5G FOR ULTRA-RELIABLE LOW-LATENCY COMMUNICATIONS

A Machine Learning Framework for Resource Allocation Assisted by Cloud Computing

Jun-Bo Wang, Junyuan Wang, Yongpeng Wu, Jin-Yuan Wang, Huiling Zhu, Min Lin, and Jiangzhou Wang

Abstract In practical networks, the overall performance

depends on how to exploit the fluctuation of
Conventionally, resource allocation is formu- wireless channels and traffic loads to efficiently
lated as an optimization problem and solved and dynamically manage the hyper-dimensional
online with instantaneous scenario information. radio resources (e.g., frequency bands, time slots,
Since most resource allocation problems are not orthogonal codes, transmit power, and trans-
convex, the optimal solutions are very difficult mit-receive beams) and fairly to support users’
to obtain in real time. Lagrangian relaxation or QoS requirements. On one hand, radio resources
greedy methods are then often employed, which are inherently scarce, since all users competitively
results in performance loss. Therefore, the con- share the common electromagnetic spectrum and
ventional methods of resource allocation are wireless infrastructures. On the other hand, wire-
facing great challenges to meet the ever increas- less services have become increasingly sophis-
ing QoS requirements of users with scarce radio ticated and various, each with a wide range of
resource. Assisted by cloud computing, a huge QoS requirements. Efficient and robust resource
amount of historical data on scenarios can be col- allocation algorithms are essential for the success
lected for extracting similarities among scenar- of future mobile networks. Conventionally, the
ios using machine learning. Moreover, optimal resource allocation problems are often formu-
or near-optimal solutions of historical scenarios lated mathematically as optimization problems.
can be searched offline and stored in advance. After collecting instantaneous channel state infor-
When the measured data of a scenario arrives, mation (CSI) and QoS requirements of users,
the current scenario is compared with historical the formulated optimization problems are solved
scenarios to find the most similar one. Then the online. That is, the solutions must be obtained
optimal or near-optimal solution in the most sim- shortly since wireless channels and traffic loads
ilar historical scenario is adopted to allocate the vary quickly. However, most of the optimization
radio resources for the current scenario. To facil- problems are not convex [3], which indicates that
itate the application of new design philosophy, the optimal solutions are often very difficult to
a machine learning framework is proposed for obtain, especially in scenarios with a lot of users
resource allocation assisted by cloud comput- and diverse radio resources. Therefore, conven-
ing. An example of beam allocation in multi-user tional Lagrangian relaxation or greedy methods
massive MIMO systems shows that the proposed are often employed to find solutions online. Inev-
machine-learning-based resource allocation out- itably, online solutions of resource allocation will
performs conventional methods. result in performance loss. With increasing QoS
requirements, conventional methods are facing
Introduction great challenges in designing more sophisticated
With the rapid development of electronic devices resource allocation schemes to further improve
and mobile computing techniques, worldwide system performance with scarce radio resource,
societal trends have demonstrated unprecedent- which motivates the exploration of a novel design
ed changes in the way wireless communications philosophy for resource allocation.
are used. It is predicted that the monthly traffic In March 2016, a five-game Go match was
of smartphones around the world will be about held between 18-time world champion Lee Sedol
50 exabytes in 2021 [1], which is about 12 times and AlphaGo, a computer Go program devel-
that in 2016. Obviously, wireless communications oped by Google DeepMind [4]. From the view of
have become indispensable to our society and are conventional computing theory, Go was previous-
involved in many aspects of our life. Many familiar ly regarded as an extremely difficult problem that
scenarios such as ultra-dense residential areas and was expected to be out of reach for the state-of-
office towers, subways, highways, and high-speed the-art technologies. Surprisingly, AlphaGo found
railways challenge the future mobile networks in its moves based on the knowledge previously
terms of ultra-high traffic volume density, ultra- “learned” from historical match records and won
high connection density, and ultra-high mobility. all but the fourth game. Inspired by the victory
Due to its ability to guarantee the users’ quality of of AlphaGo, the application of machine learning
service (QoS) and optimize the usage of facilities techniques to address the challenges in future
to maximize operators’ revenue, how to allocate communications has attracted great attention and
radio resources more efficiently is always a hot been discussed widely [5, 6].
topic for future wireless communications [2]. In practical wireless communications, the radio

Digital Object Identifier: Jun-Bo Wang (corresponding author) is with Southeast University; Junyuan Wang is with Edge Hill University; Huiling Zhu and Jiangzhou Wang are with the
10.1109/MNET.2018.1700293 University of Kent; Yongpeng Wu is with Shanghai Jiao Tong University; Jin-Yuan Wang and Min Lin are with Nanjing University of Posts and Telecommunications.

144 0890-8044/18/$25.00 © 2018 IEEE IEEE Network • March/April 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on November 29,2023 at 00:01:53 UTC from IEEE Xplore. Restrictions apply.
 resources are dynamically allocated according to
the instantaneous information including CSI and
QoS requirements of users. Inexpensive cloud
storage makes it very easy to save the information Cloud
as data on historical scenarios that previously we
would have ignored and trashed. Recent investi-
gations have found that these data convey a lot of
similarities between current and historical scenari-
os on user requirements and wireless propagation k
lin
ul

Ba
environments [7]. Using the similarities among a

ckh
ckh

Backhaul link
scenarios, the solutions of resource allocation in Ba

au
l li
historical scenarios can be exploited to improve

nk
the resource allocation of the current scenario.
More specifically, the solutions of resource alloca-
tion in historical scenarios can be searched offline
and stored in advance. When the measured data
of the current scenario arrives, it is not neces- Base station
sary to use conventional Lagrangian relaxation or Base station
Base station
greedy methods to solve the resource allocation
problem online. Instead, we only need to com-
pare the current scenario with historical scenar-
ios and find the most similar one. Then we use
the solution of the most similar historical scenar-
io to allocate the radio resources for the current
scenario. Interestingly, the offline characteristic
makes it possible to use advanced cloud com- FIGURE 1. Wireless communications assisted by cloud computing.
puting techniques to find optimal or near-optimal
solutions of resource allocation for historical sce- the system parameters or wireless propagation
narios, which can improve the performance of parameters, such as the bandwidth, subcarrier
resource allocation accordingly. number, and background noise level. {gi}m i=1 and
{H i} m
i=1 are used to define the specific scenario
Mathematical Modeling of and the limitations on resource allocation, such
as the available amount of radio resources, users’
Resource Allocation QoS requirements, and the impacts from all kinds
As illustrated in Fig. 1, the architecture of wire- of interference and noise. The objective function
less communications assisted by cloud computing describes the characteristics of the best possible
consists of three main components: solution and reveals the design objective, that is,
• Configurable computing resources clustered the key performance metrics for resource alloca-
as a cloud with high computational and stor- tion. For a specified scenario described by a, the
age capabilities optimal solution of resource allocation x* is the
• Base station (BS) with wireless access func- vector that obtains the best value of objective
tions function among all possible vectors and satisfies
• Backhaul links that deliver the measured data all constraints.
of real scenarios from the BS to the cloud
and deploy the machine-learning-based A Machine Learning Framework
resource allocation schemes at the BS. For existing wireless systems assisted by cloud
More details are discussed in the next section. computing, a huge amount of data on historical
In general, the resource allocation performed at scenarios may have been collected and stored
the BS can be formulated as a mathematical opti- in the cloud. The strong computing capability of
mization problem [3], given by the cloud is exploited to search for optimal or
minimize f (x,a) near-optimal solutions for these historical scenar-
x∈S ios. By classifying these solutions, the similarities
subject to gi (x,a) ≤ 0 i = 1, !, m hidden in these historical scenarios are extracted
as a machine-learning-based resource allocation
hi (x,a) = 0 i = 1, !, p (1)
scheme. The machine-learning-based resource
where x is the variable vector of the problem, f(·,·) allocation scheme will be forwarded to guide a
is the objective function to be minimized over BS in allocating radio resource more efficiently.
the vector x, a is the parameter vector that spec- When a BS is deployed in a new area, there is
ifies the problem instance, {g i}m m
i=1 and {H i} i=1 are usually no available data about historical scenar-
called inequality and equality constraint functions, ios. In this case, the initial historical data can be
respectively, and S is called a constraint set. By generated from an abstract mathematical model
convention, the standard form defines a minimi- with realistic BS locations, accurate building foot-
zation problem. A maximization problem can be prints, presumptive user distribution and require-
treated by negating the objective function. ments, and wireless propagation models. When
If a resource allocation problem is formulated the new BS emerges into service, the measured
in the form of Eq. 1, all elements in vector x are data of real-time scenarios will be collected from
referred to as variables that describe the allocat- practical systems and later used as historical data
ed amount or configuration of radio resources, for learning.
such as the transmit power level and the assigned The proposed machine learning framework is
subcarrier index. All elements in vector a are shown in Fig. 2. In the cloud, a huge amount of

IEEE Network • March/April 2018 145

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on November 29,2023 at 00:01:53 UTC from IEEE Xplore. Restrictions apply.
 ated solution of the lth class is selected to allocate
Through feature selection, some key attributes are selected from the historical data and presented as a radio resources for the scenario depicted by FT.
feature vector. However, there may exist some operation faults in the data measurement, transmission, Before deploying the model, the recently built
predictive model is evaluated by the test set and
and storage, which results in the abnormal, incomplete or duplicate values in feature vectors. further optimized until the evaluation results are
satisfactory.
Using the backhaul links, the built predictive
historical data on scenarios are stored using the model and the associated solutions of all classes
cloud storage. The historical data have a lot of are transmitted to the BS. At the BS, the measured
attributes, including the user number, the CSI of data of a real-time scenario is first used to form its
users, international mobile subscriber identifica- new feature vector. Then the new feature vector
tion numbers (IMSIs) of users, and so on. Some will be input into the built predictive model to
attributes, such as IMSIs of users, may be irrele- allocate radio resource. Meanwhile, the new fea-
vant for the specific resource allocation; hence, ture vector will be collected and stored temporar-
these irrelevant attributes are not included in ily at the BS and forwarded to the cloud later for
the parameter vector a in the optimization prob- updating the dataset, which is very important for
lem of Eq. 1. Learning from a large number of tracing the evolution of real scenarios, including
raw data with many attributes generally requires user behaviors and wireless propagation environ-
a large amount of memory and computation ments.
power, and it may influence the learning accu- Although a lot of computing resources are
racy [8]. Therefore, the irrelevant attributes can consumed to build a predictive model, the com-
be removed without incurring much loss of data puting work can be carried out offline during off-
quality. In order to reduce the dimensionality of peak time. Moreover, the dataset updating and
the data and enable the learning process to oper- the model deployment can also be accomplished
ate faster and more effectively, feature selection is during off-peak time. Therefore, the cloud can be
carried out to identify and remove as many irrel- shared with multiple BSs, and the computing tasks
evant attributes as possible, which is discussed in can be flexibly scheduled to make full use of the
the next section. available computing resources.
Through feature selection, some key attributes
are selected from the historical data and present-
ed as a feature vector. However, there may be
Application of Supervised Learning to
some operation faults in the data measurement, Resource Allocation
transmission, and storage, resulting in abnormal, In the proposed machine learning framework, a
incomplete, or duplicate values in feature vectors. machine learning algorithm is adopted to build a
Therefore, necessary preprocessing is required predictive model. Generally speaking, a machine
to delete erroneous or duplicate feature vectors. learning algorithm is usually categorized as either
Then all remaining feature vectors are collected supervised or unsupervised [7]. In supervised
to form a very large dataset. Further, all feature learning, the goal is to learn from training data that
vectors in a dataset are split randomly into a train- are labeled with nonnegative integers or classes in
ing and a test set. Normally, 70–90 percent of the order to later predict the correct response when
feature vectors are assigned into the training set. dealing with new data. The supervised approach
With the training set, a supervised learning is indeed similar to human learning under the
algorithm in machine learning is adopted to find supervision of a teacher. The teacher provides
the similarities hidden in historical data. By doing good training examples for the student, and the
so, a predictive model can be built that will be student then derives general rules from these spe-
used to make a resource allocation decision for cific examples. In contrast to supervised learning,
a future unexpected scenario. More specifical- the data for unsupervised learning have no labels,
ly, with the aid of cloud computing, advanced so the goal instead is to organize the data and
computing techniques can be used to search the find hidden structures in unlabeled data. Most
solutions for the optimization problem of Eq. 1 machine learning algorithms are supervised. In
with more computational time. Compared to con- the following, we discuss how to apply supervised
ventional Lagrangian relaxation or greedy meth- learning to solve the resource allocation problem.
ods, the performance of searched solutions can
be improved significantly. Therefore, a high-per- Feature Selection
formance solution of resource allocation can be In machine learning, feature selection, also known
searched offline and associated with each training as attribute selection, is the process of selecting a
feature vector, as discussed in the next section. subset of relevant attributes in historical data to
All training feature vectors with the same solu- form a feature vector for building predictive mod-
tions are classified into one class, and each class els. The selection of an appropriate feature vector
is associated with its own solution. The resource is critical due to the phenomenon known as “the
allocation problem is now transformed into a mul- curse of dimensionality” [9]. That is, each dimen-
ticlass classification problem, outlined in the next sion that is added to the feature vector requires
section. In order to solve the multiclass classifica- exponentially increasing data in the training set,
tion problem, a predictive model will be built with which usually results in significant practical perfor-
two functions. The first is to predict the class for mance degradation. Therefore, it is necessary to
a future scenario, which can be mathematically find a low dimension of feature vectors that cap-
described as a classifier l = Classifier(FT). FT is the tures the essence of resource allocation in practi-
input feature vector extracted from the scenario, cal scenarios.
and l is the output class index showing that the In order to reduce the dimensionality of fea-
scenario belongs to the lth class. Then the associ- ture vectors, only valuable information for the

146 IEEE Network • March/April 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on November 29,2023 at 00:01:53 UTC from IEEE Xplore. Restrictions apply.
 resource allocation can be selected as features.
Test set
After modeling the resource allocation as the
Training Model
optimization problem in Eq. 1, all valuable infor- set Model Model deployment
Dataset
mation is included in the parameter vector a. building Model evaluation
Observing the elements of a, it can be found that optimization
they can be further divided into two categories:
time-variant (dynamic) or time-invariant (static). Historical
data
Feature
Feature vector Machine
Some elements are constants and thus labeled selection Dataset learning
as time-invariant parameters, such as subcarrier Cloud
number, maximum transmit power, and anten- Model
na number. Other elements that change quickly deployment
and are required to be measured and fed back New
all the time for making decisions on resource feature Solution of the
predicted class
allocation are labeled as time-variant parameters, New data Feature
selection
vector Predictive
model
such as user number, CSI of all users, and interfer-
Base station
ence levels. As the time-invariant parameters stay
unchanged, in order to minimize the dimension of
New Predicted
the feature vectors, only the time-variant parame- feature class Solution of the
ters can be considered to be features. Moreover, Feature vector Classifier index Solution predicted class
selection selector
some time-variant parameters cannot be selected
as features since it may be redundant in the pres-
ence of another relevant feature with which it is
strongly correlated. In short, an individual feature FIGURE 2. A machine learning framework of resource allocation.
vector specifies a unique scenario for resource
allocation. However, it should be noted that fea- resource allocation issues can be modeled as an
ture selection is a process of trial and error, which integer optimization problem. When the num-
can be time consuming and costly, especially with ber of integer variables in an integer optimization
very large datasets. problem is very small, the optimal solution can be
found by exhaustive search. However, if there are
Solutions of Optimization Problems many integer variables, finding an optimal solu-
To facilitate the application of supervised learning, tion of resource allocation is extremely computa-
the solution of the resource allocation problem tionally complex because they are known to be
specified by each training feature vector should NP-hard [10]. In this case, it is more feasible to
be obtained in advance. Then each training fea- search the near-optimal solutions for all training
ture vector is associated with its solution. Accord- feature vectors. Moreover, the offline characteristic
ing to the associated solutions, all feature vectors of building model and the strong cloud computing
are labeled into multiple classes. More specifically, and storage capabilities make it possible to spend
all training feature vectors with the same solution more computation time using the metaheuristics to
are placed with the same class label, indexed by a search near-optimal solutions. Some famous meta-
nonnegative integer. In other words, each class is heuristics algorithms [11], such as particle swam
associated with its unique solution. The class label optimization (PSO) and ant colony optimization
information of all training feature vectors will be (ACO), have been applied to the solution of many
used to build a predictive model. In practice, the classic combination optimization problems. For
measured data of a real-time scenario is select- most of these applications, the results show that
ed as a new feature vector. Then the predictive these metaheuristics algorithms outperform other
model will predict the class for the new feature algorithms, including conventional Lagrangian
vector and output the associated solution of the relaxation or greedy algorithms.
predicted class, that is, how to allocate the radio
resource for the real-time scenario. Obviously, if Multiclass Classification Problem and Classifier
too many training feature vectors are associated When the class label information is ready for all
with low-performance solutions, the built predic- training feature vectors, it starts to look for the
tive model cannot supply high-performance solu- similarities hidden in labeled feature vectors.
tions for practical resource allocation. Therefore, Mathematically, the set of all possible feature vec-
finding optimal or near-optimal solutions of all tors constitutes a feature space. As a special case,
training feature vectors is crucial for building a if the order of the feature vectors is 2, the feature
high-performance predictive model. space is a two-dimensional space, as shown Fig.
In the resource allocation problem 1, all 3. Note that f1 and f2 are the first and second ele-
elements in vector x are used to describe how ments of feature vectors, respectively. When all
to allocate the radio resources. Mathematically, labeled feature vectors are shown in the feature
the allocation of many radio resources can be space, it can be observed that the feature vec-
described by integer variables, such as subcarriers, tors with the same class label are often distributed
time slots, and modulation and coding schemes. very closely. Accordingly, the feature space can
Intuitively, the transmit power level can be adjusted be divided into several subspaces, and most fea-
arbitrarily between the maximum transmit power ture vectors with the same class label are located
and zero. It seems that only a continuous variable within the same subspace. Then the hidden sim-
can be used to describe the transmit power allo- ilarities can be exploited by building a classifier,
cation. However, in order to simplify the system which predicts the class of a new feature vector
complexity, transmitters in practical systems are by determining the subspace in which it is locat-
usually allowed to transmit signals with only a few ed. In supervised learning, this learning process
pre-fixed power levels. Therefore, most practical is often called a multiclass classification problem

IEEE Network • March/April 2018 147

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on November 29,2023 at 00:01:53 UTC from IEEE Xplore. Restrictions apply.
 As the number of beams N is much larger than
More advanced techniques are required to design low-complexity multiclass classifiers. One promising the number of users K, and each user is served by
technique is to transform the multiclass classification problem into a set of binary classification prob- one beam, only some of the beams will be active
to serve users. Therefore, we first need to decide
lems that are efficiently solved using binary classifiers. which beams are active. This can be solved by
applying our machine learning framework. Spe-
cifically, the active beam solution serves as the
output of the predictive model. By assuming a
f2 Class 1
Class 2
line of sight (LOS) channel, as the beam gains
Class 3 of K users from N beams are determined by the
K users’ locations, the user layout u = [(r 1, q 1),
(r2, q2),…, (rK, qK)] should serve as the input data,
which contains both the radial distance and phase
information. Since the beam gains from various
beams for a user significantly vary with its phase,
as shown in Fig. 5, the achievable sum rate with a
beam allocation solution is mainly determined by
the phase information of K users. Therefore, the
feature vector Fu of a user layout data u is select-
ed as
f1
FT = [cosq(1), cosq(2), …, cosq(K)], (2)
FIGURE 3. Multiclass classification.
where q (1) ≤ q (2) ≤ … ≤ q (K) is the order statistics
obtained by arranging q1, q2, …, qK.
[12]. So far, many machine learning algorithms Before performing resource allocation, we first
have been used for designing multiclass classi- need to train the predictive model by learning
fiers. Selecting a machine learning algorithm is from a large amount of training user layout data,
also a process of trial and error. It is a trade-off which can be generated by computer according
between specific characteristics of the algorithms, to its distribution. For each training user layout, its
such as computational complexity, speed of learn- feature vector formed according to Eq. 2 is asso-
ing, memory usage, predictive accuracy on new ciated with its active beam solution, which can be
feature vectors, and interpretability. Therefore, the obtained by employing offline beam allocation
design of a multiclass classifier is an essential task. algorithms. Thanks to the strong cloud computing
Here, we briefly introduce the k-nearest neigh- capability, optimal exhaustive search or near-op-
bor (k-NN) algorithm, which is known to be very timal metaheuristics algorithms can be adopted
simple to understand but works incredibly well in as mentioned earlier. In this section, exhaustive
practice. As shown in Fig. 4, whenever a new fea- search is applied for demonstration by assuming
ture vector arrives, the k-NN algorithm picks up a smaller number of users and beams. After asso-
all k nearest neighbors of the new feature vector ciating each feature vector in the training set with
from the training set. Then the new feature vector its active beam solution, all the training feature
is judged to belong to the most common class vectors are naturally classified into a variety of
among its k nearest neighbors. If k = 1, the new classes according to their active beam solutions.
feature vector is simply categorized to the class of Specifically, the feature vectors sharing the same
its nearest neighbor. active beam solution are in the same class. A pre-
dictive model of active beam solution can then be
Example: Beam Allocation in built by applying a simple k-NN algorithm,1 which
can then be evaluated and optimized to guaran-
Multiuser Massive MIMO tee its performance, as Fig. 2 shows. For instance,
In this section, the beam allocation problem in a it can be improved by adding more training data.
single-cell multiuser massive multiple-input multi- The effect of the size of training set is discussed
ple-output (MIMO) system considered in [13] is by presenting Fig. 6b.
used as an example to demonstrate the efficiency The built predictive model is then deployed at
of our proposed machine learning framework of the BS for beam allocation. For a new user layout
resource allocation. ui, by forming its feature vector Fui and defining
In the single-cell system, it is assumed that the the distance from its feature vector Fui to a stored
BS is located at the center of the circular cell and training feature vector FTJ, dui,Tj, as
K users are uniformly distributed within the cell
with unit radius, and each user is equipped with dui,Tj =||Fui – FTJ||2, (3)
a single antenna. A massive number of N >> K
fixed beams are formed by deploying the But- k nearest neighbot feature vectors with k small-
ler network [14] with a linear array of N identical est distances are picked. According to the k-NN
isotropic antenna elements at the BS. In such a algorithm, the most common class among these
fixed-beam system, a user is served by a beam k neighbors is chosen as the predictive class of
allocated to it, as shown in Fig. 5, where each the input user layout ui, and the predictive model
user is served by the beam in the same color, and outputs the associated active beam solution of its
1 The k-NN algorithm is
(rk, qk) denotes the polar coordinate of user k. To predicted class. Based on the active beam infor-
employed in this section serve multiple users simultaneously, the key prob- mation, each active beam is allocated to its best
for illustration due to its lem is how to efficiently allocate beams to users user with the highest received signal-to-interfer-
simplicity. to maximize the sum rate. ence-plus-noise ratio (SINR) by assuming equal

148 IEEE Network • March/April 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on November 29,2023 at 00:01:53 UTC from IEEE Xplore. Restrictions apply.
 power allocation among users. In addition, the
Training data
new feature vectors F ui are collected to further Class 1
update the dataset and trace the evolution of user Class 2
layout. Distance
Figure 6 presents the average sum rate with K=3
our proposed machine learning framework of K=1
beam allocation vs. the transmit signal-to-noise
ratio (SNR) and size of training set. For compar-
ison, the average sum rate with both optimal ?
exhaustive search and low-complexity beam allo-
cation (LBA) proposed in [13] are also plotted. New feature
It can be seen from Fig. 6b that as the number vector
of training data increases, the average sum rate
achieved by our proposed machine learning
framework increases and gradually approaches
that with the optimal exhaustive search. It can
also be observed from Fig. 6 that with a larg-
er training set, our algorithm outperforms the FIGURE 4. k-NN algorithm.
LBA algorithm proposed in [13], indicating that
our proposed machine learning framework of
resource allocation outperforms conventional
techniques.
Note that for the aforementioned k-NN algo
rithm, the distances between new data and exist-
ing training data are calculated in real time. As a User k
result, with a large number of training data, the θk
computation complexity would become very
high in practical systems. It is therefore important
to design a low-complexity multiclass classifier,

k
which is discussed below.

Research Challenges and Open Issues

Machine learning offers a plethora of opportu-
nities for the research in resource allocation for  0
future wireless communications. There are many
open issues still not being studied that need to be FIGURE 5. Illustration of beam allocation in multiuser massive MIMO systems.
further explored. This section outlines some of the Each cross represents a user. A beam is allocated to the user in the same
most important ones from our viewpoint. color. (rk, qk) denotes the polar coordinate of user k.

L ow-Complexity Classifier in a crowd” scenario focuses on providing reason-

More advanced techniques are required to design able experiences even in crowded areas including
low-complexity multiclass classifiers. One of the stadiums, concerts, and shopping malls. For each
promising techniques is to transform the multi- typical scenario, the hidden common features of
class classification problem into a set of binary user behaviors and wireless propagation environ-
classification problems that are efficiently solved ments may reduce the number of classes, which
using binary classifiers. So far, the support vector can be exploited to further reduce the complexity
machine (SVM) has been regarded as one of the of classifiers. Recently, deep learning has shown
most robust and successful algorithms to design significant advantages in exploiting the hidden
low-complexity binary classifiers, which deter- common features [16, 17].
mines the class of a new feature vector by using
linear boundaries (hyperplanes) in high-dimen- Multi -BS Cooperation
sional spaces. More specifically, the two classes In practical networks, there may be some users
are divided by only a few hyperplanes. According- located at the edge of BS coverage. If the edge
ly, the class is determined based on which sides user is served by only one BS, the signal quality
of hyperplanes the new feature vector falls into. may be very poor due to the long transmission
Compared with the k-NN algorithm, the com- distance. Cooperative transmission among mul-
plexity of SVM-based binary classifiers is very low. tiple nearby BSs has been shown to be able to
For the aforementioned beam allocation exam- improve the edge user’s performance significant-
ple, the total number of active beam solutions is ly [18]. In this case, the radio resources at mul-
2N. In other words, there are at most 2N classes, tiple BSs should be allocated jointly. Compared
which indicates that the complexity is O(2N) for to a single BS scenario, the resource allocation
determining the class of scenario. Meanwhile, the problem of multi-BS cooperative transmissions
complexity of exhaustive search is O(NK). Obvi- requires more information among multiple BSs.
ously, our proposed machine learning framework Accordingly, more attributes in historical data
of resource allocation can approach the optimal will be selected into feature vectors, which
performance of exhaustive search with low com- makes the learning process more complicated
plexity. It is worth mentioning that several typical and challenging. How to use historical data to
scenarios have been defined with different QoS improve the resource allocation with multi-BS
requirements for future fifth generation (5G) com- cooperative transmissions is very challenging and
munications [15]. For example, the “great service needs to be studied.

IEEE Network • March/April 2018 149

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on November 29,2023 at 00:01:53 UTC from IEEE Xplore. Restrictions apply.
 35 24

30 22

20

Average sum rate (b/s/Hz)

Average sum rate (b/s/Hz)

25
18
20
16
15
14
10
12
Exhaustive search K = 4, 8 Exhaustive search
5 10
Proposed Proposed
LBA LBA
0 8
-30 -20 -10 0 10 20 30 40 50 10 102 103 104 105 106
Transmit SNR (dB) Size of training set
(a) (b)

FIGURE 6. Average sum rate with our proposed machine learning framework, optimal exhaustive search, and
the LBA algorithm proposed in [13]. For the employed k-NN algorithm, k = 1: a) N = 16, K = 8, 106
training data; b) N = 8, transmit SNR = 20 dB.

similar historical scenario is adopted to allocate

More attributes in historical data will be selected into feature vectors, which makes the learning the radio resources for the current scenario. An
process more complicated and challenging. How to use historical data to improve the resource example of beam allocation in multi-user massive
MIMO systems is presented to verify that our pro-
allocation with multi-BS cooperative transmissions is very challenging and needs to be studied. posed machine-learning based resource allocation
performs better than conventional methods. In a
nutshell, machine-learning-based resource alloca-
Fast Evolution of Scenarios tion is an exciting area for future wireless commu-
In many real scenarios, the user behaviors nications assisted by cloud computing.
and wireless environments are essentially time
evolving [19]. That is, the characteristic hidden Acknowledgments
in historical scenarios is also dynamic. In most This work has received funding from the Europe-
cases, such evolution is too slow and gentle to an Union’s Horizon 2020 research and innovation
be noticed. Such slow evolution can be traced programme under the Marie Sklodowska-Curie
easily by constantly collecting data and periodi- grant agreement (No. 709291), supported by the
cally updating the dataset for learning. However, National Natural Science Foundation of China
in some special situations, the evolution may be (Nos. 61571115, 61602235, 61701301, and
very sudden and significant. For example, emer- 61701254), the Natural Science Foundation of
gency maintenance is carried out for a very busy Jiangsu Province (No. BK20170901), the open
road, which changes the distribution of user research fund of the National Mobile Communi-
locations and mobility characteristics greatly; a cations Research Laboratory, Southeast University
high building is demolished by blasting, which (No. 2017D06), the open research fund of the
changes the propagation environment signifi- Key Lab of Broadband Wireless Communication
cantly. Since the predictive model is built with and Sensor Network Technology (Nanjing Univer-
outdated historical data, such fast evolution may sity of Posts and Telecommunications), Ministry
result in significant performance loss of resource of Education (Nos. JZNY201701, JZNY201706),
allocation. In machine learning, this issue can and the Key International Cooperation Research
be addressed by updating the predictive model Project (No. 61720106003).
whenever new data are available. However,
since cloud computing is shared by many appli-
References
[1] E. AB, Traffic exploration, interactive online tool; https://
cations, the new data can only be stored tempo- www.ericsson.com/TET/trafficView/loadBasicEditor.erics-
rarily at BSs and forwarded to update the dataset son.
later. How to deal with the fast evolution of [2] E. Castaneda et al., “An Overview on Resource Allocation
Techniques for Multi-User MIMO Systems,” IEEE Commun.
scenarios in resource allocation is a challenging Surveys & Tutorials, vol. 19, no. 1, 2017, pp. 239–84.
topic for future research. [3] Z.-Q. Luo and W. Yu, “An Introduction to Convex Optimiza-
tion for Communications and Signal Processing,” IEEE JSAC,
Conclusion vol. 24, no. 8, Aug. 2006, pp. 1426–38.
[4] https://deepmind.com/.
In future wireless communications, the conven- [5] Z. M. Fadlullah et al., “State-of-the-Art Deep Learning: Evolv-
tional methods of resource allocation are facing ing Machine Intelligence Toward Tomorrow’s Intelligent
great challenges to meet the ever increasing QoS Network Traffic Control Systems,” IEEE Commun. Surveys &
requirements of users with scarce radio resource. Tutorials, vol. 19, no. 4, 2017, pp. 2432–55.
[6] B. Mao et al., “Routing or Computing? The Paradigm Shift
Inspired by the victory of AlphaGo, this article Towards Intelligent Computer Network Packet Transmission
proposes a machine learning framework for Based on Deep Learning,” IEEE Trans. Computers, vol. 66,
resource allocation and discusses how to apply no. 11, Nov. 2017, pp. 1946–60.
supervised learning to extract the similarities [7] S. Bi, R. Zhang, Z. Ding, and S. Cui, “Wireless Communica-
tions in the Era of Big Data,” IEEE Commun. Mag., vol. 53,
hidden in a great amount of historical data on no. 10, Oct. 2015, pp. 190–99.
scenarios. By exploiting the extracted similarities, [8] L. Bottou, “Traffic Engineering”; http://www.cs.princeton.
the optimal or near-optimal solution of the most edu/courses/archive/spring10/cos424/slides/18-feat.pdf.

150 IEEE Network • March/April 2018

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on November 29,2023 at 00:01:53 UTC from IEEE Xplore. Restrictions apply.
 [9] D. Francois, High-Dimensional Data Analysis: From Opti-
mal Metric to Feature Selection, Ph.D. dissertation, Univ.
Ottignies- Louvain-la-Neuve, Belgium, 2007. Since the cloud computing is shared by many applications, the new data can only be stored temporar-
[10] C. H. Papadimitriou and K. Steiglitz, Combinatorial Opti- ily at BSs and forwarded to update dataset later. How to deal with the fast evolutions of scenarios in
mization: Algorithms and Complexity, Courier Corp., 1982.
[11] X.-S. Yang, Nature-Inspired Metaheuristic Algorithms, Luniv- resource allocation is a challenging topic in future research.
er Press, 2008.
[12] E. Alpaydin, Introduction to Machine Learning, MIT Press,
2014; http://ieeexplore.ieee.org/xpl/articleDetails. jsp?ar- is the Lead Guest Editor for an upcoming Special Issue of the
number=6917138 IEEE Journal on Selected Areas in Communications. He is cur-
[13] J. Wang et al., “Low-Complexity Beam Allocation for rently an Editor of IEEE Communications Letters. He has been a
Switched-Beam Based Multiuser Massive MIMO Systems,” Technical Program Committee member of various conferences,
IEEE Trans. Wireless Commun., vol. 15, no. 12, Dec. 2016, including GLOBECOM, ICC, VTC, PIMRC, and so on.
pp. 8236–48.
[14] J. Buter and R. Lowe, “Beam-Forming Matrix Simplifies J in -Y uan W ang [M’16] ([email protected]) received his
Design of Electrically Scanned Antennas,” Electronic Design, Ph.D. degree from the National Mobile Communications
Apr. 1962. Research Laboratory, Southeast University in 2015. He is now
[15] A. Osseiran et al., “Scenarios for 5G Mobile and Wire- a lecturer at the Peter Grünberg Research Center, Nanjing
less Communications: The Vision of the Metis Project,” IEEE University of Posts and Telecommunications, China. His cur-
Commun. Mag., vol. 52, no. 5, May 2014, pp. 26–35. rent research interest is visible light communications. He has
[16] N. Kato et al., “The Deep Learning Vision for Heteroge- authored/coauthored over 70 journal/conference papers. He
neous Network Traffic Control: Proposal, Challenges, and was a Technical Program Committee member of ISWTA 2014 in
Future Perspective,” IEEE Wireless Commun., vol. 24, 2017. Kota Kinabalu, ICCN 2014 in Wuhan, IEEE ICC 2015 in London,
[17] F. Tang et al., “On Removing Routing Protocol from Future WTS 2017 in Chicago, and IEEE/CIC ICCC 2017 in Qingdao.
Wireless Networks: A Real-Time Deep Learning Approach He also serves as a reviewer for many journals, such as IEEE
for Intelligent Traffic Control,” IEEE Wireless Commun., 2017. Transactions on Wireless Communications, IEEE Transactions on
[18] X. Chen, H. H. Chen, and W. Meng, “Cooperative Com- Communications, IEEE Transactions on Vehicular Technology, and
munications for Cognitive Radio Networks: From Theory To Optics Express.
Applications,” IEEE Commun. Surveys & Tutorials, vol. 16, no.
3, pp. 1180–92. H uiling Z hu [M’14] ([email protected]) received her B.S.
[19] A. Rakhlin, “Online Methods in Machine Learning”; http:// degree from Xidian Univeristy and her Ph.D. degree from Tsin-
www.mit.edu/rakhlin/6.883/. ghua University. She is currently a reader (associate professor)
with the School of Engineering and Digital Arts, University
Biographies of Kent. Her research interests are in the area of broadband
Jun-Bo Wang [M’11] ([email protected]) received his B.S. wireless mobile communications, covering topics such as radio
degree in computer science from Hefei University of Technol- resource management, distributed antenna systems, MIMO,
ogy, China, in 2003, and his Ph.D. degree in communications cooperative communications, device-to-device communications,
engineering from Southeast University, Nanjing, China, in 2008. and small cells and heterogeneous networks. She has partici-
From 2008 to 2013, he was with Nanjing University of Aero- pated in a number of European and industrial projects in these
nautics and Astronautics. From 2011 to 2013, he was a post- topics. She held a European Commission Marie Curie Fellowship
doctoral fellow with the National Laboratory for Information from 2014 to 2016. She received the best paper award from
Science and Technology, Tsinghua University, Beijing, China. IEEE GLOBECOM 2011, Houston, Texas. She served as the
He is currently an associate professor with the National Mobile Publication Chair of IEEE WCNC 2013, Shanghai, the Operation
Communications Research Laboratory, Southeast University. Chair of IEEE ICC 2015, London, the Symposium Co-Chair of
Since 2016, he has held a European Commission Marie Curie IEEE GLOBECOM 2015, San Diego, and the Track Co-Chair of
Fellowship and has been a research fellow with the University of IEEE VTC-Spring 2016, Nanjing. She currently serves as an Editor
Kent, United Kingdom. His current research interests are cloud of IEEE Transactions on Vehicular Technology.
radio access networks, mmWave communications, and wireless
optical communications. Min Lin [M’13] ([email protected]) received his B.S. degree
from the National University of Defense Technology, Changsa,
Junyuan Wang [M’15] ([email protected]) received China, in 1993, his M.S. degree from Nanjing Institute of Com-
her B.S. degree in telecommunications engineering from Xid- munication Engineering in 2000, and his Ph.D. degree from
ian University, Xi’an, China, in 2010, and her Ph.D. degree in Southeast University in 2008, all in electrical engineering. He
electronic engineering from the City University of Hong Kong, is a professor with Nanjing University of Posts and Telecom-
China, in 2015. She is currently a lecturer at the Department munications. He has authored or coauthored over 100 papers.
of Computer Science, Edge Hill University, Ormskirk, United His current research interests include wireless communications
Kingdom. She was a research associate at the School of Engi- and array signal processing. He has served as a TPC member
neering and Digital Arts, University of Kent, Canterbury, United of many international conferences, including ICC, GLOBECOM,
Kingdom. Her research interests include performance analysis, and others.
distributed antenna systems, and massive MIMO.
J iangzhou W ang [F’17] ([email protected]) is currently
Yongpeng Wu [SM’17] ([email protected]) received a professor of telecommunications with the School of Engi-
his Ph.D. degree in communication and signal processing from neering and Digital Arts, University of Kent. He has authored
the National Mobile Communications Research Laboratory, more than 200 papers in international journals and confer-
Southeast University in November 2013. During his doctoral ences in the areas of wireless mobile communications, as well
studies, he conducted cooperative research at the Department as three books. He was the Executive Chair of IEEE ICC 2015,
of Electrical Engineering, Missouri University of Science and London, United Kingdom, and the Technical Program Chair
Technology, Rolla. He was the Humboldt Research Fellow and a of IEEE WCNC 2013, Shanghai. He serves/has served as an
senior research fellow with the Institute for Digital Communica- Editor for a number of international journals. For example,
tions, Universitat Erlangen-Nürnberg, Germany. He was a senior he was an Editor of IEEE Transactions on Communications
research fellow with the Institute for Communications Engineer- from 1998 to 2013, and has been a Guest Editor of the IEEE
ing, Technical University of Munich, Germany. He is currently Journal on Selected Areas in Communications, IEEE Commu-
a tenure-track associate professor with the Department of Elec- nications Magazine, and IEEE Wireless Communications. He
tronic Engineering, Shanghai Jiao Tong University. His research is currently an Editor for Science China Information Sciences.
interests include massive MIMO/MIMO systems, physical layer He received the Best Paper Award at IEEE GLOBECOM 2012
security, and signal processing. He was an Exemplary Reviewer and was an IEEE Distinguished Lecturer from 2013 to 2014.
of IEEE Transactions on Communications in 2015 and 2016. He He is a Fellow IET.

IEEE Network • March/April 2018 151

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE MINAS GERAIS. Downloaded on November 29,2023 at 00:01:53 UTC from IEEE Xplore. Restrictions apply.