Machine Learning Approaches of Large Intelligent Surface To Assist 6G-IoT

Applications: A Holistic Review

P inakshi De1 , Biswaraj Sen2 , Sourav Dhar3 , Saumya Das 4

1 Department of Computer Application, Dinabandhu Andrews Institute of Technology and

Management,Kolkata, West Bengal, India.,
2 Department of Information Technology, Sikkim Manipal Institute of Technology, Sikkim Manipal
University, Sikkim, India,
3 Department of Electronics and Communication Engineering, Sikkim Manipal Institute of
Technology, Sikkim, India,
3 Director, Centre for Distance and Online Education (CDOE), Sikkim Manipal University,
Sikkim, India
4 Department of Computer Science and Engineering-Artificial Intelligence, Brainware University,
Barasat, West Bengal, India.

Abstract

Wireless networks of 5G and beyond require to accommodate a wide range of services and appli-
cations, a rapidly increasing data traffic flow across handheld devices, and a growing density of
mobile users. In future networks, high-frequency waves will be crucial, but they are weak over
long distances and can be easily blocked by objects. Through a tutorial, a comparison of the im-
plementation and capabilities of IRS, LIS, and RIS is carried out. Future wireless networks may
be able to increase their coverage and data rate by utilizing large intelligent surfaces (LISs). This
is how the idea of a smart radio environment is realized.As such, LISs have the infrastructure to
support the sixth generation of communication networks. The potential benefits of RIS-assisted
communication systems can be effectively maximized through machine learning (ML), especially
when the computational complexity of LIS deployment and operation rises quickly as the number
of user-infrastructure interactions begins to increase.
Keywords: 6G, LIS, Machine Learning, Deep Learning, Federated Learning, RIS, MIMO.

1
[email protected]
2
[email protected]
3
[email protected]
4
[email protected]

1
1 Introduction
Mobile communication is a wireless technology refers to the exchange of voice, data and multi-
media content while users are on the move. It defines the connectivity between different devices
such as smart phones, laptops, tablets and other wireless equipment by using different technolo-
gies such as 2G, 3G, 4G and now 5G. [25] Now-a-days mobile becomes as an essential part of
our daily life. The development of 5G communication has made significant changes and the sys-
tem will work completely in millimeter (mmWave) wave frequency and capable to handle Inter-
net of Things (IoT) services. Industry sector as well as academia already started the testing of
sixth generation (6G) [27] communication system. Till now, the fifth generation (5G) communi-
cation is in deployment phase throughout the world. Latest outlook define that 6G system will
deliver outstanding performance by incorporating multiple number of devices and users through
reliable wireless systems. It is contemplate that when a new design, architecture is being clar-
ified, the requirement of the users should be meet. Moreover, to provide the best quality of
service (QoS) [6] for clients, we need more bandwidth, higher data rates, minimum interference
and low installation costs for the service providers. It is outcome that 6G communication is capa-
ble to do monochromic-defense, monochromic-mending, monochromic-enhancement, monochromic-
governing, monochromic-plan and monochromic-supervise their assets in such a way to achieve the
maximum key performance indicators (KPIs) by controlling network activities and functionalities
to power hurdles, bandwidth utilization,[10]enironmental status and other factors. It is strenuous
to install different technologies simultaneously on a unified platform, since they have harsh stan-
dards that must be qualify in terms of latency, data throughput and trustworthiness. However,
fluctuating wireless channels represent consequential barriers to the development of ultra-reliable
wireless communications due to users’ mobility. Customary approaches to this problem include
either tailoring to channel fading through revived power, rate management and beamforming tech-
niques or recompense for it with numerous modulation, coding and diversity arrangements. The
number of nodes connected in wireless ecosystem is envisioned that in the next few years, there will
be tens of billions of nodes. Universal coverage, low latency, high energy efficiency, extremely high
data rates, and high trustworthiness are all features of 6G wireless networks. According to recent
literature, the main concepts for 6G include reconfigurable intelligent surfaces, large intelligent
surfaces (LISs), and intelligent reflecting surfaces (IRSs).[14]
In addition, by providing remarkable solutions like accurate traffic and mobility prediction,

2
the ultimate instrument for analyzing traffic latency and reliability for allowing mURLLC [21] in
6G-IoT is created by advances in machine learning (ML). By coordinating the distinguish phases
of various tracks or providing unusual reflections, blockage diffraction and signal strength enhance-
ment are achieved. In comparison to conventional mMIMO, LIS can be widely positioned all over
the terminals and devices as a surface of reflection, improving line-of-sight (LoS) [32] in the prop-
agation channel and drastically lowering the overhead for channel state information (CSI). This
improves communication reliability through minimizing the degree of smaller-scale fading. Due to
the rapidly increasing the variety of wireless gadgets being utilized and the volume of information
that every gadget uses, the demand for data traffic is increasing exponentially. However, the limi-
tations of having only a handful of transmitters at ground stations (BS) preclude existing MIMO
systems from achieving those requirements. [2] Compared to massive MIMO transmission set up,
ultra-mMIMO interacting systems have more transmitters and a larger antenna array, which results
in special channel properties. By appropriately altering signal propagation through the adjustment
of numerous passive reflecting components, a LIS [18] is a promising method for boosting wireless
network communication capacity whilst staying cost-effective and energy-efficient. The publica-
tion’s goal is to identify the basic capacity limitation of LIS-assisted dedicated MIMO articulation
networks with multiple antenna communicators and acceptors. [23]

Table 1: Comparison between IRS, RIS, LIS, LIM, SDM

Technology Role in 6G Key Characteristics Control Mechanism Hardware

Complexity

Intelligent Re- Enhances sig- Passive, low-cost, Discrete phase shifts Low to mod-
flecting Surface nal strength phase-shifting ele- via PIN diodes, var- erate
(IRS) and coverage ments that redirect actors
by passively signals without power
reflecting consumption [28]
signals

3
 Technology Role in 6G Key Characteristics Control Mechanism Hardware
Complexity

Dynamically Programmable, tun- Discrete or continu- Moderate

Reconfigurable
manipulates able surface; real-time ous tuning via elec-
Intelligent Sur-
wireless signals adaptation to channel tronics
face(RIS) [1]
to improve link conditions to large sur-
reliability faces for macro-level
control [19]

Large Intelligent Extends Large-scale arrays, Not usually reconfig- High (if ac-
Surface (LIS) IRS/RIS con- seamless wave ma- urable [30] tive); low (if
cept to large nipulation over large passive)
surfaces for areas
macro-level
control [4]

Software- Programmable Software-defined be- Full software control High.

Defined Meta- metasur- havior, can alter of each unit cell [26]
surface (SDM) face enabling reflection/refraction
wavefront /absorption dynami-
engineering [5] cally

Large Intelligent Combines Electromagnetic Time-varying modu- High

Metasurface metasurface control at scale, engi- lation of properties
(LIM) technology neered meta-atoms for (e.g., capacitance)
with large- precise manipulation
scale de- [11]
ployment for
complex tasks

4
 This is how the rest of the paper is structured. A summary of the LIS’s structural design is
covered in Section II. Section III introduces the machine learning architectures used in the literature.
Section IV discusses related works. Section V provides a description of the ML applications in LISs.
An overview of the complete work is provided at the end of this article in Section VI.

2 Outline of large intelligent surfaces

Occasionally synchronized sub wavelength components are known as metamaterials, made up of
unit cells and enable total control over the EM activities of the metasurface, are the main component
used to generate LIS models. With the help of integrated electronics, this artificial electromagnetic
material surface may be electrically manipulated and offers special wireless communication capa-
bilities. An IRS is basically an inexpensive, multipurpose surface, like a thin coating of intelligently
constructed composite material. In a much smaller or larger metasurface, wireless intercell com-
munication works well, but dedicated path communication is a superior alternative because it is
simpler to amalgamate the controller on the unique chip. The design rules for inter-communication
processes must be followed, with stringent energy latency and robustness criteria. Three separate
layers may make up a typical LIS architecture. The first layer is made of a conductive substrate,
such plastic or glass, which aids in the structural stability of the LIS and offers mechanical support.
It is constructed of several metallic plates that may be adjusted and rearranged to directly control
incoming waves. The subsequent layer consists of a copper grain sheet with a number of tiny an-
tennas that can reflect incoming radio waves in order to stop LIS reflection from causing controlled
power transmission losses. In order to ensure the longevity and durability of duo scenarios and
continuous monitoring of the plane of radiating dimensions and changes in period of the radiating
elements mobilized by the sophisticated LIS supervisor, the third layer is a protective covering
known as a control circuit board.

5
 Figure 1: LIS Architecture

The key components of the LIS that affect the metasurface reconfiguration rate are often soft-
ware and controllers that are focused on meta-atoms. Because static metasurfaces don’t need
active electrical circuits and can be entirely passive, their corresponding power consumption dif-
fers greatly from that of reconfigurable metasurfaces. Because energy is required for managing
received signalling and valves for restructuring, respectively, only metasurfaces with attributes that
can be adjusted are eligible to be considered essentially quiet. Following proper calibration of the
metasurface, however, signal transmission does not require a particular power supply. A LIS hard-
ware design consists of a flat arrangement with several incredibly thin meta particles, sometimes
identified to as units or particles. Each every meta-particles incorporates many digital presenta-
tions reflecting various electro magnetic reactions. Research on tunable or tweakable metaatoms
is ongoing and includes a variety of methods centered on the operating frequency. Beamforming
is the process by which several antennas generate duplicates of the lagged signal that are indistin-
guishable. The terrain where the precise position of broadcast copies being captured concurrently,
constructive interference takes place; in other locations, destructive interference takes place. When
many antennas are utilized, then acceptor will receive improved broadcasts with respect to an single
antenna only broadcasts information while using the exact equivalent amount of energy. Construc-
tive interference at the receiver is produced by the temporal delays at the transmitting antennas.
A collection of emitters is used in beam shaping, a type of filtering technique, to collect or release
radiation across their aperture in a specific direction. The LIS facilitates communication between
the user and the BS by passively beamforming by reflecting the signals.

6
3 Machine learning and its fundamental
Machine Learning (ML) is a discipline of research that focuses on the properties and hypotheses of
learning computations, as well as how they work and the systems that go along with them. Machine
learning is a broad, multifaceted field which embraces ideas beginning out of many other fields,
including as statistics, AI, information theory, optimization theory, optimal control, and many
other scientific, mathematical, and engineering fields. Due to its use in a wide range of applications,
machine learning has impacted almost every scientific field, having a big impact on both research
and society. These days, data mining, informatics, autonomous devices, recommendation engines,
and systems for identification all make extensive use of machine learning. The machine learning
technique generally classified into duplet considerable steps: training as well as testing. In training
module of the machine learning model, system is trained with dataset to learn the pattern and
makes the prediction. The machine learning algorithm’s learning phase involves using datasets to
teach the system patterns and generate predictions. The trained model is used to forecast output
against each input supplied to the system during the testing phase. Fig. 2 shows the categorization
of machine learning.

Figure 2: Groups of Machine Learning

3.1 Supervised Learning

A key approach to artificial intelligence and machine learning is supervised machine learning. Us-
ing labeled data—each input having a proportionately correct output—a model is trained. By
contrasting its predictions with the actual responses from the training data, the model acquires
knowledge. It gradually adapts to reduce fallacy and increase precision. Modeling precise predic-
tions in the face of fresh, unknown data is the aim of supervised learning. One of the foremost

7
techniques in the fields of artificial intelligence and supervised data mining is supervised learning,
which may be used in many different ways, such as supervised learning regression and supervised
learning classification.

Figure 3: Supervised Learning

3.2 Unsupervised learning

When a model is trained on data without labeled outcomes, it is referred to as unsupervised
learning. In contrary to supervised learning, which pairs each training sample with an output or
label, unsupervised learning looks for concealed structures, groups, or patterns in the input data.
Unsupervised learning is a subset of machine learning that functions differently from supervised
learning because the machine is not guided by a supervisor. With this method, data without labels
or categories is fed into the system. Without any prior knowledge, it independently scrutinizes the
data to identify trends, clusters, or connections.

Figure 4: Unsupervised Learning

3.3 Reinforcement learning

A particular field of machine learning called reinforcement learning (also known as RL) explores
how agents could learn to make decisions by making mistakes in order to maximize cumulative

8
rewards. RL allows robots to learn by interacting with an environment and gaining feedback based
on what they do. This feedback is given as incentives or sanctions. An agent acts in an environment,
receives rewards or penalties for those actions, and modifies its behavior in response.

Figure 5: Reinforcement Learning

3.4 Deep learning

Deep learning(DL), a portion of machine learning, provides an algorithm to build classifications
and projections based on the decisions made by the data being input without the need for explicit
programming. By replicating the neural networks found in the human brain, deep learning permits
computers to effortlessly identify patterns and generate accurate conclusions from an immense
amount of unstructured data. In order to process incoming information, a neural network is s
structure made up of layers of connected nodes, or neurons.

3.5 Federated learning

Federated learning is a machine learning environment that aims to learn a model across several
dispersed edge devices or servers that keep local data samples without sending them. With contrast,
all data is uploaded to a centralized server with classic centralized machine learning algorithms.
Rather than sending raw data to a central server, federated learning uses various points of data
origin to train the model frequently.

9
 Figure 6: Federated Learning

4 Classifying literature studies of large

intelligent surfaces
We right now categorize current LIS research work. In the literature, LISs are often get referred
to by a variety of phrases and acronyms. All forms of intelligent surfaces that have been used in
various research will be addressed to in this study by the term LIS. As a result, a overview of the
current literature on LISs is included in this section. The authors M.A.Aygul [3] et.al.in their paper
described a MCP model consists of 3 hidden layers, sampled channel is used as a input and output
is optimum phase, their model is used for single user. The authors A.M.Elbir et.al. [7] highlighted
a model named as Twin CNN by using 3 convolutional layers along with 2 fully connected layers.
They have used received pilot signals as input and gives channel matrices as output. The authors C.
Jia et. al. [12] explained a DNN model with 3 hidden layers.They have taken measured coordinates
as input and beam management as output. The authors B.Yang et.al. [31]described a CNN model
by using 2 convolutional and 2 fully connected layers. They have choosen incident RF signal as
input and gives set of interfering users as output. The authors S.Liu et.al. [15]highlighted a model
named CV-Dn CNN, by using 15 convolutional layers. In this model recieved signal is used as
input, and as a output residual noise is generated.

10
 Table 1: Analyzing and contrasting several machine learning methods for large intelligent sur-
faces

Algorithm Ref System Model Number of Goal

User
3 OFDM based antenna Single User Maximum achievable
Supervised Learning rate
7 mMIMO Multiple User Maximum perfor-
mance achievable
rate
10 mmWave beam man- Multiple user Performance maxi-
agement mization
22 MISO Multiple user Rate maximization
12 mmWave MIMO Single user Performance maxi-
mization

Because of the RIS phase changes and the interaction between active and passive beamforming,
a successful joint optimization design is challenging. Song et al. [22] proposed a NN system of
unsupervised approach to effectively address the joint optimization problem. To reduce the model
problems in RIS-aided networks, Gao et al. [9] employed DL approaches. With the goal to reduce
the labeling cost of supervised learning, the authors presented an unsupervised label-free approach
to RIS phase shift optimization. Onsite training with a customized DNN is suggested by establish-
ing a DNN parameters and optimizing the target function. Five completely integrated layers make
up the NN’s structural design. Ma et al. [17] addressed the privacy challenges in RIS-aided associ-
ation by recommending an algorithm based on FL that might allow in order to communicate at fast
speeds although while preserving an individual’s privacy. It is possible to train the model locally
using individual participant’s native database in FL. All of the local models have been combined to
create a global model. Each device’s initial setup for the subsequent [29]the global approach is used
to obtain the learning process. Until the worldwide model output converges, every step is repeated
in order to arrive at the optimal global model. Jiao et al. [13] improved the KGR of the CSI-based
material layer key generation by developing a conversational quantized threshold forecasting model
employing the machine learning technique. The authors looked at a wireless system that used RIS
to help two users with single antennas. In order to build a secret key, users communicate in the

11
Time Division Duplexing mode by investigating and acquiring unpredictability from bidirectional
channels.The authors of [16]suggested a perishing DQN (DDQN)-deployed method to project the
change in period of the RIS and the path of the aerial vehicle without an operator(UAV) while
assuring such the client’s information requirements are met in order to address combining the mo-
mentary shift and trajectory layout problem. The challenges of joint implementation of the system,
power distribution, identifying flexible decoding order, and phase shift of RIS control in a RIS-aided
system with multiple users while maintaining the individual data rate enhancement as needed were
examined by Liu et al. [20] in order to maximize the computational effectiveness of networks with
non-orthogonal multiple access (NOMA). To exclude marked information that can alter a RIS’s
transitions in phase utilizing truncated-overhead training, Taha et al. proposed a DRL shape [24].
The phase transition model in downstream MISO networks for communication with the help of
RIS has been enhanced by Feng et al. [8] in order to optimize the received signal-to-noise ratio
(SNR). The suggested approach evaluate the initial training data and monitors new instantaneous
wideband monitoring information using a function. The model’s FNN, which has duo unrevealed
layers, uses a role of the surface that is concealed and a purelin method for the resulting surface as
the functions for initiation. The forecasting model uses high discretiztion levels to reduce the bit
dissent ratio when the SNR is high and low levels of quantization to maintain the lowest possible
bit conflict rate at low SNR.

Algorithm Ref System Model Number of Goal

User
16 MISO downlink Multiple User Sum rate maximiza-
UnSupervised Learning tion
9 Single recieving an- Single user Gain enhancement
tenna
17 Single receiving an- Multiple user Data rate enhance-
Fedearated Learning tenna ment
28 mMIMO Multiple user UAV trajectory
13 TDD based single Multiple user Enhancement of
Deep Learning antenna KGR

12
 Algorithm Ref System Model Number of Goal
User
15 MISO downlink Multiple User Optimized beam-
Reinforcement Learning forming
20 MISO downlink Multiple User Energy efficiency
maximization
24 OFDM based single Multiple user Achievable rate max-
antenna imization
8 MISO Single user SNR maximization

5 Conclusion
We present a thorough analysis of the introduction of machine learning (ML) into LIS technologies
in this work. After giving a brief introduction to RIS, we went into how ML is implemented in LIS
and its drawbacks. Last but not least, the possible challenges and unresolved research issues with
applying ML in LIS have been noted. To properly handle the significant architectural needs in future
networks, significant technology challenges must be resolved. It is anticipated that the application
of LIS in the distribution of resources, processing signals, the practice of localization, visualization,
and dissemination will yield ground-breaking outcomes during the coming years. Numerous subjects
that could significantly affect unguided communication for the upcoming inception of networks can
be the subject of additional LIS and ML research.

References
[1] Ahmed Alkhateeb. “DeepMIMO: A generic deep learning dataset for millimeter wave
and massive MIMO applications”. In: arXiv preprint arXiv:1902.06435 (2019).

[2] Muhammad Alrabeiah and Ahmed Alkhateeb. “Deep learning for TDD and FDD
massive MIMO: Mapping channels in space and frequency”. In: 2019 53rd asilomar
conference on signals, systems, and computers. IEEE. 2019, pp. 1465–1470.

13
 [3] Mehmet Ali Aygül, Mahmoud Nazzal, and Hüseyin Arslan. “Deep Learning-Based Op-
timal RIS Interaction Exploiting Previously Sampled Channel Correlations”. In: 2021
IEEE Wireless Communications and Networking Conference (WCNC). 2021, pp. 1–6.
doi: 10.1109/WCNC49053.2021.9417591.

[4] JungSook Bae et al. “Overview of RIS-enabled secure transmission in 6G wireless

networks”. In: Digital Communications and Networks (2024).

[5] Ertugrul Basar et al. “Reconfigurable intelligent surfaces for 6G: Emerging hardware
architectures, applications, and open challenges”. In: IEEE Vehicular Technology Mag-
azine (2024).

[6] Robin Chataut and Robert Akl. “Massive MIMO systems for 5G and beyond net-
works—overview, recent trends, challenges, and future research direction”. In: Sensors
20.10 (2020), p. 2753.

[7] Ahmet M. Elbir et al. “Deep Channel Learning for Large Intelligent Surfaces Aided
mm-Wave Massive MIMO Systems”. In: IEEE Wireless Communications Letters 9.9
(2020), pp. 1447–1451. doi: 10.1109/LWC.2020.2993699.

[8] Keming Feng et al. “Deep reinforcement learning based intelligent reflecting surface
optimization for MISO communication systems”. In: IEEE Wireless Communications
Letters 9.5 (2020), pp. 745–749.

[9] Jiabao Gao et al. “Unsupervised learning for passive beamforming”. In: ieee commu-
nications letters 24.5 (2020), pp. 1052–1056.

[10] Fengxian Guo et al. “Enabling massive IoT toward 6G: A comprehensive survey”. In:
IEEE Internet of Things Journal 8.15 (2021), pp. 11891–11915.

[11] Meng Hua et al. “Intelligent reflecting surface assisted localization: Performance anal-
ysis and algorithm design”. In: IEEE Wireless Communications Letters (2023).

[12] Chenglu Jia et al. “Machine learning empowered beam management for intelligent re-
flecting surface assisted MmWave networks”. In: China Communications 17.10 (2020),
pp. 100–114. doi: 10.23919/JCC.2020.10.007.

14
[13] Long Jiao et al. “Machine learning-assisted wireless PHY key generation with recon-
figurable intelligent surfaces”. In: Proceedings of the 3rd ACM Workshop on Wireless
Security and Machine Learning. 2021, pp. 61–66.

[14] Sherif A Khaleel et al. “MTM-inspired graphene-based THz MIMO antenna config-
urations using characteristic mode analysis for 6G/IoT applications”. In: Electronics
11.14 (2022), p. 2152.

[15] Shicong Liu et al. “Deep denoising neural network assisted compressive channel estima-
tion for mmWave intelligent reflecting surfaces”. In: IEEE Transactions on Vehicular
Technology 69.8 (2020), pp. 9223–9228.

[16] Xiao Liu, Yuanwei Liu, and Yue Chen. “Machine learning empowered trajectory and
passive beamforming design in UAV-RIS wireless networks”. In: IEEE Journal on
Selected Areas in Communications 39.7 (2020), pp. 2042–2055.

[17] Donghui Ma et al. “Distributed rate optimization for intelligent reflecting surface
with federated learning”. In: 2020 IEEE International Conference on Communications
Workshops (ICC Workshops). IEEE. 2020, pp. 1–6.

[18] M Rezwanul Mahmood et al. “A comprehensive review on artificial intelligence/machine

learning algorithms for empowering the future IoT toward 6G era”. In: IEEE Access
10 (2022), pp. 87535–87562.

[19] Faisal Naeem et al. “Security and privacy for reconfigurable intelligent surface in 6G:
A review of prospective applications and challenges”. In: IEEE Open Journal of the
Communications Society 4 (2023), pp. 1196–1217.

[20] REM Non-orthogonal. “Multiple access networks: Deployment and passive beamform-
ing design”. In: arXiv preprint arXiv:2001.10363 (2020).

[21] Cunhua Pan et al. “An overview of signal processing techniques for RIS/IRS-aided
wireless systems”. In: IEEE Journal of Selected Topics in Signal Processing 16.5 (2022),
pp. 883–917.

15
[22] Haobo Song et al. “Unsupervised learning-based joint active and passive beamforming
design for reconfigurable intelligent surfaces aided wireless networks”. In: ieee commu-
nications letters 25.3 (2020), pp. 892–896.

[23] Abdelrahman Taha, Muhammad Alrabeiah, and Ahmed Alkhateeb. “Enabling large
intelligent surfaces with compressive sensing and deep learning”. In: IEEE access 9
(2021), pp. 44304–44321.

[24] Abdelrahman Taha et al. “Deep reinforcement learning for intelligent reflecting sur-
faces: Towards standalone operation”. In: 2020 IEEE 21st international workshop on
signal processing advances in wireless communications (SPAWC). IEEE. 2020, pp. 1–5.

[25] Faisal Tariq et al. “A speculative study on 6G”. In: IEEE Wireless Communications
27.4 (2020), pp. 118–125.

[26] Ruiqi Wang et al. “A wideband reconfigurable intelligent surface for 5G millimeter-
wave applications”. In: IEEE transactions on antennas and propagation (2024).

[27] Qingqing Wu and Rui Zhang. “Towards smart and reconfigurable environment: Intel-
ligent reflecting surface aided wireless network”. In: IEEE communications magazine
58.1 (2019), pp. 106–112.

[28] Qingqing Wu et al. “Intelligent reflecting surface-aided wireless communications: A

tutorial”. In: IEEE transactions on communications 69.5 (2021), pp. 3313–3351.

[29] Qingqing Wu et al. “Intelligent reflecting surface-aided wireless communications: A

tutorial”. In: IEEE transactions on communications 69.5 (2021), pp. 3313–3351.

[30] Qingqing Wu et al. “Intelligent surfaces empowered wireless network: Recent advances
and the road to 6G”. In: Proceedings of the IEEE (2024).

[31] Bo Yang et al. “Intelligent spectrum learning for wireless networks with reconfigurable
intelligent surfaces”. In: IEEE Transactions on Vehicular Technology 70.4 (2021),
pp. 3920–3925.

[32] Xianghao Yu et al. “Smart and reconfigurable wireless communications: From IRS
modeling to algorithm design”. In: IEEE Wireless Communications 28.6 (2021), pp. 118–
125.

16