1

Multiple Access Techniques for Intelligent and

Multi-Functional 6G: Tutorial, Survey, and Outlook
Bruno Clerckx, Fellow, IEEE, Yijie Mao, Member, IEEE, Zhaohui Yang, Member, IEEE, Mingzhe
Chen, Member, IEEE, Ahmed Alkhateeb, Member, IEEE, Liang Liu, Senior Member, IEEE, Min
Qiu, Member, IEEE, Jinhong Yuan, Fellow, IEEE, Vincent W.S. Wong, Fellow, IEEE, and Juan Montojo, Senior
Member, IEEE

(invited paper)

Abstract—Multiple access (MA) is a crucial part of any multimodal sensing-aided communications, multimodal sensing
arXiv:2401.01433v1 [cs.IT] 2 Jan 2024

wireless system and refers to techniques that make use of the and digital twin-assisted communications, and communication-
resource dimensions (e.g., time, frequency, power, antenna, code, aided sensing/localization systems. We finish with studying MA
message, etc) to serve multiple users/devices/machines/services, for emerging intelligent applications such as semantic communi-
ideally in the most efficient way. Given the increasing needs cations, metaverse, virtual reality, smart radio and reconfigurable
of multi-functional wireless networks for integrated communica- intelligent surfaces, and massive connectivity and random access
tions, sensing, localization, computing, coupled with the surge of in Internet-of-Things, before presenting a roadmap toward 6G
machine learning / artificial intelligence (AI) in wireless networks, standardization. Throughout the text, we also point out numerous
MA techniques are expected to experience a paradigm shift in directions that are promising for future research.
6G and beyond. In this paper, we provide a tutorial, survey
and outlook of past, emerging and future MA techniques and Index Terms—Multiple Access, Orthogonal Multiple Access,
pay a particular attention to how wireless network intelligence Non-Orthogonal Multiple Access, Space Division Multiple Ac-
and multi-functionality will lead to a re-thinking of those cess, Code Domain Multiple Access, Rate-Splitting Multiple Ac-
techniques. The paper starts with an overview of orthogonal, cess, Universal Multiple Access, Artificial Intelligence, Machine
physical layer multicasting, space domain, power domain, rate- Learning, Integrated Sensing and Communications, Semantic
splitting, code domain MAs, and MAs in other domains, and Communications, Reconfigurable Intelligent Surfaces, Metaverse,
highlight the importance of researching universal multiple access Augmented Reality, Internet-of-Things, 6G.
to shrink instead of grow the knowledge tree of MA schemes
by providing a unified understanding of MA schemes across all
I. I NTRODUCTION
resource dimensions. It then jumps into rethinking MA schemes
in the era of wireless network intelligence, covering AI for A. From Communication-centric 5G to Intelligent and Multi-
MA such as AI-empowered resource allocation, optimization, functional 6G
channel estimation, receiver designs, user behavior predictions
Ext generation wireless networks, such as 6G and be-
for different MA schemes, and MA for AI such as federated
learning/edge intelligence and over the air computation. We then
discuss MA for network multi-functionality and the interplay
N yond, will face mounting challenges such as higher data
throughput and (spectral/ energy) efficiency, enhanced relia-
between MA and integrated sensing, localization, and commu- bility, massive connectivity, global coverage across terrestrial
nications, covering MA for joint sensing and communications,
and non-terrestrial networks, and a growing heterogeneity in
Bruno Clerckx is with the Department of Electrical and Electronic the quality of service (QoS). Addressing those challenges is
Engineering, Imperial College London, London SW7 2AZ, U.K (email: essential to meet the demands of further-enhanced mobile
[email protected]). broadband (FeMBB) for augmented reality (AR) / virtual
Yijie Mao is with the School of Information Science and Tech-
nology, ShanghaiTech University, Shanghai 201210, China (e-mail: reality (VR), extremely ultra reliable and low-latency commu-
[email protected]). nication (eURLLC) for full automation, control, and operation
Zhaohui Yang is with College of Information Science and Elec- in industrial environment and connected robotics, ultra massive
tronic Engineering, Zhejiang University, Hangzhou 310027, China, (e-mail:
yang [email protected]). machine type communication (umMTC) for Internet-of-Things
Mingzhe Chen is with the Department of Electrical and Computer Engi- (IoT), etc [1].
neering and Institute for Data Science and Computing, University of Miami, Importantly, future wireless networks will not only provide
Coral Gables, FL, 33146 USA (Email: [email protected]).
Ahmed Alkhateeb is with the School of Electrical, Computer and Energy communications services but also offer a wide range of
Engineering, Arizona State University, Tempe, AZ 85287 USA (e-mail: new functionalities such as sensing, intelligence, computation,
[email protected]). localization/navigation, powering. This tendency for multi-
Liang Liu is with the Department of Electronic and Information Engineer-
ing, The Hong Kong Polytechnic University, Hong Kong, SAR, China (e-mail: functional networks is well exemplified by the increasing
[email protected]). number of research areas in the past decade that have looked
Min Qiu and Jinhong Yuan are with the School of Electrical Engineering at various forms of wireless systems integration. Indeed, radio
and Telecommunications, University of New South Wales, Sydney, NSW
2052, Australia (e-mail: [email protected]; [email protected]). waves can be used for numerous applications, most commonly
Vincent W.S. Wong are with the Department of Electrical and Computer communications, but also power in the form of wireless
Engineering, The University of British Columbia, Vancouver, BC V6T 1Z4, energy harvesting and wireless power transfer, sensing in
Canada (e-mail: [email protected]).
Juan Montojo is with Corporate R&D, Qualcomm Inc., San Diego, CA the form of radar, localization, etc. All of those areas have
(e-mail: [email protected]). traditionally been studied separately and have led to different
 2

disciplines within Electrical and Electronic Engineering but [8]. Classifying MA techniques has been challenging due the
have also evolved into vastly different industry sectors [2]. In proliferation of new schemes in the past decade. Unfortunately,
the past decade, the community has progressively experienced the widely used classification into non-orthogonal MA vs
a paradigm shift in wireless network design, namely, unifying orthogonal MA, i.e., NOMA vs OMA, is over-simplistic and
transmission and processing of many quantities and functional- tends to amalgamate many different MA schemes under the
ities, such as information, power, sensing, localization etc so as non-orthogonal umbrella without contrasting them or truly
to make the best use of the radio frequency (RF) spectrum and understanding the essence of all those schemes [9]. Such
radiation as well as the network infrastructure for the multi- classification has caused unnecessary confusions and misun-
purpose of communicating, energizing, sensing, locating, com- derstandings in the past few years [10]. Instead of contrasting
puting, but also for the synergies that all those disciplines can orthogonal vs non-orthogonal, [9] suggested that a different
bring to each other once properly integrated. This quest for classification should be considered in next generation wireless
integration, convergence, and multi-functionality has led to the networks and showed that the fundamental question behind
new research areas of integrated sensing and communications MA design should instead be how to manage multi-user
(ISAC), integrated sensing, localization, and communications, interference. Answering this question shed the light on the
wireless information and power transfer (WIPT), edge com- differences between various non-orthogonal approaches to MA
puting and intelligence, and integrated artificial intelligence designs and on a new classification of MA schemes based on
(AI) and communications [1], [3]. Additionally, network in- how the interference is managed. Importantly, this exercise
telligence, using AI and machine learning (ML), will become brought to light the powerful and emerging Rate-Splitting
pervasive in the design, control and optimization of the multi- Multiple Access (RSMA) that unifies into a single MA scheme
functional networks and the network itself will become the four seemingly unrelated strategies, namely OMA, power
underpinning tool to enable AI applications [4]. domain (PD)-NOMA, SDMA, and physical-layer multicas-
ting [11]. The capability of RSMA to unify and therefore
be more universal than other MA schemes makes practical
B. The Crucial Role of Multiple Access
implementation and operation easier. Indeed, one could claim
Capturing multi-functionality and intelligence in future that a single unified and general MA scheme would be easier
wireless network design will enable using wireless to its full to implement and optimize than a combination of multiple
potential, hence enabling trillions of future intelligent users to MA schemes, each optimized for specific conditions. This can
sense, compute, connect, energize, analyze anywhere, anytime, be increasingly important in multi-functional 6G and beyond
and on the move [5], [6]. One major challenge and opportunity networks where the range and diversity of services, use cases,
that such intelligent multi-functional 6G and beyond network and deployments explode.
brings is that the notion of “wireless networks” and “users”
should be understood in a much more broader context com-
pared to 1G–5G era. With the emergence of integrated wireless C. Objectives and Organization
communications/power/sensing/computing/AI networks, users This paper has three objectives: 1) a tutorial paper to educate
refer to communication devices, sensing targets, devices to the readers about the fundamentals of a wide range MA
be charged, AI nodes, training devices, or any other form of schemes, 2) a survey paper to give the readers access to the
services/functionalities that the network could provide. state-of-the-art MA schemes and literature, and 3) an outlook
At the core of wireless network design lies the multiple paper to guide the readers with new research directions.
access (MA) technique whose pivotal role is to serve and Specifically the paper contributes in the following ways.
process all these “users” and decide how to allocate them First, we provide a tutorial and survey (in Section II)
resources, including time, frequency, power, space (e.g., an- of a wide range of MA techniques: OMA in the form
tennas, beams), signal (e.g., messages, codes, etc), in the of TDMA/FDMA/OFDMA, SDMA [12], and PD-NOMA
most efficient way. The design of future intelligent and multi- [13]; RSMA [9], [14]; code-domain (non-orthogonal) mul-
functional networks brings new challenges and opportunities tiple access (CD-NOMA) departing from traditional 3G
for wireless network designers and in particular when it comes CDMA [15] including low-density spreading multiple access
to MA. It is crucial to comprehend how MA techniques (LDS-CDMA), sparse code multiple access (SCMA), multi-
can address these future demands and how they need to be user shared multiple access (MUSA), successive interference
re-thought in light of the network multi-functionality and cancellation amenable multiple access (SAMA); and other
intelligence paradigms. multiple access schemes exploiting other domains such as
Time/frequency domain multiple access (TDMA/FDMA) interleave-division multiple access (IDMA), pattern division
were popular in 2G, code-division multiple access (CDMA) in multiple access (PDMA), compute-forward multiple access
3G, orthogonal frequency division multiple access (OFDMA) (CFMA), lattice-partition-based multiple access (LPBMA),
coupled with space-division multiple access (SDMA) in 4G spatially coupled multiple access, layered-division multiplex-
and 5G. Though SDMA-OFDMA has remained the dominant ing (LDM), index modulation multiple access (IMMA), delay-
MA in the past 20 years, the past decade has also seen a Doppler domain multiple access (DDMA), etc. The advantages
wide interest in other forms of MA schemes, often classified and disadvantages of those MA schemes and the interplay
into non-orthogonal multiple access (NOMA), making use of between them are discussed, before drawing observations and
the power domain, the code domain, or other domains [7], conclusions on how to rethink the role and design of MA
 3

empowered receiver designs for advanced MA schemes, and

AI for user behavior predictions in MA. We then investigate
FDMA the converse, namely how to design MA schemes for AI
TDMA applications (in Section IV). We here touch upon MA for fed-
…… erated learning/edge intelligence and over the air computation.
OFDMA We conclude the discussion by identifying further research
…… avenues on the interplay between MA and AI. Such treatment
DDMA significantly departs from other tutorials [9], [10], [14], [16]–
[19]. This is the first tutorial paper providing a comprehensive
review of the interplay between AI and MA, addressing both
SDMA Multicasting AI for MA and MA for AI.
CDMA SCMA Fourth, we discuss MA in the era of network multi-
functionality and the interplay between MA and integrated
PD-NOMA ……
sensing, localization, and communications (in Section V).
We elaborate on how MA designs should be tailored for
MUSA joint sensing and communications, multimodal sensing-aided
communications, for multimodal sensing and digital twin-
Unify as you move assisted communications, and for communication-aided sens-
from leaves, to ing/localization systems. We identify the shortcoming of exist-
UMA
branches, to trunk ing MA schemes in multi-functional networks and interesting
research topics for future works. This treatment differs from
other tutorials such as [18] that focuses on demonstrating
Fig. 1. Shrinking the knowledge tree of MA by unifying MA schemes as we RSMA superiority over SDMA and NOMA in ISAC or [21]
move from leaves, to the branches, to the trunk. UMA, not yet found, would
be the holy grail of MA scheme unification.
that does not focus on MA designs for ISAC.
Fifth, we discuss MA designs for emerging intelligence
applications (in Section VI) such as semantic communications,
metaverse, virtual reality, smart radio and reconfigurable intel-
schemes. This overview departs from the conventional discus- ligent surfaces, and massive connectivity and random access
sion on orthogonal versus non-orthogonal approaches found in Internet-of-Things, before presenting a roadmap toward
in 5G [9] and recent tutorials [9], [10], [13], [14], [16]–[20] 6G standardization (in Section VII). This differs from other
whose focus was on RSMA and its sub-MA schemes OMA, tutorials such as [9], [14], [16] that focus on RSMA for some
SDMA, PD-NOMA, but not on code domain approaches nor of those applications.
on other domains MA. Table I details the main abbreviations used throughout this
Second, we discuss and motivate (in Section II-H) the work.
importance of shrinking the knowledge tree of the MA lit-
erature in order to identify MA schemes that can exploit II. A N OVERVIEW OF M ULTIPLE ACCESS T ECHNIQUES
multiple dimensions and unify them. Inspired by RSMA to
that shrinks the knowledge tree by providing a unified and In this section, we provide an overview and discuss pros
conceptually simple understanding of a morass of results and cons of OMA, Physical Layer Multicasting, SDMA,
on OMA, SDMA, PD-NOMA, physical-layer multicasting, PD-NOMA, RSMA, CD-NOMA, and other MAs exploiting
we motivate future research toward universal multiple access other dimensions. We draw some general observations before
(UMA). UMA should further shrink the knowledge tree of MA identifying possible paths toward UMA.
schemes by unifying RSMA with all other dimensions, such
as code domain MAs, and ultimately provide a unified and A. Orthogonal Multiple Access
conceptually simple understanding of the current and future Orthogonal multiple access (OMA) is a fundamental MA
morass of MA schemes as illustrated in Fig. 1. Clearly, there technique that has been widely used in mobile communication
is no UMA yet that optimally suits all applications and use systems. In OMA, the radio resources at hand are strategically
cases, but such scheme and research directions are expected divided into distinct, non-overlapping frequency bands, time
to become increasingly important in multi-functional 6G and slots, or codes, each meticulously allocated to an individual
beyond networks where the range and diversity of services, user. OMA adheres to a typical principle of interference
use cases, and deployments explode. This research avenue and management, which is to prevent multi-user interference.
vision has not been discussed in prior works. There are four well-established OMA strategies specifically
Third, we discuss MA in the era of network intelligence tailored for 1G to 4G communication systems, namely, fre-
and the interplay between AI and MA (in Sections III and quency division multiple access (FDMA), time division mul-
IV). We first identify how AI can be leveraged to enhance MA tiple access (TDMA), code division multiple access (CDMA),
designs (in Section III). This includes AI-empowered resource and orthogonal frequency division multiple access (OFDMA)
allocation and optimization for different MA schemes, AI- [22]. Specifically, FDMA divides the available spectrum into
empowered channel estimation for different MA schemes, AI- non-overlapping frequency bands, each accommodating one
 4

TABLE I
L IST OF ABBREVIATIONS .

BC Broadcast Channel (u)mMTC (ultra) massive Machine-Type Communication

CDMA Code Division Multiple Access MU–LP Multi-User Linear Precoding
CFMA Compute-Forward Multiple Access MU-MIMO Multi-User Multiple-Input Multiple-Output
CoMP Coordinated Multi-Point MUSA Multi-User Shared Multiple Access
CSI Channel State Information NOMA Non-Orthogonal Multiple Access
CSIT/R Channel State Information at the Transmitter/Receiver NOUM Non-Orthogonal Unicast and Multicast
C-RAN Cloud-Radio Access Networks OFDMA Orthogonal Frequency Division Multiple Access
DDMA Delay-Doppler Domain Multiple Access OMA Orthogonal Multiple Access
DoF Degree-of-Freedom PDMA Pattern Division Multiple Access
DPC Dirty Paper Coding QoS Quality of Service
DPCRS Dirty Paper Coded Rate-Splitting RF Radio Frequency
(F)eMBB (further-)enhanced Mobile Broadband Service RIS Reconfigurable Intelligent Surfaces
FDD Frequency Division Duplex RS Rate-Splitting
FDMA Frequency Division Multiple Access RSMA Rate-Splitting Multiple Access
F-RAN Fog-Radio Access Networks SAMA Successive Interference Cancellation Amenable Multiple Access
GRS Generalized Rate-Splitting SC Superposition Coding
HRS Hierarchical Rate-Splitting SCMA Sparse Code Multiple Access
IDMA Interleave-Division Multiple Access SDMA Space Division Multiple Access
IMMA Index Modulation Multiple Access SIC Successive Interference Cancellation
ISAC Integrated Sensing and Communications SISO Single-Input Single-Output
IRS Intelligent Reconfigurable Surface SNR Signal-to-Noise Ratio
LDM Layered-Division Multiplexing SWIPT Simultaneous Wireless Information and Power Transfer
LDS-CDMA Low-Density Spreading Multiple Access SIMO Single-Input Multiple-Output
LPBMA Lattice-Partition-Based Multiple Access TDD Time Division Duplex
MA Multiple Access TDMA Time-Division Multiple Access
MAC Multiple Access Channel UMA Universal Multiple Access
MIMO Multiple-Input Multiple-Output UAV Unmanned Aerial Vehicles
MISO Multiple-Input Single-Output (e)URLLC (extremely) Ultra-Reliable Low-Latency Communicati

user. TDMA partitions time into time slots allocated to dif- 2) Low capacity: OMA allocates each orthogonal
ferent users. CDMA utilizes orthogonal, user-specific codes radio resource to an individual user. The system ca-
to spread the modulated user symbols, serving multiple users pacity is restricted by the total number of available
simultaneously in the same time-frequency resources without radio resources. This limitation hinders its ability to
causing multi-user interference (under ideal propagation con- accommodate the surging user demand experienced
ditions). OFDMA divides the frequency and time resources in modern communication systems.
into narrow subcarriers and time slots, which are grouped into 3) High signaling overhead: To enhance system per-
resource units and allocated to the users. formance of OMA, it is essential to implement well-
While OMA has garnered widespread acceptance in past designed user scheduling, which typically results in
communication systems, its efficacy becomes restricted in a significant increase in signaling overhead.
light of the explosive expansion of wireless communication
worldwide. Here we outline the advantages and disadvantages B. Physical Layer Multicasting
of OMA:
Multicasting usually refers to the transmission of a message
• Advantages: intended to multiple users, i.e., one-to-all. Popular example are
1) Widespread adoption: OMA is well-established radio and television where the message of interest to multiple
and extensively utilized in current communication users is decoded by those users. However, multicasting can
systems, making it easier for facilitating smooth and also be understood in a wider context where the transmitter
effortless network expansion. transmits multiple unicast (i.e., one-to-one) messages, each
2) Simplicity: OMA simplifies transceiver design, im- intended to one user, by encoding them jointly into one stream
plementation, and management. to be decoded by all users. Users would then retrieve from
3) Interference free: OMA prevents interference be- the decoded stream the part intended to them. The encoding
tween users, enabling interference-free transmis- and transmission over the air is effectively a physical layer
sions and thereby enhancing the overall quality and multicasting since all messages are encoded into one multicast
reliability of communications. It excels in managing stream to be decoded by all users. The difference with conven-
low to moderate user loads effectively. It is effective tional multicasting is that only part of the multicast stream is
for handling low to moderate user loads. intended to a given user, instead of the entire stream. Physical-
• Disadvantages: layer multicasting contrasts with OMA where messages are
1) Inefficient spectrum utilization: In OMA, there is encoded in independent streams and transmitted on orthogonal
a risk that even low-rate users, such as IoT sensors resources.
with minimal resource requirements, may occupy an • Advantages:
entire resource block, leading to inefficient spectrum 1) Coding gain: By combining messages and encoding
utilization. jointly multiple medium-size packets together into
 5

a single stream, the encoded stream is longer and application is hindered by the impracticality arising from its
the coding gain is increased, which leads to higher high computational demands. In contrast, multi-user linear
reliability. Such feature is particularly useful in non- precoding (MU-LP) offers a more practical alternative by
terrestrial systems, such as geostationary satellite leveraging linear precoding techniques at the transmitter while
communications based on DVB-S2X technology regarding multi-user interference as noise at the receivers [25],
[23], where each spot beam of the satellite serves [26]. Although MU-LP cannot achieve the capacity region
more than one user simultaneously by transmitting achieved by DPC, it is particularly useful when users possess
a single coded frame. Since different beams illumi- semi-orthogonal channels and comparable signal strengths
nate different group of users, such satellite system (and also perfect channel state information at the transmitter-
follows a physical layer multigroup multicast trans- CSIT). Therefore, it plays a pivotal role in many transmission
mission [24]. techniques underpinning both 4G and 5G networks, including
2) Low latency: Since all messages are encoded multi-user MIMO (MU-MIMO), massive MIMO, networked
jointly into one stream to be decoded by all, users MIMO, and other advanced techniques.
can decode their messages simultaneously and do While SDMA remains crucial in modern wireless networks,
not have to wait for their message to be transmit- the ongoing evolution of wireless technologies compels us to
ted consecutively as in TDMA, therefore reducing scrutinize SDMA critically. In the following, we delineate the
latency in the network. advantages and disadvantages of SDMA:
3) Interference free: Multicasting prevents interfer- • Advantages:
ence between users since there is only one stream
transmitted, enabling interference-free transmission 1) Enhanced spectrum efficiency: By utilizing the
and thereby enhancing the overall quality and reli- spatial domain, SDMA allows multiple users to
ability of communications. efficiently share the same time-frequency resources,
thereby enhancing the spectrum efficiency when
• Disadvantages:
CSIT is perfect and the network is underloaded.
1) Inefficient spectrum utilization: In multicasting, 2) Interference mitigation: With perfect CSIT and
all users, from the weakest to the strongest, need an underloaded network, SDMA effectively elim-
to be able to decode the stream. Consequently the inates or suppresses multi-user interference, achiev-
transmission rate of the stream is always determined ing the maximum degrees-of-freedom in under-
by the weakest user, therefore leading to inefficient loaded multi-antenna BC [28].
spectrum and resource utilization. This can be frus- 3) Low transceiver complexity: Thanks to the utiliza-
trating to stronger users who could receive at higher tion of linear precoding at the transmitter and each
rates but are constrained by the weakest user in the receiver’s ability to directly decode the intended
pool. message while treating interference as mere noise,
SDMA exhibits a relatively low level of hardware
C. Space Division Multiple Access complexity at both the transmitter and receivers.
In response to the limitations inherent in OMA and the • Disadvantages:
growing demand for higher data rates, improved QoS, and 1) High sensitivity to CSIT inaccuracy: The ef-
increased network capacity, a new resource dimension - space fectiveness of SDMA is highly affected by the
- has been introduced in modern wireless networks. This gives inaccuracies in CSIT [29], [30]. While SDMA ex-
rise to the widespread integration of multiple antennas in most cels with perfect CSIT in underloaded network, its
wireless access points, signifying the advent of the multiple- performance significantly decreases with imperfect
input multiple-output (MIMO) paradigm since 4G networks. CSIT due to residual interference from imprecise
MIMO has become indispensable in modern and future wire- interference mitigation strategies that are originally
less networks, finding inclusion in nearly all high-rate wireless designed for perfect CSIT.
standards such as WiMAX, 4G LTE, IEEE 802.11n, 5G NR, 2) Limited network load tolerance: SDMA performs
etc. By leveraging the spatial dimension, MIMO networks effectively in underloaded networks but experiences
introduce a novel MA known as SDMA [22], [25], [26]. degradation in overloaded systems due to the con-
SDMA empowers multiple users to share the same time- straints posed by limited spatial resources. To ad-
frequency resources, adhering to the interference management dress this issue, user grouping is usually employed
principle of precanceling interference at the transmitter and in overloaded scenarios, but at the cost of reduced
treating interference as noise at the receivers. QoS and increased latency.
To mitigate interference at the transmitter, SDMA intro- 3) Limited user deployment flexibility: SDMA is
duces precoding techniques, which are typically classified into sensitive to user deployment such as the angles
two primary categories: non-linear and linear precoding. One and strengths of user channels. It works well for
of the most renowned non-linear techniques is dirty paper users with orthogonal channels and similar signal
coding (DPC) [25], [27]. It attains the capacity region of strengths. However, its performance significantly
MIMO Gaussian broadcast channel (BC) when perfect channel degrades when user channels become nearly aligned
state information is available at the transmitter. However, its or exhibit substantial variations in signal strengths.
 6

4) Complex scheduling: Due to the limited user de- 2) Enhanced user fairness: As PD-NOMA requires
ployment flexibility, precise scheduling is impera- power allocation favoring users with weaker chan-
tive for SDMA, leading to extra complexity in user nel strength to enable successful interference can-
scheduling. Moreover, the scheduling algorithms cellation, it is capable of enhancing user fairness
to achieve the (near) optimal performance can be than SDMA in extremely overloaded network with
challenging to implement and maintain. closely aligned users and large channel strength
5) High signaling overhead: The requirements for disparities.
channel estimation, scheduling, and interference • Disadvantages:
management in SDMA leads to substantial signaling
overhead, particularly in dynamic and densely pop-
ulated network environments or when employing a 1) Inefficient use of spatial dimensions: As shown in
massive number of antennas at the transmitter. [10], the sum multiplexing gain of MIMO NOMA
always falls below or equals that of SDMA. This
implies that the slope of the sum-rate of MIMO
D. Power-domain Non-Orthogonal Multiple Access NOMA at high signal-to-noise ratio (SNR) will
PD-NOMA leverages the concept of superposition coding be lower than that of SDMA. This phenomenon
(SC) at the transmitter and successive interference cancella- indicates that PD-NOMA makes an inefficient use
tion (SIC) at the receivers to facilitate simultaneous sharing of the multiple antennas.
of common resources, i.e., time, frequency, code, or space 2) Inefficient use of SIC: In MIMO NOMA, the
among users [13], [31], [32]. This is achieved by allocating number of SIC deployed at each user scales propor-
users with varying power levels, and enabling signals from tionally with the number of users in that user group.
different users superposed in the power domain. PD-NOMA Consequently, compared to SDMA, MIMO NOMA
ensures the effective decoding of these signals at users by introduces two notable issues: a loss in multiplexing
empowering those with weaker power levels to decode the gain and an increase in receiver complexity. This
messages of users with stronger power levels. This approach observation underscores MIMO NOMA’s inefficient
is also referred to as SC-SIC, adhering to the interference utilization of SIC.
management principle of decoding interference. 3) High sensitivity to CSIT inaccuracy: As the inter-
It is well-established in the literature that PD-NOMA based group interference management in MIMO NOMA
on SC–SIC achieves the capacity region for single-input follows the approach employed in SDMA, akin to
single-output (SISO) BC [33], and for the SISO multiple SDMA, MIMO NOMA is also sensitive to CSIT
access channel (MAC) [22], it is also capacity-achieving with inaccuracy.
time-sharing. The performance merits of PD-NOMA over 4) Limited network load tolerance: PD-NOMA has
OMA demonstrated in SISO BC/MAC has driven research into been demonstrated as a capacity-achieving scheme
MIMO NOMA. In the multi-antenna BC, to exploit the spatial in SISO BC, rendering it effective in extremely
domain, MIMO NOMA typically separates users into distinct overloaded scenarios. However, as the number of
groups. Interference from users within the same group is transmit antenna increases, MIMO NOMA suffers
managed by SC-SIC while interference from users in different from performance degradation primarily due to its
groups is treated as noise. When there exists only a single user inefficient use of spatial dimensions. Therefore, PD-
group, MIMO NOMA reduces to SC-SIC, requiring users to NOMA is sensitive to the network load.
decode and remove all interference [32]. 5) Limited user deployment flexibility: With perfect
CSIT and an underloaded network, SDMA effec-
PD-NOMA has not yet been incorporated into emerging
tively eliminates or suppresses multi-user interfer-
wireless standards, though it has been investigated as part of
ence, achieving the maximum degrees-of-freedom
a study item in 5G but was not considered any further in
in underloaded multi-antenna BC.
5G because its gains compared to SDMA/MU-MIMO were
6) High transceiver complexity: In addition to the
not found convincing [34]. As highlighted in [10], there exists
hardware complexity introduced by SIC receivers,
many confusions and misconceptions about PD-NOMA, which
PD-NOMA imposes substantial computational com-
impels us to look into its advantages and disadvantages in the
plexity at both transmitter and receivers. Specifi-
following:
cally, at the transmitter, the joint optimization of
• Advantages: user scheduling, user grouping, decoding orders,
1) Enhanced spectrum efficiency: By utilizing the and precoders is imperative for enhancing perfor-
power domain and advanced receiver techniques, mance. However, solving such resource allocation
PD-NOMA allows multiple users with closely problem is typically challenging with high compu-
aligned channels and diverse channel strengths in tational complexity.
the same time-frequency resources to efficiently 7) High signaling overhead: In PD-NOMA, extra
share the same time-frequency resources, thereby signaling overhead is required to convey information
enhancing the spectrum efficiency when the network such as the decoding order, user grouping result, and
is extremely overloaded. power levels each user should use.
 7

E. Rate-Splitting Multiple Access

RSMA has recently emerged as a promising non-orthogonal
transmission strategy for multi-antenna wireless networks,
owning to its capability to enhance the system performance
in a wide range of network loads, user deployment, and
CSIT qualities [14], [35]–[37]. RSMA is a generalized MA
scheme originally proposed in [35]–[37] for downlink and
in [38] for uplink. The key concept of RSMA is splitting
each user message into sub-messages at the transmitter. For
downlink RSMA, sub-messages are categorized into common
and private with the common sub-messages designed to be
decoded by multiple users, whereas private sub-messages are
intended to be decoded exclusively by their respective users.
For uplink RSMA, sub-messages of a given transmitter can
be decoded at the receiver in a non-consecutive manner.
This message split capability endows RSMA with a flexible
interference management strategy of partially decoding the
interference and partially treating the interference as noise.
Depending on the specific approaches used for message
splitting and combining, RSMA has a range of transmission
models, including linearly precoded 1-layer RS, 2-layer hi-
erarchical RS (HRS), and generalized RS (GRS), each of
which can be further extended to their respective non-linear
counterparts [14], [37], [39], and even to space-time designs
[40].
The most straightforward and practical downlink RSMA
scheme is 1-layer RS [35], [36], which is also the basic Fig. 2. Illustration of the downlink transmission frameworks for 1-layer RS,
building block of almost all existing RSMA schemes. In 1- HRS, and GRS with 3 users.
layer RS, each user message is split into one common sub-
message and one private sub-message (Fig. 2). All the common
sub-messages are combined and jointly encoded as a single RS, the user grouping requires to be considered in HRS to
common stream to be decoded by all users, whereas the private achieve a more flexible interference management capability
sub-messages are encoded individually as private streams. 1- than 1-layer RS.
layer RS requires only one layer of SIC at each receiver. User Unlike 1-layer RS and 2-layer HRS, which maintain a
grouping and decoding order design is unnecessary since all constant number of message splits for each user message
users decodes the unique common stream before decoding its regardless of the number of users, GRS takes a more com-
private stream. prehensive approach by utilizing all possible message splitting
HRS represents a more encompassing scheme compared and combining strategies to achieve generalized transmission
to 1-layer RS, as it introduces additional common streams framework [37]. In GRS, during the message splitting and
tailored for specific user groups (Fig. 2). In HRS, all users are combining process, all potential user grouping are explored
divided into separate groups, each comprising one or multiple and the groups containing different number of users are
users. At the transmitter, each user message is split into categorized into different layers (Fig. 2). For example, the
three sub-messages: an inter-group common sub-message, an group in the first layer contains all users while each group in
inner-group common sub-message, and a private sub-message the last layer only contains a single users. Each user message
[41]. The inter-group common sub-messages and private sub- is split into a number of common sub-messages. The common
messages respectively follow the common and private submes- sub-messages in the same group are combined and encoded
sages in 1-layer RS. The primary difference between HRS and following the basic principle of HRS. At the user sides, a
1-layer RS lies in the inner-group common sub-messages. The number of SIC layers are employed at each user to decode all
inner-group sub-messages for users within the same group are common streams and the intend private stream. As all possible
merged into a group-specific common message, subsequently user grouping has been considered in the message splitting and
encoded into a inner-group common stream. The inner-group combining process, only the decoding order among the groups
common stream is decoded exclusively by users within the in the same layer need to be carefully designed. In Fig. 2,
corresponding user group, following their decoding of the a straightforward comparison of the three-user transmission
inter-group common stream. HRS requires two layers of SIC frameworks among 1-layer RS, HRS, and GRS is illustrated.
at each receiver. Decoding order design in unnecessary since RSMA has been shown to generalize multicasting, OMA,
each user follows the decoding order of inter-group stream, SDMA, and NOMA schemes [11], [14]. Specifically, when
inner-group stream, and private stream. In contrast to 1-layer all transmit power is allocated exclusively to the common
 8

Fig. 3. Illustration of the two-user downlink transmission frameworks and beam shapes for multicasting, OMA, SDMA, PD-NOMA, and RSMA.

stream or one private stream, RSMA respectively simplifies to are illustrated in Fig. 4.
the classical multicasting or OMA. When the transmit power The main advantages and disadvantages of RSMA are
is solely allocated to the private streams, RSMA reduces to summarized as following:
SDMA. Furthermore, in the case where each user message
• Advantages:
is totally encoded into distinct layers of common streams
and private stream within the GRS framework, RSMA re- 1) Universality: RSMA is a comprehensive MA
duces to PD-NOMA. For this reason, RSMA is a superset framework encompassing OMA, SDMA, NOMA,
and therefore always achieves equal or better performance and multicasting as its constituent sub-schemes [11],
compared to OMA, SDMA, and NOMA. A simple two-user [37]. The universality of RSMA obviates the need
downlink transmission framework comparison among multi- for a system to consider switching between OMA,
casting, OMA, SDMA, NOMA, and RSMA is delineated in SDMA, NOMA, and multicasting.
Fig. 3, and the corresponding uplink transmission frameworks 2) Flexibility: RSMA is highly adaptable to varying
network loads, whether they are underloaded or
 9

the achievable rate region of RSMA is larger than

other MA schemes, and it approaches the capacity
region achieved by DPC [37]. When CSIT is im-
perfect, RSMA can achieve a larger rate region than
DPC [30]. RSMA’s enhanced spectrum efficiency is
not limited to conventional multi-user multiple-input
single-output (MISO) unicast transmissions, but also
holds for MIMO unicast [53], [54], multigroup
multicast [24], [42], [55], [56], non-orthogonal uni-
cast and multicast [57], non-terrestrial [24], [58],
[59], network slicing [60]. The benefits have been
demonstrated using link-level simulations [54], [61]
and real-world experimentation [62].
5) Enhanced energy efficiency: RSMA not only
boosts spectral efficiency but also energy efficiency
and their trade-offs across various applications with
different network loads and user deployments, as
evidenced in many existing works [57], [63].
6) Enhanced QoS and user fairness: RSMA achieves
the optimal max-min fair spatial multiplexing gain
in multi-antenna BC with both perfect and imperfect
CSIT. This advantage is reflected in the finite SNR
regime, where RSMA demonstrates its superior
max-min rate compared to the aforementioned MA
schemes. Furthermore, the substantial spectral effi-
ciency gain achieved by RSMA becomes even more
pronounced when stringent QoS rate constraints are
imposed on each individual user.
Fig. 4. Illustration of the two-user uplink transmission frameworks for OMA, 7) Low complexity: 1-layer RSMA stands out for its
NOMA, and RSMA. simplicity at both transmitter and receivers design.
Its adaptability to varying network loads and user
deployments obviates the need for user ordering,
overloaded, as well as to diverse user deployments grouping, and scheduling at the transmitter. Each
characterized by varying channel directions and user only necessitates a single SIC layer to decode
strengths [37], [42], [43]. This remarkable advan- and cancel the common stream. This contrasts with
tage of RSMA comes from its capability to flex- MIMO NOMA, which imposes stringent require-
ibly handle multi-user interference through partial ments for user grouping and ordering at the trans-
interference decoding and treating the remaining mitter along with multiple layers of SIC at each
interference as noise. user.
3) Robustness: The rapid growth of RSMA in multi- 8) Coverage extension: RSMA with user relaying,
antenna networks is primarily attributed to its capac- also known as cooperative rate-splitting (CRS) [64],
ity to achieve an optimal spatial multiplexing gain [65], has been demonstrated as a promising strategy
in the multi-antenna BC with imperfect CSIT [36], to amplify the data rates for users located at the
[40], [44]–[47]. A multitude of extant research en- cell edge, offering substantial coverage extension
deavors has substantiated RSMA’s resilience when benefits [66]. CRS involves collaborative user re-
confronted with CSIT uncertainties stemming from laying, enabling one user to decode and relay the
diverse sources of impairment, such as quantized common stream to other users, thereby enhancing
feedback [40], [48], pilot contamination [49], [50], user fairness particularly when jointly serving users
channel estimation inaccuracies, latency in feed- with substantial discrepancies in channel strengths.
back, user mobility [51], and even RF impairments 9) Low latency: RSMA has demonstrated its prowess
such as phase noise [52]. in improving throughput compared to the afore-
4) Enhanced spectrum efficiency: RSMA is guar- mentioned MA schemes when employing finite-
anteed to achieve the best sum-degree of freedom length (e.g. polar) codes, as highlighted in [67]–
(DoF) in both perfect and imperfect CSIT, which [69]. In other words, RSMA achieves the same
translates into superior spectral efficiency as well. In transmission rate as SDMA and NOMA but with
other words, RSMA outperforms other MA schemes shorter block lengths, resulting in reduced latency.
in terms of spectral efficiency in both perfect and Therefore, RSMA is a promising enabling technol-
imperfect CSIT [36], [37]. When CSIT is perfect, ogy for significantly reducing latency in URLLC
 10

services. particular, the multi-user interference pattern at the receiver

10) Security enhancement: RSMA can adjust the level entails a low-density graph, where message passing algorithm
of confidentiality of its messages and consequently (MPA) or belief propagation (BP) [75] can be adopted for
trade off spectral efficiency with secrecy [70], [71]. efficient symbol detection [76]. To design LDS sequences,
This is not possible with PD-NOMA since the a structural approach was proposed in [77]. Furthermore,
message of the weaker user is always decoded by a the capacity region of LDS-CDMA was investigated in [78],
stronger user, which creates a secrecy threat. where the impacts of spreading sequence density factor and
• Disadvantages: the maximum number of users associated with each chip on
1) SIC requirement at receiver: Some RSMA the capacity were revealed. In addition, the idea of LDS-
schemes, such as GRS, exhibit a characteristic CDMA was extended to OFDM (LDS-OFDM) [79] where
where the number of SIC layers at each user in- each user’s data symbols are spread across a number of care-
creases with the number of users. This trend not fully selected subcarriers. Significant improvement of peak-
only leads to considerable burden on receiver com- to-average-power ratio (PAPR) the link-level performance was
plexity but also introduces the error propagation is- reported for LDS-OFDM over OFDM [80], [81].
sues. Promising SIC-free RSMA architectures have 2) Sparse Code Multiple Access (SCMA): The concept of
recently been developed [61]. code domain multiplexing in LDS-CDMA was extended to
2) High encoding complexity: Compared with the SCMA [82]. In SCMA, each user is assigned a sparse code-
aforementioned MA schemes, RSMA requires more word according to its message. In other words, the modulation
date streams to be encoded due to the additional symbol mapping and the spreading sequences are merged
common streams obtained from message splitting together such that the bits are directly mapped to a sparse
and recombining. vector of a multi-dimensional constellation. Hence, SCMA
3) High signaling burden: In RSMA, additional sig- improves the spectral efficiency of LDS through shaping gains
naling overhead is necessary to facilitate alignment of multi-dimensional constellations. Meanwhile, it still inherits
between the transmitter and receivers, ensuring they the benefits of LDS in terms of overloading and moderate
possess a shared understanding of the methodology complexity of detection. An example of a SCMA system with
for splitting and combining user messages. 6 users and 4 resources is illustrated in Fig. 5. Each user maps
4) High optimization burden: In terms of resource its two bits to its SCMA codeword such that the superimposed
allocation and precoder design, RSMA requires transmitted block spread over the 4 resource blocks as shown
the precoders (accounting for beamformer direc- in Fig. 5(a). The resultant SCMA system can be represented
tions and power allocation) of the common and by a factor graph depicted in Fig. 5(b), where every circle rep-
private streams to be jointly optimized with the resents a user or a variable node (VN), every block represents
common rate allocation. This joint optimization is a resource or factor node (FN), and the edge between VN i
instrumental in unlocking the complete benefits of and a FN j means that user i’s data is mapped to the j-th
RSMA. However, it’s worth noting that this ex- resource block j for i ∈ {1, . . . , 6} and j ∈ {1, . . . , 4}. The
panded optimization space places a more demanding factor graph is regular in the sense that each VN is of degree-2
computational burden on RSMA in comparison to while each FN is of degree-3. The SCMA codebook design
conventional MA schemes. based on lattice constellations [83] was investigated in [84],
[85]. Similar to LDS-CDMA, the SCMA receiver adopts MPA.
F. Code-Domain Multiple Access To reduce the detection complexity of MPA, various detection
algorithms such as discretized MPA and sphere decoding
Code-domain NOMA (CD-NOMA) is inspired by the tradi-
can be employed [85], [86]. Potential applications of SCMA
tional CDMA [15]. It allows multiple users to share the same
in 6G wireless communications systems were discussed in
time/frequency resources while dedicated interleavers and/or
[87]. However, the overloading performance is limited by the
code sequences are employed to multiplex users. However, un-
number of sparse codewords and the sparsity.
like CDMA, the spreading sequences in CD-NOMA are usu-
ally sparse or non-orthogonal low cross-correlation sequences. 3) Multi-User Shared Access (MUSA): MUSA was intro-
In general, CD-NOMA can be divided into sparse NOMA duced in [88] to support grant-free Internet of things [89].
and dense NOMA [72], [73], depending on the sparsity of the In MUSA, the data symbols of each user are spread with
spreading sequences. In this section, we review some popular a short-length spreading sequence. The key principle is that
CD-NOMA schemes and discuss their main advantages and non-orthogonal complex spreading sequences are chosen by
disadvantages. multiple users autonomously for enabling grant-free trans-
1) Low-Density Spreading (LDS) MA: One of the early missions on the same resources. It is also worth noting that
CD-NOMA techniques is the LDS-CDMA [74], an extension the same user can choose different spreading sequences for
to the classical CDMA. In an LDS-CDMA system, the data different symbols to benefit from interference averaging. The
symbols of each user are spread by a unique LDS with receiver exploits the low cross-correlation properties of spread-
a small fraction of non-zero entries and then superimposed ing sequences and uses SIC decoding. However, unlike LDS
before transmission. Compared to the classical CDMA, the multiple access schemes and SCMA, the spreading sequences
interference on each chip can be reduced in LDS-CDMA. In of MUSA are dense.
 11

User 1 User 2 User 3 User 4 User 5 User 6

users to choose their spreading sequences au-
tonomously to reduce the signaling overhead and
latency. In this case, the spreading sequences need
00 10 10 01 11 01 to be carefully designed for the underlying multi-
user detectors to mitigate collisions.
• Disadvantages:
1) High receiver complexity: CD-NOMA requires
complex multi-user detection techniques. In partic-
(a)
ular, to achieve the promised gain of CD-NOMA,
iterative receiver architecture is often required. The
VN 1 VN 2 VN 3 VN 4 VN 5 VN 6
complexity scales proportionally with the number of
users and iteration numbers.
2) High design complexity: The key ingredient of
CD-NOMA is the spreading code sequence. How-
ever, finding the optimal design for the spreading
sequence is still a difficult problem, particularly
when the number of users is large.
3) High channel estimation complexity: Most ex-
FN 1 FN 2 FN 3 FN 4

(b) isting CD-NOMA schemes assume perfect CSI

knowledge. To accurately estimate the channel for
Fig. 5. Illustration of a SCMA system with 6 users and 4 resource blocks:
(a) mapping between each user’s two bits and SCMA codewords; (b) factor each user, channel estimation is often performed
graph corresponding to the SCMA system. jointly with multi-user detection, which incurs a
high complexity.
4) High sensitivity to synchronization: Most existing
4) Successive Interference Cancellation Amenable Multiple CD-NOMA schemes assume perfect synchroniza-
Access (SAMA): SAMA is based on the joint design of tion at the receiver. However, this is not always the
the system signature matrix and the SIC-based MPA [90]. case in practice. A delayed signal from a user could
The symbols of different users are judiciously spread in the disturb the message exchange in the iterative multi-
frequency, which can be effectively exploited by the SIC- user detection, leading to performance loss.
based MPA and to obtain the diversity gain. Different from
the aforementioned LDS schemes, the spreading sequences in
G. Multiple Access in Other Domains
SAMA have variable sparsity.
We summarize the main advantages and disadvantages of In addition to the prominent SDMA, PD-NOMA, RSMA,
CD-NOMA schemes as follows. and CD-NOMA, a range of MA schemes in other domains
have been proposed in the literature. In particular, a full list
• Advantages: of MA schemes proposed during the study phase of 5G NR
1) Enhanced spectrum efficiency: CD-NOMA uti- were included in [7]. In what follows, we will review some
lizes the code domain to allow multiple users share typical examples from [7] as well as some newly proposed
the same resources, e.g., time and frequency, ef- MA schemes.
ficiently. With advanced multi-user detectors, CD- 1) Interleave-Division Multiple Access (IDMA): IDMA
NOMA improves the spectrum efficiency when was introduced in [92] in 2003 and has gained considerable
CSIT is perfect and the network can be either interest in recent years. IDMA can be considered as a spectral
underloaded or overloaded. case of CDMA with all signature sequences reducing to
2) Enhanced user fairness: User fairness in CD- a single chip [93]. In other words, IDMA relies on user-
NOMA can be achieved by detecting strong users’ specific interleaving as the only means to distinguish the
signals first in the multi-user detection. This is signals from different users [94]. Moreover, the spreading
because the early detected users only collect less operations in CDMA can be replaced by low-rate forward error
extrinsic information compared to the late detected correction codes in IDMA to provide increased coding gain.
users [91]. In this way, more extrinsic information It is also possible to have non-orthogonal interleavers. The
together with interference cancellation performed in IDMA receivers adopt a chip-by-chip elementary signal esti-
the early detection steps can improve the detection mation, such that the total receiver complexity only increases
performance of weak users. linearly with the number of users [95]. Advantages of IDMA
3) Flexibility: CD-NOMA is adaptable to various net- over CDMA in terms of performance and complexity under
work loads and channel conditions. This is owing to practical considerations were demonstrated in [96]. Integration
the effective design of spreading code sequences in of IDMA with other technologies such as OFDM and massive
terms of the correlation and sparseness for reducing MIMO were investigated in [97] and [98], respectively.
inter-user interference. 2) Pattern Division Multiple Access (PDMA): In PDMA
4) Low signaling overhead: CD-NOMA can allow [99], the transmitted data symbols are mapped to a resource
 12

group that can consist of time, frequency, and spatial resources User 1's
Constellation

or any combination of these resources according to a pattern.

Data of multiple users can be multiplexed onto the same re-
source group with a different pattern to realize non-orthogonal User 1 Decoder

transmission. In addition, the pattern is designed with disparate (no SIC)

diversity order and sparsity such that the BP algorithm can be

efficiently employed for detection. To achieve the best possible
performance, an iterative turbo receiver architecture [100] can Base Station
Superimposed
User 2 Decoder
(no SIC)
Encoder
be adopted, where the outer iteration process between the Constellation

detector and decoder is performed on top of the inner iteration

of the detection and decoder themselves. A gain of 500% in User 2's

terms of the number of supported users under the given system Constellation

packet drop rate of 1% was reported for PDMA over OFDMA Fig. 6. An illustration of the lattice-partition-based multiple access with
in the uplink [101]. In [102], the patterns of different users are single-user TIN decoding.
judiciously designed to exhibit appropriate diversity disparity
at the symbol level and power disparity at the physical resource
element level. In this way, the appropriate disparity in diversity structure. It was proved that the scheme based on discrete
and power can be effectively exploited by the low-complexity signaling and TIN is capable of achieving the whole capacity
SIC-based BP detector with reduced error propagation during region within a constant gap independent of the number
interference cancellation. of users and channel parameters. Several new MA schemes
3) Compute-Forward Multiple Access (CFMA): Compute- evolving from lattice-partition-based multiple access were
and-forward (C&F) is a relaying strategy introduced in [103], proposed for various channels, e.g., the Gaussian interference
where the relay decodes a noisy linear combination of lattice channel, the Gaussian BC with heterogeneous blocklength and
codewords. In contrast to the classic relaying strategies such as error probability constraints, etc. [110]–[113], where carefully
amplify-and-forward and decode-and-forward, C&F exploits designed discrete signalings with TIN decoding are shown to
interference to obtain significantly higher rates between users be close to Gaussian signalings with perfect SIC decoding
in a network. Motivated by the benefits of C&F, [104] intro- analytically and numerically.
duced CFMA based on a modified C&F technique for the 5) Spatially Coupled Multiple Access (SC-MA): Spatial
Gaussian MAC. The receiver first decodes the sum of the coupling [114] is a code construction technique to boost the
linear combination of lattice codewords. Upon recovering the performance of uncoupled (weak) codes, e.g., regular LDPC
sum codewords, other users’ codewords can be successfully codes and turbo codes, all the way to capacity-achieving
decoded by using the sum as side information. It was proved [115]–[118]. Motivated by this, [119], [120] introduced a new
in [104] that whole capacity region of the two-user Gaussian signaling formats for the Gaussian MAC, where the modulated
MAC is achievable by CFMA under lattice decoding √ [105], data streams are repeated, permuted, and transmitted with
provided that the SNR of both user is larger than 1 + 2. It is regular time offsets (delays). Since the relations between
worth mentioning that any rate points on the dominant face of the data bits and modulation symbols transmitted over the
the Gaussian MAC capacity region can be achieved by CFMA, channel can be represented by a sparse graph, the receiver
without the need of time-sharing [106, Ch. 15] or rate-splitting observes spatial coupling of the individual graphs which
[38]. However, random lattice codes were used in [104] as enables efficient demodulation/decoding. It was proved in
the proof techniques, which can be difficult to implement in [119], [120] that coupling data transmission with a two-stage
practice. Practical implementation of CFMA based on off-the- demodulation/decoding in which iterative demodulation based
shelf binary low-density parity-check (LDPC) codes and sum- on symbol detection and interference cancellation followed
product decoding was conducted in [107]. by parallel decoding achieves the Gaussian MAC capacity
4) Lattice-Partition-Based Multiple Access (LPBMA) with- asymptotically such that the gap to capacity vanishes as the
out SIC: In [108], A lattice-partition framework of MA system’s SNR increases. Note that the feedback between the
without SIC, i.e., with single-user treating interference as noise decoder and the demodulator is not required. In contrast to
(TIN) decoding, was introduced for the K-user Gaussian BC. most NOMA receivers that have to demodulate/decode the
Each user adopts a linear code [109] in conjunction with whole symbol block, sliding window demodulation/decoding
an appropriately designed constellation carved from a multi- can be adopted thanks to the spatially coupled graph struc-
dimensional lattice which includes the commonly adopted ture [121]. As a result, the demodulation/decoding latency is
one- and two-dimensional constellations, i.e., pulse ampli- limited, which can be suitable for streaming services. [122]
tude modulation (PAM) and quadrature amplitude modulation extended the spatially coupled transmission scheme to the
(QAM), as special cases. Essentially, the design ensures that case of transmitted blocks with irregular delays. An example
the superposition of all users’ signalings still preserves the for the transmission scheme is illustrated in Fig. 7, where
lattice structure, which can be exploited to hardness inter- each user divides its transmission frame into sub-blocks and
user interference in TIN decoding. A two-user example is delays the transmission by τ . However, the spatially coupled
illustrated in Fig. 6, where both users’ signal constellations transmission scheme can be sensitive to synchronization errors
and the superimposed constellation preserve the same lattice and fractional delay.
 13

Stream 7
9) Other Multiple Access: All the above discussion on MA
User 4 techniques is not exhaustive with new MA strategies disclosed
Stream 5 on a regular basis such as fluid antenna multiple access
User 3 (FAMA) [133], location division multiple access (LDMA) for
Stream 2 Stream 4 near-field communication [134], or orbital angular momentum
User 2 (OAM) multiplexing and multiple access.
Stream 1 Stream 3 Stream 6
User 1
H. Technology Outlook and Future Works
2 3 4 5 6 7 time We draw several observations from the past subsections on
common challenges in MA designs and on MA dimensions
Fig. 7. Examples of spatially coupled multiple access where data streams are
transmitted with different delays. before giving an outlook of MA technology development
toward Universal Multiple Access.
1) Common Challenges in MA designs: Modern and emerg-
ing MA schemes share a number of common challenges,
6) Layered-Division Multiplexing (LDM): LDM is a non-
namely their optimization complexity (of precoders and re-
orthogonal multiplexing technology adopted in the Advanced
source allocation), channel estimation complexity, receiver
Television Systems Committee standard (ATSC 3.0) [123],
complexity (involving highly complex operations at the re-
which has never been implemented in previous broadcast
ceivers such as SIC), and sensitivity to various impairments
and broadband systems [124]. In an LDM system, a lay-
(can be synchronization or RF impairments).
ered transmission structure is used to simultaneously transmit
multiple signals with different power levels and technologies 2) MA dimensions: Previous discussions highlight that
(channel coding, interleaving, modulation, multiple-antenna, there are numerous dimensions that a MA scheme can exploit,
etc.) for different services [125]. At the receiver, the decoding starting from the usual time, frequency, space, power, code, but
of signals from multiple layers requires SIC. The principle new dimensions have also appeared in recent years including
of LDM is similar to PD-NOMA. Hence, its gain over the message, index, pattern, interleaver, etc. We have decided to
conventional OMA schemes has been well understood. Note classify them all into five categories: time, frequency, power,
that LDM is implemented on top of OFDM such that the space, signal. Signal is a broad category that encompasses
superposition of signals take place before OFDM modulations. multiple sub-dimensions such as message split and combiner,
channel coding, modulation, interleaver, spreading sequence,
7) Index Modulation Multiple Access (IMMA): Index mod- etc. Table II displays a non-exhaustive list of dimensions
ulation is a technique of using the indices of resources to carry exploited by the aforementioned MA schemes and how afore-
extra information bits [126]. It can achieve higher spectral mentioned MA schemes exploit those various dimensions.
and energy efficiency than conventional modulation schemes. Table II also highlights that MA techniques exploit one
Motivated by this, IMMA was proposed to further improve or multiple of those dimensions. For instance, SDMA and
the spectral and energy efficiency of the traditional NOMA PD-NOMA exploit the space domain and the power domain,
scheme [127]. In IMMA, each user transmits additional infor- respectively, while RSMA by opening the door to the mes-
mation bits (index bits) by partial resources, where the index sage dimension exploits the message combiner, message split,
patterns can be the activation status of time slots, subcarriers, power, and space domains to generalize and unify SDMA,
transmit or receive antennas, spreading codes, power levels, NOMA, and physical layer multicasting. PDMA exploits
etc [128]. However, the detectors of these IMMA schemes are space, time and frequency domains and CD-NOMA schemes
joint maximum-likelihood detectors which can be challenging exploit the spreading code domain together with time and
to implement. frequency. All five domains would be exploited by OFDM(A)-
8) Delay-Doppler Domain Multiple Access (DDMA): Or- RSMA [135], [136].
thogonal time-frequency space (OTFS) modulation [129]– 3) RSMA - A First Step Toward Unification in 6G: The
[131] was introduced for high-mobility wireless applications, beauty of RSMA is that it unifies into a single MA scheme
which multiplexes data in the delay-Doppler (DD) domain the seemingly unrelated strategies of PD-NOMA, SDMA,
rather than the time-frequency domain as in conventional and physical layer multicasting. This unification capability
multi-carrier modulations. OTFS has also stimulated new translates in RSMA to be a superset of those three strategies
research on delay-Doppler plane modulations. For instance, and can boil down to any of them by turning off some of the
the newly proposed orthogonal delay-Doppler division multi- streams. This was illustrated in Fig. 3 for a two user scenario
plexing (ODDM) [132] was proved to achieve orthogonality on and has been discussed extensively in the NOMA and RSMA
the DD plane’s fine resolutions with realizable pulses. Owing literature [9], [10], [14] and demonstrated by the well known
to the orthogonality, ODDM itself can be potentially employed message-to-stream mapping of [9], [11]. Consequently, RSMA
as an OMA scheme on the DD domain. Another promising can softly bridge those three strategies, explore operating
research area is that other MA schemes such as SDMA and points that are not achievable by any of them, and outperforms
RSMA can be combined with OTFS/ODDM, which can offer them all. This gives RSMA an edge over other MA schemes as
high spectral efficiency with robust performance in high- demonstrated in over 40 applications in 6G [9], [14], including
mobility channels. in multi-functional networks such as ISAC [137].
 14

TABLE II
MA DIMENSION CLASSIFICATION AND MAPPING OF MA SCHEMES TO MA DIMENSIONS . RSMA BASIC SCHEMES EXPLOIT SPACE , POWER , AND
MESSAGE SPLITTING DIMENSIONS , THOUGH SPACE - TIME AND SPACE - FREQUENCY RSMA SCHEMES EXIST [9], HENCE EXPLOITING A FOURTH
DIMENSION . IMMA HAS PRIMARILY TWO DIFFERENT SCHEMES , AND THE CURRENT DESIGNS UTILIZE THREE DIMENSIONS AT THE SAME TIME TO
DISTINGUISH USERS .

Dimension Signal
Time Frequency Space Power Message Message Channel Spreading
Modulation Interleaver
MA combiner split coding sequence
TDMA ✓
FDMA ✓
OFDMA ✓ ✓
SDMA ✓
PD-NOMA ✓
PHY Multicasting ✓
RSMA ✓ ✓ ✓✓✓ ✓ ✓✓ ✓ ✓✓ ✓✓✓
CD-NOMA ✓ ✓ ✓
IDMA ✓ ✓
PDMA ✓ ✓ ✓
CFMA ✓ ✓ ✓
LPBMA ✓ ✓ ✓
SC-MA ✓ ✓
LDM ✓ ✓ ✓
IMMA ✓✓ ✓ ✓ ✓✓
DDMA ✓ ✓ ✓

The capability of RSMA to unify MA schemes is crucial standing of the current and future morass of MA schemes.
for the long term research in the theory and practice of This question, vision, and research philosophy is illustrated
communications and wireless systems. Recall indeed the wise in Fig. 8 and Fig. 1. Following the above question, come the
words in the acknowledgments section of the book of Prof. natural and important questions “How to design UMA?” and
David Tse and Prof. Pramod Viswanath “Bob Gallager’s “Why and when do we need UMA?”.
research and teaching style have greatly inspired our writing There are no answers to those quesitons yet since UMA
of this book. He has taught us that good theory, by providing does not exist. Those questions nevertheless aim to trigger
a unified and conceptually simple understanding of a morass discussions, give research directions, and motivate further
of results, should shrink rather than grow the knowledge research. Addressing them is very timely for 6G and beyond
tree.” [22]. In our context, as illustrated in Fig. 8 and by the for two main reasons. First, the literature and the number of
RSMA branch in Fig. 1, RSMA, by providing a unified and MA schemes have exploded in the past decade and grown
conceptually simple understanding of a morass of results on non-organically, fueled by the sheer interest in making new
SDMA, PD-NOMA, physical layer multicasting, shrinks rather or better use of MA dimensions and the limited resources
grows the knowledge tree of MA schemes based on space, but also by the large number of new wireless services of-
power, signal dimensions. The capability of RSMA to unify fered by multi-functional and intelligent 6G. Many of those
and therefore be more universal than other MA schemes makes schemes claim to exploit new dimensions and presenting new
practical implementation and operation easier. Indeed a single ideas, though it remains to be seen whether this is true or
unified and general MA scheme would be easier to implement whether those are a rebranding or twist of known concepts.
and optimize than a combination of multiple MA schemes, Second, the multi-functionality of wireless networks calls for
each optimized for specific conditions. This is increasingly a more efficient, flexible and robust use of resources, better
important in multi-functional 6G and beyond networks where management of interference, better handling of heterogeneity
the range and diversity of services, use cases, and deployments of wireless services, unified and simplified hardware and
explode. software architectures. As illustrated in Fig. 9, UMA should
4) Toward Universal Multiple Access in Beyond 6G: shrink the knowledge tree by truly understanding the essence
Though RSMA provides a good example of a unified theory of a unified MA design at exploiting in the most simple
of MA schemes, RSMA does not unify all MA schemes and way all five dimensions (time, frequency, space, power, sig-
therefore does not exploit all dimensions of time, frequency, nal) to efficiently provide intelligence and multi-functionality
power, space (e.g. antennas, beams), signal (e.g. messages, (communications, sensing, localization, computation, energy
codes, etc). transfer and harvesting) in 6G and beyond network.
This would then bring the central question for future re- Many MA schemes exploiting the code domain have been
search on MA in beyond 6G: “What is Universal Multiple classified as CD-NOMA with LDS and SCMA being instance
Access (UMA)?”. Following the same wise philosophy of Prof. of sparse CD-NOMA and MUSA and SAMA of dense CD-
Gallager, Prof. Tse, and Prof. Viswanath, UMA should further NOMA. However, CD-NOMA here only stands for a col-
shrink the knowledge tree of MA schemes by unifying RSMA lection MA schemes exploiting various properties of codes
with all other dimensions, such as code domain MAs, and but CD-NOMA is not a MA scheme itself that unifies LDS,
ultimately provide a unified and conceptually simple under- SCMA, MUSA, SAMA, i.e., there is no MA scheme that
 15

Fig. 8. From RSMA to UMA: How to further shrink the knowledge tree.

remain largely unexplored. Indeed, a number of works have

attempted to combine CD-NOMA with other MA schemes.
The combination of PD-NOMA and SCMA was investigated
in [138]–[140]. The motivation is to allow the same SCMA
codeword to be used by multiple users simultaneously, thereby
increasing the number of supported users. At the receiver,
MPA combined with SIC decoder is employed. However, the
complexity would be higher than that of a single SCMA
or a PD-NOMA scheme. In [141], CD-NOMA is combined
with SDMA in the massive MIMO setting, where the cases
of CD-NOMA offering spectral efficiency improvement were
identified. The integration of CD-NOMA into RSMA was
suggested in [9]. Since the split common and private messages
in RSMA can be seen as virtual users, each virtual user can
be assigned a dedicated spreading sequence for enabling code-
domain multiplexing. However, such an idea has not been
Fig. 9. UMA exploiting all five dimensions to enable multi-functional 6G.
explored further.

unifies LDS, SCMA, MUSA, SAMA and enables to softly

III. A RTIFICIAL I NTELLIGENCE FOR M ULTIPLE ACCESS
bridge them all. Hence CD-NOMA does not have the same
unification capability as RSMA. Nevertheless, it is important
to observe that RSMA and CD-NOMA span different MA In this section, we demonstrate how AI techniques can be
dimensions, with e.g., RSMA not exploiting the spreading used to address some of the challenges highlighted in the
code dimension. Shrinking the knowledge tree of MA schemes previous section. We start by providing an overview of AI
to identify UMA would require a better understanding of how methods and delve into how they can be applied for resource
all CD-NOMA schemes are linked to each other and whether allocation and optimization for different MAs. We then discuss
they form particular instances of a more general class of MA AI-empowered channel estimation, receiver design, and user
schemes, and a better understanding of the interplay between behavior predictions, and finish the section by providing some
RSMA and CD-NOMA. Unfortunately such research avenues outlook of AI for UMA.
 16

A. AI-empowered MA Resource Allocation and Optimization systems. In this application, SVM is used to classify segments
of the spectrum as either occupied or available for secondary
In every wireless communication system, there is a constant users based on received signal characteristics, enabling effi-
challenge: harnessing limited resources to achieve improved cient spectrum utilization [142]. KNN works by finding the K
performance while satisfying various QoS requirements such closest data points in the training data to a test data point, then
as rate, latency, and reliability. Therefore, the allocation of predicting or classifying based on the majority label or aggre-
wireless resources is of critical importance in the pursuit gated values of those nearest neighbors. It’s non-parametric,
of optimized network performance and the fulfillment of so it doesn’t make any assumptions about the data distribution.
the multifaceted needs inherent to wireless communication In applications like user grouping, KNN helps clustering
systems. Different MA schemes introduce distinct resources or categorizing users based on their similarity in terms of
for allocation. For instance, OFDMA introduces the challenge channel conditions or interference patterns [143]. SVM and
of subcarrier allocation among users, TDMA involves time slot KNN are both supervised learning algorithms, and they share
allocation, NOMA introduces user pairing, grouping, decoding common challenges when dealing with large datasets due to
order, and power allocation, SDMA incorporates beamform- their high computational complexity. For KNN, it suffers the
ing, power allocation, and user scheduling, RSMA centers “curse of dimensionality” wherein data points tend to become
on beamforming, power allocation, common rate allocation equidistant from each other in high-dimensional spaces, posing
design, and SCMA deals with sparse code allocation among difficulties in identifying meaningful nearest neighbors. In
users. Essentially, the choice of the optimal resource allocation contrast, both K-means clustering and PCA are unsupervised
approach relies on various factors, including the specific MA learning algorithms. K-means clustering can be applied to
scheme, available resources, system requirements, and the wireless resource allocation problems such as user grouping,
intricate interplay of these key factors. spectrum allocation. For example, a K-means based user
Conventional resource allocation algorithms typically rely clustering algorithm is developed in [144] for NOMA by
on optimization theory and game theory. While these methods exploiting the channel correlation among users. PCA is a
can yield mathematically optimal, sub-optimal, or Nash equi- dimensionality reduction technique and itself is not a direct
librium solutions, they often come with a high computational resource allocation algorithm for wireless networks. However,
cost. In the future 6G wireless networks, the substantial it plays a crucial role as a preprocessing and analysis tool
proliferation of communication devices and antennas, along that enhances the quality of data used by resource allocation
with the exponential growth in data traffic, brings formidable algorithms [145].
optimization challenges at a large scale. As a result, these con- All the aforementioned non-neural network-based ML al-
ventional resource allocation algorithms will face an increased gorithms offer straightforward interpretations of results. How-
computational burden. ever, their effectiveness is pronounced when dealing with
The rapid advancement of AI in the past decades has data characterized by straightforward relationships. They may
opened up a new approach to tackling the complex, non- encounter difficulties in capturing intricate hierarchical repre-
convex problems often encountered in resource allocation. sentations within complex data relationships. In such case, DL
This approach allows the resource allocation scheme to be or neural works demonstrate superior performance.
learned directly from data samples or environment, eliminating 2) Deep Learning: The DL approach, powered by neural
the need for complex mathematical models. Fig. 10 shows a networks, is one of the most renowned techniques in the field
non-exhaustive search of the potential AI methods that can of AI for its success in addressing various challenges, such
be applied for resource allocation. Broadly, AI-based resource as image recognition and natural language processing. Within
allocation methods fall into three categories: traditional ML, the realm of wireless resource allocation, DL has shown to be
reinforcement learning (RL), and deep learning (DL). In the valuable thanks to its capacity to efficiently process vast and
following, we provide a concise overview of these three intricate datasets. It facilitates intelligent and adaptive resource
methodologies and how those AI-based approaches can be allocation, continually optimizing network performance in
applied to diverse resource allocation scenarios, followed by real-time. Popular neural networks mainly consist of dense
a summary of the advantages and disadvantages for using AI- neural networks (DNN), convolution neural networks (CNN),
based approaches. recurrent neural networks (RNN), and graph neural networks
1) Machine learning: The term “machine learning (ML)” (GNN). Each of these architectures has demonstrated its
originally referred to the development of algorithms that prowess in addressing specific resource allocation challenges
enabled machines to learn and solve specific problems. With for different MA schemes.
the growing development of neural network-related methods, DNN: DNN is the most fundamental neural network ar-
ML now primarily refers to traditional learning approaches that chitecture, often referred to as multi-layer perception (MLP).
do not rely on neural network. Common ML methods include Within this architecture, neurons are associated with input
support vector machine (SVM), K nearest neighbors (KNN), weights and incorporate an activation function to produce out-
and K-means clustering and principal component analysis puts. The computation unfolds layer by layer, with parameter
(PCA). updates facilitated through the process of backpropagation. It’s
SVM is often used for resource allocation scenarios that worth noting that this structural framework finds application
involve binary data classification. For instance, it can be ap- in various other types of neural networks as well. DNNs
plied to spectrum sensing in dynamic spectrum access (DSA) have been employed to approximate the weighted minimum
 17

Fig. 10. A non-exhaustive search of AI methods that can be applied for resource allocation.

mean square error (WMMSE) algorithm, a versatile non- challenges [147].

convex optimization technique for resource allocation, in an CNN: In contrast to DNNs, CNNs leverage convolutional
end-to-end manner [146]. This pioneering approach signifies a and pooling layers for specialized operations. By stacking
paradigm shift in optimization known as “learn to optimize”. It these convolutional and pooling layers alongside dense layers,
has demonstrated that a three-layer DNN can closely approxi- CNNs significantly reduce the number of parameters required
mate WMMSE while achieving significant computational time to achieve equivalent performance to that of DNNs. CNNs
reductions, often spanning orders of magnitude. Considering excel in feature extraction and are notably easier to train com-
the widespread adoption of WMMSE in wireless resource pared to DNNs. This advantageous has led to the widespread
allocation tasks such as beamforming and power allocation for adoption of CNNs in addressing resource allocation chal-
different MA schemes, this achievement implies that DNNs lenges. CNN has been trained to approximate WMMSE for
can effectively address a broad range of resource allocation power allocation in [148], and for SDMA beamforming design
 18

in [149]. In [150], the power control problem of NOMA is also efficiency of resource allocation methods like power control,
addressed by CNN. channel assignment, or user scheduling by giving decision-
RNN: RNNs take into account the impact of inputs from making algorithms a more concise and informative input. The
the previous time steps. This distinctive feature allows RNNs fundamental concept behind transfer learning involves extract-
to maintain memory of prior inputs through their unique ing essential features from the source domain and fine-tuning
network architecture. Therefore, RNNs are well-suited for the pre-trained model for application in the target domain. By
handling time-dependent data. This characteristic of RNNs leveraging its capability to transfer valuable prior knowledge
make them highly valuable to capture the temporal correla- to new scenarios, transfer learning offers a promising solution
tions of data. In [151], RNN is used by each base station to address the challenge of task mismatch encountered in
to predict the spectrum allocation. Additionally, in [152], a practical wireless communication systems. The key to meta-
graph attention RNN is proposed for traffic prediction. The learning is the ability to “learn-to-learn” so that the trained
RNN family includes various common variants, including long model can improve its learning ability and adapt to new tasks
short-term memory (LSTM) and gated recurrent unit (GRU), or domains with minimal or no human intervention. This
among others. These variants have found highly effective in approach facilitates rapid adaptation to new environments, thus
tackling various resource allocation challenges. LSTM is used enabling more effective resource allocation decisions. In this
to forecast the user mobility, and enhance the performance context, meta-learning was found to be a powerful compu-
of resource allocation [153]. In terms of beamforming design tationally efficient alternative to WMMSE-based optimization
for SDMA, LSTM emerges as a powerful tool for integrating techniques to optimize precoders and resources allocation in
information from multiple preceding steps in order to refine the RSMA [163]. The performance obtained with meta-learning
current step, as evidenced in [154]. This contrasts to conven- is comparable to convex optimization-based approaches but
tional alternative optimization algorithms, which typically rely achieves a significantly lower running time. This is partic-
solely on information from the present and the immediately ularly helpful for large scale RSMA network settings with
preceding step to influence their output. For GRU, it can serve many antennas and users, where WMMSE approaches are not
as a robust predictive model capable of consistently delivering feasible due to the astronomic computational time, and low
high levels of accuracy in forecasting resource requests [155]. complexity precoding algorithms achieve poor and clearly sub-
GNN: GNNs are specifically built upon graph structures, optimal performance.
offering the capability to handle non-Euclidean data. GNNs Additionally, algorithm-driven methods like deep unfold-
exhibit remarkable flexibility in accommodating input data ing/unrolling and learning to branch and bound (BB) have
of varying dimensionality, making them a natural fit for the also played significant roles in wireless resource allocation for
dynamic nature of the wireless communication domain, where different MA schemes. As the resource allocation problems
the number of users and antennas can vary significantly. By are typically non-convex, one line of algorithm design is to
constructing directed/undirected graphs based on the structure approximate the original non-convex problem to a sequence
of the considered wireless networks with devices as nodes and of convex problems, and use an iterative algorithm to solve
channels as edges, GNNs are capable to learn the wireless the approximated problems. Deep unfolding/unrolling extends
network and enhance resource allocation [156], [157]. As an this idea by unfolding the iterations of an iterative resource al-
alternative, GNNs can construct graphical models based on location algorithm into a neural network architecture, allowing
the mathematical formulation of specific resource optimization for end-to-end training and optimization. Such approach has
problems by defining node sets and incorporating parameters been widely studied in beamforming design, user scheduling,
as features, as evidenced in [158], [159]. By such means, power allocation, etc [164], [165]. Another standard method
GNNs are promising for addressing large-scale resource al- to deal with the non-convex optimization problem is to find
location problems for various MA schemes while enjoying a the global optimal solution based on BB, which is of high
high computational efficiency. computational complexity. To address this issue, a learning-
Other Methods: In addition to the previously mentioned based BB is proposed that use neural network to learning the
standard neural network architecture, numerous advanced DL BB algorithm [166].
approaches have gained widespread adoption in addressing 3) Reinforcement Learning: RL involves the interaction
resource allocation problems. These encompass a range of between agents and the environment, with the goal of optimiz-
techniques, such as the transformer [160], auto-encoder [161], ing a long-term objective. Among the most widely embraced
transfer learning [162], meta-learning [154]. The transformer model-free RL algorithms, Q-learning is the most well-known
takes the advantage of the attention mechanism, which sets it algorithm for computing an optimal policy that maximizes the
apart from RNNs. It is suitable for parallel computing, leading long-term reward. However, as system complexity deepens,
to substantial performance enhancements when compared to particularly when there exists hidden system states, calculating
RNNs. An autoencoder, a neural network for unsupervised and maintaining all Q-values becomes exceedingly imprac-
learning, contains two parts: an encoder that compresses input tical. This challenge, often referred to as the “curse of di-
data into a simpler form, and a decoder that recreates the mensionality”, makes Q-learning less suitable for the intricate
original input from this simpler representation. In wireless demands of ultra-dense and complex wireless networks [167].
resource allocation, autoencoders are used to learn efficient To address this issue, deep RL (DRL) emerges as a powerful
data representations that capture important patterns in com- solution that integrates traditional RL with deep learning
munication data. These learned representations improve the techniques.
 19

TABLE III
AI- EMPOWERED RESOURCE ALLOCATION FOR DIFFERENT MULTIPLE ACCESS SCHEMES .

MAs AI Method Reference Resource Allocation Problem

Deep Transfer Reinforcement Learning [172] beamforming and subcarrier assignment
OFDMA [173] subcarrier assignment
DNN [174] power allocation
[175] user scheduling
DRL [176] subcarrier assignment and power allocation
RNN (LSTM) [177] radio resource assignment
TDMA
DRL [178] time-frequency resource block allocation
DRL (Q-learning) [179] power allocation
K Nearest Neighbors [180] beamforming
K-means [181] user grouping
[182] hybrid beamforming
SDMA
CNN [183]
[149]
[184] beamforming
GNN [185]
[186]
GNN [187] user scheduling
NOMA DQN [188] user pairing and power allocation
DNN [189] joint beamforming and SIC decoding
Transformer [190]
RSMA DNN and autoencoder [191] beamforming
Meta-Learning [163]
DRL [192] power allocation

Recent advances in DRL has made it a promising frontier these algorithms are generally proficient in tackling resource
for resource optimization in wireless network for different allocation challenges of diverse sizes and complexities.
MA schemes [168]–[171]. DRL has found applications in Low computational complexity: AI inference demands only
a wide range of resource allocation problems, such as user a limited set of basic operations and can be executed in real-
scheduling, power control, beamforming design, and band- time. This appealing characteristic allows these algorithms
width allocation. DRL demonstrates remarkable capability to reduce processing time and require fewer computational
in tackling the intricate challenges of sequential decision- resources, making them practical and cost-effective solutions
making within dynamic and large-scale networks, all without for both online and offline resource allocation problems.
relying on explicit models of the transmission environment. Robustness: As AI-empowered algorithms are data-driven,
This unique feature empowers agents to continually refine they are typically robust to handle uncertainty in data or envi-
their decision policies through interactions with the unknown ronment, allowing them to make reliable decisions despite in-
environment. complete or noisy information. This robustness is particularly
4) Advantages and Disadvantages: A non-exhaustive valuable in resource allocation problems because it ensures
search of existing works that use AI methods to allocate that the algorithm can adapt and make sound decisions even
resources for difference MA schemes is summarized in TA- when operating in unpredictable or dynamic environments.
BLE.III. The advantages and disadvantages of these AI- Although AI-empowered resource allocation solutions offer
empowered algorithms are summarized as follows. numerous advantages, they also come with certain disadvan-
AI-empowered resource allocation algorithms offer the fol- tages and challenges:
lowing advantages: Lack of Interpretability: Directly replacing the conventional
Superior Performance: Thanks to their data-driven features resource allocation algorithm with a generic neural network
and the well developed neural networks, AI-empowered re- in an end-to-end fashion is often referred to as a “black-box”
source allocation can discover communication patterns and approach, which lacks interpretability regarding the decisions
relationships that are too complex to identify precise math- it makes. To address the issue, one can combine data-driven
ematical model. Therefore, AI-empowered algorithms often and model driven algorithms, i.e., by unfolding an iterative
achieve better performance compared to traditional model- algorithm, it becomes possible to enhance interpretability.
based algorithms. Data Dependency: AI-empowered algorithms have a strong
Generalization: AI-empowered resource allocation algo- dependence on data for both training and inference. When the
rithms can generalize from historical data to make predictions training data is biased, incomplete, or not representative, it can
or decisions on uncharted data. This capability empowers result in suboptimal resource allocation decisions. As the wire-
these algorithms to flexibly adapt to rapidly changing wireless less environment is changing rapidly, user channels, typically
environment and efficiently allocate resources. used as input data, may shift to different distributions over
Scalability: AI-empowered resource allocation algorithms time. This can significantly impact the ability of the trained
can efficiently handle and process larger volumes of data. neural network to generalize effectively to new conditions and
They possess the ability to accommodate growing demands scenarios.
without compromising performance or efficiency. Therefore, Training Overhead: Developing and training AI models
 20

for wireless resource allocation can be time-consuming and OFDM system, where a DNN is employed to learn the char-
resource-intensive. This is especially evident when utilizing acteristics of frequency selective wireless channels. After that,
traditional resource optimization algorithms, known for their many existing works focus on using neural networks, such as
high computational complexity, to generate the training data, CNN, generative adversarial network (GAN) to tackle channel
leading to substantial time requirements. Furthermore, to estimation of MIMO networks [199]–[201]. As summarized in
maintain optimal performance, periodic model updates may the survey paper [202], most of existing works on channel esti-
also be a necessity. mation consider point-to-point communication networks, only
limited studies have delved into the application of deep learn-
ing to uplink channel estimation in TDD multi-user MIMO
B. AI-empowered MA Channel Estimation systems for achieving less pilot-aided channel estimation. In
CSI plays a pivotal role in wireless communications. On [203], a deep learning based joint pilot design and channel
one hand, precise CSIT facilitates the implementation of estimation scheme is proposed for uplink multi-user MIMO.
adaptive resource allocation and user scheduling, leading to To mitigate inter-user interference received at the base station,
substantial improvements in performance and efficiency. On uplink NOMA is considered at the base station, which employs
the other hand, accurate CSI at the receiver (CSIR) empowers SIC for channel estimation. To jointly design pilot and channel
coherent detection and decoding processes. For acquiring estimation schemes, a two-layer neural network is proposed for
CSIR, downlink channel estimation is a necessity, where non-orthogonal pilot design and another deep neural network
the transmitter sents pilot signals to the receivers. This task is constructed for non-linear channel estimation. The proposed
becomes particularly vital and challenging in OFDM systems approach is shown to achieve better performance than the
due to the temporal variability and frequency-selective char- traditional MMSE estimator and the deep learning approach
acteristics of wireless channels. In terms of CSIT acquisition, without SIC. [204] further simplified the joint pilot signals
in time division duplex (TDD) systems, CSI is obtained at the and channel estimator design by removing the use of SIC at
transmitter through the exploitation of reciprocity. In this case, the base station. Furthermore, considering a millimeter wave
uplink channel estimation is required, where users send pilot massive multi-user MIMO systems with limited number of RF
signals, enabling the base station to estimate the uplink CSI. chains, a prior-aided Gaussian mixture learned approximate
However, tackling this task in massive access is challenging message passing (GM-LAMP) approach is proposed in [205]
due to the limited number of orthogonal pilot resources [193]. to exploit the sparsity of beamspace channels and deal with
In frequency division duplex (FDD) systems, acquiring CSIT the beamspace channel estimation problem. Simulation results
involves sending the CSI estimated at the receiver back to the show that the proposed GM-LAMP network considering the
transmitter through a feedback link. However, this becomes prior distribution can enhance estimation accuracy with a low
challenging in massive MIMO setups due to the large dimen- pilot overhead. Considering a grant-free NOMA transmission
sions of the CSI, resulting in a substantial increase in feedback network, the authors in [206] propose to deploy a DNN at the
overhead. To mitigate the overhead, it becomes imperative base station for identifying the active devices and estimating
to efficiently compress the CSI estimate at the receiver or corresponding channels. Numerical results show that such
employ a codebook-based approach for quantizing the CSI approach achieves a higher user activity detection success
[194], [195]. probability and more accurate channel estimation compared
AI can play a pivotal role in tackling the aforementioned to the conventional approaches, especially in the small packet
challenges and improving the quality of CSI obtained at either transmission scenarios for the grant-free NOMA system.
the transmitter or receiver. This can be achieved through two AI-enabled CSI compression: In FDD massive MIMO sys-
primary directions [196]: tems, to reduce the CSI feedback overhead, one line of
works focuses on CSI compression methods and codebook
1) Channel estimation: By modeling the channel estimation designs. While compression theories such as compressive
problem as a regression task, we can create a neural sensing have been applied to simply the CSI feedback or
network-based channel estimator. codebook design [207], [208], the feedback signaling overhead
2) CSI compression: By modeling CSI compression as in these approaches remains heavy since the overhead typically
a dimension reduction problem, we can develop AI increases linearly with the number of antennas, imposing
architectures specifically tailored to address this issue. practical limitations. Deep learning has shown to be highly
These two directions in general apply to different MA effective in addressing data compression challenges across var-
techniques. Note that for different MA enabled channel estima- ious domains, including images, audio, and video [209], [210].
tion, some additional problems might introduce. For example, Recognizing CSI as an additional data source for compression,
in TDD uplink NOMA, the same pilot might be shared by deep learning techniques have been successfully applied to the
multiple users. Additional power control issue is introduced, compression and feedback of CSI for massive MIMO systems
which can also be addressed by AI methods [197]. In the [211]–[217]. In [211], an autoencoder-based deep learning net-
following, we will respectively discuss the existing works on work called CsiNet is proposed to handle CSI compression and
these two directions for different systems with different MA feedback. By deploying an encoder at each user to compress
schemes. the CSI into a unique codeword for efficient feedback and
AI-enabled channel estimation: Deep learning enabled chan- a decoder at the base station to reconstruct the CSI from the
nel estimation is first studied in [198] for a point-to-point received codeword, CsiNet shows significant compression effi-
 21

ciency enhancement compared with conventional compressive algorithms and model-agnostic DL approaches, model-based
sensing-based approaches. Built upon CsiNet, various deep deep learning (MBDL) has been introduced. This innovative
learning based approaches have been investigated for massive approach aims to combine the simplicity inherent in traditional
MIMO CSI compression and feedback to enhance CSI recon- model-based algorithms with the model-agnostic nature of DL
struction accuracy while aligning with practical requirements. techniques, as highlighted in [218]. To realize this integration,
Taking temporal dynamics into account, [212] further extends MBDL systems are implemented by substituting specific steps
CsiNet to handle time-varing channels by proposing a LSTM- and computations in model-based algorithms, which rely on
based deep learning network called CsiNet-LSTM. To address precise channel model knowledge, with compact neural net-
the difficulties of transmitting continuous codeword values, works. These neural networks demand smaller training data
recent works incorporate feedback quantization into network sets compared to conventional DL systems. In the realm of
training so as to improve the network performance with low- wireless receivers, prior adaptations of the SIC receiver using
resolution feedbacks [213], [216], [217]. MBDL have been proposed for uplink and downlink NOMA
in [220], [221], as well as for downlink RSMA in [222].
In these studies, specialized DNNs are employed for tasks
C. AI-empowered MA Receiver Design
such as interference cancellation and symbol classification.
Aforementioned MA schemes often require advanced re- The findings demonstrate that the MBDL approach surpasses
ceivers. For instance, PD-NOMA and RSMA commonly rely the performance of the conventional model-based SIC receiver
on SIC, though other types of receivers can be used, as specifically in the context of NOMA and RSMA. The illus-
demonstrated recently in [61] for RSMA where the tradeoff tration in Fig. 11 showcases an MBDL receiver designed for
between performance and receiver complexity of various in- RSMA, offering a visual comparison with the conventional
terference cancellation and joint demodulation and decoding SIC approach.
architectures have been compared.
One commonality in the design of advanced receivers for
RSMA or other types of MAs is the reliance on models D. AI for MA User Behavior Predictions
and assumption for the noise, interference, signal propagation, Due to the multi-functionality of wireless networks and
etc [218]. It is nevertheless challenging to mathematically the heterogeneity of users, various user and communication
relax those assumptions while maintaining a tractable model. service requirements exist. Different services usually lead to
Indeed interference is often non-Gaussian with finite constel- different QoS requirements and different resource manage-
lations, SIC is imperfect leading to error propagation, CSIR ment schemes. However, conventional resource management
is imperfect, decoding delay and latency increases chance of strategy adopts static average resource allocation methods and
errors, etc. Those model-based receiver algorithms have so far ignores the disparity of service from different users, which
performed quite well [61] but take the risk to rely on accurate makes it hard to meet the QoS requirements of all services with
prior model knowledge and potentially perform poorly if it is limited wireless resource. Thus, it is important to investigate
not accurately acquired [218]. the user service behavior prediction to match the wireless
AI/BL-based receivers are able to directly extract mean- communication system configuration with the diverse service
ingful information from the unknown channel solely on ob- QoS demands, thus improving system efficiency, decreasing
servations, which is a major advantage. Therefore, AI/DL network delay, and guaranteeing the QoS demands of each
is well suited for scenarios in which the underlying math- user.
ematical channel model is unknown, its parameters cannot According to [223], a base station captures user behavior
be acquired with precision, or when it is too complex to be information such as user identity number, user Internet access
studied by model-based algorithms with low computational traffic costs, terminal brands, mobile product brands, user call
resources [219]. Nevertheless, employing conventional AI/DL fees, network request base station latitude, network request
techniques for wireless receivers presents substantial chal- base station longitude, value-added service fee, mobile phone
lenges. Firstly, the appeal of model agnosticism in AI/DL operating system, call start time, call end time, call base station
approaches necessitates a complex network with numerous latitude, call base station longitude, and network request time.
nodes and layers, along with a sizable training set to acquire Through analyzing the obtained history user mobile broadband
a specific mapping. Consequently, this results in a substantial information with the help of ML technology, the future user
computational burden at the receiver during the training phase. behavior can be predicted and then future communication QoS
Secondly, sharing a training set between the transmitter and requirements can be further obtained [224]. Based on the
receiver introduces significant training overhead, potentially predicted communication QoS requirements, resource man-
causing delays in transmitting actual useful data to commu- agement such as bandwidth allocation, base station on-off
nication users. Lastly, due to the time-varying and dynamic control, and transmit beamforming, can be configured in time,
nature of wireless channels, periodic training of the DNNs thus providing seamless coverage. Moreover, to support a
at the receivers becomes essential to accommodate channel wide diversity and number of users, different MA schemes
variations. However, this requirement renders the transmission (including OMA, PD-NOMA, SDMA RSMA, CD-NOMA)
of large training sets highly impractical. could be applied, which also pose design challenges in ML for
In an attempt to tackle the aforementioned challenges and user behavior predictions due to different signal transmission
leverage the advantages of both conventional model-based schemes in different MA schemes.
 22

detection

detection
LPR Decoder Splitter
Equalizer LLR Decoder Splitter

+ Interference cancellation
- Encoder
Channel Precoder Encoder

detection
LPR Decoder Combiner
Equalizer LLR Decoder Combiner

SIC receiver Interference cancellation /

detection

MBDL receiver

Fig. 11. 1-Layer Model-Based Deep Learning RSMA receiver and relationship with 1-Layer SIC RSMA receiver [222].

Based on the types of predicted user behavior information, 2) Wireless Traffic Prediction: The wireless traffic includes
the main research areas in ML for user behaviour predictions the traffic duration time and traffic data requirements, and
in MA mainly include four aspects: user activity prediction, can be explored via reinforcement learning and other AI
wireless traffic prediction, content popularity prediction, and techniques. Using reinforcement learning, in [233], the Duel-
user mobility prediction. In the following, we introduce these ing deep-Q network is applied to obtain the wireless traffic
four aspects with pointing the design difference in the aspect prediction and server load balance in OMA system. The
of MA. wireless traffic prediction models with AI are categorized
1) User Activity Prediction: User activity indicates that into statistical, ML, and DL models [234]. The ML models
the user is active or not. In the communication system, include random forest, LightGBM, Gaussian progress regres-
the user activity information can reflect the total number of sion, multiple linear regression (MLR), and Prophet, while the
served users. User activity prediction has first been studied in deep learning models [235], [236] include feed-forward neural
cognitive radio networks, where, to efficiently utilize dynamic networks (FFNNs), CNNs, and RNNs.
spectrum resource in such networks, ML techniques can be For random access (see Section VI-E), supervised learning-
used to predict the primary users’ activity and the secondary based method is proposed to improve the access performance,
user can choose the idle channel [225]–[227]. ML techniques which uses the history data to obtain the time-varying trend
for user activity prediction in OMA cognitive radio networks of wireless traffic [237], [238]. Besides, the LSTM-based
such as TDMA are mainly supervised ML, including artificial prediction algorithm can be used to characterize the time
neural networks (ANN) and SVM [228]. Furthermore, for related property of traffic, which is helpful in predicting the
the multi-channel multi-hop mobile cognitive radio networks, locations of aerial base stations, such as unmanned aerial ve-
the Probabilistic Suffix Tree based ML algorithm is used hicles (UAVs) [239]–[241]. With the predicted UAV locations,
to learn the primal user’s activity pattern [229]. Using the various uplink MA schemes are considered in [241], such as
history space-time access records in mobile broadband data, TDMA, FDMA, NOMA, and RSMA. To minimize the total
the scalable online expectation-maximization based algorithm transmit power, it is found that RSMA can achieve the best
to fast learn latent Dirichlet allocation has been studied for performance among all schemes since the RSMA can achieve
user activity level prediction [230]. With the location-based all points in the capacity region.
social network data, [231] shows the feasibility of using 3) Content Popularity Prediction: The basic principle of
ML algorithms for user activity prediction, which include content-based content prediction is to obtain the user’s in-
gradient-boosted trees, deep learning, logistic regression, and terest preferences based on the user’s historical behavior,
generalized linear model. and recommend objects similar to his interest preferences to
Different from OMA, there exists interference among users the user. Content prediction generally requires three steps,
with NOMA. In particular, users need to be grouped for namely constructing user feature representation based on user
forming NOMA transmission, which require the user activity information and user operation behavior, constructing subject
information. However, due to the dynamic channel gain, the feature representation based on subject information, and repre-
user grouping and user activity are always dynamic and senting user recommended subject based on user and subject
coupled, thus making it hard to predict the user activity in characteristics. Having predicted the content popularity, the
NOMA. Through taking user activity as an unknown mapping, most preferred content can be cached near the intended users
the work in [232] uses the LSTM to approximate this mapping. in advance, thus improving the system performance including
The framework that integrates LSTM into NOMA is shown to delay and energy [242].
improve the user activity performance in terms of reliability To design a comprehensive popularity prediction scheme
and complexity. for OMA systems, a multi-head attention based popularity
 23

prediction model is introduced in [243] to extract the multi- same reasons as RSMA but also because of the additional
dimension features. Since multiple users can be served si- complexity increase incurred by the the exploitation of the
multaneously in NOMA and the content storage capacity is code dimension.
always limited, these exists a tradeoff between the resource AI for user behaviour predictions will also be helpful to
for NOMA transmission and content caching [244]. To solve UMA to predict in advance the user activity, wireless traffic,
this tradeoff, an autoencoder-based algorithm is proposed content popularity and user mobility and accordingly feed
in NOMA enabled fog radio access networks for content those information to the AI-empowered resource allocation
popularity prediction in NOMA system [244]. and receiver architecture for most efficient use of all MA
4) User Mobility Prediction: The user mobility information dimensions of UMA.
including speed, direction, and angle, plays a crucial role in
the prediction of the wireless environment [245]. Due to the IV. M ULTIPLE ACCESS FOR A RTIFICIAL I NTELLIGENCE
mobility of users and limited coverage area of one base station, We now switch the role and assess how MA schemes can
handover is needed to maintain the QoS requirement during be designed and tailored to serve AI applications. We focus
the whole moving process of the user. In [245], state-of-art on MA for federated learning (FL), edge intelligence and over
ML approaches for user mobility prediction are summarized the air computation.
including the movement predictability, prediction outputs, and
performance metrics. Considering the wireless communication
protocol and standard, a ML based mobility prediction frame- A. MA for Federated Learning and Edge Intelligence
works is proposed in [246] in 5G core network. Further for FL represents a decentralized method for training ML
supporting the emerging applications such as automotive, the models. It eliminates the need for transferring data from
framework of RNN together with LSTM is proposed in [247]. client devices to global servers. Instead, the model is trained
The time-varying property of user movements can be viewed locally using raw data on edge devices, thereby enhancing
as a time-varying graph and a graph based ML network can data privacy. One challenge of deploying FL [252] over
be used in [248] for user mobility prediction. wireless networks is the communication bottleneck, which
Since the user movement is always related to the physi- arises from several edge devices transmitting large sized
cal environment, high-accuracy prediction of user movement model parameters to a parameter server or a base station.
requires AI technology to obtain the relationship between Moreover the heterogeneity of edge devices computing ca-
the user movement and environment information [249]–[251]. pabilities, communication rates, and amount and quality of
In order to solve the high complexity and spatio-temporal data can affect the training performance in terms of accuracy,
correlation of user mobile behavior, multi-dimensional and fairness and convergence time. Researchers have attempted
multi-scale user behavior prediction method based on ML aims to reduce the resultant communication latency using different
to fully capture the dynamic and complex spatio-temporal rela- approaches such as excluding slow devices (“stragglers”) [253]
tionship between individual mobile behavior and the physical or compressing FL model parameters by exploiting their
environment. sparsity. In this section, we provide a detailed literature review
In summary, by analyzing the history user behavior data on the use of MA schemes for supporting FL deployment
in different MA schemes, the ML approaches can predict over realistic wireless networks, with a specific emphasis on
various kinds of behavior information, such as user activity, OMA/SDMA/PD-NOMA/RSMA-based networks for wireless
wireless traffic, content popularity, and user mobility, as shown FL performance optimization.
in Fig. 12. 1) OMA for FL: Using OMA, devices must use different
spectrum resource for FL parameter transmission. Due to
limited resources in wireless networks, the wireless resource
E. Technology Outlook and Future Works (i.e., power, computational power, spectrum) allocated to each
AI-empowered resource allocation, optimization, channel user is limited, which may significantly limit the number of
estimation, user behaviour predictions have been investigated devices that can participate in FL and increase FL convergence
partially for some relatively simple MA schemes such as time [254]–[256]. Consequently, it is necessary to optimize
OMA, but more needs to be done for more advanced MA resource allocation such that devices can efficiently complete
schemes, such as RSMA and CD-NOMA, for which the the FL training process. In [257], the authors implemented
literature still remains quite limited with a lot of new research FL over a hierarchical network architecture and they showed
challenges and opportunities on the horizon. that local training with the help of small base stations (SBSs)
AI-empowered resource allocation, optimization, and chan- can not only speed up the learning process but also increase
nel estimation is likely to play an even bigger role for UMA for communication efficiency by frequency reuse across multiple
which the optimization space is enlarged due to the exploita- SBSs. The works in [258], [259] investigated the trade-off be-
tion of all five dimensions (and sub-dimensions). Moreover, tween local and global ML model updates to minimize the total
AI could also be helpful to complement communication and energy consumption used for local ML model transmission and
information theory to discover and identify key ingredients of update. In [260], the authors used gradient statistics to select
UMA designs and architectures. the devices to join FL training per communication round. The
Building upon RSMA and CD-NOMA, UMA is likely to authors in [261] analyzed how user scheduling affects the FL
benefit even more from AI-empowered receiver design for the convergence and then optimized spectrum resource for the
 24

User Activity Wireless Traffic Content Popularity User Mobility

ML algorithm

On Content 1
Off

Content 2
Traffic rate
History data
Predicted

Time

Fig. 12. An illustration of ML for user behavior predictions.

selected users. The work in [262] designed a FL algorithm for channel hardening and favorable propagation, and designed
devices with non independent and identically distributed data a novel communication design for FL in a MIMO wireless
and then optimized resource allocation to improve FL conver- system. In [272], the authors considered the application of FL
gence and training loss. In [263], the authors jointly optimized in wireless networks featuring uplink multi-user MIMO, and
resource allocation and device scheduling to maximize the aimed at optimizing the communication efficiency during the
model accuracy within a training time constraint. The authors aggregation of client-side updates by exploiting the inherent
in [264] designed a multi-armed bandit based method to select superposition of RF signals.
FL participating devices without the knowledge of wireless
channel state information and data information of devices. 3) PD-NOMA for FL: To further improve the FL parameter
To further improve FL performance in OMA based networks, transmission efficiency, one can use NOMA to separate edge
current researchers also study the joint optimization of wireless devices into distinct groups and managed the interference
resource allocation and other advanced technologies such as in each group by SC-SIC. Currently, the authors in [273]
RIS [265], [266], MIMO [267], compression [268]. The main exploited PD-NOMA together with adaptive gradient quan-
drawback of OMA is that it does not scale well with the tization and sparsification to facilitate efficient FL parameter
number of devices. Specifically, the required radio resources transmission from edge devices to the parameter server. In
increase linearly with the number of transmitters, or else the [274], the authors studied power allocation of PD-NOMA
latency will grow linearly. This calls for the integration of FL for distributed gradient descent in wireless FL with the aim
with other MA schemes. of minimizing the learning optimality gap under privacy and
power constraints. The work in [275] jointly optimized IoT
2) SDMA for FL: To solve OMA drawback and enhance device scheduling, transmit power allocation, and computation
the applications of FL, one can use MIMO and spatial do- frequency allocation for FL in a PD-NOMA and Relay enabled
main techniques, which enable edge devices to use the same wireless network. In [276], the authors studied the PD-NOMA
time-frequency resources for FL parameter transmission thus assisted FL in which a group of edge devices form a PD-
improving wireless resource utilization and FL convergence NOMA cluster to send their locally trained models to the
speed. The authors in [269] studied the deployment of FL cellular base station for model aggregation and the base station
over a cell-free MIMO based network and jointly optimized adopts wireless power transfer to power edge devices for their
the local FL training accuracy, transmit power, data rate, and data transmission and local training. They jointly optimized the
users’ processing frequency to minimize FL convergence time. wireless power transfer of the base station, PD-NOMA based
The authors in [270] studied the joint use of MIMO and FL parameter transmission schemes, the processing rates of the
compression techniques for FL parameter transmission and base station and edge devices, as well as the training accuracy
analyzed the reconstruction error of the designed method and of the FL, with the objective of minimizing the system-wise
its impact on the FL convergence rate. The work in [271] cost accounting for the total energy consumption as well as
considered two unique characteristics of massive MIMO – the FL convergence latency.
 25

Model Update at Edge Device Model Update at Edge Device

Rate Splitting
Local
Local
Distributed Dataset
Gradient Compression Gradient
Dataset
Gradient Updates
and Modulation Computation Distributed Gradient
Over the Air
Gradient Updates Computation
. Over the Air
. .
.
.
Model Update at Edge Device .
Parameter Server Local Model Update at Edge Device
Dataset
Parameter
Gradient Compression Gradient Local
Broadcast of and Modulation Computation Server
Global Gradient Broadcast of Dataset
Edge
Virtual Devices Global Gradient Gradient
Devices Computation
Edge
Devices
Fig. 13. An illustration of an RSMA-based FL.
Fig. 14. An illustration of an analog AirComp FL.

B. Over The Air Computation

4) RSMA for FL: Since PD-NOMA decodes the interfer-
ence of all other users, it lacks in flexibility and the designed In this section, we introduce an alternative MA approach –
FL algorithms in NOMA based networks may not be used over-the-air computation (AirComp) for wireless FL perfor-
for scenarios where each user needs to transmit with a strict mance optimization. In particular, AirComp enables edge
rate demand due to the high dimension of model gradient. devices to achieve over-the-air FL parameter aggregation via
Moreover, since the FL process requires multiple transmission the waveform superposition property of a multi-access channel
rounds and accurate knowledge of the CSI in all those time [280]. Hence, when edge devices implement AirComp, the FL
slots is unlikely, the use of MA schemes robust to imperfect transmission delay per iteration will not depend on the number
CSI is important in FL applications. However, NOMA cannot of edge devices thus improving FL convergence speed. Note
be well adapted to the imperfect CSI scenarios [10]. To that AirComp uses “interference” among device transmission
address these issues and benefit from the space and power for functional computation via uplink devices’ cooperation.
domains, one can use RSMA to efficiently balance the signal This contrasts with the usual uplink MA schemes that aim
and interference and cope with imperfect CSI. In particular, at suppressing interference among transmitted messages as
due to the flexibility of signal design and decoding, RSMA devices transmit independent data and do not collaborate. In
can achieve robust performance with imperfect CSI, which is this section, we first introduce the basic concept AirComp.
suited to the long-term transmission in FL. Fig. 13 shows an Then, we introduce the state-of-art of using AirComp for FL.
illustration of an RSMA-based FL. In this figure, each device 1) Analog AirComp: The main idea of AirComp is sum-
is considered as two virtual devices that jointly transmit local marized as follows. First, AirComp requires all edge devices
FL model with different data rates. to simultaneously transmit FL parameters to a parameter
server (PS) such that their transmitted FL parameters can
Currently, several works have focused on the use of RSMA be superimposed over-the-air and their weighted sum (global
for FL performance optimization. In particular, the authors in FL parameters) is received by the PS and the weights of
[277] designed an RSMA based uplink radio access scheme each edge device’s FL parameters is determined by channel
for wireless clustering FL where the edge devices performing coefficients, as shown in Fig. 14. To make the weights of edge
local training are closely located in a cluster. By exploiting the devices uniform and eliminate the bias of model aggregation,
message split capability, RSMA has been shown not only to AirComp based FL requires each device to modulate its
significantly reduce the latency of clustering FL in comparison signal using linear analog modulation and to invert its fading
to systems employing TDMA and NOMA, but also to be more channel by transmit power control. AirComp based FL can
energy efficient, reaching a lower latency with less power than compute a broad class of nomographic functions [281] such
the latter alternatives. Another promising scenario of RSMA as polynomial, Euclidean norm, arithmetic mean, weighted
for FL is in a fog radio access network (F-RANs). F-RAN sum, and geometric mean. It was shown in [282] that Analog
leverages proximity edge nodes (ENs) to collect local model AirComp can be optimal in terms of minimizing the mean
parameters of wireless devices [278]. By splitting the message squared error (MSE) distortion if all the multi-access channels
in two parts in uplink RSMA as in Fig. 13, one part is and sources are Gaussian and independent.
intended to be decoded by the ENs and the other part by A key requirement for implementing AirComp based FL
the cloud. By adjusting the content and the transmit power is time synchronization of devices’ transmissions. To address
of each part, one can softly combine the edge-decoding and this challenge, one can use timing advance methods which
cloud-decoding approaches and consequently achieve a better have been used in several existing practical communication
tradeoff between fronthaul overhead (communication cost) and systems (e.g., LTE). The timing advance technique requires
learning accuracy. The work in [279] designed an RSMA- each edge device to estimate the FL parameter propagation
based Internet of Vehicles solution that jointly optimized delay and then transmitting in advance to “cancel” the delay.
platoon control and FL training. Thereby, different signals can overlap with sufficiently small
 26

Model Update at Edge Device of channel codes to get aggregation results. An illustration of
Local a digital AirComp FL is shown in Fig. 15.
Distributed
Dataset Several works have studied the design of novel digital
Gradient Compression Gradient
Gradient Updates and Modulation Computation AirComp FL. Specifically, the authors in [288] analyzed the
Over the Air convergence rate of digital AirComp based FL. In [289], the
.
. authors studied the use of MIMO and beamforming techniques
. to aggregate FL model parameters over the air. Different
Model Update at Edge Device from the works on analog AirComp based FL that do not
Parameter
Server
Local process the FL model signals that will be transmitted to the
Dataset
Broadcast of server, the work in [289] considered parameter quantization
Global Gradient Gradient Compression Gradient
and Modulation Computation and modulation before FL signal transmission and hence it can
Edge
Devices
be considered as digital AirComp based FL. The authors in
[290] presented an orthogonal frequency-division multiplexing
Fig. 15. An illustration of a digital AirComp FL. (OFDM)-based digital AirComp system for wireless FL, where
multiple edge devices transmit model parameters simultane-
ously using non-orthogonal OFDM subcarriers, and the server
misalignment. The signal distortion of AirComp stems from aggregates data directly from the superimposed signal. The
channel noise and interference of analog modulated signals. work in [291] designed a digital broadband over-the-air aggre-
Therefore, channel noise and interference will affect the model gation, which features one-bit gradient quantization followed
updates of AirComp based FL. The impacts of channel noise by digital quadrature amplitude modulation (QAM) at edge
and interference on FL model updates can be evaluated by the devices and over-the-air majority-voting based decoding at
corresponding FL performance metrics. edge server.
The authors in [283] analyzed the convergence and opti-
mized the performance of analog AirComp based FL over
device-to-device communication networks. In [284], [285], C. Technology Outlook and Future Works
the authors optimized device selection and power allocation
Different MA techniques have their own advantages and
for analog AirComp based FL. The work in [286] exploited
disadvantages as discussed in Section II. For FL, the selection
the use of intelligent reflecting surface to assist FL model
of MA techniques depend on the number of participating edge
parameter transmission for analog AirComp based FL. In
devices, wireless resource (i.e., spectrum, transmit power),
[287], the authors designed the FL model parameter retrans-
computational power, and hardware (i.e., CPU or GPU), FL
mission schemes for analog AirComp FL. Finally, the authors
model architecture of each device. To find the optimal MA
in [271] introduced a novel random orthogonalization scheme
schemes, one must first analyze the convergence of FL and
for FL parameter transmission while considering two unique
figure out how different MA schemes affect FL convergence.
characteristics of massive MIMO - channel hardening and
However, in practice, the selection of MA schemes must
favorable propagation.
be performed before FL implementation. Hence, the central
2) Digital AirComp: Analog AirComp uses discrete-time
controller (i.e., base station or parameter sever) may not
analog transmission for FL parameter transmission and ag-
be able to know channel state information and FL model
gregation. However, the performance of analog AirComp
setting information (i.e., gradient information), which may
is sensitive to the phase offsets among the super imposed
increase the difficulty of MA scheme selection. Therefore, MA
signal. In particular, if the signals cannot be synchronized,
scheme selection also depends on the available wireless and
the errors can also be accumulated thus leading to significant
FL setting information. The above discussion further motivates
aggregation errors. To address this problem, digital AirComp
the design of unified and universal schemes such as RSMA
is proposed which enables multiple devices to use non-
and ultimately UMA, which would completely eliminate the
orthogonal subcarriers for data transmission. Different from
need for selection of MA schemes and whose design and
analog AirComp that relies on accurate channel precoding,
optimization could be fully integrated with FL. Also, the use
digital AirComp leverages digital modulation and channel
of over the air computation opens the door to other types of
codes to combat channel impairments and phase misalign-
uplink MA schemes that brings back analog computation into
ments, thereby achieving accurate model aggregation even
MA design picture.
when the signal phases of multiple edge devices are misaligned
at the parameter server. In particular, for digital AirComp
based FL, local FL model parameters must be preprocessed by V. M ULTIPLE ACCESS FOR I NTEGRATED S ENSING AND
quantization, source coding, channel coding, and modulation. C OMMUNICATIONS
Then, the pre-processed FL parameters will be transmitted
over wireless channels, and superimposed in the air. The In this section, we discuss the use and design of MA
receiver will receive the aggregated FL parameters and decode schemes for ISAC, including joint sensing and commu-
them instead of decoding the FL parameters transmitted from nications, multimodal sensing-aided communications, digi-
several edge devices. Here, the receiver will first demodulate tal twin-assisted communications, communication-aided sens-
phase-asynchronous symbols, and then removes the protection ing/localization systems
 27

A. Joint Sensing and Communications

ISAC has gained increasing interest in recent years. This
is primarily motivated by the potential of sharing the same
spectrum and, hence, providing access to more bands for
both the communication and sensing systems. Operating at
the same band, however, requires careful coordination of the
various wireless resources and necessitates the development
of new communication and sensing systems as well as joint
processing frameworks. For multiple access, and when a base
station or access point is jointly serving communication users Fig. 16. RSMA-assisted dual-functional radar-communication systems with
one base station serving K users while tracking a target’s direction [300].
and performing sensing functions, it is critical to efficiently
allocate the available resources in time, frequency, space,
code, power, etc., and optimize the waveform for both the based on PD-NOMA can provide a better trade-off between
communication and sensing objectives. Next, we summarize communication and sensing performance compared to the con-
a few key directions for multiple access in ISAC systems and ventional ISAC schemes based on OMA. However, the number
highlight important challenges in the ISAC multiple access of SIC scales proportionally with the number of users, which
design. can increase the decoding latency and complexity and may
1) SDMA-Assisted ISAC: Based on the waveform diversity not be suitable for delay-sensitive sensing and communication
of MIMO radar and the spatial degrees of freedom of SDMA, applications. In addition, SIC by fully decoding interference
[292] embeds information stream into radar pulses for target as in PD-NOMA may not achieve the optimal multiplexing
detection at the mainlobe and performing communication gain when combined with multiple antennas [10], which leads
tasks at the sidelobe. The communication symbols at the to a performance loss of PD-NOMA Assisted ISAC compared
sidelobe level can be modulated via amplitude shift keying to other multi-antenna MA schemes [18].
(ASK) [292] or phase shift keying (PSK) [293]. That said, 3) RSMA-Assisted ISAC: Inspired by the advantages of
the communication rate is limited by the Pulse Repetition RSMA in robust interference management, spectral efficiency,
Frequency (PRF) in those works. To overcome this limitation, and user fairness, the interplay between RSMA and DFRC sys-
[294] proposed a joint MIMO radar-communication system, tems was proposed in [137]. As illustrated in Fig. 16, a radar
where a joint waveform shared by both radar and commu- sequence and RSMA-processed common and private commu-
nication users is transmitted to serve both sensing and com- nication streams are transmitted to achieve the dual function
munication purposes. To design the probing beampattern, the of probing targets and communications. The precoder and
precoders are designed using zero-forcing (ZF) beamforming common stream of RSMA were carefully designed to achieve
under the communication SINR constraints. [295] extended optimal sensing and communication performance. In [18], it
the framework of [294] for probing multiple radar targets was shown that RSMA-assisted ISAC has been extended to
and communicating to multiple users simultaneously, where the scenario with partial CSIT and users with mobility, satellite
individual communication and radar waveforms are jointly systems with integration of communications and moving target
precoded to achieve the maximum DoF of the MIMO radar sensing in space networks [298], and the DFRC systems with
waveform. The advantages of integrated dual-functional radar- low-resolution digital-to-analog converter (DAC) on each RF
communication (DFRC) MIMO transmission over separated chain [299]. These works demonstrate that RSMA-assisted
antenna deployment for communication and sensing in terms ISAC is very effective in managing multi-user/inter-beam in-
of better trade-off between communication and sensing per- terference while simultaneously performing radar functionality
formance have been demonstrated in [21] and the references with the required mean square error constraints. Importantly,
therein. However, the ISAC systems based on SDMA may the RSMA-assisted DFRC framework unifies and embraces
not perform well in overloaded scenarios and the impacts of existing SDMA and NOMA-based DFRC frameworks as sub-
imperfect channel estimation on sensing and communication schemes. Consequently, RSMA-assisted DFRC outperforms
performance trade-off have not been fully investigated. the SDMA and NOMA counterparts [18]. However, the joint
2) PD-NOMA Assisted ISAC: The concept of PD-NOMA optimization of precoding, common, and private stream rate
has been applied to ISAC for the uplink multiple access allocation can be a demanding computational burden at the
channel and the downlink broadcast channel [296]. ISAC transmitter side.
schemes based on PD-NOMA can be characterized into two 4) Technology Outlook and Future Works: The design of
types. The first type exploits the transmitted superimposed efficient MA techniques for ISAC systems is a challeng-
signals intended for communications for sensing. As a design ing task. First, the joint optimization of the resources and
example, [297] considered maximizing the weighted sum waveforms is typically associated with high computational
of the communication throughput and the effective sensing complexity. This complexity mainly stems from the differences
power by designing the beamformers. The second type further in the communication and sensing objectives and from the
exploits sensing signals for communications, where extra in- various hardware and operation constraints on these systems.
formation is included in the sensing signals whose interference This becomes more complicated in large-scale MIMO systems
to the communication users is removed by SIC. ISAC schemes which are main components in 5G, 6G, and beyond. Going
 28

5G systems, and with the potential use of sub-THz bands in

6G and beyond. The key trends, while promising high data
rates, pose multiple challenges for the design and operation
of future communication systems. For example, the use of
large numbers of antennas increases the overhead of estimating
the channels or training the beams, making it hard for these
systems to support highly mobile applications. Further, the
use of high frequency bands makes the signal propagation
very sensitive to blockages, which challenges the reliability
and latency requirements for these networks.
While the use of large antenna arrays and high frequency
bands impose new challenges on these communication sys-
tems, they also make the communication links very dependent
Fig. 17. The figure illustrates the key idea of multi-modal sensing-aided
multiple access, where infrastructure-side and user-side multi-modal sensing
on the user location and on the geometry of the surrounding
information from cameras, LiDARs, radars, position sensors, etc., is leveraged environment. Therefore, the knowledge of the user position
to optimize the multiple access decisions. and surrounding environment could potentially aid the com-
munication system decisions. This motivated the use of various
multi-modal sensory data, collected for example from visual
from nodes to networks, it becomes critical to design multiple
cameras, LiDAR, radar, position/IMU sensors, etc., to assist
access approaches for networks of integrated sensing and com-
future communication systems [303]. Over the last few years,
munication nodes. This requires accounting for the possible
researchers investigated the potential gain of leveraging multi-
interference between different nodes in both the sensing and
modal sensing data in multiple problems. For example, [304]–
communication signals. With that, it becomes interesting to
[311] showed that position, visual, radar, and LiDAR data
explore the trade-off between coordination performance gains
can guide the beam selection/alignment at the base station
and complexity in multiple access ISAC networks. This results
and user, leading to significant reduction in the beam training
in a set of approaches that range from no coordination to
overhead. This was then extended to beam tracking where the
fully coherent ISAC operation where the full network acts
sensing data is utilized for predicting future beams, providing
as one cell, leading to cell-free ISAC systems [301]. How to
a practical solution for relaxing the data capture, processing,
design multiple access in these coordinating ISAC networks
and prediction latency requirements [312], [313]. Multi-modal
is an interesting research direction. Finally, the fundamental
sensing has also been leveraged to proactively predict future
trade-off between sensing and communication performance of
blockages [314]–[320], predict channel statistics [321], design
general ISAC systems with multiple access has not been fully
codebooks [322], design MIMO precoding without traditional
understood. Only a few prior works have characterized the
CSI acquisition and feedback [320] among other interesting
trade-off for very basic channel models with small numbers
applications.
of users, e.g., [302]. Hence, to gain more insights into coding,
Multi-modal sensing provides some perception about the
signaling, and waveform designs for more general channels, it
location and geometry of the user devices, infrastructure nodes,
is also important to study the fundamental trade-off in terms
and various elements of the environment. This perception
of achievability and converse for these channels. Last but
can enable more efficient multiple access operation in several
not least, though RSMA-assisted ISAC has been shown to
ways. First, multi-modal sensing may enable the base stations
outperform SDMA/NOMA-assisted ISAC thanks to RSMA
and access points to intelligently select and schedule users
universality and flexibility advantages, it would be important
to maximize the multiple access system performance. For
to explore how other MA domains could be exploited for ISAC
example, leveraging the absolute or relative user position
and get a complete picture of what UMA-assisted ISAC would
information, obtained by GPS, radar, LiDAR, or camera, the
look like.
base station may schedule the users that are well-separated on
the same time-frequency resource to have less interference.
B. Multimodal Sensing-aided Communications In dense deployment, the sensing information may also be
Two key trends in current and future communication sys- utilized to assign users to base stations/access points or to
tems are (i) the use of large numbers of antennas and (ii) the proactively predict handover, which yields higher reliabil-
increasing dependence on high frequency bands. The moti- ity, less latency, and better session continuity. Beyond user
vation for using large antenna arrays ranges from increasing selection and scheduling, the underlying information cap-
the multiplexing gain, which is a main goal for sub-6GHz tured by multi-modal sensing can guide the multiple access
massive MIMO, to ensuring high the array gain, which is a resource allocation decisions. For example, in PD-NOMA,
key objective for millimeter wave (mmWave) and sub-terahertz multi-modal sensing data can potentially aid the optimization
(THz) MIMO. Beyond the base stations, the use of large of the transmit power of the different devices. Further, in
antenna arrays is also a feature of the new network nodes OMA, this sensing data may enable adaptive allocation of
such as reconfigurable intelligent surfaces and holographic the time, frequency, and spatial resources based on the users
MIMO repeaters. The other key trend is captured by the relative positions, mobility patterns, and dynamic scatterers
increasing adoption of the mmWave band as a key pillar for in the environment. All that motivates investigating the var-
 29

ious promising applications of multi-modal sensing in next- many other applications. Towards this objective, multiple
generation multiple access systems. access techniques play a central role in communicating the
Realizing the potential gains of multi-modal sensing aided data from the distributed sensors in autonomous vehicles, IoT
multiple access in practice, however, requires overcoming a devices, distributed infrastructure side units, etc., needed for
number of challenges. First, a fundamental challenge in using constructing and maintaining the digital twins. For example,
sensors such as cameras, LiDARs, and radars is identifying it is important to develop digital twin-specific multiple access
and differentiating the communication users in the sensing approaches that optimize the resource allocations among dif-
scene. For example, for a given visual scene (photo) taken by ferent sensors and modalities with the required digital twin
a camera installed at a base station, how could the base station fidelity and freshness as objectives [326]. This is an inter-
identify the mobile user from the other objects in the scene? esting and challenging research problem since these different
This task, which is defined as user identification [323], is a communication channels for the different sensors/modalities
key for enabling multi-modal sensing-aided communication in may have different QoS requirements, especially when condi-
realistic environments. Further, with multiple access, the scene tioned on the final digital twin construction objectives. This
could typically have multiple users, which makes it impor- optimization needs also to be done in a way that ensures the
tant for the communication system to differentiate between scalability of the network and the adaptability to the varying
the different users. Initial approaches for user identification digital twin requirements for the different tasks.
rely on augmenting other user-specific measurements such as On the other side, these environment digital twins could also
sparse/limited channel or beam power measurements [305], be very beneficial for the optimization of the communication
[324]. In addition to user identification and differentiation, networks and the design of multiple access solutions. In par-
the coverage mismatch between the multi-modal sensing and ticular, the digital twins provide the communication network
communication systems raises critical challenges. Overcoming with comprehensive perception, that is potentially near real-
that motivates the extension to distributed multi-modal sensing time, about its surrounding environment including the loca-
which involves the design and coordination of distributed tions of their users, the nearest stationary/dynamic scatterers,
sensors to optimize the multiple access objectives. Address- sources of interference/blockages, etc. This perception can
ing these important challenges defines interesting directions assist the various resource allocation decisions. For example,
for future research in the interplay between multiple access the digital twins could be leveraged to decide on what users to
techniques and multi-modal sensing systems. schedule in a specific MA transmission based on their relative
locations, mobility patterns, and interaction with the nearest
blockages. They can also be leveraged to develop adaptive
C. From Multimodal Sensing to Digital Twin-Assisted Com-
and site-specific MA techniques that dynamically change their
munications
parameters based on the specific site characteristics. The
Multi-modal sensing information provides the mobile de- perception provided by these digital twins can also enable the
vices and communication network with some awareness about communication networks to make proactive multiple access
what is happening in the environment. This awareness could and resource allocation decisions as they can extrapolate to
be leveraged, as discussed in Section V-B, to aid the various how the environment will change in the next few tens or
communication decisions. Building upon this and targeting a hundreds of milliseconds. This highlights the promising gains
more comprehensive perception of the communication envi- when leveraging environment digital twins to assist multiple
ronment, the concept of (near) real-time digital twins of the access in future communication systems.
physical environment has been proposed [325]. In these digital
twins, the multi-modal sensing information is fused with the
3D maps to construct real-time or near real-time maps of D. Communication-aided sensing/localization systems
the surrounding environment. Running real-time ray-tracing, Localization technologies have been embedded in mobile
which could potentially be machine learning enhanced, on communication systems for decades [327]. It is well known
these 3D maps leads to (near) real-time digital twins of the that satellite-based positioning technologies, e.g., Global Navi-
physical communication system. Using these digital twins, gation Satellite Systems (GNSS), have been popular. However,
the communication systems can pre-train ML models for the they suffer from severe degradation in indoors and urban
different communication tasks, optimize the various network areas. To address these challenges, cellular-based localization
procedures, or directly make predictions for the real system. methods have been considered due to their natural advantages
This is particularly valuable when these digital twins are of wide coverage, easy deployment, and low cost [328]. As we
sufficiently calibrated with the site-specific measurements. have seen in the previous sections, multiple access schemes are
The relationship between wireless communication and digi- capable of providing larger spectral and energy efficiencies in
tal twins is mutually beneficial. With more efficient communi- multi-user communications. Hence, under limited resources,
cation of the sensory data, more accurate and updated digital multiple access techniques can contribute to the potential im-
twins can be constructed. Further, as presented in [325], the provement of positioning accuracy and coverage for multiple
digital twins with real-time 3D maps of the environment can be users via efficient resource and interference management. In
jointly leveraged to optimize the communication networks and what follows, we review the latest multiple access schemes
to assist other applications such as autonomous driving, proac- for communication-aided localization and discuss the research
tive collision prediction, traffic management, surveillance, and challenges.
1498 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 41, NO. 5, MAY302023

1) SDMA-Assisted Localization: MIMO positioning [329]

B. Contributions
becomes popular since it provides high-angle resolution, which
In thisresults
paper,inwe propose
a more a Pareto
accurate optimization
estimation frameworkand
of angle-of-arrival
for transmission scheme optimization in the CoISAC
angle-of-departure. In [330], 3D MIMO was employed system, to
where multiple BSs cooperatively
simultaneously communicate
support high-rate with multiple
data communication and mo-
users andbile 3D positioning
directly estimate for
the multiple
target’s users in the downlink
3D location [330].
simultane-
ously overThethetransmit beamforming
same resources was optimized
through by minimizing
RSMA framework. The the
sum-localization
main contributions error
of this under
work are data rate, totalastransmit
summarized follows.power
constraints, and imperfect CSI assumption. However, individ-
• CoISAC System: We
ual localization propose
accuracy maya novel
not beCoISAC
guaranteed framework
in this case.
in C-RAN
Moreover, equipped with RSMA
[330] assumes techniques
that localization and support-
is performed during
ing data
direct localization
transmission afterservices. RSMA hasIn been
pilot transmission. [331],shown
mmWave
to achieve
MIMO system significant withadvantages
cooperative in ISAC
base systems
stations [13],
is considered
for joint communication and
[14], [15]. Cooperative RSMA in C-RAN can further localization under imperfect
CSIthe
unveil and potential
hybrid beamforming
of RSMAassumptions,
for unifiedwhere localization
communica- Fig.Fig.
1. 18.The CoISAC
RSMA system
assisted direct with RSMAwith
localization transmission
N cooperativescheme and direct
base stations
is performed together with channel
tion and sensing considering the tradeoffs between estimation during
dual thelocalization.
serving K users while estimating a target’s location [334].
pilot transmission
functionalities, whichperiod. A jointthe
also enables beamforming
direct localizationand power
allocation optimization problem for maximizing the weightedmatrix A. Re {·} and Im {·} stand for real part and imaginary
to improve the positioning accuracy.
sum rate under individual localization accuracy constraints cooperative ISAC system.
• Pareto Optimization: To explore the optimal tradeoff part operator. ⊗ and ⊙ stand for Kronecker product and
was formulated and solved. A larger rate-accuracy region was 4) Technology Outlook and Future Works: Using multiple
demonstrated in [331] compared to the conventional schemes.Element-wise
between cooperative communication and direct local- product. IL means identity matrix with L × L
access for assisting localization is a new research area. In light
ization,2) aPD-NOMA-Assisted
Pareto optimization mechanism
Localization: [10], [11]
PD-NOMA wasiscon-dimension.
of the above 1L means
works,

−1
all
a one matrix
number with L ×
of challenges dimension.
areLidentified.
adopted.
sideredDue for to the normality and compactness
a communication-localization integration of system
the ΠAFirst, = AtheAabove H
A works is the orthogonal
AH assume static users. However, matrix
projection the
CoISAC
in [332]. performance
Specifically, region, two constrained
a multi-scale PD-NOMAoptimiza- system wasonto the column space of A
impacts of user movements and synchronizationand Π⊥ A = I−Π A . ∥W∥
errors in MA- ∗ and
tionintroduced
problems, in the
i.e.,multi-cell scenario for supporting
communication-centric and continuous
radar- ∥W∥ 2 denote
assisted the nuclear
localization have not norm beenand norm respectively.
ℓ2 investigated.
fully Second,
positioning waveform and flexible
centric optimization problems, can be constructed to configurations for users the performance metric for measuring positioning accuracy is
withfor
search different
the completelocationsPareto
to obtain higher positioning
boundary, in which we accu- generally based on the II. Cramér–Rao
S YSTEM Mbound. ODEL Explicit designs of
racy and less resource consumption.
cooperatively optimize the splitting ratios, TBF and BS A convex optimization practical localization estimators to achieve this bound are still
problem of power allocation was formulated to minimize the In this section,
lacking. Third, thewe introduce
MA-assisted thelocalization
CoISAC MISO schemes network
based as
scheduling policies under various practical constraints.
error rate for communication and the ranging accuracy fordepicted in Fig. 1, where a CP is connected
on SIC could cause some decoding latency, which may not be to N BSs each
• Near-Optimal Pareto Boundary Searching: To effectively
positioning. Performance gains over the benchmark TDMA-with L antennas via high-speed and capacity-limited
desirable for delay-sensitive applications. To fully exploit the fronthaul
solve the localization
based challengingwere communication
demonstrated. and radar-centric
In [333], PD-NOMAlinks. N BSs
benefits cooperatively
of MA-assisted serve K insingle-antenna
localization wider communication users and
problems with guaranteed
was combined with mmWave performance,
MIMO for we proposetosuc-
localization furtherestimate
scenarios, the these
location of a target
challenges 1
need tosimultaneously.
be addressed inLet future
N =
cessive
improve convex approximation
the positioning (SCA)-based
accuracy, where thealgorithms
users within{1,works.
. . . , NAdditionally,
} and K = it{1,is . crucial. . , K}todenote
understandthe BS howindices
one canand
to address
one beamthe form nonconvexities
a local PD-NOMA and nonsmoothness
group. However, in theseuserexploit
indices, all MA dimensions
respectively. Letinqthe = design
[qx , qy ,ofqz ]MA-assisted
T
and bn =
the works assume problem.
optimization perfect SIC, By which
varying always be the nlocalization
maythenotperformance  beyondthe the position
SDMA/PD-NOMA/RSMA
bx , bny , bnz denote of target and theapproaches position of
case in practice.
requirement of oneSIC of also
the introduces
functions, latency
we canin find localization.
the theinvestigated so far and ultimately aim for a UMA-assisted
In addition,
n-th BS, respectively. The data to be transmitted to the K
optimal Pareto PD-NOMA by fully decoding interference does localization.
boundary approximately. users are split, precoded and designated to cooperative BSs
not achieve the optimal multiplexing gain in multi-antenna
• Performance Verification: Simulation results demonstrate
settings [10]. Hence, further improvements on the schemes, by CP, thus enabling cooperative transmission. The reflected
tremendous success of the proposed CoISAC system signals VI. M ULTIPLE ACCESS FOR E MERGING I NTELLIGENT
e.g., by using RSMA, are possible. collected at cooperative BSs are jointly processed at
compared with the SDMA, FDRC, TDRC and the A PPLICATIONS
3) RSMA-Assisted Localization: The use of RSMA to assistthe CP to estimate the target location directly. We consider
schemes
direct without
localization cooperative scheduling
was investigated under
in [334], different
where A. Semantic
multipleperfect channelCommunications
state information (CSI)2 of all users at the
system settings. It shows that the integration
base stations cooperatively communicate with multiple usersCP and of RSMA, Semantic
assume communication
cooperative BSs (SeCom),
adoptregarded
uniformasplanner
an advance- arrays
ISACandand C-RAN
directly can afulfill
estimate their
target’s 3Drespective potentials (UPA)
locations simultaneously mentinbeyond the Shannon
3 paradigm,
full-duplex mode to support ISAC technology. strives to successfully
to lead
through to athefully
RSMA updated ISACviasystem
framework efficientwith improved
interference man- transmit semantic information from the source. Unlike the
overall performance.
agement. An example with N cooperative base stations and conventional Shannon paradigm that focuses on precision in
A. Communication Signal Model
The rest Kof users
paperis is shown in Fig. 18, where the splitting of common receiving individual symbols or bits regardless of the meaning,
organized as follows. In Section II, we
and private messages are illustrated, and all base stations We SeCom greatlyRSMA
consider enhances the communication
transmission mode, where efficiency theand mes-
present the signal models and performance metrics for coop-
are connected to a central processor via high-speed links. Asage reliability by prioritizing the transmission of meaningful infor-
vk of user k is split into a common part vk and a private c
erative communication
Pareto optimization andframework
direct localization,
is formulatedrespectively.
to characterizepartmation [335]–[345]. In SeCom,
vkp . The common parts ofit is alltypical
usersfor{vboth the trans-
1 , . . . , vK } are
c c
In SectiontheIII,
sum werate
introduce
of multiple the Pareto
users andboundary to describe
the positioning error the mitter into
boundencoded and receiver to possess a shared knowledge base. This of
a common message s , while the private parts
c
CoISAC of performance
the radar target. tradeoffs and formulate
In particular, two constrainedand shared knowledge
communication-centric p penables the extraction of compact semantic
all users {v1 , . . . , vK } are encoded individually into K private
optimization problems to optimization
localization-centric search for problems
the Pareto boundary.
in their proposed information at the transmitter and the reconstruction of the
message {sp1 , · · · , spK from }. Let Pn = [p c p
. . . , patpn,K
n , pn,1 ,signal ] ∈
In Section IV, we propose
cooperative ISAC scheme two cooperative
are investigateddesign algorithms
to find the optimal original information compressed semantic the
C receiver [346]. To handle additional knowledge information,n itand
denote the precoding matrix at BS where
L×(K+1) c
to solve performance
the constrained trade-offs. It was shown
problems, that the Finally,
respectively. RSMA-assisted
the n, p
p
extensivecooperative
simulationISAC resultssystemare can achieveintheSection
provided V andper-pn,k
best trade-off are the transmit
is important to designprecoders for the common
the communication system toand private
smartly
p p T
formance among
conclusions are drawn in Section VI. various baselines, e.g., the SDMA-assisted message
synergize for user
MA k, respectively.
schemes and SeCom. Let Besides,
s = [sc considering
, s 1 , · · · , s the
K] ∈
Notation: Upper case and lower case bold face letters denote C K+1
denote the transmit messages at BS n. Then the transmit
matrices and vectors, respectively. A = Diag (α) stands 1 The extension to multi-target case is discussed in Section IV-C.
for the diagonal matrix with vector α on the diagonal and
 31

OMA SDMA

Knowledge Knowledge

Data Data User 2 as

interference
Resource block 1 User 1
User 1

Resource block 2

Base station Base station Same resource block

No interference User 1 as
interference
User 2 User 2

NOMA RSMA
Knowledge Knowledge Common message
Data Weak user 2 as Data Private message
interference

User 1
User 1

Same resource block

Base station Same resource block Base station

Successively cancel the First decode common message,

interference of strong user then private message
User 2 User 2

Fig. 19. An illustration of SeCom with multiple MA schemes.

process of SeCom, proper MA design can greatly enhance the of multiplying the reciprocal of the semantic compression ratio
performance of SeCom. is that one bit of obtained semantic information can convey
more than one bit in the original information. In the second
SeCom typically involves AI-related design. Unlike con-
aspect, the SINR expression is modified by reducing the com-
ventional communication systems, it usually lacks a well-
munication power. This adjustment is necessary as some power
defined performance metric, such as the Shannon capacity.
needs to be allocated for the computation involved in semantic
To determine the appropriate MA for SeCom, it is important
information extraction. For a fair comparison between the
to identify the key performance indicator (KPI) for SeCom.
conventional communication and SeCom, it is essential to
Broadly, there are two main kinds of KPI, i.e., simulation-
maintain a fixed total power including both communication
based KPI and analysis-based KPI. In simulation-based KPI,
and computation power. The precise SINR expression in in
the semantic rate expression is regraded as a function of
SeCom is therefore obtained by theoretically demonstrating
SINR. This function is obtained through data fitting in the
that the computation power for semantic information extrac-
simulation results about semantic rate versus SINR [347]–
tion follows a linear segment function concerning the semantic
[349]. In analysis-based KPI for SeCom, one involves utilizing
compression ratio. Furthermore, the communication power
the probability graph shared by both the transmitter and the
in semantic communication can be derived by reducing the
receiver, referred as the knowledge base [350], [351]. The
computation power within the total power budget.
transmitter extracts the small-size semantic information based
on the probability graph, while the receiver recovers the Based on the aforementioned KPIs, various existing works
original information based on the shared probability graph. delve into the MA design for SeCom. In general, the cur-
By employing the probability graph [350], the semantic rate rect MA schemes for SeCom encompass OMA, PD-NOMA,
is derived by modifying the Shannon capacity expression in SDMA, RSMA, and some customized MA schemes, which
the following two aspects. In the first aspect, the reciprocal are summarized as follows:
of the semantic compression ratio is multiplied in front of 1) OMA-aided SeCom: In OMA-aided SeCom, the base
the Shannon capacity expression. Here, semantic compression station uses orthogonal resources to transmit both common
ratio is defined as the ratio of the bits of the extracted semantic knowledge and data information to different users as shown in
information to the bits of the original information. The reason Fig. 19. For OMA-aided SeCom with simulation-based KPI,
 32

the spectral efficiency is maximized through optimizing the quality of experience (QoE) for image transmission in the
channel assignment and number of semantic symbols [347], RSMA-aided SeCom system. In RSMA-aided SeCom, the
[352]. Additionally, to address the energy efficiency issues, common knowledge is transmitted in multicast, while the
[353], [354] explores the optimization of total energy con- individual data are transmitted in unicast. Due to the flexible
sumption under latency constraints. Combining with mobile combination of multicast and unicast, the system performance
edge computing (MEC), the energy consumption is minimized in RSMA-aided SeCom is superior over NOMA-aided SeCom
through jointly optimizing the semantic-aware division factor, and SDMA-aided SeCom.
as well as communication and computation resource man- 5) MDMA-aided SeCom: A novel MA called model di-
agement [355], [356]. For OMA-aided SeCom with analysis- vision multiple access (MDMA) is proposed in [366]. This
based KPI [350], [357], the total energy of the whole system innovative approach involves simultaneously serving multiple
can be optimized through optimizing the semantic compres- users with distinct model information space resources. In
sion ratio. The advantage of OMA-aided SeCom is easy to MDMA-aided SeCom, through analysing the resource of the
deploy, while the disadvantage is the number of served users semantic domain in high-dimensional semantic space, the
is limited by the wirelress resource. information of different users can be recovered in the model
2) PD-NOMA-aided SeCom: In PD-NOMA-aided SeCom, domain [366]. It is shown that MDMA-aided SeCom can
the base station can simultaneously transmit the semantic improve the performance over traditional FDMA and NOMA.
information to different users as shown in Fig. 19. Moreover, Due to the superior performance of MDMA, MDMA-aided
as illustrated in [348], [358], PD-NOMA can support the SeCom is applied in a point cloud based SeCom system to
coexistence of conventional bit-interested user and semantic solve the three-dimensional representation problem [367]. The
information-interested user. In particular, the far-user exploits realization of MDMA requires learning-based design and the
the SeCom, while the near-user employs the conventional PD- authors in [368] utilize deep learning-based multiple access
NOMA-aided communication [358]. Furthermore, a unified (DeepMA) method by training SeCom models to construct
framework for NOMA-aided SeCom with general datasets MDMA-aided SeCom.
and data modalities is presented in [359]. The PD-NOMA- 6) Technology Outlook and Future Works: In general,
aided SeCom system achieves the system performance gain OMA can be easily applied to SeCom and satisfy the require-
through jointly considering both conventional and SeCom ments of knowledge information transmission and semantic
transmissions. computation. By leveraging multiple antennas, SDMA-aided
3) SDMA-aided SeCom: In SDMA-aided SeCom, the se- SeCom can be used to further enhance the system perfor-
mantic information can be simultaneously transmitted to mul- mance. For the weak user with low transmission rate, PD-
tiple users through the strategic design of precoding in the NOMA-aided SeCom can be flexibly applied. Recognizing
multiple-antenna assisted system as shown in Fig. 19. To the inefficiency of knowledge transmission in PD-NOMA
solve the semantic performance optimization problem for and the superior performance gain of RSMA over OMA,
SDMA-aided SeCom, the generalized power iteration pre- SDMA, and PD-NOMA, RSMA-aided SeCom is shown to be
coding algorithm is applied in [360]. Considering the signal more effective in balancing the interference during knowledge
transmission characteristic of wireless communication, the information and data information transmissions. Moreover,
multimodal SDMA-aided SeCom system is investigated in leveraging the unique properties of semantic space, new MA
[361]. In SDMA-aided SeCom, the precoding scheme enables schemes such as MDMA can be tailored for SeCom. Future
efficient semantic information transmission for multiple users directions of MA schemes for SeCom should explore the
through exploiting the space domain characteristics. integration of MA design in semantic information extraction
4) RSMA-aided SeCom: In RSMA-aided SeCom, the and resource allocation for emerging MA schemes including
knowledge intended to multiple users can be encoded into NOMA-aided SeCom, RSMA-aided SeCom, MDMA-aided
the common stream, while the individual data intended to SeCom, and ultimately UMA-aided SeCom to exploit all MA
a specific user is encoded into the private stream [362], as dimensions.
shown in Fig. 19. In NOMA-aided SeCom and SDMA-aided
SeCom, the common knowledge for each user is transmitted
and received in a unicast way, even though the common knowl- B. Metaverse
edge is the same for all users. In RSMA-aided SeCom, the The Metaverse is considered to be the new generation of
common knowledge is transmitted in a multicast way. Using the Internet, aiming to build a digital world where people can
multicast, the resource allocated for common knowledge can meet and interact in real-time by integrating various emerging
be greatly reduced in RSMA-aided SeCom than NOMA-aided technologies, such as digital twin (DT), extended reality (XR),
SeCom and SDMA-aided SeCom. As a result, the overall and holographic [369]–[372]. In particular, the physical de-
performance of RSMA-aided SeCom outperforms NOMA- vices in the physical world transmit the sensed information to
aided SeCom and SDMA-aided SeCom [362]. Taking into the mobile edge servers, which partial/completely compute the
account the URLLC, the weighted sum semantic information received information [373], [374]. Then, mobile edge servers
transmission rate maximization problem is formulated in [363] transmit the computed information to the cloud server for
through jointly optimizing the semantic information extraction, further inference. In Metaverse, the information of the digital
delivery duration, rate splitting, and transmit beamforming. world is updated in the cloud server to meet the real condition
Besides, the authors in [364], [365] focus on optimizing the of the physical world. Besides, the simulated information of
 33

the digital world in Metaverse including the communication in particular for high-mobility scenarios.
system schedules and physical devices configurations can be 5) Technology Outlook and Future Works: The Metaverse
transmitted to the physical devices through communication applications require high-data transmission, immerse experi-
system. To support the immerse and intelligent-related re- ences, and intelligent services. To meet the unique characteris-
quirements of emerging applications in Metaverse including tics of Metaverse, MA schemes can be applied to support mas-
education, military, healthcare, real estate, and manufacturing, sive connectivity, enhance low-latency immerse experience,
MA schemes must be considered [375]. Existing MA schemes high spectral efficiency, and robustness. Comparing different
for the Metaverse applications include OMA, NOMA, SDMA, MA schemes for Metaverse, NOMA can support more physi-
and RSMA, which are respectively specified as follows. cal devices than OMA at the cost of additional decoding com-
1) OMA-aided Metaverse: In OMA-aided Metaverse, users putation complexity. RSMA-aided Metaverse is more suited
are supported with orthogonal resource. To support the im- to the Metaverse applications than other MA schemes espe-
merse applications in Metaverse, a blockchain based OMA cially for high-mobility scenarios. Future directions of MA-
scheme is proposed for ubiquitous access controls [376]. For aided Metaverse include the cross-layer design of considering
VR applications in the Metaverse, a learning-based system both the physical-layer OMA/NOMA/SDMA/RSMA resource
can be applied [377]. Further considering the security in the allocation and the application-layer scheduling, the design
Metaverse [378], the covert communication technique is used of MA schemes that account for considering multiple time
for OMA-aided Metaverse. slots in the Metaverse, as well as MA schemes that consider
2) NOMA-aided Metaverse: Considering multiple physical both low-latency and prediction accuracy in the digital world.
devices and limited wireless resource, NOMA-aided Meta- Ultimately, the goal is to design UMA-assisted metaverse.
verse can serve more physical devices than OMA-aided Meta-
verse. The QoE model for Metaverse application is explored C. Virtual Reality
in [379] by incorporating factors such as the virtual distance In VR video streaming, an immersive viewing experience
and network effect. To maximize the QoE utility within the is provided for users by rendering 360-degree videos through
Metaverse system, a two-step joint resource allocation for their head-mounted displays (HMDs). To accommodate the
NOMA user channel allocation and Metaverse service selec- growing demand for VR streaming and the increasing num-
tion scheme is proposed. In the first step, the resource alloca- ber of VR users, different challenges should be addressed
tion for NOMA is determined by the one-to-many matching in wireless systems. The bit rate of the 360-degree videos
game, while a hedonic coalition formation game in the second transmitted in VR streaming is much higher than that of
step is applied to solve the Metaverse service selection. The conventional two-dimensional (2D) videos. In VR streaming,
two-step scheme is verified to improve the average QoE utility users may experience motion sickness due to a motion-to-
of the system. In NOMA-aided Metaverse, the massive access photon delay (from the user’s head movement to the time at
problem in Metaverse can be solved through serving multiple which the image gets updated on the VR screen) larger than
users with the same time and frequency resource. 20 milliseconds (ms). This delay occurs when a user turns
3) SDMA-aided Metaverse: In SDMA-aided Metaverse, the its head to watch another part of the 360-degree video frame,
beamforming technique is utilized to support the communi- but that part has not been rendered for the user yet [382].
cation between the physical devices in the physical world To meet the required low latency and high bit rate for VR
and the digitial devices in the Metaverse. Considering the video streaming, different components of wireless systems,
advantages of RIS, the RIS-assisted Metaverse system is such as radio resource allocation algorithms, MA schemes,
explored in [380]. Given the substantial communication rate and network infrastructure, must be optimized and integrated.
demands in the Metaverse, mmWave and THz communication Three MA schemes are widely used in the literature for VR
systems are required. To solve the beam prediction problem video streaming, namely OMA, NOMA, and RSMA.
introduced by the high-frequency system, the authors in [381] 1) OMA-aided VR: Both TDMA and OFDMA have been
propose a multimodal deep learning framework based on three- used as OMA in VR video streaming. OMA scheme provides
dimensional convolutional transformers using both optical and VR users with a simple receiver, which can subsequently lead
rad data. In SDMA-aided Metaverse, the huge communication to lower weight HMDs and extended battery life. A number
rate demand can be solved through proper beamforming design of works have focused on TDMA and OFDMA VR systems
in the mmWave and THz communication systems. discussed in the sequel.
4) RSMA-aided Metaverse: Through dynamically treating TDMA was investigated in [383] for a 360-degree VR video
the power of power of interference and noise, RSMA-aided streaming with multicast opportunities. The authors aimed at
Metaverse is a promising way to flexibly handle the dynamic minimizing energy consumption by optimizing transmission
channel conditions between the physical world and digitial resource allocation and quality level selection. In [384] a
world. In uplink RSMA-aided Metaverse, the multi-user ca- downlink TDMA system for multicast transmission of VR
pacity region can be approximated with lower complexity than video tiles from a server to multiple users is considered. A
the conventional scheme such as dirty paper coding, which globally optimal closed-form solutions for two formulated
is promising in the systems with a large number of users. non-convex optimization problems is provided: one aimed at
In downlink RSMA-aided Metaverse, thanks to robustness of minimizing average transmission energy while maintaining
RSMA towards CSIT inaccuracy and user mobility, RSMA video quality, and the other focused on maximizing received
achieved by RSMA can well support Metaverse applications video quality within a given transmission energy budget.
 34

Metaverse
(Digital world)

OMA SDMA NOMA RSMA

Space Space Space Space

Frequent Frequency Frequency
Frequency Frequency
All users
information All users

exchange User 1
with MA …
User 1
User 2
User 3
User 4

User 4

Time Time Time Time

Only individual message Private + common message
/message from two virtual users

Physical world

Fig. 20. An illustration of Metaverse over communication systems with multiple MA schemes.

OFDMA was studied in [385] for 360-degree VR video and the live video is streamed from the MEC to users using
streaming across wireless networks. A multi-group multi- OFDMA. The authors aimed to maximize the QoS of the
cast optimization problem was formulated aiming at finding live video for the users while maintaining an acceptable
the optimal beamforming, subcarrier allocation, transmission QoS for other existing traffic types in the network. In [390],
power, and rate allocation. In [386], the resource allocation OFDMA for VR video streaming was used in a UAV-aided
problem in a buffer-aware OFDMA VR streaming network MEC wireless network, which includes one macro base station
was formulated as a stochastic game, where users bid for their and multiple UAVs. The authors considered that each VR
desired data rates to enhance their VR viewing experience. The video comprises a base layer and several enhancement layers,
authors proposed a DRL algorithm for each user to solve this and aimed to maximize the users’ QoE by optimizing the
problem in a distributed manner. In [387] OFDMA was used UAV placement, VR video layer assignment, and MEC and
in a DT-enabled wireless VR system. The authors formulated radio resource allocation. In [391], the authors considered live
an optimization problem to maximize the max-min QoE and multicast streaming of 360-degree videos to mobile users by
ensure fairness of the resource allocation among the users. allocating the available resource blocks in an OFDMA-based
In [388], a VR video streaming using cooperative multicast wireless network. The authors first partition the users into mul-
and unicast in an OFDMA-based heterogeneous cloud-radio ticast groups and allocate the available resource blocks among
access network (H-CRAN) was studied, consisting of one them. Once the resource blocks are allocated to each group,
macro base station and several remote radio heads (RRHs). the users’ QoE is maximized by solving a tile quality selection
The authors formulated a mixed-integer nonlinear problem to problem. In the aforementioned works, each orthogonal radio
maximize users’ QoE when the macro base station transmits resource (i.e., time slot or time-frequency resource block) is
a specific version of the VR video to all users via multicast, allocated to only one VR user for unicast transmission. Thus,
and the RRHs transmit enhanced versions of tiles based the number of VR users that can be supported simultaneously
on the predicted viewport to their corresponding users. The is restricted by the number of available orthogonal radio
authors in [389] proposed a ML-based scheduling and resource resources.
allocation algorithm when an unmanned aerial vehicle (UAV)
records live events with its 360-degree spherical camera, sends 2) PD-NOMA-aided VR: In NOMA, VR video tiles can be
the recorded video to a mobile edge computing (MEC) server, transmitted to multiple users using the same radio resource.
Thus, compared to the OMA scheme, NOMA can support a
 35

greater number of VR users and has been the focus of several Common and private
messages Downlink
User 1’s FoV Data retrieved by User 1
works discussed next. transmission
m1 m0 m1
In [392], a downlink PD-NOMA system is considered for
Overlapping User 1
transmitting VR frames from a Metaverse server to multiple FoV
m0 Data retrieved by User 2
VR users. The authors proposed a multi-agent DRL algorithm m2 Downlink m0 m2
called user-centric critic with heterogeneous actors (UCHA) User 2’s FoV transmission
User 2
for optimizing channel access and downlink power allocation.
m0 Common message
In [393], PD-NOMA is used to deliver VR videos in an
edge-enhanced wireless network, employing a multi-agent m1 Private message for user 1
DRL algorithm as a cooperative caching scheme in edge m2 Private message for user 2
base stations. To enhance content delivery efficiency, the
authors considered a combination of unicast and multicast, and Fig. 21. An example of using RSMA for VR applications. Users 1 and 2
proposed a base station-multicast group matching algorithm. have overlapping field of view (FoV). The video tiles with overlapping FoV
are sent via the common messages. The video tiles with non-overlapping FoV
In [394], a wireless network consisting of a base station, are sent via the private messages.
UAVs, and VR users is considered. In this network, PD-
NOMA is employed to transmit video tiles from the base
station to UAVs, and also from the UAVs to VR users. The parameters such as the rates of common and private messages
authors aimed to optimize the selection of tiling patterns at in RSMA, as well as transmission beamforming vectors.
the base station, the grouping of VR users and UAVs, as well In [399], the users in a 360-degree video streaming system
as the allocation of computational resources and the duration are clustered into different groups based on their overlapped
of NOMA transmissions. In [395], authors consider a UAV- field-of-view (FoV). Then, to effectively manage interference
assisted cellular network in which PD-NOMA is employed among the users, the authors adopted a hierarchical RSMA
for content delivery from the UAVs to users. In particular, approach in which the messages are split and encoded into
two users that request the same content for their AR and three streams: super common stream, group common stream,
normal multimedia applications are grouped for PD-NOMA and private stream. All those works demonstrate the benefits
transmission. In this work, the authors aimed to minimize of RSMA over other MA schemes, such as SDMA and PD-
the content delivery delay by optimizing the user association, NOMA, in VR applications.
power allocation of PD-NOMA, and UAVs deployment. Since 4) Technology Outlook and Future Works: Some potential
the PD-NOMA scheme manages the multi-user interference research challenges in MA design for VR streaming sys-
with SIC, the receiver complexity is increased as the number tems include designing a low-complexity receiver for VR
of VR users increases. Additionally, the grouping of VR users users while simultaneously preserving the VR users’ privacy
and the decoding order play a crucial role in the performance and considering the HMD characteristics. For example, it
of the PD-NOMA scheme. Furthermore, due to the use of is important to consider that HMDs have limited battery
SIC, the VR users’ physical layer privacy may be at risk in and computational resources. HMDs should be lightweight.
a PD-NOMA system, necessitating additional privacy protec- Additionally, they may have multiple antennas, which raises
tion measures. Therefore, most of the existing works in the the challenge of designing RSMA for receivers with multiple
literature have focused on utilizing PD-NOMA for VR video antennas.
multicasting. However, finding the optimal user grouping is a For cache-enabled wireless networks, placing VR content
challenging problem because it requires optimizing the number in the cache requires a larger timescale than delivering VR
of groups and the number of users within each group. content using an MA scheme. To enhance VR video streaming
3) RSMA-aided VR: RSMA manages multi-user interfer- performance, content placement and delivery should be jointly
ence by decoding a part of the interference, considering it optimized, leading to a multi-timescale optimization problem.
as the common message of VR users, while treating the In UAV-assisted cellular networks, the deployment and tra-
remaining part as noise, representing the private messages jectory of UAVs should be designed according to the mobility
of other VR users. An example of using RSMA for VR of VR users. This design is essential to manage interference
applications is shown in Fig. 21. RSMA has been the focus in NOMA and RSMA schemes and to prevent degradation in
of recent VR studies. In [396], the authors proposed iterative VR streaming performance.
algorithms to optimize beamforming vectors and rates of In IRS-aided VR streaming systems, it is important to
sub-messages in a multicast RSMA VR streaming system. consider the impact of imperfectly estimated CSI for both the
In [397], RSMA is employed in an reconfigurable intelligent IRS-user and BS-IRS links on the performance of the MA
surface-aided VR streaming system. The authors proposed a schemes.
deep deterministic policy gradient (DDPG) algorithm with In VR video streaming, particularly in the case of live
imitation learning to jointly optimize beamforming vectors, VR streaming, multiple users may request the same part of
IRS phase shifts, rate-splitting parameters, and video tile a 360-degree video frame due to their shared interests. This
bit rate selection. In [398], RSMA is exploited for adaptive correlated data presents the opportunity to design a more
streaming of tiled 360-degree videos from a multi-antenna efficient RSMA for VR streaming, in contrast to typical
base station to multiple single antenna users. The authors scenarios in which the data requested by users are independent
utilized the concave-convex procedure (CCCP) to optimize and uncorrelated. However, RSMA should also perform well
 36

when the requested tiles by users do not overlap and the users In general, RIS can be classified into D-RIS and BD-RIS
watch distinct parts of the video frames. Therefore, the RSMA depending on the nature of the scattering matrix. Conventional
parameters need to be optimized, considering all possible cases D-RIS is typically modeled as a diagonal phase shift matrix,
in VR video streaming systems. limiting its functionality to a reflection mode where incident
The integration of RSMA with other MA schemes, such as signals are exclusively reflected from one side of RIS. In
OFDMA, has not yet been investigated to determine whether contrast, BD-RISs, with beyond diagonal scattering matrices,
RSMA performance can be further enhanced in VR video are capable of creating group/fully-connected reconfigurable
streaming systems. This approach provides greater flexibility, impedance networks based on the same reflection model
especially when dealing with multigroup multicasting prob- as D-RIS. Moreover, to achieve full-space coverage, hybrid
lems, by pairing a group of users and serving them on a and multi-sector modes are proposed for single/group/fully-
designated resource block using RSMA. Beyond OFDMA- connected BD-RIS [402]. The hybrid mode enables partial
RSMA, integration of CD-NOMA and RSMA would also be reflection and transmission of incident signals to different sides
of interest to address the need for bit rate in VR. of RIS, which is also known as intelligent omni-surface (IOS)
In VR streaming systems, certain information, such as users’ or simultaneous transmitting and reflecting RIS (STAR-RIS)
head movements, needs to be transmitted from the users to the [404]. The multi-sector mode extends the hybrid model by
server. Using RSMA, users can share the uplink channel to dividing the full space into sectors, allowing incident signals
send their information simultaneously. Furthermore, federated from one sector to be partially reflected and scattered to
learning can be applied within a VR streaming system to different sectors.
obtain a viewport prediction model for the users. In particular, Thanks to its powerful capability for controlling wireless
federated learning enables the users to train a shared viewport propagation environments and manipulating signals spatially,
prediction model, while their privacy is preserved without considerable research efforts have been dedicated to RIS-
sharing their historical head movement data [386], [400]. assisted MA schemes, such as OMA, SDMA, NOMA, and
RSMA can be incorporated in federated learning to facilitate RSMA.
model aggregation without compromising model accuracy. 1) RIS-assisted OMA: The deployment of RISs was ini-
Higher frequency bands, such as mmWave and THz, can tially studied and implemented in OMA, focusing on point-
be employed in VR streaming systems to fulfill the bitrate to-point communication with a single or multiple resource
requirements of high-resolution 360-degree videos. The use blocks [405]–[419]. In this simplified context, the primary
of RSMA in combination with beamforming design in these research focus involved passive beamforming design [405]–
high-frequency bands represents an interesting research area. [408], channel modeling [409], channel estimation [407],
Moreover, communication in mmWave and THz bands faces [410], [411], theoretical limits analysis [405], [412], [413], and
high path loss and blocking issues. It would be interesting to prototype development [414]–[419]. Despite the introduction
explore how RSMA can be integrated with cooperative com- of additional challenges related to channel estimation and pas-
munication to tackle these challenges in VR video streaming sive beamforming design, RIS-assisted OMA (OMA) shows
systems. significant potential to cost-effectively enhance the spectrum
6G wireless systems will provide various services, including efficiency and extend coverage without requiring additional
an evolved version of eMBB, known as eMBB+. The high energy consumption.
data rate and low latency requirements for VR video streaming 2) RIS-assisted SDMA: In multi-user scenarios, the most
position it as one of the key eMBB+ services provided by 6G commonly considered MA is SDMA. It is worth noting
systems. Whether RSMA is a promising MA scheme to meet that in RIS-assisted SDMA, maximizing the channel gain
the hybrid requirements of VR users remains unknown and a is not the only objective due to the introduction of multi-
potential research direction. user interference. To achieve objectives, such as maximizing
spectral or energy efficiency and minimizing energy con-
sumption, transceiver design in RIS-assisted SDMA typically
D. Smart Radio and Reconfigurable Intelligent Surfaces involves the joint optimization of passive beamforming at the
Reconfigurable intelligent surface (RIS) has garnered sig- RIS and active beamforming at the transmitter that require
nificant recognition as a revolutionary technique for 6G. By tailored optimization methods and channel estimation meth-
incorporating multiple passive reconfigurable elements with ods. Numerous optimization techniques have been proposed
tunable amplitudes and phases, RIS is capable of effectively in existing literature addressing joint beamforming design
manipulating the direction and strength of the scattered signals, with continuous or discrete D-RIS phase shifts [420], or for
thereby exerting a transformative impact on the propagation BD-RIS [401], [403], [421]. Concerning channel estimation,
environment. Recent advances on RIS has witnessed its evo- multi-user channel estimation in RIS-assisted SDMA brings
lution from convention diagonal RIS (D-RIS)–where each additional research challenges due to the presence of inter-
element is independently controlled by a tunable impedance user interference. Existing literature has focused on advanced
connected to ground, resulting in a diagonal scattering matrix– channel estimation techniques tailored to unstructured chan-
to the generalized beyond diagonal RIS (BD-RIS) [401], [402]. nels, which model low-frequency rich-scattering scenarios,
BD-RIS facilitates interconnections among elements, leading as well as structured channels designed for high-frequency
to a beyond diagonal scattering matrix. Figure 22 illustrates sparse channels [420], [422]. In terms of prototyping, most
the state-of-the-art RIS architectures as outlined in [403]. of existing research on RIS prototypes considered a single-
 37

Fig. 22. Illustration of the family of RISs.

user case. The design of RIS-aided multi-user prototypes is 3) RIS-assisted PD-NOMA: In conventional PD-NOMA
still in its early stages. Recent work [423] introduced a novel without RIS, the efficacy of SIC decoding relies on a substan-
blind beamforming strategy for RIS-assisted SDMA, which tial channel strength disparity between users. In RIS-assisted
designs the phase shifts without relying on channel informa- PD-NOMA, the aid of an additional reflection/transmission
tion. Real-world trails have been conducted to demonstrate link helps to enhance the channel strength disparity, thereby
the effectiveness of this strategy. In [424], a new prototype augmenting the benefits of SIC decoding. However, RIS intro-
that incorporates both channel estimation and beamforming duces challenges in the resource allocation as the optimal SIC
design is proposed. Experimental results highlight a notable decoding order and user grouping depends on both the active
enhancement in spectrum efficiency with the proposed scheme, beamforming at the transmitter and the passive beamforming
marking a significant advancement of RIS-assisted SDMA at the RIS. The interplay between NOMA and RISs have
prototypes. been summarized in [20]. The primary research focus in RIS-
 38

assisted NOMA is on addressing the above challenges. Be- to more efficient energy utilization for all MA schemes.
sides, RIS offers significant advantages in addressing security • Improved coverage: RISs introduce a new paradigm
concerns within conventional NOMA by enhancing wireless for enhancing coverage in wireless networks. Instead of
communication covertness at the physical layer [425]. An deploying additional energy-consuming infrastructure for
interesting observation from [426] is that in the presence of relaying and forwarding signals, RISs adaptively redirect
RIS, NOMA is not always a better option than OMA. Indeed signals to navigate obstacles like trees and buildings,
the minimum power needed to achieve a target rate may be thereby extending coverage for all MA schemes.
lower for TDMA than NOMA, i.e. NOMA may perform worse • Low complexity: RISs contribute to reduce the computa-
than TDMA. This contrasts with the non-RIS deployment tional and hardware complexity for various MA schemes.
where PD-NOMA is superior to TDMA. In terms of computational complexity, the ability of RIS
4) RIS-assisted RSMA: The potential advantages of RSMA to generate narrow beams and mitigate multi-user inter-
and RIS have sparked interest in their combined use in recent ference allows for simplified signal processing techniques
years [427], [428]. The synergy between D-RIS and RSMA at the transceivers for all MA schemes. As demonstrated
has been widely studied, and revealing substantial performance in [421], the transmitter employing classical regularized
advantages over RIS-aided SDMA, NOMA, OMA in terms zero-forcing (RZF) precoding can attain comparable per-
of spectrum efficiency [428]–[433], energy efficiency [434], formance to a scenario where the transmitter utilizes
[435], user fairness [436], [437], outage probability [438], optimized precoders (based on fractional programming),
[439] etc. Recent works have also delved into the interplay provided that the number of RIS elements is sufficiently
between BD-RIS and RSMA, covering scenarios such as large. From the hardware complexity perspective, as
STAR-RIS aided RSMA [440], [441], RSMA with group/fully RIS can enhance the spectral efficiency, it is capable
connected RIS in the reflection mode [442], [443], and RSMA of reducing the number of antennas at the transceivers
with multi-sector BD-RIS [444]. BD-RIS aided RSMA has without influencing system performance. Moreover, for
demonstrated significant advantages in extending coverage, RSMA, the integration of RIS empowers the use of sim-
further boosting both spectrum and energy efficiency. Given plified schemes, i.e., 1-layer RS, to achieve performance
the absence of an RF chain at RISs, channel estimation is comparable to HRS/GRS without RIS. Therefore, RIS
a widely recognized challenge in RIS-assisted links. Con- contributes to reduce the receiver complexity [445].
sidering the resilience of RSMA to CSI inaccuracies and
user mobility as shown in many existing works, RIS-assisted
E. Internet-of-Things Networks, Massive Connectivity, and
RSMA emerges as a promising new paradigm to effectively
Random Access
address and compensate for the channel estimation limitations
inherent in RIS [428]. RIS-assisted RSMA therefore offers For IoT communication network, the key is how to co-
a mutually beneficial solution to both RSMA and RIS. This ordinate the transmission of small packets between the base
advantage is distinctive to RIS-assisted RSMA, as RSMA station and a huge number of IoT devices in an reliable and
uniquely demonstrates robustness to CSI, which is not shared low-latency manner. This is in sharp contrast to the human-
by other conventional MA schemes. type communication, which targets at high-speed transmission
5) Technology Outlook and Future Works: RIS can signifi- to/from a medium number of users. Such a difference poses
cantly improve the performance of various MA schemes from new challenges for the MA design under IoT applications.
the following perspectives: Specifically, if the number of users is huge, then access
collision is more likely to occur, which leads to long delay.
• Enhanced spectrum efficiency: By dynamically adjust-
In IoT network, it is of paramount importance to design
ing the phase and amplitude of incident signals, RISs pro-
innovative MA schemes that can greatly reduce the access
vide narrow beams directed towards users. This versatile
delay so as to facilitate the small-packet transmission from a
capability brings about various advantages in different
massive number of IoT devices. Generally speaking, in the
MA schemes:
literature, there are two solutions for MA in IoT network: the
– OMA: enhances the signal strength, thereby improve grant-based MA technique and the grant-free MA technique,
the spectrum efficiency. as shown in Fig. 23.
– SDMA: effectively redirects intended signals to their Grant-based MA technique are conventional for eMBB-like
respective users and mitigates multi-user interfer- services and relies on the base station to grant users the access
ence, thereby enhancing spectrum efficiency. of the radio spectrum for data transmission.
– NOMA: creates an accurate beam towards the serv- In grant-free MA, or also called grant-free random access
ing users and enhances the channel strength disparity (RA), the devices can access the channel without any prior
between the users, and thereby enhancing the spec- resource requests [89], [446]. In addition, the set of active
trum efficiency. users is unknown to the receiver. Grant-free RA is suitable
– RSMA: unifies the benefits for OMA, SDMA and for supporting a massive number of devices with sporadic
NOMA, leading to higher spectrum efficiency. traffic, e.g., mMTC as in IoT, and can be divided into two
• Enhanced energy efficiency: RISs employ beamforming different paradigms, namely sourced and unsourced RA [447],
techniques to concentrate signals on specific users. This [448]. For sourced RA, the receiver is interested in both
avoids spreading energy in multiple directions and leads messages and user identities. Since the active device detection
 39

IoT User Base Station

1. Random Preamble

2. Random Access Response

3. Connection Request
Metadata
Data
4. Contention Resolution (Preamble)

(a) Grant-based Random Access (b) Grant-Free Random Access

Fig. 23. Grant-based versus grant-free random access schemes.

for sourced random access is a typical sparse signal processing 2) Grant-Free Sourced Random Access based on Com-
problem, numerous compressed sensing (CS) and AI-based pressed Sensing: Second, one may ask whether it is possible
approaches have been employed to handle the detection and to reduce the collision probability, rather than resolving the
user identification problems [449]–[453]. For unsourced RA, collision, under IoT networks. The goal can be achieved via
the receiver is interested in the transmitted messages only (user the grant-free random access scheme [446]. Under the grant-
identity is recovered at the higher layers of the communication based random access scheme, collision arises from the case
protocol rather than the physical layer). when multiple users select the same preamble for transmission.
To mitigate collision, a natural idea is that each user is
In the sequel, we discuss grant-based MA, grant-free
allocated with a fixed preamble, while the preambles allocated
sourced random access (based on compressed sensing and AI),
to different users are different. In this case, the preamble of
and grant-free unsourced random access.
each user serves as the identification of this user. In other
1) Grant-based Random Access: First, with human-type words, after receiving the preambles sent from the active
communication, the contention-based grant-based access IoT devices, the base station can detect which preambles are
schemes, e.g., ALOHA, are widely used. Under this scheme, received, and then know which users are active. This scheme is
there is a preamble pool consisting of lots of orthogonal called grant-free scheme because the base station just detects
preambles. When a user becomes active, it will randomly which users are active, instead of sending grants to permit
select a preamble from the poor. If an active users selects a some users to be active. However, detecting which preambles
unique preamble that is not selected by other active users, are received is quite challenging in massive IoT connectivity,
the base station will grant this user the access for data because it is not possible to assign orthogonal preambles
transmission. Otherwise, if some active users select the same to a huge number of users. Recently, a breakthrough was
preambles, the base station will detect their access collision made to tackle the above challenge arising from grant-based
and not grant them the access for data transmission. Despite random access based on the compressed sensing technique.
of their simplicity, such schemes do work quite well in current Specifically, although there is a huge number of IoT devices,
cellular network, where the number of users is moderate. most of them are in the sleeping model to save the energy at
Therefore, it is not surprised that some efforts have been each time slot. Thanks to this activity sparsity, it was shown
paid in the literature to investigate how to modify the existing in [449] that detecting the active devices, i.e., deciding which
grant-based access schemes according to the requirements of preambles are received, can be cast into a compressed sensing
IoT applications. Here, the main issue to directly apply grant- problem. More importantly, based on the state evolution,
based schemes in IoT network lies in delay - when there is a [449] rigorously proved that in the asymptotic massive MIMO
huge number of IoT devices to compete for the transmission regime where the number of antennas at the base station goes
grant from the base station, the possibility for collision is very to infinity, the activity detection error probability can go down
high. As a result, in the literature, some works have been done to zero when the approximate message passing algorithm is
to propose efficient methods for resolving the collisions. For used. This result theoretically justifies the effectiveness of the
example, in [454], a strongest-user collision resolution scheme compressed sensing based grant-free random access scheme
was proposed for grant-free random access in IoT network. for IoT networks.
The idea is that after a collision occurs, each colliding user will 3) Grant-Free Sourced Random Access based on AI: In
make a decision about whether it is with the strongest channel massive IoT connectivity setup, the number of IoT devices
among all the colliding users. If a colliding user believes that is huge. As a result, low-complexity multiple access design
its channel is the strongest, it will send the access request to suitable for large-scale IoT network is appealing. In this
the base station. Otherwise, it will be quiet for some time and sense, recently, AI-empowered multiple access schemes have
re-compete for the access. It was shown in [454] that under the attracted more and more attention in the area of massive
massive MIMO regime, the strongest-user collision resolution machine-type communications.
scheme can resolve most of the access collisions in massive On one hand, the performance of compressed sensing algo-
connectivity. rithms introduced in the above heavily depend on the choice
 40

of the sensing matrix, while most of the works on compressed they become active, they transmit their preambles at the same
sensing based device activity detection assume that the sensing time. In practice, it is impossible to perfect synchronize a huge
matrix, which is determined by the pilot signals assigned number of low-cost IoT devices. In this case, how to make
to IoT devices, is given. For example, many works assume the strongest-user collision resolution scheme work under the
Gaussian pilots [449], [455], [456] such that the approximate grant-based random access scheme? How to formulate the
message passing (AMP) algorithm can be adopted to detect the device activity detection and synchronization error estimation
active devices and analyze the detection performance based on problem as a compressed sensing problem under the grant-
the state evolution. Moreover, Reed-muller sequences based based random access scheme? Whether the sample covariance
pilots have been utilized for device activity detection in [457]. matrix of the received sequence is still a sufficient statistic
However, how to design the pilot signals that work the best for device activity detection? All these questions should be
for device activity detection is still open. This difficulty lies carefully addressed before the MA schemes can work in IoT
in the unknown relation between the sensing matrix and the networks. Moreover, there exist correlation in user activity
detection performance. It is well-known that AI techniques are along the temporal and spatial domain. For example, if there
a good option to learn the complicated and unknown relations. is a fire, several sensors around the fire site tend to be active
Therefore, recently, there is a growing interest in applying together, and each of these sensors tend to be active for some
AI to design the pilot signals for massive IoT connectivity. time. How to utilize these correlation as side information to
For example, it was found in [458], [459] that if the deep improve the activity detection accuracy is also a challenge.
learning technique is used to design the pilot signals, the
device detection performance can be improved compared to VII. ROADMAP TO 6G S TANDARDIZATION
using Gaussian pilots and Reed-muller sequences based pilots. 3GPP is the standards body that introduced UMTS/W-
On the other hand, deep learning technique can also be CDMA for 3G, LTE for 4G and NR for 5G. 3GPP also
applied to design efficient device activity detection algorithms, introduced GSM for 2G even before being called 3GPP. While
when the dimension of the problem is huge due to the large 6G work has not yet started in 3GPP, it is expected to quickly
number of IoT devices. In general, these deep learning based intensify in the 2nd half of 2025.
device activity detection methods can be classified into data- At the highest conceptual level, 3GPP has delivered over
driven based methods and model-driven based methods. Under its history systems based on TDMA (2G GSM), CDMA
data-driven based methods, basic neural network architectures, (3G UMTS) and (O)FDMA (4G LTE and 5G NR) closely
e.g., fully connected neural networks, CNNs, are utilized to following any textbook description of basic MA schemes. In
detect active devices [459], [460]. Under model-driven based this section we look ahead into what 3GPP may consider for
methods, existing algorithms, e.g., AMP, group LASSO, etc., MA for 6G by looking back first into previous generations.
are utilized to train the neural networks for better activity The work of 3GPP is structured in so-called Releases, the
detection performance [461]–[464]. first version of the 3G specifications was introduced in Release
To summarize, there is a theoretical gain of the grant-free 99 towards the end of last millennium. Release 4 was the
random access scheme over the grant-based random access subsequent Release after which Releases have been labeled
scheme in IoT networks, because the former scheme can with increasing index. For example, the first version of 4G
mitigate almost all the collisions in the massive MIMO region, LTE was introduced in Release 8 while the first version of
but the latter scheme can merely tackle some of the collisions. 5G NR was introduced in Release 15. Release 18, the first
But the theoretical gain is achieved with complicated device Release of 5G-Advanced, was finalized late 2023 and Release
activity detection algorithms, under both the compressed sens- 19 started in early 2024. The study phase for 6G is expected
ing based scheme and the covariance based scheme. to launch in Release 20 in the second half of 2025. In turn,
4) Grant-Free Unsourced Random Access: For unsourced the first 6G specifications are expected in Release 21 with a
RA [465], the receiver is interested in the transmitted messages completion date anticipated to be between the end of 2028
only (user identity is recovered at the higher layers of the com- and the end of 2029 aiming at commercial 6G deployments
munication protocol rather than the physical layer). Moreover, sometime around 2030.
all users share a common codebook, which is motivated by MA is one of the main characteristics of any wireless
the practical scenario where millions of low-cost IoT devices system. Indeed, it is one of the first decisions made when the
have their codebook hardwired at the moment of production. In design of a new generation takes place as it relates closely to
addition, the number of users grows with the blocklength. Only the underlying waveform. It is important to note that, typically,
per-user error probability is of interest. The fundamental limits there is one main MA scheme which is supplemented by
of the unsourced RA channel have been studied in [465]– other schemes for certain specific functions as we will discuss
[467]. Various popular achievable schemes are proposed in the shortly.
literature, which are based on concatenated coding structures
[468], [469], ALOHA [470], random spreading [471], [472],
etc. A. MA in 3G and 4G
5) Technology Outlook and Future Works: There are still In 3G systems, the MA scheme in the downlink is based
many challenges for MA in IoT network that have not been on CDMA with Hadamard or Walsh codes applied to coded
addressed. For example, most of the works assume that all modulation symbols and orthogonally separating the various
the IoT devices are perfectly synchronized, such that when channels in a given cell. Different codes can be used to
 41

transmit data of different users and multiple codes can be used domain gating function which did not exist in 3G days.
to increase the data rate for a given user. The application of While FDMA/TDMA was the main MA scheme for the
these codes effectively expands the transmission bandwidth by uplink enabling orthogonal access for users with the same
a factor called spreading factor. The spreading factor offers serving cell, it is not the only MA scheme used in 4G LTE.
the ability to boost the SINR at the receiver helping to cope Indeed, random access is performed via the Physical Ran-
with multiple sources of interference, i.e., inter-cell, intra-cell dom Access Channel (PRACH) which, while orthogonalized
(ISI). However, the spreading factor decreased over time as the in time and frequency with the other uplink Physical Layer
target operating data rates increased. This trend increased the (PHY) channels, PRACH transmissions from different users
need to use equalizers at the receiver, which, in turn, made overlap with each other yielding a CDMA-like MA for this
OFDM based transmissions more appealing as they would physical channel with possibility of recoverable or unrecover-
not require equalization at the receiver. The application of a able inter-user collisions.
cell-dependent scrambling based on long pseudo-noise (PN) Furthermore, SDMA was envisioned to enable MU-MIMO
sequences onto the aggregated transmissions of a given cell operation from the start of 4G. In this scenario, the spatial
makes it appear as noise when received at neighboring cells dimensionality of the channel is exploited by enabling data
in full frequency reuse deployments. transmissions from different users to overlap in time and
For 4G, OFDMA was introduced in the downlink main- frequency with each other relying on spatially separating them
taining intra-cell orthogonality while being more resilient to at the multi-antenna base-station receiver. As a result, while
ISI caused by the delay spread of the channel. In full fre- the data portion of the uplink transmission is merely spatial-
quency reuse deployments, transmissions from different cells division-multiplexed (SDM’d) across users, their reference
are randomized through the application of a cell-dependent signal transmissions are orthogonalized by virtue of different
scrambling based on long PN sequences similar to 3G. cyclic shifts of the underlying Zadoff-Chu sequences used
In the transition from 3G to 4G, for the uplink we went from to modulate the demodulation reference signals (DMRS). In
a non-orthogonal MA with CDMA in 3G to orthogonal MA turn, that orthogonalization enables the possibility to perform
with DFT-spread OFDM (DFT-S-OFDM) in 4G. Transitions channel estimation for each user’s transmission off cleaner
are typically somewhat blurred with inroads into new schemes DMRS samples without inter-user interference.
by way of extensions of the current scheme before an entirely Moreover, uplink control channel transmissions typically
new scheme is introduced. For example, in the second half bear few bits of information which attempting to orthogonalize
of 3G days we experienced the migration from circuit switch solely in the frequency domain would very rapidly run out of
(CS) to packet switch (PS) communications. With that mi- dimensions in the frequency domain (resource limitation). In
gration, we also moved from continuous downlink and uplink order to alleviate that limitation, the possibility to orthogonally
transmissions to packetized transmissions which would need to multiplex within the same time/frequency resources via differ-
be scheduled by the network. Therefore, Release 5 introduced ent cyclic shifts of underlying Zadoff-Chu sequences (similar
scheduled downlink with much lower multiplexing capability to DMRS) or time-domain orthogonal covers applied to each
in the code domain (lower spreading factor) and aiming uplink symbol was also introduced to increase the capacify
more at time-domain-multiplexing (TDM’ing) transmissions of the uplink control channel. Release 10 introduced the
to multiple users with the goal to increase the instantaneous possibility for SU-MIMO in the uplink of LTE still resorting
individual user data rate with the so-called high speed down- to DFT-S-OFDM for each of the transmitted layers.
link packet access (HSDPA). Similarly, Release 6 introduced at Towards the latter part of LTE, multi-user superposition
the time scheduled UL with what was called enhanced uplink transmission (MUST) was standardized for the downlink of
(eUL) or high speed uplink packet access (HSUPA). Release 7, LTE in Release 14 after the conclusion of the corresponding
in turn, introduced discontinuous uplink transmissions which, study item in the preceding release. The project attempted to
for the first time, enabled uplink transmissions to be gated jointly optimize multi-user operation from both transmitter and
in time and hence favoring TDM’ing of transmissions from receiver’s perspective to improve the MU system capacity even
different users supplemented with the possibility of non- if the transmission/precoding was non-orthogonal. The out-
orthogonally multiplexing multiple users’ transmissions in the come of the study is captured in the technical report (TR) for
code domain (CDMA). MUST in [8]. Three MUST categories were identified during
The main driver for the switch of MA from 3G to 4G the study: 1) MUST Category 1: Superposition transmission
was the uplink performance for cell-edge users for which with adaptive power ratio on component constellations and
the introduction of frequency domain multiplexing (FDM’ing) non-Gray-mapped composite constellation; 2) MUST Cate-
brought important performance gains in terms of uplink user gory 2: Superposition transmission with adaptive power ratio
throughput [473]. The realization of FDM in the uplink of on component constellations and Gray-mapped composite con-
4G was done via the introduction of DFT-S-OFDM, which stellation; MUST Category 3: Superposition transmission with
due to its single carrier properties was key to reduce the user label-bit assignment on Gray-mapped composite constellation.
transmissions’ peak-to-average-power-ratio (PAPR) which di-
rectly entails a coverage gain for the uplink. Also, because of
its underlying block-based construction of OFDM, subframe B. MA in 5G and 6G
transmissions of 1ms simply meant transmissions of 14 sym- With the exception of MUST, all the schemes mentioned for
bols facilitating TDM’ing of transmissions with a built-in time- 4G LTE were also adopted for 5G NR. Indeed, for the uplink
 42

data channel, the earlier mentioned DFT-S-OFDM is only used whether this approach will still hold for 6G with the pro-
for single layer transmissions, while multi-layer transmissions liferation of schemes to cope with eMBB and mMTC but
always resort to OFDMA to improve link efficiency at the cost also and mainly to enable new use cases fueled by emerging
of an increased PAPR. The DMRS of the uplink data channel intelligent and multi-functional applications. Aside the ones
for different layers of the same user and from different users already discussed as part of 5G, more recent MA schemes
in the same cell are orthogonalized in time/frequency domain have made their way toward standardization. This is the
by corresponding orthogonal covers. case of RSMA whose ”universal” capability has been found
5G NR while introduced in Release 15, had a project in attractive by various industries and recently been proposed
Release 16 to study NOMA [474] as a follow up to the for the first time for 6G for various use cases such as 1) to
discussions during the 5G NR study of Release 14 [475]. boost the performance of unicast transmissions beyond what
The TR for NOMA is a good reference [7] to review the has been achieved traditionally in 4G and 5G by SDMA-
schemes that were considered during the study with some of multi-user MIMO [477], and 2) to enable non-orthogonal
them presented in this paper. The Release 16 NOMA study unicast and multicast transmissions where multicast traffic
investigated the application of NOMA techniques to mMTC, are superimposed over unicast traffic instead of being served
URLLC, and eMBB communications. in orthogonal resources as in 4G and 5G [478]. It will be
The uplink NOMA transmitter was generalized so that it interesting to see in the coming years whether 3GPP will
would include all the different flavors being considered. Also, favour a single unified and general MA scheme as main MA
different receivers were considered tailored to the different scheme instead of a combination of multiple MA schemes,
schemes and incurring various levels of complexity. each optimized for specific conditions.
The outcome of the Release 16 study was deemed to be
non-conclusive and a follow up work item was not approved
C. AI/ML in 5G and 6G
at the time. Part of the reason was the proliferation of schemes
during the study without a clear winner amongst them which A Study Item for AI/ML applied to the NR Air Interface was
made it impossible to define the scope of a normative project agreed for Release 18 [479]. The TR for this study is available
for only one scheme. Perhaps the scope of the project was too at [480] and constitutes a good reference to review what is be-
broad aiming at NOMA for mMTC, URLLC and eMBB use ing studied (project completion planned for December 2023).
cases. Note that the mMTC use case of IMT-2020 is actually While AI/ML first appeared in 3GPP for the purpose of users’
served by LTE CatM (eMTC) and NBIOT originally developed data collection, this project is the first time in 3GPP where the
in Release 13 and evolved thereafter. direct application of AI/ML is sought for some fundamental
Indeed, it was discussed that NOMA could be enabled air interface problems. In order to avoid too abstract, high
in NR without specification impact provided that there was level conceptual discussions and to give some shape to the
adequate receiver processing. The possibility to expand the project, a number of pilot use cases was identified to be
number of orthogonal DMRS ports was discussed which, in analyzed for this project. Namely, CSI enhancements, beam
combination with multi-user detectors at the receiver, would management (BM) and positioning accuracy enhancements.
enable the possibility for processing fully overlapping data Each of those three use cases was further refined into two
transmissions from different users. Those transmissions from sub-use cases. Namely, CSI (spatial-frequency) compression
multiple users could be resulting from dynamic scheduling and CSI (temporal) prediction for CSI enhancement, Spatial
from the network (with individual grants to each of the users) beam prediction and Temporal beam prediction for BM, direct
or from overlapping configured grants (pre-configured via L2 AI/ML positioning and AI/ML assisted positioning for Posi-
or L3 allocations). tioning accuracy enhancements. All sub use cases but CSI
In the absence of a normative projects for NOMA specifica- compression rely on single-sided models at the user terminal
tion, a minor improvement to the Random Access procedure or the network, while CSI compression resorts to two-sided
which used to take 4-steps was approved with the goal to AI/ML models with one part of the model running at the
shorten it to 2-steps [476]. It is, however, likely that the user terminal side and the other at the network side. For the
investigations that took part during Release 16 will be revisited CSI compression sub use case, the user terminal side performs
when NOMA is considered again in 3GPP in the future. the encoding or compression of the CSI information, and the
Clearly, there is no universal MA (yet) that optimally suits network side performs the decoder or decompression to obtain
all applications and use cases. The MA scheme aiming at op- the terminal’s CSI.
timizing individual user’s high throughput operation (eMBB- It is expected that normative work to address the specifica-
like use case) is not the same as the one aiming at maximizing tion impact of these use cases will be carried out as part of
multiplexing capability for low-rate applications of many, Release 19. Moreover, further use cases, for example Mobility
many users (mMTC-like use case). To some extent, that trade- enhancements, may be additionally investigated.
off is already seen by looking into the nuanced MA scheme As indicated by the ITU-R WP 5D framework document for
used today for random access, uplink control channel, uplink IMT-2030 [481], Integrated AI and communication has been
data channel, etc. identified as one of the usage scenarios for IMT-2030. As a
6G will see a resurgence of MA discussions. Though result, it is expected that 6G will be AI/ML native.
traditionally there is one main MA scheme supplemented While nobody knows precisely what that will imply from
by other schemes for certain specific functions, it is unclear 3GPP perspective, the foundational work on AI/ML for NR air
 43

interface that is underway will for sure play a very important [3] C. De Lima, D. Belot, R. Berkvens, A. Bourdoux, D. Dardari,
role on how native 6G will be specified [482] [483] [484] M. Guillaud, M. Isomursu, E.-S. Lohan, Y. Miao, A. N. Barreto,
M. R. K. Aziz, J. Saloranta, T. Sanguanpuak, H. Sarieddeen, G. Seco-
[485] [486]. Granados, J. Suutala, T. Svensson, M. Valkama, B. Van Liempd, and
The application of AI/ML techniques to MA may be inves- H. Wymeersch, “Convergent communication, sensing and localization
tigated for 6G with the aim at figuring out the optimal MA in 6g systems: An overview of technologies, opportunities and chal-
lenges,” IEEE Access, vol. 9, pp. 26 902–26 925, 2021.
scheme for the given application and use case. Indeed, channel [4] K. B. Letaief, W. Chen, Y. Shi, J. Zhang, and Y. J. A. Zhang, “The
access policies that fluently transition from contention- to roadmap to 6G: AI empowered wireless networks,” IEEE Communi-
schedule-based depending on the use case and environment cations Magazine, vol. 57, no. 8, pp. 84–90, August 2019.
[5] B. Clerckx, K. Huang, L. R. Varshney, S. Ulukus, and M.-S. Alouini,
are deemed desirable [484]. “Wireless power transfer for future networks: Signal processing, ma-
Similar to the Rel-18 study where single-sided and two- chine learning, computing, and sensing,” IEEE Journal of Selected
sided (or cross-node) models are being considered depending Topics in Signal Processing, vol. 15, no. 5, pp. 1060–1094, 2021.
[6] B. Clerckx, J. Kim, K. W. Choi, and D. I. Kim, “Foundations
on the use case, different AI/ML functions for 6G will require of wireless information and power transfer: Theory, prototypes, and
one or another approach. In particular, how the AI/ML engines experiments,” Proceedings of the IEEE, vol. 110, no. 1, pp. 8–30, 2022.
running at the terminals and the network will figure out the [7] 3GPP, “Study on non-orthogonal multiple access (NOMA) for NR,”
3rd Generation Partnership Project (3GPP), Technical Report TR
optimal MA scheme is subject of research and will for sure 38.812 V16.0.0, Dec. 2018.
open up new opportunities for improved system performance [8] ——, “Study on Downlink Multiuser Superposition Transmission
and user experience. Adaptability of the models through (MUST) for LTE,” 3rd Generation Partnership Project (3GPP), Tech-
nical Report TR 36.859 V13.0.0, Dec. 2015.
further offline or even online training will also play a key [9] B. Clerckx, Y. Mao, E. A. Jorswieck, J. Yuan, D. J. Love, E. Erkip, and
role in the end-to-end performance optimization. D. Niyato, “A primer on rate-splitting multiple access: Tutorial, myths,
As a result, the current investigations on join training and frequently asked questions,” IEEE Journal on Selected Areas in
Communications, vol. 41, no. 5, pp. 1265–1308, 2023.
of two-sided models, as well as, their interoperability and [10] B. Clerckx, Y. Mao, R. Schober, E. A. Jorswieck, D. J. Love, J. Yuan,
performance testing will be very relevant for the future native L. Hanzo, G. Y. Li, E. G. Larsson, and G. Caire, “Is NOMA efficient
AI work in 6G. in multi-antenna networks? a critical look at next generation multiple
access techniques,” IEEE Open Journal of the Communications Society,
vol. 2, pp. 1310–1343, 2021.
[11] B. Clerckx, Y. Mao, R. Schober, and H. V. Poor, “Rate-splitting
VIII. C ONCLUSIONS unifying SDMA, OMA, NOMA, and multicasting in MISO broadcast
Wireless networks experience a paradigm shift with the channel: A simple two-user rate analysis,” IEEE Wireless Commun.
Lett., vol. 9, no. 3, pp. 349–353, Mar. 2020.
integration of multi-functionality and the rise of AI and ML. [12] D. Gesbert, M. Kountouris, R. W. Heath, C.-b. Chae, and T. Salzer,
6G and beyond will be AI-native and multi-functionality- “Shifting the mimo paradigm,” IEEE Signal Processing Magazine,
native, spurring the need to re-think multiple access techniques vol. 24, no. 5, pp. 36–46, 2007.
[13] L. Dai, B. Wang, Y. Yuan, S. Han, I. Chih-lin, and Z. Wang, “Non-
to make use of the available time, frequency, space, power, orthogonal multiple access for 5g: solutions, challenges, opportunities,
signal dimensions to serve users, devices, machines, services, and future research trends,” IEEE Communications Magazine, vol. 53,
training nodes in the most efficient way and cope with those no. 9, pp. 74–81, 2015.
[14] Y. Mao, O. Dizdar, B. Clerckx, R. Schober, P. Popovski, and H. V.
new network advances. The paper highlighted that much work Poor, “Rate-splitting multiple access: fundamentals, survey, and future
is left for researchers in this area. Aside many other schemes, research trends,” IEEE Commun. Surv. Tutor., vol. 24, no. 4, pp. 2073–
RSMA has emerged in the past few years by providing a 2126, Jul. 2022.
[15] J. G. Proakis, Digital Communications. New York, NY, USA:
unified and conceptually simple understanding of SDMA, PD- McGraw-Hill, 2008.
NOMA, physical layer multicasting, and uniquely shrinks [16] J. Park, B. Lee, J. Choi, H. Lee, N. Lee, S.-H. Park, K.-J. Lee, J. Choi,
rather grows the knowledge tree of MA schemes based on S. H. Chae, S.-W. Jeon, K. S. Kwak, B. Clerckx, and W. Shin, “Rate-
splitting multiple access for 6g networks: Ten promising scenarios and
space, power, signal dimensions. Consequently RSMA has applications,” IEEE Network, pp. 1–1, 2023.
found numerous applications for 6G. Moving next, UMA [17] A. Mishra, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple
should further shrink the knowledge tree of MA schemes access for 6g—part i: Principles, applications and future works,” IEEE
Communications Letters, vol. 26, no. 10, pp. 2232–2236, 2022.
by unifying RSMA with all other dimensions, such as code [18] L. Yin, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple
domain MAs, and ultimately provide a unified and conceptu- access for 6G—part II: Interplay with integrated sensing and commu-
ally simple understanding of the current and future morass of nications,” IEEE Communications Letters, vol. 26, no. 10, pp. 2237–
2241, 2022.
MA schemes. Such UMA does not exist yet. It is hoped that [19] H. Li, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple ac-
the techniques and outlook presented in this article will help cess for 6g—part iii: Interplay with reconfigurable intelligent surfaces,”
inspiring future research in this exciting and important area IEEE Communications Letters, vol. 26, no. 10, pp. 2242–2246, 2022.
[20] Z. Ding, L. Lv, F. Fang, O. A. Dobre, G. K. Karagiannidis, N. Al-
and pave the way for designing and implementing efficient Dhahir, R. Schober, and H. V. Poor, “A state-of-the-art survey on re-
multiple access techniques in future wireless systems. configurable intelligent surface-assisted non-orthogonal multiple access
networks,” Proceedings of the IEEE, vol. 110, no. 9, pp. 1358–1379,
2022.
R EFERENCES [21] F. Liu, Y. Cui, C. Masouros, J. Xu, T. X. Han, Y. C. Eldar, and
S. Buzzi, “Integrated sensing and communications: Toward dual-
[1] W. Saad, M. Bennis, and M. Chen, “A vision of 6G wireless systems: functional wireless networks for 6G and beyond,” IEEE Journal on
Applications, trends, technologies, and open research problems,” IEEE Selected Areas in Communications, vol. 40, no. 6, pp. 1728–1767,
Network, vol. 34, no. 3, pp. 134–142, May 2020. 2022.
[2] B. Clerckx, R. Zhang, R. Schober, D. W. K. Ng, D. I. Kim, and H. V. [22] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
Poor, “Fundamentals of wireless information and power transfer: From Cambridge University Press, 2005.
rf energy harvester models to signal and system designs,” IEEE Journal [23] “Second generation framing structure, channel coding and modulation
on Selected Areas in Communications, vol. 37, no. 1, pp. 4–33, 2019. systems for broadcasting interative services, news gathering and other
 44

broadband satellite applications; part 2: Dvb-s2 extensions (dvbs2x), [48] S. Park, J. Choi, J. Park, W. Shin, and B. Clerckx, “Rate-splitting
document etsi en 302-307-2 v1.1.1, european broadcasting union (ebu), multiple access for quantized multiuser mimo communications,” IEEE
grand-saconnex, switzerland, oct. 2014.” Transactions on Wireless Communications, vol. 22, no. 11, pp. 7696–
[24] L. Yin and B. Clerckx, “Rate-splitting multiple access for multigroup 7711, 2023.
multicast and multibeam satellite systems,” IEEE Transactions on [49] A. Mishra, Y. Mao, L. Sanguinetti, and B. Clerckx, “Rate-splitting
Communications, vol. 69, no. 2, pp. 976–990, 2021. assisted massive machine-type communications in cell-free massive
[25] B. Clerckx and C. Oestges, MIMO wireless networks: Channels, mimo,” IEEE Communications Letters, vol. 26, no. 6, pp. 1358–1362,
techniques and standards for multi-antenna, multi-user and multi-cell 2022.
systems. Academic Press, 2013. [50] A. Mishra, Y. Mao, C. K. Thomas, L. Sanguinetti, and B. Clerckx,
[26] Q. Spencer, A. Swindlehurst, and M. Haardt, “Zero-forcing methods “Mitigating intra-cell pilot contamination in massive mimo: A rate
for downlink spatial multiplexing in multiuser mimo channels,” IEEE splitting approach,” IEEE Transactions on Wireless Communications,
Transactions on Signal Processing, vol. 52, no. 2, pp. 461–471, 2004. vol. 22, no. 5, pp. 3472–3487, 2023.
[27] M. Costa, “Writing on dirty paper (corresp.),” IEEE Trans. Inf. Theory, [51] O. Dizdar, Y. Mao, and B. Clerckx, “Rate-splitting multiple access to
vol. 29, no. 3, pp. 439–441, May 1983. mitigate the curse of mobility in (massive) mimo networks,” IEEE
[28] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacity Transactions on Communications, vol. 69, no. 10, pp. 6765–6780,
limits of MIMO channels,” IEEE J. Sel. Areas Commun., vol. 21, no. 5, 2021.
pp. 684–702, June 2003. [52] A. Papazafeiropoulos, B. Clerckx, and T. Ratnarajah, “Rate-splitting to
[29] N. Jindal, “MIMO broadcast channels with finite-rate feedback,” IEEE mitigate residual transceiver hardware impairments in massive mimo
Trans. Inf. Theory, vol. 52, no. 11, pp. 5045–5060, Nov. 2006. systems,” IEEE Transactions on Vehicular Technology, vol. 66, no. 9,
[30] Y. Mao and B. Clerckx, “Beyond dirty paper coding for multi-antenna pp. 8196–8211, 2017.
broadcast channel with partial CSIT: A rate-splitting approach,” IEEE [53] A. R. Flores, R. C. de Lamare, and B. Clerckx, “Linear precoding and
Trans. Commun., vol. 68, no. 11, pp. 6775–6791, Nov. 2020. stream combining for rate splitting in multiuser mimo systems,” IEEE
[31] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and Communications Letters, vol. 24, no. 4, pp. 890–894, 2020.
K. Higuchi, “Non-orthogonal multiple access (noma) for cellular future [54] A. Mishra, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple
radio access,” in 2013 IEEE 77th Vehicular Technology Conference access for downlink multiuser mimo: Precoder optimization and phy-
(VTC Spring), 2013, pp. 1–5. layer design,” IEEE Transactions on Communications, vol. 70, no. 2,
[32] M. Vaezi, Z. Ding, and H. V. Poor, Multiple Access Techniques for 5G pp. 874–890, 2022.
Wireless Networks and Beyond. Springer, 2019, vol. 159. [55] H. Chen, D. Mi, B. Clerckx, Z. Chu, J. Shi, and P. Xiao, “Joint power
[33] T. Cover, “Broadcast channels,” IEEE Trans. Inf. Theory, vol. 18, no. 1, and subcarrier allocation optimization for multigroup multicast sys-
pp. 2–14, Jan. 1972. tems with rate splitting,” IEEE Transactions on Vehicular Technology,
vol. 69, no. 2, pp. 2306–2310, 2020.
[34] B. Makki, K. Chitti, A. Behravan, and M.-S. Alouini, “A survey
[56] A. Z. Yalcin, M. Yuksel, and B. Clerckx, “Rate splitting for multi-
of noma: Current status and open research challenges,” IEEE Open
group multicasting with a common message,” IEEE Transactions on
Journal of the Communications Society, vol. 1, pp. 179–189, 2020.
Vehicular Technology, vol. 69, no. 10, pp. 12 281–12 285, 2020.
[35] B. Clerckx, H. Joudeh, C. Hao, M. Dai, and B. Rassouli, “Rate splitting
[57] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting for multi-antenna
for MIMO wireless networks: A promising PHY-layer strategy for LTE
non-orthogonal unicast and multicast transmission: Spectral and energy
evolution,” IEEE Commun. Mag., vol. 54, no. 5, pp. 98–105, May 2016.
efficiency analysis,” IEEE Transactions on Communications, vol. 67,
[36] H. Joudeh and B. Clerckx, “Sum-rate maximization for linearly pre-
no. 12, pp. 8754–8770, 2019.
coded downlink multiuser miso systems with partial csit: A rate-
[58] L. Yin and B. Clerckx, “Rate-splitting multiple access for satellite-
splitting approach,” IEEE Transactions on Communications, vol. 64,
terrestrial integrated networks: Benefits of coordination and coopera-
no. 11, pp. 4847–4861, 2016.
tion,” IEEE Transactions on Wireless Communications, vol. 22, no. 1,
[37] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting multiple access pp. 317–332, 2023.
for downlink communication systems: bridging, generalizing, and [59] Z. W. Si, L. Yin, and B. Clerckx, “Rate-splitting multiple access for
outperforming SDMA and NOMA,” EURASIP J. Wireless Commun. multigateway multibeam satellite systems with feeder link interfer-
Netw., vol. 2018, no. 1, p. 133, May 2018. ence,” IEEE Transactions on Communications, vol. 70, no. 3, pp. 2147–
[38] B. Rimoldi and R. Urbanke, “A rate-splitting approach to the Gaussian 2162, 2022.
multiple-access channel,” IEEE Transactions on Information Theory, [60] Y. Liu, B. Clerckx, and P. Popovski, “Network slicing for embb, urllc,
vol. 42, no. 2, pp. 364–375, 1996. and mmtc: An uplink rate-splitting multiple access approach,” IEEE
[39] A. R. Flores, R. C. De Lamare, and B. Clerckx, “Tomlinson-harashima Transactions on Wireless Communications, pp. 1–1, 2023.
precoded rate-splitting with stream combiners for mu-mimo systems,” [61] S. Zhang, B. Clerckx, D. Vargas, O. Haffenden, and A. Murphy, “Rate-
IEEE Transactions on Communications, vol. 69, no. 6, pp. 3833–3845, splitting multiple access: Finite constellations, receiver design, and sic-
2021. free implementation,” arXiv:2305.17178, 2023.
[40] C. Hao, Y. Wu, and B. Clerckx, “Rate analysis of two-receiver miso [62] X. Lyu, S. Aditya, J. Kim, and B. Clerckx, “Rate-splitting multiple
broadcast channel with finite rate feedback: A rate-splitting approach,” access: The first prototype and experimental validation of its superiority
IEEE Transactions on Communications, vol. 63, no. 9, pp. 3232–3246, over sdma and noma,” arXiv 2305.07361, 2023.
2015. [63] G. Zhou, Y. Mao, and B. Clerckx, “Rate-splitting multiple access for
[41] M. Dai, B. Clerckx, D. Gesbert, and G. Caire, “A rate splitting strategy multi-antenna downlink communication systems: Spectral and energy
for massive mimo with imperfect csit,” IEEE Transactions on Wireless efficiency tradeoff,” IEEE Transactions on Wireless Communications,
Communications, vol. 15, no. 7, pp. 4611–4624, 2016. vol. 21, no. 7, pp. 4816–4828, July 2022.
[42] H. Joudeh and B. Clerckx, “Rate-splitting for max-min fair multigroup [64] J. Zhang, B. Clerckx, J. Ge, and Y. Mao, “Cooperative rate splitting for
multicast beamforming in overloaded systems,” IEEE Transactions on MISO broadcast channel with user relaying, and performance benefits
Wireless Communications, vol. 16, no. 11, pp. 7276–7289, 2017. over cooperative NOMA,” IEEE Signal Process. Lett., vol. 26, no. 11,
[43] Y. Mao, E. Piovano, and B. Clerckx, “Rate-splitting multiple access pp. 1678–1682, Nov. 2019.
for overloaded cellular internet of things,” IEEE Transactions on [65] Y. Mao, B. Clerckx, J. Zhang, V. O. K. Li, and M. A. Arafah, “Max-
Communications, vol. 69, no. 7, pp. 4504–4519, 2021. min fairness of K-user cooperative rate-splitting in MISO broadcast
[44] H. Joudeh and B. Clerckx, “Robust transmission in downlink multiuser channel with user relaying,” IEEE Trans. Wireless Commun., vol. 19,
miso systems: A rate-splitting approach,” IEEE Transactions on Signal no. 10, pp. 6362–6376, Oct. 2020.
Processing, vol. 64, no. 23, pp. 6227–6242, 2016. [66] H. Cho, B. Ko, B. Clerckx, and J. Choi, “Coverage increase at thz
[45] C. Hao and B. Clerckx, “Miso networks with imperfect csit: A topolog- frequencies: A cooperative rate-splitting approach,” IEEE Transactions
ical rate-splitting approach,” IEEE Transactions on Communications, on Wireless Communications, vol. 22, no. 12, pp. 9821–9834, 2023.
vol. 65, no. 5, pp. 2164–2179, 2017. [67] Y. Xu, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple ac-
[46] E. Piovano and B. Clerckx, “Optimal dof region of the k -user miso cess with finite blocklength for short-packet and low-latency downlink
bc with partial csit,” IEEE Communications Letters, vol. 21, no. 11, communications,” IEEE Transactions on Vehicular Technology, vol. 71,
pp. 2368–2371, 2017. no. 11, pp. 12 333–12 337, 2022.
[47] C. Hao, B. Rassouli, and B. Clerckx, “Achievable dof regions of [68] ——, “Max-min fairness of rate-splitting multiple access with finite
mimo networks with imperfect csit,” IEEE Transactions on Information blocklength communications,” IEEE Transactions on Vehicular Tech-
Theory, vol. 63, no. 10, pp. 6587–6606, 2017. nology, vol. 72, no. 5, pp. 6816–6821, 2023.
 45

[69] J. Xu, O. Dizdar, and B. Clerckx, “Rate-splitting multiple access for [92] L. Ping, L. Liu, K. Wu, and W. Leung, “Interleave-division multiple-
short-packet uplink communications: A finite blocklength analysis,” access (IDMA) communications,” in Proc. 3rd International Sympo-
IEEE Commun. Lett., pp. 1–1, 2022. sium on Turbo Codes & Related Topics, 2003, pp. 173–180.
[70] A. Salem, C. Masouros, and B. Clerckx, “Secure rate splitting multiple [93] L. Ping, “Interleave-division multiple access and chip-by-chip iterative
access: How much of the split signal to reveal?” IEEE Transactions multi-user detection,” IEEE Communications Magazine, vol. 43, no. 6,
on Wireless Communications, vol. 22, no. 6, pp. 4173–4187, 2023. pp. S19–S23, 2005.
[71] H. Xia, Y. Mao, X. Zhou, B. Clerckx, S. Han, and C. Li, [94] L. Ping, L. Liu, K. Wu, and W. Leung, “Interleave division multiple-
“Weighted sum-rate maximization for rate-splitting multiple access access,” IEEE Transactions on Wireless Communications, vol. 5, no. 4,
based multi-antenna broadcast channel with confidential messages,” pp. 938–947, 2006.
arXiv 2202.07328, 2023. [95] P. Wang, J. Xiao, and L. Ping, “Comparison of orthogonal and
[72] Z. Liu and L.-L. Yang, “Sparse or dense: A comparative study of code- non-orthogonal approaches to future wireless cellular systems,” IEEE
domain NOMA systems,” IEEE Transactions on Wireless Communi- Vehicular Technology Magazine, vol. 1, no. 3, pp. 4–11, 2006.
cations, vol. 20, no. 8, pp. 4768–4780, 2021. [96] K. Kusume, G. Bauch, and W. Utschick, “IDMA vs. CDMA: Analysis
[73] B. M. Zaidel and S. S. Shitz, “Sparse and dense: An achievable region and comparison of two multiple access schemes,” IEEE Transactions
for code-domain NOMA with mixed users,” in IEEE International on Wireless Communications, vol. 11, no. 1, pp. 78–87, 2012.
Symposium on Information Theory (ISIT), 2021, pp. 2942–2947. [97] L. Ping, Q. Guo, and J. Tong, “The OFDM-IDMA approach to wireless
[74] R. Hoshyar, F. P. Wathan, and R. Tafazolli, “Novel low-density sig- communication systems,” IEEE Wireless Communications, vol. 14,
nature for synchronous CDMA systems over AWGN channel,” IEEE no. 3, pp. 18–24, 2007.
Transactions on Signal Processing, vol. 56, no. 4, pp. 1616–1626, 2008.
[98] C. Xu, Y. Hu, C. Liang, J. Ma, and L. Ping, “Massive MIMO, non-
[75] F. Kschischang, B. Frey, and H.-A. Loeliger, “Factor graphs and the
orthogonal multiple access and interleave division multiple access,”
sum-product algorithm,” IEEE Transactions on Information Theory,
IEEE Access, vol. 5, pp. 14 728–14 748, 2017.
vol. 47, no. 2, pp. 498–519, 2001.
[76] D. Guo and C.-c. Wang, “Multiuser detection of sparsely spread [99] S. Kang, X. Dai, and B. Ren, “Pattern division multiple access for 5G,”
CDMA,” IEEE Journal on Selected Areas in Communications, vol. 26, Telecommun. Netw. Technol., vol. 5, no. 5, pp. 43–47, 2015.
no. 3, pp. 421–431, 2008. [100] X. Wang and H. Poor, “Iterative (turbo) soft interference cancellation
[77] J. van de Beek and B. M. Popovic, “Multiple access with low-density and decoding for coded CDMA,” IEEE Transactions on Communica-
signatures,” in IEEE Global Telecommunications Conference, 2009, pp. tions, vol. 47, no. 7, pp. 1046–1061, 1999.
1–6. [101] S. Chen, B. Ren, Q. Gao, S. Kang, S. Sun, and K. Niu, “Pattern
[78] R. Razavi, R. Hoshyar, M. A. Imran, and Y. Wang, “Information division multiple access—a novel nonorthogonal multiple access for
theoretic analysis of LDS scheme,” IEEE Communications Letters, fifth-generation radio networks,” IEEE Transactions on Vehicular Tech-
vol. 15, no. 8, pp. 798–800, 2011. nology, vol. 66, no. 4, pp. 3185–3196, 2017.
[79] R. Hoshyar, R. Razavi, and M. Al-Imari, “LDS-OFDM an efficient [102] X. Dai, Z. Zhang, B. Bai, S. Chen, and S. Sun, “Pattern division
multiple access technique,” in IEEE Vehicular Technology Conference, multiple access: A new multiple access technology for 5G,” IEEE
2010, pp. 1–5. Wireless Communications, vol. 25, no. 2, pp. 54–60, 2018.
[80] M. AL-Imari, M. A. Imran, and R. Tafazolli, “Low density spreading [103] B. Nazer and M. Gastpar, “Compute-and-forward: Harnessing inter-
for next generation multicarrier cellular systems,” in International ference through structured codes,” IEEE Transactions on Information
Conference on Future Communication Networks (ICFCN), 2012, pp. Theory, vol. 57, no. 10, pp. 6463–6486, 2011.
52–57. [104] J. Zhu and M. Gastpar, “Gaussian multiple access via compute-and-
[81] M. AL-Imari, M. A. Imran, R. Tafazolli, and D. Chen, “Performance forward,” IEEE Transactions on Information Theory, vol. 63, no. 5,
evaluation of low density spreading multiple access,” in Interna- pp. 2678–2695, 2017.
tional Wireless Communications and Mobile Computing Conference [105] U. Erez and R. Zamir, “Achieving 1/2 log (1+SNR) on the AWGN
(IWCMC), 2012, pp. 383–388. channel with lattice encoding and decoding,” IEEE Transactions on
[82] H. Nikopour and H. Baligh, “Sparse code multiple access,” in Proc. Information Theory, vol. 50, no. 10, pp. 2293–2314, 2004.
IEEE Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC), Sep. [106] T. M. Cover and J. A. Thomas, Elements of Information Theory.
2013, pp. 332–336. Hoboken, NJ, USA: Wiley, 2006.
[83] J. Boutros, E. Viterbo, C. Rastello, and J.-C. Belfiore, “Good lattice [107] E. Sula, J. Zhu, A. Pastore, S. H. Lim, and M. Gastpar, “Com-
constellations for both rayleigh fading and Gaussian channels,” IEEE pute–forward multiple access (CFMA): Practical implementations,”
Transactions on Information Theory, vol. 42, no. 2, pp. 502–518, 1996. IEEE Transactions on Communications, vol. 67, no. 2, pp. 1133–1147,
[84] M. Taherzadeh, H. Nikopour, A. Bayesteh, and H. Baligh, 2019.
“SCMA codebook design,” in IEEE Vehicular Technology Conference [108] M. Qiu, Y.-C. Huang, S.-L. Shieh, and J. Yuan, “A lattice-partition
(VTC2014-Fall), 2014, pp. 1–5. framework of downlink non-orthogonal multiple access without SIC,”
[85] M. Vameghestahbanati, I. D. Marsland, R. H. Gohary, and IEEE Trans. Commun., vol. 66, no. 6, pp. 2532–2546, Jun. 2018.
H. Yanikomeroglu, “Multidimensional constellations for uplink SCMA
[109] M. Qiu, L. Yang, Y. Xie, and J. Yuan, “On the design of multi-
systems—a comparative study,” IEEE Communications Surveys &
dimensional irregular repeat-accumulate lattice codes,” IEEE Trans-
Tutorials, vol. 21, no. 3, pp. 2169–2194, 2019.
actions on Communications, vol. 66, no. 2, pp. 478–492, 2018.
[86] S. Chaturvedi, Z. Liu, V. A. Bohara, A. Srivastava, and P. Xiao, “A
tutorial on decoding techniques of sparse code multiple access,” IEEE [110] M. Qiu, Y.-C. Huang, J. Yuan, and C.-L. Wang, “Lattice-partition-based
Access, vol. 10, pp. 58 503–58 524, 2022. downlink non-orthogonal multiple access without SIC for slow fading
[87] L. Yu, Z. Liu, M. Wen, D. Cai, S. Dang, Y. Wang, and P. Xiao, “Sparse channels,” IEEE Trans. Commun., vol. 67, no. 2, pp. 1166–1181, Feb.
code multiple access for 6g wireless communication networks: Recent 2019.
advances and future directions,” IEEE Communications Standards [111] M. Qiu, Y.-C. Huang, and J. Yuan, “Downlink non-orthogonal multiple
Magazine, vol. 5, no. 2, pp. 92–99, 2021. access without SIC for block fading channels: An algebraic rotation
[88] Z. Yuan, G. Yu, W. Li, Y. Yuan, X. Wang, and J. Xu, “Multi-user approach,” IEEE Trans. Wireless Commun., vol. 18, no. 8, pp. 3903–
shared access for internet of things,” in IEEE Vehicular Technology 3918, Aug. 2019.
Conference (VTC Spring), 2016, pp. 1–5. [112] ——, “Discrete signaling and treating interference as noise for the
[89] M. B. Shahab, R. Abbas, M. Shirvanimoghaddam, and S. J. Johnson, Gaussian interference channel,” IEEE Trans. Inf. Theory, vol. 67,
“Grant-free non-orthogonal multiple access for IoT: A survey,” IEEE no. 11, pp. 7253–7284, Nov. 2021.
Communications Surveys & Tutorials, vol. 22, no. 3, pp. 1805–1838, [113] ——, “Downlink transmission with heterogeneous URLLC services:
2020. Discrete signaling with single-user decoding,” IEEE Journal on Se-
[90] X. Dai, S. Chen, S. Sun, S. Kang, Y. Wang, Z. Shen, and J. Xu, “Suc- lected Areas in Communications, vol. 41, no. 7, pp. 2261–2277, 2023.
cessive interference cancelation amenable multiple access (SAMA) for [114] A. Jimenez Felstrom and K. Zigangirov, “Time-varying periodic con-
future wireless communications,” in IEEE International Conference on volutional codes with low-density parity-check matrix,” IEEE Trans-
Communication Systems, 2014, pp. 222–226. actions on Information Theory, vol. 45, no. 6, pp. 2181–2191, 1999.
[91] L. Ma, S. Tong, H. Zheng, B. Bai, and X. Dai, “Edgewise serial [115] M. Lentmaier, A. Sridharan, D. J. Costello, and K. S. Zigangirov,
message passing detection of uplink SCMA systems for better user “Iterative decoding threshold analysis for LDPC convolutional codes,”
fairness and faster convergence rate,” IEEE Wireless Communications IEEE Transactions on Information Theory, vol. 56, no. 10, pp. 5274–
Letters, vol. 8, no. 4, pp. 1285–1288, 2019. 5289, 2010.
 46

[116] S. Kudekar, T. J. Richardson, and R. L. Urbanke, “Threshold saturation [138] M. Moltafet, N. Mokari, M. R. Javan, H. Saeedi, and H. Pishro-Nik,
via spatial coupling: Why convolutional LDPC ensembles perform so “A new multiple access technique for 5G: Power domain sparse code
well over the BEC,” IEEE Transactions on Information Theory, vol. 57, multiple access (PSMA),” IEEE Access, vol. 6, pp. 747–759, 2018.
no. 2, pp. 803–834, 2011. [139] N. M. Balasubramanya, A. Gupta, and M. Sellathurai, “Combining
[117] S. Moloudi, M. Lentmaier, and A. Graell i Amat, “Spatially coupled code-domain and power-domain NOMA for supporting higher number
turbo-like codes,” IEEE Transactions on Information Theory, vol. 63, of users,” in IEEE Global Communications Conference (GLOBECOM),
no. 10, pp. 6199–6215, 2017. 2018, pp. 1–6.
[118] M. Qiu, X. Wu, J. Yuan, and A. Graell i Amat, “Generalized spatially- [140] S. Sharma, K. Deka, V. Bhatia, and A. Gupta, “Joint power-domain
coupled parallel concatenated codes with partial repetition,” IEEE and SCMA-based NOMA system for downlink in 5G and beyond,”
Transactions on Communications, vol. 70, no. 9, pp. 5771–5787, 2022. IEEE Communications Letters, vol. 23, no. 6, pp. 971–974, 2019.
[119] D. Truhachev, “Universal multiple access via spatially coupling data [141] M. T. P. Le, L. Sanguinetti, E. Björnson, and M.-G. D. Benedetto,
transmission,” in IEEE International Symposium on Information The- “Code-domain NOMA in massive MIMO: When is it needed?” IEEE
ory, 2013, pp. 1884–1888. Transactions on Vehicular Technology, vol. 70, no. 5, pp. 4709–4723,
[120] D. Truhachev and C. Schlegel, “Coupling data transmission for 2021.
multiple-access communications,” IEEE Transactions on Information [142] M. Jia, X. Zhang, J. Sun, X. Gu, and Q. Guo, “Intelligent resource
Theory, vol. 65, no. 7, pp. 4550–4574, 2019. management for satellite and terrestrial spectrum shared networking
[121] A. Karami and D. Truhachev, “Coupling information transmission with toward b5g,” IEEE Wireless Communications, vol. 27, no. 1, pp. 54–
window decoding,” IEEE Wireless Communications Letters, vol. 8, 61, 2020.
no. 2, pp. 560–563, 2019. [143] M. Hmila, M. Fernández-Veiga, M. Rodrı́guez-Pérez, and S. Herrerı́a-
[122] ——, “Self-coupling data transmission for random multiple access Alonso, “Energy efficient power and channel allocation in underlay
communications,” IEEE Communications Letters, vol. 24, no. 6, pp. device to multi device communications,” IEEE Transactions on Com-
1168–1172, 2020. munications, vol. 67, no. 8, pp. 5817–5832, 2019.
[123] L. Zhang, W. Li, Y. Wu, X. Wang, S.-I. Park, H. M. Kim, J.-Y. Lee, [144] J. Cui, Z. Ding, P. Fan, and N. Al-Dhahir, “Unsupervised machine
P. Angueira, and J. Montalban, “Layered-division-multiplexing: Theory learning-based user clustering in millimeter-wave-noma systems,”
and practice,” IEEE Transactions on Broadcasting, vol. 62, no. 1, pp. IEEE Transactions on Wireless Communications, vol. 17, no. 11, pp.
216–232, 2016. 7425–7440, 2018.
[124] L. Zhang, Y. Wu, W. Li, B. Rong, K. Salehian, S. Lafleche, X. Wang, [145] P. M. Santos, B. R. Manoj, M. Sadeghi, and E. G. Larsson, “Universal
S. I. Park, H. M. Kim, J.-y. Lee, N. Hur, P. Angueira, and J. Montal- adversarial attacks on neural networks for power allocation in a massive
ban, “Layered-division-multiplexing for high spectrum efficiency and mimo system,” IEEE Wireless Communications Letters, vol. 11, no. 1,
service flexibility in next generation ATSC 3.0 broadcast system,” IEEE pp. 67–71, 2022.
Wireless Communications, vol. 26, no. 2, pp. 116–123, 2019. [146] H. Sun, X. Chen, Q. Shi, M. Hong, X. Fu, and N. D. Sidiropoulos,
[125] L. Zhang, Y. Wu, W. Li, K. Salehian, S. Lafleche, X. Wang, S. I. Park, “Learning to optimize: Training deep neural networks for interference
H. M. Kim, J.-y. Lee, N. Hur, P. Angueira, and J. Montalban, “Layered- management,” IEEE Transactions on Signal Processing, vol. 66, no. 20,
division multiplexing: An enabling technology for multicast/broadcast pp. 5438–5453, 2018.
service delivery in 5G,” IEEE Communications Magazine, vol. 56, [147] J. Kim, H. Lee, S.-E. Hong, and S.-H. Park, “Deep learning methods
no. 3, pp. 82–90, 2018. for universal MISO beamforming,” IEEE Wireless Communications
[126] E. Basar, u. Aygulu, E. Panayırcı, and H. V. Poor, “Orthogonal Letters, vol. 9, no. 11, pp. 1894–1898, 2020.
frequency division multiplexing with index modulation,” IEEE Trans- [148] W. Lee, M. Kim, and D.-H. Cho, “Deep power control: Transmit
actions on Signal Processing, vol. 61, no. 22, pp. 5536–5549, 2013. power control scheme based on convolutional neural network,” IEEE
[127] Y.-T. Lai, Y.-R. Ciou, and J.-M. Wu, “Index modulation multiple Communications Letters, vol. 22, no. 6, pp. 1276–1279, 2018.
access,” in IEEE Annual International Symposium on Personal, Indoor [149] H. Huang, Y. Peng, J. Yang, W. Xia, and G. Gui, “Fast beamforming
and Mobile Radio Communications (PIMRC), 2018, pp. 1–5. design via deep learning,” IEEE Transactions on Vehicular Technology,
[128] J. Li, S. Dang, M. Wen, Q. Li, Y. Chen, Y. Huang, and W. Shang, vol. 69, no. 1, pp. 1065–1069, 2020.
“Index modulation multiple access for 6G communications: Principles, [150] A. B. M. Adam, L. Lei, S. Chatzinotas, and N. U. R. Junejo, “Deep
applications, and challenges,” IEEE Network, vol. 37, no. 1, pp. 52–60, convolutional self-attention network for energy-efficient power control
2023. in NOMA networks,” IEEE Transactions on Vehicular Technology,
[129] R. Hadani, S. Rakib, M. Tsatsanis, A. Monk, A. J. Goldsmith, A. F. vol. 71, no. 5, pp. 5540–5545, 2022.
Molisch, and R. Calderbank, “Orthogonal time frequency space modu- [151] M. Chen, W. Saad, and C. Yin, “Virtual reality over wireless networks:
lation,” in IEEE Wireless Communications and Networking Conference Quality-of-service model and learning-based resource management,”
(WCNC), 2017, pp. 1–6. IEEE Transactions on Communications, vol. 66, no. 11, pp. 5621–
[130] Z. Wei, W. Yuan, S. Li, J. Yuan, G. Bharatula, R. Hadani, and 5635, 2018.
L. Hanzo, “Orthogonal time-frequency space modulation: A promising [152] K. He, X. Chen, Q. Wu, S. Yu, and Z. Zhou, “Graph attention spatial-
next-generation waveform,” IEEE Wireless Communications, vol. 28, temporal network with collaborative global-local learning for citywide
no. 4, pp. 136–144, 2021. mobile traffic prediction,” IEEE Transactions on Mobile Computing,
[131] Y. Hong, T. Thaj, and E. Viterbo, Delay-Doppler Communicationss: vol. 21, no. 4, pp. 1244–1256, 2022.
Principles and Applications. London, U.K.: Elsevier Academic Press, [153] R. Li, C. Wang, Z. Zhao, R. Guo, and H. Zhang, “The LSTM-based
2022. advantage actor-critic learning for resource management in network
[132] H. Lin and J. Yuan, “Orthogonal delay-doppler division multiplexing slicing with user mobility,” IEEE Communications Letters, vol. 24,
modulation,” IEEE Transactions on Wireless Communications, vol. 21, no. 9, pp. 2005–2009, 2020.
no. 12, pp. 11 024–11 037, 2022. [154] J.-Y. Xia, S. Li, J.-J. Huang, Z. Yang, I. M. Jaimoukha, and D. Gündüz,
[133] K.-K. Wong and K.-F. Tong, “Fluid antenna multiple access,” IEEE “Metalearning-based alternating minimization algorithm for nonconvex
Transactions on Wireless Communications, vol. 21, no. 7, pp. 4801– optimization,” IEEE Transactions on Neural Networks and Learning
4815, 2022. Systems, vol. 34, no. 9, pp. 5366–5380, 2023.
[134] Z. Wu and L. Dai, “Multiple access for near-field communications: [155] Y. Lu, L. Liu, J. Panneerselvam, B. Yuan, J. Gu, and N. Antonopoulos,
Sdma or ldma?” IEEE Journal on Selected Areas in Communications, “A GRU-based prediction framework for intelligent resource manage-
vol. 41, no. 6, pp. 1918–1935, 2023. ment at cloud data centres in the age of 5G,” IEEE Transactions on
[135] M. M. Şahin, O. Dizdar, B. Clerckx, and H. Arslan, “Multicarrier rate- Cognitive Communications and Networking, vol. 6, no. 2, pp. 486–498,
splitting multiple access: Superiority of ofdm-rsma over ofdma and 2020.
ofdm-noma,” IEEE Communications Letters, vol. 27, no. 11, pp. 3088– [156] M. Lee, G. Yu, and G. Y. Li, “Graph embedding-based wireless link
3092, 2023. scheduling with few training samples,” IEEE Transactions on Wireless
[136] M. M. Sahin, O. Dizdar, B. Clerckx, and H. Arslan, “Ofdm-rsma: Ro- Communications, vol. 20, no. 4, pp. 2282–2294, 2021.
bust transmission under inter-carrier interference,” arXiv 2310.01686, [157] Y. Shen, Y. Shi, J. Zhang, and K. B. Letaief, “Graph neural networks
2023. for scalable radio resource management: Architecture design and theo-
[137] C. Xu, B. Clerckx, S. Chen, Y. Mao, and J. Zhang, “Rate-splitting retical analysis,” IEEE Journal on Selected Areas in Communications,
multiple access for multi-antenna joint radar and communications,” vol. 39, no. 1, pp. 101–115, 2021.
IEEE Journal of Selected Topics in Signal Processing, vol. 15, no. 6, [158] M. Lee, S. Hosseinalipour, C. G. Brinton, G. Yu, and H. Dai, “A fast
pp. 1332–1347, 2021. graph neural network-based method for winner determination in multi-
 47

unit combinatorial auctions,” IEEE Transactions on Cloud Computing, multi-agent DRL approach,” IEEE Transactions on Network Science
vol. 10, no. 4, pp. 2264–2280, 2022. and Engineering, vol. 9, no. 6, pp. 4426–4443, 2022.
[159] M. Lee, G. Yu, H. Dai, and G. Y. Li, “Graph neural networks [179] M. Chen, W. Saad, and C. Yin, “Echo state networks for self-organizing
meet wireless communications: Motivation, applications, and future resource allocation in lte-u with uplink–downlink decoupling,” IEEE
directions,” IEEE Wireless Communications, vol. 29, no. 5, pp. 12– Transactions on Wireless Communications, vol. 16, no. 1, pp. 3–16,
19, 2022. 2017.
[160] Z. Gao, L. Yang, and Y. Dai, “Fast adaptive task offloading and resource [180] J.-B. Wang, J. Wang, Y. Wu, J.-Y. Wang, H. Zhu, M. Lin, and J. Wang,
allocation via multiagent reinforcement learning in heterogeneous “A machine learning framework for resource allocation assisted by
vehicular fog computing,” IEEE Internet of Things Journal, vol. 10, cloud computing,” IEEE Network, vol. 32, no. 2, pp. 144–151, 2018.
no. 8, pp. 6818–6835, 2023. [181] W. V. Mauricio, D. C. Araujo, F. Hugo Neto, F. Rafael Lima, and T. F.
[161] Z. Ke and H. Vikalo, “Real-time radio technology and modulation Maciel, “A low complexity solution for resource allocation and SDMA
classification via an lstm auto-encoder,” IEEE Transactions on Wireless grouping in massive MIMO systems,” in 15th International Symposium
Communications, vol. 21, no. 1, pp. 370–382, 2022. on Wireless Communication Systems (ISWCS), 2018, pp. 1–6.
[162] Y. Yuan, G. Zheng, K.-K. Wong, B. Ottersten, and Z.-Q. Luo, “Transfer [182] A. M. Elbir, “A deep learning framework for hybrid beamforming
learning and meta learning-based fast downlink beamforming adapta- without instantaneous csi feedback,” IEEE Transactions on Vehicular
tion,” IEEE Transactions on Wireless Communications, vol. 20, no. 3, Technology, vol. 69, no. 10, pp. 11 743–11 755, 2020.
pp. 1742–1755, 2021. [183] W. Xia, G. Zheng, Y. Zhu, J. Zhang, J. Wang, and A. P. Petropulu,
[163] R. Loli Cerna and B. Clerckx, “A meta-learning based pre- “A deep learning framework for optimization of miso downlink beam-
coder optimization framework for rate-splitting multiple access,” forming,” IEEE Transactions on Communications, vol. 68, no. 3, pp.
arXiv:2307.08822, 2023. 1866–1880, 2020.
[164] Q. Hu, Y. Cai, Q. Shi, K. Xu, G. Yu, and Z. Ding, “Iterative [184] Z. Zhang, M. Tao, and Y.-F. Liu, “Learning to beamform in joint mul-
algorithm induced deep-unfolding neural networks: Precoding design ticast and unicast transmission with imperfect CSI,” IEEE Transactions
for multiuser MIMO systems,” IEEE Transactions on Wireless Com- on Communications, vol. 71, no. 5, pp. 2711–2723, 2023.
munications, vol. 20, no. 2, pp. 1394–1410, 2021. [185] J. Kim, H. Lee, S.-E. Hong, and S.-H. Park, “A bipartite graph
[165] C. Xu, Y. Jia, S. He, Y. Huang, and D. Niyato, “Joint user scheduling, neural network approach for scalable beamforming optimization,” IEEE
base station clustering and beamforming design based on deep unfold- Transactions on Wireless Communications, vol. 22, no. 1, pp. 333–347,
ing technique,” IEEE Transactions on Communications, pp. 1–1, 2023. 2023.
[166] M. Lee, G. Yu, and G. Y. Li, “Learning to branch: Accelerating resource [186] S. Liu, J. Guo, and C. Yang, “Multidimensional graph neural networks
allocation in wireless networks,” IEEE Transactions on Vehicular for wireless communications,” IEEE Transactions on Wireless Commu-
Technology, vol. 69, no. 1, pp. 958–970, 2020. nications, pp. 1–1, 2023.
[167] Z. Xiong, Y. Zhang, D. Niyato, R. Deng, P. Wang, and L.-C. Wang, [187] X. Xu, Y. Liu, Q. Chen, X. Mu, and Z. Ding, “Distributed auto-learning
“Deep reinforcement learning for mobile 5g and beyond: Funda- gnn for multi-cell cluster-free NOMA communications,” IEEE Journal
mentals, applications, and challenges,” IEEE Vehicular Technology on Selected Areas in Communications, vol. 41, no. 4, pp. 1243–1258,
Magazine, vol. 14, no. 2, pp. 44–52, 2019. 2023.
[168] F. B. Mismar, B. L. Evans, and A. Alkhateeb, “Deep reinforcement
[188] X. Gao, Y. Liu, X. Liu, and L. Song, “Machine learning empowered
learning for 5G networks: Joint beamforming, power control, and
resource allocation in IRS aided MISO-NOMA networks,” IEEE Trans-
interference coordination,” IEEE Transactions on Communications,
actions on Wireless Communications, vol. 21, no. 5, pp. 3478–3492,
vol. 68, no. 3, pp. 1581–1592, 2020.
2022.
[169] C. He, Y. Hu, Y. Chen, and B. Zeng, “Joint power allocation and
[189] J.-M. Kang, I.-M. Kim, and C.-J. Chun, “Deep learning-based MIMO-
channel assignment for NOMA with deep reinforcement learning,”
NOMA with imperfect SIC decoding,” IEEE Systems Journal, vol. 14,
IEEE Journal on Selected Areas in Communications, vol. 37, no. 10,
no. 3, pp. 3414–3417, 2020.
pp. 2200–2210, 2019.
[190] M. Wu, Z. Wan, Y. Wang, S. Liu, and Z. Gao, “Deep learning-
[170] N. Q. Hieu, D. T. Hoang, D. Niyato, and D. I. Kim, “Optimal power
based rate-splitting multiple access for massive MIMO-OFDM systems
allocation for rate splitting communications with deep reinforcement
with imperfect CSIT,” in 2022 International Symposium on Wireless
learning,” IEEE Wireless Communications Letters, vol. 10, no. 12, pp.
Communication Systems (ISWCS), 2022, pp. 1–6.
2820–2823, 2021.
[171] J. Ji, L. Cai, K. Zhu, and D. Niyato, “Decoupled association with rate [191] M. R. Camana, C. E. Garcia, and I. Koo, “Deep learning-assisted power
splitting multiple access in UAV-assisted cellular networks using multi- minimization in underlay MISO-SWIPT systems based on rate-splitting
agent deep reinforcement learning,” IEEE Transactions on Mobile multiple access,” IEEE Access, vol. 10, pp. 62 137–62 156, 2022.
Computing, pp. 1–15, 2023. [192] N. Q. Hieu, D. T. Hoang, D. Niyato, and D. I. Kim, “Optimal power
[172] X. Wang, G. Sun, Y. Xin, T. Liu, and Y. Xu, “Deep transfer allocation for rate splitting communications with deep reinforcement
reinforcement learning for beamforming and resource allocation in learning,” IEEE Wireless Communications Letters, vol. 10, no. 12, pp.
multi-cell MISO-OFDMA systems,” IEEE Transactions on Signal and 2820–2823, 2021.
Information Processing over Networks, vol. 8, pp. 815–829, 2022. [193] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski,
[173] T. Zhang, S. Wang, Y. Zhuang, C. You, M. Wen, and Y.-C. Wu, “Five disruptive technology directions for 5G,” IEEE Communications
“Reconfigurable intelligent surface assisted ofdm relaying: Subcarrier Magazine, vol. 52, no. 2, pp. 74–80, 2014.
matching with balanced SNR,” IEEE Transactions on Vehicular Tech- [194] D. J. Love, R. W. Heath, V. K. N. Lau, D. Gesbert, B. D. Rao, and
nology, vol. 72, no. 2, pp. 2216–2230, 2023. M. Andrews, “An overview of limited feedback in wireless communi-
[174] J. Zhou, S. Dang, B. Shihada, and M.-S. Alouini, “Power allocation cation systems,” IEEE Journal on Selected Areas in Communications,
for relayed OFDM with index modulation assisted by artificial neural vol. 26, no. 8, pp. 1341–1365, 2008.
network,” IEEE Wireless Communications Letters, vol. 10, no. 2, pp. [195] H. Xie, F. Gao, and S. Jin, “An overview of low-rank channel estimation
373–377, 2021. for massive MIMO systems,” IEEE Access, vol. 4, pp. 7313–7321,
[175] J. Feng, S. Han, H. Chen, and C. Yang, “Uplink scheduling for MIMO- 2016.
OFDMA systems with rate constraints by deep learning,” in IEEE [196] D. Gündüz, P. de Kerret, N. D. Sidiropoulos, D. Gesbert, C. R. Murthy,
Wireless Communications and Networking Conference (WCNC), 2023, and M. van der Schaar, “Machine learning in the air,” IEEE Journal
pp. 1–6. on Selected Areas in Communications, vol. 37, no. 10, pp. 2184–2199,
[176] Y. Hu, M. Chen, Z. Yang, M. Chen, and G. Jia, “Optimization of 2019.
resource allocation in multi-cell OFDM systems: A distributed rein- [197] G. Gui, H. Huang, Y. Song, and H. Sari, “Deep learning for an
forcement learning approach,” in 2020 IEEE 31st Annual International effective nonorthogonal multiple access scheme,” IEEE Transactions
Symposium on Personal, Indoor and Mobile Radio Communications, on Vehicular Technology, vol. 67, no. 9, pp. 8440–8450, 2018.
2020, pp. 1–6. [198] H. Ye, G. Y. Li, and B.-H. Juang, “Power of deep learning for channel
[177] Y. Zhou, Z. M. Fadlullah, B. Mao, and N. Kato, “A deep-learning-based estimation and signal detection in OFDM systems,” IEEE Wireless
radio resource assignment technique for 5G ultra dense networks,” Communications Letters, vol. 7, no. 1, pp. 114–117, 2018.
IEEE Network, vol. 32, no. 6, pp. 28–34, 2018. [199] M. Soltani, V. Pourahmadi, A. Mirzaei, and H. Sheikhzadeh, “Deep
[178] S. Zhang, Z. Ni, L. Kuang, C. Jiang, and X. Zhao, “Load-aware learning-based channel estimation,” IEEE Communications Letters,
distributed resource allocation for MF-TDMA ad hoc networks: A vol. 23, no. 4, pp. 652–655, 2019.
 48

[200] M. S. Safari, V. Pourahmadi, and S. Sodagari, “Deep UL2DL: Data- [222] R. C. Loli, O. Dizdar, B. Clerckx, and C. Ling, “Model-based deep
driven channel knowledge transfer from uplink to downlink,” IEEE learning receiver design for rate-splitting multiple access,” IEEE Trans-
Open Journal of Vehicular Technology, vol. 1, pp. 29–44, 2020. actions on Wireless Communications, pp. 1–1, 2023.
[201] H. Ye, L. Liang, G. Y. Li, and B.-H. Juang, “Deep learning-based [223] A. Balachandran, G. M. Voelker, P. Bahl, and P. V. Rangan, “Charac-
end-to-end wireless communication systems with conditional GANs as terizing user behavior and network performance in a public wireless
unknown channels,” IEEE Transactions on Wireless Communications, LAN,” in Proc. ACM SIGMETRICS Int. Conf. Measurement on Model-
vol. 19, no. 5, pp. 3133–3143, 2020. ing of Computer Syst., Las Vegas, NV, USA, June 2002, pp. 195–205.
[202] B. Ozpoyraz, A. T. Dogukan, Y. Gevez, U. Altun, and E. Basar, [224] Z. Yang, M. Chen, K.-K. Wong, H. V. Poor, and S. Cui, “Federated
“Deep learning-aided 6G wireless networks: A comprehensive survey learning for 6G: Applications, challenges, and opportunities,” Engi-
of revolutionary PHY architectures,” IEEE Open Journal of the Com- neering, vol. 8, pp. 33–41, Jan. 2022.
munications Society, vol. 3, pp. 1749–1809, 2022. [225] A. Agarwal, S. Dubey, M. A. Khan, R. Gangopadhyay, and S. Debnath,
[203] C.-J. Chun, J.-M. Kang, and I.-M. Kim, “Deep learning-based joint “Learning based primary user activity prediction in cognitive radio
pilot design and channel estimation for multiuser mimo channels,” networks for efficient dynamic spectrum access,” in Proc. Int. Conf.
IEEE Communications Letters, vol. 23, no. 11, pp. 1999–2003, 2019. Signal Process. Commun. (SPCOM), Bangalore, India, Nov. 2016, pp.
[204] X. Ma and Z. Gao, “Data-driven deep learning to design pilot and 1–5.
channel estimator for massive MIMO,” IEEE Transactions on Vehicular [226] D. Roy, T. Mukherjee, M. Chatterjee, and E. Pasiliao, “Primary user
Technology, vol. 69, no. 5, pp. 5677–5682, 2020. activity prediction in dsa networks using recurrent structures,” in
[205] X. Wei, C. Hu, and L. Dai, “Deep learning for beamspace channel Proc. IEEE Int. Symp. Dynamic Spectrum Access Networks (DySPAN),
estimation in millimeter-wave massive MIMO systems,” IEEE Trans- Newark, NJ, USA, Dec. 2019, pp. 1–10.
actions on Communications, vol. 69, no. 1, pp. 182–193, 2021. [227] J. Xiao, C. Chen, C. Hou, and X. Yuan, “Short-term user activity
[206] H. Yu, Z. Fei, Z. Zheng, N. Ye, and Z. Han, “Deep learning-based prediction with massive mobile broadband data,” in Proc. IEEE Int.
user activity detection and channel estimation in grant-free NOMA,” Conf. Mobile Data Management (MDM), Daejeon, Korea (South), June
IEEE Transactions on Wireless Communications, vol. 22, no. 4, pp. 2017, pp. 348–351.
2202–2214, 2023. [228] M. Bkassiny, Y. Li, and S. K. Jayaweera, “A survey on machine-
[207] X. Rao and V. K. N. Lau, “Distributed compressive CSIT estimation learning techniques in cognitive radios,” IEEE Commun. Surveys &
and feedback for FDD multi-user massive MIMO systems,” IEEE Tut., vol. 15, no. 3, pp. 1136–1159, Third Quarter, 2012.
Transactions on Signal Processing, vol. 62, no. 12, pp. 3261–3271, [229] Y. Xue, C. Tang, F. Tang, Y. Yang, J. Li, M. Guo, and J. Wu, “Primary
2014. user activity prediction based joint topology control and stable routing
[208] M. S. Sim, J. Park, C.-B. Chae, and R. W. Heath, “Compressed in mobile cognitive networks,” in 2016 IEEE Wireless Communications
channel feedback for correlated massive MIMO systems,” Journal of and Networking Conference, Doha, Qatar, Apr. 2016, pp. 1–6.
Communications and Networks, vol. 18, no. 1, pp. 95–104, 2016. [230] C. Luo, J. Zeng, M. Yuan, W. Dai, and Q. Yang, “Telco user activity
[209] J. Ballé, V. Laparra, and E. P. Simoncelli, “End-to-end optimized level prediction with massive mobile broadband data,” ACM Trans.
image compression,” in 5th International Conference on Learning Intelligent Syst. Technology (TIST), vol. 7, no. 4, pp. 1–30, May 2016.
Representations (ICLR), 2017. [231] N. U. Khan, W. Wan, R. Riaz, S. Jiang, and X. Wang,
[210] S. Lombardo, J. Han, C. Schroers, and S. Mandt, “Deep generative “Prediction and classification of user activities using machine
video compression,” Advances in Neural Information Processing Sys- learning models from location-based social network data,” Applied
tems, vol. 32, 2019. Sciences, vol. 13, no. 6, Mar. 2023. [Online]. Available: https:
[211] C.-K. Wen, W.-T. Shih, and S. Jin, “Deep learning for massive MIMO //www.mdpi.com/2076-3417/13/6/3517
CSI feedback,” IEEE Wireless Communications Letters, vol. 7, no. 5, [232] G. Gui, H. Huang, Y. Song, and H. Sari, “Deep learning for an effective
pp. 748–751, 2018. nonorthogonal multiple access scheme,” IEEE Trans. Veh. Technol.,
[212] T. Wang, C.-K. Wen, S. Jin, and G. Y. Li, “Deep learning-based CSI vol. 67, no. 9, pp. 8440–8450, Sept. 2018.
feedback approach for time-varying massive MIMO channels,” IEEE [233] S. Song, Z. Fang, Z. Zhang, C.-L. Chen, and H. Sun, “Semi-online
Wireless Communications Letters, vol. 8, no. 2, pp. 416–419, 2019. computational offloading by dueling deep-q network for user behavior
[213] M. B. Mashhadi, Q. Yang, and D. Gündüz, “Distributed deep con- prediction,” IEEE Access, vol. 8, pp. 118 192–118 204, June 2020.
volutional compression for massive MIMO CSI feedback,” IEEE [234] W. Jiang, “Cellular traffic prediction with machine learning: A survey,”
Transactions on Wireless Communications, vol. 20, no. 4, pp. 2621– Expert Systems with Applications, vol. 201, p. 117163, Sept. 2022.
2633, 2021. [235] J. Zhang, F. Ye, and Y. Qian, “Intelligent and application-aware network
[214] X. Ma, Z. Gao, F. Gao, and M. Di Renzo, “Model-driven deep traffic prediction in smart access gateways,” IEEE Network, vol. 34,
learning based channel estimation and feedback for millimeter-wave no. 3, pp. 264–269, Apr. 2020.
massive hybrid MIMO systems,” IEEE Journal on Selected Areas in [236] Y. Li, X. Sun, H. Zhang, Z. Li, L. Qin, C. Sun, and Z. Ji, “Cellular
Communications, vol. 39, no. 8, pp. 2388–2406, 2021. traffic prediction via a deep multi-reservoir regression learning network
[215] Z. Yin, W. Xu, R. Xie, S. Zhang, D. W. K. Ng, and X. You, “Deep for multi-access edge computing,” IEEE Wireless Commun., vol. 28,
CSI compression for massive MIMO: A self-information model-driven no. 5, pp. 13–19, Oct. 2021.
neural network,” IEEE Transactions on Wireless Communications, [237] A. D. Shoaei, D. T. Nguyen, and T. Le-Ngoc, “Traffic prediction for
vol. 21, no. 10, pp. 8872–8886, 2022. reconfigurable access scheme in correlated traffic mtc networks,” in
[216] X. Liang, H. Chang, H. Li, X. Gu, and L. Zhang, “Changeable rate Proc. Annual Int. Symposium Personal, Indoor Mobile Radio Commun.
and novel quantization for CSI feedback based on deep learning,” IEEE (PIMRC), Helsinki, Finland, Sept. 2021, pp. 953–958.
Transactions on Wireless Communications, vol. 21, no. 12, pp. 10 100– [238] N. Jiang, Y. Deng, and A. Nallanathan, “Traffic prediction and ran-
10 114, 2022. dom access control optimization: Learning and non-learning-based
[217] S. Zhang, W. Xu, S. Jin, X. You, D. W. K. Ng, and L.-C. Wang, approaches,” IEEE Commun. Magazine, vol. 59, no. 3, pp. 16–22, Mar.
“Dual-propagation-feature fusion enhanced neural CSI compression 2021.
for massive MIMO,” IEEE Transactions on Communications, vol. 71, [239] Q. Zhang, M. Mozaffari, W. Saad, M. Bennis, and M. Debbah,
no. 9, pp. 5182–5198, 2023. “Machine learning for predictive on-demand deployment of uavs for
[218] N. Shlezinger, J. Whang, Y. C. Eldar, and A. G. Dimakis, “Model- wireless communications,” in Proc. IEEE Global Commun. Conf.
based deep learning: Key approaches and design guidelines,” in 2021 (GLOBECOM), Abu Dhabi, United Arab Emirates, Dec. 2018, pp. 1–6.
IEEE Data Science and Learning Workshop (DSLW), 2021, pp. 1–6. [240] L. Lu, Y. Hu, Y. Zhang, G. Jia, J. Nie, and M. Shikh-Bahaei, “Machine
[219] H. Huang, S. Guo, G. Gui, Z. Yang, J. Zhang, H. Sari, and F. Adachi, learning for predictive deployment of UAVs with rate splitting multiple
“Deep learning for physical-layer 5g wireless techniques: Opportuni- access,” in Proc. IEEE Global Commun. Conf. Workshops (GC Wkshps,
ties, challenges and solutions,” IEEE Wireless Communications, vol. 27, Taipei, Taiwan, Dec. 2020, pp. 1–6.
no. 1, pp. 214–222, 2020. [241] F. Guo, L. Lu, Z. Zang, and M. Shikh-Bahaei, “Machine learning
[220] M. A. Aref and S. K. Jayaweera, “Deep learning-aided successive for predictive deployment of UAVs with multiple access,” IEEE Open
interference cancellation for mimo-noma,” in GLOBECOM 2020 - 2020 Journal of the Communications Society, vol. 4, pp. 908–921, Apr. 2023.
IEEE Global Communications Conference, 2020, pp. 1–5. [242] H. Feng, Y. Jiang, D. Niyato, F.-C. Zheng, and X. You, “Content
[221] J.-M. Kang, I.-M. Kim, and C.-J. Chun, “Deep learning-based mimo- popularity prediction via deep learning in cache-enabled fog radio ac-
noma with imperfect sic decoding,” IEEE Systems Journal, vol. 14, cess networks,” in Proc. IEEE Global Commun. Conf. (GLOBECOM).
no. 3, pp. 3414–3417, 2020. Waikoloa, HI, USA: IEEE, December 2019, pp. 1–6.
 49

[243] J. Liang, D. Zhu, H. Liu, H. Ping, T. Li, H. Zhang, L. Geng, and [263] W. Shi, S. Zhou, Z. Niu, M. Jiang, and L. Geng, “Joint device
Y. Liu, “Multi-head attention based popularity prediction caching in scheduling and resource allocation for latency constrained wireless
social content-centric networking with mobile edge computing,” IEEE federated learning,” IEEE Transactions on Wireless Communications,
Commun. Lett., vol. 25, no. 2, pp. 508–512, Oct. 2021. vol. 20, no. 1, pp. 453–467, Jan. 2021.
[244] S. Yan, L. Qi, Y. Zhou, M. Peng, and G. S. Rahman, “Joint user access [264] W. Xia, T. Q. S. Quek, K. Guo, W. Wen, H. H. Yang, and H. Zhu,
mode selection and content popularity prediction in non-orthogonal “Multi-armed bandit-based client scheduling for federated learning,”
multiple access-based F-RANs,” IEEE Trans. Commun., vol. 68, no. 1, IEEE Transactions on Wireless Communications, vol. 19, no. 11, pp.
pp. 654–666, Jan. 2019. 7108–7123, Nov. 2020.
[245] H. Zhang and L. Dai, “Mobility prediction: A survey on state-of-the-art [265] Z. Wang, Y. Zhou, Y. Zou, Q. An, Y. Shi, and M. Bennis, “A graph
schemes and future applications,” IEEE Access, vol. 7, pp. 802–822, neural network learning approach to optimize ris-assisted federated
Dec. 2019. learning,” IEEE Transactions on Wireless Communications, vol. 22,
[246] J. Jeong, D. Roeland, J. Derehag, Å. A. Johansson, V. Umaashankar, no. 9, pp. 6092–6106, 2023.
G. Sun, and G. Eriksson, “Mobility prediction for 5g core networks,” [266] J. Mao and A. Yener, “Roar-fed: Ris-assisted over-the-air adap-
IEEE Commun. Standards Magazine, vol. 5, no. 1, pp. 56–61, Dec. tive resource allocation for federated learning,” arXiv preprint
2021. arXiv:2309.09883, 2023.
[247] U. Fattore, M. Liebsch, B. Brik, and A. Ksentini, “AutoMEC: LSTM- [267] A. M. Elbir and S. Coleri, “Federated learning for channel estimation
based user mobility prediction for service management in distributed in conventional and ris-assisted massive mimo,” IEEE Transactions on
MEC resources,” in Pro. Int. ACM Conf. Modeling, Analysis and Wireless Communications, vol. 21, no. 6, pp. 4255–4268, 2022.
Simulation of Wireless and Mobile Syst., Nov. 2020, pp. 155–159. [268] Y. Oh, Y.-S. Jeon, M. Chen, and W. Saad, “Fedvqcs: Federated
[248] F. Sun, P. Wang, J. Zhao, N. Xu, J. Zeng, J. Tao, K. Song, C. Deng, learning via vector quantized compressed sensing,” IEEE Transactions
J. C. Lui, and X. Guan, “Mobile data traffic prediction by exploiting on Wireless Communications, pp. 1–1, 2023.
time-evolving user mobility patterns,” IEEE Trans. Mobile Comput., [269] T. T. Vu, D. T. Ngo, N. H. Tran, H. Q. Ngo, M. N. Dao, and R. H.
vol. 21, no. 12, pp. 4456–4470, Dec. 2022. Middleton, “Cell-free massive MIMO for wireless federated learning,”
[249] Z. Cheng, N. Chen, B. Liu, Z. Gao, L. Huang, X. Du, and M. Guizani, IEEE Transactions on Wireless Communications, vol. 19, no. 10, pp.
“Joint user association and resource allocation in hetnets based on user 6377–6392, Oct. 2020.
mobility prediction,” Computer Networks, vol. 177, p. 107312, Aug. [270] Y.-S. Jeon, M. Mohammadi Amiri, and N. Lee, “Communication-
2020. efficient federated learning over mimo multiple access channels,” IEEE
[250] E. Chaalal, S.-M. Senouci, and L. Reynaud, “A new framework for Transactions on Communications, vol. 70, no. 10, pp. 6547–6562, Oct.
multi-hop abs-assisted 5g-networks with users’ mobility prediction,” 2022.
IEEE Trans. Veh. Technol., vol. 71, no. 4, pp. 4412–4427, Apr. 2022. [271] X. Wei, C. Shen, J. Yang, and H. Vincent Poor, “Random orthogo-
[251] J. Ding, H. Liu, L. T. Yang, T. Yao, and W. Zuo, “Multiuser multivariate nalization for federated learning in massive MIMO systems,” IEEE
multiorder markov-based multimodal user mobility pattern prediction,” Transactions on Wireless Communications, to appear, 2023.
IEEE Internet Things J., vol. 7, no. 5, pp. 4519–4531, May 2020. [272] T. Huang, B. Ye, Z. Qu, B. Tang, L. Xie, and S. Lu, “Physical-layer
arithmetic for federated learning in uplink MU-MIMO enabled wireless
[252] M. Chen, N. Shlezinger, H. V. Poor, Y. C. Eldar, and S. Cui, “Com-
networks,” in Proc. IEEE Conference on Computer Communications,
munication efficient federated learning,” Proceedings of the National
Toronto, ON, Canada, July 2020, pp. 1221–1230.
Academy of Sciences of the United States of America, vol. 118, no. 17,
[273] H. Sun, X. Ma, and R. Q. Hu, “Adaptive federated learning with gra-
Apr. 2021.
dient compression in uplink NOMA,” IEEE Transactions on Vehicular
[253] J. Chen, X. Pan, R. Monga, S. Bengio, and R. Jozefowicz, “Revisiting
Technology, vol. 69, no. 12, pp. 16 325–16 329, Dec. 2020.
distributed synchronous SGD,” [Online] Available: https:// arxiv.org/
[274] D. Liu and O. Simeone, “Privacy for free: Wireless federated learning
abs/ 1604.00981, 2017.
via uncoded transmission with adaptive power control,” IEEE Journal
[254] M. Chen, Z. Yang, W. Saad, C. Yin, H. V. Poor, and S. Cui, “A joint on Selected Areas in Communications, vol. 39, no. 1, pp. 170–185,
learning and communications framework for federated learning over Jan. 2021.
wireless networks,” IEEE Transactions on Wireless Communications, [275] M. S. Al-Abiad, M. Z. Hassan, and M. J. Hossain, “Energy-efficient
vol. 20, no. 1, pp. 269–283, Jan. 2021. resource allocation for federated learning in NOMA-enabled and relay-
[255] Z. Yang, M. Chen, W. Saad, C. S. Hong, and M. Shikh-Bahaei, “Energy assisted Internet of Things networks,” IEEE Internet of Things Journal,
efficient federated learning over wireless communication networks,” vol. 9, no. 24, pp. 24 736–24 753, Dec. 2022.
IEEE Transactions on Wireless Communications, vol. 20, no. 3, pp. [276] Y. Wu, Y. Song, T. Wang, L. Qian, and T. Q. S. Quek, “Non-orthogonal
1935–1949, Mar. 2021. multiple access assisted federated learning via wireless power transfer:
[256] M. Chen, D. Gunduz, K. Huang, W. Saad, M. Bennis, A. V. Feljan, A cost-efficient approach,” IEEE Transactions on Communications,
and H. V. Poor, “Distributed learning in wireless networks: Recent vol. 70, no. 4, pp. 2853–2869, April 2022.
progress and future challenges,” IEEE Journal on Selected Areas in [277] N. Führling, K. Ando, H. S. Rou, and G. T. F. D. Abreu, “A rate
Communications, vol. 39, no. 12, pp. 3579–3605, Dec. 2021. splitting multiple access interface for clustered wireless federated
[257] M. S. H. Abad, E. Ozfatura, D. Gündüz, and O. Ercetin, “Hierarchical learning,” IEEE Access, vol. 11, pp. 82 652–82 664, August 2023.
federated learning across heterogeneous cellular networks,” in Proc. [278] S.-H. Park and H. Lee, “Completion time minimization of fog-ran-
IEEE International Conference on Acoustics, Speech and Signal Pro- assisted federated learning with rate-splitting transmission,” IEEE
cessing (ICASSP), Barcelona, Spain, May 2020. Transactions on Vehicular Technology, vol. 71, no. 9, pp. 10 209–
[258] N. H. Tran, W. Bao, A. Zomaya, M. N. H. Nguyen, and C. S. Hong, 10 214, 2022.
“Federated learning over wireless networks: Optimization model design [279] S. Zhang, S. Zhang, W. Yuan, Y. Li, and L. Hanzo, “Efficient rate-
and analysis,” in Proc. IEEE Conference on Computer Communica- splitting multiple access for the internet of vehicles: Federated edge
tions, Paris, France, May 2019. learning and latency minimization,” IEEE Journal on Selected Areas
[259] S. Wang, T. Tuor, T. Salonidis, K. K. Leung, C. Makaya, T. He, and in Communications, vol. 41, no. 5, pp. 1468–1483, May 2023.
K. Chan, “Adaptive federated learning in resource constrained edge [280] G. Zhu, Y. Wang, and K. Huang, “Broadband analog aggregation for
computing systems,” IEEE Journal on Selected Areas in Communica- low-latency federated edge learning,” IEEE Transactions on Wireless
tions, vol. 37, no. 6, pp. 1205–1221, June 2019. Communications, vol. 19, no. 1, pp. 491–506, 2020.
[260] R. Balakrishnan, M. Akdeniz, S. Dhakal, and N. Himayat, “Resource [281] M. Goldenbaum and S. Stanczak, “Robust analog function computation
management and fairness for federated learning over wireless edge via wireless multiple-access channels,” IEEE Transactions on Commu-
networks,” in Proc. IEEE International Workshop on Signal Processing nications, vol. 61, no. 9, pp. 3863–3877, Sep. 2013.
Advances in Wireless Communications (SPAWC), Atlanta, GA, USA, [282] M. Gastpar, “Uncoded transmission is exactly optimal for a simple
May 2020. Gaussian sensor network,” IEEE Transactions on Information Theory,
[261] H. H. Yang, Z. Liu, T. Q. S. Quek, and H. V. Poor, “Scheduling policies vol. 54, pp. 2008–2017, Nov. 2008.
for federated learning in wireless networks,” IEEE Transactions on [283] H. Xing, O. Simeone, and S. Bi, “Federated learning over wire-
Communications, vol. 68, no. 1, pp. 317–333, Jan. 2020. less device-to-device networks: Algorithms and convergence analysis,”
[262] C. T. Dinh, N. H. Tran, M. N. H. Nguyen, C. S. Hong, W. Bao, IEEE Journal on Selected Areas in Communications, vol. 39, no. 12,
A. Y. Zomaya, and V. Gramoli, “Federated learning over wireless pp. 3723–3741, 2021.
networks: Convergence analysis and resource allocation,” IEEE/ACM [284] W. Guo, R. Li, C. Huang, X. Qin, K. Shen, and W. Zhang, “Joint
Transactions on Networking, vol. 29, no. 1, pp. 398–409, Feb. 2021. device selection and power control for wireless federated learning,”
 50

IEEE Journal on Selected Areas in Communications, vol. 40, no. 8, [307] G. Charan, A. Hredzak, C. Stoddard, B. Berrey, M. Seth, H. Nunez, and
pp. 2395–2410, 2022. A. Alkhateeb, “Towards real-world 6g drone communication: Position
[285] N. Zhang and M. Tao, “Gradient statistics aware power control for and aided beam prediction,” in GLOBECOM 2022-2022 IEEE Global
over-the-air federated learning,” IEEE Transactions on Wireless Com- Communications Conference. IEEE, 2022, pp. 2951–2956.
munications, vol. 20, no. 8, pp. 5115–5128, 2021. [308] J. Morais, A. Behboodi, H. Pezeshki, and A. Alkhateeb, “Position aided
[286] Z. Wang, J. Qiu, Y. Zhou, Y. Shi, L. Fu, W. Chen, and K. B. beam prediction in the real world: How useful gps locations actually
Letaief, “Federated learning via intelligent reflecting surface,” IEEE are?” arXiv preprint arXiv:2205.09054, 2022.
Transactions on Wireless Communications, vol. 21, no. 2, pp. 808– [309] M. Zecchin, M. B. Mashhadi, M. Jankowski, D. Gündüz, M. Koun-
822, 2022. touris, and D. Gesbert, “Lidar and position-aided mmwave beam selec-
[287] H. Hellström, V. Fodor, and C. Fischione, “Federated learning over- tion with non-local cnns and curriculum training,” IEEE Transactions
the-air by retransmissions,” IEEE Transactions on Wireless Communi- on Vehicular Technology, vol. 71, no. 3, pp. 2979–2990, 2022.
cations, pp. 1–1, 2023. [310] S. Rezaie, C. N. Manchón, and E. de Carvalho, “Location- and
[288] Z. Lin, X. Li, V. K. N. Lau, Y. Gong, and K. Huang, “Deploying orientation-aided millimeter wave beam selection using deep learning,”
federated learning in large-scale cellular networks: Spatial convergence in ICC 2020 - 2020 IEEE International Conference on Communications
analysis,” IEEE Transactions on Wireless Communications, vol. 21, (ICC), 2020, pp. 1–6.
no. 3, pp. 1542–1556, 2022. [311] V. Va, T. Shimizu, G. Bansal, and R. W. Heath, “Online learning for
[289] S. Wang, M. Chen, C. Shen, C. Yin, and C. G. Brinton, “Cross-layer position-aided millimeter wave beam training,” IEEE Access, vol. 7,
federated learning optimization in mimo networks,” arXiv preprint pp. 30 507–30 526, 2019.
arXiv:2302.14648, 2023. [312] S. Jiang and A. Alkhateeb, “Computer vision aided beam tracking in
[290] L. You, X. Zhao, R. Cao, Y. Shao, and L. Fu, “Broadband digital a real-world millimeter wave deployment,” in 2022 IEEE Globecom
over-the-air computation for wireless federated edge learning,” IEEE Workshops (GC Wkshps). IEEE, 2022, pp. 142–147.
Transactions on Mobile Computing, pp. 1–16, 2023. [313] H. Luo, U. Demirhan, and A. Alkhateeb, “Millimeter wave v2v beam
[291] G. Zhu, Y. Du, D. Gündüz, and K. Huang, “One-bit over-the- tracking using radar: Algorithms and real-world demonstration,” arXiv
air aggregation for communication-efficient federated edge learning: preprint arXiv:2308.01558, 2023.
Design and convergence analysis,” IEEE Transactions on Wireless [314] S. Wu, M. Alrabeiah, C. Chakrabarti, and A. Alkhateeb, “Blockage
Communications, vol. 20, no. 3, pp. 2120–2135, 2021. prediction using wireless signatures: Deep learning enables real-world
[292] A. Hassanien, M. G. Amin, Y. D. Zhang, and F. Ahmad, “Dual-function demonstration,” IEEE Open Journal of the Communications Society,
radar-communications: Information embedding using sidelobe control vol. 3, pp. 776–796, 2022.
and waveform diversity,” IEEE Transactions on Signal Processing, [315] S. Wu, C. Chakrabarti, and A. Alkhateeb, “Proactively predicting
vol. 64, no. 8, pp. 2168–2181, 2016. dynamic 6g link blockages using lidar and in-band signatures,” IEEE
[293] ——, “Phase-modulation based dual-function radar-communications,” Open Journal of the Communications Society, vol. 4, pp. 392–412,
IET Radar, Sonar &Amp; Navigation, vol. 10, pp. 1411–1421, 2016. 2023.
[294] F. Liu, C. Masouros, A. Li, H. Sun, and L. Hanzo, “MU-MIMO [316] G. Charan, M. Alrabeiah, and A. Alkhateeb, “Vision-aided dynamic
communications with MIMO radar: From co-existence to joint trans- blockage prediction for 6g wireless communication networks,” in 2021
mission,” IEEE Transactions on Wireless Communications, vol. 17, IEEE International Conference on Communications Workshops (ICC
no. 4, pp. 2755–2770, 2018. Workshops). IEEE, 2021, pp. 1–6.
[295] X. Liu, T. Huang, N. Shlezinger, Y. Liu, J. Zhou, and Y. C. Eldar,
[317] G. Charan and A. Alkhateeb, “Computer vision aided blockage pre-
“Joint transmit beamforming for multiuser MIMO communications and
diction in real-world millimeter wave deployments,” in 2022 IEEE
MIMO radar,” IEEE Transactions on Signal Processing, vol. 68, pp.
Globecom Workshops (GC Wkshps). IEEE, 2022, pp. 1711–1716.
3929–3944, 2020.
[318] U. Demirhan and A. Alkhateeb, “Radar aided 6g beam prediction:
[296] X. Mu, Z. Wang, and Y. Liu, “NOMA for integrating sensing and
Deep learning algorithms and real-world demonstration,” in 2022
communications towards 6G: A multiple access perspective,” IEEE
IEEE Wireless Communications and Networking Conference (WCNC).
Wireless Communications, pp. 1–8, 2023.
IEEE, 2022, pp. 2655–2660.
[297] Z. Wang, Y. Liu, X. Mu, Z. Ding, and O. A. Dobre, “NOMA empow-
ered integrated sensing and communication,” IEEE Communications [319] ——, “Radar aided proactive blockage prediction in real-world mil-
Letters, vol. 26, no. 3, pp. 677–681, 2022. limeter wave systems,” in ICC 2022-IEEE International Conference
[298] L. Yin and B. Clerckx, “Rate-splitting multiple access for dual- on Communications. IEEE, 2022, pp. 4547–4552.
functional radar-communication satellite systems,” in IEEE Wireless [320] M. Nerini and B. Clerckx, “Overhead-free blockage detection and
Communications and Networking Conference (WCNC), 2022, pp. 1–6. precoding through physics-based graph neural networks: Lidar data
[299] O. Dizdar, A. Kaushik, B. Clerckx, and C. Masouros, “Energy efficient meets ray tracing,” IEEE Wireless Communications Letters, vol. 12,
dual-functional radar-communication: Rate-splitting multiple access, no. 3, pp. 565–569, 2023.
low-resolution DACs, and RF chain selection,” IEEE Open Journal [321] W. Xu, F. Gao, J. Zhang, X. Tao, and A. Alkhateeb, “Deep learning
of the Communications Society, vol. 3, pp. 986–1006, 2022. based channel covariance matrix estimation with user location and
[300] C. Xu, B. Clerckx, S. Chen, Y. Mao, and J. Zhang, “Rate-splitting scene images,” IEEE Transactions on Communications, vol. 69, no. 12,
multiple access for multi-antenna joint communication and radar trans- pp. 8145–8158, 2021.
missions,” in 2020 IEEE International Conference on Communications [322] J. Chen, F. Gao, X. Tao, G. Liu, C. Pan, and A. Alkhateeb, “Computer
Workshops (ICC Workshops), 2020, pp. 1–6. vision aided codebook design for mimo communications systems,”
[301] U. Demirhan and A. Alkhateeb, “Cell-free isac mimo systems: Joint IEEE Transactions on Wireless Communications, 2022.
sensing and communication beamforming,” 2023. [323] G. Charan and A. Alkhateeb, “User identification: The key enabler
[302] M. Ahmadipour, M. Kobayashi, M. Wigger, and G. Caire, “An for multi-user vision-aided wireless communications,” arXiv preprint
information-theoretic approach to joint sensing and communication,” arXiv:2210.15652, 2022.
IEEE Transactions on Information Theory, pp. 1–1, 2022. [324] G. Charan, M. Alrabeiah, T. Osman, and A. Alkhateeb, “Camera
[303] A. Alkhateeb, G. Charan, T. Osman, A. Hredzak, J. Morais, based mmwave beam prediction: Towards multi-candidate real-world
U. Demirhan, and N. Srinivas, “Deepsense 6g: A large-scale real-world scenarios,” arXiv preprint arXiv:2308.06868, 2023.
multi-modal sensing and communication dataset,” IEEE Communica- [325] A. Alkhateeb, S. Jiang, and G. Charan, “Real-time digital twins: Vision
tions Magazine, vol. 61, no. 9, pp. 122–128, 2023. and research directions for 6g and beyond,” IEEE Communications
[304] U. Demirhan and A. Alkhateeb, “Radar aided 6g beam prediction: Magazine, pp. 1–7, 2023.
Deep learning algorithms and real-world demonstration,” in 2022 [326] C. Ruah, O. Simeone, and B. Al-Hashimi, “Digital twin-based mul-
IEEE Wireless Communications and Networking Conference (WCNC). tiple access optimization and monitoring via model-driven bayesian
IEEE, 2022, pp. 2655–2660. learning,” 2023.
[305] G. Charan, T. Osman, A. Hredzak, N. Thawdar, and A. Alkhateeb, [327] S. Dwivedi, R. Shreevastav, F. Munier, J. Nygren, I. Siomina,
“Vision-position multi-modal beam prediction using real millimeter Y. Lyazidi, D. Shrestha, G. Lindmark, P. Ernstrom, E. Stare, S. M.
wave datasets,” in 2022 IEEE Wireless Communications and Network- Razavi, S. Muruganathan, G. Masini, A. Busin, and F. Gunnars-
ing Conference (WCNC). IEEE, 2022, pp. 2727–2731. son, “Positioning in 5G networks,” IEEE Communications Magazine,
[306] S. Jiang, G. Charan, and A. Alkhateeb, “Lidar aided future beam vol. 59, no. 11, pp. 38–44, 2021.
prediction in real-world millimeter wave v2i communications,” IEEE [328] J. A. del Peral-Rosado, R. Raulefs, J. A. Lopez-Salcedo, and G. Seco-
Wireless Communications Letters, vol. 12, no. 2, pp. 212–216, 2022. Granados, “Survey of cellular mobile radio localization methods: From
 51

1G to 5G,” IEEE Communications Surveys & Tutorials, vol. 20, no. 2, [350] Z. Zhao, Z. Yang, Q.-V. Pham, Q. Yang, and Z. Zhang, “Semantic
pp. 1124–1148, 2018. communication with probability graph: A joint communication and
[329] N. Garcia, H. Wymeersch, E. G. Larsson, A. M. Haimovich, and computation design,” arXiv preprint arXiv:2310.00015, 2023.
M. Coulon, “Direct localization for massive MIMO,” IEEE Transac- [351] Z. Yang, M. Chen, G. Li, Y. Yang, and Z. Zhang, “Secure seman-
tions on Signal Processing, vol. 65, no. 10, pp. 2475–2487, 2017. tic communications: Fundamentals and challenges,” arXiv preprint
[330] S. Jeong, O. Simeone, and J. Kang, “Optimization of massive full- arXiv:2301.01421, 2023.
dimensional MIMO for positioning and communication,” IEEE Trans- [352] L. Yan, Z. Qin, R. Zhang, Y. Li, and G. Ye Li, “Qoe-aware resource
actions on Wireless Communications, vol. 17, no. 9, pp. 6205–6217, allocation for semantic communication networks,” in GLOBECOM
2018. 2022 - 2022 IEEE Global Communications Conference, 2022, pp.
[331] G. Kwon, A. Conti, H. Park, and M. Z. Win, “Joint communication and 3272–3277.
localization in millimeter wave networks,” IEEE Journal of Selected [353] M.-D. Nguyen, Q.-V. Do, Z. Yang, Q.-V. Pham, and W.-J. Hwang,
Topics in Signal Processing, vol. 15, no. 6, pp. 1439–1454, 2021. “Joint communication and computation framework for goal-oriented
[332] L. Yin, J. Cao, Q. Ni, Y. Ma, and S. Li, “Design and performance semantic communication with distortion rate resilience,” arXiv preprint
analysis of multi-scale NOMA for future communication-positioning arXiv:2309.14587, 2023.
integration system,” IEEE Journal on Selected Areas in Communica- [354] Z. Yang, M. Chen, Z. Zhang, C. Huang, and Q. Yang, “Performance
tions, vol. 40, no. 4, pp. 1333–1345, 2022. optimization of energy efficient semantic communications over wire-
[333] L. Han, R. Liu, Z. Wang, X. Yue, and J. S. Thompson, “Millimeter- less networks,” in 2022 IEEE 96th Vehicular Technology Conference
wave MIMO-NOMA-based positioning system for Internet-of-things (VTC2022-Fall). IEEE, 2022, pp. 1–5.
applications,” IEEE Internet of Things Journal, vol. 7, no. 11, pp. [355] Y. Cang, M. Chen, Z. Yang, Y. Hu, Y. Wang, C. Huang, and Z. Zhang,
11 068–11 077, 2020. “Online resource allocation for semantic-aware edge computing sys-
[334] P. Gao, L. Lian, and J. Yu, “Cooperative ISAC with direct localization tems,” IEEE Internet of Things Journal, 2023.
and rate-splitting multiple access communication: A pareto optimiza- [356] Y. Cang, M. Chen, Z. Yang, Y. Hu, Y. Wang, Z. Zhang, and K.-K.
tion framework,” IEEE Journal on Selected Areas in Communications, Wong, “Resource allocation for semantic-aware mobile edge computing
vol. 41, no. 5, pp. 1496–1515, 2023. systems,” arXiv preprint arXiv:2309.11736, 2023.
[335] G. Shi, Y. Xiao, Y. Li, and X. Xie, “From semantic communication to [357] Z. Zhao, Z. Yang, Y. Hu, L. Lin, and Z. Zhang, “Semantic information
semantic-aware networking: Model, architecture, and open problems,” extraction for text data with probability graph,” in 2023 IEEE/CIC
IEEE Communications Magazine, vol. 59, no. 8, pp. 44–50, 2021. International Conference on Communications in China (ICCC Work-
[336] J. Bao, P. Basu, M. Dean, C. Partridge, A. Swami, W. Leland, and J. A. shops). IEEE, 2023, pp. 1–6.
Hendler, “Towards a theory of semantic communication,” in 2011 IEEE [358] X. Mu and Y. Liu, “Exploiting semantic communication for non-
Network Science Workshop. IEEE, 2011, pp. 110–117. orthogonal multiple access,” IEEE Journal on Selected Areas in Com-
[337] Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y. Li, “Semantic communi- munications, 2023.
cations: Principles and challenges,” arXiv preprint arXiv:2201.01389, [359] W. Li, H. Liang, C. Dong, X. Xu, P. Zhang, and K. Liu, “Non-
2021. orthogonal multiple access enhanced multi-user semantic communica-
[338] D. Gündüz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, tion,” IEEE Transactions on Cognitive Communications and Network-
K. K. Wong, and C.-B. Chae, “Beyond transmitting bits: Context, se- ing, pp. 1–1, 2023.
mantics, and task-oriented communications,” IEEE Journal on Selected [360] M. Kim, S. Kim, J. Kwak, and J. Lee, “Semantic-based precoding
Areas in Communications, vol. 41, no. 1, pp. 5–41, 2022. design for multi-user MISO networks,” IEEE Wireless Communications
[339] W. Yang, H. Du, Z. Q. Liew, W. Y. B. Lim, Z. Xiong, D. Niyato, Letters, 2023.
X. Chi, X. S. Shen, and C. Miao, “Semantic communications for [361] X. Luo, R. Gao, H.-H. Chen, S. Chen, Q. Guo, and P. N. Suganthan,
future internet: Fundamentals, applications, and challenges,” IEEE “Multi-modal and multi-user semantic communications for channel-
Communications Surveys & Tutorials, 2022. level information fusion,” IEEE Wireless Communications, 2022.
[340] H. Tong, Z. Yang, S. Wang, Y. Hu, W. Saad, and C. Yin, “Federated [362] Z. Yang, M. Chen, Z. Zhang, and C. Huang, “Energy efficient semantic
learning based audio semantic communication over wireless networks,” communication over wireless networks with rate splitting,” IEEE
in 2021 IEEE Global Communications Conference (GLOBECOM). Journal on Selected Areas in Communications, vol. 41, no. 5, pp. 1484–
IEEE, 2021, pp. 1–6. 1495, 2023.
[341] K. Niu, J. Dai, S. Yao, S. Wang, Z. Si, X. Qin, and P. Zhang, [363] C. Zeng, J.-B. Wang, M. Xiao, C. Ding, Y. Chen, H. Yu, and J. Wang,
“Towards semantic communications: A paradigm shift,” arXiv preprint “Task-oriented semantic communication over rate splitting enabled
arXiv:2203.06692, 2022. wireless control systems for urllc services,” IEEE Transactions on
[342] Y. Lin, H. Du, D. Niyato, J. Nie, J. Zhang, Y. Cheng, and Z. Yang, Communications, 2023.
“Blockchain-aided secure semantic communication for ai-generated [364] Y. Cheng, D. Niyato, H. Du, J. Kang, C. Miao, and D. I. Kim, “Interest-
content in metaverse,” IEEE Open Journal of the Computer Society, based semantic information transmission with rsma in smart cities,” in
vol. 4, pp. 72–83, 2023. ICC 2023-IEEE International Conference on Communications. IEEE,
[343] X. Luo, H.-H. Chen, and Q. Guo, “Semantic communications: 2023, pp. 82–87.
Overview, open issues, and future research directions,” IEEE Wireless [365] Y. Cheng, D. Niyato, H. Du, J. Kang, Z. Xiong, C. Miao, and D. I. Kim,
Communications, vol. 29, no. 1, pp. 210–219, 2022. “Resource allocation and common message selection for task-oriented
[344] K. Lu, Q. Zhou, R. Li, Z. Zhao, X. Chen, J. Wu, and H. Zhang, semantic information transmission with rsma,” IEEE Transactions on
“Rethinking modern communication from semantic coding to semantic Wireless Communications, 2023.
communication,” IEEE Wireless Communications, vol. 30, no. 1, pp. [366] P. Zhang, X. Xu, C. Dong, K. Niu, H. Liang, Z. Liang, X. Qin, M. Sun,
158–164, 2022. H. Chen, N. Ma et al., “Model division multiple access for semantic
[345] Y. Ding, Z. Yang, Q.-V. Pham, Z. Zhang, and M. Shikh-Bahaei, communications,” Frontiers of Information Technology & Electronic
“Distributed machine learning for uav swarms: Computing, sensing, Engineering, pp. 1–12, 2023.
and semantics,” arXiv preprint arXiv:2301.00912, 2023. [367] X. Liu, H. Liang, Z. Bao, C. Dong, and X. Xu, “Semantic commu-
[346] W. Xu, Z. Yang, D. W. K. Ng, M. Levorato, Y. C. Eldar, and nications system with model division multiple access and controllable
M. Debbah, “Edge learning for b5g networks with distributed signal coding rate for point cloud,” arXiv preprint arXiv:2307.06027, 2023.
processing: Semantic communication, edge computing, and wireless [368] W. Zhang, K. Bai, S. Zeadally, H. Zhang, H. Shao, H. Ma, and V. C. M.
sensing,” IEEE journal of selected topics in signal processing, vol. 17, Leung, “Deepma: End-to-end deep multiple access for wireless image
no. 1, pp. 9–39, 2023. transmission in semantic communication,” IEEE Transactions on Cog-
[347] L. Yan, Z. Qin, R. Zhang, Y. Li, and G. Y. Li, “Resource allocation nitive Communications and Networking, pp. 1–1, 2023.
for text semantic communications,” IEEE Wireless Communications [369] Y. Wang, Z. Su, N. Zhang, R. Xing, D. Liu, T. H. Luan, and X. Shen,
Letters, vol. 11, no. 7, pp. 1394–1398, 2022. “A survey on metaverse: Fundamentals, security, and privacy,” IEEE
[348] X. Mu, Y. Liu, L. Guo, and N. Al-Dhahir, “Heterogeneous semantic and Communications Surveys & Tutorials, 2022.
bit communications: A semi-noma scheme,” IEEE Journal on Selected [370] J. Kang, J. He, H. Du, Z. Xiong, Z. Yang, X. Huang, and S. Xie,
Areas in Communications, vol. 41, no. 1, pp. 155–169, 2022. “Adversarial attacks and defenses for semantic communication in
[349] Z. Hu, T. Liu, C. You, Z. Yang, and M. Chen, “Multiuser resource vehicular metaverses,” arXiv preprint arXiv:2306.03528, 2023.
allocation for semantic-relay-aided text transmissions,” arXiv preprint [371] M. Xu, W. C. Ng, W. Y. B. Lim, J. Kang, Z. Xiong, D. Niyato,
arXiv:2311.06854, 2023. Q. Yang, X. S. Shen, and C. Miao, “A full dive into realizing the edge-
 52

enabled metaverse: Visions, enabling technologies, and challenges,” and delivery framework for edge-enhanced next-generation wireless
IEEE Communications Surveys & Tutorials, 2022. networks,” IEEE J. Sel. Areas Commun., vol. 40, no. 5, pp. 1615–1631,
[372] H. Du, B. Ma, D. Niyato, J. Kang, Z. Xiong, and Z. Yang, “Rethink- May 2022.
ing quality of experience for metaverse services: A consumer-based [394] Y. Li, C. Dou, Y. Wu, W. Jia, and R. Lu, “NOMA assisted two-tier
economics perspective,” IEEE Network, 2023. VR content transmission: A tile-based approach for QoE optimization,”
[373] X. Luo, J. Wen, J. Kang, J. Nie, Z. Xiong, Y. Zhang, Z. Yang, and IEEE Trans. Mobile Comput., Early Access, May 2023.
S. Xie, “Privacy attacks and defenses for digital twin migrations in [395] T. Zhang, Z. Wang, Y. Liu, W. Xu, and A. Nallanathan, “Joint resource,
vehicular metaverses,” IEEE Network, 2023. deployment, and caching optimization for AR applications in dynamic
[374] S. K. Jagatheesaperumal, Z. Yang, Q. Yang, C. Huang, W. Xu, UAV NOMA networks,” IEEE Trans. Wireless Commun., vol. 21, no. 5,
M. Shikh-Bahaei, and Z. Zhang, “Semantic-aware digital twin for pp. 3409–3422, May 2022.
metaverse: A comprehensive review,” arXiv preprint arXiv:2305.18304, [396] L. Zhao, Y. Cui, S. Yang, and S. S. Shitz, “An optimization framework
2023. for general rate splitting for general multicast,” IEEE Trans. Wireless
[375] B. Brik, H. Moustafa, Y. Zhang, A. Lakas, and S. Subramanian, “Guest Commun., vol. 22, no. 3, pp. 1573–1587, Mar. 2022.
editorial: Multi-access networking for extended reality and metaverse,” [397] R. Huang, V. W.S. Wong, and R. Schober, “Rate-splitting for intelligent
IEEE Internet of Things Magazine, vol. 6, no. 1, pp. 12–13, 2023. reflecting surface-aided multiuser VR streaming,” IEEE J. Sel. Areas
[376] K. Gai, S. Wang, H. Zhao, Y. She, Z. Zhang, and L. Zhu, “Blockchain- Commun., vol. 41, no. 5, pp. 1516–1535, May 2023.
based multisignature lock for uac in metaverse,” IEEE Transactions on [398] L. Zhao, Y. Cui, Z. Liu, Y. Zhang, and S. Yang, “Adaptive streaming of
Computational Social Systems, 2022. 360 videos with perfect, imperfect, and unknown FoV viewing prob-
[377] M. Xu, D. Niyato, J. Kang, Z. Xiong, C. Miao, and D. I. Kim, “Wireless abilities in wireless networks,” IEEE Trans. Image Process., vol. 30,
edge-empowered metaverse: A learning-based incentive mechanism for pp. 7744–7759, Sept. 2021.
virtual reality,” in IEEE International Conference on Communications, [399] N. Q. Hieu, D. N. Nguyen, D. T. Hoang, and E. Dutkiewicz, “When
2022, pp. 5220–5225. virtual reality meets rate splitting multiple access: A joint communi-
[378] H. Du, D. Niyato, C. Miao, J. Kang, and D. I. Kim, “Optimal cation and computation approach,” IEEE J. Sel. Areas in Commun.,
targeted advertising strategy for secure wireless edge metaverse,” in vol. 41, no. 5, pp. 1536–1548, May 2023.
GLOBECOM 2022 - 2022 IEEE Global Communications Conference, [400] M. Setayesh and V. W.S. Wong, “A content-based viewport prediction
2022, pp. 4346–4351. framework for 360◦ video using personalized federated learning and
[379] Y. Jiang, J. Kang, X. Ge, D. Niyato, and Z. Xiong, “QoE analysis and fusion techniques,” in Proc. of Int’l Conf. on Multimedia and Expo
resource allocation for wireless metaverse services,” IEEE Transactions (ICME), Brisbane, Australia, Jul. 2023.
on Communications, 2023. [401] H. Li, S. Shen, and B. Clerckx, “Beyond diagonal reconfigurable
[380] M. H. Rahman, M. A. S. Sejan, M. A. Aziz, D.-S. Kim, Y.-H. You, and intelligent surfaces: From transmitting and reflecting modes to single-
H.-K. Song, “Spectral efficiency analysis for irs-assisted miso wireless , group-, and fully-connected architectures,” IEEE Transactions on
communication: A metaverse scenario proposal,” Mathematics, vol. 11, Wireless Communications, vol. 22, no. 4, pp. 2311–2324, 2023.
no. 14, p. 3181, 2023.
[402] ——, “Beyond diagonal reconfigurable intelligent surfaces: A multi-
[381] J. Nie, Q. Zhou, J. Mu, and X. Jing, “Vision and radar multimodal aided
sector mode enabling highly directional full-space wireless coverage,”
beam prediction: Facilitating metaverse development,” in Proceedings
IEEE Journal on Selected Areas in Communications, vol. 41, no. 8,
of the 2nd Workshop on Integrated Sensing and Communications for
pp. 2446–2460, 2023.
Metaverse, 2023, pp. 13–18.
[403] M. Nerini, S. Shen, H. Li, and B. Clerckx, “Beyond diagonal reconfig-
[382] M. S. Elbamby, C. Perfecto, M. Bennis, and K. Doppler, “Toward low-
urable intelligent surfaces utilizing graph theory: Modeling, architec-
latency and ultra-reliable virtual reality,” IEEE Netw., vol. 32, no. 2,
ture design, and optimization,” arXiv preprint arXiv:2305.05013, 2023.
pp. 78–84, Apr. 2018.
[404] H. Zhang and B. Di, “Intelligent omni-surfaces: Simultaneous re-
[383] K. Long, Y. Cui, C. Ye, and Z. Liu, “Optimal wireless streaming of
fraction and reflection for full-dimensional wireless communications,”
multi-quality 360 VR video by exploiting natural, relative smoothness-
IEEE Communications Surveys & Tutorials, vol. 24, no. 4, pp. 1997–
enabled, and transcoding-enabled multicast opportunities,” IEEE Trans.
2028, 2022.
Multimedia, vol. 23, pp. 3670–3683, Oct. 2021.
[384] C. Guo, Y. Cui, and Z. Liu, “Optimal multicast of tiled 360 VR video,” [405] E. Björnson, O. Özdogan, and E. G. Larsson, “Intelligent reflecting
IEEE Commun. Lett., vol. 8, no. 1, pp. 145–148, Feb. 2019. surface versus decode-and-forward: How large surfaces are needed to
[385] C. Guo, L. Zhao, Y. Cui, Z. Liu, and D. W. K. Ng, “Power-efficient beat relaying?” IEEE Wireless Communications Letters, vol. 9, no. 2,
wireless streaming of multi-quality tiled 360 VR video in MIMO- pp. 244–248, 2020.
OFDMA systems,” IEEE Trans. Wireless Commun., vol. 20, no. 8, [406] Y. Yang, B. Zheng, S. Zhang, and R. Zhang, “Intelligent reflecting
pp. 5408–5422, Aug. 2021. surface meets OFDM: Protocol design and rate maximization,” IEEE
[386] R. Zhang, J. Liu, F. Liu, T. Huang, Q. Tang, S. Wang, and F. R. Yu, Transactions on Communications, vol. 68, no. 7, pp. 4522–4535, 2020.
“Buffer-aware virtual reality video streaming with personalized and [407] B. Zheng and R. Zhang, “Intelligent reflecting surface-enhanced
private viewport prediction,” IEEE J. Sel. Areas in Commun., vol. 40, OFDM: Channel estimation and reflection optimization,” IEEE Wireless
no. 2, pp. 694–709, Feb. 2022. Communications Letters, vol. 9, no. 4, pp. 518–522, 2020.
[387] J. Feng, L. Liu, X. Hou, Q. Pei, and C. Wu, “QoE fairness resource [408] P. Wang, J. Fang, X. Yuan, Z. Chen, and H. Li, “Intelligent reflecting
allocation in digital twin-enabled wireless virtual reality systems,” surface-assisted millimeter wave communications: Joint active and
IEEE J. Sel. Areas in Commun., Early Access, Sept. 2023. passive precoding design,” IEEE Transactions on Vehicular Technology,
[388] J. Yang, J. Luo, J. Wang, and S. Guo, “CMU-VP: Cooperative multicast vol. 69, no. 12, pp. 14 960–14 973, 2020.
and unicast with viewport prediction for VR video streaming in 5G H- [409] M. A. ElMossallamy, H. Zhang, L. Song, K. G. Seddik, Z. Han, and
CRAN,” IEEE Access, vol. 7, pp. 134 187–134 197, Sept. 2019. G. Y. Li, “Reconfigurable intelligent surfaces for wireless communi-
[389] I.-S. Comşa, G.-M. Muntean, and R. Trestian, “An innovative machine- cations: Principles, challenges, and opportunities,” IEEE Transactions
learning-based scheduling solution for improving live UHD video on Cognitive Communications and Networking, vol. 6, no. 3, pp. 990–
streaming quality in highly dynamic network environments,” IEEE 1002, 2020.
Trans. Broadcast., vol. 67, no. 1, pp. 212–224, Mar. 2021. [410] C. You, B. Zheng, and R. Zhang, “Channel estimation and passive
[390] L. Zhang and J. Chakareski, “UAV-assisted edge computing and beamforming for intelligent reflecting surface: Discrete phase shift and
streaming for wireless virtual reality: Analysis, algorithm design, and progressive refinement,” IEEE Journal on Selected Areas in Commu-
performance guarantees,” IEEE Trans. Veh. Teclmol., vol. 71, no. 3, nications, vol. 38, no. 11, pp. 2604–2620, 2020.
pp. 3267–3275, Mar. 2022. [411] B. Zheng, C. You, and R. Zhang, “Fast channel estimation for IRS-
[391] O. Eltobgy, O. Arafa, and M. Hefeeda, “Mobile streaming of live 360- assisted OFDM,” IEEE Wireless Communications Letters, vol. 10,
degree videos,” EEE Trans. Multimedia, vol. 22, no. 12, pp. 3139–3152, no. 3, pp. 580–584, 2021.
Dec. 2020. [412] Q. Tao, J. Wang, and C. Zhong, “Performance analysis of intelligent re-
[392] W. Yu, T. J. Chua, and J. Zhao, “User-centric heterogeneous-action flecting surface aided communication systems,” IEEE Communications
deep reinforcement learning for virtual reality in the Metaverse over Letters, vol. 24, no. 11, pp. 2464–2468, 2020.
wireless networks,” IEEE Trans. Wireless Commun., Early Access, May [413] A. Zappone, M. Di Renzo, X. Xi, and M. Debbah, “On the optimal
2023. number of reflecting elements for reconfigurable intelligent surfaces,”
[393] H. Xiao, C. Xu, Z. Feng, R. Ding, S. Yang, L. Zhong, J. Liang, IEEE Wireless Communications Letters, vol. 10, no. 3, pp. 464–468,
and G.-M. Muntean, “A transcoding-enabled 360 VR video caching 2021.
 53

[414] X. Tan, Z. Sun, J. M. Jornet, and D. Pados, “Increasing indoor spectrum [434] Z. Yang, J. Shi, Z. Li, M. Chen, W. Xu, and M. Shikh-Bahaei, “Energy
sharing capacity using smart reflect-array,” in IEEE International efficient rate splitting multiple access (RSMA) with reconfigurable
Conference on Communications (ICC), 2016, pp. 1–6. intelligent surface,” in IEEE International Conference on Communi-
[415] X. Tan, Z. Sun, D. Koutsonikolas, and J. M. Jornet, “Enabling indoor cations Workshops (ICC Workshops), 2020, pp. 1–6.
mobile millimeter-wave networks based on smart reflect-arrays,” in [435] R. Zhang, K. Xiong, Y. Lu, P. Fan, D. W. K. Ng, and K. B. Letaief,
IEEE Conference on Computer Communications (INFOCOM), 2018, “Energy efficiency maximization in RIS-assisted swipt networks with
pp. 270–278. RSMA: A PPO-based approach,” IEEE Journal on Selected Areas in
[416] W. Tang, J. Y. Dai, M. Z. Chen, K.-K. Wong, X. Li, X. Zhao, S. Jin, Communications, vol. 41, no. 5, pp. 1413–1430, 2023.
Q. Cheng, and T. J. Cui, “MIMO transmission through reconfigurable [436] K. Zhao, Y. Mao, Z. Yang, L. Lian, and B. Clerckx, “Reconfigurable
intelligent surface: System design, analysis, and implementation,” IEEE intelligent surfaces empowered cooperative rate splitting with user
Journal on Selected Areas in Communications, vol. 38, no. 11, pp. relaying,” in International Symposium on Wireless Communication
2683–2699, 2020. Systems (ISWCS), 2022, pp. 1–6.
[417] A. Araghi, M. Khalily, M. Safaei, A. Bagheri, V. Singh, F. Wang, [437] T. Zhang and S. Mao, “Joint beamforming design in reconfigurable
and R. Tafazolli, “Reconfigurable intelligent surface (RIS) in the sub- intelligent surface-assisted rate splitting networks,” IEEE Transactions
6 GHz band: Design, implementation, and real-world demonstration,” on Wireless Communications, pp. 1–1, 2023.
IEEE Access, vol. 10, pp. 2646–2655, 2022. [438] A. Bansal, K. Singh, B. Clerckx, C.-P. Li, and M.-S. Alouini, “Rate-
[418] X. Pei, H. Yin, L. Tan, L. Cao, Z. Li, K. Wang, K. Zhang, and splitting multiple access for intelligent reflecting surface aided multi-
E. Björnson, “RIS-aided wireless communications: Prototyping, adap- user communications,” IEEE Transactions on Vehicular Technology,
tive beamforming, and indoor/outdoor field trials,” IEEE Transactions vol. 70, no. 9, pp. 9217–9229, 2021.
on Communications, vol. 69, no. 12, pp. 8627–8640, 2021. [439] A. Bansal, K. Singh, and C.-P. Li, “Analysis of hierarchical rate
[419] J. Tang, M. Cui, S. Xu, L. Dai, F. Yang, and M. Li, “Transmissive splitting for intelligent reflecting surfaces-aided downlink multiuser
RIS for B5G communications: Design, prototyping, and experimental MISO communications,” IEEE Open Journal of the Communications
demonstrations,” IEEE Transactions on Communications, pp. 1–1, Society, vol. 2, pp. 785–798, 2021.
2023. [440] S. Dhok and P. K. Sharma, “Rate-splitting multiple access with
[420] C. Pan, G. Zhou, K. Zhi, S. Hong, T. Wu, Y. Pan, H. Ren, M. D. STAR RIS over spatially-correlated channels,” IEEE Transactions on
Renzo, A. Lee Swindlehurst, R. Zhang, and A. Y. Zhang, “An overview Communications, vol. 70, no. 10, pp. 6410–6424, 2022.
of signal processing techniques for RIS/IRS-aided wireless systems,” [441] C. Meng, K. Xiong, W. Chen, B. Gao, P. Fan, and K. B. Letaief,
IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 5, “Sum-rate maximization in STAR-RIS assisted RSMA networks: A
pp. 883–917, 2022. PPO-based algorithm,” IEEE Internet of Things Journal, pp. 1–1, 2023.
[421] T. Fang and Y. Mao, “A low-complexity beamforming design for [442] T. Fang, Y. Mao, S. Shen, Z. Zhu, and B. Clerckx, “Fully connected
beyond-diagonal RIS aided multi-user networks, (early access),” IEEE reconfigurable intelligent surface aided rate-splitting multiple access for
Communications Letters, 2023. multi-user multi-antenna transmission,” in IEEE International Confer-
[422] B. Zheng, C. You, W. Mei, and R. Zhang, “A survey on channel ence on Communications Workshops (ICC Workshops), 2022, pp. 675–
estimation and practical passive beamforming design for intelligent 680.
reflecting surface aided wireless communications,” IEEE Communica- [443] M. Soleymani, I. Santamaria, E. Jorswieck, and B. Clerckx, “Optimiza-
tions Surveys & Tutorials, vol. 24, no. 2, pp. 1035–1071, 2022. tion of rate-splitting multiple access in beyond diagonal ris-assisted
[423] S. Ren, K. Shen, Y. Zhang, X. Li, X. Chen, and Z.-Q. Luo, “Con- urllc systems,” IEEE Transactions on Wireless Communications, pp.
figuring intelligent reflecting surface with performance guarantees: 1–1, 2023.
Blind beamforming,” IEEE Transactions on Wireless Communications, [444] H. Li, S. Shen, and B. Clerckx, “Synergizing beyond diagonal recon-
vol. 22, no. 5, pp. 3355–3370, 2023. figurable intelligent surface and rate-splitting multiple access,” arXiv
[424] Z. Zhou, H. Yin, L. Tan, R. Zhang, K. Wang, and Y. Liu, “Multi- preprint arXiv:2303.06912, 2023.
user beamforming in RIS-aided communications and experimental [445] H. Li, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple ac-
validations,” arXiv preprint arXiv:2309.09460, 2023. cess for 6G–Part III: Interplay with reconfigurable intelligent surfaces,”
[425] L. Lv, H. Jiang, Z. Ding, Q. Ye, N. Al-Dhahir, and J. Chen, “Secure IEEE Communications Letters, vol. 26, no. 10, pp. 2242–2246, 2022.
non-orthogonal multiple access: An interference engineering perspec- [446] L. Liu, E. G. Larsson, W. Yu, P. Popovsk, C. Stefanovic, and E. D.
tive,” IEEE Network, vol. 35, no. 4, pp. 278–285, 2021. Carvalho, “Sparse signal processing for grant-free massive connectiv-
[426] B. Zheng, Q. Wu, and R. Zhang, “Intelligent reflecting surface- ity: A future paradigm for random access protocols in the Internet of
assisted multiple access with user pairing: Noma or oma?” IEEE Things,” IEEE Signal Processing Magazine, vol. 35, no. 5, pp. 88–99,
Communications Letters, vol. 24, no. 4, pp. 753–757, 2020. Sep. 2018.
[427] A. S. de Sena, P. H. J. Nardelli, D. B. da Costa, P. Popovski, [447] Y. Wu, X. Gao, S. Zhou, W. Yang, Y. Polyanskiy, and G. Caire,
and C. B. Papadias, “Rate-splitting multiple access and its interplay “Massive access for future wireless communication systems,” IEEE
with intelligent reflecting surfaces,” IEEE Communications Magazine, Wireless Communications, vol. 27, no. 4, pp. 148–156, 2020.
vol. 60, no. 7, pp. 52–57, 2022. [448] X. Chen, D. W. K. Ng, W. Yu, E. G. Larsson, N. Al-Dhahir, and
[428] H. Li, Y. Mao, O. Dizdar, and B. Clerckx, “Rate-splitting multiple R. Schober, “Massive access for 5G and beyond,” IEEE Journal on
access for 6G—part III: Interplay with reconfigurable intelligent sur- Selected Areas in Communications, vol. 39, no. 3, pp. 615–637, 2021.
faces,” IEEE Communications Letters, vol. 26, no. 10, pp. 2242–2246, [449] L. Liu and W. Yu, “Massive connectivity with massive MIMO-Part I:
2022. Device activity detection and channel estimation,” IEEE Transactions
[429] M. Soleymani, I. Santamaria, and E. A. Jorswieck, “Rate splitting in on Signal Processing, vol. 66, no. 11, pp. 2933–2946, Jun. 2018.
MIMO RIS-assisted systems with hardware impairments and improper [450] Z. Chen, F. Sohrabi, and W. Yu, “Sparse activity detection for massive
signaling,” IEEE Transactions on Vehicular Technology, vol. 72, no. 4, connectivity,” IEEE Transactions on Signal Processing, vol. 66, no. 7,
pp. 4580–4597, 2023. pp. 1890–1904, 2018.
[430] H. Niu, Z. Lin, K. An, J. Wang, G. Zheng, N. Al-Dhahir, and K.- [451] L. Liu and W. Yu, “Massive connectivity with massive MIMO—part I:
K. Wong, “Active RIS assisted rate-splitting multiple access network: Device activity detection and channel estimation,” IEEE Transactions
Spectral and energy efficiency tradeoff,” IEEE Journal on Selected on Signal Processing, vol. 66, no. 11, pp. 2933–2946, 2018.
Areas in Communications, vol. 41, no. 5, pp. 1452–1467, 2023. [452] M. Ke, Z. Gao, Y. Wu, X. Gao, and R. Schober, “Compressive
[431] R. Huang, V. W. Wong, and R. Schober, “Rate-splitting for intelligent sensing-based adaptive active user detection and channel estimation:
reflecting surface-aided multiuser VR streaming,” IEEE Journal on Massive access meets massive MIMO,” IEEE Transactions on Signal
Selected Areas in Communications, vol. 41, no. 5, pp. 1516–1535, Processing, vol. 68, pp. 764–779, 2020.
2023. [453] X. Shao, X. Chen, D. W. K. Ng, C. Zhong, and Z. Zhang, “Cooperative
[432] M. Katwe, K. Singh, B. Clerckx, and C.-P. Li, “Rate-splitting multiple activity detection: Sourced and unsourced massive random access
access and dynamic user clustering for sum-rate maximization in paradigms,” IEEE Transactions on Signal Processing, vol. 68, pp.
multiple riss-aided uplink mmwave system,” IEEE Transactions on 6578–6593, 2020.
Communications, vol. 70, no. 11, pp. 7365–7383, 2022. [454] E. Björnson, E. de Carvalho, J. H. Sørensen, E. G. Larsson, and
[433] ——, “Rate splitting multiple access for sum-rate maximization in P. Popovski, “A random access protocol for pilot allocation in crowded
irs aided uplink communications,” IEEE Transactions on Wireless massive MIMO systems,” IEEE Transactions on Wireless Communica-
Communications, vol. 22, no. 4, pp. 2246–2261, 2023. tions, vol. 16, no. 4, pp. 2220–2234, Apr. 2017.
 54

[455] A. Fengler, S. Haghighatshoar, P. Jung, and G. Caire, “Non-Bayesian [476] ZTE, “Revised wid on 2-step rach for NR,” 3rd Generation Partnership
activity detection, large-scale fading coefficient estimation, and un- Project (3GPP), document RP-200085, 3GPP TSG RAN Meeting #87e,
sourced random access with a massive MIMO receiver,” IEEE Trans- March 2020.
actions on Information Theory, vol. 67, no. 5, pp. 2925–2951, May [477] Viavi, “High-level release 19 proposal,” 3GPP Rel. 19 Workshop,
2021. Taiwan, June 2023.
[456] Z. Chen and W. Yu, “Sparse activity detection for massive connec- [478] BBC, “Rsma for mu-mimo – enhancement of unicast and joint uni-
tivity,” IEEE Transactions on Signal Processing, vol. 66, no. 7, pp. cast/multicast delivery,” 3GPP Rel. 19 Workshop, Taiwan, June 2023.
1890–1904, Apr. 2018. [479] Qualcomm, “Study on artificial intelligence (AI)/machine learning
[457] J. Wang, Z. Zhang, and L. Hanzo, “Joint active user detection and (ML) for NR air interface,” 3rd Generation Partnership Project (3GPP),
channel estimation in massive access systems exploiting Reed-Muller document RP-221348, 3GPP TSG RAN Meeting #96, June 2022.
sequences,” IEEE Journal on Selected Topics in Signal Processing, [480] 3GPP, “Study on artificial intelligence (AI)/machine learning (ML) for
vol. 13, no. 3, pp. 739–752, Jun. 2019. NR air interface,” 3rd Generation Partnership Project (3GPP), Technical
[458] Y. Cui, S. Li, and W. Zhang, “Jointly sparse signal recovery and support Report TR 38.843 V1.0.0, Sept. 2023.
recovery via deep learning with applications in MIMO-based grantfree [481] W. P. 5D, “The ITU-R (framework) for IMT-2030,” International
random access,” IEEE Journal on Selected Areas in Communications, Telecommunication Union (ITU), Tech. Rep., July 2023.
vol. 39, no. 3, pp. 788–803, May 2021. [482] Qualcomm, “Vision, market drivers, and research directions on the path
[459] S. Li, W. Zhang, Y. Cui, H. V. Cheng, and W. Yu, “Joint design of to 6G,” Tech. Rep., Dec. 2022.
measurement matrix and sparse support recovery method via deep auto- [483] Ericsson, “6G – Connecting a cyber-physical world,” Tech. Rep., Feb.
encoder,” IEEE Signal Processing Letters, vol. 26, no. 12, pp. 1778– 2022.
1782, Dec. 2019. [484] Nokia, “Toward a 6G AI-Native Air Interface,” Tech. Rep., March
[460] X. Wu, S. Zhang, and J. Yan, “A CNN architecture for learning device 2021.
activity from MMV,” IEEE Communications Letters, vol. 25, no. 9, [485] Huawei, “6G: The Next Horizon,” Tech. Rep., Nov. 2021.
pp. 2933–2937, Sep. 2021. [486] Samsung, “6G: The Next Hyper-Connected Experience for All,” Tech.
[461] Y. Cui, S. Li, and W. Zhang, “Jointly sparse signal recovery and support Rep., July 2020.
recovery via deep learning with applications in MIMO-based grant-free
random access,” IEEE Journal on Selected Areas in Communications,
vol. 39, no. 3, pp. 788–803, Mar. 2021.
[462] Y. Bai, W. Chen, B. Ai, Z. Zhong, and I. J. Wassell, “Prior information
aided deep learning method for grant-free NOMA in mMTC,” IEEE
Journal on Selected Areas in Communications, vol. 40, no. 1, pp. 112–
126, Jan. 2022.
[463] J. Johnston and X. Wang, “Model-based deep learning for joint activity
detection and channel estimation in massive and sporadic connectivity,”
IEEE Transactions on Wireless Communications, vol. 21, no. 11, pp.
9806–9817, Nov. 2022.
[464] W. Zhu, M. Tao, X. Yuan, and Y. Guan, “DeeplLearned approximate
message passing for asynchronous massive connectivity,” IEEE Trans-
actions on Wireless Communications, vol. 20, no. 8, pp. 5434–5448,
Aug. 2021.
[465] Y. Polyanskiy, “A perspective on massive random-access,” in 2017
IEEE International Symposium on Information Theory (ISIT), 2017,
pp. 2523–2527.
[466] S. S. Kowshik and Y. Polyanskiy, “Fundamental limits of many-
user MAC with finite payloads and fading,” IEEE Transactions on
Information Theory, vol. 67, no. 9, pp. 5853–5884, 2021.
[467] J. Gao, Y. Wu, S. Shao, W. Yang, and H. V. Poor, “Energy efficiency of
massive random access in MIMO quasi-static rayleigh fading channels
with finite blocklength,” IEEE Transactions on Information Theory,
vol. 69, no. 3, pp. 1618–1657, 2023.
[468] V. K. Amalladinne, J.-F. Chamberland, and K. R. Narayanan, “A coded
compressed sensing scheme for unsourced multiple access,” IEEE
Transactions on Information Theory, vol. 66, no. 10, pp. 6509–6533,
2020.
[469] A. Fengler, P. Jung, and G. Caire, “SPARCs for unsourced random
access,” IEEE Transactions on Information Theory, vol. 67, no. 10,
pp. 6894–6915, 2021.
[470] E. Marshakov, G. Balitskiy, K. Andreev, and A. Frolov, “A polar
code based unsourced random access for the gaussian MAC,” in IEEE
Vehicular Technology Conference (VTC-Fall), 2019, pp. 1–5.
[471] A. K. Pradhan, V. K. Amalladinne, K. R. Narayanan, and J.-F. Cham-
berland, “Polar coding and random spreading for unsourced multiple
access,” in IEEE International Conference on Communications (ICC),
2020, pp. 1–6.
[472] ——, “LDPC codes with soft interference cancellation for uncoordi-
nated unsourced multiple access,” in IEEE International Conference
on Communications, 2021, pp. 1–6.
[473] 3GPP, “Feasibility study for evolved universal terrestrial radio access
(UTRA) and universal terrestrial radio access network (UTRAN),” 3rd
Generation Partnership Project (3GPP), Technical Report TR 35.912
V8.0.0, Dec. 2008.
[474] ZTE, “Study on non-orthogonal multiple access (NOMA) for NR,” 3rd
Generation Partnership Project (3GPP), document RP-181403, 3GPP
TSG RAN Meeting #80, June 2018.
[475] 3GPP, “Study on new radio (NR) access technology,” 3rd Generation
Partnership Project (3GPP), Technical Report TR 38.912 V15.0.0, June
2016.