section 7 demonstrates the way ANN-based model will be
incorporated into MN, Section 8 shows results of the ML model, Section 9 discusses contributions of
the proposed scheme and Section 10 provides a conclusion of the discussion.
Electronics 2020, 9, 1660 4 of 19

2. Background
As already discussed, this paper combines two diverse fields of computers, i.e., MN and ML.
The framework is conceptually divided into two sections for better understanding of the proposed
scheme: (a) Network module (b) ML module.

2.1. Network Module

Mobile communication systems have made a lot of things easier for people, but then, on the other
hand, they have posed some tough research questions as well. QoS provision and efficient spectrum
utilization are extremely important in mobile communication systems. QoS provision usually relies on
proper resource allocation. The spectrum is generally allocated on the basis of channel conditions of
CPS users. The channel conditions depend on SINR of the CPS.
For resource scheduling in LTE, an uplink SRS (Sounding Reference Signal) is sent, which intends
to estimate channel conditions at different frequencies. 5G will also follow the same pattern except
that its SRS will be one, two or four symbols long in the time domain [15]. These SRS are generated
on the basis of the Zadoff–Chu sequence [16]. An SRS when received at BS is used to estimate SINR
using different state-of-the-art techniques, one such novel method is discussed in [17]. Here is how
SRS sequence works mathematically:

rαu,v = ejαn ru,v (n), 0 ≤ n ≤ MRS

SC − 1 (1)

In the above equation, MRS SC

represents length of the reference signal sequence and ru,v (n) is
generated using the Zadoff-Chu sequence and has Constant Amplitude Zero Autocorrelation (CAZAC)
characteristics and hence allows orthogonal code multiplexing by cyclic shift α, where α is demonstrated
as follows,
nCS
α = 2π SRS (n = 0, 1, 2, 3, 4, 5, 6, 7), (2)
8
Here, n shows the number of SC-FDMA symbols used in uplink. The two indices CS and SRS
refer to Cyclic Shift and Sounding Reference Signal, respectively.
Radio resource scheduling is then done on the basis of this calculated SINR. 5G will also be using
the same configurations and features as that of LTE SRS [16] and 5G is also envisaged to give a further
role of MIMO configuration to SRS [18]. SRS is sent in gaps of milliseconds. It can be sent as frequently
as after 2 ms and as seldom as after 160 ms [16]. Figure 3 shows resource grid of LTE-A. LTE-A and 5G
have the same numerology except that the subcarrier in 5G can be up to 30, 60 and 120 KHz and 5G
can also have up to 275 PRBs (Physical Resource Blocks) in its resource grid. Different terminologies
are explained in Figure 3. Figure 4 shows SRS configuration on a resource grid. It is evident from
the aforementioned figures that SRS uses large bandwidth and occupies almost 4.1% of the spectrum.
However, in some cases like high Path Loss (PL) and fast fading, SRS may have poor performance on
consecutive frequency resources [16]. Frequency hopping is used to overcome this problem.
A Base Station (BS) also acquires a CPS location information using Observed Time Difference of
Arrival (OTDOA), Positioning Reference Signal (PRS) or Global Positioning System (GPS) [20]. For the
purpose of simulating our network and acquiring required parameters, a Vienna Simulator [21] has
been used, while for the purpose of mobility of CPSs, a Random Waypoint model has been used [22].
Electronics 2020, 9, 1660 5 of 19
Electronics 2020, 9, x FOR PEER REVIEW 5 of 19

Electronics 2020, 9, x FOR PEER REVIEW 6 of 19

Figure3.3.Resource
Figure grid[19].
Resource grid [19].

A Base Station (BS) also acquires a CPS location information using Observed Time Difference of
Arrival (OTDOA), Positioning Reference Signal (PRS) or Global Positioning System (GPS) [20]. For
the purpose of simulating our network and acquiring required parameters, a Vienna Simulator [21]
has been used, while for the purpose of mobility of CPSs, a Random Waypoint model has been used
[22].

Figure
Figure 4. Sounding
4. Sounding ReferenceSignals
Reference Signals (SRS)
(SRS)configuration
configurationonon
a subframe [16].[16].
a subframe

2.2. ML
2.2.Module
ML Module
As already discussed,
As already discussed,ML
MLhas
has revolutionized several
revolutionized several research
research areas.
areas. It mayIthardly
may have
hardly
lefthave
any left
area of
any area of the
thedigital
digitalworld
worlduninfluenced.
uninfluenced. Recently, the research
Recently, area of area
the research ML isofproving
ML is to be of great
proving to be of
significance in numerous fields. It is achieving popularity in fields like image processing,
great significance in numerous fields. It is achieving popularity in fields like image processing, medical
diagnostics, self-driving cars, etc. With the advent of ML, devices or things are becoming automated,
requiring less labor and hence less manual control is needed. In some cases, ML is also proving to be
more efficient and reliable than statistical methods [13].
Several techniques with a lot of innovations and revamps have so far been employed for
accomplishing AI and ML tasks. However, there is a generic categorization of ML algorithms:
Electronics 2020, 9, 1660 6 of 19

medical diagnostics, self-driving cars, etc. With the advent of ML, devices or things are becoming
automated, requiring less labor and hence less manual control is needed. In some cases, ML is also
proving to be more efficient and reliable than statistical methods [13].
Several techniques with a lot of innovations and revamps have so far been employed for
accomplishing AI and ML tasks. However, there is a generic categorization of ML algorithms:

2.2.1. Supervised Learning

It is the category of ML algorithms that works on such data that are pre-labeled. This means
that there are pre-defined inputs, we call features, and pre-defined output(s) for them called label(s).
Cases where data are continuous and do not belong to a few classes, falls under regression. When the
output label is/are discrete class (es) then it is classification.

2.2.2. Unsupervised Learning

These types of algorithms use unlabeled data. These models work in a way that they first learn
features from the fed data and then the validation step is done, and after that, data are tested for the
already learned features.

2.2.3. Semi-Supervised Learning

It is a kind of hybrid algorithm of supervised and unsupervised algorithms where the data are a
mix of labeled and unlabeled data.

2.2.4. Reinforcement Learning

Electronics 2020, 9, x FOR PEER REVIEW 7 of 19
In reinforcement, self-learning behavior of humans and animals is co-opted. It works on the
principle of
In“Action and Punishment”.
reinforcement, self-learning behavior of humans and animals is co-opted. It works on the
principle of “Action and Punishment”.
2.2.5. Artificial Neural Networks
2.2.5. Artificial Neural Networks
Artificial Neural Networks (ANNs) work on the principle of Human Neural Networks (HNNs).
They haveArtificial Neural Networks
certain inputs, multiplied (ANNs) work on and
with weights the principle of Humanan
passed through Neural Networks
activation (HNNs).
function for one
They have certain inputs, multiplied with weights and passed through an activation function for one
or more hidden layers and hence reach the output layer. In the O/P layer, results are compared to
or more hidden layers and hence reach the output layer. In the O/P layer, results are compared to
targets and error is rectified through back propagation for defined epochs and in the last accuracy with
targets and error is rectified through back propagation for defined epochs and in the last accuracy
best adjusted
with bestweight is weight
adjusted calculated. A simple
is calculated. ANN is
A simple shown
ANN in Figure
is shown 5. 5.
in Figure

Figure
Figure 5. simple
5. A A simple ArtificialNeural
Artificial Neural Networks
Networks (ANN).
(ANN).

Commonly used algorithms in wireless networks are Deep Learning, Ensemble Learning and K-
Nearest Neighbors [13,23,24].
An MIT Review has traced the trails of research papers published at arXive.org since 1993. Their
findings show the fact that the number of research papers published at arXive.org in this decade has
increased more than two-fold as vis-à-vis the last decade [25].
Electronics 2020, 9, 1660 7 of 19

Commonly used algorithms in wireless networks are Deep Learning, Ensemble Learning and
K-Nearest Neighbors [13,23,24].
An MIT Review has traced the trails of research papers published at arXive.org since 1993.
Their findings show the fact that the number of research papers published at arXive.org in this decade
has increased more than two-fold as vis-à-vis the last decade [25].
However, neural networks have gained a lot more popularity than other ML algorithms. An MIT
review has analyzed research articles of arXive since 1993 of how ANNs have gained popularity
especially in the last half of the current decade [25].
For this research work, ANN with a Levenberg–Marquardt algorithm (LBM) has been taken
into consideration. An ANN with LBM was selected on the basis of non-linearity of data as ANN
and LBM processes and optimizes non-linear data. LBM is the training/ optimization algorithm for
ANN to minimize the error. ANN has an input layer that takes the input, multiplies weights to the
input, adds bias, and processes it through neurons. The result(s) acts as input to the next layer of
neurons. The output layer retracts the values back again for optimization. Here is how it happens
mathematically:
Xn
z= ((xi )(wi )) + b (3)
i=1

where, xi is input, wi is weight and b is bias.

z is then passed through a squashing/activation function, in our case, for hidden layers, we have
used the Sigmoid function also known as logistic function. Mathematically it is shown as follows:

1
y = (4)
1 + e−z

For the output layer, a linear function has been used. Mathematically, it is shown as follows:

y = cx, (5)

In the above equation, x is the input from the preceding layer and c is the constant.
Optimization is followed with LBM as follows:
 
Jt J + λ I δ = Jt E, (6)

where J is the Jacobian matrix, λ is Levenberg’s damping factor, δ is the weight update while E is the
error matrix.

3. Related Work
Significant research has been carried out so far in wireless networks integration with ML methods
and a lot of work is still going on. Here, a few selected pieces of work on ML integration with wireless
networks that have been discussed in the research community are revisited.
K. Saija et al. [26] have investigated the delay DL-only (Downlink only) CSI-RS (Channel State
Information- Reference Signal). The authors have tried to mitigate the long delay of CSI feedback.
A BS transmits CSI-RS in the DL and then the UE sends feedback of this CSI-RS in the uplink to
the BS. Authors have considered three ML approaches for building their research, Batch Learning,
Online Learning, and Deep Learning. The features they have used are time and SNR while the target
output they have considered is CQI (Channel Quality Indicator) and for a part of the experiment they
also have added SNR as a target. The ML performance measure they have considered is accuracy
improvement of SNR prediction and the network performance has been measured in terms of error
mitigation of SNR calculation.
F. D. Calabrese et al. [27] have discussed the problems and opportunities in ML based
Radio Resource Management (RRM). The authors have considered two different ML approaches,
Electronics 2020, 9, 1660 8 of 19

namely Reinforcement Learning (RL) and Artificial Neural Networks (ANN). In this research,
the authors have proposed an architecture for training and learning of ML models for 5G networks.
The authors have focused on different RRM parameter predictions like power control, cell and BW, etc.
M. Chen et al. [23] have investigated the issue of spectrum management and multiple access
technologies. The authors have adopted an ANN model for resource management. The network
condition has been used as input to the model frequency bands as target. This model also has the
capability of switching among different frequency bands, which will be a key feature of 5G networks.
The task of switching among different frequency bands is based on the availability of Line of Sight
(LoS) and congestion avoidance.
D. Kumar et al. [28] have dealt with the upper layers functionalities of an OSI (Open System
International) model. The authors have proposed an ML model at the top of HTTP (Hyper Text
Transfer Protocol) that would ensure fast and smooth transmission of video data. The authors have
proposed an unsupervised (clustering) model. With the said model they were able to achieve up to
7% improvement in peak signal-to-noise-ratio and up to 25% improvement in video quality metrics.
The model performance was validated in a live video streaming session.
P. Casas [24] in his research has addressed a couple of smart phone traffic specific problems.
The first problem the author has addressed is detecting anomalies generated by smartphones as a
result of a large number of data traffic being generated by apps. The other problem addressed is the
prediction of Quality-of-Experience (QoE) for the most used apps. An Ensemble Learning (EL) model
has been preferred where six different algorithms have been ensembled. The ensembled algorithms
are decision trees, Naïve Bayes, Multi Layers Perceptron (MLP), Support Vector Machines (SVM),
Random Forests (RF) and K Nearest Neighbors (K-NN). In this research, the authors have achieved an
accuracy of up to 99% using EL stacking.
Some other issues addressed by researchers in this field are as follows:

• Handover prediction [29];

• Link status prediction [29];
• Intrusion detection system [29];
• Modulation classification [30];
• Network optimization using different techniques [30];
• QoE enhancement for 5G [31];
• AI as Micro Service (AIMS) for data-driven intelligent traffic over 5G networks [32];
• Cognitive autonomous 5G networks [33];
• Wireless network monitoring and analysis [24];
• Fault Diagnosis for autonomous vehicles using SVM [34].

It can safely be pronounced that several problems/issues regarding ML integration with wireless
networks have been addressed but still there is a huge gap to be addressed. ML is anticipated to play a
key role in 5G and later systems. Hence, there is a dire need of this area to be addressed. Here are some
opportunities, for mobile and wireless networks in general and 5G in specific, described in literature:

• Accurate channel state information estimation [35];

• Network optimization modeling [35];
• Approaches to select best (less complex) machine learning algorithms for wireless networks [13];
• Network resource allocation is attributed to be one of the key issues in 5G [35];
• Spatio-temporal mobile data mining [13].

There are a handful of research gaps in 5G for AI and ML, albeit, there are many more opportunities
being discussed in the literature as in [13,23,29]. Table 1 gives an insight of AI integration into MN and
why they could not solve the problem.
Electronics 2020, 9, 1660 9 of 19

Table 1. A glimpse of the literature survey.

Power Throughput
Reference Problem(s) Addressed BW Efficiency AI Integration
Saving Increase
[36] Power saving at user end Yes No No Yes
[37] Saving Power at BS Yes No No Yes
Increased capacity with
[38] Yes No Yes No
optimized power
[39] Resource allocation for NOMA Yes No Yes No
[40] Power optimization Yes No No No

4. Data Acquisition
As a rule of thumb for ML algorithms, they need data for training, which includes features and
labels (in case of supervised learning). In our case, features are geographical positions of the user, i.e.,
X and Y coordinates and the labels are the SINR values. To acquire SINR, MATLAB based traces from a
Vienna simulator are obtained and a Random Waypoint mobility model is simulated for 10,000 s to get
the mobility patterns, returning almost 200,000 instances of SINR against X, Y coordinates. These traces
were acquired for a single node. The speed interval of the node was set to 0.02 m/s and the pause
interval was 0 s (no pause). The time step for trace collection was 0.001 s. Figure 6 shows the simulation
process of how data are collected. Where a CPS sends SRS and position signal to a BS and BS estimates
SINR from2020,
Electronics that,
9, xand
FORthen
PEERallocates
REVIEW the scheduling scheme accordingly. 10 of 19

Figure 6. Data acquisition.

Figure 6. Data acquisition.
5. Simulation Setup
5. Simulation Setup
The footprints of channel conditions were obtained using a Vienna simulator. A MATLAB-based
systemThe footprints
level of channel
simulator conditions
was developed were obtained
for simulating using anetworks.
advanced Vienna simulator.
In order A toMATLAB-based
obtain the link
quality (i.e., SINR), the model uses three parameters. The first parameter is the Macro obtain
system level simulator was developed for simulating advanced networks. In order to the (L),
path-loss link
quality
which (i.e., SINR),
models both Lthe
duemodel uses three
to distance andparameters. TheDistance
antenna gain. first parameter
based Lis is
thecalculated
Macro path-loss
using the(L),
which models
following both L due to distance and antenna gain. Distance based L is calculated using the
formula.
following formula. L = 128.1 + 37.6 log10 R [Km] (7)
L = 128.1 + 37.6 log10 R[Km] (7)
In the above equation, R [Km represents the distance between transmitter and receiver in
kilometers, whereas 128.1 is a constant and 37.6 is the constant coefficient of the second term in the
equation.
The second parameter is the shadow fading, which estimates the loss of the signal due to
obstacles in the way of signal traveling from transmitter to receiver. It is referred to as the
introduction change in geographical properties with respect to the Macro path-loss. The third
Electronics 2020, 9, 1660 10 of 19

In the above equation, R [Km] represents the distance between transmitter and receiver in
kilometers, whereas 128.1 is a constant and 37.6 is the constant coefficient of the second term in
the equation.
The second parameter is the shadow fading, which estimates the loss of the signal due to obstacles
in the way of signal traveling from transmitter to receiver. It is referred to as the introduction change in
geographical properties with respect to the Macro path-loss. The third parameter is channel modeling.
In this part, J. C. Ikuno et al. [41] have considered small scale fading as a time varying function.
Channels have been modeled for both multiple-input and multiple-output (MIMO) and single-input
and single-output (SISO) modes of transmission. Finally, the link quality traces are collected from the
Vienna simulator. Different scenarios have different SINR estimation methods. Our SINR of interest
uses SISO as the uplink. Systems such as LTE have only SISO support, but the model could be extended
to MIMO as well. SISO-based SINR is estimated using the following formula.

Ptx
SINR = (8)
PNint|h1 |2
1
σ2 + P
i=1 |h0 |2 tx,i
|h0 |2

1 |h1 |2
where Ptx is the transmission power, 2, are noise and interference traces, Ptx,i shows transmission
|h0 | |h0 |2
power at the ith transmitter.

6. Training and Testing

As it has already been discussed, the acquired dataset has two input features, i.e., the coordinates
Electronics 2020, 9, x FOR PEER REVIEW 11 of 19
of the user location and the SINR. So, in light of the available data, our problem lies in the
scope of supervised
supervised learning. As learning. As label
the output the output labelofiscontinuous
is a vector a vector of continuous
values of SINR, values of SINR,
the regression
the regression technique has been used for curve fitting. The MATLAB
technique has been used for curve fitting. The MATLAB environment is used for the experiment. environment is usedThe for
the experiment.
ANN model that Thehas
ANN model
been used thatfor
hasthisbeenexperiment
used for this is experiment
a Non-Linear is a Non-Linear Auto
Auto Regressive
Regressive External/Exogenous (NARX) model. The NARX model has
External/Exogenous (NARX) model. The NARX model has two architectures—one is series-parallel two architectures—one is
series-parallel architecture, which is also called an open-loop and the other is a parallel
architecture, which is also called an open-loop and the other is a parallel architecture also referred to architecture
also
as referred
closed loop.toWe
as closed
preferredloop.
theWe
openpreferred the open loop
loop architecture architecture
for our simulationsfor mainly
our simulations
because themainly
true
because the true output is provided to the network, unlike a closed loop that feeds
output is provided to the network, unlike a closed loop that feeds an estimated output. Another an estimated output.
Another advantage
advantage of using ofanusing
openanloop
openis loop
that isit that
usesitauses a simple
simple feed feed forward
forward network,
network, which
which allows
allows to
to incorporate efficient training algorithms [42]. The two different architectures of
incorporate efficient training algorithms [42]. The two different architectures of NARX are depicted NARX are depicted
in Figure
in Figure 7.
7.

Non-LinearAuto
Figure7.7.Non-Linear
Figure AutoRegressive
RegressiveExternal/Exogenous
External/Exogenous (NARX)
(NARX) [42].
[42].

Mathematically, open loop NARX is shown by the following formula:

y(t) = f(y(t − 1), y(t − 2),…,y(t − ny), x(t − 1), x(t − 2),…, x(t − nx)), (9)
In the above equation, y(t) is the predicted value based on the previous value of the input vector
u(t − 1) as well as that of the output vector y(t − 1). F(.) is the mapping function, ny represents
output delays, while nx shows input delays.
Results show that the best performance/accuracy (𝑅 = 0.87112) is achieved when the training
Electronics 2020, 9, 1660 11 of 19

Mathematically, open loop NARX is shown by the following formula:

   
ŷ(t) = f y(t − 1), y(t − 2), . . . , y t − n y , x(t − 1), x(t − 2), . . . , x(t − nx ) , (9)

In the above equation, ŷ(t) is the predicted value based on the previous value of the input vector
u(t − 1) as well as that of the output vector y(t − 1). F( .) is the mapping function, n y represents output
delays, while nx shows input delays.
Results show that the best performance/accuracy (R = 0.87112) is achieved when the training data
are spread across the time domain. The spreading of data in the time domain means that the input
data are not fed at once along with the delays.
The dataset is split into three sets including training, validation and testing in the proportion
as shown by Table 2. The configuration of the proposed neural network comprises 15 hidden layers
with 5 delays and 20 epochs. The results show that the training process stops if the generalization, i.e.,
cross validation does not improve beyond six iterations.

Table 2. Distribution of the dataset.

Dataset Proportion
Training 70%
Validation 15%
Testing 15%

7. Working
Initially, it is not desired to totally exterminate the SRS; nonetheless, the goal is to reduce
its transmission rate by half. As already discussed, SRS is normally transmitted after every 2 ms,
but because of this experiment it is transmitted after 4 ms and for the rest of the 2 ms, BS predicts SRS
based values of SINR with the trained ML algorithm.
Normally, a regular communication of SRS and geo-position signals between the user and BS
exists. In this case, it is proposed to reduce the rate of the communication of SRS to half. In the deployed
systems, SRS is transmitted almost after every 2 ms and it is proposed to send SRS after 4 ms in order
to save two sub frames, which are especially configured for SRS. The CPS regularly communicates its
position to the BS in the form of PRS.
In the 2 ms of the middle, where no SRS is being communicated, ML helps in predicting the
channel conditions. In this case, only the position signal is to be configured for transmission on the basis
of geo-location of the CPS. Since the machine has been trained for predicting SINR, BS schedules DL
resources for the CPS without SRS being transmitted. Figure 8 shows this concept through a flowchart.
It is inferable from Figure 8 that first the CPS will be configured in such a manner that it will
decide whether to use ML or the classical statistical method. If it is ML’s turn, then the CPS sends only
its position signal rather than sending SRS and at the BS the position coordinates are provided to the
ML model in order to predict the SINR. DL scheduling is decided on the basis of the calculated SINR.
Alternatively, if it is not ML’s turn, then the CPS will send SRS as well as position signal to the BS.
The BS will estimate SINR on the basis of the received SRS (already discussed in Section 2). Therefore,
the scheduling is achieved based on this estimated SINR.
 It is inferable from Figure 8 that first the CPS will be configured in such a manner that it will
decide whether to use ML or the classical statistical method. If it is ML’s turn, then the CPS sends
only its position signal rather than sending SRS and at the BS the position coordinates are provided
to the ML model in order to predict the SINR. DL scheduling is decided on the basis of the calculated
SINR.
Electronics 2020, 9, 1660 12 of 19

Figure 8. Flow chart for implementation of the proposed scheme.

Figure 8. Flow chart for implementation of the proposed scheme.
For the decision of when it will be the turn of the ML model calling and the classical method
Alternatively,
calling, we proposed if the
it isfollowing
not ML’s Algorithm
turn, then the
1. CPS will send SRS as well as position signal to the
BS. The BS will estimate SINR on the basis of the received SRS (already discussed in Section 2).
Therefore, the scheduling is achieved
Algorithmbased on thisdecision
1: Method estimated SINR.
For the decision of when it will be the turn of the ML model calling and the classical method
start
calling, we proposed the following for Algorithm 1.
t equals 0 to_max with t + 2ms
{
CPS.Classical_Method
}
for t equals 1 to_max with t + 2ms
{
CPS.ML_Method
}
end

The BS configures a CPS on the basis of the above algorithm. For the time slot (t = 0), the classical
method is called where both SRS as well as position signals are transmitted. For the next slot, the ML
method is invoked and only a position signal is being sent and an ML model is applied.
Here, only the time domain is taken into account due to the fact that the already proposed/allocated
bandwidth for SRS is not being decreased but in fact its transmission in the time domain is intended to
be decreased. SRS will still stretch over a large bandwidth but not as frequent in time as it used to
be previously.

8. Results
After completing the training, validation and testing task, different matrices of performance
measurement are acquired. However, the parameter considered here is the loss function and accuracy
measure obtained at the end of the simulation. The loss function preferred for our experiment is Mean
Electronics 2020, 9, 1660 13 of 19

Squared Error (MSE) and the accuracy measure is Regression (R). The MSE achieves a value of 12.88
and the accuracy measure of the experiment is 0.87, as shown in Figure 9. Mathematically, MSE works
as follows:
n
1X 2
MSE = ( yi − e
yi ) (10)
n
i=1

Here, y shows target/real output, while e

Electronicsi 2020, 9, x FOR PEER REVIEW
yi is the predicted/ estimated output. 14 of 19

Figure
Figure 9.
9. Machine
Machine Learning
Learning (ML)
(ML) model
model accuracy.
accuracy.

Table
The 3 shows a comparison
mathematical formula of of
R isaccuracy of NARX with a Non-Linear Input Output technique
as follows.
and SVM with 10-fold cross validation under tR
v accuracy measure and for the same data. It can be
Pn 2
inferred that NARX outperforms the latter. i=1 (eyi − y)
R= Pn 2
(11)
i=1 ( yi − y )
Table 3. Comparison of accuracy of NARX with Non-Linear Input Output and Support Vector
Machines
The (SVM).equation shows mean value of e
y in above y. i
There are four different results in Figure 9 obtained at different
Model Accuracy stages of the experiment. Overall,
this figure shows the comparison of the real
NARX output (target) 0.87 the predicted values (output).
with
The training results are those thatNon-Linear
have been obtained during the0.79
Input Output training phase. The model achieves an
SVM 0.47
Equation (12) shows how the Non-Linear Input Output technique works mathematically.

y(t) = f(x(t − 1),…,x(t − d)) (12)

Electronics 2020, 9, 1660 14 of 19

accuracy of 0.87 for training. The validation results are obtained during the validation step (ensuring
that there is no overfitting) and yielded an accuracy of 0.86. The testing results are those obtained
after feeding the test part of the dataset. Accuracy of the test phase is 0.86. The last result shows a
generic model obtained after running all of the above steps. Output = 0.77 × Target + 1.7 is the final
model obtained, it shows that the input ( f ( y, x)) is multiplied with the coefficient 0.77 (weight) and
then added with 1.7 (bias). The fit and Y = T lines show the convergence of the real output and target
and the data legend shows the distribution of data instances against the fitness function.
Table 3 shows a comparison of accuracy of NARX with a Non-Linear Input Output technique and
SVM with 10-fold cross validation under R accuracy measure and for the same data. It can be inferred
that NARX outperforms the latter.

Table 3. Comparison of accuracy of NARX with Non-Linear Input Output and Support Vector
Machines (SVM).

Model Accuracy
NARX 0.87
Non-Linear Input Output 0.79
SVM 0.47

Equation (12) shows how the Non-Linear Input Output technique works mathematically.

ŷ(t) = f (x(t − 1), . . . , x(t − d)) (12)

Table 4 shows performance parameters at a glance. The RMSE (Root MSE) is acquired by taking
the square root of MSE. The r2 shows the R-squared value that has been acquired by taking the square
of the regression value R.

Table 4. Performance measures.

Parameter Value Percentage

R 0.87 87%
MSE 12.88 N/A
r2 0.75 75%
RMSE 3.58 N/A

9. Contribution
The increase in the amount of maximum throughput/good-put capacity for 16 QAM is depicted
in Figure 10. Payload throughput is obtained after the exclusion of reference signals overhead,
reference signals take up to 25% of the total capacity of a channel BW. Here, 10 MHz BW may not seem
better for some higher frequencies in 5G because it will not satisfy the higher data rate requirements;
however, it is still considered by standardizing bodies and operators for Non-Standalone (NSA). NSA is
the initial phase of 5G, where it will work as an extension of 4G and will use LTE packet core and
lower frequencies in Standalone (SA). SA will be the full fledge 5G system requiring no LTE or LTE-A
support [43,44]. For this reason, this paper also considers 10 MHz BW. Besides that, 15 and 20 MHz
BW are stronger candidates for uplink in 5G, and therefore considered in this work [45]. 256 QAM and
MIMO have not been considered because they are not supported at the CPS.
 Standalone (NSA). NSA is the initial phase of 5G, where it will work as an extension of 4G and will
use LTE packet core and lower frequencies in Standalone (SA). SA will be the full fledge 5G system
requiring no LTE or LTE-A support [43,44]. For this reason, this paper also considers 10 MHz BW.
Besides that, 15 and 20 MHz BW are stronger candidates for uplink in 5G, and therefore considered
in this work
Electronics 2020, 9,[45].
1660 256QAM and MIMO have not been considered because they are not supported at19
15 of
the CPS.

Throughput Comparison
60
51 52.43
50

37.9 39.24
Throughput (Mbps)

40

30 25.5 26.8
Old Throughput
20 Increased Throughput

10

Electronics 2020, 9, x FOR PEER REVIEW 16 of 19

0
10 15 20
Here, 𝐶 is the capacity in bits/sec of a channel for a given BW while 𝐶 is the reference signal
overhead capacity in bits/sec and 𝑆 is the
BWthroughput
(MHz) in bits/sec.

𝑆 = 𝑆 + 𝐶(Quadrature
/2 (14)
Figure10.
Figure 10.Throughput
Throughput comparison
comparison for 16 QAM
QAM (QuadratureAmplitude
AmplitudeModulation).
Modulation).
𝑆 is the enhanced throughput. It has been obtained by adding 𝑆 and 𝐶 /2. 𝐶 /2 is
the The
Theimpact
impact
channel of the
of proposed
capacity the mechanism
proposed
dedicated SRS inonbits/sec.
for mechanism throughput
The with
2 in 64
on throughput QAM
thewith scheme,
64 QAMwhere
denominator 6 bits/symbols
scheme,
shows where
that in the6
are allowedresults,
bits/symbols
obtained is are
depicted
allowed in Figure 11.has
is depicted
SRS transmission inbeen
Figure 11.
reduced to half.
As seen in the results section, the ML model has been trained, now its incorporation into the
network allows us to reduce the number of transmissions of SRS. This increases the throughput
capacity. Throughput Comparison
Throughput
90 is the measure of real user data (bits) transfer capacity of a BW in one second. This
is acquired by subtracting the overhead from the total capacity 77.1 of a BW. The increased throughput is
80 75
obtained by adding the saved part of SRS to the payload/ throughput of the BW.
70
𝑆 =𝐶 −𝐶 (13)
Throughput (Mbps)

60 55.1 56.7
50
40 36.7 37.77 Old Throughput
30 Increased Throughput

20
10
0
10 15 20
B.W (MHz)

Figure 11.
Figure 11. Throughput comparison
comparison for
for64
64QAM.
QAM.

As
Theseen in the results
exclusion section,
of overhead the MLthe
increases model has been
efficiency of BWtrained, now Normally,
utilization. its incorporation into of
almost 25% the
network allows
the channel us to for
is used reduce the number
reference signals,offor
transmissions
example, SRS, of SRS. This increases
Demodulation the throughput
Reference capacity.
Signal (DMRS),
Throughput
overhead and the is rest
the measure
of the 75%ofisreal userfor
usable data (bits)
real usertransfer capacity
data transfer. of a BW
Among thein25%
oneof
second. This is
the channel
acquired
BW that by subtracting
is occupied the overhead
by reference from
signals, SRSthe total
only capacity
occupies of aofBW.
4.1% The increased throughput is
the BW.
So, by
obtained in the proposed
adding solution,
the saved part the BW to
of SRS efficiency for real
the payload/ user data throughput
throughput of the BW. increases by about
2% when SRS transmission is reduced to half and the BW efficiency increases approximately by 4%
when SRS transmission is totally replaced by = ML
Soldthe Cch − Coh and hence the total usable part of BW for
model (13)
user data becomes 77% and 79% respectively. Thus, a remarkable BW efficiency improvement can be
effectively realized. Figure 12 shows a graphical representation of BW efficiency.
Electronics 2020, 9, 1660 16 of 19

Here, Cch is the capacity in bits/sec of a channel for a given BW while Coh is the reference signal
overhead capacity in bits/sec and Sold is the throughput in bits/sec.

Snew = Sold + CSRS /2 (14)

Snew is the enhanced throughput. It has been obtained by adding Sold and CSRS /2. CSRS /2 is the
channel capacity dedicated for SRS in bits/sec. The 2 in the denominator shows that in the obtained
results, SRS transmission has been reduced to half.
The exclusion of overhead increases the efficiency of BW utilization. Normally, almost 25% of
the channel is used for reference signals, for example, SRS, Demodulation Reference Signal (DMRS),
overhead and the rest of the 75% is usable for real user data transfer. Among the 25% of the channel
BW that is occupied by reference signals, SRS only occupies 4.1% of the BW.
So, in the proposed solution, the BW efficiency for real user data throughput increases by about
2% when SRS transmission is reduced to half and the BW efficiency increases approximately by 4%
when SRS transmission is totally replaced by the ML model and hence the total usable part of BW for
user data becomes 77% and 79% respectively. Thus, a remarkable BW efficiency improvement can be
effectively realized. Figure 12 shows a graphical representation of BW efficiency.
Electronics 2020, 9, x FOR PEER REVIEW 17 of 19

BW efficiency comparision

80% 79.10%
79%
78%
77%
Old efficiency
76% 75%
Total extermination of SRS
75%
74%
73%
72%

Figure
Figure 12.
12. Bandwidth
Bandwidth (BW)
(BW) efficiency.
efficiency.

The typical
typical power
powerconsumption
consumptionforfora aCPS
CPSininuplink
uplinkis is
2323
dBmdBmperper physical
physical resource
resource block,
block, as
as per
per
3 GPP3 GPP
[46].[46].
ThisThis becomes
becomes 199 mW
199 mW on aon a linear
linear scale.scale.
SinceSince the number
the number of transmissions
of transmissions of SRSofhas
SRSbeen
has
been significantly
significantly reduced,
reduced, a decrease
a decrease in power inconsumption
power consumption cannoticed
can also be also bebynoticed
the same byproportion.
the same
proportion. It saves
It saves us power of us power
almost 4.1of almost
mW, which4.1inmW, whichbecomes
log scale in log scale becomes
6.2 dBm. 6.2 useful
Hence, dBm. Hence,
power canuseful
be
power
conservedcan by
be consumption
conserved by reduction.
consumption reduction.

10. Conclusions
Conclusions

In this
thispaper,
paper,a different paradigm
a different is investigated
paradigm for introducing
is investigated artificial intelligence
for introducing in mitigating
artificial intelligence in
resource utilization
mitigating in a physical
resource utilization in layer of SRS.
a physical Weofproposed
layer SRS. We aproposed
novel method
a novel formethod
predicting an uplink
for predicting
SINR,
an which
uplink is based
SINR, whichonisSRS
basedusing an ANN-based
on SRS scheme. This
using an ANN-based research
scheme. Thisholds potential
research holds value for
potential
enhancing the most precious aspects of RRM. We attempted to mitigate the research gap
value for enhancing the most precious aspects of RRM. We attempted to mitigate the research gap by by increasing
the throughput,
increasing saving uplinksaving
the throughput, power,uplink
and increasing
power, BWand efficiency forBW
increasing 5G networks
efficiencythrough
for 5G prediction
networks
through prediction rather than estimation via pilot signals. This research is important for exploring
new future ideas, such as the prediction of PL, fading, QoS and power control.
Author Contributions: Conceptualization, methodology, software, writing—original draft, R.U.; supervision,
validation, S.N.K.M.; formal analysis, A.M.A.; investigation, S.A.; writing—review and editing, A.H.; resources,
T.K.; project administration, M.T. All authors have read and agreed to the published version of the manuscript.
Electronics 2020, 9, 1660 17 of 19

rather than estimation via pilot signals. This research is important for exploring new future ideas,
such as the prediction of PL, fading, QoS and power control.

Author Contributions: Conceptualization, methodology, software, writing—original draft, R.U.; supervision,

validation, S.N.K.M.; formal analysis, A.M.A.; investigation, S.A.; writing—review and editing, A.H.; resources,
T.K.; project administration, M.T. All authors have read and agreed to the published version of the manuscript.
Funding: This research is a part of the Higher Education Commission of Pakistan funded projects, grant number
No.9-5(Ph-IV-MG-2)/Peridot/R&D/HEC/2017.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Kelechi, G.E.; Matthew, N.O.S.; Sarhan, M.M. 5G Wireless Technology: A Primer. Int. J. Sci. Eng. Technol.
2018, 7, 62–64.
2. Egham, G. Gartner Says 8.4 Billion Connected ”Things” Will Be in Use in 2017, up 31 Percent from 2016,
14 January 2018. Available online: https://www.gartner.com/newsroom/id/3598917 (accessed on 8 June 2020).
3. Zhang, N.; Yang, P.; Ren, J.; Chen, D.; Yu, L.; Shen, X. Synergy of Big Data and 5G Wireless Networks:
Oppurtunities. Challenges and Approaches. IEEE Wireless Commun. 2018, 25, 12–18. [CrossRef]
4. Mohammed, S.H.; Ahmed, Q.L.; Taisir, E.H.E.-G.; Jaafar, M.H.E. Big Data Analytics for Wireless and Wired
Network Design: A Survey. Comput. Netw. 2018, 132, 180–199.
5. IBM. 10 Key Marketing Trends for 2017. Paul Writer. Available online: https://paulwriter.com/10-key-
marketing-trends-2017 (accessed on 2 February 2020).
6. Li, S.; Xu, L.D.; Zhao, S. 5G Internet of Things: A Survey. J. Ind. Inf. Integr. 2018, 10, 1–9. [CrossRef]
7. Huawei. 5G: A Technology Vision; Huawei: Shenzhen, China, 2019.
8. Uchino, T.; Kai, K.; Toeda, T.; Takahashi, H. Specifications of NR Higher Layer in 5G. NTT DOCOMO Tech. J.
2019, 20, 3.
9. Marimuthu, S. 5G NR White Paper. In New Era for Enhanced Mobile Broadband and 5G. A Future beyond
Imagination; Available online: http://www.u5gig.ae/MediaTek-5G-NR-White-Paper-PDF5GNRWP.pdf
(accessed on 8 June 2020).
10. Ghosh, A. 5G New Radio (NR): Physical Layer Overview and Performance. In Proceedings of the IEEE
Communication Theory Workshop. Available online: http://ctw2018.ieee-ctw.org/files/2018/05/5G-NR-CTW-
final.pdf (accessed on 8 June 2020).
11. Jakobsson. The 5G Future Will Be Powered By AI. Network Computing, 14 Mar 2019. Available online:
https://www.networkcomputing.com/wireless-infrastructure/5g-future-will-be-powered-ai (accessed on
14 February 2020).
12. Cayamcela, M.E.M.; Lim, W. Artificial Intelligence in 5G Technology: A Survey. In Proceedings of the 2018
International Conference on Information and Communication Technology Convergence (ICTC), Jeju, Korea,
17–19 October 2018.
13. Zhang, C.; Patras, P.; Hadded, H. Deep Learning in Mobile and Wireless Networking: A Survey. IEEE Commun.
Surv. Tutor. 2019, 21, 2224–2287. [CrossRef]
14. GSMA. Spectrum Handbook; GSMA. Available online: https://www.gsma.com/spectrum/wp-content/uploads/
2012/09/Spectrum-GSMA-Spectrum-Handbook-2012.pdf (accessed on 8 June 2020).
15. Lin, X.; Li, J.; Baldemair, R.; Cheng, T.; Parkvall, S.; Larsson, D.; Koorapaty, H.; Frenne, M.; Falahati, S.;
Grövlen, A.; et al. 5G New Radio: Unveiling the Essentials of the Next Generation Wireless Access Technology.
IEEE Commun. Stand. Mag. 2019, 3, 30–37. [CrossRef]
16. Mathworks. MathWorks. Available online: https://www.mathworks.com/help/lte/ug/sounding-reference-
signal-srs.html (accessed on 9 July 2019).
17. Tian, H.; Yang, L.; Li, S. SNR Estimation based on Sounding Reference Signal in LTE Uplink. In Proceedings of
the 2013 IEEE International Conference on Signal Processing, Communication and Computing (ICSPCC 2013),
KunMing, China, 5–8 August 2013.
Electronics 2020, 9, 1660 18 of 19

18. Giordano, L.G.; Campanalonga, L.; P’erez, D.L.; Rodriguez, A.G.; Geraci, G.; Baracca, P.; Magarini, M.
Uplink Sounding Reference Signal Coordination to Combat Pilot Contamination in 5G Massive MIMO.
In Proceedings of the 2018 IEEE Wireless Communications and Networking Conference (WCNC), Barcelona,
Spain, 15–18 April 2018.
19. Marwat, S.N.K. Future Machine-to-Machine Communications: LTE-A Optimization for M2M Applications;
WiSa Jessica Haunschild/Christian Schön GbR; Ibidem-Verlag: Stuttgart, Germany, 2018; ISBN 9783955380304.
20. Mahyuddin, M.F.M.; Isa, A.A.M.; Zin, M.S.I.M.; H, A.M.A.; Manap, Z.; Ismail, M.K. Overview of Positioning
Techniques for LTE Technology. J. Telecommun. Electron. Comput. Eng. 2017, 9, 2–13.
21. TU Wien, Vienna LTE-A Simulators. 2019. Available online: https://www.nt.tuwien.ac.at/research/mobile-
communications/vccs/vienna-lte-a-simulators/ (accessed on 11 June 2020).
22. Camp, T.; Boleng, J.; Davies, V. A survey of Mobility Models for Ad Hoc Network Research. WCMC Spec.
Issue Mob. Ad. Hoc. Netw. Res. Trends Appl. 2002, 2, 483–502. [CrossRef]
23. Chen, M.; Challita, U.; Saad, W.; Yin, C.; Debbah, M. Artificial Neural Networks-Based Machine Learning for
Wireless Networks: A Tutorial. IEEE Commun. Surv. Tutor. 2019, 21, 3039–3071. [CrossRef]
24. Casas, P. Machine Learning Models for Wireless Network Monitoring and Analysis. In Proceedings of the
2018 IEEE Wireless Communications and Networking Conference Workshops (WCNCW): International
Workshop on Big Data with Computational, Barcelona, Spain, 15–18 April 2018.
25. Hao, K. MIT Technology Review. 25 January 2019. Available online: https://www.technologyreview.com/s/
612768/we-analyzed-16625-papers-to-figure-out-where-ai-is-headed-next/ (accessed on 15 February 2019).
26. Saija, K.; Nethi, S.; Chaudhuri, S.; M, K.R. A Machine Learning Approach for SNR Prediction in 5G Systems.
In Proceedings of the 2019 IEEE International Conference on Advanced Networks and Telecommunications
Systems (ANTS), GOA, India, 16–19 December 2019.
27. Calabrese, F.D.; Wang, L.; Ghadimi, E.; Peters, G.; Hanzo, L.; Soldati, P. Learning Radio Resource Management
in RANs:Framework, Opportunities and Challenges. IEEE Commun. Mag. 2018, 56, 138–145. [CrossRef]
28. Dhananjay, K.; Logganathan, N.; Kafle, V.P. Double Sarsa Based Machine Learning to Improve Quality of
Video Streaming over HTTP through Wireless Networks. In Proceedings of the ITU Kaleidoscope Academic
Conference, Santa Fe, Argentina, 26–28 November 2018.
29. Bhutani, G. Application of Machine-Learning Based Prediction Techniques in Wireless Networks. Int. J.
Commun. Netw. Syst. Sci. 2014, 7, 131–140.
30. Wang, J.; Jiang, C. Machine Learning Paradigms in Wireless Network Association; Shen, X., Lin, X., Zhang, K., Eds.;
Encyclopedia of Wireless Networks; Springer: Cham, Switzerland, 2018.
31. Ibarrola, E.; Davis, M.; Voisin, C.; Cristobo, L. A Machine Learning Management Model for QoE Enhancement
in Next-Generation Wireless Ecosystems. In Proceedings of the Itu Academic Conference, Santa Fe, Argentina,
26–28 November 2018.
32. Lee, G.M.; Um, T.W.; Choi, J.K. AI as a Microservice (AIMS) over 5G Networks. In Proceedings of the ITU
Kaleidoscope Academic Conference, Santa Fe, Argentina, 26–28 November 2018.
33. Mwanje, S.S.; Mannweiler, C. Towards Cognitive Autonomous Networks in 5G. In Proceedings of the ITU
Academic Conference, Santa Fe, Argentina, 26–28 November 2018.
34. Shi, Q.; Zhang, H. Fault diagnosis of an autonomous vehicle with an improved SVM algorithm subject to
unbalanced datasets. IEEE Trans. Ind. Electron. 2020. [CrossRef]
35. You, X.; Zhang, C.; Tan, X.; Jin, S.; Wu, H. AI for 5G: Research directions and paradigms. Sci. China Inf. Sci.
2019, 62, 21301. [CrossRef]
36. Memon, M.L.; Maheshwari, M.K.; Saxena, N.; Roy, A.; Shin, D.R. Artificial Intelligence-Based Discontinuous
Reception for Energy Saving in 5G Networks. Electronics 2019, 8, 778. [CrossRef]
37. Sesto-Castilla, D.; Garcia-Villegas, E.; Lyberopoulos, G.; Theodoropoulou, E. Use of Machine Learning for
energy efficiency in present and future mobile networks. In Proceedings of the IEEE Wireless Communications
and Networking Conference, Marrakesh, Morocco, 15–18 April 2019; pp. 1–6.
38. Bohli, A.; Bouallegue, R. How to Meet Increased Capacities by Future Green 5G Networks: A Survey.
IEEE Access 2019, 7, 2220–42237. [CrossRef]
39. Song, Z.; Ni, Q.; Sun, X. Spectrum and Energy Efficient Resource Allocation with QoS Requirements for
Hybrid MC-NOMA 5G Systems. IEEE Access 2018, 6, 37055–37069. [CrossRef]
40. ABROL, A.; Jha, R.K. Power Optimization in 5G Networks: A Step towards GrEEn Communication.
IEEE Access 2016, 4, 1355–1374. [CrossRef]
Electronics 2020, 9, 1660 19 of 19

41. Ikuno, J.C.; Wrulich, M.; Rupp, M. System level simulation of LTE networks. In Proceedings of the 2010 IEEE
71st Vehicular Technology Conference, Taipei, Taiwan, 16–19 May 2010.
42. Boussaada, Z.; Curea, O.; Remaci, A.; Camblong, H.; Bellaaj, N.M. A Nonlinear Autoregressive Exogenous
(NARX) Neural Network Model for the Prediction of the Daily Direct Solar Radiation. Energies 2018, 11, 620.
[CrossRef]
43. Nokia. IEEE 5G Summit Rio; Nokia: Rio de Janeiro, Brazil, 2018.
44. Tabbane, S. 5G Networks and 3GPP Release 15; ITU PITA Workshop on Mobile Network Planning and Security:
Nadi, Fiji, 2019.
45. ETSI. 3GPP TS 38.101-1 Version 15.2.0 Release 15. Available online: https://www.etsi.org/deliver/etsi_ts/138100_
138199/13810101/15.02.00_60/ts_13810101v150200p.pdf (accessed on 13 June 2020).
46. Haider, A.; Hwang, S.H. Maximum Transmit Power for UE in an LTE Small Cell Uplink. Electronics 2019,
8, 796. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).