This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier

Performance Enhancement of mmWave

MIMO Systems using Machine Learning
FAWAD AHMAD1 , WAQAS BIN ABBAS2 , SALMAN KHALID1 , FARHAN KHALID1 (MEMBER,
IEEE), IBRAR KHAN1 , AND FAHAD ALDOSARI3
1
National University of Computer and Emerging Sciences, Islamabad, Pakistan (e-mail: [email protected], [email protected], [email protected],
[email protected] )
2
University of Huddersfield, Huddersfield, UK (e-mail: [email protected])
3
Umm Al-Qura University Meccah KSA (e-mail: [email protected])

ABSTRACT For future wireless communication, millimeter wave (mmWave) coupled with the massive
multiple-input multiple-output (MIMO) are key technologies to overcome the huge data rate requirements.
Although massive MIMO greatly improves the spectral efficiency (SE) of the system, the use of large
antenna arrays not only increases the computational complexity it may also decrease the energy efficiency.
Focusing on improvement in energy efficiency, we propose a low-complexity solution for joint transmit
antenna selection and hybrid precoder design for multi-user mmWave Massive MIMO communication sys-
tems. Particularly, considering a partially connected hybrid architecture, binary particle swarm optimization
and deep neural network (DNN) algorithms are employed for transmit antenna selection and analog precoder
design, respectively. Results show that the proposed solution performs very close, in terms of spectral
efficiency, to the optimal exhaustive search based antenna selection and singular value decomposition based
precoder design with lower computational complexity. It is also shown that the proposed solution also
improves the energy efficiency of the system. Finally, the proposed solution is not very sensitive to channel
imperfections.

INDEX TERMS Spectral Efficiency, Deep Learning, Antenna Selection, Hybrid Precoding, Energy
Efficiency.

I. INTRODUCTION systems operating in a congested microwave band are unable

he number of wireless connected devices such as cell- to meet the future data rate requirements. Hence, researchers
T phones, tablets, laptops, smart-watches are increasing
exponentially every year. According to a global forecast
turned their attention to the millimeter wave (mmWave)
communication spectrum which has a very large available
report by CISCO [1], it is predicted that approximately bandwidth. These electromagnetic waves operate at 30GHz
67% of the total global population will have access to the to 300GHz and have wavelength of λ = 10mm to 1mm.
internet by the year 2023. In 2018, there were 3.9 billion Due to a short wavelength, mmWave suffers from a high
internet users and by 2023 it is estimated that this will path loss, penetration loss, rain, and, oxygen absorption [5].
increase up to 5.3 billion at a compound annual growth rate However, at the same time, the reduction in antenna size
(CAGR) of 6%. Besides this, machine-to-machine (M2M) makes it possible to patch hundreds of antennas on the same
communication, internet-of-things (IoT) for industries and physical size and thus enable the advantages of massive
smart cities, and the use of electronic gadgets such as smart- multiple input multiple output technology which can combat
bracelet, smart-watch, robots, drones, and CCTV cameras the above mentioned losses. Hence, massive MIMO coupled
are all interconnected through wireless networks [2], [3]. with mmWave is a key technology for future generation
These new technologies cause an increase in the data traffic wireless communication systems [6], [7].
which imposes a severe challenge to increase the bandwidth In conventional MIMO systems, fully digital beamforming
to support the future data rate requirements. Future fifth is used in which each antenna is connected with a dedicated
generation (5G) and beyond 5G (B5G) wireless generation RF chain. Digital beamforming can provide benefits such as
systems aim to support a very high data rate at a very spatial multiplexing gain, diversity gain, and beamforming
low latency [4]. Existing fourth generation (4G) wireless capabilities. However, in a massive MIMO system, digital

VOLUME x, 2022 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

beamforming is not suitable owing to have high power con- ing [24], [25]. DL has made it possible to design complex
sumption and complexity [8]. One way to reduce the power communication system models without knowing the actual
consumption is to use analog beamforming which contains a mathematical model which was impossible in the context of
massive number of antennas attached to a single RF chain. existing wireless design principles. Moreover, in traditional
However, it lacks the ability to transmit parallel data streams wireless systems, each module of communication is designed
to provide spatial multiplexing gain. To effectively exploit separately while DL trains all parameters as a whole (end to
digital processing gain and energy consumption for mmWave end network training). They have a very low computational
massive MIMO systems, a hybrid beamforming technique complexity even for highly complex DL models due to
is considered which is a combination of both an analog parallel computing. Hence, DL based solutions are suitable
beamformer and a digital beamformer. There are two types for large scale wireless systems. DL has applications in wire-
of hybrid architectures: (a) a fully connected hybrid architec- less communication such as channel estimation [26], angle
ture in which all antenna elements are connected with each of departure (AOD) estimation [27], analog beam selection
RF chain, and (b) a partially connected hybrid architecture [28]. Designing a hybrid precoder using machine learning is
where an individual RF chain is connected with a subset of also an active area of research. Authors in [29] considered
antennas. In contrast to a fully connected hybrid structure, a a multiuser scenario and proposed DL based solution for a
partially connected hybrid structure is more energy efficient fully connected hybrid architecture. Authors in [30] proposed
and bears less complexity [11]. Focusing on energy efficiency an unsupervised DL based precoding for a fully connected
and low computational complexity, in this paper, a joint hybrid architecture. The authors proposed a convolutional
antenna selection and precoding solution is proposed for a neural network based approach for a partially connected
partially connected hybrid massive MIMO architecture. hybrid structure in [31]. Most of the DL approaches are
Researchers have proposed several precoding techniques focused on a fully connected hybrid architecture whereas
for hybrid architectures. Author in [9] applied hybrid beam- the work on a partially connected hybrid architecture is very
forming using accurate deep learning-based CSI prediction to limited.
have better network connectivity for users even in the pres- A large number of antennas in a massive MIMO system
ence of blockages. Authors in [10] use hybrid beamforming still poses power consumption (PC) and complexity issues
architecture with finite resolution phase shifters to reduce the regardless of using a partially connected hybrid architecture.
cost and power consumption of a traditional digital beam- Antenna selection is an effective way to overcome the power
former for mmWave MIMO systems. Orthogonal Matching consumption and complexity issues [32], [33]. An increase
Pursuit (OMP) which is applicable on a fully connected in the number of antennas results in a higher spectral effi-
hybrid structure is based on spatially sparse precoding as ciency. However, beyond a certain limit, a further increase
discussed in [12]. Alternating minimization techniques for in the number of antennas does not have any significant
phase extraction (PE) and manifold optimization (MO) are improvement in the spectral efficiency (SE) rather only an
discussed in [13] to optimize the analog and digital precoder. increase in power consumption which results may reduce
OMP and MO approaches are optimal in terms of spectral energy efficiency (EE). Authors in [34] showed that 90% of
efficiency but have a high computational complexity which the ergodic rate achieved by a full set of antennas can be
makes them impractical. Authors in [14], [15] proposed a achieved by selecting a subset of antennas. The exhaustive
predefined code book based precoding. Hybrid regularized search algorithm (ESA) discussed in [35] is the optimal way
channel diagonalization (HRCD) precoding scheme is pro- of selecting a subset of antennas. It is an iterative algorithm
posed in [16]. Designing a precoding scheme for a par- that evaluates all combinations of antennas and chooses those
tially connected hybrid structure is also an active area of antennas which maximize the spectral efficiency. However,
research. Singular value decomposition (SVD) based pre- this algorithm is computationally inefficient. The genetic
coder is discussed in [17], [18]. Precoding for a partially algorithm (GA) discussed in [36] [37] follows the natural
connected hybrid architecture using successive interference selection process that mimics biological evolution for both
cancellation (SIC) is discussed in [19], [20], [21]. SIC in- constrained and unconstrained optimization problems. For
tegrates the idea of canceling interference iteratively from practical purposes, there should be a fast selection algorithm
RF chains. Evolutionary algorithms are also utilized for the with minimum computational complexity. Authors in [38]
computation of precoders. Authors [22], [23] utilized the proposed low complexity conventional binary particle swarm
idea of particle swarm optimization (PSO) and artificial bee optimization (BPSO) antenna selection algorithm for a par-
colony (BEE) for a partially connected architecture. From a tially connected hybrid structure.
practical point of view, where there are a massive number In order to reap maximum benefits of a mmWave massive
of antenna elements, precoding with a partially connected MIMO system, precoding to tackle path loss and antenna
hybrid architecture is a feasible option owing to have a low selection to reduce computational complexity and increase
power consumption and complexity. energy efficiency are essential. Authors in [38] proposed
In recent years, researchers are tilted toward deep learn- a single user, low complexity joint antenna selection and
ing (DL) in many areas of research such as visual ob- precoding for a massive MIMO partially connected hybrid
ject recognition, language processing, and speech process- structure. Authors applied conventional BPSO for antenna
2 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

selection and SIC for precoding. In this paper, we have considered that the base station is equipped with Nt transmit
proposed a modified form of conventional BPSO i.e., cyclic antennas serving K number of users by means of NRF
BPSO (CBPSO) for the antenna selection and a fully con- number of RF chains such that Nt >> K. Each RF chain
nected DL based precoding algorithm for a partially con- at the base station is transmitting Ns data streams, where
nected hybrid architecture. In addition, we have also consid- Ns = NRF = K. Each user contains a single receive
ered a multiuser scenario. This work is the enhancement of antenna and RF chain. On the transmitter side, each RF chain
already published work [38]. Work of [38] is enhanced from a is connected with sub antennas M = Nt /NRF as shown
single-user to multiuser, from conventional precoding to DL in the Figure 1. A perfect channel state information (CSI)
based precoding, and from conventional antenna selection is considered between the transmitter and the receiver.
BPSO scheme to CBPSO. Proposed CBPSO has the same In mmWave, the transmission channel can be represented
computational complexity as BPSO, however, has a high by a geometric model having limited scatters, and therefore,
performance in terms of spectral efficiency. The motivation it is considered that the channel matrix for the kth user is Hk
of this work is to devise a computationally efficient scheme comprising of L paths. In this work, we considered a uniform
that improves energy efficiency without compromising the linear array (ULA), however, it can be extended to uniform
spectral efficiency of the system. The DL approach performs planar array (UPA) as well. With ULA, the channel can be
close to optimal and has a lower computational complexity modeled as
than conventional SIC, OMP, and other precoding schemes
L
discussed in [20]. The proposed system improves the spectral X
efficiency, energy efficiency and reduces the computational Hk = γ αk,l ar (Θk,l )aH
t (ϕk,l ) (1)
l=0
complexity of multiuser system.
The main contributions of this work are as follows: where αk,l represents the complex gain associated with the
• Proposed a low complexity and energy efficient joint lth path and the kth user. Θk,l and ϕk,l is the angle of arrival
transmit antenna selection and precoding solution for (AoA) and angle of departure (DoA), respectively. Similarly,
a mmWave massive MIMO partially connected hybrid ar and at represent array responses of the receiver and the
architecture where the transmit antenna selection is transmitter, respectively. Number of paths is represented by
performed using CBPSO and precoding using DL al- L. Normalization factor γ which is a scalar quantity can be
gorithm. written as
• In terms of spectral efficiency, it is shown that the
p
proposed DL based precoding performs very close to γ= (Nt K) /∈ L (2)
SVD based precoding with a very low computational
complexity. where ∈ represents pathloss. The array responses can be
• It is shown that the proposed CBPSO outperforms the written as
conventional BPSO algorithm [38] in terms of spectral
efficiency. 1 2π 2π

• Joint antenna selection and DL based precoding solution

ar (θ) = √ [1, expj λ d sin θ ..... expj(K−1) λ d sin ϕ ]T (3)
K
with imperfect CSI performs close to perfect CSI.
Furthermore, we have also evaluated the performance of
1 2π 2π
the proposed joint transmit antenna selection and DL based at (ϕ) = √ [1, expj λ d sin θ ..... expj(Nt −1) λ d sin ϕ ]T
precoding solution for a multiuser scenario. Nt
(4)
where λ is the wavelength of signal and d is spacing between
A. NOTATIONS AND ORGANIZATION
arrays of antenna. Antenna selection reduces the size of the
Through out this paper matrices and vectors written in by
channel matrix H by selecting only the subset of transmit
boldface upper A and lower case a symbols, respectively.
antennas NtSel having good channel conditions. A set con-
Frobenius norm is donated by ∥ · ∥F . The rest of the draft
taining all transmit antenna combinations is donated
by ψ.
is formulated as: Section II explains the system model.
Total available combinations can be written as NNSel t
. Set
Section III discusses antenna selection techniques. Section t
containing all selected transmit antennas can be written as Ψ.
IV discusses hybrid precoder design. Section V discusses
The new channel matrix between a subset of selected transmit
deep neural network DNN details. Section VII provides
antennas NtSel and receive antennas is donated by HΨ .
simulation results and complexity analysis. Finally, Section
Ns data streams are first precoded by a digital precoder
VIII provides conclusions and future research directions.
and then analog precoding is realized by phase shifters. The
II. SYSTEM MODEL
signal that is to be transmitted can be written as
This section explains the system model of the downlink x = FRF FBB s (5)
single cell multi-user mmWave massive MIMO scenario.
Partially connected hybrid architecture is considered due to where s ∈ CNs ×1 are transmitted symbols, FBB ∈
Ns ×NRF
its low complexity and high energy efficiency [19]. It is C is a digital baseband precoder and FRF ∈ CNt ×K
VOLUME x, 2022 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

FIGURE 1. Multi User System Model

is an RF precoder. Digital precoder is in the form of a matrix single receive antenna. nk is the independent and identically
and is written as: distributed (i.i.d) complex Gaussian noise having N (0, σ 2 ).

fBB1 0 ··· 0
 Finally, the achievable rate R for kth user is given as
 0 fBB2 · · · 0  K
X
FBB =  . (6)
 
. . ..  R= log2 (1 + SIN Rk ) (10)
 .. .. .. .  k=1
0 0 · · · fBBK
III. ANTENNA SELECTION ALGORITHM
Similarly analog precoder can be written as:
  In this section, we have discussed different antenna selection
fRF1 0 ··· 0 algorithms such as ESA, conventional BPSO, CBPSO, and
 0 fRF2 · · · 0  random search (RS). Exhaustive search algorithms (ESA) is
FRF =  . (7)
 
. . .. an optimal antenna selection algorithm that enumerates and
 .. .. .. 
. 
calculates all possible combinations and picks up one which
0 0 ··· fRFK
gives optimal performance. In real time scenarios where
The received signal with antenna selection for the kth user the channel changes quickly, ESA is not a feasible solu-
can be written as tion for antenna selection owing to its high computational
complexity. Therefore, a low complexity solution is required
K
X for practical implementation. BPSO and its variant CBPSO
yk = Hψ
k FRF fBBk sk + Hψ
k FRF fBBn sn + nk (8)
are population based transmit antenna selection algorithms
n=1̸=k that have lower computational complexity than ESA. CBPSO
Ns ×1 performs close to ESA in terms of spectral efficiency with
where yk ∈ C for kth user. For a multi-user scenario,
the same computational complexity as conventional BPSO.
the sum rate can be calculated by computing the signal to
RS technique selects an antenna subset from each RF chain
interference and noise ratio (SINR). It is a ratio of the signal
randomly. Due to random in nature spectral efficiency is
power to the interference plus noise power. SINR for kth user
much lower as compared to other algorithms.
with the reduced channel Hψ can be written as
P ψ 2
K |Hk FRF fBBk | A. CONVENTIONAL BPSO ALGORITHM
SIN Rk = log2 PK ψ
(9)
σ2 P 2 Conventional BPSO is a meta-heuristic population based
+ n=1,n̸=k K |Hk FRF fBBn |
binary version of a particle swarm optimization (PSO) al-
Sel
where Hψ k ∈C
K×Nt
is the channel matrix having num- gorithm. It is motivated by the idea of searching for a food
ber of selected transmit antennas as NtSel and K number of source by a bird swarm or a school of fishes. The objective
4 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

TABLE 1. Conventional BPSO Notation

Notation Description
Is Solution Space
OF Objective Function
Npop Population Size
Niter Number of Iterations
DM Dimension of Particle String
χi Position of Particle at ith Iteration
PBi Best Position of Particles of Swarm
GiB Global Best Position
Ui Velocity of Particle

of PSO is to maximize the objective function or minimize its

cost function. It can solve optimization problems very effi-
ciently. Each particle in PSO is known as the solution and the
solutions (referred here as selected antennas) when grouped
together are known as the population. It is a population based
algorithm in which every particle has its own position known
as local position and velocity. They modify their own position
by using the best position exploited by themselves or by the
particles present in the swarm using their own experience or
by the experience of other particles. The governing equations
for velocity and position update are as follows

uikj = ui−1
kj + α + β (11)
i−1
α = c1 × U (0, 1) × (gkj − xi−1
kj ) (12)

β = c2 × U (0, 1) × (pi−1 i−1

bkj − xkj ) (13)
where U(0, 1) is a uniformly distributed random variable
within the interval (0,1). c1 and c2 are social and cognitive
parameters to control the movement of particles within a
specific direction. FIGURE 2. Flow Chart of CBPSO

PSO finds the optimal solution by combining local search

and global search. In conventional BPSO, a string of ones
and zeros are used to indicate the position of every particle.
TABLE 2. Conventional BPSO Having Random Feed
By bit manipulation, every particle can change its present
position to a new position in the search space. Notations used Feed A1 A2 A3 A4 A5 A6
for conventional BPSO algorithm are mentioned in Table 1. Random Feed 1 0 1 1 0 1
For the ith iteration, position of a particle can be written as Random Feed 1 0 0 1 1 0
Random Feed 1 0 1 0 1 1
χi = [X1i , X2i ...XN i
pop
], where Xki = [xik,1 , xik,2 ....xik,DM ] Random Feed 1 0 1 1 1 0
shows the position of a particle at k index and k = Random Feed 0 0 1 1 1 1
1, 2, 3..Npop .
The objective function determines the best position of a
i
particle at the ith iteration and can be written as Xbk =
i
max(Xk ). Similarly, for all particles in a swarm, the best TABLE 3. Cyclic Feed BPSO
position can be written as PBi = [Xb1 i i
, Xb2 ...XbNi
pop
]. Fur-
thermore, velocity of particles in swarms can be written as Feed A1 A2 A3 A4 A5 A6
Random Feed 1 0 1 1 0 1
U i = [U1i , U2i ...UN
i
pop
]. Cyclic Shift 1 1 0 1 1 0
Cyclic Shift 0 1 1 0 1 1
B. CYCLIC BPSO ALGORITHM Cyclic Shift 1 0 1 1 0 1
Cyclic Shift 1 1 0 1 1 0
Cyclic BPSO (CBPSO) is a modified version of conventional Cyclic Shift 0 1 1 0 1 1
BPSO. The flow chart of CBPSO is given in Figure 2. In Random Feed 0 1 1 1 0 1
the conventional BPSO, the initial population generated by Cyclic Shift 1 0 1 1 1 0
the algorithm is random ones and zeros as shown in Table 2. Cyclic Shift 0 1 0 1 1 1
This table illustrates that on every iteration, a random feed
VOLUME x, 2022 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

is generated and there may be a chance one of the antennas digital precoder can be written as F = FRF FBB .
remains deactivated all the time. For example, antenna A2 Furthermore, the hybrid precoder design has to meet the
remains deactivated for the whole time and that may be the following constraints:
best antenna. This results in a sub optimal spectral efficiency. • C1: All non-zero elements of the diagonal matrix should
Therefore, to solve this issue, a modification is required to the have the same amplitude.
conventional BPSO. In a modified version, i.e., CBPSO, the • C2: Forbenious norm of hybrid matrix F should have
initial feed is generated randomly the same as in conventional ∥F∥2F ≤ Ns . This is also known as the power constraint.
BPSO, however, in the next iteration previously generated Designing a hybrid precoder is an optimization problem.
feed is given a cyclic shift as shown in Table 3. This table According to literature [12] Euclidean distance between
shows that the first feed is random and in the next feed cyclic Fopt and F should be minimum.
shift is given till the last available antenna. After that again
random feed is generated and the same process is followed.
This ensures that every antenna gets a chance to be active. arg min ∥Fopt − FRF FBB ∥2F
FRF FBB
Complete steps of how BPSO is converted in to CBPSO is
given in Algorithm 1.
subject to ∥F∥2F ≤ Ns ,
Algorithm 1 CBPSO Algorithm
1: Initialize Number of Iterations,Agents, Antennas

2: (Step 1) Initialize position and velocity to particles ran- It has been shown that minimizing the objective function
domly and perform cyclic permutation as described in (shown above) approximately leads to the maximization of
Table 3 . At i = 0 each particle has its current position the spectral efficiency. The optimal solution for the above
0
as best position as Pb1 = X10 mentioned problem can be obtained by singular value decom-
position (SVD) performed on a reduced channel matrix Hϕ .
3: (Step 2) Update iteration counter Although SVD gives an optimal solution it is a computation-
4: (Step 3) Updating the velocity and position of each ally complex technique. To address this, a deep learning algo-
particle as given by: rithm can be applied. It is highlighted that the deep learning
algorithm can be trained in an offline mode. Furthermore, the
uikj = ui−1
kj + α + β (14) DNN model is trained with different channel realizations and
i−1 their corresponding precoder using SVD. After that, during
α = c1 × U (0, 1) × (gkj − xi−1
kj ) (15)
the online mode, trained network is given with the channel
β = c2 × U (0, 1) × (pi−1
bkj − xi−1
kj ) (16) realization to predict the precoder efficiently.

5: (Step 4) Update position of particle V. DNN TRAINING AND DATASET GENERATION

6: (Step 5) According to fitness function F calculates the DNN is supervised learning and it requires several input
population. Store these values in temporary variables as and output combinations to train its model. Input X to the
T emp
PB1 T emp
= [Pb1 T emp
, Pb2 . . . . . . PbTNemp ] network is of size K × Nt × c where c = 3 channels. The
pop
first channel to input is the imaginary value of Hψ , second
7: (Step 6) If fitness function converges then terminate and third are real and absolute respectively.
otherwise go to next step
X(K : Nt : 1) = ℑ(Hψ ) (17)
8: (Step 7) Update personal best position as: ψ
T emp
If F (Pbk
(i−1)
) > F (Pbk ) X(K : Nt : 2) = ℜ(H ) (18)
i T emp ψ
Pbk = Pbk X(K : Nt : 3) = abs|H | (19)
Else
i (i−1) To train the network, the training dataset is prepared us-
Pbk = Pbk
ing several unknown channel realizations NR . Against each
9: (Step 8) Update global best position as GiB = channel realization, NL number of noisy channel realizations
i
maxF (Pbk ) corrupted by synthetic noise is generated [29]. Hence, the
10: (Step 9) If program does not terminate then repeat from total number of realizations generated will be K × Nt ×
step 1 again. NR NL . Singular value decomposition (SVD) is applied to
these channel realizations to generate FRF precoder. Against
each FRF , RF label is generated in order to establish a link
between input and output. Channel matrix is the input to the
IV. PARTIALLY CONNECTED HYBRID PRECODER DNN and the FRF label will be output.
After performing antenna selection, reduced channel Hψ is The proposed network is shown in Figure 3 which contains
precoded in order to maximize the sum rate R. A hybrid input, fully connected, and output layers. There are two
precoder which is a combination of an analog precoder and a fully connected layers having 512 nodes each. The activation
6 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

be increased to further tune up the results but this also in-

creases the computational complexity. The antenna selection
is also performed using either the Genetic Algorithm (GA)
or conventional BPSO and the results are compared with
the CBPSO algorithm. After the antenna selection, the DNN
scheme is employed to predict the hybrid precoder efficiently.
Comparison is done with cyclic BPSO antenna selection
along with the optimal precoder design using SVD. For
comparison, the following combinations of antenna selection
and precoding algorithms are considered.
• Cyclic BPSO AS + OPT: Cyclic BPSO antenna selec-
tion along with SVD based optimal precoder.
• Cyclic BPSO AS + DBF: Cyclic BPSO antenna selec-
tion along with deep learning based hybrid precoder.
FIGURE 3. DNN Architecture • Conventional BPSO AS + DBF Conventional BPSO
antenna selection along with deep learning based hybrid
precoder.
function used in this network is ReLU. Output layer is of
• Random AS + DBF Instead of BPSO selection, a
size Nt × 1 and is a regression layer that produces non zero
random selection is performed along with deep learning
elements FRF precoder. After the predicting stage, the block
based hybrid precoder.
diagonal structure of FRF is obtained by appending zeros
at appropriate positions. Digital precoder FBB is obtained
by the equivalent channel approach. Simulation is done us- A. SPECTRAL EFFICIENCY WITH VARIABLE TRANSMIT
ing MATLAB simulation software. Network parameters are ANTENNAS
updated by a stochastic gradient descent (SGD) algorithm. In this subsection, the performance of CBPSO, conventional
Negative log-likelihood or cross entropy is used as a loss BPSO, GA and random antenna selection along with the
function. The learning rate is kept 0.005 with a batch size of deep learning based hybrid precoder is carried out to in-
500 along with 200 epochs. From total data, 30% data is kept vestigate the effect of a variable number of transmit anten-
for the validation stage and 70% data is kept for the training nas while considering different antenna selection techniques.
stage. Validation data is used to check the performance of the Furthermore, DBF is compared with the conventional SVD
DNN model, 100 Monte Carlo simulations are performed to approach. Different scenarios are considered by increasing
analyze the performance and is discussed in the next section. the total number of transmit antennas to 32, 64, and 256 while
keeping half of them active by applying an antenna selection
VI. COMPLEXITY ANALYSIS algorithm.
The computational complexity in terms of complex addi- In the first scenario, there are Nt = 32 transmit anten-
tions and multiplications is considered. CBPSO and con- nas connected with NRF = 2 RF chains. Each RF chain
ventional BPSO both require 1/3 × Npop × Nt3 complex contains M = 16 transmit antennas. The antenna selection
additions and multiplications. The number of complex de- algorithm selects NtSel = 8 antennas from each RF chain
terminants required by ESA are NNSel


t
. For n × n matrix, and results in activating 16 out of the 32 transmit antennas.
t
the number of complex operations required are 1/3 × n3 . The number of users considered is K = NRF . Result in
Similarly, deep learning based hybrid precoding requires Figure 4 shows that both variant of BPSO outperforms GA
O(L2 Ns Nt2 ) computations and SIC based precoding re- and random antenna selection algorithms, however, among
quires O(2Nr Ns (Nt2 (K + Ns Nr ))) computations. There- the BPSO variants, CBPSO has higher performance than the
fore, the proposed joint solution of CBPSO and deep learning conventional BPSO in terms of spectral efficiency because
has lower complexity as compared to other techniques. of having cyclic feed rather than random feed. Random
antenna selection has a lower performance due to the random
VII. SIMULATION RESULTS AND PERFORMANCE nature of selection and the performance gap increases as
ANALYSIS SNR increases. Furthermore, the CBPSO antenna selection
In this section, we will present the simulation results for joint technique along with the deep learning based precoder is
antenna selection and precoding. Simulations are performed compared with SVD optimal precoder. The performance gap
to evaluate the performance in terms of spectral efficiency remains the same between CBPSO + DBF and CBPSO +
with respect to the, 1) change in the total number of transmit OPT at both the low and high SNR. Results clearly show that
antennas, and 2) change in the number of selected transmit among antenna selection techniques CBPSO performs better
antennas, while keeping the number of RF chains and a total in terms of spectral efficiency. SVD based precoding is an
number of transmit antennas fixed. For CBPSO, the number optimal precoding scheme, however, SVD based scheme is
of iterations and agents is considered to be 20 each. They can not feasible for practical implementation due to its high com-
VOLUME x, 2022 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

putational complexity. The deep learning approach performs

close to SVD with a very low computational complexity.
In the second scenario, there are Nt = 64 total available
transmit antennas connected with NRF = 8 RF chains.
Each RF chain contains M = 8 transmit antennas. The
antenna selection algorithm selects NtSel = 4 antennas
from each RF chain. The number of users considered is
K = NRF . The proposed CBPSO outperforms all other
options. Moreover, with the increase in total available trans-
mit antennas from 32 to 64 and selecting half of them active
through the antenna selection CBPSO algorithm, the spectral
efficiency increases. At SNR = 15 dB, when compared
with the first scenario where total available transmit antennas
were Nt = 32 proposed solution has a spectral efficiency
of around 9 bps/Hz whereas in the second scenario where
total available transmit antennas were Nt = 64 spectral
efficiency of system increases to 15 bps/Hz. In this scenario,
FIGURE 4. Simulations with Nt = 32, NRF = K = 2, NtSel = 4
the proposed solution is also compared with GA + DBF.
Performance in terms of the spectral efficiency of GA is
lower than both conventional BPSO and CBPSO. Proposed
solution CBPSO + DBF performs close to optimal as shown
in Figure 5. Owing to the lower performance of GA antenna
selection, comparison among conventional BPSO, CBPSO
and random selection techniques is performed in rest of the
simulation results.
Similarly, in the third scenario, there are Nt = 256 total
available transmit antennas connected with NRF = 16 RF
chains. Each RF chain contains M = 16 transmit antennas.
Antenna selection algorithm selects NtSel = 8 antennas from
each RF chain and results in activating 128 antennas out of
256. Number of users considered are K = NRF . As shown
in Figure 6, with total available transmit antennas Nt = 256
spectral efficiency increases to 20 bps/Hz at SNR = 15 dB.
From all the above scenarios, the spectral efficiency of
the proposed algorithm increases as the number of antennas
increases with no effect on the performance of the proposed
FIGURE 5. Simulations with Nt = 64, NRF = K = 8, NtSel = 4
algorithm. Moreover, the CBPSO based antenna selection
algorithm outperforms GA based antenna selection as well.

B. SPECTRAL EFFICIENCY WITH VARIABLE SELECTED

TRANSMIT ANTENNAS
In this subsection, the spectral efficiency analysis against
SNR is performed in which selected antennas from each RF
chain are increased while keeping the total RF chains and
number of the available transmit antennas fixed. Different
scenarios are considered in which the number of selected
antennas is increased.
In the first scenario, there are Nt = 64 transmit antennas
connected with NRF = 2 RF chains. Each RF chain contains
M = 32 transmit antennas. The antenna selection algorithm
selects NtSel = 4 antennas from each RF chain and results
in activating 8 antennas out of 64. The number of users
considered is K = NRF . CBPSO based selection scheme
is outperforming all other selection schemes. Specifically,
at SNR = 15 dB, the spectral efficiency of the proposed FIGURE 6. Simulations with Nt = 256, NRF = K = 16, NtSel = 8
algorithm is around 9 bps/Hz as shown in Figure 7.
8 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

In the second scenario, there are Nt = 64 transmit an-

tennas connected with NRF = 2 RF chains. Each RF chain
contains M = 32 transmit antennas. The antenna selection
algorithm selects NtSel = 8 antennas from each RF chain
and results in activating 16 antennas out of 64. The number
of users considered is K = NRF . At SNR = 15 dB, the
spectral efficiency of the proposed algorithm is around 9.5
bps/Hz. The spectral efficiency of the system increases as
the number of selected antennas is increased from 4 to 8 as
shown in Figure 8.
Similarly, in the third scenario, there are Nt = 64 transmit
antennas connected with NRF = 2 RF chains. Each RF chain
contains M = 32 transmit antennas. The antenna selection
algorithm selects NtSel = 16 antennas from each RF chain
and results in activating 32 antennas out of 64. The number
of users considered is K = NRF . At SNR = 15 dB, the
spectral efficiency of the proposed algorithm is around 9.8
FIGURE 7. Simulations with Nt = 64, NRF = K = 2, NtSel = 4
bps/Hz. This shows that the spectral efficiency of the system
increases with the increase in the number of selected anten-
nas. Results also show that the performance gap of CBPSO
and conventional BPSO also increases with an increase in
selected antennas as shown in Figure 9. Results from the
above scenarios show that by keeping the number of total
transmit antennas fixed, if we start increasing the number
of selected antennas from each RF chain, spectral efficiency
goes on increasing. However, at a certain level of increasing
selected transmit antennas while approaching total transmit
antennas, the increase in spectral efficiency is small. At this
level, a further increase in the number of selected antennas
only imparts an increase in power consumption with a slight
increase in spectral efficiency. From the results, it is clear
that by increasing the selected number of antennas from 4
to 8 from each RF chain, the increase in performance is
prominent. However, an increase in selected antennas from
8 to 16 from each RF chain results in a slight increase
in spectral efficiency. This slight increase only imparts an
FIGURE 8. Simulations with Nt = 64, NRF = K = 2, NtSel = 8
increase in power consumption along with computational
complexity.

C. ENERGY EFFICIENCY (EE) ANALYSIS

Energy efficiency analysis is performed in this subsection. It
is calculated using the energy consumption model given in
[38].
SE
η= (20)
Ptotal
where Ptotal = Pt + NRF PRF + NP S PP S is overall total
power consumption. NP S , NRF is the number of phase
shifters and the total number of RF chains, respectively. Pt ,
PRF , PP S is transmitted power, power used by each RF
chain, and power consumed by the phase shifter, respectively.
Energy efficiency analysis is performed with antenna selec-
tion and without antenna selection of deep learning based
precoder. Without antenna selection, a scenario is considered
in which NRF = K = 8, Nt = 32 and total active an- FIGURE 9. Simulations with Nt = 64, NRF = K = 2, NtSel = 16
tennas are 32, however, with cyclic BPSO antenna selection,
VOLUME x, 2022 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

NRF = K = 8, Nt = 64 and total active antennas are 32 by

selecting 4 antennas from each RF chain. Results in Figure
10 show that the energy efficiency of cyclic BPSO is higher
than without antenna selection.

FIGURE 11. Comparison of Perfect Channel State Information with Imperfect

Channel State Information

ACKNOWLEDGMENT
The authors would like to thank the Deanship of Scientific
FIGURE 10. Energy Efficiency Analysis
Research at Umm Al-Qura University for supporting this
work by Grant Code: (22UQU4210128DSR02)

D. SPECTRAL ANALYSIS WITH IMPERFECT CSI REFERENCES

In this subsection, we perform an analysis of our proposed [1] Cisco Forecast and Trends, 2018–2023, White Paper c11-741490 [Online].
algorithm in case of imperfect channel state information Available: https://www.cisco.com, Accessed on: Jan. 26, 2022
[2] N. Ahmed, D. De, and I. Hussain, "Internet of Things (IoT) for smart
(CSI). Let H represent the estimated channel matrix with precision agriculture and farming in rural areas,” IEEE Internet of Things,
imperfect CSI and that can be modeled as [19]: vol. 5, pp. 4890-4899, Dec. 2018.
[3] Y. Mehmood, N. Haider, W. Afzal, U. Younas, I. Rashid and M. Imran,
√ "Impact of Massive MIMO systems on future M2M communication" IEEE
H̃ = αH + 1 − αE (21) 11th Malaysia International Conference on Communications (MICC),
2013, pp. 534-537.
[4] P. Yang, Y. Xiao, M. Xiao and S. Li, "6G Wireless Communications:
where H̃ is a channel matrix with imperfect CSI. E represents Vision and Potential Techniques," IEEE Network, vol. 33, no. 4, pp. 70-75,
the error matrix having independent and identically dis- July/August 2019, doi: 10.1109/MNET.2019.1800418.
tributed Gaussian. Figure 11 shows an analysis of imperfect [5] Alsharif, M.H.; Nordin, R, "A evolution towards fifth generation (5G)
wireless networks: Current trends and challenges in the deployment of
CSI for mmWave massive MIMO channel. With imperfect millimetre wave, massive MIMO, and small cells", Telecommun. Syst.
channel conditions, the proposed solution still performs close 2017, 64, 617–637.
to channel conditions having perfect CSI. [6] S. Talwar, D. Choudhury, K. Dimou, E. Aryafar, B. Bangerter, and K.
Stewart, “Enabling technologies and architectures for 5G wireless,” Proc.
IEEE MTT-S International Microwave Symposium (IMS), pp.1-4, Jun.
VIII. CONCLUSION 2014.
In this work, we presented joint CBPSO based transmit an- [7] L. Wei, R. Q. Hu, Y. Qian, and G. Wu, “Key elements to enable millimeter
wave communications for 5G wireless systems,” IEEE Wireless Commun.,
tenna selection and deep learning based hybrid precoder for vol. 21, no. 6, pp. 136–143, Dec. 2014.
a multiuser mmWave massive MIMO partially connected hy- [8] S. Sun, T. S. Rappaport, R. W. Heath, A. Nix, and S. Rangan, “MIMO
brid architecture. This proposed solution efficiently reduced for millimeter-wave wireless communications: beamforming, spatial mul-
tiplexing, or both?” IEEE Communications Magazine, vol. 52, no. 12, pp.
the computational complexity and increased the spectral and 110–121, Dec. 2014.
energy efficiency of the system. Performance of CBPSO [9] B. Alsunbuli, et al. "Hybrid beamforming with relay and dual-base stations
is higher than conventional BPSO scheme with the same blockage mitigation in millimetre-wave 5G communication applied in
(VIOT)." Computers and Electrical Engineering 100 (2022): 107953.
computational complexity and is close to optimal selection [10] A. Lu, et al.,. "Reduced Complexity Hybrid Beamforming for Time-
scheme i.e., exhaustive search. The deep learning approach Varying Channels in Millimeter Wave MIMO Systems." Wireless Personal
performs close to the SVD based optimal precoding scheme Communications (2022): 1-20.
[11] Abbas, Waqas Bin, and Michele Zorzi. "Towards an appropriate receiver
having much lesser computational complexity than the con- beamforming scheme for millimeter wave communication: A power con-
ventional precoding scheme. Future extensions of this work sumption based comparison." European Wireless 2016; 22th European
may include, a) antenna selection at both the transmitter and Wireless Conference. VDE, 2016.
[12] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, R. Heath, Spatially
the receiver using a machine learning approach, and b) joint sparse precoding in millimeter wave MIMO systems, IEEE Trans. Wirel.
solution for antenna selection, user selection, and precoding. Commun. (3) (2014) 1499–1513

10 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

[13] X. Yu, J. Shen, J. Zhang, K.B. Letaief, Alternating minimization algorithm [35] A. Molisch, M. Win, Y. Choi, J. Winters, Capacity of MIMO systems with
for hybrid precoding in millimeter wave MIMO systems, IEEE J. Sel. Top. antenna selection, IEEE Trans. Wirel. Commun. 4 (4) (2005) 1759–1772
Signal Process. 10 (3) (2016) 485–500. [36] H.-Y. Lu and W.-H. Fang, “Joint transmit/receive antenna selection in
[14] He, Shiwen, et al. "Codebook-based hybrid precoding for millimeter MIMO systems based on the priority-based genetic algorithm,” IEEE
wave multiuser systems." IEEE Transactions on Signal Processing 65.20 Antennas and Wireless Propagation Lett., vol. 6, pp. 588-591, 2007.
(2017): 5289-5304. [37] Xin-SheYang, Book: Nature-Inspired Optimization Algorithms 2014,
[15] Kim C., Son J.S., Kim T., Seol J.Y. On the hybrid beamforming with Pages 77-87
shared array antenna for mmWave MIMO-OFDM systems; Proceedings [38] Salman Khalid, Rashid Mehmood, Waqas Bin Abbas, Farhan Khalid,
of the 2014 IEEE Wireless Communications and Networking Conference Muhammad Naeem, Joint transmit antenna selection and precoding for
(WCNC); Istanbul, Turkey, 6–9 April 2014. millimeter wave massive MIMO systems, Physical Communication, Vol-
[16] Khalid F. Hybrid Beamforming for Millimeter Wave Massive Multiuser ume 42, 2020.
MIMO Systems Using Regularized Channel Diagonalization. IEEE Wirel.
Commun. Lett. 2019;8:705–708. doi: 10.1109/LWC.2018.2886882
[17] Li, Ang, and Christos Masouros. "Hybrid precoding and combining de-
sign for millimeter-wave multi-user MIMO based on SVD." 2017 IEEE
International Conference on Communications (ICC). IEEE, 2017.
[18] Zhang, Dan, et al. "SVD-based low-complexity hybrid precoding for
millimeter-wave MIMO systems." IEEE Communications Letters 22.10 FAWAD AHMAD completed Bachelor’s degree in
(2018): 2176-2179. Electrical Engineering from Air University (AU),
[19] Gao X., Dai L., Han S., Chih-Lin I., Heath R.W. Energy-efficient hybrid Islamabad in 2014, and completed a Master’s de-
analog and digital precoding for mmWave MIMO systems with large gree in Electrical Engineering from the National
antenna arrays. IEEE J. Sel. Areas Commun. 2016;34:998–1009 University of Computer and Emerging Science
[20] Dai L., Gao X., Quan J., Han S., Chih-Lin I. Near-optimal hybrid analog (NUCES), Islamabad in 2021. During Masters,
and digital precoding for downlink mmWave massive mimo systems, Pro- his research focused on hybrid precoder design
ceedings of the 2015 IEEE International Conference on Communications for multi-user millimeter-wave massive MIMO
(ICC); London, UK. 8–12 June 2015; pp. 1334–1339. wireless communication systems. His research in-
[21] Zhang D., Wang Y., Li X., Xiang W. Hybridly Connected Structure for terests include millimeter-wave communication,
Hybrid Beamforming in mmWave Massive MIMO Systems. IEEE Trans. machine learning, beamforming, and energy efficiency systems for next 5G
Commun. 2018;66:662–674. doi: 10.1109/TCOMM.2017.2756882
wireless systems.
[22] L. Hou, Y. Liu, X. Ma, Y. Li, S. Na and M. Jin, "Particle Swarm Op-
timization Inspired Low-complexity Beamforming for mmWave Massive
MIMO Systems," 2020 IEEE Wireless Communications and Networking
Conference (WCNC), 2020, pp. 1-6
[23] Khalid, S.; Abbas, W.B.; Kim, H.S.; Niaz, M.T. Evolutionary Algorithm
Based Capacity Maximization of 5G/B5G Hybrid Precoding Systems.
Sensors 2020, 20, 5338. https://doi.org/10.3390/s20185338 WAQAS BIN ABBAS received the Bachelors and
[24] Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. the Masters degree from National University of
7553, pp. 436–444, 2015 Computer and Emerging Sciences (NUCES), Is-
[25] D. Yu and L. Deng, “Deep learning and its applications to signal and in- lamabad, Pakistan in 2008 and 2012, respectively,
formation processing [exploratory DSP],” IEEE Signal Process. Mag.,vol. and the Ph.D. in Information Engineering in 2017
28, no. 1, pp. 145–154, Jan. 2011
from the University of Padova, Italy. Currently,
[26] H. Ye, G. Y. Li, and B. Juang, “Power of deep learning for channel esti-
he is associated with University of Huddersfield,
mation and signal detection in OFDM systems,” IEEE Wireless Commun.
UK. During Masters, his research was focused in
Lett., vol. 7, no. 1, pp. 114–117, Feb. 2018
[27] H. Huang, J. Yang, H. Huang, Y. Song, and G. Gui, “Deep learning for
underwater wireless communication, while during
super-resolution channel estimation and DOA estimation based massive Ph.D., his research was mostly focused on energy
MIMO system,” IEEE Trans. Veh. Technol., vol. 67, no. 9, pp. 8549–8560, efficiency in wireless networks. His current research interests include energy
Sep. 2018 efficiency in 5G millimeter wave cellular networks, MIMO communication
[28] Y. Long, Z. Chen, J. Fang, and C. Tellambura, ”Data-driven-based analog and multi-hop wireless networks.
beam selection for hybrid beamforming under mm-Wave channels,” IEEE
Journal of Selected Topics in Signal Processing, vol. 12, no. 2, pp. 340-
352, 2018
[29] A. M. Elbir and A. K. Papazafeiropoulos, "Hybrid Precoding for Multiuser
Millimeter Wave Massive MIMO Systems: A Deep Learning Approach,"
IEEE Transactions on Vehicular Technology, vol. 69, no. 1, pp. 552-563,
Jan. 2020, doi: 10.1109/TVT.2019.2951501. SALMAN KHALID received the bachelor’s de-
[30] H. Hojatian, J. Nadal, J. -F. Frigon and F. Leduc-Primeau, "Unsupervised gree from the University of Engineering and Tech-
Deep Learning for Massive MIMO Hybrid Beamforming," IEEE Trans- nology (UET), Taxila, Pakistan in 2008, master’s
actions on Wireless Communications, vol. 20, no. 11, pp. 7086-7099, Nov. degree from the Center of Advanced Studies in En-
2021, doi: 10.1109/TWC.2021.3080672. gineering (CASE), Islamabad, Pakistan, in 2014,
[31] Salman Khalid, Waqas bin Abbas, Farhan Khalid "Deep Learning based and currently pursuing the Ph.D. degree in elec-
Joint Precoder Design and Antenna Selection for Partially Connected trical engineering from the National University
Hybrid Massive MIMO Systems", arXiv:2102.01495v1 of Computer and Emerging Sciences (NUCES),
[32] Salman Khalid, Rashid Mehmood, Waqas Bin Abbas, Farhan Khalid, Islamabad, Pakistan. During master’s degree, his
Muhammad Naeem,Joint transmit antenna selection and precoding research was focused on wireless communication,
for millimeter wave massive MIMO systems,Physical Communica- compressed sensing and digital signal processing, while during Ph.D. de-
tion,Volume 42, 2020,101137,ISSN 1874-4907, gree, his research is mostly focused on hybrid precoder design, resource
[33] Li, Hui, Lingyang Song, and Mérouane Debbah. "Energy efficiency of allocation in MIMO communication networks and energy efficiency in wire-
large-scale multiple antenna systems with transmit antenna selection." less networks. His current research interests include energy efficiency in 5G
IEEE Transactions on Communications 62.2 (2014): 638-647. millimeter-wave cellular networks, MIMO communication, and millimeter-
[34] S. Asaad, A. M. Rabiei and R. R. Müller, "Massive MIMO With Antenna wave hybrid beamforming.
Selection: Fundamental Limits and Applications," IEEE Transactions on
Wireless Communications, vol. 17, no. 12, pp. 8502-8516, Dec. 2018.

VOLUME x, 2022 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3190388

Author et al.

FARHAN KHALID received the B.S. degree in

electronic engineering from Ghulam Ishaq Khan
Institute of Engineering Sciences and Technology,
Topi, Pakistan, in 1999, the M.Sc. degree in elec-
trical engineering from Blekinge Institute of Tech-
nology, Karlskrona, Sweden, in 2005, and the Dr.-
Ing. (Ph.D.) degree in electrical engineering from
the University of Stuttgart, Stuttgart, Germany, in
2012. Additionally, he holds a master’s degree in
telecommunication management awarded by the
Institut National des Telecommunications, France, in 2001. He started his
professional career as an engineer, working for various organizations in
the telecom sector. Since 2006, he has been involved with academia and
research in the domains of digital communications and signal processing
and is currently associated with the Department of Electrical Engineering,
National University of Computer and Emerging Sciences (NUCES), Islam-
abad, Pakistan. His current research interests lie in the area of wireless com-
munications, with particular focus on transceiver design and optimization
for multiuser massive MIMO and millimeter wave systems.

IBRAR KHAN received his B.S degree in en-

gineering from Comsats Institute of Information
Technology, Abbottabad, Pakistan in 2007, the
M.S degree in Telecommunication Engineering
from National University of Computer and Emerg-
ing Sciences(NUCES), Islamabad, Pakistan in
2009. He is currently PhD scholar and assistant
professor at the department of electrical engineer-
ing, NUCES FAST Islamabad Pakistan. His re-
search interests include IoT security at physical
layer, Memristive Aided Logic and Networks, and AI techniques for IoT
security

FAHAD ALDOSARI has Ph.D. in computer net-

works and M.Sc. in mobile and satellite com-
munications from Bradford university; and B.Sc.
in computer engineering from KFUPM. He had
served as the dean of computer and information
systems college and the dean of IT in Umm
AlQura university in Saudi Arabia. His research
focuses on cloud computing, computer networks
and digital transformation.

12 VOLUME x, 2022

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/