978-1-5386-5041-7/18/$31.00 ©2018 IEEE 860 ICTC 2018
Artificial Intelligence in 5G Technology: A Survey
Manuel Eugenio Morocho Cayamcela*, Wansu Lim†
Department of Electronic Engineering, Kumoh National Institute of Technology
Gumi, Gyeongsangbuk-do, 39177, South Korea
Email: *email@example.com, †firstname.lastname@example.org
Abstract—A fully operative and efficient 5G network cannot be
complete without the inclusion of artificial intelligence (AI) routines.
Existing 4G networks with all-IP (Internet Protocol) broadband
connectivity are based on a reactive conception, leading to a poorly
efficiency of the spectrum. AI and its sub-categories like machine
learning and deep learning have been evolving as a discipline, to the
point that nowadays this mechanism allows fifth-generation (5G)
wireless networks to be predictive and proactive, which is essential
in making the 5G vision conceivable. This paper is motivated by the
vision of intelligent base stations making decisions by themselves,
mobile devices creating dynamically-adaptable clusters based on
learned data rather than pre-established and fixed rules, that will
take us to a improve in the efficiency, latency, and reliability of the
current and real-time network applications in general.
exploration of the potential of AI-based solution approaches in
the context of 5G mobile and wireless communications
technology is presented, evaluating the different challenges and
open issues for future research.
Index Terms—5G Networks, Artificial Intelligence, IT
Convergence, Machine Learning, Deep Learning, Next Generation
Artificial Intelligence is great for problems for which existing
solutions require a lot of hand-tuning or long lists of rules, for
complex problems for which there is no good solution at all using
traditional approaches, for adaptation to fluctuating
environments, to get insights about complex problems that use
large amounts of data, and in general to notice the patterns that a
human can miss . Hard-coded software can go from a long list
of complex rules that can be hard to maintain to a system that
automatically learn from previous data, detect anomalies, predict
future scenarios, etc. These problems can be tackled adopting the
capability of learn offered by AI along with the dense amount of
transmitted data or wireless configuration datasets.
We have witnessed AI, mobile and wireless systems becoming
an essential social infrastructure, mobilizing our daily life and
facilitating the digital economy in multiple shapes . However,
somehow 5G wireless communications and AI have been
perceived as dissimilar research fields, despite the potential they
might have when they are fused together.
Certain applications available in this intersection of fields
have been addressed within specific topics of AI and next-
generation wireless communication systems. Li et al. ,
highlighted the potentiality of AI as an enabler for cellular
networks to cope with the 5G standardization requirements.
Bogale et al. , discussed the Machine Learning (ML)
techniques in the context of fog (edge) computing architecture,
aiming to distribute computing power, storage, control and
networking functions closer to the users. Jiang et al. , focused
on the challenges of AI in assisting the radio communications in
intelligent adaptive learning, and decision-ma kin g.
The next generation of mobile and wireless communication
technologies also requires the use of optimization to minimize or
maximize certain objective functions. Many of the problems in
mobile and wireless communications are not linear or
polynomial, in consequence, they demand to be approximated.
Artificial neural networks (ANN) is an AI technique that has
been suggested to model the objective function of the non- linear
problem that requires optimization .
In this article, we will introduce the potential of AI from the
basic learning algorithms, ML, deep learning, etc., into the
next generation wireless networks, that help fulfilling the
coming diversified requirements of the 5G standards to operate
in a fully automated fashion, meeting the increased capacity
demand and to serve users with superior quality of experience
(QoE). The article is divided according to the level of supervision
the AI technique requires on the training stage. The major
categories boarded on the following sections are in supervised
learning, unsupervised learning, and reinforcement learning. To
understand the difference between these three learning
subcategories, a quintessential concept of learning can be
invoked: "A computer program is said to learn from experience
E with respect to some class of tasks T and performance measure
P, if its performance at tasks in T, as measured by P, improves
with experience E" .
Supervised Learning comprises looking at several examples of
a random vector x and its label value of vector y, then learning to
predict y from x, by estimating p(y | x), or particular properties of
that distribution. Unsupervised Learning implicates observing
different instances of a random vector x and aiming to learn the
probability distribution p(x), or its properties. Reinforcement
Learning interacts with the environment, getting feedback loops
between the learning system and its experiences, in terms of
rewards and penalties .
In supervised learning, each training example has to be fed
along with their respective label. The notion is that training a
learning model on a sample of the problem instances with known
optima, and then use the model to recognize optimal solutions to
new instances. A typical task on supervised
learning is to predict a target numeric value, given a set of
features, called predictors. This description of the task is called
Transfer Learning is a popular technique often used to classify
vectors. Essentially, one would train a convolutional neural
network (CNN) on a very large dataset, for example ImageNet
, and then fine-tune the CNN on a different vector dataset. The
good part here is, training on the large dataset is already done by
some people who offer the learned weights for public research
use. The dataset can change during the implementation, but the
strength of AI is that it does not depend on fixed rules; therefore
adapting the model to changes in time is done by retraining the
model with the augmented or modified dataset.
Another typical task of Supervised Learning is regression or
prediction, where the task is to predict a target numerical value,
given a set of features/attributes, also called predictors. The key
difference between classification is that with ML algorithms like
Logistic Regression, the model can output the probability of that
certain value belongs to a given class. This type of system is
trained with multiple examples of a class, along with their label,
and the model must learn how to classify new instances.
LTE small cells are increasingly being deployed in 5G
networks to cope with the high traffic demands. These small-
scale cells are characterized by its unpredictable and dynamic
interference patterns, expanding the demand for self-optimized
solutions that can lead to lower drops, higher data rates, and
lower cost for the operators. Self-organizing networks (SON) are
expected to learn and dynamically adapt to different
environments. For the selection of optimal network
configuration in SONs, several AI-based fixes had been
discussed. In , machine learning and statistical regression
techniques are evaluated (bagging tree, boosted tree, SVM, linear
regressors, etc), gathering radio performance metrics like path
loss and throughput for particular frequencies and bandwidth
settings from the cells, and adjusting the parameters using
learning-based approaches to predict the performance that a user
will experience, given previous performance measurement
instances/samples. The authors showed that the learning-based
dynamic frequency and bandwidth allocation (DFBA) prediction
methods yield outstanding performance gains, with bagging tree
prediction method as the most promising approach to increase
the capacity of next-generation cellular networks. An extensive
interest in path-loss prediction has raised since researchers
noticed the power of AI to model more efficient
path-loss models based on publicly available datasets . The
use of AI has been proved to provide adapt- ability to network
designers who rely on signal propagation models. Timoteo et al.
, proposed a path loss prediction model for urban
environments using support vector regression to ensure an
acceptable level of quality of service (QoS) for wireless network
users. They employed different kernels and parameters over the
Okumura-Hata model and Ericsson 9999 model, and obtained
similar results as a complex neural network, but with a lower
Wireless communications count actively on channel state
information (CSI) to make an acquainted decision in the
operations of the network, and during signal processing. Liu et
al. , investigated the unobservable CSI for wireless
communications and proposed a neural-network-based
approximation for channel learning, to infer this unobservable
information, from an observable channel. Their framework was
built upon the dependence between channel responses and
location information. To build the supervised learning
framework, they train the network with channel samples, where
the unobservable metrics can be calculated from traditional pilot-
aided channel estimation. The applications of their work can be
extended to cell selection in multi-tier networks, device
discovery for device-to-device (D2D) communications, or end-
to-end user association for load balancing, among others.
Sarigiannidis et al. , used a machine-learning framework
based on supervised learning on a Software-Defined-Radio-
enabled hybrid optical wireless network. The machine-learning
framework receives the traffic-aware knowledge from the SDN
controllers and adjusts the uplink-downlink configuration in the
LTE radio communication. The authors argue that their
mechanism is capable of determining the best configuration
based on the traffic dynamics from the hybrid network, offering
significant network improvements in terms of jitter and latency.
A commonly AI architecture used to model or approximate
objective functions for existing models or to create accurate
models that were impossible to represent in the past without the
intervention of learning machines, is Artificial Neural Networks
(ANN). ANNs have been proposed to solve propagation loss
estimation in dynamic environments, where the input parameters
can be selected from the information of the transmitter, receiver,
buildings, frequency, and so on, and the learning network will
train on that data to learn to estimate the function that best
approximates the propagation loss for next-generation wireless
networks –. In the same context, Ayadi et al. ,
proposed a multi-layer perceptron (MLP) architecture to predict
coverage for either short or long distance, in multiple
frequencies, and in all environmental types. The MLP presented
uses feed-forward training with back propagation to update the
weights of the ANN. They used the inputs of the ITU-R P.1812-
4 model , to feed their network composed by an input layer,
a hidden layer, and one output layer. They showed that the ANN
model is more accurate to predict coverage in outdoor
environments than the ITU model, using the standard deviation
and correlation factor as a comparison measure.
Among other AI techniques with potential for wireless
communications, there are K-Nearest Neighbors, Logistic
Regression, Decision Trees and Random Forests. Table I, shows
a summary of the potential applications of supervised learning in
5G wireless communication technologies.
In unsupervised learning, the training data is unlabeled, and the
system attempts to learn without any guidance. This technique is
particularly useful when we want to detect groups of similar
characteristics. At no point, we tell the algorithm to try to detect
groups of related attributes; the algorithm solves this connection
without intervention. However, in some cases, we can select the
number of clusters we want the algorithm to create.
Balevi et al. , incorporated fog networking into
heterogeneous cellular networks and used an unsupervised soft-
clustering algorithm to locate the fog nodes that are upgraded
from low power nodes (LPNs) to high power nodes (HPNs). The
authors showed that by applying machine learning clustering to
a priori known data like the number of fog nodes and location
of all LPNs within a cell, they were able to determine a clustering
configuration that reduced latency in the network. The latency
calculation was performed with open- loop communications,
with no ACK for transmitted packets, and compared to the
Voronoi tessellation model, a classical model based on Euclidean
A typical unsupervised learning technique is K-means
clustering; numerous authors have investigated the applications
of this particular clustering technique in the next generation
wireless network system. Sobabe et al. , proposed a
cooperative spectrum-sensing algorithm using a combination of
an optimized version of K-means clustering, Gaussian mixture
model and expectation maximization (EM) algorithm. They
proved that their learning algorithm outperformed the energy
vector-based algorithm. Song et al. , discussed how K-means
clustering algorithm and its classification capabilities can aid in
selecting an efficient relay selection among urban vehicular
networks. The authors investigated the methods for multi-hop
wireless broadcast and how K-means is a key factor in the
decision-making and learning steps of the base stations, that
learn from the distribution of the devices and chooses
automatically which are the most suitable devices to use as a
When a wireless network experience unusual traffic demand
at a particular time and location, it is often called an anomaly,
To help identify these anomalies, Parwez et al. , used mobile
network data for anomaly detection purposes, with the help of
hierarchical clustering to identify this kind of inconsistencies.
The authors claim that the detection of this data deviations helps
to establish regions of interest in the network that require special
actions, such as resource allocation, or fault avoidance
Ultra-dense small cells (UDSC) is expected to increase the
capacity of the network, spectrum and energy efficiency. To
consider the effects of cell switching, dynamic interference,
time-varying user density, dynamic traffic patterns, and
changing frequencies, Wang et al. , proposed a data-driven
resource management for UDSC using Affinity Propagation, an
unsupervised learning clustering approach, to perform data
analysis and extract the knowledge and behavior of the system
under complex environments. Later they introduced a power
control and channel management system based on the results of
the unsupervised learning algorithm. They conclude their
research stating that by means of simulation, their data-driven
resource management framework significantly improved the
efficiency of the energy and throughput in UDSC.
Alternative clustering models like K-Means, Mini- Batch K-
Means, Mean-Shift clustering, DBSCAN, Agglomerative
Clustering, etc., can be used to associate the users to a certain
base station in order to optimize the user equipment (UE) and
base stations (BS) transmitting/receiving power. Table II, shows
a summary of the potential applications of unsupervised learning
in 5G wireless communication technologies.
The philosophy of Reinforcement Learning scheme is based
on a learning system often called agent, that reacts to the
environment. The agent performs actions and gets rewards or
penalties (negative rewards) in return for its actions. That means
that the agent has to learn by itself creating a policy that defines
the action that the agent should choose in a certain situation. The
aim of the reinforcement-learning task is to maximize the
aforementioned reward over time.
Resource allocation in Long Term Evolution (LTE) net- works has
been a dilemma since the technology was introduced. To
overcome the wireless spectrum scarcity in 5G, novel deep
learning approaches that account the coexistence of LTE and
LTE-Unlicensed, to model the resource allocation problem in
LTE-U small base stations (SBS), has been introduced in .
To accomplish their contribution, the authors introduced a
reinforcement-learning algorithm based on long short-term
memory (RL-LSTM) cells to allocate pro- actively the resources
of LTE-U over the unlicensed spectrum. The problem formulated
reassembles a non-cooperative game between the SBSs, where a
RL-LSTM framework enables the SBSs to learn automatically
which of the unlicensed channels to use, based on the probability
of future changes in terms of the WLAN activity and the LTE-U
traffic loads of the unlicensed channels. This work takes into
account the value of LTE-U as a proposal that allows cellular
network operators to offload some of their data traffic, and the
connotation of AI in the form of RL-LSTM, as a promising
solution to long-term dependencies learning, sequence, and time-
series problems. Nevertheless, researchers should be warned that
this deep learning architecture is one of the most difficult to train,
due to the vanishing and exploding gradient problem in Recurrent
Neural Networks (RNN) , the speed of activation functions,
as well as the initialization of parameters for LSTM systems .
Reinforcement Learning has also played an important role on
Heterogeneous Networks (HetNets), enabling Femto Cells (FCs)
to autonomously and opportunistically sense the radio
environment and tune their parameters accordingly to satisfy
specific pre-set quality-of-service requirements. Alnwaimi et al.
, proved that by using reinforcement learning for the
femtocells self-configuration, based on dynamic-learning games
for a multi-objective fully-distributed strategy, the intra/inter-tier
interference can be reduced significantly. The collision and
reconfiguration measurements were used as a "learning cost
during training. This self-organizing potential, empower FCs to
identify available spectrum for opportunistic use, based on the
BASED SCHEMES FOR
Machine Learning and statistical logistic
Dynamic frequency and bandwidth allocation in self-organized LTE dense small
cell deployments (as in ).
Support Vector Machines (SVM).
Path loss prediction model for urban environments (as in [ 12]).
Neural-Netw ork-based a pproximation.
Channel Learning to infer unobservable channel state information (CSI) from
an observable channel (as in ).
Supervised Machine Learning Frameworks.
Adjustment of the TDD Uplink-Downlink configuration in XG-PON-LTE
Systems to maximize the network performance base d on the ongoing traffic
conditions in the hybrid optical-wirele ss networ k (as in ).
Artificial Neural Networks (ANN), and
Multi-Layer Perceptrons (MLPs).
Modelling and approximations of objective functions for link budget and
propagation loss for next-generation wireless networks (as in –).
BASED SCHEMES FOR
clustering, Gaussian Mixture
Model (GMM), and Expectation
Cooperative spectrum sensing (as in ). Relay node selection in vehicular
networks (as in ).
Anomaly/Fault/Intrusion detection in mobile wireless networks (as i n ).
Latency reduction by clustering fog nodes to automatically decide which low
power node (LPN) is upgraded to a high power node (HPN) in heterogeneous
cellular networks. (as in ).
Affinity Propagation Clustering.
Data-Driven Re source Ma nagement for Ultra-Dense Small Cells (as in ).
BASED SCHEMES FOR
Reinforce ment Learning
Reinforce ment Learning algorithm based on
long short-term memory (RL-LSTM) cells.
Proactive resource allocation in LTE-U Networks, formulated as a non-
cooperative game which enables SBSs to learn which unlicensed channel, given
the long-term WLAN activity in the channels and LTE-U traffic loads (as in
Gradient f ollower (GF), the modified Roth-
Erev (MRE), and the modified Bush and
Enable Femto-Cells (FCs) to autonomously and opportuni stically sense the
radio environment a nd tune t heir parameters i n HetNets, to reduce intra/inter-
tier interference (as in ).
Reinforcement Learning with Network
Heterogene ous Radio Access Technologies (RATs) selection (as in ).
5G wireless networks will also contain multiple radio access
technologies (RAT). However, selecting the right RAT is a latent
problem in terms of speed, exploration times, and
convergence. Nguyen et al. , developed a feedback
framework using limited network-assisted information from the
base stations (BS), to improve the efficiency of distributed
algorithms for RAT selection. The framework used
reinforcement learning with network-assisted feedback to
overcome the aforementioned problems. Table III, shows a
summary of the potential applications of reinforcement learning
in 5G wireless communication technologies.
After exploring some of the successful cases where AI is used
as a tool to improve 5G technologies, we strongly believe that
the convergence between these two knowledge expertises will
have an enormous impact in the development of future
generation networks. The era where wireless networks
researchers were afraid to use AI-based algorithms due to the
lack of understanding of the artificial-learning process, has been
left in the past. Nowadays, with the power and ubiquity of
information, numerous researchers are adapting their knowledge
and expanding their tools arsenal with AI-based models,
algorithms and practices, especially in the 5G world, where even
a few milliseconds of latency can make a difference. A reliable
5G system requires extremely low latency, which is why not
everything can be stored in remote cloud servers far away.
Latency increases with distance and congestion of network links.
Base stations have limited storage size, so they have to learn to
predict user needs by applying a variety of artificial intelligence
tools. With these tools, every base station will be able to store a
reduced but adequate set of files or contents. This is one example
why our future networks must be predictive, and how Artificial
Intelligence becomes crucial in optimizing this kind of
problems in the network. An additional goal of linking AI with
5G networks would be to obtain significant improvements in the
context of edge caching just by applying off-the-shelf machine
learning algorithms. We have shown how AI can be a solution
that can fill this gap of requirements in 5G mobile and wireless
communications, allowing base stations to predict what kind of
content users nearby may request in the near future, allocating
dynamic frequencies in self-organized LTE dense small cell
deployments, predicting path loss/link budget with
approximated NN models, inferring the unobservable channel
state information from an observable channel, adjusting the
TDD uplink-downlink configuration in XG-PON-LTE systems
based on ongoing network conditions, sensing the spectrum
using unsupervised models, reducing the latency by
automatically configuring the clusters in Het-Nets, detecting
anomalies/faults/intrusions in mobile wireless networks,
managing the resources in ultra-dense small cells, selecting the
relay nodes in vehicular networks, allocating the resources
in LTE-U networks, enabling autonomous and opportunistic
sensing of the radio environment in femto-cells, selecting the
optimal radio access technology (RAT) in HetNets, among
This work was supported by the Global Excellent Technology
Innovation Program (10063078) funded by the Ministry of
Trade, Industry and Energy (MOTIE) of Korea; and by the
National Research Foundation of Korea (NRF) grant funded by
the Korea government (MSIP; Ministry of Science, ICT &
Future Planning) (No. 2017R1C1B5016837).
A. Geron, “Hands-on machine learning with Scikit-Learn and
TensorFlow: concepts, tools, and techniques to build intelligent
systems,” p. 543, 2017. [Online]. Available:
A. Osseiran, J. F. Monserrat, and P. Marsch, 5G Mobile and Wireless
Communications Technology, 1st ed. United Kingdom: Cambridge
University Press, 2017. [Online]. Available: www.cambridge.org/
R. Li, Z. Zhao, X. Zhou, G. Ding, Y. Chen, Z. Wang, and H. Zhang,
“Intelligent 5G: When Cellular Networks Meet Artificial
Intelligence,” IEEE Wireless Communications, vol. 24, no. 5, pp.
T. E. Bogale, X. Wang, and L. B. Le, “Machine Intelligence
Techniques for Next-Generation Context-Aware Wireless
Networks,” ITU Special Issue: The impact of Artificial Intelligence
(AI) on communication networks and services., vol. 1, 2018.
C. Jiang, H. Zhang, Y. Ren, Z. Han, K. C. Chen, and L. Hanzo,
“Machine Learning Paradigms for Next-Generation Wireless
Networks,” IEEE Wireless Communications, 2017.
G. Villarrubia, J. F. De Paz, P. Chamoso, and F. D. la Prieta,
“Artificial neural networks used in optimization problems,”
Neurocomputing, vol. 272, pp. 10–16, 2018.
T. M. Mitchell, Machine Learning, 1st ed. McGraw-
Hill Science/Engineering/Math, 1997.
line]. Available: https://www.cs.ubbcluj.ro/~gabis/ml/ml-books/
I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning, 1st
T. Dietterich, Ed. London, England: The MIT Press, 2016. [Online].
P. U. Stanford Vision Lab, Stanford University, “ImageNet.”
 B. Bojovic´, E. Meshkova, N. Baldo, J. Riihijärvi, and M.
“Machine learning-based dynamic frequency and
bandwidth allocation in self-organized LTE dense small cell
deployments,” Eurasip Journal on Wireless Communications and
Networking, vol. 2016, no. 1, 2016. [Online]. Available:
S. Y. Han, N. B. Abu-ghazaleh, and S. Member, “Efficient and Con-
sistent Path loss Model for Mobile Network Simulation,”
IEEE/ACM Transactions on Networking, vol. PP, no. 99, pp. 1–1,
R. D. a. Timoteo, D. C. Cunha, and G. D. C. Cavalcanti, “A Proposal
for Path Loss Prediction in Urban Environments using Support
Vector Re- gression,” Advanced International Conference on
Telecommunications, vol. 10, no. c, pp. 119–124, 2014.
J. Liu, R. Deng, S. Zhou, and Z. Niu, “Seeing the unobservable:
Channel learning for wireless communication networks,” 2015
IEEE Global Communications Conference, GLOBECOM 2015,
P. Sarigiannidis, A. Sarigiannidis, I. Moscholios, and P.
Zwierzykowski, “DIANA: A Machine Learning Mechanism for
Adjusting the TDD Uplink-Downlink Configuration in XG-PON-
LTE Systems,” Mobile Information Systems, vol. 2017, no. c, 2017.
S. P. Sotiroudis, S. K. Goudos, K. A. Gotsis, K. Siakavara, J. N.
Sahalos, and L. Fellow, “Application of a Composite Differential
Evolution Algorithm in Optimal Neural Network Design for
Propagation Path-Loss Prediction in Mobile Communication
Systems,” IEEE ANTENNAS AND WIRELESS PROPAGATION
LETTERS, vol. 12, pp. 364–367, 2013.
J. M. Mom, C. O. Mgbe, and G. A. Igwue, “Application of Artificial
Neural Network For Path Loss Prediction In Urban Macrocellular
Envi- ronment,” American Journal of Engineering Research
(AJER), vol. 03, no. 02, pp. 270–275, 2014.
I. Popescu, D. Nikitopoulos, I. Nafornita, and P. Constantinou,
“ANN prediction models for indoor environment,” IEEE
International Confer- ence on Wireless and Mobile Computing,
Networking and Communica- tions 2006, WiMob 2006, pp. 366–
S. P. Sotiroudis, K. Siakavara, and J. N. Sahalos, “A Neural Network
Approach to the Prediction of the Propagation Path-loss for Mobile
Communications Systems in Urban Environments,” PIERS Online,
vol. 3, no. 8, pp. 1175–1179, 2007.
M. Ayadi, A. Ben Zineb, and S. Tabbane, “A UHF Path Loss Model
Us- ing Learning Machine for Heterogeneous Networks,” IEEE
Transactions on Antennas and Propagation, vol. 65, no. 7, pp.
International Telecommunication Union, “A path-specific
propagation prediction method for point-to-area terrestrial services
in the VHF and UHF bands,” ITU P-Series Radiowave propagation,
no. P.1812-4, pp. 1–35, 2015. [Online]. Available:
E. Balevi and R . D. Gitlin, “Unsupervised machine learning in 5G
networks for low latency communications,” 2017 IEEE 36th
national Performance Computing and Communications
Conference, IPCCC 2017, vol. 2018-Janua, pp. 1–2, 2018.
 G. Sobabe, Y. Song, X. Bai, and B. Guo, “A cooperative spectrum
sensing algorithm based on unsupervised learning,” 10th International
Congress on Image and Signal Processing, BioMedical Engineering
and Informatics (CISP-BMEI 2017), vol. 1, pp. 198–201, 2017.
 W. Song, F. Zeng, J. Hu, Z. Wang, and X. Mao, “An Unsupervised-
Learning-Based Method for Multi-Hop Wireless Broadcast Relay Se-
lection in Urban Vehicular Networks,” IEEE Vehicular Technology
Conference, vol. 2017-June, 2017.
 M. S. Parwez, D. B. Rawat, and M. Garuba, “Big data analytics for
user-activity analysis and user-anomaly detection in mobile wireless
network,” IEEE Transactions on Industrial Informatics, vol. 13, no. 4,
pp. 2058–2065, 2017.
 L.-C. Wang and S. H. Cheng, “Data-Driven Resource Management
for Ultra-Dense Small Cells: An Affinity Propagation Clustering
Approach,” IEEE Transactions on Network Science and Engineering,
vol. 4697, no. c, pp. 1–1, 2018. [Online]. Available:
 U. Challita, L. Dong, and W. Saad, “Deep learning for proactive
resource allocation in LTE-U networks,” in European Wireless 2017-
23rd European Wireless Conference, 2017. [Online]. Available:
 R. Pascanu, T. Mikolov, and Y. Bengio, “On the difficulty
of training recurrent neural networks,” Tech. Rep., 2013.
 Q. V. Le, N. Jaitly, and G. E. Hinton Google, “A Simple Way to
Initialize Recurrent Networks of Rectified Linear Units,” Tech. Rep.,
2015. [Online]. Available: https://arxiv.org/pdf/1504.00941v2.pdf
 G. Alnwaimi, S. Vahid, and K. Moessner, “Dynamic heterogeneous
learning games for opportunistic access in LTE-based
macro/femtocell deployments,” IEEE Transactions on Wireless
Communications, vol. 14, no. 4, pp. 2294–2308, 2015.
 D. D. Nguyen, H. X. Nguyen, and L. B. White, “Reinforcement
Learning with Network-Assisted Feedback for Heterogeneous RAT
Selection,” IEEE Transactions on Wireless Communications, vol. 16,
no. 9, pp. 6062–6076, 2017.