Available online at www.sciencedirect.

com

ScienceDirect
ICT Express 9 (2023) 341–355
www.elsevier.com/locate/icte

Enabling technologies for AI empowered 6G massive radio access networks

Md. Shahjalala , Woojun Kimb , Waqas Khalidc , Seokjae Moond , Murad Khane , ShuZhi Liuf ,
Suhyeon Limg , Eunjin Kimg , Deok-Won Yunh , Joohyun Leei , Won-Cheol Leeh , Seung-Hoon Hwangf ,
Dongkyun Kime , Jang-Won Leed , Heejung Yuj , Youngchul Sungb , Yeong Min Janga ,∗
a Department of Electronics Engineering, Kookmin University, Seoul, Republic of Korea
b School of Electrical Engineering, Korea Advanced Institute of Science and Technology, Daejeon, Republic of Korea
c Institute of Industrial Technology, Korea University, Sejong, Republic of Korea
d School of Electrical and Electronic Engineering, Yonsei University, Seoul, Republic of Korea
e School of Computer Science and Engineering, Kyungpook National University, Daegu, Republic of Korea
f Department of Electronic and Electrical Engineering, Dongguk University, Seoul, Republic of Korea
g Department of Electronic and Electrical Engineering, Hankyong National University, Anseong, Republic of Korea
h Department of Electronic Engineering, Soongsil University, Seoul, Republic of Korea
i School of Electrical Engineering, Hanyang University, Ansan, Republic of Korea
j Department of Electronics and Information Engineering, Korea University, Sejong, Republic of Korea

Received 29 October 2021; received in revised form 25 May 2022; accepted 9 July 2022
Available online 16 July 2022

Abstract
Predictably, the upcoming six generation (6G) networks demand ultra-massive interconnectivity comprising densely congested sustainable
small-to-tiny networks. The conventional radio access network (RAN) will be redesigned to provide the necessary intelligence in all areas to
meet required network flexibility, full coverage, and massive access. In this respect, this paper focuses on intelligent massive RAN (mRAN)
architecture and key technologies fulfilling the requirements. Particularly, we investigate potential artificial intelligence algorithms for network
and resource management issues in 6G mRAN. Furthermore, we summarize the research issues in edge technologies and physical layer
intelligence on 6G network architecture.
© 2022 The Author(s). Published by Elsevier B.V. on behalf of The Korean Institute of Communications and Information Sciences. This is an open
access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Keywords: 6G; Massive radio access networks; Edge AI; AI-assisted networking

Contents

1. Introduction........................................................................................................................................................................... 342
2. Enabling technologies for 6G mRAN...................................................................................................................................... 344
2.1. New spectrum ............................................................................................................................................................. 344
2.2. Ultra-massive MIMO and beamforming ........................................................................................................................ 345
2.3. Reconfigurable intelligent surfaces ................................................................................................................................ 345
2.4. Non-terrestrial communication networks........................................................................................................................ 346
2.5. Ultra-dense networks ................................................................................................................................................... 346
2.6. Cell-free networking .................................................................................................................................................... 346
2.7. Quantum communication.............................................................................................................................................. 346
∗ Corresponding author.
E-mail addresses: mdshahjalal26@ieee.org (M. Shahjalal), woojun.kim@kaist.ac.kr (W. Kim), waqas283@korea.ac.kr (W. Khalid),
sjmoon@yonsei.ac.kr (S. Moon), mkhan@knu.ac.kr (M. Khan), shuzhiliu@dongguk.edu (S. Liu), 03070226@hknu.ac.kr (S. Lim), gate1180@hknu.ac.kr
(E. Kim), dhtor@naver.com (D.-W. Yun), joohyunlee@hanyang.ac.kr (J. Lee), wlee@ssu.ac.kr (W.-C. Lee), shwang@dongguk.edu (S.-H. Hwang),
dongkyun@knu.ac.kr (D. Kim), jangwon@yonsei.ac.kr (J.-W. Lee), heejungyu@korea.ac.kr (H. Yu), ycsung@kaist.ac.kr (Y. Sung), yjang@kookmin.ac.kr
(Y.M. Jang).
Peer review under responsibility of The Korean Institute of Communications and Information Sciences (KICS).

https://doi.org/10.1016/j.icte.2022.07.002
2405-9595/© 2022 The Author(s). Published by Elsevier B.V. on behalf of The Korean Institute of Communications and Information Sciences. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

3. Key stipulations for 6G mRAN .............................................................................................................................................. 347

3.1. Network flexibility for massive RAN ............................................................................................................................ 347
3.2. Energy-efficient massive access technologies ................................................................................................................. 347
3.3. Advanced technologies for massive interconnectivity...................................................................................................... 347
3.3.1. Cooperative massive MIMO-NOMA................................................................................................................. 347
3.3.2. Rate splitting multiple access........................................................................................................................... 348
4. Advanced AI/ML techniques for 6G mRAN............................................................................................................................ 348
4.1. Reinforcement learning ................................................................................................................................................ 348
4.2. Transfer learning ......................................................................................................................................................... 348
4.3. Federated learning ....................................................................................................................................................... 349
4.4. Quantum machine learning........................................................................................................................................... 349
4.5. Standardization timeline for AI/ML integrated 6G ......................................................................................................... 349
5. Edge technologies for intelligent 6G mRAN............................................................................................................................ 349
5.1. Intelligent hyper-connected edge networking ................................................................................................................. 350
5.2. Split computing with the hyper-connected edge networks............................................................................................... 350
5.3. Edge-offloading technology .......................................................................................................................................... 350
6. Intelligent PHY layer conceptualization for 6G........................................................................................................................ 350
6.1. Cognitive intelligent-based autonomous radio resource management................................................................................ 351
6.2. Intelligent channel coding and modulation .................................................................................................................... 351
6.3. AI-based channel estimation......................................................................................................................................... 351
6.4. Intelligent multiple access and spectrum sharing............................................................................................................ 351
7. Conclusion ............................................................................................................................................................................ 351
Declaration of competing interest ........................................................................................................................................... 351
Acknowledgments.................................................................................................................................................................. 351
References ............................................................................................................................................................................ 352

challenging and sophisticated requirements in performance,

1. Introduction reliability, and power/cost efficiency (see Table 1).
While communication service providers in leading coun- The evolution of radio access networks (RANs) up to the
tries are competing to validate and deploy commercial fifth- 5G is described in [9]. 5G RAN comes with a cloud computing
generation (5G) networks, research efforts have begun for concept and advanced antenna system that centralizes the radio
next-generation communications [1]. Smart societies are even access technology (RAT) processing and increases the compu-
more focused on data-centric-based use cases, such as virtual tational and data transfer speed. The so-called “split options”
and augmented reality (VAR), holographic projection, event- have been applied to the RAT protocol layer. The higher-layer
based communications, video-on-demand, remote surgery, and split supports multiple radio sites, frequency bands, and dual
brain-computer interface posing ultra-high data rate (Tbps connectivity with fourth generation, whereas the lower-layer
level), ultra-low latency (<1 ms), and ultra-high reliability split interfaces between the main and remote units. However,
(99.99999%). However, 5G networks ultimately can run into 6G will continue to modernize the existing RANs using arti-
several technical limitations and challenges when supporting ficial intelligence (AI) at the edge and cloud by establishing
such ultra-massive interconnectivity with immense computa- intelligent 6G massive RAN (mRAN). AI with powerful abil-
tional requirements. Thus, extensive research on the future ities in learning, sensing, reasoning, optimizing, and adapting
sixth-generation (6G) networking areas has been proposed. allows 6G networks to intelligently cause performance opti-
Concerning the 6G key performance indicators [2,3], the con- mization and adapt themselves to diverse service requirements.
sensus is that the requirements are higher than those of the In intelligent 6G mRANs, the traditional centralized AI will
5G. 6G envisions 1 Tbps peak data rate which is 100 times gradually evolve to decentralized and collaborative models
greater than the 5G. The control plane end-to-end latency is where the models are trained in decentralized edge servers.
considered to be maximum 1 ms which is 10 times lower Such advanced system will provide more privacy in train-
than the 5G. The connection density is expected to increase ing data because it shares only the parameters to the global
10 to 100 times (10–100 million devices/km3 ) for the 6G. model; hence efficiently reduce transmission delay and storage
6G will support 1000 km/h mobility and 100 to 1000 times overhead. Therefore, the intelligent 6G mRANs will establish
(0.1–10 Gbps/m3 ) increased traffic capacity. It should be noted RAN intelligent controller at the central cloud server and
that only a partial requirements can be satisfied at a spe- distributed edge AI at the edge cloud server. Providing network
cific 6G use case. 6G will construct a new 3D network flexibility to dynamic and adaptive network operations is the
paradigm with space–aerial–terrestrial–ocean integration and key stipulation for such complex and multidimensional 6G
bring emerging technologies, such as ultra-massive multiple- mRAN. Network flexibility can be achieved by providing a
input multiple-output (um-MIMO) [4], reconfigurable intelli- flexible frame format with scalable subcarrier spacing, adap-
gent surfaces (RISs) [5], holographic beamforming [6], and tive cell sizing, and customized network and radio resource
new spectrum (terahertz (THz) [7], optical [8]), to fulfill the management [2].
342
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

Table 1
List of acronyms used in the paper.
3D Three dimensions NOMA Non-orthogonal multiple access
5G Fifth-generation NTN Non-terrestrial network
6G Sixth-generation OFDM Orthogonal frequency division multiplexing
AI Artificial intelligence OWC Optical wireless communication
BS Base station QKD Quantum key distribution
CPU Central processing unit QL Quantum learning
CSI Channel state information QML Quantum ML
DL Deep learning QNN Quantum neural network
DM-MIMO Distributed massive MIMO QoS Quality of service
eMBB Enhanced mobile broadband QRL Quantum RL
FL Federated learning QSA Quantum search algorithms
GEO Geosynchronous equatorial orbit RAN Radio access network
GHz Gigahertz RAT Radio access technology
HAP High-altitude platform RF Radio frequency
LDPC Low-density parity-check codes RIS Reconfigurable intelligent surface
LEO Low Earth orbit RL Reinforcement learning
LOS Line-of-sight RSMA Rate splitting multiple access
MEC Mobile edge computing SDMA Space division multiple access
MISO Multiple-input single-output SWIPT Simultaneous wireless information and power transfer
MIMO Multiple-input multiple-output THz Terahertz
ML Machine learning TL Transfer learning
mMTC Massive machine-type communications UAV Unmanned aerial vehicle
mmWave Millimeter wave UDN Ultra-dense network
mRAN Massive radio access network um-MIMO Ultra-massive multiple-input multiple-output
MUD Massive user detection URLLC Ultra-reliable low latency communications
mURLLC Massive URLLC VAR Virtual and augmented reality
NN Neural network VLC Visible light communication

Recently, several studies have been articulated on the 6G specified. A clear distinction has been drawn between the
visions, enabling technologies, and requirements [3,10,11]. As primitive optimization techniques and potential AI/ML-based
an early-stage study, [10] provides trendy and novel research approaches in the future wireless communications. Addition-
directions on 6G use cases and key enabler technologies, ally, the standardization timeline for the integration of AI/ML
such as pervasive AI, contextual communications, quantum into 6G has been addressed. Moreover, for the intelligent 6G
communications and networks, and blockchain technologies. mRAN, the edge technologies and the physical layer chal-
Authors in [11], provided a forward-looking vision on 6G, lenges with potential solutions are provided in detail. The main
technological trends driven by exciting and underlying ser- contributions are summarized below:
vices, and concrete recommendations for the roadmap toward
• The key architectural innovations in 6G mRAN have
6G. Moreover, the emergence of AI and machine learning
been surveyed with an in-depth analysis of current and
(ML) algorithms based on big data, especially the application
future research trends.
of deep reinforcement learning (RL), federated learning (FL),
• We envision the technology requirements that support
and quantum learning (QL) have created new opportunities
future driving applications, including network flexibility,
for the 6G communication networks [3,12–15]. A compre-
energy-efficient massive access technologies, and tech-
hensive study on AI networks and models has been studied
nologies for massive interconnectivity.
in [3] that provided detailed discussions on AI-empowered
• We present potential applications of advanced AI and
6G network solutions in radio interface, traffic control, man-
ML techniques in 6G mRAN, such as RL, FL, transfer
agement, optimization, and security. Table 2 summarizes the
learning (TL), and QL.
current research progresses from various categories, such as
• We discuss developments and research issues on edge
visions toward 6G RAN, cloud-RAN, fog-RAN, aerial access
technologies, such as split computing and intelligence for
networks in 6G, RIS-assisted networks, and AI-empowered
hyper-connected edge networking and edge-offloading
6G. However, further research improvements are needed on
techniques.
several critical aspects of intelligent mRAN, such as key re-
• Some key physical layer aspects for next-generation
quired technologies, AI-based network management, edge and wireless communications have been addressed, including
cloud technologies, and PHY layer intelligence. Therefore, this intelligent channel coding, modulation, estimation, and
survey fills a gap of the key technological requirements and prediction.
research issues associated with the intelligent mRAN architec-
ture for 6G communications. We have quantitatively surveyed The rest of this paper is organized as follows. Section 2
the recent technological advancements and their future chal- presents the key technologies for the 6G mRAN architecture.
lenges with possible solutions. Besides, the key technology Section 3 discusses the requirements and potential AI tech-
requirements to support 6G mRAN architecture have been niques for the network management are given in Section 4.
343
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

Table 2
Summary of current research progresses in 6G.
References Main focus Research objective and contributions
[3,10,11] 6G vision, use Provides a philosophical view of the 6G wireless system,
cases, technologies networking trends, and research directions on the 6G use
cases, key enabler technologies, and requirements in a
system-level perspective.
[2,16] 6G RAN Presents some key driving applications for 6G that
architecture necessitates RAN transformation. Discusses key
requirements and enabling technologies in 6G RAN
architecture.
[10] Challenges in 6G Discusses the open research issues and challenges.
networking
[17,18] Study on cloud-RAN Cloud-RAN virtualization, benefits and challenges,
minimizing power consumption in cloud-RAN.
[1,19,20] Study on fog-RAN Establishing delay-sensitive fog-RAN, its design
architecture, and key problems in the centralized learning
paradigm.
[3,12–15] AI-empowered 6G Studies AI networks and models, provide detailed
discussions on AI-empowered 6G network solutions in
radio interface, traffic control, management, optimization,
and security.
[5,21,22] RIS-assisted 6G Minimizing the total transmission power and maximizing
networks the sum and secrecy rates of the downlink MISO system.
[2,23–25] 6G aerial access Integration of non-terrestrial networks to provide broadband
network services to the under-served and remote areas and
maximize the coverage.
This article Intelligent massive Provides an architecture of 6G intelligent massive RAN,
RAN architecture in key technology requirements, AI applications in the 6G
6G network solutions, intelligence at the edge technologies,
and physical layer.

Sections 5 and 6 present the intelligent edge technologies and capacity. The main part of THz bands used for cellular com-
physical layer intelligence, respectively. Finally, we draw our munications, as recommended by the international telecom-
conclusion in Section 7.
munication unit is 275 GHz to 3 THz. To exploit higher
mmWave and THz frequencies, new architectural designs with
2. Enabling technologies for 6G mRAN
denser deployments of tiny cells and high-frequency mobility
This section provides an in-depth discussion on the poten- management techniques are required. In THz network-based
tial key enabling technologies to be considered for developing VR experience, data rates of 18.3 Gbps were achieved with
6G mRANs, as shown in Fig. 1. These enabling technologies
99.999% reliability [26]. Recent studies show that optical
empower 6G mRANs with flexibility, massive interconnectiv-
ity, and spectral and energy efficiency. Below we discuss the wireless communications (OWCs) [27] contribute to the 6G
potential, technological advancements, quantitative require- device-to-access networks and network-to-backhaul connectiv-
ments, challenges, and possible solutions for these enabling ity. Since well-known OWC technologies, such as free-space
technologies.
optical [28], visible light communication (VLC) [29], and
light-fidelity [30], have been extensively used in inter-satellite
2.1. New spectrum
communication, under-water communication, VAR, vehicle-
As data rate requirements grow higher for 6G, new spectra to-everything technology, and indoor and outdoor communica-
with higher bandwidth support are necessary. Thus, further tions, they will be further used in 6G with the support of band
development of millimeter wave (mmWave) (18–100 GHz) requirement in both terrestrial and non-terrestrial areas. An
um-MIMO technologies is required along with its high rate ultra-high data rate (100 Gbps) VLC services can be realized
of transmission. THz provides tens of wireless Gbps trans-
via differential wavelength division multiplexing [31]. The
mission rates that ensure feasible high data rate, short-range
broadband wireless communication. Moreover, it has a strong data rate can be further enhanced by using parallelization of
anti-interference ability and good directionality with a large future microLED arrays to fulfill the 6G requirements.
344
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

Fig. 1. Proposed intelligent mRAN system for 6G wireless communications.

2.2. Ultra-massive MIMO and beamforming environments can unleash the full potential of future 6G
networks. In this respect, RISs have emerged as a promis-
THz bands will complement mmWave communications in ing greener physical layer technology for reconfiguring the
future 6G networks. The high-frequency bands offer ultraw- propagation environment and providing supplementary links to
ide bandwidths and enable ultra-fast, ultra-low-latency com- assist transmissions, thereby improving both the QoS and radio
munications, superior reliability, and massive connectivity.
connectivity [36]. With passive reflecting elements, a typi-
However, these bands offer high transmission loss. To tackle
this problem, um-MIMO arrays have been introduced which cal RIS reflects the signals independently using controllable
enable multi-beam radiation and flexible beamforming and en- phase shifts and enables directional signal enhancement or
hance spectral efficiency and system capacity. The hybrid and nulling. Through intelligent placement and passive beamform-
full-digital beamforming arrays are promising candidates for ing, RIS provides extraordinary benefits to the 6G wireless
mmWave systems [32]. For mmWave communications, [33] paradigm, including signal boosting, interference suppres-
proposed a multi-user scheduling algorithm based on block sion, secure transmission, and wireless information and power
diagonalization precoding to effectively reduce the bit error transfer [37]. In [38], the RIS-assisted transmissions are com-
rate and improve the system spectrum efficiency of the 6G pared to the direct point-to-point LOS transmission without
V2X um-MIMO systems. In general, the THz band brings small-scale fading. In an indoor environment, a RIS exhibits
crucial benefits, it also poses tremendous research and engi- significant results in error performance and achievable data
neering challenges. The antenna design and beamforming in
rates. Further, it is theoretically possible to achieve the per-
THz bands remain immature and unreliable compared with
those operating at mmWave bands and require extensive in- formance of point-to-point line-of-sight (LOS) communication
vestigation owing to stringent design requirements [34]. The when the number of reflecting elements is increased suf-
rapid developments of physical layer technologies with novel ficiently. It is demonstrated that a RIS with a sufficiently
design solutions and precise channel models should properly large number of reflecting elements can also provide an ef-
characterize the THz propagation over both the hardware fective solution in an outdoor environment at a 2.4 GHz
and air. [35] showed that um-MIMO antenna arrays with operating frequency even if the transmission link is blocked.
proper configuration can overcome the distance and power It is shown that the RIS-assisted systems provide enhanced
limitations of the THz-communications. The AI/ML-based performances even with system imperfections, such as discrete
intelligent beamforming technology can be used to overcome phase-shift/amplitude, imperfect channel phase estimation, and
the THz propagation challenges in future 6G networks.
phase-dependent amplitude. Although studies on RIS cover
2.3. Reconfigurable intelligent surfaces several aspects, efforts should tackle some challenges, includ-
ing channel estimation, robust design using imperfect channel
The 6G infrastructure is challenged by the THz and parameters, security, and privacy-related issues, joint active
mmWave propagation characteristics and will require inno- and passive beamforming designs, mobility management, and
vative solutions. The controllable and programmable radio AI/ML-enabled design and optimization.
345
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

2.4. Non-terrestrial communication networks knowledge assumed over the underlying ultra-dense environ-
ments. Since the traditional iterative algorithms for solving op-
Non-terrestrial communication networks (NTNs) will com- timization problems are inefficient due to the high-complexity
plement the terrestrial networks in future 6G networks. The of large-scale UDNs, efficient mathematical techniques, such
integration of satellite constellation, high-altitude platforms, as game theory and real-time optimization, for optimizing
and unmanned aerial vehicles (UAVs), etc., into the existing the 6G UDNs under stringent radio resource constraints are
dense terrestrial networks, will improve the radio access ca- discussed [45].
pability and will unlock the cost-effective and high-capacity
ubiquitous 3D global connectivity. The aerial base stations can 2.6. Cell-free networking
deliver unified and affordable high-level QoS to the end-users,
provide an unprecedented degree of freedom for real-time Recently, um-MIMO and distributed massive MIMO (DM-
control and computation, and offload core networks in heavy- MIMO) have been introduced to provide continuous connec-
load situations [39]. The development of NTNs will be favored tivity anywhere and anytime using large collocated antenna
by recent technological innovations, e.g., multi-layered archi- arrays and spread-out antennas, respectively [46]. Deploy-
tecture, and antenna and spectrum advancements. For example, ing THz waves with both um-MIMO and DM-MIMO incurs
compared to a standalone deployment, a multi-layered sce- several challenges, including environmental conditions. There-
nario with a high-altitude platform (HAP) as a bridge for fore, networking technology is needed to bring the antennas
the low earth orbit (LEO) communication offers a much- closer to the UE to provide a high data rate and fewer ef-
improved capacity broadcasting capability and a performance fects from the environmental conditions. Thus, a candidate
boost of +250% [40]. The results in [41] also show that networking technology called “cell-free” was recently intro-
the GEO-to-HAP-to-Earth configuration (i.e., HAP relays) in duced to implement 6G technology using THz waves and
multi-layered hierarchical NTNs ensures a 6x better capac- jointly serving a UE with multiple access points (APs), thereby
ity than point-to-point GEO transmissions. However, several providing a data rate of 1 to 3 Tb/s [47]. The APs in cell-free
open issues related to NTNs require long-term research and networking technologies are connected to a central processing
proper protocol design. The inherent challenges include chan- unit (CPU) via a fronthaul link. Furthermore, the CPUs are
nel modeling, multi-dimensional air traffic management, ar- connected using a backhaul link. To implement 6G using cell-
chitecture technologies, higher-layer design, synchronization free networking, Ericsson introduced a cell-free technology
called the radio stripes [46]. Here, the antennas and associ-
issues, spectrum sharing, mobility management, constellation
ated antenna processing units are serially located inside the
management, and PHY/antenna design [42]. In sum, the effec-
same cable, which provides synchronization, data transfer,
tive integration and interoperability between NTNs and exist-
and power supply through a shared bus. Similarly, several
ing terrestrial networks are essential requirements to achieve
research teams from Sweden, China, and the United Kingdom
the 6G coverage and capacity goals.
published several research articles on protocol designs for
enabling physical layer communication in cell-free technolo-
2.5. Ultra-dense networks gies [47,48]. [49] addressed the system throughput and user
equipment fairness trade-off problem in cell-free um-MIMO
Ultra-dense networks (UDNs) will be a key technology systems and developed a novel scheme to jointly optimize
for 6G to serve the massive number of users, enhance the the access point, power allocation, and beamforming design,
coverage/quality of service (QoS), and increase the network outperforming existing methodologies, including conventional
capacity. Recently, numerous studies have been proposed for ones sum-rate maximization and max–min worst case.
6G UDNs to deal with the challenges of the deployment
of large-scale networks. In mmWave UDNs, accurate instan- 2.7. Quantum communication
taneous CSI is difficult to estimate owing to the densifi-
cation of mmWave BSs, thereby posing stiff challenges to Quantum communication opens a new era of envisioning
user association. To solve these challenges, [43] proposed a extremely high data rates, reduced computational complex-
novel ML-based user association approach that supports multi- ity, and cyber-security requirements in the future 6G and
connectivity in densely deployed mmWave networks. Test next-generation communications [12]. In addition to the exist-
results of [43] show that the proposed approach achieved good ing classical communication networks, quantum-assisted com-
performance using a few training samples and without CSI. munications can enhance various aspects of next-generation
In [44], a deep RL-based mobility load-balancing algorithm telecommunications, such as optimal massive user detection
was proposed with a two-layer architecture to solve large-scale (MUD), channel estimation, and incorporating many more
load-balancing problems for UDN. The two-layer architecture functionalities. Quantum communication can be realized via
alleviates the global traffic variations by dynamically grouping quantum internet that connects quantum machines to solve
small cells into self-organized clusters using their historical certain computational problems substantially faster than clas-
loads. An off-policy DR-based MLB algorithm autonomously sical computers. Here, information is exchanged within an
learns the optimal mobility load-balancing policy under an environment that harnesses the principles of quantum mechan-
asynchronous parallel learning framework, without any prior ics. Governed by the principle of quantum mechanics, the
346
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

linear combination of the logical values is used to transmit RANs. For instance, by measuring network requirements, such
data through the quantum nanoscale objects (i.e., electrons as throughput, latency, and reliability, the subcarrier spac-
and photons) called a qubit. Thus quantum communication ing can be controlled intelligently. Wider subcarrier spacing
exploits the collective properties of quantum states, such as is suitable for high-reliability and low-latency applications,
superposition and entanglement, which significantly boost the whereas narrower spacing is used for low-rate machine-type
computational capacity. communications [2].
For 6G um-MIMO channels, the first requirement is the
performance enhancement of the system by improving the 3.2. Energy-efficient massive access technologies
channel estimation and MUD techniques. The complexity of
channel estimation using massive user data can be reduced by To support massive access to 6G mRAN with limited
employing quantum search algorithms (QSA). Authors in [50, time–frequency resources, advanced technologies such as um-
51] stated that for unsorted big data storage with N entries,
(√ ) MIMO and non-orthogonal multiple access (NOMA) using the
the computational complexities can be reduced to O N spatial and power or code domains, respectively, are consid-
using QSA, such as Dürr-Høyer algorithm, quantum-assisted ered. However, both um-MIMO and non-orthogonal multiple
genetic optimization, and quantum-assisted repeated weighted access (NOMA) inevitably involve extremely high energy
boosting search. The proposed QSA in [50], achieved not only consumption due to their highly complex signal processing
a lower complexity but a better performance than did clas- and hardware architectures. Thus, the energy efficiency should
sical approaches in MUD, and channel estimation problems be considered to effectively support 6G mRAN. One so-
in various MIMO-orthogonal frequency division multiplexing lution is to exploit RISs in wireless communications [37].
(OFDM) systems. Since RIS does not require radio frequency (RF) chains and
complex signal processing, energy consumption significantly
3. Key stipulations for 6G mRAN decreases. Another energy-efficient technology is energy har-
vesting since energy can be obtained from ambient sources.
In this section, the key stipulations based on the key tech- Recently, simultaneous wireless information and power trans-
nologies are studied to effectively support the 6G mRAN. fer (SWIPT), where RF signals are exploited for both transfer-
In specific, three key requirements including network flexi- ring energy and transmitting data, has been studied [52]. Since
bility, energy-efficient massive access technologies, and ad- signal processing and resource allocation in um-MIMO incur
vanced technologies for massive interconnectivity are provided high complexity, sophisticated resource allocation algorithms
with discussion of the up-to-date researches. In the follow- should be studied. Alternatively, energy-efficient resource al-
ing, we envision several possible technological solutions in location algorithms that achieve a small gap from the optimal
detail to support the future driving applications for these key resource allocation should be further studied in RIS-aided
requirements. SWIPT systems.

3.1. Network flexibility for massive RAN 3.3. Advanced technologies for massive interconnectivity

Except for the enhanced mobile broadband (eMBB), ultra- Massive interconnectivity to support a large amount of
reliable low latency communications (URLLC), and massive data transmissions, computing, and data storing should be
machine-type communications (mMTC) use cases in 5G, 6G considered in 6G mRAN, unlike in the 5G mMTC where
is envisaged to support more diverse requirements and so- various users intermittently transmit small-size data. UDN
phisticated service types. These candidate service types are is a key approach where low-power BSs are deployed ex-
reliable eMBB, massive URLLC (mURLLC), 3D-integrated tremely dense [53]. However, exploiting UDN alone cannot
communications, ultra-high-speed with low-latency communi- fully support massive interconnectivity owing to greatly in-
cations, extremely low-power communications, long-distance creased inter-cell interference. Thus, advanced technologies
and high-mobility communications, human-centric services, should be employed to ensure reliable communication while
etc. [16]. Those services are more inclusive to serve the re- supporting massive interconnectivity.
quired applications in 6G. Several aspects that require support
in the 6G RAN flexibility include adaptive cell sizing, flexible 3.3.1. Cooperative massive MIMO-NOMA
frame structure, dynamic radio resource management, adap- In the um-MIMO system, hybrid precoding is exploited
tive bandwidth scaling, and virtualizing network functions. because of hardware cost and power consumption. However,
Adaptive cell sizing supports massive access in the UDN. because of multiple RF chains and highly complex signal pro-
It is realized by the shift from small-to-tiny cell networks cessing in the hybrid precoder technique, um-MIMO cannot
with an increase in data transmission rate and high-frequency support massive interconnectivity for 6G mRAN. Another key
user mobility support. For example, cell sizes are changed candidate technology that supports massive interconnectivity
adaptively for areas of high user densities to increase spec- is NOMA. However, NOMA involves interference cancelation
tral efficiency and decrease cell range bias [2]. The flexible at the receiver, which incurs high computational complexity.
frame structure, realized by scalable OFDM numerology, be- Thus, instead of separately employing um-MIMO and NOMA,
comes the underlying foundation for flexible, intelligent 6G the network capacity is effectively increased by cooperatively
347
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

applying both um-MIMO and NOMA [54]. Authors in [55], 4.1. Reinforcement learning
proposed coordinated beamforming techniques to serve users
with increased network capacity while controlling inter-cell Recently, several RL algorithms that adopt deep neural
interference for cell-edge users. Furthermore, authors in [56], networks and other techniques for challenges in RL, such as
studied an optimization problem to maximize the number exploration, have been proposed [61,62]. Additionally, multi-
of admissible users under the total transmission power and agent RL (MARL), an extended framework of RL to multi-
individual-rate constraints. In [57], the authors showed that agent systems, is gaining increasing attention [63,64]. RL has
NOMA techniques enabling the transmission of both unicast been considered and investigated as a major tool for optimizing
and multicast achieved higher network capacity under the numerous aspects of 6G networks from network slicing to
um-MIMO system. However, high complexity still exists in radio access control. This is because the network architecture
cooperative um-MIMO-NOMA, leading to greatly increased and operation in 6G networks are expected to be complicated,
energy consumption. Thus, efficient resource allocation algo- thereby precise modeling of such networks is very challenging
rithms and signal processing techniques need further study to and network optimization based on conventional model-based
support massive interconnectivity. approaches will become inefficient. Thus, RL can provide
an effective solution to the optimal operation of such 6G
3.3.2. Rate splitting multiple access networks. For instance, RL has been applied to optimize net-
The rate splitting multiple access (RSMA) is considered work access problems [65–72], rate control problems [73,74],
a promising candidate for massive interconnectivity. Unlike
resource management problems [75–78], caching and offload-
NOMA and space division multiple access (SDMA), RSMA
ing [79–89]. Authors in [65] developed dynamic multichannel
partially decodes the interference and treats it as noise. Re-
access problems as a partially observable RL setting, where
sults have shown that RSMA outperforms both NOMA and
a single agent chooses one of N channels to succeed the
SDMA [58]. Moreover, it is robust to imperfect channel in-
transmission using partial information of the channel state.
formation [59], and pilot contamination [60]. Furthermore, it
Also, authors in [66] modeled multiple transmitters as multiple
provides low computation complexity, thereby achieving high
energy efficiency. RSMA combined with RIS supports massive agents and maximize their contributions to the downlink spec-
interconnectivity and reduces computational complexity. Since tral efficiency minus interference to other transmitters. [90]
RSMA is robust to imperfect channel information, the RIS- proposed a Q-learning-based transmission scheduling method.
aided communication that is vulnerable to channel estimation The proposed method is to maximize the system throughput
errors benefits from RSMA. Since it is a flexible scheme by learning the proper strategy to transmit packets from the
that can be jointly exploited using various technologies and different buffers. [91] applies deep reinforcement learning to
outperforms NOMA or um-MIMO when used alone, it is one optimize the quality of experience (QoE) of the dynamic adap-
of the strongest candidates to be exploited for 6G mRAN. tive streaming over HTTP. The architecture proposed in [91]
An efficient resource allocation algorithm for RSMA in the combines feed-forward and recurrent deep neural networks
um-MIMO system requires further study. that outperforms the state-of-the-art algorithms. In addition
to the aforementioned examples, RL has been adopted to
4. Advanced AI/ML techniques for 6G mRAN improve the performance of various optimization problems
in 6G network. In [92], adopted multi-agent reinforcement
In 6G wireless systems, complicated optimization problems
learning where each user is trained by DQN, to maximize
should be considered under variable network topology and
the network utility in multichannel wireless networks. The
application scenarios. In such cases, we can find a solution
with two different approaches. The first approach is to use algorithm facilitates multiple user to train to maximize the
an analytical optimization. To this end, the separation of the objective in distributed manner. As described above, the afore-
original complicated problem into several simple problems mentioned requirements of 6G, such as network flexibility and
and/or the relaxation, which approximates the original problem performance improvement, can be satisfied by leveraging RL
by an easier problem, e.g., a convex problem, may be needed. algorithms.
However, the obtained solution can be far from the optimal
one because of such simplification and relaxation. The second 4.2. Transfer learning
approach is to use an AI/ML approach. Although most of
the problems in 6G networks with complex topology and TL refers to transferring learned knowledge from expe-
nonlinear components can be solved with the AI/ML methods, rienced/source tasks to new/target tasks. By leveraging the
the careful design of an AI/ML model, selection of appropriate learned knowledge, TL can dramatically improve data effi-
training data set, and ways to guarantee the convergence and ciency, which is a core issue in many practical problems,
increase the convergence rate of the designed algorithm are especially ML-based approaches.
required. Consequently, the selection of an approach to solving TL has been investigated and successfully applied to com-
the problem highly depends on a given condition and we munication system optimization, especially in ML-based opti-
cannot clearly sort out them. mization problems. For example, [93] proposed an RL
Therefore in this section, the AI-based methods for 6G are framework for energy-saving problems in RANs and combined
discussed in detail. We have also discussed about the integra- the proposed framework with TL to speed up the learning. The
tion of AI/ML into the 6G standardization timeline in brief. policy trained in the source task was used to update that for the
348
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

target task. Furthermore, authors in [94] proposed to optimize leveraging FL, the tail distribution of the network-wide queues
resource management using RL by minimizing the total power is trained accurately and the extreme events are reduced. In
consumption in fog RANs. Also, they integrated the proposed this regard, FL improves network flexibility resulting that the
RL with TL to improve the learning speed. They transferred requirements of 6G mRAN can be fulfilled.
the RL model parameters, which were trained on source envi-
ronment with certain caching service capabilities, to a model to 4.4. Quantum machine learning
be trained for the environments with different caching service
capabilities. Authors in [95] adopted TL technique to predict Ultra-large data exchanges that incorporate more sophisti-
the channel quality indicator across different wireless chan- cated applications envisioned in 6G put security and privacy
nels. They transfer the trained model by adding new layer to in greater threats. To tackle the wide range of challenges in
the previous model while fixing the first layer. Thus, by lever- 6G security, quantum computing-assisted communication is
aging TL, the energy efficiency, which is one of requirements employed to achieve high reliability in 6G networks by devel-
of 6G mRAN, can be improved. Added to the aforemen- oping the quantum key distribution (QKD) technique. QKD,
tioned examples, TL has been widely applied to optimize also known as quantum cryptography, exchanges a secret
communication systems including resource management [96– key between two lawful parties through the quantum chan-
100], channel estimation [95,100,101], caching [102–104], nel by providing absolute certainty. Potentially challenging
and localization [105–107] for improving the efficiency of research paradigms are open to ML integration and quan-
learning. tum cryptography. Quantum ML (QML) will enhance cyber
security by providing absolute randomness through the com-
4.3. Federated learning munication links in 6G. QML can assist in several areas by
enhancing security in terrestrial networks and NTNs, ocean
FL learns ML models using data from distributed devices communication, THz, and optical communications [12]. Au-
without sharing local data with other devices. Contrary to thors in [14], investigated a user grouping problem in NOMA
traditional distributed optimization methods, FL considers four by applying quantum neural network (QNN) and RL-aided
key properties [108]. First, the distribution of data is non-IID. QNN (RL-QNN) to reduce complexity and improve perfor-
In practice, data distributions from different devices depend on mance. RL-QNN showed better results considering the average
the local situation, and thus, non-IID data distribution is con- loss in such a problem. How an agent treats the environment
sidered in the FL setting. Second, data are unbalanced; i.e., the in the RL algorithm has been discussed in Section IV(A).
amount of data collected can be different depending on each Records show that it uses feedback and action to enhance
device use. Third, there exist massively distributed devices. performance. Using the quantum superposition and parallelism
Lastly, the distributed devices have limited communication. By properties, the quantum RL (QRL) can speed up performance
considering these aforementioned properties, FL assumes the trade-offs, such as convergence, balancing, and optimality.
system consists of one centralized server and a fixed set of Also, recent advancements of QRL techniques in spectrum as-
clients. signments, management, and allocation have been investigated
For future wireless communication systems, which provide in [120].
seamless connectivity of massive devices, FL has been in-
vestigated as a solution for distributed optimization problems 4.5. Standardization timeline for AI/ML integrated 6G
in communication systems considering privacy and limited
communication resources. In particular, FL has been applied 5G Advanced will be standardized in 3GPP Rel-18 phase
to optimize mobile edge computing (MEC) and IoT systems, and as a bridge between 5G and 6G technologies. AI/ML
such as mobile network optimization [109–111], edge caching technologies based on 5G Advanced will also be applicable
optimization [112–114], privacy [115,116], and vehicular net- to 6G. Whereas, AI/ML-Air Interface, as one of the Rel-18
work optimization [117–119]. Authors in [112] addressed the RAN1 items, will be used for performance enhancement or
cache replacement problem in all edge devices based on RL complexity/overhead reduction. In addition, AI-enabled NG
and integrated RL with FL for intelligent resource manage- RAN will address data collection enhancements and signaling
ment in MEC systems. Furthermore, authors in [115] proposed support. The study item for 6G will be launched in Rel-20
an FL-based content caching scheme that keeps local private phase (2025), which requires higher performance compared
data from leaking by adopting hierarchical architecture. [119] to 5G to meet future service requirements. Therefore, the
introduced two-dimensional contract theory to enable the in- standardization of AI/ML for 5G Advanced in Rel-18/19 will
teractions between the centralized server and vehicular clients help to advance the process of AI/ML standardization for 6G.
for addressing the resource asymmetry. In addition to the
aforementioned applications, optimization problems for MEC,
5. Edge technologies for intelligent 6G mRAN
including edge caching and vehicular network optimization
have been adopted in the FL framework for achieving collab- Intelligent edge computing services bring new and valu-
orative learning of edge devices. Authors in [117] proposed able services to the end-users. These include high-speed data
a FL framework to minimize the power consumption of ve- uploading and downloading, real-time service provision, and
hicular users with reliability and low latency constraints. By intelligent solutions to smart systems such as factories and
349
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

industries. However, at the edge, it is a challenging job to computing has been adopted into the image classification do-
upload and process the data in real-time. Therefore, intelli- main using deep neural networks where the target entity is an
gent solutions are needed which could be embedded with a input, which creates various problems, such as compression-
hyper-connected edge. related problems. Moreover, many other unexplored domains
can take the advantage of split computing. For instance, real-
5.1. Intelligent hyper-connected edge networking time monitoring of smart health care systems through wearable
sensors is a good example of adopting split computing to share
In this high-speed networking and connectivity era, edge significant information between edges and sensors, such as
computing has shown promising results in data processing, home routers and cellular phones.
storing, and prompt delivery to end-users. Several limitations
abound in cloud computing, one critical issue being how to
send data to long distances for processing and storing [121]. 5.3. Edge-offloading technology
However, edge computing solves this challenge by bringing
the computing nearer to the devices, such as autonomous Wireless edge-offloading technology can be classified into
cars, medical devices, smartphones, and computers. Because two parts, namely, the network traffic offloading and compu-
of these advantages, edge computing approaches have been tation offloading. Femtocaching introduced a novel wireless
adopted in various telecommunication companies, automobile edge caching architecture and proposed an efficient algorithm
industries, smart factories, etc. With the recent launch of for proactively caching popular content. The architecture in-
5G, real-time connectivity can be addressed even with an cludes a set of edge caches deployed at BSs, which refresh
enormous number of users connected to the internet daily. popular content. Several research papers [124,125] extended
Thus, providing services to these users in real-time needs
the femtocaching architecture. Lyu et al. [124] designed co-
the deployment of many edge devices nearer to these users.
operative edge caching where multiple BSs jointly optimized
Furthermore, this increasing number of users connected to
content placement. Sadeghi et al. [125] considered a hierarchi-
the internet has necessitated intelligent and hyper-connected
cal cloud–edge caching model where the cloud and edge stores
edge networking. Mobile phones, remote work sites, sensors,
computers, etc. generate a huge amount of data, and sending files according to global and local file popularities, respec-
all data to the edge is an inefficient idea. Recently, new tively. They modeled the Spatio-temporal popularity variations
lightweight techniques such as FL and TL are introduced to using a Markov chain model and solved them with RL. Similar
overcome the challenge of training a ML algorithm locally to cloud computing, computation offloading in edge networks
and with fewer data samples. However, owing to their recent handles computation tasks for end devices, in which the end-
introduction, these solutions require extensive experimentation to-end delays are smaller than central clouds. Also, Kwak
and testing. Furthermore, the idea of hyper-converged edge et al. [126] considered a mobile code offloading framework
computing has been developed to reduce the network strain with heterogeneous task types, which minimizes energy con-
by reducing the number of samples sent to the edge server. sumption in end devices. Next, Chen et al. [127] studied code
Therefore, integrating intelligence to hyper-converged edge offloading decisions among multiple mobile users with com-
computing can reduce the network load and deployment of mon wireless access using a game-theoretical approach. Kim
edge devices nearer to the sources. et al. [128] optimized an integrated edge-offloading frame-
work using the Lyapunov drift-plus-penalty technique in both
5.2. Split computing with the hyper-connected edge networks competition and cooperative scenarios among the offloading
service provider and mobile users. For future 6G networks,
Split computing is an intermediary technology between learning-based resource allocation is the fundamental idea for
edge and mobile computing. Its main concept involves split- solving dynamical changes in mobile traffic demands and
ting the application process into the head and tail segments computations. Also, where to deploy edge devices and how
that are divided and deployed at the edge and mobile devices. much resources to invest in the edge devices are important
For instance, authors in [122] split an artificial neural network issues to study.
model among network entities to optimize model performance.
The total conclusion time of a task in split computing consists
of three factors: edge processing time, device processing time, 6. Intelligent PHY layer conceptualization for 6G
and communication delay. To reduce the processing time to
those of edge and mobile computing, split computing faces the Facing the higher peak rate, capacity requirements, and
main challenge of minimizing the communication delay with spectrum efficiency requirements of 6G mobile communica-
a tiny segment of computation left on the device to transfer tions, it becomes an inevitable choice to address limited spec-
the information to the server. In some exceptional cases, split trum resources problem in edge computing, future multiple
computing fails to meet the criteria as constructed in [123]. access and new spectrum sharing technologies, new channel
The main challenge is in designating a part of the overall coding technologies, and propose highly accurate intelligent
task to a low computational power device with the result of channel prediction technologies. In this section, more detail
transmitting a small amount of information. Currently, split discussion is provided on those issues.
350
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

6.1. Cognitive intelligent-based autonomous radio resource 6.4. Intelligent multiple access and spectrum sharing
management
Multiple access technology defines how to share the band-
The 6G RAN technology must consider efficient ways to width resources between mobile users. Presently, 5G mobile
manage the limited radio resources to provide stable services communication systems have used 3–6 GHz and 24–50 GHz
when converting from 5G RAN with a centralized structure
frequency bands. Facing the higher peak rate, capacity require-
to an edge computing structure. One key approach is cogni-
ments, and spectrum efficiency requirements of 6G mobile
tive intelligent-based autonomous radio resource management,
communications, it becomes an inevitable choice to be more
including learning, reasoning, and optimization engines. How-
suitable for future multiple access technologies and the ap-
ever, the optimization process for interference and coexistence
between neighboring cells makes allocating resources quickly plication of new spectrum sharing technologies. To meet the
within a limited time difficult and wastes valuable-optimized demand to use spectrum resources in the future 6G system,
solution information [129]. Therefore, the performance of expanding the available spectrum, such as the THz and visible
cognitive intelligent-based autonomous radio resource man- light spectra, is necessary. Furthermore, the spectrum usage
agement is improved using Monte Carlo algorithm-based inter- rules must be changed and the current status of authorized
ference analysis and data mining technology to provide stable carrier usage be broken. Also, the allocation and use of spec-
and rapid service [130]. trum in a more flexible way must be sustained to improve
the use of spectrum resources. Spectrum sharing can play
6.2. Intelligent channel coding and modulation an important role in the upcoming 6G era. In particular, the
spectrum sharing of unlicensed mmWave bands, especially
Channel coding methods based on turbo codes, low-density the sharing of unlicensed spectrum above 60 GHz, can meet
parity-check codes (LDPC), and polar codes have received the high-performance requirements of the 6G ultra-large-scale
widespread attention. They not only approach the Shannon internet-of-things [135].
limit infinitely but have achieved great error correction ability
and code rate and code length reconstruction and support
7. Conclusion
hybrid automatic re-transmission requests, and the complex-
ity [131]. LDPC codes have been employed to protect 5G
The recent evolution of the internet-of-everything expands
standards owing to their flexibility in coding rate, coding
to a new connection paradigm of industrial internet-of-things,
length, and decoding delay and their ability to easily support
internet-of-people, and internet globalization services. The ser-
hybrid automatic repeat request [132]. Recently, improving the
vices will create diverse and multidimensional requirements
performance of communication systems using AI technology
of near-instant, ubiquitous, and unlimited connectivity sig-
has become a research hotspot, and is used in data-driven
analysis, inference, and decision-making technology [133]. nificantly increasing mobile data traffic in future wireless
Traditional channel coding mechanism design mainly uses networks. Therefore, in this paper, we investigated the driving
coding theory to optimize coding performance, while AI- issues associated with the intelligent 6G mRANs architecture,
driven channel coding mechanism adopts AI-based schemes, main stipulations supporting novel technologies, and network-
DL, NN, and genetic algorithms. ing solutions at the edge and physical layer. Furthermore, we
prioritized network intelligence and brought a depth analysis
6.3. AI-based channel estimation on the potentials of AI and ML models in network manage-
ment, resource allocation, spectrum sharing, edge networking,
To meet the requirements of greater communication capac- and security. Also, we highlighted key research issues and
ity, 6G will adopt the ultra-large-scale MIMO technology with challenges in each aspect of the 6G mRANs.
accurate channel state information for the efficient application
of MIMO. However, the channel estimation problem of the
Declaration of competing interest
ultra-large-scale antenna array is more complicated owing to
several antennas. The proposed channel estimation algorithms The authors declare that they have no known competing
mainly include support detection-based, approximate message financial interests or personal relationships that could have
passing based on sparse analysis, and distributed compressed appeared to influence the work reported in this paper.
sensing-based [134]. The performance of channel estimation
can be improved using AI technology, and AI-based channel
estimation technology can be implemented using a neural net- Acknowledgments
work. Specifically, the input signal is divided into the pilot and
data parts, and the data bits are equally divided and inputted This work was supported by Institute of Information &
into the neural network. Upon outputting the corresponding communications Technology Planning & Evaluation (IITP)
data sequence, the channel characteristics can be estimated grant funded by the Korea government (MSIT) (No. 2022-0-
using the input–output relationship of the signal. 00590, Industrial small cell system supporting 5G multi-band).
351
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

References [23] N.N. Dao, et al., Survey on aerial radio access networks: Toward a
comprehensive 6G access infrastructure, IEEE Commun. Surv. Tutor.
[1] M. Kaneko, I. Randrianantenaina, H. Dahrouj, H. Elsawy, M.S. 23 (2) (2021) 1193–1225.
Alouini, On the opportunities and challenges of NOMA-based [24] Z. Jia, M. Sheng, J. Li, D. Niyato, Z. Han, LEO-satellite-assisted
fog radio access networks: An overview, IEEE Access 8 (2020) UAV: Joint trajectory and data collection for internet of remote things
205467-205476. in 6G aerial access networks, IEEE Internet Things J. 8 (12) (2021)
[2] Y.L. Lee, D. Qin, L.C. Wang, G.H. Sim, 6G massive radio access 9814–9826.
networks: Key applications, requirements and challenges, IEEE Open [25] W. Khalid, H. Yu, S. Noh, Residual energy analysis in cognitive
J. Veh. Technol. 2 (2021) 54–66. radios with energy harvesting UAV under reliability and secrecy
[3] S. Zhang, D. Zhu, Towards artificial intelligence enabled 6G: State constraints, Sensors 20 (10) (2020) 2998.
of the art, challenges, and opportunities, Comput. Netw. 183 (2020) [26] C. Chaccour, M.N. Soorki, W. Saad, M. Bennis, P. Popovski, Can
107556. terahertz provide high-rate reliable low latency communications for
[4] I.F. Akyildiz, J.M. Jornet, Realizing ultra-massive MIMO (1024 wireless VR? IEEE Internet Things J. http://dx.doi.org/10.1109/JIOT.
×1024) communication in the (0.06-10) terahertz band, Nano 2022.3142674.
Commun. Netw. 8 (2016) 46–54. [27] M.Z. Chowdhury, M.K. Hasan, M. Shahjalal, M.T. Hossan, Y.M.
[5] M. Fu, Y. Zhou, Y. Shi, Reconfigurable intelligent surface empowered Jang, Optical wireless hybrid networks: Trends, opportunities, chal-
downlink non-orthogonal multiple access, IEEE Trans. Commun. 69 lenges, and research directions, IEEE Commun. Surv. Tutor. 22 (2)
(6) (2021) 3802–3817. (2020) 930–966.
[6] R. Deng, B. Di, H. Zhang, Y. Tan, L. Song, Reconfigurable [28] F.V. Corral, C. Cuenca, I. Soto, Design of an optical wireless
holographic surface: Holographic beamforming for metasurface-aided network using free space optics technology (FSO), in 5G/6G net-
wireless communications, IEEE Trans. Veh. Technol. 70 (6) (2021) works environment, in: 2021 IEEE International Conference on
6255–6259. Automation/XXIV Congress of the Chilean Association of Automatic
[7] K.M.S. Huq, J. Rodriguez, I.E. Otung, 3D network modeling for Control, ICA-ACCA, 2021, pp. 1–5.
THz-enabled ultra-fast dense networks: A 6G perspective, IEEE [29] N. Chi, Y. Zhou, Y. Wei, F. Hu, Visible light communication in 6G:
Commun. Stand. Mag. 5 (2) (2021) 84–90. Advances, challenges, and prospects, IEEE Veh. Technol. Mag. 15
[8] M. Shahjalal, M.K. Hasan, M.M. Islam, M.Z. Chowdhury, Y.M. Jang, (4) (2020) 93–102.
A two-stage power allocation-based NOMA architecture for optical [30] https://www.techuk.org/asset/3F1DC099-6D50-4116-
camera communication, IEEE Syst. J. 15 (3) (2021) 4421–4430. AFEE5DDB3B86E435/.
[9] P. Rost, et al., Benefits and challenges of virtualization in 5G radio [31] E.C. Strinati, et al., 6G: The next frontier: From holographic
access networks, IEEE Commun. Mag. 53 (12) (2015) 75–82. messaging to artificial intelligence using subterahertz and visible light
[10] G. Liu, et al., Vision, requirements and network architecture of 6G communication, IEEE Veh. Technol. Mag. 14 (3) (2019) 42–50.
mobile network beyond 2030, China Commun. 17 (9) (2020) 92–104. [32] S. Payami, K. Nikitopoulos, M. Khalily, R. Tafazolli, A signal
[11] M.Z. Chowdhury, M. Shahjalal, S. Ahmed, Y.M. Jang, 6G wireless processing framework for agile RF beamforming: From RF-chain-
communication systems: Applications, requirements, technologies, free to hybrid beamformers, IEEE Trans. Commun. 69 (6) (2021)
challenges, and research directions, IEEE Open J. Commun. Soc. 4038–4053.
1 (2020) 957–975. [33] S. He, J. Du, Y. Liao, Multi-user scheduling for 6G V2X
[12] S.J. Nawaz, S.K. Sharma, S. Wyne, M.N. Patwary, M. Asaduzzaman, ultra-massive MIMO system, Sensors 21 (20) (2021) 6742.
Quantum machine learning for 6G comm. netw.: State-of-the-art and [34] C.D. Lima, et al., Convergent communication, sensing and localiza-
vision for the future, IEEE Access 7 (2019) 46317–46350. tion in 6G systems: An overview of technologies, opportunities and
[13] H.T.H. Giang, P.D. Thanh, I. Koo, Deep Q-learning-based resource challenges, IEEE Access 9 (2021) 26902–26925.
[35] A. Faisal, H. Sarieddeen, H. Dahrouj, T.Y. Al-Naffouri, M.-S.
allocation for solar-powered users in cognitive radio networks, ICT
Alouini, Ultra-massive MIMO systems at terahertz bands: Prospects
Express 7 (2021) 49–59.
and challenges, IEEE Veh. Techn. Mag. 15 (4) (2020) 33–42.
[14] B. Narottama, S.Y. Shin, Quantum neural networks for resource
[36] W. Khalid, H. Yu, D.T. Do, Z. Kaleem, S. Noh, RIS-aided physical
allocation in wireless communications, IEEE Trans. Wirel. Commun.
layer security with full-duplex jamming in underlay D2D networks,
(2021) 1, (Early Access).
IEEE Access 9 (2021) 99667–99679.
[15] J. Mei, et al., Intelligent radio access network slicing for service
[37] W. Khalid, H. Yu, On the impact of transceiver impairments and
provisioning in 6G: A hierarchical deep reinforcement learning
reflecting elements for RIS-aided communications, in: 2021 Int. Conf.
approach, IEEE Trans. Commun. 69 (9) (2021) 6063–6078.
ICT Converg., Jeju, 2021, pp. 1–3.
[16] A. Yazar, S.D. Tusha, H. Arslan, 6G vision: An ultra-flexible [38] I. Yildirim, A. Uyrus, E. Basar, Modeling and analysis of recon-
perspective, ITU J. Future Evol. Technol. 1 (1) (2020). figurable intelligent surfaces for indoor and outdoor applications in
[17] R.J. Reifert, A.A. Ahmad, Y. Mao, A. Sezgin, B. Clerckx, Rate- future wireless networks, IEEE Trans. Commun. 69 (2) (2021)
splitting multiple access in cache-aided cloud-radio access networks, 1290–1301.
Front. Comms. Net. 2 (2021) 40. [39] Y.J. Guo, M. Ansari, N.J.G. Fonseca, Circuit type multiple beam-
[18] R.N. Mitra, D.P. Agrawal, 5G mobile technology: A survey, ICT forming networks for antenna arrays in 5G and 6G terrestrial and
Express 1 (2015) 132–137. non-terrestrial networks, IEEE J. Microw. 1 (3) (2021) 704–722.
[19] Y. Shih, W. Chung, A. Pang, T. Chiu, H. Wei, Enabling low-latency [40] M. Giordani, M. Zorzi, Non-terrestrial networks in the 6G era:
applications in fog-radio access networks, IEEE Netw. 31 (1) (2017) Challenges and opportunities, IEEE Network 35 (2) (2021) 244–251.
52–58. [41] D. Wang, M. Giordani, M.-S. Alouini, M. Zorzi, The potential of
[20] Z. Zhao, C. Feng, H.H. Yang, X. Luo, Federated-learning-enabled multilayered hierarchical nonterrestrial networks for 6G: A compar-
intelligent fog radio access networks: Fundamental theory, key ative analysis among networking architectures, IEEE Veh. Technol.
techniques, and future trends, IEEE Wirel. Commun. 27 (2) (2020) Mag. 16 (3) (2021) 99–107.
22–28. [42] M. Giordani, M. Zorzi, Non-terrestrial networks in the 6G era:
[21] M.M. Zhao, Q. Wu, M.J. Zhao, R. Zhang, Intelligent reflecting Challenges and opportunities, IEEE Netw. 35 (2) (2021) 244–251.
surface enhanced wireless networks: Two-timescale beamforming [43] H.B. Chikha, A. Almadhor, W. Khalid, Machine learning for 5G
optimization, IEEE Trans. Wirel. Commun. 20 (1) (2021) 2–17. MIMO modulation detection, Sensors 21 (5) (2021) 1556.
[22] Q. Wu, R. Zhang, Beamforming optimization for wireless network [44] Y. Xu, W. Xu, Z. Wang, J. Lin, S. Cui, Load balancing for ultra-
aided by intelligent reflecting surface with discrete phase shifts, IEEE dense networks: A deep reinforcement learning-based approach, IEEE
Trans. Commun. 68 (3) (2020) 1838–1851. Internet Things J. 6 (6) (2019) 9399–9412.
352
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

[45] B.T. Tinh, L.D. Nguyen, H.H. Kha, T.Q. Duong, Practical optimiza- [67] C. Man, et al., Reinforcement learning-based multiaccess control
tion and game theory for 6G ultra-dense networks: Overview and and battery prediction with energy harvesting in IoT systems, IEEE
research challenges, IEEE Access 10 (2022) 13311–13328. Internet Things J. 6 (2) (2018) 2009–2020.
[46] P. Frenger, J. Hederen, M. Hessler, G. Interdonato, Improved antenna [68] Y. Hao, et al., Deep reinforcement learning for resource allocation
arrangement for distributed massive MIMO, 2018, Sweden Patent in V2V communications, in: 2018 IEEE International Conference on
WO/2018/103897. Communications, ICC, 2018.
[47] G. Interdonato, E. Bjornson, H.Q. Ngo, P. Frenger, E.G. Larsson, [69] M. Chen, W. Saad, C. Yin, Liquid state machine learning for resource
Ubiquitous cell-free massive MIMO communications, EURASIP J. allocation in a network of cache-enabled LTE-U UAVs, in: 2017 IEEE
Wireless Commun. Networking 2019 (1) (2019) 1–13. Global Communications Conference, 2017, pp. 1–6.
[48] S.-E. Hong, On the effect of shadowing correlation and pilot [70] U. Challita, et al., Proactive resource management for LTE in unli-
assignment on hybrid precoding performance for cell-free mmwave censed spectrum: A deep learning perspective, IEEE Trans. Wireless
massive MIMO UDN system, ICT Express 7 (1) (2021) 60–70. Commun. 17 (7) (2018) 4674–4689.
[49] S. Fukue, H. Iimori, G.T.F. De Abreu, K. Ishibashi, Joint access [71] S. Liu, X. Hu, W. Wang, Deep reinforcement learning based dynamic
configuration and beamforming for cell-free massive MIMO systems channel allocation algorithm in multibeam satellite systems, IEEE
with dynamic TDD, IEEE Access 10 (2022) 40130–40149. Access 6 (2018) 15733–15742.
[50] P. Botsinis, D. Alanis, Z. Babar, S.X. Ng, L. Hanzo, Joint quantum- [72] Z Nan, et al., Deep reinforcement learning for user association and
assisted channel estimation and data detection, IEEE Access 4 (2016) resource allocation in heterogeneous networks, in: 2018 IEEE Global
7658–7681. Communications Conference, GLOBECOM, 2018.
[51] A. Malossini, E. Blanzieri, T. Calarco, Quantum genetic optimization, [73] H. Mao, R. Netravali, M. Alizadeh, Neural adaptive video streaming
IEEE Trans. Evol. Comput. 12 (2) (2008) 231–241. with pensieve, in: Proc. of the Conference of the ACM Special
[52] T.D.P. Perera, D.N.K. Jayakody, S.K. Sharma, S. Chatzinotas, J. Interest Group on Data Communication, 2017, pp. 197–210.
Li, Simultaneous wireless information and power transfer (SWIPT): [74] P.V.R. Ferreira, et al., Multiobjective reinforcement learning for cog-
Recent advances and future challenges, IEEE Commun. Surv. Tutor. nitive satellite communications using deep neural network ensembles,
20 (1) (2018) 264–302. IEEE J. Sel. Areas Commun. 36 (5) (2018) 1030–1041.
[53] M. Kamel, W. Hamouda, A. Youssef, Ultra-dense networks: A survey,
[75] H. Mao, M. Alizadeh, I. Menache, S. Kandula, Resource management
IEEE Commun. Surv. Tutor. 18 (4) (2016) 2522–2545.
with deep reinforcement learning, in: 15th ACM Workshop on Hot
[54] W. Khalid, H. Yu, Security improvement with QoS provisioning using Topics in Networks, Atlanta, 2016, pp. 1–7.
service priority and power allocation for NOMA-IoT networks, IEEE
[76] H. Peng, X. Shen, Multi-agent reinforcement learning based resource
Access 9 (2021) 9937–9948.
management in MEC- and UAV-assisted vehicular networks, IEEE J.
[55] W. Shin, et al., Coordinated beamforming for multi-cell MIMO-
Sel. Areas Commun. 39 (1) (2021) 131–141.
NOMA, IEEE Commun. Lett. 21 (1) (2017) 84–87.
[77] Z. Xu, Y. Wang, J. Tang, J. Wang, M.C. Gursoy, A deep reinforce-
[56] S.H. Amin, A.H. Mehana, Y.A. Fahmy, S, S. Soliman, Maximizing
ment learning based framework for power-efficient resource allocation
the number of users in clustered MIMO-NOMA systems under rate
in cloud RANs, in: 2017 IEEE Int. Conf. on Commun., ICC, Paris,
constraints, Mobile Netw. Appl. 24 (2019) 618–629.
2017, pp. 1–6.
[57] L. Zhang, et al., Layered-division multiplexing: An enabling technol-
[78] T. Li, et al., Model-free control for distributed stream data processing
ogy for multicast/broadcast service delivery in 5G, IEEE Commun.
using deep reinforcement learning, Proc. VLDB Endow. 11 (6) (2018)
Mag. 56 (3) (2018) 82–90.
705–718.
[58] B. Clerckx, Y. Mao, R. Schober, H.V. Poor, Rate-splitting unifying
[79] Y. He, F.R. Yu, N. Zhao, V.C.M. Leung, H. Yin, Software-defined
SDMA, OMA, NOMA, and multicasting in MISO broadcast channel:
networks with mobile edge computing and caching for smart cities:
A simple two-user rate analysis, IEEE Wirel. Commun. Lett. 9 (3)
A big data deep reinforcement learning approach, IEEE Commun.
(2020) 349–353.
Mag. 55 (12) (2017) 31–37.
[59] M. Dai, B. Clerckx, D. Gesbert, G. Caire, A rate splitting strategy for
massive MIMO with imperfect CSIT, IEEE Trans. Wirel. Commun. [80] Y. He, et al., Software-defined networks with mobile edge computing
15 (7) (2016) 4611–4624. and caching for smart cities: A big data deep reinforcement learning
[60] A. Papazafeiropoulos, B. Clerckx, T. Ratnarajah, Rate-splitting to approach, IEEE Commun. Mag. 55 (12) (2017) 31–37.
mitigate residual transceiver hardware impairments in massive MIMO [81] C. Zhong, et al., A deep reinforcement learning-based framework for
systems, IEEE Trans. Veh. Technol. 66 (9) (2017) 8196–8211. content caching, in: 52nd Annual Conference on Information Sciences
[61] T. Haarnoja, A. Zhou, P. Abbeel, S. Levine, Soft actor-critic: and Systems, CISS, 2018, pp. 1–6.
Off-policy maximum entropy deep reinforcement learning with a [82] Y. He, et al., A big data deep reinforcement learning approach
stochastic actor, in: 2018 Int. Conf. Mach. Learn., ICML, Stockholm, to next generation green wireless networks, in: 2017 IEEE Global
pp. 1–10. Communications Conference, 2017, pp. 1–6.
[62] S. Han, Y. Sung, Diversity actor-critic: Sample-aware entropy regu- [83] Y. He, et al., Optimization of cache-enabled opportunistic interference
larization for sample-efficient exploration, in: 2021 Int. Conf. Mach. alignment wireless networks: A big data deep reinforcement learning
Learn., ICML, Stockholm, pp. 1–12. approach, in: IEEE International Conference on Communications,
[63] J.N. Foerster, Y.M. Assael, N.D. Freitas, S. Whiteson, Learning to ICC, 2017, pp. 1–6.
communicate with deep multi-agent reinforcement learning, in: 2016 [84] X. He, et al., Green resource allocation based on deep reinforcement
30th Conf. Neural Inf. Process. Syst., NIPS, Barcelona, 2016, pp. learning in content-centric IoT, IEEE Trans. Emerg. Top. Comput. 8
1–9. (3) (2018) 781–796.
[64] H. Lee, E. Kim, H. Kim, J.H. Na, H.-H. Choi, Multi-agent Q-learning [85] Y. He, N. Zhao, H. Yin, Integrated networking, caching, and comput-
based cell breathing considering SBS collaboration for maximizing ing for connected vehicles: A deep reinforcement learning approach,
energy efficiency in B5G heterogeneous networks, ICT Express IEEE Trans. Veh. Technol. 67 (1) (2017) 44–55.
(2021) http://dx.doi.org/10.1016/j.icte.2021.09.006, in press. [86] C. Zhang, et al., A deep reinforcement learning based approach
[65] S. Wang, H. Liu, P.H. Gomes, B. Krishnamachari, Deep reinforce- for cost-and energy-aware multi-flow mobile data offloading, IEICE
ment learning for dynamic multichannel access in wireless networks, Trans. Commun. E101.B (7) (2018) 1625–1634.
IEEE Trans. Cogn. Commun. Netw. 4 (2) (2018) 257–265. [87] J. Li, H. Gao, T. Lv, Y. Lu, Deep reinforcement learning based
[66] Y.S. Nasir, D. Guo, Multi-agent deep reinforcement learning for computation offloading and resource allocation for MEC, in: IEEE
dynamic power allocation in wireless networks, IEEE J. Sel. Areas Wireless Communications and Networking Conference, WCNC,
Commun. 37 (10) (2019) 2239–2250. 2018, pp. 1–6.
353
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

[88] X. Chen, H. Zhang, C. Wu, S. Mao, Y. Ji, M. Bennis, Perfor- [108] H.B. McMahan, E. Moore, D. Ramage, S. Hampson, B.A. Arcas,
mance optimization in mobile-edge computing via deep reinforcement Communication-efficient learning of deep networks from decentral-
learning, in: IEEE 88th Vehicular Technology Conference, VTC-Fall, ized data, in: 2017 Int. Conf. Artif. Intell. Stat., AISTATS, Florida,
2018, pp. 1–6. 2017, pp. 1–10.
[89] X. Chen, et al., Optimized computation offloading performance in [109] J. Kang, et al., Reliable federated learning for mobile networks, IEEE
virtual edge computing systems via deep reinforcement learning, Wirel. Commun. 27 (2) (2020) 72–80.
IEEE Internet Things J. 6 (3) (2018) 4005–4018. [110] S. Savazzi, M. Nicoli, V. Rampa, Federated learning with cooperating
[90] J. Zhu, Y. Song, D. Jiang, H. Song, A new deep-Q-learning-based devices: A consensus approach for massive IoT networks, IEEE
transmission scheduling mechanism for the cognitive internet of Internet Things J. 7 (5) (2020) 4641–4654.
things, IEEE Internet Things J. 5 (4) (2017) 2375–2385. [111] L. Liu, et al., Client-edge-cloud hierarchical federated learning, in:
[91] M. Gadaleta, F. Chiariotti, M. Rossi, A. Zanella, D-DASH: A deep ICC 2020-2020 IEEE International Conference on Communications,
Q-learning framework for DASH video streaming, IEEE Trans. Cogn. ICC, 2020, pp. 1–6.
Commun. Networking 3 (4) (2017) 703–718. [112] X. Wang, et al., In-edge ai: Intelligentizing mobile edge computing,
[92] O. Naparstek, K. Cohen, Deep multi-user reinforcement learning caching and communication by federated learning, IEEE Network 33
for dynamic spectrum access in multichannel wireless networks, in: (5) (2019) 156–165.
GLOBECOM 2017-2017 IEEE Global Communications Conference, [113] J. Ren, et al., Federated learning-based computation offloading
2017, pp. 1–7. optimization in edge computing-supported internet of things, IEEE
[93] R. Li, Z. Zhao, X. Chen, J. Palicot, H. Zhang, TACT: A transfer Access 7 (2019) 69194–69201.
actor-critic learning framework for energy saving in cellular ra- [114] Z. Yu, et al., Federated learning based proactive content caching in
dio access networks, IEEE Trans. Wirel. Commun. 13 (4) (2014) edge computing, in: 2018 IEEE Glob. Commun. Conf., GLOBECOM,
2000–2011. Abu Dhabi, 2018, pp. 1–6.
[94] Y. Sun, M. Peng, S. Mao, Deep reinforcement learning-based mode [115] Y. Qian, et al., Privacy-aware service placement for mobile edge
selection and resource management for green fog radio access computing via federated learning, Inform. Sci. 505 (2019) 562–570.
networks, IEEE Internet Things J. 6 (2) (2019) 1960–1971. [116] Y. Qu, et al., Decentralized privacy using blockchain-enabled feder-
[95] C. Parera, A.E.C. Redondi, M. Cesana, Q. Liao, I. Malanchini, ated learning in fog computing, IEEE Internet Things J. 7 (6) (2020)
Transfer learning for channel quality prediction, in: 2019 IEEE 5171–5183.
International Symposium on Measurements & Networking, M & N, [117] S. Samarakoon, et al., Federated learning for ultra-reliable low-
2019, pp. 1–6. latency V2V communications, in: 2018 IEEE Global Communications
[96] A. Zappone, M. Di Renzo, M. Debbah, T.T. Lam X. Qian, Model- Conference, GLOBECOM, 2018, pp. 1–7.
aided wireless artificial intelligence: Embedding expert knowledge in
[118] J. Posner, et al., Federated learning in vehicular networks:
deep neural networks for wireless system optimization, IEEE Veh.
Opportunities and solutions, IEEE Network 35 (2) (2021) 152–159.
Technol. Mag. 14 (3) (2019) 60–69.
[119] D. Ye, et al., Federated learning in vehicular edge computing:
[97] R. Dong, C. She, W. Hardjawana, Y. Li, B. Vucetic, Deep learning for
A selective model aggregation approach, IEEE Access 8 (2020)
radio resource allocation with diverse quality-of-service requirements
23920–23935.
in 5G, IEEE Trans. Wireless Commun. 20 (4) (2020) 2309–2324.
[120] V. Dunjko, J.M. Taylor, H.J. Briegel, Advances in quantum reinforce-
[98] Q. Zhao, D. Grace, A. Vilhar, T. Javornik, Using k-means clus-
ment learning, in: 2017 IEEE Int. Conf. Syst. Man, Cybern., SMC,
tering with transfer and Q learning for spectrum, load and energy
Banff, 2017, pp. 282–287.
optimization in opportunistic mobile broadband networks, in: 2015
[121] G. Lu, W.H. Zeng, Cloud computing survey, Appl. Mech. Mater. 530
International Symposium on Wireless Communication Systems,
(2014) 650–661.
ISWCS, 2015, pp. 116–120.
[99] C. Parera, et al., Transfer learning for multi-step resource utiliza- [122] J. Chen, X. Ran, Deep learning with edge computing: A review, Proc.
tion prediction, in: IEEE 31st Annual International Symposium on IEEE 107 (8) (2019) 1655–1674.
Personal, Indoor and Mobile Radio Communications, 2020, pp. 1–6. [123] Y. Kang, et al., Neurosurgeon: Collaborative intelligence between the
[100] Q. Zeng, et al., Traffic prediction of wireless cellular networks based cloud and mobile edge, in: 2017 Int. Conf. Archit. Support Program.
on deep transfer learning and cross-domain data, IEEE Access 8 Lang. Oper. Syst., ASPLOS, Xian, 2017, pp. 615–629.
(2020) 172387-172397. [124] X. Lyu, et al., Distributed online learning of cooperative caching in
[101] C. Zhang, et al., Deep transfer learning for intelligent cellular traffic edge cloud, IEEE. Trans. Mob. Comput. 20 (8) (2021) 2550–2562.
prediction based on cross-domain big data, IEEE J. Sel. Areas [125] A. Sadeghi, F. Sheikholeslami, G.B. Giannakis, Optimal and scal-
Commun. 37 (6) (2019) 1389–1401. able caching for 5G using reinforcement learning of space–time
[102] B.B. Nagaraja, K.G. Nagananda, Caching with unknown popularity popularities, IEEE J. Sel. Top. Signal Process. 12 (1) (2018) 180–190.
profiles in small cell networks, in: IEEE Global Communications [126] J. Kwak, Y. Kim, J. Lee, S. Chong, DREAM: Dynamic resource
Conference, GLOBECOM, 2015, pp. 1–6. and task allocation for energy minimization in mobile cloud systems,
[103] T. Hou, et al., Proactive content caching by exploiting transfer IEEE J. Sel. Areas Commun. 33 (12) (2015) 2510–2523.
learning for mobile edge computing, in: GLOBECOM 2017-2017 [127] X. Chen, Decentralized computation offloading game for mobile
IEEE Global Communications Conference, 2017, pp. 1–6. cloud computing, IEEE Trans. Parallel Distrib. Syst. 26 (4) (2015)
[104] B.N. Bharath, K.G. Nagananda, H.V. Poor, A learning-based ap- 974–983.
proach to caching in heterogenous small cell networks, IEEE Trans. [128] Y. Kim, J. Kwak, S. Chong, Dual-side optimization for cost-delay
Commun. 64 (4) (2016) 1674–1686. tradeoff in mobile edge computing, IEEE Trans. Veh. Technol. 67
[105] K. Liu, et al., Toward low-overhead fingerprint-based indoor local- (2) (2017) 1765–1781.
ization via transfer learning: Design, implementation, and evaluation, [129] G. Ding, et al., Spectrum inference in cognitive radio networks:
IEEE Trans. Ind. Inform. 14 (3) (2017) 898–908. Algorithms and applications, IEEE Commun. Surv. Tutor. 20 (1)
[106] H. Zou, et al., A transfer kernel learning based strategy for adaptive (2018) 150–182.
localization in dynamic indoor environments: Poster, in: Proceedings [130] D.W. Yun, W.C. Lee, Intelligent dynamic spectrum resource man-
of the 22nd Annual International Conference on Mobile Computing agement based on sensing data in space–time and frequency domain,
and Networking, 2016, pp. 462–464. Sensors 21 (16) (2021) 5261.
[107] Z. Sun, et al., Adaptive localization through transfer learning in [131] D.M. Shin, S.C. Lim, K. Yang, Mapping selection and code con-
indoor wi-fi environment, in: Seventh International Conference on struction for 2m-ary polar-coded modulation, IEEE Commun. Lett.
Machine Learning and Applications, 2008, pp. 331–336. 16 (6) (2012) 905–908.
354
M. Shahjalal, W. Kim, W. Khalid et al. ICT Express 9 (2023) 341–355

[132] M. Mondelli, S.H. Hassani, R. Urbanke, Construction of polar codes [134] A.N. Uwaechia, N.M. Mahyuddin, Spectrum-efficient distributed
with sublinear complexity, in: 2017 IEEE Int. Symp. Inf. Theory., compressed sensing based channel estimation for OFDM systems
ISIT, Aachen, 2017, pp. 1853–1857. over doubly selective channels, IEEE Access 7 (2019) 35072–35088.
[133] R. Shafin, et al., Artificial intelligence-enabled cellular networks: A [135] M.M. Blue, S. Yrjola, P. Ahokangas, Spectrum management in the
critical path to beyond-5G and 6G, IEEE Wirel. Commun. 27 (2) 6G era: The role of regulation and spectrum sharing, in: 2020 2nd
(2020) 212–217. 6G Wirel. Summit. 6G SUMMIT, Levi, 2020, pp. 1–5.

355