SPECIAL SECTION ON INNOVATIVE TRENDS IN 6G ECOSYSTEMS

Received 29 March 2023, accepted 18 April 2023, date of publication 12 May 2023, date of current version 18 May 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3275755

RIS-Assisted Multi-User MIMO Systems Exploiting

Extreme Learning Machine
M. REZWANUL MAHMOOD 1, MOHAMMAD ABDUL MATIN 1, (Senior Member, IEEE),
AND SOTIRIOS K. GOUDOS 2, (Senior Member, IEEE)
1 Department of Electrical and Computer Engineering, North South University, Dhaka 1229, Bangladesh
2 ELEDIA@AUTH, Department of Physics, Aristotle University of Thessaloniki, 54124 Thessaloniki, Greece
Corresponding author: Mohammad Abdul Matin ([email protected])

ABSTRACT Massive multiple-input multiple-output (mMIMO), assisted by reconfigurable intelligent

surface (RIS), can ensure reliable and energy-efficient data transmission. However, the receiver design for
large-scale networks based on traditional mathematical approaches requires complex statistics. Therefore,
in this paper, machine learning (ML) approaches are investigated to design receivers for the RIS-assisted
multi-user MIMO (muMIMO) systems to avoid complicated channel information requirements. Extreme
learning machine (ELM) is an effective ML tool for MIMO receiver design because it simplifies the learning
process. However, the learning performance of the ELM can get affected by the random choice of its hidden
layer size. To address this issues, this paper proposes an incremental ELM (I-ELM) based receiver for the
RIS-mu-MIMO system. The proposed receiver computes the weights between the hidden and the output
layer based on the automated incremental addition of hidden neurons and provided conditions. The suggested
receiver is contrasted with the multilayer perceptron (MLP), conventional ELM, and minimum mean square
error (MMSE) receivers. The simulation results show that the throughput performance of the proposed
receiver is satisfactory.

INDEX TERMS Extreme learning machine, intelligent surface, machine learning, massive MIMO, multi-
layer perceptron.

I. INTRODUCTION provides the opportunity to improve the energy of received

Massive multiple-input multiple-output (mMIMO) is signal as well as coverage area, and prevent signal leakage
acknowledged as a key enabler for future wireless network to the undesired user. Compared to conventional relays, such
systems [1], [2]. The deployment of large number of anten- as amplify-and forward (AF), decode-and-forward (DF), RIS
nas in mMIMO provides improved spectral efficiency and consumes less amount of power to amplify the signal without
massive IoT device connectivity with high throughput. These introducing any noise ideally [4]. Thus, RIS assisted MIMO
large antenna arrays can assist in producing directional beams system provides cost and energy efficient 6G network sys-
that can reduce propagation loss at high frequency spectrum, tems. Our research is focused on to design a receiver for RIS-
but the directed beams suffer from blockage. To overcome multi-user MIMO (RIS-muMIMO) system.
this phenomenon, additional relays and base stations (BSs)
can be implemented, which in turn increase the implementa- A. RELATED WORKS
tion cost and power consumption. In general, channel estimation (CE) is required to design
In recent times, reconfigurable intelligent surface (RIS) has a MIMO receiver [5]. In an RIS-muMIMO system,
attracted a lot of attention due to its capability in changing CE becomes more crucial since its mathematical modeling
EM wave properties, such as amplitude and phase [3]. This involves amplitude attenuation and phase shifts [3]. As a
result, much research efforts have been provided to estimate
The associate editor coordinating the review of this manuscript and RIS-MIMO channels that would aid in designing a receiver.
approving it for publication was Vladimir Poulkov. A two stage channel estimation approach is presented in [6],
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 47129
 M. R. Mahmood et al.: RIS-Assisted Multi-User MIMO Systems Exploiting ELM

where first stage of the proposed CSI estimation approach applications, which are observed in [20], [21], and [22].
uses the traditional uplink training method to determine the In [20], the ELM is trained offline first with the training
direct MIMO channel between the access point (AP) and data and then deployed online. However, the ELM algorithm
the users and the second stage of the proposed approach can be both trained and deployed online, which is demon-
estimates the AP-RIS channel and the RIS-user channel. strated in [21]. Least square (LS)-based channel estimation is
In [7], the CE involves estimation of the direction-of-arrival conducted first and the symbols are equalized, based on the
(DoA) and path gains. Channel parameters such as angle estimated channel. Then the online trained ELM is used to
of departures (AoDs) of BS-RIS channels, angle of arrivals refine the equalization process [22]. Therefore, in this paper,
(AoAs) of the RIS-user end channels, angle differences and the online ELM algorithm is explored for designing RIS-
product of the path gains in a RIS-mmWave-MIMO system MIMO receiver. The traditional ELM, implemented in the
are estimated in [8]. In [9], the CE problem is formulated above studies, is not immune to the random choice of the
as DoA estimation problem. In [10], maximum achievable hidden layer size. It will not provide satisfactory performance
rate of the RIS-MIMO system is attempted to obtain by if the hidden layer size is too small or large. Therefore,
means of the proposed CE and the passive beamforming [11] to address this issue, an incremental constructive approach
schemes. The reflection coefficients of the RIS are consid- for the ELM algorithm-based method is presented.
ered to be pre-designed and the effective superposed channel
is estimated. The beamformer is designed on the basis of B. CONTRIBUTION OF THE PAPER
the estimated superposed channel. In [12], minimum mean In an incremental ELM (I-ELM) algorithm, the hidden neu-
square error (MMSE)-based approaches are studied for CE rons are randomly chosen. Based on given specifications,
and the design of phase shifts and equalizer. Estimation of it is able to adjust the hidden layer size and provide descent
the composite channel, which is the direct channel between prediction performance. Therefore, an I-ELM algorithm is
the user and BS and cascaded channel among the user, RIS proposed in this paper to design a receiver for an RIS-
and BS, is also analyzed in the study. muMIMO system. The main contributions of the paper is
With the increase of the number of users, complexities summarized as follows.
in network performance analysis will also increase. Diffi- • An I-ELM receiver is proposed for an RIS-assisted
culties in ensuring QoS demands of the users will also rise. muMIMO systems with a single BS, a single RIS unit
Machine learning (ML)-based methods is seen as a potential and several users.
tool for the future network systems, including RIS-based • The proposed receiver is compared to the MMSE, the
networks [13]. A deep learning (DL) framework, presented traditional ELM and the multilayer perceptron (MLP)
in [14], consists of a twin convolutional neural network receivers in terms of throughput performance for a
(CNN) architecture that estimates both the direct link and cas- given RIS-muMIMO system configuration. Further-
caded link channels by analyzing the received pilot signals. more, comparative studies are also conducted for differ-
The research work conducted in [15] uses a three-stage train- ent BS antennas and RIS elements. It is shown that the
ing approach with deep neural network (DNN) to estimate proposed I-ELM provide satisfactory performance with
the CSI of direct communication link as well as the cascaded respect to the other receivers.
CSI for both the active and inactive RIS components. A deep The paper is organized as follows. The RIS-muMIMO sys-
residual learning-based CSI estimation approach is presented tem model and receiver design with traditional mathematical
in [16]. A convolutional neural network architecture is used approach in section II. ELM based receiver is discussed in
that utilizes the spatial properties of the noisy channel matri- section III. Architecture of I-ELM is described in section IV.
ces as well as the cumulative behavior of the noise. In [17], the Simulation results are provided in V. The paper is concluded
training of a deep neural network with the aid of unsupervised with section VI.
learning technique is performed for designing passive beam-
forming [11] matrices in an RIS-enabled network system. II. SYSTEM MODEL OF RIS-muMIMO
The joint design of the transmit beamformer and phase shift In this paper, an uplink communication system is consid-
matrix is performed by means of deep reinforcement learning ered, as shown in Fig. 1, where the single BS contains NR
(DRL) algorithm in [18]. The transmit beamforming and antennas, the RIS unit contains M elements and K users
phase shift matrices are both said to be obtained concurrently contain NT antennas. The user-base station direct link, the
by the proposed approach at the neural network’s output. user-RIS link and the RIS-base station link are denoted by
However, DL-based approaches require complicated param- H d,k ∈ CNR ×NT , Fk ∈ CM ×NT and G ∈ CNR ×M respec-
eter adjustment through backpropagation algorithm, which tively. These channel matrices follow Rayleigh distribution.
can be expensive in terms of duration. In such case, extreme A total of N subcarrier frequencies are used for transmitting
learning machine (ELM) can be implemented. data symbols, where Nk = N K orthogonal subcarriers are
An ELM-based receiver is designed in [19] for a allocated to each users. Cyclic prefix-orthogonal frequency
light-emitting diode (LED)-based MIMO system. The division multiplexing (CP-OFDM) protocol with 120 kHz of
ELM receiver is constructed on real domain base. ELM- subcarrier spacing is followed by the system. The channel
based approach can also be extended to complex domain coded bits are digitally modulated, based on the normalized
47130 VOLUME 11, 2023
M. R. Mahmood et al.: RIS-Assisted Multi-User MIMO Systems Exploiting ELM

A. RECEIVER DESIGN BY MEANS OF CHANNEL

INFORMATION
With the help of the estimated channel information, tradi-
tional linear receiver(s) equalizes the received symbols. If the
transmitted reference symbols are X ρ,k and the received
reference symbols are Y ρ,k , then the CE by means of LS
estimation technique [24] is

ρ,k .
Ĥ k,LS = Y ρ,k X −1 (5)
To minimize the mean square error (MSE) of the LS chan-
nel estimation, a Bayesian-based MMSE estimation method
can be used [25], [26], which gives the following equation.
FIGURE 1. System model of RIS-assisted mu-MIMO system.
Ĥ k = σ 2 (σ 2 + σk2 )−1 Ĥ k,LS . (6)
Ĥ k,LS Ĥ k,LS

Here, the estimated channel variance is denoted by σ 2 .

modulation power. The inverse discrete Fourier transform Ĥ k,LS
operation (IDFT) is executed on these normalized modulated With the help of MMSE-based CE, that is, Ĥ k , channel
data and transmitted through the direct and cascaded channel equalization is performed on the received data.
links. The received time domain signal at the base station is ∗ ∗
X̂ k,MMSE = (Ĥ k Ĥ k + σk2 I)−1 Ĥ k Y k = V k Y k . (7)
K
X To extract the transmitted bits, digital demodulation on
y= (H d,k + G8Fk )xk + z (1) the equalized symbols and channel decoding operations are
k=1
performed.
Here, the transmitted signal by the user k is represented
by xk . The noise is denoted by z with the characteristics B. THROUGHPUT DETERMINATION
CN (0, σ 2 ), where σ 2 is the noise power. The amplitude The throughput of the user k is determined from the following
attenuation αi and the phase shifts φi , occurred at each RIS expression.
element (i = 1, 2, . . . , M ), are respectively represented by
Rk = B log2 (1 + SINRk ). (8)
the following diagonal matrix.
Here, the allocated bandwidth for each user is denoted
8 = diag[α1 ejφ1 , α2 ejφ2 , . . . , αM ejφM ], (2) by B. The signal-to-noise ratio is represented by SINRk ,
where αi ∈ [0, 1] and φi ∈ [0, 2π ). No signal attenuation which is generally defined as follows.
on the RIS is considered, thus αi = 1 is considered in |V ∗k Ĥ k |2
this paper [23]. Furthermore, the phase shifts are derived SINRk = PK
( i=1,i̸=k |V ∗k Ĥ i |2 + K ∗ 2
P
according to [6]. i=1 V k C k V k + ||V k || )
At the receiving end, the CP is removed from the received (9)
signal and discrete Fourier transform operation (DFT) is Here, the corresponding estimation error covariance matrix
executed. The received symbol Y k ∈ CNR ×NUT at the base is denoted by C k . However, the ELM algorithm does not
station corresponding to the UT k is given by explicitly determine channel information. Therefore, SINRk
is determined by adopting the following equation.
Y k = DFT (H d,k + G8Fk )X k
X K SINRk = −[PAPR + 20 log10 (EVMk /100%)]. (10)
+ DFT (H d,i + G8Fi )X i + DFT (Z) (3)
Here, PAPR represents peak-to-average-power ratio and
i=1,i̸=k
EVMk represents the error vector magnitude. Peak-to-
The above equation is further simplified as follows. average-power ratio value is chosen according to the digital
K
modulation strategy. The error vector magnitude is deter-
X mined by calculating the Euclidean distance between the
Y k = Hk Xk + H i X i + Z. (4)
originally transmitted and received modulated symbols.
i=1,i̸=k

Here, H k = DFT (H d,k + G8Fk ), H i = DFT (H d,i + III. PRELIMINARY KNOWLEDGE OF THE ELM
G8Fi ), X k = DFT (xk ) and X i = DFT (xi ). In this paper, ALGORITHM
the composite channel, Ĥ k , is aimed to be estimated for the In this section, a three-layered traditional extreme learning
user k by using the reference symbols (transmitted with the architecture is described, as shown in Fig. 2. The input layer,
data symbols), which is known at the receiving end (i.e. base hidden layer and output layer contains NR , L and NT neurons
station). respectively. The ELM is assigned with random input weights

VOLUME 11, 2023 47131

 M. R. Mahmood et al.: RIS-Assisted Multi-User MIMO Systems Exploiting ELM

Algorithm 1 Summary of ELM-Based Receiver Design

Mechanism
Training pilot symbol sets T k = Y ρ,k , X ρ,k for user k,
hidden layer with L number of neurons and activation
function a(.)

TRAINING PHASE
• Random generation of W k and bk .
• Determination of Dρ,k from (11).
• Determination of β k from (12).
TESTING PHASE
• Determination of Dk from (13).
• Estimation of X ELM ,k from (14).

The equalized symbols by means of ELM algorithm is

FIGURE 2. Architecture of traditional extreme learning machine.
denoted by X ELM ,k ∈ CNT ×Nk . The ELM algorithm is sum-
marized in 1.
W k ∈ CL×NR and biases bk ∈ CL×1 . The input layer
IV. PROPOSED RECEIVER FOR THE RIS-muMIMO
is fed with the received signals received at each antenna.
SYSTEM
On the basis of a chosen activation function, the hidden layer
In this section, the working mechanism of the I-ELM is
output is computed from the input data. The desired output
discussed with a view to designing RIS-muMIMO receiver.
is obtained at the output layer after the ELM network is
Fig. 3 represents the training architecture of the I-ELM
trained. Different from the backpropagation approach based
receiver. Let us assume that the ELM network has Li initially
ML networks to update the weights W k and biases bk (bk =
assigned neurons. The initial input weights and biases would
b1 , b2 , . . . , bL ), the ELM uses an output weight β k ∈ CL×NT
be W k,i ∈ CLi ×NR and bk,i ∈ CLi ×1 respectively. The
(between the hidden layer and the output layer). It is calcu-
hidden layer output for received reference symbols would be
lated in training phase from the hidden layer output and the
expressed as follows.
desired output (from the training data).
The transmitted and received reference symbols are con- Dρ,k,i = a((W k,i Y ρ,k + bk,i )T ). (15)
sidered as training datasets, expressed as
The output weight for Li hidden layer neurons would be
T k = (Y ρ,k , X ρ,k ), † †
β k,i = (Dρ,k,i Dρ,k,i )−1 Dρ,k,i X Tρ,k . (16)
for the ELM training. The hidden layer output Dρ,k for the The equations (13) and (14) would be used to estimate the
received reference symbols is calculated from Y ρ,k (received transmitted reference symbols for Li hidden neurons.
reference symbols), W k and bk as follows.
Dk,i = a(W k,i Y ρ,k + bk,i ). (17)
Dρ,k = a((W k Y ρ,k + bk )T ). (11)
X I −ELM ,ρ,k,i = β Tk,i Dk,i . (18)
Here, a(.) is the activation function. The output weights
If the number of hidden neurons is not appropriately cho-
β k can be determined from Dρ,k and transmitted reference
sen, then an error (say, MSE = ξρ ) will exist between X ρ,k
symbols X ρ,k , by finding the least square solution of the
and X I −ELM ,ρ,k,i . The I-ELM network will then keep adding
problem Dρ,k β k = X Tρ,k .
hidden layer neurons until ξρ is minimized to an expected
† † value ξρ,exp . A maximum allowable number of hidden layer
β k = (Dρ,k Dρ,k )−1 Dρ,k X Tρ,k . (12)
neurons, Lmax is also assigned to control the addition of
The received data symbols Y k is then provided into the hidden neurons.
input layer of the trained ELM receiver. The hidden layer After the training of I-ELM, input weights W k,new , biases
output Dk for the received data symbols is then computed as bk,new and output weights β k,new for the hidden layer with
follows. Lnew neurons. The testing of I-ELM is similar to the tradi-
tional ELM. The I-ELM receiver mechanism is summarized
Dk = a(W k Y k + bk ). (13) in 2.
By using the computed output weight β k , the transmitted With the increase of hidden layer neurons, the prediction
symbols is determined from Dk . accuracy improves. However, too many hidden layer neurons
may increase the load on the computational resources. In this
X ELM ,k = β Tk Dk . (14) case, I-ELM algorithm is advantageous for automatically

47132 VOLUME 11, 2023

M. R. Mahmood et al.: RIS-Assisted Multi-User MIMO Systems Exploiting ELM

FIGURE 3. Block diagram of the training phase of I-ELM receiver. This phase of the proposed receiver continues until ξρ and Li criteria are
met.

handling of the hidden layer size and consequently deliver Algorithm 2 I-ELM-Based Receiver Mechanism
satisfactory prediction performance. Training pilot symbol sets T k = Y ρ,k , X ρ,k for user k,
activation function a(.), initial hidden layer neurons Li ,
V. SIMULATION RESULTS AND DISCUSSIONS maximum hidden layer neurons Lmax and ξρ = ξρ,exp .
In this section, the simulation results obtained for the
RIS-muMIMO system with center frequency of 30 GHz is TRAINING PHASE
presented. The spacing among the subcarrier frequencies is • Initialization of W k,i and bk,i for given Li .
120 kHz. Four users (K = 4, with NT = 2 transmitting • Computation of β k,i from (15) and (16).
antenna each), single BS with NR = 64 and single RIS unit • Estimation of X I −ELM ,ρ,k,i from (17) and (18).
with M = 64 are initially considered for the study. The • Determination of ξ between X ρ,k and X I −ELM ,ρ,k,i .
number of total subcarrier is N = 960, which is divided • While Li ≤ Lmax and ξρ ≥ ξρ,exp
equally among the 4 users. In other words, the number – Li is expanded to Li+1 for adding single hidden layer
of allocated subcarrier for each user is Nk = 240. Low- neuron each time.
density parity check (LDPC) with 1/2 code rate is applied – Expansion of W k,i and bk,i to W k,i+1 and
to the data bits. The modulation format is elected to be bk,i+1 for the newly added hidden layer neuron.
16-quadrature amplitude modulation (16-QAM). The signal-
2
– Computation of β k,i+1 by following (15) and
k| ]
to-noise ratio is defined as E[|x
σ2
. The ELM algorithms (both (16).
traditional and proposed) are trained with the transmitted – Estimation of X I −ELM ,ρ,k,i+1 for the newly deter-
and received symbols (corresponding to the user k) at one mined β k,i+1 according to (17)
OFDM symbol time, out of 14 OFDM symbol times. Acti- and (18) and computation of ξρ between X ρ,k
vation function tanh is chosen for the hidden layer of both and X I −ELM ,ρ,k,i+1 .
of the ELM receivers. The input weights and the biases are TESTING PHASE
assigned by random complex numbers within [−10−2 , 10−2 ].
• Given, Lnew , W k,new ∈ CLnew ×NR , bk,new ∈ CLnew ×1
This interval is chosen such that the weight and the bias
and β k,new .
values fall within the activation function’s region of con-
– Dk,new = a(W k,new Y k + bk,new ).
vergence [22]. Based on the study conducted in [27], the
– X I −ELM ,k,new = β Tk,new Dk,new .
number of hidden layer neurons in the traditional ELM is
considered to be L = max(NR , NT ). In case of I-ELM
receiver, the initial hidden layer size Li and the maximum
number of hidden layer neurons, i.e. Lmax , are considered to
be equal to 10 and max(NR , NT ) respectively. The expected A. ANALYSIS OF THROUGHPUT PERFORMANCES
minimum MSE value ξρ,exp for I-ELM receiver is chosen to The performance of the proposed I-ELM receiver is com-
be 0.01. pared to that of the MMSE, traditional ELM and MLP

VOLUME 11, 2023 47133

 M. R. Mahmood et al.: RIS-Assisted Multi-User MIMO Systems Exploiting ELM

TABLE 1. Simulation parameters.

FIGURE 5. Average throughput as a function of BS antennas by means of

MMSE, traditional ELM, I-ELM and MLP receivers.

FIGURE 4. Average throughput obtained by means of the MMSE,

traditional ELM, I-ELM and MLP receiver.

receiver. An MLP algorithm [28], which is a DNN-based

algorithm, is used to design a receiver for comparing its per-
formance with the ELM-based receivers. According to [25],
the MLP algorithm uses each neuron for real and imaginary
part of the complex numbers respectively. Therefore, the
MLP-based receiver is constructed such that the real and
the imaginary part of the received symbols are separated
for training and equalization. The input layer of the MLP FIGURE 6. Average throughput as a function of RIS elements by means of
contains 2 NR neurons and the output layer contains 2 NT MMSE, traditional ELM, I-ELM and MLP receivers.
neurons. The MLP receiver has 4 hidden layers. The number
of neurons in the successive hidden layers are 64, 32, 16
and 8. The activation function chosen for each of the hidden complexities. It can adapt the hidden layer size as necessary
layers is tanh. To update the weight parameters, the learning based on the provided criteria, such as Lmax and ξρ,exp .
rate of 0.005 and momentum of 0.9 are assigned. In Fig. 4,
both of the ELM receivers have shown better throughput B. EVALUATION OF THE RECEIVERS FOR RIS-muMIMO
performance with respect to the MMSE and the MLP receiver CONFIGURATION
for the RIS-muMIMO system with 2 transmitting antennas In this section, the presented RIS-muMIMO receivers are
and 64 receiving antennas. The MMSE receiver achieves evaluated by varying the BS antennas and the RIS ele-
low throughput performance at lower SNR values due to ments respectively, which can increase the throughput of the
imperfect channel estimation. At SNR = 30 dB, about RIS-muMIMO systems [3], [29]. Here, the expected MSE, ξρ
6 Mbits/s more is achieved by the ELM receivers than the is kept 0.01. In Fig. 5, the MMSE, traditional ELM, I-ELM
MMSE receiver. For SNR values between 0 − 28 dB, both and MLP receivers are evaluated in terms of throughput per-
the ELM receivers utilize L = 64 neurons in the hidden layers formance for 64 RIS elements and different BS antennas (as
for all user data. However, for the SNR value above 28 dB, mentioned in Table 1) at SNR = 22. As the BS antennas are
the lowest and highest values of L are 39 and 64 respectively increased from 16 to 32, noticeable improvement is observed
for the I-ELM receiver, whereas, the traditional ELM utilizes for both of the ELM receivers. However, by increasing BS
L = 64 neurons. This demonstrates that the I-ELM receiver antennas from 32 to 128, the throughput performance is
has enough throughput performance that is comparable to the increased by small margin. On the other hand, the through-
traditional ELM receiver while also reducing computational put performance marginally increases by means of MMSE

47134 VOLUME 11, 2023

M. R. Mahmood et al.: RIS-Assisted Multi-User MIMO Systems Exploiting ELM

receiver. In Fig. 6, the MMSE, traditional ELM, I-ELM [6] J. Mirza and B. Ali, ‘‘Channel estimation method and phase shift design for
and MLP receivers are evaluated for the RIS-muMIMO sys- reconfigurable intelligent surface assisted MIMO networks,’’ IEEE Trans.
Cognit. Commun. Netw., vol. 7, no. 2, pp. 441–451, Jun. 2021.
tem with 64 receiving antennas and different RIS elements, [7] X. Chen, J. Shi, Z. Yang, and L. Wu, ‘‘Low-complexity channel estimation
as mentioned in Table 1, at SNR = 22. With the increase for intelligent reflecting surface-enhanced massive MIMO,’’ IEEE Wire-
of RIS elements, the throughput performance achieved by less Commun. Lett., vol. 10, no. 5, pp. 996–1000, May 2021.
employing the MMSE and both the ELM receivers improved. [8] J. He, H. Wymeersch, and M. Juntti, ‘‘Channel estimation for RIS-aided
mmWave MIMO systems via atomic norm minimization,’’ IEEE Trans.
The ELM receivers prevail its superiority over the MMSE Wireless Commun., vol. 20, no. 9, pp. 5786–5797, Sep. 2021.
and the MLP receiver. The complexity of the receivers are [9] K. Ardah, S. Gherekhloo, A. L. F. de Almeida, and M. Haardt,
discussed in the following subsection. ‘‘TRICE: A channel estimation framework for RIS-aided millimeter-
wave MIMO systems,’’ IEEE Signal Process. Lett., vol. 28, pp. 513–517,
2021.
1) COMPLEXITY OF THE RECEIVERS [10] J. An, C. Xu, L. Gan, and L. Hanzo, ‘‘Low-complexity channel estimation
and passive beamforming for RIS-assisted MIMO systems relying on dis-
Following the expression, KNk (3NT2 NR + 2NR2 NT − NR2 + crete phase shifts,’’ IEEE Trans. Commun., vol. 70, no. 2, pp. 1245–1260,
2 3
3 NT + NT NR ), the MMSE receiver requires a total of Feb. 2022.
12661760 mathematical operations for symbol detection, [11] W. Yan, X. Yuan, and X. Kuai, ‘‘Passive beamforming and information
transfer via large intelligent surface,’’ IEEE Wireless Commun. Lett., vol. 9,
whereas, 2KNk (La 2NR +La Lb +Lb Lc +Lc Ld +Ld 2NT ) math- no. 4, pp. 533–537, Apr. 2020.
ematical operations are required for trained MLP receiver. [12] O. T. Demir and E. Bjornson, ‘‘Is channel estimation necessary to select
Here, La = 64, Lb = 32, Lc = 16 and Ld = 8. Therefore, phase-shifts for RIS-assisted massive MIMO?’’ IEEE Trans. Wireless
Commun., vol. 21, no. 11, pp. 9537–9552, Nov. 2022.
a total of 20951040 mathematical operations are required
[13] M. R. Mahmood, M. A. Matin, P. Sarigiannidis, and S. K. Goudos, ‘‘A com-
for MLP receiver. The expression for complexity of the prehensive review on artificial intelligence/machine learning algorithms
trained ELM receivers is KNk (2LNR + L + 2LNT − NT ). The for empowering the future IoT toward 6G era,’’ IEEE Access, vol. 10,
traditional ELM receiver performs 8169600 mathematical pp. 87535–87562, 2022.
[14] A. M. Elbir, A. Papazafeiropoulos, P. Kourtessis, and S. Chatzinotas,
operations at all SNR values, whereas, at SNR values above ‘‘Deep channel learning for large intelligent surfaces aided mm-wave
28 dB (Fig. 4), the I-ELM receiver performs less number massive MIMO systems,’’ IEEE Wireless Commun. Lett., vol. 9, no. 9,
of operations with 39 neurons (lowest value of L), which pp. 1447–1451, Sep. 2020.
is 4977600. This demonstrates that the I-ELM receiver has [15] S. Gao, P. Dong, Z. Pan, and G. Y. Li, ‘‘Deep multi-stage CSI acquisition
for reconfigurable intelligent surface aided MIMO systems,’’ IEEE Com-
comparable throughput to the traditional ELM receiver while mun. Lett., vol. 25, no. 6, pp. 2024–2028, Jun. 2021.
also reducing computational complexities. [16] C. Liu, X. Liu, D. W. K. Ng, and J. Yuan, ‘‘Deep residual learning for chan-
nel estimation in intelligent reflecting surface-assisted multi-user commu-
nications,’’ IEEE Trans. Wireless Commun., vol. 21, no. 2, pp. 898–912,
VI. CONCLUSION Feb. 2022.
In this paper, an I-ELM-based receiver is proposed for the [17] J. Gao, C. Zhong, X. Chen, H. Lin, and Z. Zhang, ‘‘Unsupervised learn-
ing for passive beamforming,’’ IEEE Commun. Lett., vol. 24, no. 5,
RIS-muMIMO system. Based on the provided specifica- pp. 1052–1056, May 2020.
tions, this ML architecture adjusts the hidden layer size and [18] C. Huang, R. Mo, and C. Yuen, ‘‘Reconfigurable intelligent surface
minimizes the overall performance error. For a given RIS- assisted multiuser MISO systems exploiting deep reinforcement learn-
muMIMO system, the I-ELM receiver is compared to the ing,’’ IEEE J. Sel. Areas Commun., vol. 38, no. 8, pp. 1839–1850,
Aug. 2020.
MMSE, traditional ELM and MLP receivers in terms of [19] D. Gao and Q. Guo, ‘‘Extreme learning machine-based receiver for
throughput. Performance evaluation is done for different BS MIMO LED communications,’’ Digit. Signal Process., vol. 95, Dec. 2019,
antennas and RIS elements. In all cases, the I-ELM receiver Art. no. 102594.
[20] C. Qing, L. Wang, L. Dong, and J. Wang, ‘‘Enhanced ELM based chan-
performs better than MMSE and MLP receivers. nel estimation for RIS-assisted OFDM systems with insufficient CP and
imperfect hardware,’’ IEEE Commun. Lett., vol. 26, no. 1, pp. 153–157,
Jan. 2022.
REFERENCES
[21] J. Liu, K. Mei, X. Zhang, D. Ma, and J. Wei, ‘‘Online extreme learning
[1] S. Khwandah, J. Cosmas, P. Lazaridis, Z. Zaharis, and I. Chochliouros, machine-based channel estimation and equalization for OFDM systems,’’
‘‘Massive MIMO systems for 5G communications,’’ Wireless Pers. Com- IEEE Commun. Lett., vol. 23, no. 7, pp. 1276–1279, Jul. 2019.
mun., vol. 120, no. 3, pp. 2101–2115, 2021. [22] D. F. Carrera, D. Zabala-Blanco, C. Vargas-Rosales, and
[2] C. Xu, J. Cosmas, Y. Zhang, P. Lazaridis, G. Araniti, and Z. D. Zaharis, C. A. Azurdia-Meza, ‘‘Extreme learning machine-based receiver for
‘‘3D MIMO radio channel modeling of a weighted linear array system multi-user massive MIMO systems,’’ IEEE Commun. Lett., vol. 25, no. 2,
of antennas for 5G cellular systems,’’ in Proc. Int. Conf. Telecommun. pp. 484–488, Feb. 2021.
Multimedia (TEMU), Jul. 2016, pp. 1–6. [23] B. Sheen, J. Yang, X. Feng, and M. M. U. Chowdhury, ‘‘A deep learning
[3] P. Mursia, V. Sciancalepore, A. Garcia-Saavedra, L. Cottatellucci, based modeling of reconfigurable intelligent surface assisted wireless
X. C. Pérez, and D. Gesbert, ‘‘RISMA: Reconfigurable intelligent surfaces communications for phase shift configuration,’’ IEEE Open J. Commun.
enabling beamforming for IoT massive access,’’ IEEE J. Sel. Areas Com- Soc., vol. 2, pp. 262–272, 2021.
mun., vol. 39, no. 4, pp. 1072–1085, Apr. 2021. [24] Y. S. Cho, J. Kim, W. Y. Yang, and C. G. Kang, MIMO-OFDM Wireless
[4] H. Zhang, S. Ma, Z. Shi, X. Zhao, and G. Yang, ‘‘Sum-rate maxi- Communications With MATLAB. Singapore: Wiley, 2010.
mization of RIS-aided multi-user MIMO systems with statistical CSI,’’ [25] D. F. Carrera, C. Vargas-Rosales, N. M. Yungaicela-Naula, and
IEEE Trans. Wireless Commun., early access, Dec. 19, 2022, doi: L. Azpilicueta, ‘‘Comparative study of artificial neural network based
10.1109/TWC.2022.3228910. channel equalization methods for mmWave communications,’’ IEEE
[5] S. Zettas, S. Kasampalis, P. Lazaridis, Z. D. Zaharis, and J. Cosmas, Access, vol. 9, pp. 41678–41687, 2021.
‘‘Channel estimation for OFDM systems based on a time domain pilot [26] D. F. Carrera, C. Vargas-Rosales, L. Azpilicueta, and J. A. Galaviz-Aguilar,
averaging scheme,’’ in Proc. 16th Int. Symp. Wireless Pers. Multimedia ‘‘Comparative study of channel estimators for massive MIMO 5G NR
Commun. (WPMC), 2013, pp. 1–6. systems,’’ IET Commun., vol. 14, no. 7, pp. 1175–1184, Apr. 2020.

VOLUME 11, 2023 47135

 M. R. Mahmood et al.: RIS-Assisted Multi-User MIMO Systems Exploiting ELM

[27] M. R. Mahmood, M. A. Matin, P. Sarigiannidis, S. K. Goudos, SOTIRIOS K. GOUDOS (Senior Member, IEEE)
and G. K. Karagiannidis, ‘‘Residual compensation-based extreme learn- received the B.Sc. degree in physics, the M.Sc.
ing machine for MIMO-NOMA receiver,’’ IEEE Access, vol. 11, degree in electronics, and the Ph.D. degree in
pp. 13398–13407, 2023. physics from the Aristotle University of Thessa-
[28] L. Noriega, ‘‘Multilayer perceptron tutorial,’’ School Comput., loniki, in 1991, 1994, and 2001, respectively, the
Staffordshire Univ., Staffordshire, U.K., Tech. Rep., 2005. master’s degree in information systems from the
[29] P. Wang, J. Fang, L. Dai, and H. Li, ‘‘Joint transceiver and large intelli- University of Macedonia, Greece, in 2005, and
gent surface design for massive MIMO mmWave systems,’’ IEEE Trans. the Diploma degree in electrical and computer
Wireless Commun., vol. 20, no. 2, pp. 1052–1064, Feb. 2021.
engineering from the Aristotle University of Thes-
saloniki, in 2011. He joined the Department of
M. REZWANUL MAHMOOD received the B.S. Physics, Aristotle University of Thessaloniki, in 2013, where he is currently
degree in electrical and electronic engineering an Associate Professor. He is the Director of the ELEDIA@AUTH and a
from East West University, Dhaka, Bangladesh, Laboratory Member of the ELEDIA Research Center Network. He is the
in 2017. He is currently pursuing the M.S. degree author of the book Emerging Evolutionary Algorithms for Antennas and
in electrical and electronic engineering with Wireless Communications (The Institution of Engineering and Technology,
North South University, Dhaka. He is a Research 2021). His research interests include antenna and microwave structures
Assistant with North South University. He has design, evolutionary algorithms, wireless communications, and semantic
published few papers in reputed journals and web technologies. He is a member of the IEICE, the Greek Physics Society,
international conferences. His research interests the Technical Chamber of Greece, and the Greek Computer Society. He is
include future wireless networks, the Internet the founding Editor-in-Chief of the Telecom open access journal (MDPI).
of Things, and machine learning. He is serving as an Associate Editor for IEEE TRANSACTIONS ON ANTENNAS AND
PROPAGATION, IEEE ACCESS, and IEEE OPEN JOURNAL OF THE COMMUNICATION
SOCIETY. He is serving as the IEEE Greece Section Secretary.
MOHAMMAD ABDUL MATIN (Senior Member,
IEEE) received the B.Sc. degree in electrical and
electronic engineering from BUET, Bangladesh,
the M.Sc. degree in digital communication from
Loughborough University, U.K., and the Ph.D.
degree in wireless communication from Newcastle
University, U.K. He has been a Professor with the
Department of Electrical and Computer Engineer-
ing, North South University (NSU), since 2008.
He has published over 120 peer-reviewed journals
and conference papers. He is the author/editor of 17 academic books and
21 book chapters. He serves as a member for the Editorial Board for
several international journals, including IEEE Communications Magazine
and IET Wireless Sensor Systems. He has received a number of prizes and
scholarships, including the Best Student Prize (Loughborough University),
Commonwealth Scholarship, and Overseas Research Scholarship (ORS)
conferred by the Committee of Vice Chancellors and Principals (CVCP) in
the U.K.

47136 VOLUME 11, 2023