SemanticCommunication/paper3.txt

IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026 2965
Resource Allocation in Wireless Semantic
Communications: A Comprehensive Survey
Chujun Zhang , Linyu Huang , Member, IEEE, and Qian Ning
Abstract—Withtheadventofsixth-generationmobilecommu- Metaverse require wireless communication networks to trans-
nicationtechnology(6G)andtheemergenceoffutureapplication mit huge amounts of data. Wireless communication networks
scenarios such as Metaverse and digital twin (DT), the exist-
mustachieveanextremelylowtransmissiondelayinscenarios
ing traditional wireless communication technology based on
such as autonomous driving and telemedicine. The emergence
Shannon’s information theory has not been able to meet the
increasing demand for data transmission. Semantic commu- of these applications presents new challenges to traditional
nications (SemCom), which greatly reduces the amount of communication systems.
information transmitted and alleviates the burden of communi- In the face of such a large communication load, how
cation by transmitting the meaning behind the information, has
can one go beyond Shannon’s limit to the future? Inspired
been considered a promising 6G enabler. SemCom’s resource
by the three levels of the previous communication problem,
allocation is critical to the system’s reliability and effectiveness.
Compared to traditional wireless communication systems, the a new communication paradigm, semantic communication
system architecture and performance metrics of SemCom have (SemCom) [2], [3], [4], has been proposed to shift the com-
undergone significant changes, making it difficult for traditional munication paradigm to the semantic and effectiveness levels.
resource allocation strategies to adapt well to this new architec-
In traditional communication systems, data is compressed
ture. However, the issue remains unresolved and inadequately
bythesourceencoder,andredundancyisaddedtothechannel
researched.Inordertoprovideresearcherswithvaluableinsight
to promote follow-up research, this paper reviews the latest encoder to improve its robustness to interference/noise in the
research results in recent years and presents an overview of channel. At the destination, a reverse process is performed to
research progress in the field of resource allocation in wireless recover the original sent data. The transmission and reception
SemCom.
of signals do not involve any intelligence and the semantic
Index Terms—Performance metrics, resource allocation, information is omitted [5].
semantic communications, semantic similarity. However, in a SemCom system, the semantic source and
destination are intelligent agents that can perform various
highly intelligent algorithms. Semantic coding replaces tradi-
I. INTRODUCTION tional source coding through deep learning (DL) and other
A. Context technologies to extract semantic information. Unlike tra-
ditional communication systems, which are easily affected
IN1949,WeaverexpandedShannon’stheorytothreelevels:
by channel conditions, SemCom performs well, especially
technical level, semantic level, and effectiveness level [1].
at low signal-to-noise ratios (SNR), because only semantic
The lowest level is the technical level, which is mainly
information is transmitted. Goal-oriented SemCom or task-
responsible for the accurate and effective transmission of
oriented SemCom is a subset of SemCom that pays more
information symbols; the middle level is the semantic level,
attention to the effectiveness level. Specifically, it focuses on
which points to the transmission of information symbols to
the efficient use of semantic information for the successful
convey the desired meaning; the upper level is the effective-
execution of tasks at a suitable time [6]. The receiver in a
ness level, which aims at effectively performing intelligent
goal-oriented SemCom is interested in the significance and
tasks and providing the needed communication efficiency on
effectiveness (semantics) of the transported source message
the lower two levels. Traditional communications operate at
to achieve a certain task or goal [7]. In summary, SemCom
the technical level, focusing on accurate bit transmission.
is becoming an excellent solution to the above questions.
However, they transmit all information, including useless and
SemCom is also regarded as a key enabling technology for
irrelevant data, to the receiver, leading to channel resource
6G, and it is an important step towards the future of wireless
waste. As sixth-generation mobile communication technology
communication.
(6G) emerges, scenarios such as Digital Twin (DT) and
Received 25 February 2025; revised 23 April 2025 and 17 June 2025;
accepted 30 July 2025. Date of publication 4 August 2025; date of current B. Resource Allocation
version 2 January 2026. This work was supported by the National Natural
In general, resource allocation refers to a set of method-
ScienceFoundationofChinaunderGrant61801318.(Correspondingauthor:
LinyuHuang.) ologies to achieve goals by efficiently allocating resources
The authors are with the College of Electronics and Information and using resource allocation methods based on resource
Engineering,SichuanUniversity,Chengdu610065,China(e-mail:zhangchu-
availability. The resource allocation problem in wireless com-
jun@stu.scu.edu.cn;lyhuang@scu.edu.cn;ningq@scu.edu.cn).
DigitalObjectIdentifier10.1109/COMST.2025.3595168 munications and SemCom is mapped into a mathematical
1553-877X(cid:2)c 2025IEEE.Allrightsreserved,includingrightsfortextanddatamining,andtrainingofartificialintelligence
andsimilartechnologies. Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seehttps://www.ieee.org/publications/rights/index.htmlformoreinformation.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2966 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
clarifying the unique contribution of this work. Table I pro-
vides a comparison between our survey and representative
prior works. Although there are some recent surveys on
resource allocation in other communication scenarios that
provided us with great insights, such as edge comput-
ing [16], fifth-generation mobile communication technology
(5G)-and-beyondmobileedgecomputing(MEC)[17],Internet
of Things (IoT) enabled vehicular edge computing [18],
energy-efficient Orthogonal Frequency Division Multiplexing
(OFDM) enabled networks [19], and ultra-dense networks
(UDNs) [20]. Resource allocation is a critical and under-
explored aspect of SemCom, it significantly differs from
traditional communication systems, as it involves unique
allocatable resources like semantic fidelity and computation
overheadforsemanticprocessing,alongsidetraditionalfactors
such as bandwidth and power. Moreover, SemCom introduces
novel performance metrics that will make the object function
Fig.1. TheintegratedframeworkofresourceallocationinSemCom. more complicated, which we will provide a more explicit
descriptioninSectionIII.Thesedifferenceshighlighttheneed
to focus specifically on resource allocation in SemCom, as
optimization problem by modeling the network structure and
existing surveys tend to overlook the unique challenges and
designing the objective function. In resource allocation, the
optimization strategies required in this domain. By dedicating
available resources for allocation are optimization variables;
our review entirely to this topic, our aim is to fill this
the availability of resources and other inherent conditions are
gap and provide a comprehensive and systematic overview
constraints; the objective function is the function to evaluate
of how resource allocation can be effectively addressed
the system performance of achieving a specific goal; the
within the context of SemCom. Therefore, we review from
resourceallocationalgorithmisacombinationofoptimization
multiple perspectives, including SemCom network models,
techniques that are used to solve this optimization problem.
performance metrics, resource allocation optimization algo-
TheresourceallocationalgorithmsinSemComcanbedivided
rithms, as well as challenges and future research directions,
into centralized and distributed ways. The centralized algo-
providing researchers with a new, comprehensive, and rich
rithm includes techniques based on convex optimization and
perspective.
other mathematical methods, and based on deep reinforce-
ment learning (DRL), etc. The distributed algorithms include
techniques based on multi-agent deep reinforcement learning D. Research Methodology
(MADRL) and matching theory, etc. The integrative frame-
In this subsection, the process followed to collect the
workofresourceallocationinSemComisillustratedinFig. 1.
references used in this study is described. The methodology
However, compared to the traditional wireless communica-
includestheselection,inclusion,andexclusioncriteriaapplied
tionsystem,theSemComsystemarchitectureandperformance
toensurethequalityandrelevanceofthereferences.Thesteps
metrics have undergone tremendous changes, making it diffi-
followed in the research process are as follows:
cult for traditional resource allocation strategies to adapt well
• Literature Search: The search was performed using
to this new architecture. In the next section, we willprovide a
databases such as Google Scholar, IEEE Xplore, and
moredetaileddescriptionofthedifferencebetweentraditional
ScienceDirect. The primary focus was on peer-reviewed
communication and SemCom in terms of resource allocation
journal articles, conference papers, books, and other
and why it is important.
reputablesourcesrelatedtoSemCom,resourceallocation,
optimization, and wireless communication networks.
C. Related Surveys and Motivation • InclusionCriteria:Tobeincludedinthestudy,references
The development of SemCom has led to the publication must meet the following criteria: 1) Published in a peer-
of numerous surveys in recent years. Existing surveys on reviewed journal or conference proceedings. Directly
SemCom may address resource allocation to some extent, related to resource allocation in SemCom networks
they mostly provide a global perspective of SemCom and or relevant areas such as optimization, deep learning
often focus on broader aspects such as system architectures, techniques, and wireless communications. 2) Except
semantic information theory, enable techniques or general for some classic and fundamental literature, references
applications of SemCom [3], [4], [5], [6], [8], [9], [10], should be published within the last 10 years to ensure
[11], [12], [13], [14], [15]. However, our work is the first to that the research is up-to-date and relevant. 3) For
present a dedicated and in-depth review of resource allocation researchpapers,theoreticalandempiricalstudiesmustbe
in SemCom systems. To highlight this distinction, we have included.
added a comparative table that outlines the resource allo- • Exclusion Criteria: References that met any of the
cation aspects covered (or not) by existing surveys, thereby following conditions were excluded from the review:
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2967
TABLEI
COMPARISONOFEXISTINGSEMCOMSURVEYSANDTHISWORK
1) Studies that were completely not related to SemCom
or resource allocation; 2) Non-peer-reviewed sources or
articleswithoutsufficientmethodologicalrigor.3)Studies
published more than 10 years ago unless they introduced
foundational theories or seminal works that remain rele-
vant to current research.
• Information Extraction and Analysis: After finalizing the
selected references, the key information was extracted
andanalyzedinresourceallocation.Thisincludedunder-
standing the research objectives, methodologies used,
findings, and how each study contributed to advance the
understanding of SemCom resource allocation.
Fig.2. Thedistributionofpaperssurveyedbyyearandsource.
E. Contributions and Organization
This paper reviews the current state of research in the
period2021-2025(February)onresourceallocationinwireless divided into end-to-end (E2E) and multi-user situations.
SemCom.Fig. 2showsthedistributionofthearticlessurveyed We also investigated the use of the next generation of
by year and source. The report encompasses arXiv articles multipleaccess(NGMA)technologiesandhybridseman-
and website articles, while the conference category includes tic/bit communications in SemCom resource allocation.
conference and symposium papers, and the journal category WegivetheoverviewofresourceallocationinSemCom,
includes journal and magazine articles. The contributions of thereby explaining the reason why resource allocation
this paper can be summarized as follows: in SemCom is important for the theoretical perspective
• We first explain the basics of resource allocation in and reality perspective, clarifying the unique specific
SemCom and introduce the network model in the cur- challenges that inherently exist in the resource allocation
rent literature on SemCom resource allocation, which is of SemCom.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2968 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
Fig.3. Roadmapofthesurvey.
• The construction of the objective function is the core of summarize the different centralized and distributed resource
the optimization problem modeling, so we introduce the allocation optimization algorithms in detail. Section VI points
performance metrics in the SemCom resource allocation out the challenges and possible future research directions.
in detail. We mainly summarized the construction meth- Finally, Section VII summarizes this survey. Fig. 3 shows the
ods into two types. One is utilizing traditional resource organization and structure of this survey paper.
allocationperformancemetrics,suchasdelayandenergy
consumption. The other type is based on the semantic
II. BASICSOFRESOURCEALLOCATIONINSEMCOM
similarity, establishing new performance metrics.
ThissectionwillexplainthebasicsoftheSemComresource
• We discuss in detail different optimization algorithms
allocation problem. We provide an overview of SemCom,
in the allocation of SemCom resources, which are
followed by a review of the fundamental network mod-
divided into centralized and distributed algorithms.
els found in various SemCom resource allocation studies.
Centralized algorithms include algorithms based on
Furthermore, we give an explicit contrast between bit-level
convex optimization and other mathematical methods,
and semantic-level modeling in Table III, which provides a
algorithms based on DRL, and heuristic algorithms.
side-by-side comparison between the two paradigms, high-
Distributed algorithms include methods based on
lightingtheirrespectivetargets,metrics,modelingapproaches,
MADRL, matching theory, and auction. These meth-
and optimization goals. Next, we provide an overview of
ods are summarized in three comprehensive tables for
resource allocation in SemCom. Besides, we give the taxon-
comparison.
omy of system framework establishment in Fig. 5. Lastly, we
• Through the analysis presented above, we propose future
summarize the literature in Table IV.
researchdirectionsandseveralchallengestobesolvedin
the field of SemCom resource allocation.
The remainder of this paper is organized as follows. A. Overview of SemCom
Section II introduces the basic architecture of the SemCom Traditional communications aim to reach the technical
resource allocation problem. Section III presents traditional level, which means achieving a high data transmission rate
performancemetrics,thedefinitionofsemanticsimilarity,and and a low symbol error rate. However, the basic idea of
new semantic-based performance metrics. Sections IV and V SemCom is to extract the “meanings” or “features” of the
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2969
Fig.4. Acomparisonbetweenthebasicend-to-endnetworkarchitectureoftraditionalcommunicationandSemCom.
source and “interpret the semantic information” at a des- infer the received information to complete the recovery
tination. Therefore, SemCom surpasses traditional bit-level of the received semantic information.
transmissiontoachievesemantic-leveltransmission,leadingto Currently, most of the research literature is focused on three
significant changes in the design of the network architecture. types of sources: text signal, image signal, and speech signal.
Moreover, thereis very littleliteraturethat points the research
1) Basic End-to-End SemCom: A comparison between the direction to multi-modal tasks [22].
basic end-to-end (E2E) network architecture of traditional Text: Text SemCom systems have been widely stud-
communication and SemCom is shown in Fig. 4. Fig. 4a ied. Various DL techniques are used to represent the
illustrates the typical traditional E2E communication archi- underlying meaning of texts. DL-enabled semantic codecs
tecture, where the source encoder receives the transmitted have been through the early Long Short Term Memory
dataandcompressesitinitially,partiallyeliminatingredundant (LSTM)-based models [23], [24], to today’s Transformer-
information through source encoding. The channel encoder basedmodels[25], [26].In2018,Farsadetal.[23]proposeda
adds redundancy in various coding ways to combat noise jointsource-channelcoding(JSCC)schemefortextSemCom,
and attenuation in the channel, thereby enhancing its anti- in which the encoder and decoder are implemented by two
interference capability and error correction ability. At the LSTM networks. Compared to the single source channel
destination, a reverse process is conducted to recover the coding (SSCC) scheme, the DL-based JSCC scheme per-
original sent data. We can see in Fig. 4b that SemCom forms better [23]. In 2021, Xie et al. [25] proposed the
primarily differs from traditional end-to-end architecture in DeepSC framework by fine-tuning the basic structure of
three key ways. Transformer[27].DeepSCcanadapttodifferentchannelenvi-
• SemanticCoding:ASemComsystemextractstheseman- ronments, perform well under low SNR, and have excellent
tic information (features) from the original data through robustness. The author of [28] proposed a semantic extraction
semantic coding enabled by technologies such as DL scheme based on the entity recognition model (NER) and
and then encodes these features for channel coding. Due LSTM that transforms the transmitted sentence into multiple
to the implicit meaning inherent in the message under triplets of semantic importance, and important triplets will be
consideration, the amount of redundant data removed is allocated more transmission resources to improve reliability.
significantly greater than that achieved by source coding. The authors of [29] introduce a life-long model updating
Not like semantic segmentation in computer vision, in approach in which the receiver can learn from previously
SemCom, all communication parties must maintain a received messages and automatically update the rules to
high degree of consistency in semantic expression and reasoningforhiddeninformationwhennewunknownsemantic
understanding, which poses a challenge to semantic entities and relations have been discovered.
compression. Image: The image SemCom system is similar to the text
• Knowledge Base: Another important feature of SemCom SemCom system, and there is much research on it. In con-
is that it is a knowledge-based system [21]. This means trast to text systems, image SemCom systems extract the
that semantic source and semantic purpose can be like original image’s features (which, in this context, represent
the human brain, through self-learning to establish their the image’s “meaning”) and extensively utilize convolutional
own background knowledge bases (KBs) to guide the neural networks (CNNs). In addition, in many task-oriented
transmitter to obtain multi-level semantic knowledge SemCom systems (such as image classification tasks), the
descriptionofsourcedata,semanticinference,estimation image does not need to be reconstructed at the receiver. In
of transmission environment, and semantic requirements 2019, Bourtsoulatze et al. [30] first proposed an end-to-end
of downstream tasks. The system performs semantic imagetransmissionsystemusingCNN’sJSCCscheme,which
coding and directs the receiver to execute the inverse has better performance than traditional image transmission
process, known as semantic decoding. methods. In 2022, Dong et al. [31] proposed a layer-based
• Semantic Decoding: Based on technologies such as semantic communication system for images (LSCI), and
semantic KBs and DL, the receiver can understand and the concept of semantic slice-models (SeSM) is proposed
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2970 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
to enable flexible model resemblance under the different samples using CNN and the gated recurrent unit (GRU)-
requirementsofthemodelperformance,channelsituation,and based bidirectional RNN (BRNN) modules. In the image
transmissiongoals.In2023,Lokumarambageetal.[32]imple- SemCom systems [30], [31], [32], [33], [39], networks
mented a semantic communication-based end-to-end image such as CNN and GAN are often used, and the input
transmission system, where a pre-trained GAN network is is a n-dimensional image, not a sequence like text or
used at the receiver as the transmission task to reconstruct speech. Simulation results show that SemCom performs
the realistic image based on the semantic segmented image well especially under the low SNR. This is because the
at the receiver input. Kadam and Kim [33] proposed a joint extractedsemanticfeaturesreduceredundancywhichwill
CNN-LSTM-based SemCom model in which the semantic use more channel resources. After semantic extraction,
encoder of a camera extracts the relevant semantics from the high-level semantic representations are less sensitive to
raw images, resulting in a novel approach to the problem of noise, which makes the SemCom system more robust.
predicting vehicle counts. • Knowledge Graph-based semantic extraction: It extracts
Speech: Unlike the previous two modes of the SemCom structured information as semantic triples (subject, pred-
system, the speech signal possesses more complex icate, object) to form a knowledge graph, which
performance characteristics, including speech speed, volume, enhances interpretability and enables reasoning but
tone, and dialect, all of which can express the same meaning. requires high construction and maintenance costs. The
The general approach is to convert the speech into text for semantic information of a knowledge graph is typically
processing. However, the same text information expressed expressed as triples in the form of (head, relation,
in different intonations will produce different meanings. tail). From a piece of text data, multiple triples can be
Therefore, the process of voice semantic transmission is more extracted, and these triples can be used to characterize a
complex and challenging to manage [34], [35]. The majority knowledge graph. The knowledge graph extracted from
of the source modes in SemCom’s resource allocation are text each sample data Tn is represented as
(cid:4) (cid:5)
and image modes. Currently, there is no relevant research on
the allocation of resources for the SemCom speech system. Gn = ε1 n ,ε2 n ,...,εm n ,...,εM n , (2)
In the following content, we will introduce and compare
these papers comprehensively and organize them in tables for
whereεm
n isthem-thtripleinknowledgegraphGn,M is
reference. the total number of triples. The triple
εm
n can be written
in the following form:
2) Mathematical System Modeling of SemCom: While the
previous section has highlighted the core components of εm n =(hn m,rn m,tn m ), (3)
semanticcommunication,itisequallyimportanttounderstand
how these elements integrate into a mathematical frame- where hn
m
is the head entity of triple
εm
n , tn
m
is the tail
m
work. We will introduce some essential parts of mathematical entity, and rn is the relation of head and tail entities.
modelinginpapers,mainlyonsemanticextractionandseman- For text, the work in [40] used an information extrac-
tic metrics (it will be discussed thoroughly in Section III). tion system to extract semantic triples from texts and
• NN features-based semantic extraction: It utilizes deep modeled as KGs, and the receiver used a graph-to-text
learning models for end-to-end semantic encoding, generative algorithm to recover the original texts based
offering strong contextual understanding but lacking on the received triples. In [41], a cognitive text semantic
interpretability and explicit semantic relationships. In communication framework is proposed by exploiting
such an approach, the encoded symbol stream can be knowledge graph. For image, the scene graph (SG) is
represented by a visual KG that describes visual relationships between
(cid:2) (cid:3) entities,theauthorsin[42]and[43]usedobjectdetection
x=Cα Sβ(s) , (1) and RE algorithms to extract SG from images.
3) Multi-User SemCom and Multiple Access Techniques:
where, Sβ(·) is the semantic encoder network with The previous section introduces several end-to-end SemCom
parameter set β and Cα(·) is the channel encoder with systems.However,alltheabovesystemsdonotinvolvemulti-
parameter set α, the specific networks are various in user transmission. In general, the connection density of 5G is
different systems. In text SemCom systems [25], [36], 106devicespersquarekilometer,whiletheconnectiondensity
networks such as Transformer, BERT, or LSTM are ofthe6Gnetworkwillincreaseto10timesthatof5G,andthe
utilized for semantic extraction, s = [w1 ,w2 ,...,wL] regional traffic density should be 100 times that of 5G, which
denotes the original sentence, wl represents the l- requires a significant improvement of spectral efficiency [5].
th word in each sentence. In speech SemCom Moreover, the knowledge base within the SemCom system
systems [34], [37], [38], ResNet, Transformer, CNN and may vary significantly. Therefore, from a more realistic point
RecurrentNeuralNetwork(RNN)areutilizedforseman- ofview,itisnecessarytodesignamulti-userSemComsystem.
ticextractionindifferentstudies,theinputs isthespeech Notably, we only survey the multi-user SemCom system in
sample sequence, s =[s1 ,s2 ,...,sW] with W samples, papers on resource allocation in SemCom, not all SemCom-
where sw is the w-th item in s and it is a scalar value. related papers.
In the DeepSC-ST system [38], text-related semantic In the resource allocation problem of multi-user SemCom,
features are extracted from the input speech spectrum the classical multiple access (MA) techniques such as
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2971
TABLEII
THECOMPARISONOFDIFFERENTNGMATECHNIQUES
frequency division multiple access (FDMA) [36], [44], [45], via SIC, user-k decodes its desired private stream sk, so the
[46],[47],[48],orthogonalfrequencydivisionmultipleaccess private rate of user k is
(OFDMA) [40], [49], [50], [51], [52], [53], [54], [55], (cid:6) (cid:8) (cid:8)hHp (cid:8) (cid:8)2 (cid:7)
[
(T
56
D
]
M
, [
A
57
)
]
[
,
61
[5
],
8]
[
,
62
[
]
59
te
]
c
,
h
[
n
6
i
0
q
]
ue
o
s
r
a
t
r
i
e
m
m
e
o
d
s
i
t
v
ly
is
u
io
s
n
ed
m
.H
u
o
lt
w
ip
e
l
v
e
er
a
,
c
w
ce
i
s
th
s Rk =Blog2 1+ (cid:8)
(cid:8)h k p
k
j
(cid:8)
(cid:8)2
k
+Nk
, (6)
the continuous development of communication technology,
so that the achievable total rate of user k is
researchers have begun to explore the application of the
combination of next-generation multiple access (NGMA) and Rk,tot =Ck +Rk . (7)
SemCom in resource allocation. Before comparing different
In [64] and [65], the authors used SDMA as the multiple
MA techniques in papers on resource allocation in SemCom,
access method and established an SDMA-based multiuser
wesummarizedthesethreekeyNGMAtechniquesinTable II.
probabilistic SemCom (PSC) framework that considers both
As in Table II, spatial division multiple access (SDMA)
transmission and computational consumption. The authors
treats the interference of other users fully as noise. Non-
of [61] proposed a new semantic-aware resource allocation
orthogonal multiple access (NOMA) will employ successive
scheme in the integration of the radio frequency energy
interference cancellation (SIC) at one user to fully decode the
harvesting (EH), cognitive radio (CR), and NOMA scenario.
interference. Rate splitting multiple access (RSMA), based on
An uplink network consisting of multiple primary users (PU)
theconceptofratesplitting(RS),isconsideredtobeapromis-
using TDMA and a secondary user (SU) using NOMA and
ing physical layer transmission paradigm for non-orthogonal
PUmultiplexingspectrumisconsidered.Inthebackgroundof
transmission, interference management, and multiple access
PSC, the work in [66] studied the joint communication and
strategies in 6G. The main idea of RSMA is to divide
computation design in the reconfigurable intelligent surface
user messages into common and private parts (sc and sk)
(RIS)-assisted industrial Internet of Things (IIoT).
and to be able to partially decode interference and partially
Compared to SDMA and NOMA, the research on resource
treat interference as noise, which is in stark contrast to the
allocation in the combination of RSMA and SemCom is
extreme interference management strategies used in SDMA
obviously more [67], [68], [69], [70], [71], [72]. In [67],
and NOMA. The flexibility of RSMA makes it perform well
the optimization problem of the energy consumption of the
at all levels of interference [63]. In RSMA, pk and pc are
downlink SemCom network with RSMA is studied. The
the power allocated to private messages and the common
authors of [69], [70] focused on the PSC framework based
message. The common stream sc is decoded first by treating
on RSMA; reference [70] expanded the work of [64], and
the interference from private streams s1 and s2 as noise. As
the multiple access mode was changed from uplink SDMA
sc contains part of the intended message as well as part of
to downlink RSMA, while the authors of [69] paid more
the message of the interferer, it enables the ability to partially
attention to the energy-saving design of the PSC system.
decode interference and partially treat interference as noise.
The simulations of the above literature [68], [70] compare
The instantaneous rates for decoding the common streams at
the SDMA and NOMA-based schemes. The results show
user-k are
(cid:6) (cid:7) that the RSMAbasedschemeperformanceisthebestinterms
|h p |2
Rc,k =Blog2 1+
|h k p 1 |2 +
k
|h k
c
p 2 |2 +Nk
. (4) of
4
to
)
ta
H
l
y
s
b
e
r
m
id
anti
S
c
em
tra
a
n
n
s
t
m
ic-
is
B
s
i
i
t
on
C
ra
o
t
m
e.
munication: While most
research on resource allocation focuses solely on SemCom
To guarantee that common message sc is decoded by both
itself, the coexistence of SemCom and bit communication
users, the common rate shall not exceed
(BitCom) modes has also received attention [44], [45], [73],
Rc =min{Rc,1 ,Rc,2 }. (5) [74], [75], [76], [77], [78]. SemCom is more suitable for low
signal-to-interference-plus-noise ratio (SINR) and resource-
Denote Ck as the common rate portion of user-k: C1+C2 = constrained scenarios, while BitCom performs well in high
Rc .Oncesc isdecodedandremovedfromthereceivedsignal SINR regions. Moreover, it is not possible to completely
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2972 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
TABLEIII
COMPARISONOFMODELINGANDFRAMEWORKS:TRADITIONALVS.SEMANTICCOMMUNICATION
replace BitCom’s current huge infrastructure and user bases (suts/s, which will be discussed in the next section), which
at once. In the future, hybrid semantic/bit communication is unified into the semantic correlation measure to measure
networks will become an inevitable and persistent example of the network performance. Compared with the combination
intermediate networks [78]. The authors of [73] proposed a mode of SemCom and BitCom in [75], [78], the works
novel multi-carrier E2E system that combines both semantic in[44], [45], [74]bothstudiedanotherformofcoexistenceof
and Shannon (bit) communications, in which both the BS and SemCom and BitCom separately in the downlink and uplink
theusercancommunicatebychoosingtoutilizeeitherbitCom transmission. A semantic relay (SemRelay)-aided system was
or SemCom on each subcarrier. For resource allocation in proposed. We use the uplink transmission scenario in [74] for
the coexistence of semantic and bit communication networks, explanation: from the users to SemRelay using BitCom, from
the focus is how to combine the measurement of the two. SemRelay to the BS using SemCom. In the User-SemRelay
us
In [78], a bit-to-message (B2M) conversion function is used link, FDMA is adopted, the achievable rate Rn is:
to convert the rate metric into the capacity of the semantic (cid:6) (cid:7)
c
p
h
e
a
r
n
u
n
n
e
i
l
t
(
t
i
i
.
m
e.
e
,
,
th
m
e
s
a
g
c
/
h
s)
i
,
ev
le
a
t
bl
R
e
i
m
j(
e
·
s
)
sa
d
g
e
e
no
ra
te
te
th
in
e
u
B
n
2
it
M
so
f
f
u
m
nc
e
t
s
io
sa
n
g
o
es
f
Rn us =Bn us log2 1+ |h
B
n u
n u
s
s
|2
N
p
0
n u , (11)
the SemCom link between mobile user (MU) i and BS j, its
instantaneous achievable message rate in time slot t should be where N0 is the power spectral density of the additive white
Mi S j(t)=β ij(t)R ij (cid:2) bij log2(1+γ ij(t)) (cid:3) . (8) G of au u s s s e ia r n n n , o h is n u e s (A d W en G ot N es ), t p h n u e d c e h n a o n t n e e s l th g e ai t n ran f s ro m m iss u io s n er po n we to r
us
Here, β ij(t), bij(t), and γ ij(t) represents the knowledge- SemRelay and Bn denotes the bandwidth allocated to the
matching degree, bandwidth, and SINR between MU i and its link. The transmission delay for each user n is given by
communicationcounterpartatslott.ComparedtotheSemCom tn us = R D u n s . Here, Dn is the volume of text data in bits. The
n
link,theinstantaneous achievable messagerateoftheBitCom computation time cost for semantic compression at SemRelay
c
link in slot t is given by is t , The achievable rate of the SemRelay-BS link is:
(cid:2) (cid:3) (cid:6) (cid:7)
Here, bij(t
M
),
i B j
an
(t
d
)
γ
=
ij(
ρ
t
ij
)
R
de
ij
no
b
t
i
e
j
s
lo
th
g
e
2(
s
1
am
+
e
γ
t
i
h
j(
in
t)
g
)
a
.
s in Eq. (
(
8
9
)
)
,
R sb =B sb log2 1+ |h
B
s
s
b
b
|
N
2 p
0
s , (12)
and ρ ij is an average B2M transformation ratio to measure
s sb
network performance with a message-related metric unified where, pn denotes the transmission power of SemRelay, hn
sb
with SemCom. denotes the channel gain from SemRelay to BS, and Bn
IftakingbothSemComandBitComintoaccount,useyij to denotes the bandwidth allocated to the link. The transmission
denote the communication mode selection (yij = 1 represents delay for SemRelay is given by t
sb
=
D
R
S
s
e
b
m.
Here, D
Sem
that the SemCom mode is selected for the link between MU is the total number of bits for the compressed semantic
c Sem
i and BS j, and yij = 0 indicates that the BitCom mode is information. The explicit expression of t and D can be
all
selected), the time-averaged message rate of each link is found in [74]. So the overall latency t is
Mij = 1 (cid:9)N (cid:10) yijMi S j(t)+ (cid:2) 1−yij (cid:3) Mi B j (t) (cid:11) . (10) t all =max{tn us,∀n}+t c +t sb. (13)
N
t=1 Another difference from [75], [78] is that [44] and [45]
In [75], the equivalent transformation method in [36] is transformthesemanticrateintothebitratetounifythesetwo
usedtotransformthebitrateintotheequivalentsemanticrate rate metrics into a bit-based metric (bit/s).
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2973
B. Overview of Resource Allocation in SemCom • Reality perspective: In the context of 6G, the amount of
data generated by terminal devices around the world is
We have given a brief description of resource allocation in
explosively increasing. Coordinating limited resources to
SectionI.Wearenowgivingamoredetaileddescriptionofthe
betterprocessthesedatarequiresanappropriateresource
differencebetweentraditionalcommunicationandSemComin
allocation strategy. Data from different application sce-
terms of resource allocation, as well as why it is important.
narios may have different service requirements. Vehicles
1) The Difference With Traditional Communications:
in autonomous driving scenarios need to process data
• Optimization Problem: Compared with traditional wire-
in milliseconds to ensure traffic safety. Therefore, ultra-
lesscommunication,SemCom’snetworkarchitecturehas
low latency is its main goal. Semantic sensing systems
changed in many aspects, from codec level to multiple
assisted by uncrewed aerial vehicles (UAVs) usually
access modes. Due to the inexplicability of neural
pay more attention to the long battery life and expect
networks, it is difficult to derive closed-form expressions
to achieve low energy consumption. In addition, some
of some objective functions or variables. Therefore, the
mobile devices and IoT devices are designed to achieve
constructed optimization problem, from the objective to
lowdataprocessingcostsorachievethebestusersatisfac-
the constraints and optimization variables, differs signif-
tion. Therefore, appropriate resource allocation strategies
icantly from the traditional architecture.
are needed to meet these diverse needs.
• Optimization Algorithm: As artificial intelligence and
machine learning technology continue to advance, an
3) Specific Challenges of Resource Allocation in SemCom:
increasing number of intelligent methods have emerged
SemCom brings fundamental shifts to the modeling, evalu-
to address resource allocation problems. For exam-
ation, and optimization of wireless communication systems.
ple, neural networks are used to approximate the
These shifts give rise to several unique challenges that are
function in which closed-form expressions cannot be
rare or nonexistent in traditional communications and funda-
obtained,anddeepreinforcementlearning(DRL)hasalso
mentally affect how resource allocation must be performed.
become a powerful tool for solving complex resource
Although Section VI will discuss open research problems
allocation problems in recent years [79], [80], [81].
and promising future directions for SemCom, this subsec-
Though traditional methods like mathematical and con-
tion focuses on the specific and practical challenges that
vex optimization-based algorithms are still widely used,
currentlyariseinexistingSemComsystemdesignsandimple-
resource allocation in SemCom is more applicable
mentations. These challenges reflect the inherent complexity
to intelligent methods, and many papers tend to use
and unique characteristics of SemCom. By clarifying these
intelligentmethod-basedalgorithms.Wewillgiveacom-
concrete issues, we lay the foundation for understanding
prehensiveintroductiontothesealgorithmsinSectionIV.
why the optimization techniques in SemCom (which will be
2) The Reason Why Resource Allocation in SemCom is introducedinSectionsIVandV)arenecessary.Thesespecific
Important: challenges can be summarized as follows:
• Theoretical perspective: Firstly, from the perspective of • Tradeoff Caused by Semantic Compression Ratio:
thenetworkmodel,SemComhasalotofnewmodulesto There are many tradeoffs, such as the energy-latency
consider,suchasthesemanticencoderandtheknowledge tradeoff and the accuracy-efficiency tradeoff, that
base.MostSemComsystemsuseDLtechniquestoadopt already exist in traditional communications. However,
semantic extraction. Neural networks will bring about a SemCom introduces the new resource type, the seman-
lotofinexplicabilityandcanresultinthelackofaclosed tic compression/extraction ratio, which directly affects
form of part of the objective function. Moreover, as it communication, computation, and semantic fidelity. For
involves unique allocatable resources such as semantic instance, a higher compression ratio reduces the data
fidelity and computation overhead for semantic pro- size for transmission and saves transmission delay and
cessing. Optimization algorithms to optimize these new energyconsumption(communicationloadreduction),but
semantic-related variables directly have a great influence it lowers the semantic fidelity and task accuracy and
on the whole system performance. Besides, traditional needs more computing resources to process the seman-
performance metrics do not consider the meaning of tic extraction and recover, which results in the local
information. Using traditional performance metrics for extraction latency and energy consumption at the trans-
resource allocation may even lead to a decrease in mitter, the recover latency and energy consumption
system performance. Therefore, developing new metrics at the receiver (computation load increase). Moreover,
that match the characteristics of SemCom and designing for intelligent tasks, a higher compression ratio (lower
proper optimization algorithms to deal with the new in value) results in higher computing cycles for task
objective functions and constraints caused by these new processing, thus increasing task computing latency and
metrics can also have a positive influence on system energy consumption. These tightly coupled tradeoffs of
performance. Recently, there has been a lot of research computing, communication, and accuracy make resource
on the new performance metrics of SemCom, such as allocation in SemCom inherently more complex. The
semantic similarity, semantic energy efficiency, and task detailed description about how the semantic compression
successrate,ofwhichwewillgiveadetaileddescription ratio influences latency and energy consumption is in
in Section III-C. Section III-A.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2974 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
• Optimization with Non-differentiable and Implicit
Objectives: Many SemCom key performance metrics
rely on semantic similarity, which is difficult to express
analytically. These objectives often lack closed-form
expressions, are non-differentiable, or even implicitly
defined through closed-box models, which make
traditional optimization methods hard to apply well.
• Highly Coupled and Non-convex Optimization Variables:
Unlikeconventionalsystemswhereresourcevariablescan
often be decomposed or linearized, SemCom involves
complexcouplingbetweenvariablessuchascomputation
capacity, transmission power, and semantic compression
ratio. The resulting optimization problems are typi-
cally non-convex and nonlinear, in both objectives and
constraints.
• Task-related Semantic Information Transmission in Task-
oriented SemCom: In task-oriented SemCom systems,
the resource allocation is closely tied to the task-related
importance of the semantic information. For example,
tasks involving safety-critical or context-rich data trans-
mission (e.g., autonomous driving) may need to acquire Fig.5. Thetaxonomyofsystemframeworkestablishment.
high semantic fidelity, while other types of tasks may
tolerate coarse-grained transmission. This task depen-
dence necessitates adaptive resource allocation schemes to SemCom, followed by an examination of key network
thatalignwithtask-relatedsemanticinformationandtheir modelsusedinSemComresourceallocation.Wethendiveinto
utility, to complete the transmission of task-related and the core aspects of resource allocation within SemCom and
high semantic-importance features. At the same time, conclude with a preliminary review of the relevant literature,
it ensures the allocation of other resources (bandwidth, summarized in Table IV, which gives a preliminary summary
power, computing resources) to jointly optimize the of the literature on resource allocation in SemCom based on
overall system performance. source modal, communication mode, multiple access mode,
These challenges motivate the development of novel and resource allocation type. In the table, one(many)-to-many
optimization formulations and solution algorithms, as will be means one(many) BS(s)/edge server(s) to many users/end
discussed in the following sections. devices(EDs).Furthermore,weusethesymbol“–”toindicate
4) ResourcetoBeAllocatedinSemCom: Generallyspeak- that this property is not presented in the paper.
ing,thecurrentresearchonresourceallocationmainlyinvolves
computing, communication, and storage resources, with the
following resources typically requiring allocation.
III. PERFORMANCEMETRICSOFRESOURCEALLOCATION
• Computing resources: The computing frequency of Building upon the previous section, this section will review
CPUs/GPUs on the BS or user side, also known as the research on performance metrics in SemCom and the
computing capacity. formation of optimization objectives in different literature.
• Communication resources: The wireless resources used Usually, we evaluate a communication system based on
by BS or clients for data transmission, including band- its accuracy and effectiveness. The traditional communica-
width, power, etc. tion method is measured by the bit error rate and the bit
• Network parameter resources: The network-parameter transmission rate. For SemCom, accuracy can be measured
resources are the parameter settings in the SemCom by task performance and quantified by semantic similarity of
system, including the semantic compression ratio, the text transmission, character error rate of speech recognition,
neural network parameters, and other parameters or pol- etc. However, the efficiency of SemCom is usually difficult
icy settings. to measure and quantify [121]. As a result, it is critical
• Storageresources:EdgeserversorBSusethesehardware and challenging to establish new performance metrics for
storage resources to cache computing tasks and popular SemCom resource allocation. At present, the research of
content (such as road monitoring). SemCom resource allocation on constructing optimization
Inthispaper,wesummarizetheresourcestobeallocatedinthe objectives is mainly divided into two methods: based on
literature in Tables IX, XI, and XIII, and we need to mention traditional resource allocation performance metrics such as
thatthestorageresourcesareomittedsinceonlyonework[82] energy consumption, delay, and utility; and establishing new
considered them. The symbol “–” in the tables indicates that semantic-related performance metrics. See below for details.
this particular resource type is not allocated. Indifferentarticles,thesymbolsforthesamevariablesmay
In this section, we outline the foundational structure of the be inconsistent. To improve the reader’s understanding of the
SemCom resource allocation. It begins with an introduction compositionoftheseperformancemetrics,thispapermodifies
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2975
TABLEIV
COMPARISONOFPAPERSFOCUSINGONDIFFERENTSOURCEMODAL,COMMUNICATIONMODES,ANDSCENARIOS
theexpressionsinsomeliteratureandunifiesthemathematical cell, etc. At this time, we follow the expressions in their work
expressions of common variables in different literature, as and provide additional descriptions.
shown in Table V.
In most of the literature, for the subscript of a single
A. Traditional Performance Metrics in Resource Allocation
variable, we use n to represent the n-th user, m to represent
them-thsubchannel,btorepresenttheindexofBS,xn,m =1 1) Energy Consumption and Time Delay: Energy con-
to represent the association of user n and subchannel m, and sumption and delay/latency are two of the most traditional
xn,m =0torepresentdisassociation.Insomeotherreferences, and commonly used performance metrics in resource
the subscript may refer to a task, a user group in a cellular allocation.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2976 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
TABLEIV
(Continued.)COMPARISONOFPAPERSFOCUSINGONDIFFERENTSOURCEMODAL,COMMUNICATIONMODES,ANDSCENARIOS
VARIABL
T
E
A
S
B
D
LE
ES
V
CRIPTION
ρ is the compression ratio, fe is the computing capacity at
thetransmitter,andF(ρ,D) istherequiredcompressionCPU
cycles, which might be different across the literature. For
instance, [60] modeled F(ρ,D) as
αD
F(ρ,D)= , (15)
ρβ
where α > 0, β > 0 are constants relevant to the tasks.
Transmission latency is
T 2 =
ρD
, (16)
R
R is the transmission rate. Computing latency is
T 3 =
ρwDG
, (17)
fr
wherew istherequiredCPUcyclesperbittoprocessthetask,
For applications sensitive to delay, the design of a resource
fr is the allocated computing capacity at the receiver. We use
G todenotetheratioofcomputationintensityofsemanticdata
allocation algorithm to reduce latency is one of the main
to that of raw data. The increase is caused by computations
concerns [42], [58], [60], [73]. Delay modeling generally
forprocessingsemanticdataandcompensationsforenhancing
includes the following parts: a) semantic extraction latency
1 2 accuracy [59]. G can be denoted as
at the transmitter (T ); b) transmission latency (T ); and
c) semantic recovery latency or task process latency at the 1
receiver (T 3 ). G = ρc , (18)
We previously mentioned in Section I-B3) that the
where c is a constant related to specific tasks. Fig. 6 shows
influence of the semantic compression ratio is the computing- the relation of compress ratio and G, where ρ min is the
transmission tradeoff in latency. For better understanding, we
minimum compression ratio to maintain the integrity of
simply model the latency of a single user in the semantic-
source information or task, which can vary from different
aware task process scenario, compression latency is
tasks/users/information modalities.
So the total latency is
T 1 =
F(ρ,D)
, (14) T =T 1 +T 2 +T 3 =
F(ρ,D)
+
αD
+
ρwDG
. (19)
fe fe ρβ fr
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2977
2
E )needstobeconsideredasaconstraintsincetheusersides
(like mobile devices) often have energy budgets. However,
there are some other scenarios like energy minimization
of energy efficient communication system, which needs to
consider the total energy consumption of both transmitter and
1 2 3
receiver (E +E +E ).
To better show the relations of compression ratio and
latency/energy consumption, we illustrate them in Fig. 7a
and Fig. 7b, where
ρ∗
t is the optimal compression ratio for
the minimum latency of single user and
ρ∗
e is the optimal
compression ratio for the minimum energy consumption of
single user. In the figures, the range of compression ratio ρ is
in[ρ min ,1]duetotheρbelowthresholdρ mincannotmaintain
theintegrityofsourceinformationortask.ρ min canvaryfrom
different tasks/users/information modalities, here we set it to
0.5 for illustration,
In Fig. 7a and Fig. 7b, we can notice that the transmission
Fig.6. TherelationofG andcompressionratioρ. delay/energy decreases with decreasing compression ratio,
the computation delay/energy increases with the decrease
of compression ratio, and the optimal compression ratio to
It is obvious that compression ratio affect all parts of the total
reach the minimum value of latency and energy consumption
1
latency,andcontrolsthetradeoffbetweencomputing(T and
is different, thus leading to a tradeoff in computing and
3 2
T ) and transmission (T ).
transmission.(Note:Thisrelationmayvaryindifferentsystem
Whenthefocusofthearticleison“EnergyEfficiency”,the
models and with different users. The relation in Fig. 7 is an
totalenergyconsumptionoftheentiresystemisoftenusedasa
example illustration of a certain user.)
performance metric [50], [59], [62], [67], [87], [114]. In most
2) Utility Function: The concept of utility in resource
cases, delay and energy consumption are contradictory. The
allocation refers mainly to the satisfaction of users under
otheroftenbecomesaconstraintwhenoneistheoptimization
a certain resource allocation scheme. Utility is generally
goal. In [50], the authors proposed a semantic-aware energy-
expressed by the utility function. According to various objec-
saving task offloading network model. The goal is to extend
tives, the utility function is represented and mathematically
thebatterylifeoflocalusers,sothesumoflocalusers’energy
transformedbydifferentqualityofserviceparameters,suchas
consumption is used as the objective function. Considering
data transmission rate, delay, energy consumption, and cost,
the power shortage of mobile devices, the study in [59] is
which can achieve a better overall effect. The mathematical
committed to the allocation of resources for semantic-aware
transformation mainly includes reciprocal, logarithmic, and
MEC systems to minimize energy consumption. As discussed
weighted summation. Finally, an effective optimization algo-
in [67], the authors modeled the delay and total energy
rithm is designed to maximize the utility [46], [48], [49],
consumptionofasingleuserthatconsistsofthesethreeparts.
[117], [118]. For example, the utility function established in
The goal is to minimize the total energy consumption of the
the literature [117] is shown in Eq. (23):
entire system, considering constraints such as delay.
We also previously mentioned in Section I-B3) that the U =β 1A−β 2T −β 3E, (23)
influence of the semantic compression ratio is the computing-
whereAisthetotaltaskaccuracy,T isthetotaltimedelay,E
transmission tradeoff in energy consumption. Similarly, we
is the total energy consumption, and β 1 ,β 2 ,β 3 are the weight
also simply model the energy consumption in the semantic-
factors.
aware task process scenario. The energy consumption of
3) Traditional QoS and QoE:
semantic compression can be denoted as
• Quality of Service (QoS): Defined by the International
E 1 =κF(ρ,D)fe 2, (20) Telecommunication Union (ITU) as “the totality of char-
acteristicsofatelecommunicationsservicethatbearonits
where κ is a constant coefficient. F(ρ,D) also denotes the
abilitytosatisfythestatedandimpliedneedsoftheuser.”
CPU cycles required to compress the data D to ρD. The
It primarily focuses on system performance measured
transmission energy is
through physical parameters [133].
E 2 =pT 2 =p
ρD
, (21) • Quality of Experience (QoE): Refers to users’ subjective
R perception of the system or service performance, influ-
where p is the transmission power. And the task computing enced by context, culture, expectations, psychological
energy can be denoted as factors, and more [133].
E 3 =κ(ρwDG)fr 2. (22) In resource allocation for wireless communications, QoS
modeling is often similar to the utility function, but the
In many one-to-many uplink wireless communication scenar- mathematical complexity is higher than the general utility
t 1
ios, only the transmitter’s energy consumption (E = E + function.In[47],theQoSmodelingbasedonthetransmission
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2978 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
Fig.7. Latencyandenergyconsumptionversussemanticcompressionratio.
delay and the number of received semantic information is manualsupervisionanddatareconstruction,suchastextsenti-
shown in Eq. (24). ment classification, image classification, and target detection,
semantic fidelity can be expressed as average classification
QoSm,n(t)=
accuracyordetectionaccuracy.Theestablishmentofmostnew
1
(cid:10) (cid:11)(cid:10) (cid:11), performance metrics for SemCom resource allocation must
1+e β T (Tm,n(t)−T th ) 1+e β H Am(t)(H th −H˜ m,n(t)) rely on the concept of semantic similarity, so this section will
detail the current definition of various types of semantic
(24)
similarity. The comparison of different types of semantic
The two terms on the right side of the equation repre- similarities is presented in Table VI.
sent the transmission delay score and the received semantic 1) Semantic Similarity of Text Signal: For text transmis-
information score of the user n, respectively, where Tth and sion, BER does not reflect the performance well. In machine
Hth are the transmission delay thresholds and the received translation,bilingualevaluationunderstudy(BLEU)scoresare
semantic information and β T ,β H are the weight factors of generally used to measure results [136]. However, the BLEU
the delay in time and the received semantic information. score can only compare the differences between words in
The primary goal of wireless communication network twosentences,butcannotcomparetheirsemanticinformation.
services is to provide a user-satisfied quality of experience BLEU outputs a number between 0 and 1, representing the
(QoE) that is more user-centric. QoS does not contain any similaritybetweentwosentences,with1representingthehigh-
human-related quality factors, which means that for two estsimilarity.However,worderrorsmaynotalterthemeaning
different users, the same level of QoS may not guarantee of sentences. For example, the two sentences “That car had
the same level of QoE [134]. Designing QoE and managing been deserted” and “That vehicle had been abandoned” have
it while providing a service is necessary for high-quality the same meaning, but their BLEU scores are different due to
experiences. This requires assessment methodologies that can the use of different words to represent “car” and “deserted”,
quantifyQoE[135].Reference[71]studiedthetransmissionof which is a flaw in BLEU’s recognition of synonyms. A
imagesemanticinformationintheMetaverse3Dconstruction. word can have different meanings in different contexts. For
Data rate, bit error rate (BER) and interest score (the degree example,“bus”canhavedifferentmeaningsintermsofpublic
of interest in the image after semantic segmentation, which transportationandamicrocomputer.Traditionalmethods,such
is related to people) are considered when modeling QoE. In asword2vec[137],cannotrecognizeapolysemy.Theproblem
SectionIII-Cofthispaper,theQoEinthecontextofMetaverse is how to represent the word with a numerical vector, which
and SemCom is introduced. is different in different contexts [25].
Therefore, based on the bidirectional encoder representa-
B. Semantic Similarity tion from transformers (BERT) model [138], Reference [25]
Semantic similarity is defined as the degree of similarity proposed a new metric, Sentence Similarity, which describes
between the sender and the receiver’s semantic information the similarity of two sentences according to their semantic
under a specific semantic task. For task-oriented SemCom, information, as shown in Eq. (25).
semantic similarity can be extended to semantic fidelity.
The specific representation of semantic fidelity varies with
ξ =
BΦ(s)·BΦ(ˆs) T
, (25)
different target tasks. For automated tasks that do not require (cid:3)BΦ(s)(cid:3)(cid:3)BΦ(ˆs)(cid:3)
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2979
TABLEVI
COMPARISONBETWEENDIFFERENTSEMANTICSIMILARITIES
whereBΦ representstheBERTmodel.Thesentencesimilarity semantic information in the image is extracted into a scene
defined in Eq. (25) is a number between 0 and 1, which graph (SG) in the form of text, which captures the objects
representsthesimilaritybetweenthedecodedsentenceandthe andtheirrelationshipsintheoriginalimage.Thisinterpretable
transmitted sentence; 1 represents the highest similarity, and semantic information can not only be directly read and
0 represents no similarity. understood by humans but also be used to generate original
Currently, to measure text semantic similarity, most of images and retrieve similar images.
the literature [36], [52], [53], [55], [89], [90], [101] uses [42] introduced a comprehensive image-to-graph semantic
sentence similarity based on the BERT model as semantic similarity (ISS) metric, which uses a pre-trained deep neural
similarity. However, the authors of [40] proposed a metric of network (DNN) to directly capture the correlation between
semantic similarity (MSS), which is a function of semantic the original image and its semantic information without
accuracy and completeness. Based on token matching [139], any reconstruction of the image. The DNN is trained by
semantic accuracy is defined as the ratio of the sum of the Webimagetext [140],adatasetof400millionimage-textpairs
correct occurrences of each token in the recovered text to the collected from the Internet. Compared with the structural
sum of the occurrences of each token in the recovered text. similarity index measure (SSIM) [141], which measures the
Semantic completeness is defined as the ratio of the sum of difference between the original image and the reconstructed
the correct occurrences of each token in the recovered text to image on a set of pixels, the DNN can be used directly to
the sum of the occurrences of each token in the original text. obtaintheimagevectorandthesemanticinformationvectorof
Due to the high complexity of the expressions, we omit the thereceivedSG.TheISSmetricisdefinedasthecosineofthe
explicitexpressionofMSS.Reference[40]includesadetailed angle between the image vector and its normalized semantic
description of these metrics. triplet vector, which is calculated by the projection of the
2) Image-to-Graph Semantic Similarity: Although most of image vector on the set of semantic information vectors. The
the current work in the resource allocation of SemCom specific calculation steps and formulas are detailed in [42].
is text and image modalities, the work of [58] and [42] 3) SemanticSimilarityofImageSignal: Thesemanticsim-
combines the two in semantic extraction and establishes an ilarity of the image signal is used to measure the similarity
image-to-text semantic information extraction method. The between the original image and the restored image. The
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2980 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
TABLEVII
MAPPINGBETWEENRESOURCETYPESANDSEMANTICPERFORMANCEMETRICS
more classical method is measured by the peak signal-to- C. New Performance Metrics for SemCom
noise ratio (PSNR), which is based on the errors between
As mentioned above, the traditional resource allocation
corresponding pixel points. In the previous section, SSIM
model is usually modeled based on Shannon capacity,
is mentioned. It is widely used in the application of image
which fails to give full play to the performance advan-
similarity measurement, including the resource allocation of
tages of SemCom to ensure the best performance of the
SemCom [47]. These two metrics are used mainly in the
SemCom network. SemCom does not require error-free trans-
imagesignalsimilarityevaluation.However,in[113],ametric
mission of bits or symbols, so the optimization problem
for image semantic transmission (MIST) is proposed, which
based on Shannon capacity construction may reduce system
combinestheimportanceweightofeachsemanticinformation
performance. Therefore, it is essential to reconsider resource
with its respective transmission quality to obtain the final
utilization from a semantic perspective to develop new
evaluationresults.Aftercapturingtheimage,theUAVsendsit
performance metrics [142].
to the user and first extracts the semantic information through
Similarly, we will give a systematic summary and com-
O
the target detector. Specifically, a total of U objects are
parison of these new metrics in Table VIII, including a
detected, where i represents the i-th object and ci represents
critical evaluation of their strengths,limitations and suitabil-
its corresponding confidence. The relationship between the
ity for different modalities and applications. Considering
importance score Δi and the confidence ci of the object i can
that most of these new metrics are based on the con-
be expressed as Δi =c
i
σ , where σ is a variable that regulates
cept of semantic similarity, we illustrate the connections
the importance between different semantic information. The
and evolution of the semantic similarity-based metrics in
final MIST can be expressed as follows:
Fig. 8.
(cid:9)U
1) Semantic Transmission Rate and Semantic Spectral
E(A,Δi ,Q(pi))=A (Δi ×Q(pi)), (26) Efficiency: Firstly, reference [36] assumes that the semantic
i=1 unit (sut), representing the basic unit of semantic information,
where A represents the accuracy of extracting semantic canmeasuresemanticinformationinthetexttransmissionsce-
information, and Q(pi) represents the SSIM value of target nario. Then, two critical semantic-based performance metrics
i before and after transmission, which is a function that is are defined: semantic transmission rate (S-R) and semantic
positively correlated with the transmission power pi [43]. spectral efficiency (S-SE).
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2981
Only SemCom that reaches the semantic similarity threshold
ξ th required by the downstream task is considered effective.
Let η n,m denote whether the user n performs an effective
semantic transmission on the subchannel m. If ξ n,m > ξ th,
then η n,m = 1; otherwise, η n,m = 0. Ψ is called ES-SE,
which can be expressed as:
(cid:9)N (cid:9)M
Ψ= xn,m η n,mΦn,m , (29)
n=1m=1
where Φn,m is the S-SE of the user n in the subchannel m.
3) Task-Oriented S-R and S-SE: The authors of [109]
integrateS-RandS-SEintothescenariooffeatureimportance-
aware image classification, and two performance metrics of
the task-oriented SemCom system are defined: task-oriented
semantictransmissionrate(TOSR)andtask-orientedsemantic
spectral efficiency (TOSSE). Unlike the definition of [36],
whichconsiderslong-termtexttransmissionratherthansingle-
sentence transmission, the work of [109] focuses on the
performance ofeach user.Whenasemantictransmissiontime
slot begins, there are S semantic features after joint source-
channel coding (JSCC). BS obtains the feature transmission
Fig.8. Metricsthatbasedonsemanticsimilarity,theirconnections. ratedecisionvectorrf
basedonthechannelconditionsandthe
f
historical data distribution of each user. Then rn is fed back
tothefeatureselectionmoduleineachuserntodeterminethe
• S-R: S-R refers to the effective transmission of semantic number of features that need to be transmitted: Sn =rn f S/2.
information per second, measured by suts/s, as follows:
Therefore, the average semantic information for each symbol
Γn,m = WI ξ n,m , (27) of user n is In,m /Sn.
knL • TOSR: TOSR refers to the amount of semantic
allsubchannelbandwidthwasallocatedequally,usingW information effectively transmitted per second for a spe-
to represent the subchannel bandwidth. Since the article cific task. The expression is as follows:
focuses on long-term text transmission rather than the
WIn,m
transmission of a single sentence, I, L should take the ψ n,m = ξ n,m . (30)
Sn
expectedvalueandnottherandomvalue,thatis,foreach
user n, I/L is a fixed value, so omit the subscript n. The Compared to [36], it is equivalent to replacing I/knL
unitofI/knLissuts/symbol,andthechannelbandwidth (unit: sut/symbol) with In,m /Sn (unit: sut/symbol) of
of the band pass transmission in the ideal state is equal Eq. (27) in this paper, while the other parts remain
to the symbol rate (unit: symbol/s), so the unit becomes unchanged.
suts/s after multiplying by W. The semantic similarity • TOSSE: TOSSE refers to the rate at which task-related
based on BERT ξ n,m depends on the structure of the semantic information is successfully transmitted through
DeepSC neural network kn and the channel conditions a single bandwidth unit. The expression is as follows:
γ n,m. It can be expressed as ξ n,m =f(kn ,γ n,m).
ψ n,m In,m
• S-SE: S-SE refers to the rate at which semantic φ n,m = = ξ n,m . (31)
information is successfully transmitted within a unit W Sn
bandwidth, measured by suts/s·Hz, as follows: 4) Semantic Energy Efficiency: Based on the concept of
Γn,m I S-R [36], semantic energy efficiency (S-EE) is introduced
Φn,m = = ξ n,m , (28) in [88] as a measure of energy efficiency in the SemCom
W knL
system, which is quantified by suts/Joule. Traditional com-
The proposal of S-R and S-SE provides an impor-
munication systems define energy efficiency as the number
tant theoretical basis for many subsequent studies such
of bits that the system can transmit per unit of consumed
as [88], [90], [109]. Based on these two metrics, they made
energy. From a semantic point of view, the feature of S-
expansions and cross-domain transformations, and we now
EE is the number of semantic symbols transmitted by unit
continue with our discussion of them.
energy consumption. It is expressed as the S-R ratio that can
2) EffectiveS-SE: Thestudyin[90]consideredtherequire-
be achieved by the total power consumed in the SemCom
ment of semantic information similarity for downstream
network. The S-EE of user n is denoted by:
semantic tasks, and the concept of effective semantic spectral
efficiency (ES-SE) is introduced. The serious deviation of Γn wnI
semantic similarity will directly lead to inaccurate results. En = pn +pc = (pn +pc)knL ξ n , (32)
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2982 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
c
pn is the transmit power of user n, p is the electrical power it can be DeepSC-VQA [145]. In order to more reasonably
that the circuit consumed, and bn is the bandwidth. The Γn reflecttheuser’sQoErequirements,reference[22]established
here represents the S-R of user n. a semantic QoE model, which is expressed as:
(cid:9)
5) SemanticEntropy: Semanticinformationreliesnotonly
on the source data, but also on the specific task, which
QoEq b = wnGn R +(1−wn)Gn A
is significantly different from the information defined by n∈G n b
Shannon. Consequently, the same data may contain different (cid:9) wn (1−wn)
= + .
amounts of semantic information for different tasks. In this 1+e βn(ϕr n eq−ϕn) 1+e λn(ξ n req−ξn)
n∈Gb
regard, the authors of [121] defined the semantic entropy as q
follows. (36)
Definition 1: GivensemanticsourceX,semanticentropyis
It should be noted that the authors of [22] modeled the
defined as the minimum average number of semantic symbols
complex situation of multi cell task and user. b denotes the
about data X ∈X that is sufficient to predict task Y, i.e.,
H(X;Y) (cid:2)minE (cid:2) dim (cid:12) Code E S(X) (cid:13)(cid:3) ,ES ∈E S c In ell E i q n . d ( e 3 x 6) a , n G d q b q d d e e n n o o t t e e s s t t h h e e q in -t d h ex us o e f r t g h r e o u u s p er in gr t o h u e p b i - n th c c e e ll l s l . .
(cid:12) E S (cid:13) wn and (1 − wn) are the weights of the semantic rate ϕ n
s.t. P Y|Code E S(X) =P(Y|X), (33) and the semantic accuracy ξ n on the user n, respectively. Gn R
A
and Gn are the semantic rate and semantic accuracy for user
where Code E S(X)) denotes the semantic symbol vector n, respectively. β n and λ n represent the growth rates of Gn R
extracted from X with the semantic encoder ES, E S is the and Gn A . In addition, ϕr n eq and ξ n req represent the minimum
set of semantic encoders, and P(Y|X) is the conditional semantic rate and semantic accuracy of 50% scores [22].
7) QoE of Metaverse Service Providers: With the support
probability of achieving the goal of Y given X.
ofvirtualreality(VR),augmentedreality(AR),andthetactile
The constraint in Definition 1 implies that the defined
Internet, Metaverse hardware devices cannot only mobilize
semanticentropyislosslessandthatitisactuallydefinedasan
expectedvaluethroughoutthedatasetX,thatis,thesemantic all senses of the user and provide an immersive experi-
ence [146], but also revolutionize the way people interact
entropy is constant for the same task and dataset. However,
∗ with each other and even with objects. Therefore, it is crucial
it is intractable to find an optimal semantic encoder, E , to
S
to design the QoE of Metaverse Service Providers (MSPs)
derive the semantic entropy [143]. To obtain a measure that is
as a performance indicator to measure the performance of
bothmeaningfulandmanipulableforsemanticcommunication
Metaverse Service [147]. In the proposed framework in refer-
systems, [121] utilize a well-designed DL model as the
ence[71],theauthorsaimtotransmitthesemanticinformation
encoder to obtain an approximate semantic entropy for a task,
of interest to each MSP. Therefore, the performance metrics
which is:
(cid:2) (cid:12) (cid:13)(cid:3) of the data rate, the BER, and the interest rating should be
H(X;Y) (cid:2)minE dim Code E DL(X) considered together. Thus, the QoE of the k-th MSP Uk can
(cid:12) (cid:13)
be defined as [148]:
s.t. P(Y|X)−P Y|Code E DL(X) <ε, (34)
(cid:9)N
k
(cid:12) (cid:13)
where the constraint indicates that the task performance Qk = J k iT 1−B k i , (37)
degradation can not exceed ε. From Eq. (34), the defined i=1
a
af
p
o
p
r
r
e
o
m
xi
e
m
nt
a
i
t
o
e
ne
s
d
em
m
a
e
n
th
ti
o
c
d,
e
t
n
h
t
e
ro
a
p
p
y
pr
i
o
s
xim
lo
a
ss
te
y.
se
A
m
c
a
c
n
o
t
r
i
d
c
in
e
g
ntro
to
py
th
o
e
f
where Nk is the number of objects that Uk is interested, J
k
i
is the normalized interest rating of Uk for the i-th object
the considered tasks can be derived based the corresponding
recommended to Uk, T is the normalized time that all MSPs
DLmodels.Therefore,[121]usesemanticentropytoconstruct
finish the transmission, and
Bi
is the BER of transmitting the
the semantic rate and semantic QoE model. We now move on k
i-th object’s semantic information to Uk.
to this semantic entropy-based metric - semantic QoE.
8) System Throughput in Message: System throughput in
6) Semantic QoE: The accuracy and efficiency of message
message (STM) represents the network performance from
transmission are different from the user’s point of view, and
a semantic point of view, proposed by [97]. In text com-
depending on the application, users may have their own
munication, an entire text sentence ending in a cycle, or
preferences for them. For example, some users prefer higher
in voice communication, a completely emitted voice signal,
accuracy but have a certain tolerance for delay, while some
can be regarded as a message. Taking this into account, the
users may want to get a higher rate but do not need high
message rate (unit: msg/s) is interpreted as the number of
accuracy [144]. The semantic rate of user based on semantic
messages transmitted or processed per unit time under the
entropy is given as
reference of the bit rate (unit: bit/s) definition. Because the
ϕ n = k H n ˜ / D W L , (35) s fr y a s m te e m wo th r r k ou b g as h e p d ut o h n a S s h a a v n e n r o y n p ’s er t f h e e c o t r e y x : pressionofthesystem
(cid:9)(cid:9) (cid:9)(cid:9)
where the meaning of W and kn is the same as in Table V. S T = xnbrnb = xnbwnblog2(1+γ nb).
H˜ DL isthesemanticentropybasedonspecificDLmodel.For n b n b
text modal task, it can be DeepSC [25]. For bi-modal task, (38)
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2983
Here, n and b represent the n-th user and the b-th BS, respec- The first attribute is usually represented by a non-decreasing
tively. Among them, wnb and γ nb represent the bandwidth time penalty function f(t) ∈ T, which includes metrics like
and SNR, respectively. The system throughput represents AoI. The second attribute is captured by an error detection
the number of bits successfully transmitted per unit time function g(Xt ,Xˆ t)∈X, typically encompassing metrics such
in the system, reflecting the network performance. Therefore, asmeansquareerror(MSE)ormeanpercentageerror(MPE).
the authors of [97] defined a general bit-to-message (B2M) The third comes from practical constraints like spectrum limit
conversion function S(·), which is related to different seman- and energy consumption, denoted by a predefined function
tic encoders, knowledge matching, and message properties. C(Xt ,dt)∈C basedonsourcestatesXt andactiondt,where
Therefore, according to the bit rate rnb given by the Shannon the latter refers to the transmission policies like generation
M
theorem, the message rate r
nb
= Sn(rnb) can be naturally decisions, code rate, and resource allocation.
defined by S(·), and the expression of STM is derived as 11) Efficiency of Semantic Information: In the context
follows: of the SemCom-Industrial Internet of Things (SemCom-
(cid:9)(cid:9) (cid:9)(cid:9)
IIoT), traditional performance metrics are no longer the best
S TM = xnbrn M b = xnbSn(rnb). (39) choice. As reported in [108], a new performance metric was
n b n b
designed at the semantic level, named Efficiency of Semantic
STMcharacterizesthenumberofmessagessuccessfullytrans- Information(EoSI).Thescenarioisrelativelydifferent,andwe
mittedinthesystemperunittime,whichcanwellcharacterize needtostatethatn,mdoesnotrefertotheuserandsubchannel
network performance from a semantic perspective. indexes only here. The intelligent sensing device (ISD) in
9) Age of Semantic Information: In traditional communi- the scene is divided into m categories, so the subscript of
cation systems, Age of Information (AoI) [149] is a popular ISDm,n meansthen-thISDinthem-thclass.Thepreliminary
measure of information importance, which is defined as expression of EoSI is as follows:
Δ AoI (t) = t − u(t) by measuring the information delay
UoSIm,n(t)
of the destination. u(t) is the generation time of the latest EoSIm,n(t)= . (44)
costm,n(t)
received data packet. In order to capture the freshness of
information and semantic loss in the SemCom system, the lit- UoSI is semantic information utility: considering both seman-
erature [89] proposed a new measurement method called Age tic timeliness and task accuracy, the expression is as follows:
ofSemanticImportance(AoSI).Beforegivingthedefinitionof
AoSI, the reference [89] first defined the semantic importance
UoSIm,n(t)=Fm d ,n(t)Fm a ,n(t). (45)
(SI): semantic loss caused by missing or incorrect semantic Among them, task accuracy Fm a ,n(t) quantifies the impact
content [150]. It can be expressed as ψ = 1−ξ. Here, ξ is d
of semantic information on task accuracy, Fm,n(t) quantifies
the semantic similarity, which we discussed in the previous
the impact of the timeliness of semantic information on the
subsection. For example, in a text transmission task, semantic
timeliness of task results, and the timeliness of task results is
importance can be denoted as
also the standard for judging whether the task is successfully
B(x)·B(xˆ) T completed. costm,n(t) represents the resource overhead of
ψ =1−ξ =1−
(cid:3)B(x)(cid:3)·(cid:3)B(xˆ)(cid:3),
(40) ISDm,n to complete intelligent tasks, which is a weighting
function of bandwidth resources, local computing resources,
where B(·) represents the BERT model. The definition of and MEC computing resources. The complete expression of
AoSI can be obtained by the definition of SI and AoI: EoSI is complex. If you are interested in the details and the
Δ AoSI (t)=Δ AoI (t)·ψ(u(t))=(t −u(t))·ψ(u(t)),(41) derivation process, see [108].
12) Success Probability of Tasks: In order to simulta-
where ψ(u(t)) is the semantic importance of the last received neously evaluate the impact of transmission and adaptive
semanticcompression(ASC)ontheperformanceofSemCom,
packet.
a new performance metric is defined in [106]: success prob-
10) Utility of Information: Reference [72] introduced a
ability of tasks. Reference [111] further improved the work
utility of information (UoI) metric. It encompasses multiple
in reference [106] and also adapted this performance metric.
contextualattributestocapturetheutilitygradeoftheupdates
According to [106], the definition of success transmission
transmitted to communication systems or services. From a
probability of users is first introduced, as follows:
mathematicalperspective,itcanbemodeledusingacomposite
(cid:6) (cid:7)
function:
2
an(1−on)−1
P(tn ≤t0)=2Q , (46)
U(t)=(Θ◦U)(D t). (42) bn δ
Here, Θ(·) : Rm → [0,M] is a non-increasing function where tn is the transmission delay of user n, P(·) is the
that converts the penalty into the corresponding utility grade. probability, and on is the semantic compression ratio. In
U :Rn →Rm,n ≥m,isanon-decreasingnon-linearpenalty practical scenarios, such as the Internet of Vehicles (IoV),
function with respect to the three attributes as follows: a large number of tasks are delay sensitive, so there are
(cid:2) (cid:3) always strict transmission delay constraints, represented by
f(t),g(Xt ,X ˆ
t
),C(Xt ,dt ) ∈T ×X ×C → F U(D
t
)∈Rm.
t0. Therefore, the user’s transmission success probability is
(43) P(tn ≤ t0). an = w d n 0 t 0 , wn is the bandwidth of user n,
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2984 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
• Semantic Fidelity: Defined as the fidelity between the
original vectorized data X and the received information
Xˆ. It is expressed as:
(cid:12) (cid:13) (cid:12) (cid:13)
SF ε,n X,Xˆ =fsa X,Xˆ , (49)
where the subscript n represents the vehicle n, ε repre-
sentstheindexoftheedgeserver,andfsa(·)isthefidelity
mapping function, which varies with the task.
• Semantic Timeliness: Semantics will evolve over time.
By modeling and tracking temporal changes, includ-
ing aggregating new semantic information as much as
possible, communication efficiency can be significantly
improved,andtheprobabilityoferrorsinsemantictrans-
mission can be reduced. The timeliness of the semantic
information extracted by the system is defined as:
(cid:14) (cid:15)
ST ε,n(·)=fst,ς
Tth −T
, (50)
Fig.9. Therelationoftaskaccuracyandsemanticcompressionratio. Tth
where fst,ς(·) is a non-linear decreasing function with
parameter ς on semantic timeliness. T is the total delay
and d0 is the initial extraction of semantic information for
users without semantic compression. bn =
N
p
0
n
Bn
, pn is the
t
o
h
f
e
th
to
e
ta
sy
l
s
d
t
e
e
l
m
ay
,
,
a
t
n
h
d
e
T
g
t
r
h
ea
i
t
s
er
th
t
e
he
de
s
l
e
a
m
y
a
c
n
o
ti
n
c
st
t
r
i
a
m
in
e
t
l
.
in
T
e
h
s
e
s.
lower
transmissionpoweroftheusern,andN0isthespectraldensity
of the noise power.
δ2
is the variance of channel gain. The Q
The following formula defines semantic utility:
function represents the tail distribution function of a standard Qa n ll =ζ n SF ε,n +χ n ST ε,n . (51)
normal distribution. So we can obtain the n-th user’s success
probability of tasks: Among them, ζ n and χ n are the preferences of semantic
fidelity and semantic timeliness, respectively.
Ωn =η(on)×P(tn ≤t0), (47) 15) SemComQoS: Semanticsimilarityisfurtherpromoted
by [53], and SemCom QoS (SC-QoS) based on Semantic
where η(on) is the probability of which task is success-
Quantization Efficiency (SQE) is created as follows:
fully executed under successful transmission. It can be seen
• SQE: In order to solve the tradeoff between semantic
from (47) that the task success probability proposed by [106]
accuracyandthenumberofbitsconsumed,anewmetric,
for evaluating SemCom performance can control the tradeoff
SQE, is proposed. This metric quantifies the ratio of the
between semantic transmission and semantic understanding.
semantic similarity gain of each semantic feature to the
13) Transmission Efficiency of Tasks: In [112], the authors
bit-relatedsemanticsimilaritygain.Duetoitsstrongcor-
modeled the physical channel as a non-trainable fully con-
relation with the novel semantic bit quantization (SBQ)
nected layer to simulate different channel states. With the
proposedintheirwork,thesecontentsarenotintroduced.
help of the curve fitting method, the mathematical relation-
See [53] for more details.
ship between compression ratio and task performance under
• SC-QoS: Defined based on SQE and transmission delay,
different channel states is explored. Then a new measurement
and the effective SC-QoS is expressed as:
standard is established in [112]: transmission efficiency of
(cid:9)N (cid:12) (cid:13)
tas T k h s. e transmission efficiency of the task is defined as the Ψ= (cid:19)(cid:16) n (cid:11) −φ G G(cid:16) n , (52)
n=1
weighted sum of the number of packets from each user and
the corresponding achievable task accuracy at the receiver. where the user’s index is n, (cid:19)(cid:16) n (cid:11) is the effective SQE
Specifically, the semantic task transmission efficiency vt in (the sum of SQE whose semantic similarity satisfies the
time slot t is defined as follows: minimum threshold), G(cid:16) n is the delay, and φ G is the
(cid:9)J (cid:9)N j balance coefficient.
vt = vt n,j ×A n t ,j. (48) 16) Semantic Score: To measure the overall semantic loss
betweentheoriginalsentences andthereconstructedsentence
j=1n=1
ˆs atthereceiver,theworkin[151]definesanewmetricnamed
Thesubscriptndenotesusernandjdenotestheintelligenttask SemanticScore(SS).whichcombinesthebestoftwodifferent
n,j
j corresponding to user n. At is the classification accuracy quantities, BLEU score and sentence similarity which uses
n,j
and vt is the number of data packets that each user can BERT. The BLEU score cannot handle word synonyms, but it
transmit in slot t. is a fast and low-cost algorithm that is language independent
14) Semantic Utility: The reference [107] proposed a and corresponds to human judgment. The sentence similarity
semantic utility measurement method that considers semantic score using BERT vectors is slow and has ratings comparable
timeliness and semantic fidelity. to the BLEU, but it also handles synonyms. Let Δλ(s,ˆs)
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2985
denote the SS between sentence s and ˆs, which is a convex
combination of BLEU(s,ˆs) and ξ(s,ˆs).
Δλ(s;ˆs)=(1−λ)BLEU(s,ˆs)+λξ(s,ˆs), (53)
where λ∈[0,1] is a parameter.
In this section, we explore the construction of the objective
function in resource allocation of SemCom, which is a key
to the modeling of optimization problems. We provide a
detailed review of performance metrics, categorizing them
into two types. The first type includes traditional metrics
such as delay and energy consumption, while the second type
focuses on new metrics based on semantic similarity. We give
two comprehensive comparative matrices to better synthesize Fig. 10. The taxonomy of centralized resource allocation algorithms in
findings across references. To further clarify the influence of SemCom.
different resource types on these performance metrics, we
provide a resource–metric mapping summary in Table VII,
(SCA) methods, and the interior point method, as well as
in which we use some clear examples in different studies to
some other mathematical algorithms based on other mathe-
illustrate this influence.
maticalalgorithms,suchastheHungarianalgorithm[153].An
optimizationalgorithmbasedonconvexoptimizationtypically
IV. CENTRALIZEDRESOURCEALLOCATIONALGORITHMS
combines several of these techniques.
In order to realize resource allocation in SemCom and
1) Lyapunov Optimization: Lyapunov optimization is a
meet the requirements of these performance metrics proposed
powerful long-term resource optimization scheme to find
above, advanced resource allocation strategies and algo-
stability or equilibrium points of dynamical systems with
rithms are essential. The optimization problem constructed
stochastic properties of nonlinear systems. It requires less
is extremely complex and differs significantly from the tra-
priorknowledgeandhaslowcomputationalcomplexity[154].
ditional communication architecture in terms of objectives,
Lyapunov optimization focuses on analyzing and optimizing
constraints, and optimization variables. It is a challenge to
stochasticnetworks(networkscharacterizedbyrandomevents,
construct a well-performing optimization algorithm that can
time-varying dynamics, and uncertainties). It is particularly
adapt well to SemCom. Currently, there are a variety of
well-suited for applications in communication systems and
centralized algorithms for resource allocation in SemCom,
queueingsystems.Theauthorsof[102]adoptedtheLyapunov
mainlyconsistingofconvexoptimization,heuristicalgorithms,
optimization method to solve the problem, which first trans-
andDRL.Fig.10showsthetaxonomyofcentralizedresource
formsthelong-termconstraintsintoqueuestabilityconditions
allocation algorithms in SemCom.
using the concept of virtual queue and then transforms the
In recent years, many researchers have summarized the
long-termobjectivefunctionandthequeuestabilityconditions
state-of-the-art resource allocation algorithms of various sce-
intosolvableshort-termsubproblems.Similarly,[59]and[118]
narios in their surveys. In [16], the authors summarized
also used Lyapunov optimization techniques to transform
different optimization methods for resource allocation in edge
the original stochastic optimization problem of multiple time
computing. The comparison tables of different papers are
slots into a series of deterministic problems in a single
designed according to the objective, brief description of
time slot. Lyapunov optimization, as a stochastic optimization
the methods, advantages, and disadvantages. Reference [152]
method, enables online decision making while maintaining
summarized different resource allocation schemes for the
sub-optimal performance. Therefore, it applies well in a
two dominant vehicular network technologies, e.g., Dedicated
long-term stochastic scenario in SemCom system, like the
Short Range Communications (DSRC) and cellular-based
semantic-aware dynamic long-term MEC systems using time
vehicular networks. In this subsection, centralized resource
division duplexing (TDD) in [59]. Lyapunov optimization can
allocation optimization algorithms from different literature in
also combine with DRL-based method, in [72], expanding on
SemCom are reviewed.
the Lyapunov transformation, the UoI minimization problem
is converted into a sequence of deterministic single time-slot
A. Algorithms Based on Convex Optimization and optimization problems. Subsequently, the DRL-based method
Mathematical Techniques PPO(willbeintroducedlaterinSectionIV-B)isusedtotackle
Because resource allocation involves a lot of variables this problem.
and constraints, the corresponding optimization problems are 2) Alternating Optimization Algorithm: The alternating
usuallycomplex,evennon-convexorNP-hard.Aconsiderable optimization(AO)algorithmistodecomposetheoptimization
part of the research transforms the non-convex problems into problem into several sub-problems, and then these sub-
near-convex or convex optimization problems, which leads problems are solved iteratively. Commonly used in the case
to feasible convex optimization methods. The main tech- of multi-variable optimization, which iteratively optimizes
niques include Lyapunov optimization techniques, alternating each variable while treating other variables as a fixed value.
optimization (AO) algorithms, successive convex approximate Depending on the specific problem, the complexity of the
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2986 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
TABLEVIII
COMPARISONBETWEENDIFFERENTNEWSEMANTICMETRICS
decomposed problem varies; the simpler case is decom- extraction strategy subproblem and the wireless resource
posed into two to three subproblems, where each subproblem allocationsubproblem,whichwillbeoptimizedalternatelyand
optimizes a single variable in [44], [45], [57], [91], [98], iteratively, where each of the two subproblems also employs
[119], [130], [132]. As the problem and the optimization the AO algorithm to optimize the corresponding parameters.
variables increase, the optimization problem is decomposed Aniterationofthealgorithmforthetotaloptimizationproblem
into three subproblems in which the subproblem has two or contains the number of iterations L1 and L2 of the AO
more optimization variables in [59], [66], [67], [74], [93], algorithm for the two sub-problems.
[100], [105]. A more complicated situation occurs in [68], 3) Successive Convex Approximate: The idea behind suc-
where the paper employs a nested AO algorithm to divide cessive convex approximate (SCA) istofindalocallyoptimal
the optimization problem into two subproblems: the semantic solution to the original problem by iteratively solving a series
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2987
Algorithm 1 Basic SCA Algorithm for Problem P maximummatchingofabipartitegraphbasedonanymatching
Find a feasible solution x ∈ X in P, choose a step size ifwehaveawaytokeepsearchingforaugmentingpathsuntil
θ ∈(0,1] and set k = 0. eventuallywefindnonewaugmentingpaths.Thecoreideaof
Repeat theHungarianalgorithmistoiterativelysearchforaugmenting
1) Compute xˆ(xk ), the solution of P xk; paths to get a maximum match.
2) Set
xk+1 =xk +θ(xˆ(xk )−)xk
; The Hungarian algorithm can solve the allocation problem
3) Set k ←k +1 in polynomial time, which can significantly reduce the algo-
Until convergence criterion is met. rithmic complexity. When it comes to the scenario of the
resource allocation problem in SemCom, it is usually used
for the subproblem of subcarrier pairing/subchannel alloca-
of convex optimization problems similar to the original non- tion after the original optimization problem is decomposed
convex problem. Consider the following optimization: by the AO algorithm above. In the literature [36], [109],
and [57], the optimization subproblem of channel allocation
P:min U(x) (54) is regarded as a bipartite graph matching problem, and then
x
the Hungarian algorithm is used to solve this optimization
s.t. gl(x)≤0, ∀l =1,...,m (54a) subproblem. Among them, the knowledge-assisted proximal
x∈K (54b) policy optimization (K-PPO) algorithm is proposed in [109],
which uses the Hungarian method to determine channel
where the objective function and constraint (54a) is smooth allocation, greatly reducing the complexity of the original
(possibly nonconvex), the feasible set is denoted as X. The proximal policy optimization (PPO) algorithm by introducing
original non-convex or non-concave function is transformed the Hungarian algorithm. The details of PPO will be intro-
into a series of convex or concave functions. The convex duced later in Section IV-B.
approximationoftheoriginalproblemcanbestatedasfollows: 5) Lagrange Methods: The Lagrange multiplier method
given xk ∈X: is a common method for solving constrained optimization
(cid:12) (cid:13)
problems. For the optimization problem with only equation
P xk:m x in U˜ x;xk (55) constraints, you can directly use the Lagrange multiplier
method to list the Lagrange function, which will be trans-
(cid:12) (cid:13) formed into an unconstrained optimization problem to solve.
s.t. g˜l x;xk ≤0, ∀l =1,...,m (55a) Fortheoptimizationproblemwithinequalityconstraints,using
x∈K (55b) the Lagrange function to optimize it must satisfy the Karush-
Kuhn-Tucker(KKT)condition,whichisanecessarycondition
whereU˜(x;xk )andg˜l(x;xk
)representtheapproximationsof for taking the optimal parameter values and a sufficient
U(x)andgl(x)atcurrentiterationxk
,respectively,thefeasible condition for some special convex optimization problems.
set is denoted as
X(xk
). We can summarize the basic SCA Problems containing inequality constraints after listing the
algorithm in Algorithm 1. Lagrangian function still have constraints that are not easy to
This process is repeated until the stopping criterion is deal with, then it can be transformed into a Lagrangian dual
satisfied. It is assumed that at each iteration, some original problem; this dual problem must be a convex optimization
functions are approximated by their upper bounds, where the problem and therefore easy to solve. But in order to make
same first-order behavior is preserved [155]. the dual problem and the original problem have the same
Since an approximate solution to the original optimization solution, it must satisfy the strong duality. The sufficient
problem is solved in each iteration, there is no condition is Slater’s condition; the necessary condition is the
guarantee that the global optimum will be obtained. KKT condition. Lagrangian methods have been employed
The convergence of the method is guaranteed due to in many works, where the problem is decomposed into
convexity/concavity [19]. subproblems and then the sub-optimization problem is solved
AO algorithms and the SCA algorithm are two methods using Lagrangian methods [57], [68], [102], or the problem
that work well with each other, and almost all the literature is transformed directly using the Lagrangian methods to
on SCA uses a combination of the two. Decomposing a solve [78].
largenon-convexoptimizationproblemintoseveralsmallnon- Summary: Traditional optimization algorithms based on
convex optimization subproblems to solve iteratively reduces convex optimization techniques and other mathematical
the difficulty/complexity of the SCA algorithm, thus allowing algorithms are applicable to small-scale solutions and high-
the difficulty and complexity of the overall problem to be reliability demand scenarios. They have the following
reduced[44],[45],[49],[59],[67],[68],[70],[91],[92],[106], advantages: a) mature and widely used; b) easy to obtain sub-
[111], [119]. optimal optimization results; c) not relying on data. However,
4) Hungarian Algorithm: The solution to the maximum algorithms based on these techniques are often too complex.
matching problem in bipartite graphs is the origin of the As a result, its complexity makes it difficult to implement in
Hungarian algorithm. Since a maximum matching of a bipar- practical systems and not suitable for large-scale problems.
tite graph necessarily exists, e.g., the upper bound is a perfect Although algorithm complexity may vary due to different
matching that contains all vertices, it is possible to get a problems and scenarios, we can still give a brief summary
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2988 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
of these algorithms. In terms of computational complexity, With the development of DL and reinforcement learning
Lyapunovoptimizationitselftypicallyhasthelowestcomplex- (RL)techniques,puredata-drivenDRLhasbecomeapowerful
ityduetoitsonlineanddynamicnature,makingitsuitablefor tooltosolvecomplexresourcemanagementproblemsinrecent
real-time systems. The complexity of Lyapunov optimization years [81], [156], [157]. By efficiently learning the dynamics
is primarily determined by the per-slot deterministic sub- of the environment, DRL can provide resource allocation
2 3
problem, and often falls in the range of O(n ) to O(n ) strategiesthatmaximizelong-termreturnsbasedonpretrained
when convex formulations are involved, making it particu- policy networks.
larly suitable for low-latency real-time systems. Alternating RL and DRL approaches can be mainly distributed in two
optimization (AO) and Lagrangian methods exhibit moderate ways:basedonvaluefunctions(1-4)andbasedonpolicygra-
complexity, with AO being effective for decomposable non- dients(5-9).Thispaperalsoprovidesabriefdescriptionofthe
convex problems and Lagrangian methods for constrained algorithms based on these techniques in various publications.
optimization. The complexity of AO is mainly determined by 1) Q-Learning: Q-Learning (QL) [158] is an off-policy
the complexity of solving each subproblem. For instance, if control method for finding the optimal policy, mainly used in
each subproblem involves convex optimization with complex- discrete action space. The core idea is to utilize a Q function
3
ity O(n ), and k such subproblems are solved per iteration, that represents the expected reward of taking an action in
thetotalcomplexityperiterationbecomesapproximatelyO(k· a particular state. The Q function updating rule satisfies the
3
n ). For Lagrangian-based methods, the overall complexity Bellman equation:
depends on both the structure of the primal problem and the (cid:17) (cid:18)
(cid:2) (cid:3)
methodusedforupdatingdualvariables.Iftheprimalproblem Q(s,a)←Q(s,a)+α r +γmaxQ s (cid:4),a (cid:4) −Q(s,a) .
admitsaclosed-formsolution,eachiterationmayinvolveonly a(cid:2)
2
dual updates with complexity around O(n ), leading to a (56)
total complexity of O(K · n 2 ), where K is the number of
iterations. However, if the primal problem requires solving In [71], the selection of public messages uses QL techniques.
a numerical optimization (e.g., quadratic programming), the The work of [124] compares the QL-based approach with
per-iteration cost may increase to O(n 3 ), resulting in a total the convex optimization-based approach under the video
complexity of O(K ·n 3 ). Successive convex approximation semantics-drivenresourceallocationproblem.Theexperimen-
(SCA)tendstohavehighercomplexityduetoiterativeconvex tal results prove that the QL-based approach performs better
approximations. Each subproblem often requires O(n 3 ) time, than the convex optimization-based approach.
and the total complexity O(T ·n 3 ) grows linearly with the 2) DeepQNetwork: Mnihetal.[159]introducedthedeep
number of iterations T. Thus, SCA is suitable for non- Qnetwork(DQN),whichpioneered thefieldofDRL.Inreal-
convex problems with a manageable size and structure. The world scenarios, the number of states can be large, making
3
Hungarianalgorithm,withacomplexityofO(n ),isefficient the construction of Q-tables computationally intractable. To
for small-scale linear assignment problems but less scalable addressthislimitation,DQNusesaneuralnetworktoestimate
for larger systems. Table IX reviews the literature using the Q-values of each state-action pair. The most important
these traditional optimization techniques, which are based feature of DQN is that it uses experience replay [160]
on convex optimization techniques and other mathematical and target networks to stabilize the training of deep neural
algorithms. networks [161]. As mentioned in the previous paper, [89]
defined the AoSI metric. In the paper, the long-term average
AoSI optimization problem is modeled as an MDP, and a
DQN-based algorithm is proposed to find the suboptimal
B. Algorithms Based on Deep Reinforcement Learning solution for source scheduling and the number of semantic
In the context of SemCom, direct modeling of the relation- symbols.Comparedwiththesimplerstatespacecasemodeled
ship between semantic accuracy (or fidelity) and optimization intheliterature[71]usingQL,mostoftheresourceallocation
variables, such as the semantic compression ratio, is often problems in SemCom have a more complex state space, so
infeasibleduetotheabsenceofexplicitanalyticalexpressions. DQNisobviouslymoresuitable.In[22],theexhaustivesearch
Then results in the non-differentiable and implicit objectives. to solve the semantic compression subproblem will lead to
To address this challenge, some authors [112], [121] use high computational complexity. Because when using exhaus-
different curve fitting techniques to approximate this implicit tivesearchtosolvethiscombinationoptimizationproblem(for
relationship. For instance, in [121], neural networks are K-users,thereareK!permutations),thecomplexitywillgrow
adopted to fit the relationship curve of semantic fidelity and exponentially with the number of users and cells. Therefore,
optimization variables (power, channel assignment, semantic in the journal version of [22], that is, the reference [121], the
compression)foreachtask(singlemodalandbi-modal).Once authorsproposedasolutionthatcombinesDQNandmatching
this approximation is obtained, the originally implicit objec- theory. The exhaustive search is replaced by the DQN-based
tive becomes differentiable or at least numerically tractable. method to improve overall QoE effectively.
However, after curve fitting, the output fitting function is still 3) Double Deep Q Network: Hasselt et al. [162] proposed
complex and non-convex. Traditional mathematical methods the double deep Q network (DDQN) to solve the over-
are often difficult to model or calculate in the face of these estimation problem in QL. The DDQN algorithm borrows
complexities. from the double-Q learning algorithm [163] and makes
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2989
TABLEIX
CENTRALIZEDALGORITHMBASEDONCONVEXOPTIMIZATIONMETHODANDMATHEMATICALTECHNIQUES
improvements to the DQN algorithm: estimating the pol- went one step further than the QL-based work [124]
icy based on the online Q-network, selecting the action, that was already mentioned. They wanted to improve the
and estimating the Q-value with the target network. Some accuracy of video semantic understanding and build a
experimental results show that DDQN finds a better strat- multidimensional resource allocation model that combined
egy than DQN in Atari games. The authors of [82] communication, computation, and caching. They designed
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2990 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
the DDQN-based algorithm, which is shown to achieve bet- have also been attempts in the literature to use TD3 instead
ter results than those achieved by the QL-based approach of DDPG as the base algorithm of the scheme [107], [108],
in [124]. in which [107] proposed a TD3-driven dynamic semantic-
4) DuelingDoubleDeepQNetwork: Duelingdoubledeep aware algorithm: dynamic semantic-aware TD3 (DSATD3)
Q network (D3QN) is a combination of Dueling DQN [164] for a federated learning-driven semantic vehicular network
and DDQN. Dueling DQN separates the computation of to guide agents in adopting accurate semantic extraction and
Q-values into two components: the value function (V) and resource allocation strategies. The simulation results showed
the advantage function (A), which enables Dueling DQN to that DSATD3 has better performance compared to DDPG-
provide more accurate Q-value estimation while needing less based approaches.
discrete action data, thus improving sample efficiency. As in In contrast to the previous two papers, the work in [75]
Table X, the authors of [90] and [94] both use D3QN for improves the TD3 algorithm and proposes the TD3-RNS
discrete action in the whole DRL framework. algorithm (TD3 with reference neuron-enhanced Softmax) to
5) Actor-Critic: The actor-critic (AC) [165] algorithm solvealong-termsemanticthroughputmaximizationproblem.
learns both the policy and the state-value function, using the The actor network uses a reference neuron technique and a
value function to reduce variance in policy updates. Actor- linearly decreasing Gaussian action noise in the output layer
criticmethodstendtobemorestablethanpurepolicygradient to enhance training efficiency and balance exploration and
methods. In the work of [58], the allocation of transmitted utilization by the agent.
semantic information and resource block (RB) was jointly 8) Proximal Policy Optimization: Proximal Policy
optimized to minimize the average transmission delay, based Optimization (PPO) is proposed by Schulman et al. [169]
ontheimprovedACalgorithm,inwhichanovelvaluefunction in 2017. PPO aims to improve and simplify previous policy
is designed to improve the probability of action exploration gradientalgorithms,suchasTrustRegionPolicyOptimization
and finding the optimal solution. In traditional model-free (TRPO).ThekeyaspectofthePPOalgorithmisthatitmakes
DRL, the value function V(s k+1) is approximated by DNN: the learning process more stable by limiting the magnitude
E s ∼P[V(sk+1)].Inthemodel-basedDRLproposedinthe of policy updates. The authors of [71] designed a power
k+1
article, due to the deterministic nature of the state transitions, allocation algorithm to maximize the total QoE based on
there is E s ∼P[V(sk+1)] = V(sk+1). Therefore, the PPO. The algorithm can appropriately allocate the power of
k+1
proposedalgorithmdoesnotneedtouseDNNtoapproximate public and private messages to maximize the total QoE while
the value function. As a result, the estimation error resulting guaranteeingindividualQoEforeachMSP.Accordingto[95],
fromtheapproximationofthevaluefunctioncanbeprevented, the authors proposed a semantic-aware resource allocation
andthestate-actionvaluefunctioncanbecomputedaccurately. framework with a flexible duty cycle co-existence mechanism
More details about the algorithm can be found in [58]. (SARADC) algorithm that utilizes PPO to optimize resource
6) Deep Deterministic Policy Gradient: Silver et al. [166] allocation in high-speed vehicular networks.
proposed the deterministic policy gradient (DPG) algorithm We mentioned in Section II-B3 that in task-oriented
for RL problems with continuous action spaces. The deter- SemCom systems, the resource allocation is closely tied
ministic policy gradient is the expected gradient of the to the task-related importance of the semantic information.
action-valued function, which integrates over the state space Thistaskdependencenecessitatesadaptiveresourceallocation
andcanbeestimatedmoreefficientlythanthestochasticpolicy schemes that align with the utility of semantic content,
gradient. Lilicrap et al. [167] proposed the deep deterministic ensuring the transmission of task-related and semantically
policy gradient (DDPG) algorithm in the continuous action importantfeatures,whilejointlyoptimizingbandwidth,power,
space by extending DQN and DPG. DQN can only handle and computing resources for overall system performance.
discrete and low-dimensional action spaces, but many cases, However, traditional PPO methods struggle to handle such
especially physical control tasks, have continuous and high- cross-layer optimization under semantic-aware constraints. To
dimensional action spaces, and DQN cannot be directly thisend,[109]and[40]madenovelimprovementstothePPO
applied to continuous domains, so DDPG adopts the AC algorithm. Reference [109] proposes a knowledge-assisted
method based on the DPG algorithm. PPO (K-PPO) algorithm, which utilizes a prior model and
Theauthorsof[112]developedajointoptimizationproblem the Hungarian algorithm to assist PPO in solving the joint
of semantic feature compression rate, transmit power, and optimization problem of importance-aware semantic feature
bandwidth for each smart device to maximize the long-term selection and channel assignment within the joint semantic-
transmission efficiency of the task. A DDPG-based wireless channel transmission (JSCT) mechanism. Meanwhile, [40]
resource allocation scheme is proposed to efficiently handle develops an attention-enhanced PPO (APPO) by introducing
the continuous action space. the attention network [27], enabling the base station to learn
7) Twin Delayed Deep Deterministic Policy Gradient: the correlation between the semantic importance distribution
Twin delayed deep deterministic policy gradient (TD3) is f i(G i)andthetaskperformancemetricMSS,thusoptimizing
proposedbyFujimotoetal.[168]basedontheimprovementof the resource block (RB) allocation and semantic information
theDDPGalgorithm.TheTD3algorithmincorporatestheidea selection strategies accordingly.
of the double Q-learning algorithm into the DDPG algorithm. 9) Soft Actor-Critic: The soft actor-critic (SAC) algo-
Adetaileddescriptioncanreferto[161].FromACandDDPG rithm[170]isamodel-freeDRLalgorithmbasedonmaximum
to TD3, with the evolution of these RL algorithms, there entropy, introducing the concept of maximum entropy on
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2991
TABLEX
results are difficult to interpret, which will affect the credi-
DIFFERENCEINCOMBINATIONOFDRLALGORITHMS
bility and acceptability of the results. c) DRL algorithms are
sensitivetotheselectionofhyperparametersandtrainingdata,
and the instability is higher. Table XI reviews the literature
using DRL-based centralized optimization algorithms.
C. Heuristic Algorithms
A heuristic algorithm is an algorithm based on an intuitive
or empirical construction that usually performs well with
limited computational resources and is suitable for scenarios
with low performance requirements to fulfill engineering
needs. They can provide effective approximations, but are not
guaranteed to find the global-optimal solution.
As we mentioned earlier, semantic similarity does not
have a closed-form expression. This can be regarded as a
closed-box optimization problem, which is difficult to solve
with traditional optimization algorithms. Heuristic algorithms
provide a feasible way to solve the closed-box problem.
top of maximizing future cumulative rewards to enhance Reference [88] proposed a variant of the Whale Optimization
the robustness and exploration ability of agents. In refer- Algorithm (WOA) [171] that introduces a penalty strategy:
ence [53], a dynamic intelligent resource allocation scheme the Whale Optimization Algorithm with a Penalty Strategy
was designed. It is based on SAC and D-SAC to realize (WOARA) to solve the optimal resource allocation problem.
real-time decision-making based on perceptual semantic tasks More details about WOA and WOARA can be seen in [88].
and channel features. Among them, D-SAC is to extend SAC Theauthorsof[53]usetheparticleswarmoptimization(PSO)
to discrete space to solve the discrete variable allocation algorithmtooptimizethecompressionratioandtheallocation
problem. The Four-Soft Actor Critical (4-SAC) algorithm is of power and bandwidth for each user jointly. In [77], the
proposed in [83]. It comprises four SAC intelligent agents, PSO algorithm is developed to determine the computation
which collectively optimize the trajectory of a UAV, number resource allocation in each step of the matching game. There
of semantic symbols, and power allocation to strike a balance are also a few other works that incorporate heuristics into
between data transmission efficiency and energy efficiency, the overall program design, such as [78], which also uses a
and QL was used to facilitate learning for the optimal policy. preference list-based heuristic algorithm for problem solving.
In fact, the challenge of highly coupled and non-convex Furthermore,[66],[70],[118],[131]haveincorporatedsimple
optimization variables is particularly critical in SemCom. heuristics such as greedy algorithms into their overall pro-
Unlikeconventionalsystems,whereoptimizationvariablescan gram design. Table IX also includes papers that use heuristic
often be easily decoupled or approximated linearly. In the algorithms.
resource allocation problem of SemCom, the composition of Summary: Heuristic algorithms have some advantages in
optimization variables is very complex and hard to decouple, terms of cost and convergence speed, but their performance
which may be both in the case of discrete action space: is relatively poor, are prone to fall into local optimal, and
semantic symbol number selection, subchannel allocation, are sensitive to parameters. Therefore, they are applicable for
communication mode selection, and some discretized vari- scenarios that only have requirements on low latency and do
ables, etc., and in the case of continuous action space: power not have high demands on other performance metrics.
allocation, bandwidth allocation, semantic compression rate, In this section, centralized resource allocation optimization
etc.Totacklethiscomplexity,recentstudieschosetocombine algorithms from different literature in SemCom are reviewed.
two or more of these methods to solve the problem [47], These algorithms are categorized into several types, includ-
[55],[61],[71],[90],[94],combiningthevaluefunction-based ing those based on mathematical optimization (Lyapunov
method and the policy gradient-based methods to form a two- optimization, AO algorithm, SCA, etc.), DRL (value-based
layer DRL framework, which is also succinctly summarized and policy-based), and heuristic methods. While previous
in Table X. sections have systematically categorized performance metrics
Summary: Centralized optimization algorithms based on and optimization strategies, it is also crucial to understand
DRL are applicable to highly dynamic scenarios. They have how these elements interact across various network scenarios.
the following advantages: a) They can deal with high- Whileexistingworksproposevariousoptimizationtechniques
dimensional, nonlinear state and action spaces, making them tailored to SemCom scenarios, there remains a lack of in-
suitable for complex decision problems. b) It can adap- depth discussion on how these methods specifically address
tivelylearntheoptimalpolicywithoutexcessivemathematical the unique challenges Table XII provides a challenge-centric
derivationandcomputation.However,italsohasthefollowing synthesis of representative works, their applied optimization
disadvantages: a) High complexity of training. b) Closed-box techniques, and directions for future hybrid or enhanced
process: the learning process is unobservable, and the output methods.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2992 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
TABLEXI
CENTRALIZEDOPTIMIZATIONALGORITHMSBASEDONDRL
V. DISTRIBUTEDRESOURCEALLOCATIONALGORITHMS
Nowadays, the network structure of wireless commu-
nications is increasingly oriented toward a multilevel
heterogeneous network structure, and efficiently managing
resource allocation in such a complex environment requires a
fundamental shift from traditional centralized mechanisms to
self-organizing and self-optimizing approaches [172]. In this
context,moreandmoredistributedmethodshavebeenutilized
to meet the increasingly complex situation. This section will Fig. 11. The taxonomy of distributed resource allocation algorithms in
SemCom.
provide an illustration of the distributed optimization algo-
rithmsusedintheliteratureonresourceallocationinSemCom, A. Multi-Agent Deep Reinforcement Learning
among them matching theory and auctions originating from
Multi-agent reinforcement learning (MARL) is the appli-
thefieldofeconomics,andaportionofreinforcementlearning
cation of reinforcement learning ideas and algorithms to
with multi-agents. Fig. 11 shows the taxonomy of centralized
multi-intelligent systems, extending MARL to deep reinforce-
resource allocation algorithms in SemCom.
ment learning is multi-agent deep reinforcement learning
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2993
TABLEXII
MAPPINGOFSEMCOM-SPECIFICCHALLENGESTOOPTIMIZATIONSTRATEGIESANDPOTENTIALEXTENSIONS
(MADRL).In[117],directDQNisgeneralizedtomulti-agent leveraging a CTDE approach. During training, agents share
DQN,andUAVsinthecoverageareaofdifferentMECservers observations and actions to learn coordinated strategies, while
are considered agents in the DQN algorithm. However, other during execution, each agent acts independently based on its
literature utilized more refined and mature multi-intelligent own policy. Reference [123], the modified MADDPG method
body deep reinforcement learning methods, as follows: is designed to optimize both global system performance
and individual agent behavior in a dynamic semantic
1) Multi-Agent PPO/DDPG Algorithm: The multi-agent
communicationenvironment.Thesimulationresultsshowthat
PPO (MAPPO) algorithm [173] is a variant of the PPO
the proposed algorithm performs better than centralized DRL
algorithm applied to multi-agent tasks, the critic can observe
methods like DDPG and TD3.
the global state, including information about other agents and
the environment. The basic idea of the MAPPO algorithm 2) MADRL Based on Value Decomposition: MADRL
is centralized training and decentralized execution (CTDE). based on Value Decomposition (VD) is one of the many
The optimization problem of joint computational resources MADRLalgorithms.Itutilizessomeconstraintstodecompose
and bandwidth allocation is established in [126] with the the joint action-value function of a multi-agent system into a
objective of maximizing semantic accuracy. The problem is specific combination of individual action-value functions and
then transformed into a DRL framework, and MAPPO is is able to effectively solve problems such as environmental
utilized to solve the problem. non-stability and exponential explosion of the action space
MADDPG (Multi-Agent Deep Deterministic Policy in multi-agent systems, ensuring the convergence of the
Gradient)isspecificallydesignedformulti-agentsystems,also algorithm.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2994 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
In [51], a VD-based DQN is used to allow users and BSs theBERT-basedmodel.Ingeneral,whentheBLEUscoreand
to work together to find a team RB allocation and partial similarity score are higher, the device has a greater incentive
semantic information transmission scheme to optimize the to pay a higher price for energy.
similarity of all users. The work of [42] proposed a VD- Summary: The advantages of the economic approaches
based entropy-maximized MARL (VD-ERL) algorithm. The is: effective in highly dynamic and complex heterogeneous
algorithm enables each server to coordinate its work with networks, practical in real-world scenarios. However, there
other servers in the training phase, perform RB allocation are some disadvantages, which include: a) the global optimal
in a distributed manner, and approximate the global-optimal solution may not be obtained; b) in the auction, the need
performance with fewer training iterations. for an additional third-party trusted organization for auction
Summary: The advantage of MADRL is the ability to management may incur additional costs.
solvecomplexmulti-intelligentcollaborationproblems,which In this section, we explore distributed resource allocation
is in line with the trend of increasingly complex real-world algorithms in SemCom, focusing on multi-agent deep rein-
networkchanges.Thedisadvantagesarethecomplexityofthe forcement learning (MADRL), matching theory and auction
training process and the difficulty of balancing collaboration methods. MADRL, including the multi-agent PPO and value
and competition. decomposition-based algorithms, is discussed for its ability
to handle complex coordination tasks. Economic methods,
such as matching theory and auction theory, are highlighted
B. Economic Methods
for their efficiency in decentralized resource allocation in
There are two main economic methods used in distributed dynamicenvironments.Table XIIIreviewstheliteratureusing
optimization algorithms in resource allocation of SemCom: distributed resource allocation optimization algorithms.
matchingtheoryandauction.Theyaresubfieldsofeconomics
and are promising concepts in distributed resource manage-
VI. OPENCHALLENGESANDFUTURERESEARCH
ment and allocation.
DIRECTIONS
1) Matching Theory: As a powerful tool for studying
Despite the significant achievements in resource allocation
the dynamics and mutually beneficial relationships formed
research in SemCom, many key issues remain unexplored.
between different types of agents, the matching theory
This section discusses several open research challenges and
is particularly well suited to develop practical and high-
future research directions.
performance, low-complexity, decentralized solutions in these
complex networks. In particular, it can effectively cope with
the high dynamics of the network, the selfish, competitive, A. Resource Allocation Under Other SemCom Network
and distributed nature of the network elements, the limited Architectures
wireless resources, and the QoS constraints of the different The diversity of SemCom network architectures reflects
elements [174]. its adaptability to a wide range of application scenarios,
The authors of [66] used many-to-many matching to from dynamic environments to multimodal communications.
solve the subproblem of the association between RIS users. However, without tailored resource allocation strategies, these
However,theauthorsof[22]utilizedmatchingtheorytosolve architectures cannot achieve their full potential. Developing
the subproblems of channel association and power allocation. solutions for specific frameworks is critical to maximize
In order to cope with the tight coupling between users in performance under real-world constraints. This progress will
multi-cell user and bimodal user pairs, a matching game pair expand the scope of SemCom’s applications, driving innova-
is constructed for modeling, and a low-complexity matching tions in areas such as smart homes, autonomous vehicles, and
algorithm is proposed to obtain stable matching in this part. immersive metaverse applications.
The authors of [77] establish a many-to-one matching game 1) NetworkArchitectureCombinedWithNGMA: Resource
to determine the joint communication mode and the channel allocation problem of multi-user SemCom is particularly
selection problem, in which the users and channels act as the critical in scenarios with dense user environments or limited
game players. The computational resource was allocated by communicationresources,mostofthepreviousmultipleaccess
the PSO algorithm in each step of the matching game. methods used FDMA, OFDMA, or TDMA. However, as
2) Auction: As a subfield of economics and business communication technology continues to develop, researchers
management, auction theory provides an interdisciplinary have begun to investigate the application of the combination
technique for the allocation of wireless resources (e.g., of NGMA and SemCom in resource allocation. Incorporating
subchannels, time slots, and transmit power levels) in wire- NGMA into SemCom architectures could lead to transforma-
less systems. Auction methods are widely used in areas tive advances in scenarios requiring high connectivity, such
such as cognitive radio, cellular networks, and wireless grid as smart cities, industrial IoT, and Metaverse. Currently, some
networks [175]. In [85], the bids of IoT devices (bidders) scholars have carried out research in this field; see Section II.
for energy and power transmitters (auctioneers) are used to A 2) of this article.
determine the winner and payment by competing for the 2) Network Architecture in a Dynamic Environment:
energy of the hybrid access point (H-AP) through an optimal Dynamicenvironments,suchasvehicularnetworksordisaster
auctionbasedonDL[176].TheIoTdeviceswillbidforenergy emergency communications, are highly unpredictable due to
based on sentence similarity and BLEU score derived from factors such as user mobility, interference, and time-varying
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2995
TABLEXIII
DISTRIBUTEDOPTIMIZATIONALGORITHMSBASEDONMADRLANDECONOMICMETHODS
channel conditions. Designing resource allocation strategies standpoint,asefficientresourceallocationisessentialtoensure
that account for these dynamic characteristics is critical seamless, high-quality interactions across different types of
to ensuring robust and reliable SemCom performance. The services. At this time, the resource allocation problem for
ability to adapt to dynamic environments directly affects multimodal SemCom networks will also become a major
the network’s ability to deliver semantically accurate and challenge.
timelycommunication.Moreover,thisadaptabilityisessential
for applications like real-time AR/VR, autonomous systems, B. Establishment of SemCom Related Theory
and telemedicine, where latency and semantic accuracy are
Carnap and Bar-Hillel first proposed Classic Semantic
paramount.
Information Theory (CSIT) in 1952, based on logical proba-
3) Network Architecture of Speech SemCom: The rise of bility[177].Inspiredbythispioneeringwork,sometheoretical
speech-basedinterfacesinconsumerelectronics,smarthomes, research has been carried out in the past two decades, such
and healthcare applications highlights the need for optimized as[2]and[178],butitisnotsufficient,especiallyinSemCom
resourceallocationinspeechSemCom.Unliketextandimage based on the DL framework. This gap in foundational theory
modalities, speech signals have unique characteristics, such presents a critical opportunity for advancing SemCom, as
as real-time requirements, continuous data streams, and high developing a solid theoretical framework will enable more
sensitivity to latency and noise. Addressing these challenges effective, robust communication systems in dynamic and
inresourceallocationwillbecrucialforimprovingthequality evolving environments.
and efficiency of speech communication systems. Improved 1) Building a Universal Semantic Information Theory
resource allocation for speech SemCom could enhance appli- Framework: Compared to traditional information theory,
cations such as real-time voice recognition, smart home which has been studied for many years, the development
automation, and voice-assisted healthcare, ensuring that these of semantic information theory is relatively weak. Mainly
systems operate smoothly with minimal delay. reflected in three aspects: a) So far, there has been no
4) NetworkArchitectureofMulti-ModalSemCom: Mostof unified theoretical method for how to represent and measure
the existing SemCom network models are developed around semantics. b) SemCom lacks a comprehensive mathematical
a single modality. However, scenarios such as Metaverse basis. c) It is difficult to extend the theory of traditional
require a multi-modal service model that includes multiple information theory to semantic information theory.
types of instant interactions, such as audio, image, video, and 2) BuildingMoreAdvancedSemanticPerformanceMetrics:
tactileservices.Thisrequiresamulti-modalSemComnetwork The diversity of different scenarios and tasks in which
to solve. As a result, resource allocation for multi-modal SemCom systems are deployed means that a single static
SemCom networks becomes a critical challenge. Traditional performance metric will not suffice. Moreover, shifting
single-modal resource allocation techniques, optimized for towards a user-centric SemCom paradigm is another critical
simple scenarios, are inadequate when it comes to managing challenge.Traditionalmetrics,suchasbiterrorrates,typically
the dynamic and diverse needs of multi-modal data streams. focus on technical performance, but do not capture the real-
This problem is even more important from a user-centric world effectiveness of SemCom systems from the user’s
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2996 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
TABLEXIV
CHALLENGESANDFUTUREDIRECTIONS
perspective. Establishing such advanced metrics will enable a are particularly useful for modeling complex, structured data,
moreaccurateevaluationofasystem’sabilitytoprovidevalue suchasnetworktopologiesorrelationshipsbetweendevicesin
to users, beyond just raw data transmission efficiency. a distributed SemCom environment. GNNs can help optimize
communication pathways and improve resource allocation by
modeling dependencies between nodes in real time, problems
C. SemCom Resource Allocation Optimization Scheme
partially modeled as combination optimization problems, or
1) CombinationofMultipleAlgorithms: Theresourceallo-
other special scenarios. A simple outline below illustrates
cation problem can be solved by combining mathematical
how GNNs could help address channel assignment as a
optimization and the RL method to save computing resources
combinatorial problem in semantic communication systems:
andachievetheoptimalsolution.Similarly,combiningheuris-
1) Graph modeling: Represent users and interference links
tic algorithms with optimization allows fast, near-optimal
as a graph.
solutions in time-sensitive situations. The impact of these
2) Problem setup: Frame channel assignment as a combi-
combined techniques is significant: improved service quality,
natorial optimization problem.
reduced energy consumption, and the ability to meet the
3) GNN encoding: Use GNNs to learn node embeddings
demands of the growing user.
that capture interference and task demands.
2) FL-Enabled and Other DL-Enabled Techniques:
4) Solution generation: Predict channel assignments
Federated learning (FL) is a machine learning technology
directly or guide heuristics with learned scores.
that can train resource scheduling algorithms on multiple
5) Training feedback: Optimize GNN using corresponding
distributed edge devices or servers that do not exchange local
metrics like S-SE.
data samples [179]. FL allows edge devices to collaboratively
Diffusion models, with their generative capabilities, may
learn resource scheduling policies without sharing raw
support joint optimization of semantic compression and
semantic data. This decentralized training paradigm naturally
resource usage under strict delay or accuracy constraints.
protects user privacy, making it well-suited for privacy-
Here are some open research issues about utilizing these DL-
sensitive SemCom applications such as telemedicine or
enabled techniques in resource allocation in SemCom:
autonomous driving. While this approach remains largely
• Communicationoverheadforfrequentmodelupdatescan
unexplored, the following outline presents one possible way
be significant in FL.
to incorporate federated learning into semantic resource
• Lightweight FL protocols are needed for resource-
allocation:
constrained devices.
1) Alocalmodelistrainedoneachdevicetomakeresource
• Scalability challenges for large-scale distributed systems
allocation schemes (e.g., semantic compression ratio,
and difficulty in integrating multiple objectives (e.g.,
offloading decision).
latency, accuracy, energy) into GNN-based models.
2) Devices send model updates (e.g., gradients or parame-
• High computational cost of diffusion models may hinder
ters) to a central aggregator.
real-time deployment.
3) The server performs model aggregation and updates the
global model.
D. Other Challenges
4) Theglobalmodelisredistributedtodevicesforthenext
training round. There are still some other challenges that need to be
Additionally, other deep learning techniques, such as dif- considered in the resource allocation of SemCom.32
fusion models and graph neural networks (GNNs), can also 1) Transformation of Focus: Resource allocation in
play a crucial role in resource allocation optimization. GNNs SemComisundergoingashiftfromsystem-leveloptimization
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2997
to a user-centric approach. Traditionally, resource allocation alsosummarizeandexplainthenetworkstructureandresource
schemes in SemCom primarily aim to maximize system- allocation types in these studies, emphasize the performance
wide performance metrics, such as throughput, latency, indicators and resource allocation optimization algorithms in
or energy efficiency. However, with the growing demand these studies, and provide detailed tables to summarize these
for personalized services and the increasing diversity of studies. We identify current research bottlenecks and chal-
user requirements, future resource allocation schemes must lenges in the allocation of SemCom resources and anticipate
prioritize individual user satisfaction. Despite its importance, further research in the future. We hope that our work can
transforming resource allocation into a user-centric paradigm provide references and insight to future researchers, as well
posessignificantchallenges.Oneofthebiggestobstaclesisthe as encourage follow-up research.
design of meaningful and measurable metrics that accurately
reflectusersatisfaction,asitrequiresaccountingforsubjective
ACKNOWLEDGMENT
factors such as preferences and context. Additionally, these
metricsneedtobedynamicandadaptabletovaryingscenarios, The authors would like to thank the anonymous reviewers
such as real-time changes in user behavior or network for their valuable comments and suggestions, which helped to
conditions. Another challenge lies in the inherent complexity improve the quality of this article.
of resource allocation to meet diverse and sometimes
conflictinguserneeds.Forexample,balancingthesatisfaction REFERENCES
of multiple users while ensuring the fairness and efficient
[1] C. E. Shannon and W. Weaver, The Mathematical Theory of
use of network resources requires advanced algorithms and
Communication.Urbana,IL,USA:Univ.IllinoisPress,1949.
computationally efficient solutions. Moreover, these resource [2] J.Baoetal.,“Towardsatheoryofsemanticcommunication,”inProc.
allocation algorithms must be scalable to accommodate the IEEENetw.Sci.Workshop,2011,pp.110–117.
[3] Z.Qin,X.Tao,J.Lu,W.Tong,andG.Y.Li,“Semanticcommunica-
huge number of devices and users envisioned in 6G.
tions:Principlesandchallenges,”2022,arXiv:2201.01389.
2) Security and Privacy Issue: In existing research on [4] D. Gündüz et al., “Beyond transmitting bits: Context, semantics, and
SemCom resource allocation, security and privacy concerns task-orientedcommunications,”IEEEJ.Sel.AreasCommun.,vol.41,
no.1,pp.5–41,Jan.2023.
have not been adequately addressed. However, the failure
[5] X. Luo, H.-H. Chen, and Q. Guo, “Semantic communications:
of a SemCom model or an attack on the system can sig- Overview,openissues,andfutureresearchdirections,”IEEEWireless
nificantly undermine the reliability and robustness of the Commun.,vol.29,no.1,pp.210–219,Feb.2022.
entire network, rendering resource allocation ineffective. It [6] Z.Luetal.,“Semantics-empoweredcommunications:Atutorial-cum-
survey,” IEEE Commun. Surveys Tuts., vol. 26, no. 1, pp.41–79, 1st
is critical to develop mechanisms to enhance the robustness
Quart.,2024.
of the system and to formulate intrusion detection strategies [7] C.Zhang,H.Zou,S.Lasaulce,W.Saad,M.Kountouris,andM.Bennis,
to protect against vulnerabilities. In real-world applications “Goal-oriented communications for the IoT and application to data
compression,” IEEE Internet Things Mag., vol. 5, no. 4, pp.58–63,
such as semantic-aware IoV, SemCom IIoT, 6G-envisioned
Dec.2022.
telemedicine,andtheMetaverse,largeamountsoftransmitted [8] S. Iyer et al., “A survey on semantic communications for intelli-
data often involve user privacy and even business-sensitive gent wireless networks,” Wireless Pers. Commun., vol. 129, no. 1,
pp.569–611,Mar.2023.
information. If these data are exposed or leaked during trans-
[9] Y.Liu,X.Wang,Z.Ning,M.Zhou,L.Guo,andB.Jedari,“Asurvey
mission and processing, it could lead to substantial financial onsemanticcommunications:Technologies,solutions,applicationsand
and reputational losses. One promising approach to address challenges,”Digit.Commun.Netw.,vol.10,no.3,pp.528–545,2024.
[10] M.KountourisandN.Pappas,“Semantics-empoweredcommunication
privacy issues is the application of federated learning (FL)
fornetworkedintelligentsystems,”IEEECommun.Mag.,vol.59,no.6,
in SemCom. By enabling devices to collaboratively train pp.96–102,Jun.2021.
semantic extraction models without sharing raw data, FL [11] G.Shi,Y.Xiao,Y.Li,andX.Xie,“Fromsemanticcommunicationto
semantic-awarenetworking:Model,architecture,andopenproblems,”
helps preserve user privacy and reduces the risk of sensitive
IEEECommun.Mag.,vol.59,no.8,pp.44–50,Aug.2021.
information leakage. This decentralized learning paradigm is [12] W. Yang et al., “Semantic communications for future Internet:
especially suitable for privacy-critical scenarios, where tradi- Fundamentals, applications, and challenges,” IEEE Commun. Surveys
Tuts.,vol.25,no.1,pp.213–250,1stQuart.,2023.
tionalcentralizedtrainingmaynotbeviable.Therefore,thisis
[13] T.M.Getu,G.Kaddoum,andM.Bennis,“Makingsenseofmeaning:
an important future research direction. This research direction A survey on metrics for semantic and goal-oriented communication,”
is crucial for the scalability and trustworthiness of SemCom IEEEAccess,vol.11,pp.45456–45492,2023.
[14] T. M. Getu, G. Kaddoum, and M. Bennis, “A survey on goal-
networks in the future, helping them meet the increasing
oriented semantic communication: Techniques, challenges, and future
demandsofemergingapplicationswithoutcompromisinguser directions,”IEEEAccess,vol.12,pp.51223–51274,2024.
privacy or system performance. [15] D.Wonetal.,“Resourcemanagement,security,andprivacyissuesin
In this section, we discuss various challenges and future semantic communications: A survey,” IEEE Commun. Surveys Tuts.,
vol.27,no.3,pp.1758–1797,Jun.2025.
research directions for resource allocation in SemCom
[16] Q.Luo,S.Hu,C.Li,G.Li,andW.Shi,“Resourceschedulinginedge
networks. These challenges and their future directions are computing: A survey,” IEEE Commun. Surveys Tuts., vol. 23, no. 4,
summarized in Table XIV. pp.2131–2165,4thQuart.,2021.
[17] A. Sarah, G. Nencioni, and M. M. I. Khan, “Resource allocation in
multi-access edge computing for 5G-and-beyond networks,” Comput.
Netw.,vol.227,May2023,Art.no.109720.
VII. CONCLUSION [18] Naren, A. K. Gaurav, N. Sahu, A. P. Dash, G. Chalapathi, and
V. Chamola, “A survey on computation resource allocation in IoT
In this survey, we provide a systematic and comprehensive
enabled vehicular edge computing,” Complex Intell. Syst., vol. 8,
overview of the resource allocation problem in SemCom. We pp.3683–3705,Oct.2022.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

2998 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
[19] B.Bossy,P.Kryszkiewicz,andH.Bogucka,“Energy-efficientOFDM [44] Z. Hu, T. Liu, C. You, Z. Yang, and M. Chen, “Multiuser resource
radio resource allocation optimization with computational awareness: allocationforsemantic-relay-aidedtexttransmissions,”inProc.IEEE
Asurvey,”IEEEAccess,vol.10,pp.94100–94132,2022. GlobecomWorkshops(GCWkshps),2023,pp.1273–1278.
[20] Y.Teng,M.Liu,F.R.Yu,V.C.M.Leung,M.Song,andY.Zhang, [45] T. Liu, C. You, Z. Hu, C. Wu, Y. Gong, and K. Huang, “Semantic-
“Resourceallocationforultra-densenetworks:Asurvey,someresearch relay-aided text transmission: Placement optimization and bandwidth
issues and challenges,” IEEE Commun. Surveys Tuts., vol. 21, no. 3, allocation,”inProc.IEEEGlobecomWorkshops(GCWkshps),2023,
pp.2134–2168,3rdQuart.,2019. pp.215–220.
[21] E. C. Strinati and S. Barbarossa, “6G networks: Beyond Shannon [46] Y. Lil, X. Zhou, and J. Zhao, “Resource allocation for semantic
towardssemanticandgoal-orientedcommunications,”Comput.Netw., communication under physical-layer security,” in Proc. IEEE Global
vol.190,no.8,pp.1–17,May2021. Commun.Conf.,2023,pp.2063–2068.
[22] L. Yan, Z. Qin, R. Zhang, Y. Li, and G. Y. Li, “QoE-aware resource [47] H. Hu, X. Zhu, F. Zhou, W. Wu, and R. Q. Hu, “Semantic-oriented
allocation for semantic communication networks,” in Proc. IEEE resource allocation for multi-modal UAV semantic communica-
GlobalCommun.Conf.,2022,pp.3272–3277. tion networks,” in Proc. IEEE Global Commun. Conf., 2023,
[23] N.Farsad,M.Rao,andA.Goldsmith,“Deeplearningforjointsource- pp.7213–7218.
channelcodingoftext,”inProc.IEEEInt.Conf.Acoust.,SpeechSignal [48] X. He, C. You, and T. Q. Quek, “Joint user association and resource
Process.(ICASSP),2018,pp.2326–2330. allocationformulti-cellnetworkswithadaptivesemanticcommunica-
[24] M. Rao, N. Farsad, and A. Goldsmith, “Variable length joint source tion,”2024,arXiv:2312.01049.
channelcodingoftextusingdeepneuralnetworks,”inProc.19thInt. [49] Y. Zheng,T. Zhang,R.Huang,andY.Wang, “Computingoffloading
WorkshopSignalProcessAdv.WirelessCommun.(SPAWC),Kalamata, and semantic compression for intelligent computing tasks in MEC
2018,pp.1–5. systems,”inProc.IEEEWirelessCommun.Netw.Conf.(WCNC),2023,
[25] H. Xie, Z. Qin, G. Y. Li, and B.-H. Juang, “Deep learning enabled pp.1–6.
semantic communication systems,” IEEE Trans. Signal Process., vol. [50] Z.JiandZ.Qin,“Energy-efficienttaskoffloadingforsemantic-aware
69,pp.2663–2675,2021. networks,”inProc.IEEEInt.Conf.Commun.,2023,pp.3584–3589.
[26] Q.Zhou,R.Li,Z.Zhao,C.Peng,andH.Zhang,“Semanticcommu- [51] M. Chen, Y. Wang, and H. V. Poor, “Performance optimization for
nicationwithadaptiveuniversaltransformer,”IEEEWirelessCommun. wireless semantic communications over energy harvesting networks,”
Lett.,vol.11,no.3,pp.453–457,Mar.2022. in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP),
[27] A.Vaswanietal.,“Attentionisallyouneed,”inAdvancesinNeural 2022,pp.8647–8651.
InformationProcessingSystems,vol.30.RedHook,NY,USA:Curran
[52] O. Marnissi, H. E. Hammouti, and E. H. Bergou, “Semantic-aware
Assoc.,Inc.,2017.
resource allocation in constrained networks with limited user partici-
[28] S. Jiang et al., “Reliable semantic communication system enabled by
pation,”inProc.IEEEWirelessCommun.Netw.Conf.(WCNC),2024,
knowledgegraph,”Entropy,vol.24,no.6,p.846,2022.
pp.1–6.
[29] J.Liang,Y.Xiao,Y.Li,G.Shi,andM.Bennis,“Life-longlearningfor
[53] L. Wang, W. Wu, F. Zhou, Z. Yang, Z. Qin, and Q. Wu, “Adaptive
reasoning-based semantic communication,” in Proc. IEEE Int. Conf.
resource allocation for semantic communication networks,” IEEE
Commun.Workshops(ICCWorkshops),2022,pp.271–276.
Trans.Commun.,vol.72,no.11,pp.6900–6916,Nov.2024.
[30] E. Bourtsoulatze, D. B. Kurka, and D. Gündüz, “Deep joint source-
[54] Y. Wang, M. Chen, W. Saad, T. Luo, S. Cui, and H. V. Poor,
channel coding for wireless image transmission,” IEEE Trans. Cogn.
“Performanceoptimizationforsemanticcommunications:Anattention-
Commun.Netw.,vol.5,no.3,pp.567–579,Sep.2019.
based learning approach,” in Proc. IEEE Global Commun. Conf.
[31] C.Dong,H.Liang,X.Xu,S.Han,B.Wang,andP.Zhang,“Semantic
(GLOBECOM),2021,pp.1–6.
communication system based on semantic slice models propagation,”
[55] L.Wang,W.Wu,F.Tian,andH.Hu,“Intelligentresourceallocation
IEEEJ.Sel.AreasCommun.,vol.41,no.1,pp.202–213,Jan.2023.
forUAV-enabledspectrumsharingsemanticcommunicationnetworks,”
[32] M.U.Lokumarambage,V.S.S.Gowrisetty,H.Rezaei,T.Sivalingam,
in Proc. IEEE 23rd Int. Conf. Commun. Technol. (ICCT), 2023,
N.Rajatheva,andA.Fernando,“Wirelessend-to-endimagetransmis-
pp.1359–1363.
sion system using semantic communications,” IEEE Access, vol. 11,
[56] G.Cheng,X.Wang,D.Li,R.Jiang,andY.Xu,“Resourceallocation
pp.37149–37163,2023.
for multi-cell semantic communication based on deep reinforcement
[33] S.KadamandD.I.Kim,“Semanticcommunication-empoweredtraffic
learning,” in Proc. IEEE 23rd Int. Conf. Communication Technol.
managementusingvehiclecountprediction,”2023,arXiv:2307.12254.
(ICCT),2023,pp.528–533.
[34] Z. Weng and Z. Qin, “Semantic communication systems for speech
[57] G.Ding,S.Liu,J.Yuan,andG.Yu,“JointURLLCtrafficscheduling
transmission,” IEEE J Sel Areas Commun, vol. 39, pp.2434–2444,
and resource allocation for semantic communication systems,” IEEE
Aug.2021.
Trans.WirelessCommun.,vol.23,no.7,pp.7278–7290,Jul.2024.
[35] Z.Weng,Z.Qin,andG.Y.Li,“Semanticcommunicationsforspeech
recognition,” in Proc. IEEE Global Commun. Conf. (GLOBECOM), [58] W. Zhang,Y. Wang, M. Chen, T. Luo, andD. Niyato, “Optimization
2021,pp.1–6. ofimagetransmissioninsemanticcommunicationnetworks,”inProc.
[36] L.Yan,Z.Qin,R.Zhang,Y.Li,andG.Y.Li,“Resourceallocationfor IEEEGlobalCommun.Conf.,2022,pp.5965–5970.
textsemanticcommunications,”IEEEWirelessCommun.Lett.,vol.11, [59] Y. Cang et al., “Online resource allocation for semantic-aware edge
no.7,pp.1394–1398,Jul.2022. computing systems,” IEEE Internet Things J., vol. 11, no. 17,
[37] Z.Weng,Z.Qin,andX.Tao,“Task-orientedsemanticcommunications pp.28094–28110,Sep.2024.
forspeechtransmission,”inProc.IEEE98thVeh.Technol.Conf.(VTC- [60] Y. Cang et al., “Resource allocation for semantic-aware mobile
Fall),2023,pp.1–5. edge computing systems,” in Proc. IEEE Globecom Workshops (GC
[38] Z.Weng,Z.Qin,X.Tao,C.Pan,G.Liu,andG.Y.Li,“Deeplearning Wkshps),2023,pp.1585–1590.
enabledsemanticcommunicationswithspeechrecognitionandsynthe- [61] H.Zhang,H.Wang,Y.Li,K.Long,andV.C.Leung,“Towardintel-
sis,” IEEE Trans. Wireless Commun., vol. 22, no. 9, pp.6227–6240, ligent resource allocation on task-oriented semantic communication,”
Sep.2023. IEEEWirelessCommun.,vol.30,no.3,pp.70–77,Jun.2023.
[39] D. Huang, X. Tao, F. Gao, and J. Lu, “Deep learning-based image [62] M. Poposka, H. A. Suraweera, G. K. Karagiannidis, and
semanticcodingforsemanticcommunications,”inProc.IEEEGlobal Z.Hadzi-Velkov, “Semantic wireless networks with minimal energy
Commun.Conf.(GLOBECOM),2021,pp.1–6. consumption,” IEEE Commun. Lett., vol. 28, no. 8, pp.1894–1898,
[40] Y. Wang et al., “Performance optimization for semantic communica- Aug.2024.
tions: An attention-based reinforcement learning approach,” IEEE J. [63] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting multiple access
Sel.AreasCommun.,vol.40,no.9,pp.2598–2613,Sep.2022. for downlink communication systems: Bridging, generalizing, and
[41] F. Zhou, Y. Li, X. Zhang, Q. Wu, X. Lei, and R. Q. Hu, “Cognitive outperforming SDMA and NOMA,” EURASIP J. Wireless Commun.
semanticcommunicationsystemsdrivenbyknowledgegraph,”inProc. Netw.,vol.2018,no.1,p.133,May2018.
IEEEInt.Conf.Commun.,2022,pp.4860–4865. [64] Z. Zhao, Z. Yang, M. Chen, Z. Zhang, and H. V. Poor, “A joint
[42] W.Zhang,Y.Wang,M.Chen,T.Luo,andD.Niyato,“Optimizationof communication and computation design for probabilistic semantic
imagetransmissionincooperativesemanticcommunicationnetworks,” communications,” Entropy, vol. 26, no. 5, p. 394, 2024. [Online].
IEEETrans.WirelessCommun.,vol.23,no.2,pp.861–873,Feb.2024. Available:https://www.mdpi.com/1099-4300/26/5/394
[43] J. Kang et al., “Personalized saliency in task-oriented semantic com- [65] Z.Zhaoetal.,“Multi-userprobabilisticsemanticcommunicationwith
munications: Image transmission and performance analysis,” IEEE J. semantic compression ratio optimization,” in Proc. IEEE Int. Conf.
Sel.AreasCommun.,vol.41,no.1,pp.186–201,Jan.2023. Commun.Workshops(ICCWorkshops),2024,pp.1647–1652.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 2999
[66] Z. Zhao et al., “A joint communication and computation design for [87] Z. Yang, M. Chen, Z. Zhang, C. Huang, and Q. Yang, “Performance
distributedRIS-assistedprobabilisticsemanticcommunicationinIIoT,” optimizationofenergyefficientsemanticcommunicationsoverwireless
IEEEInternetThingsJ.,vol.11,no.16,pp.26568–26579,Aug.2024. networks,” in Proc. IEEE 96thVeh. Technol. Conf. (VTC-Fall), 2022,
[67] Z.Yang,M.Chen,Z.Zhang,andC.Huang,“Energyefficientsemantic pp.1–5.
communicationoverwirelessnetworkswithratesplitting,”IEEEJ.Sel. [88] A.Xiao,K.Zhao,Z.Liu,andC.Liang,“Energyefficiencyinsemantic
AreasCommun.,vol.41,no.5,pp.1484–1495,May2023. networks:Aheuristicoptimizationapproachforresourceallocation,”in
[68] C. Zeng et al., “Task-oriented semantic communication over rate Proc.28thAsia–PacificConf.Commun.(APCC),2023,pp.219–224.
splittingenabledwirelesscontrolsystemsforURLLCservices,”IEEE [89] L.ChenandJ.Gong,“Multi-sourceschedulingandresourceallocation
Trans.Commun.,vol.72,no.2,pp.722–739,Feb.2024. forage-of-semantic-importanceoptimizationinstatusupdatesystems,”
[69] R.Xu,Z.Yang,Z.Zhao,Q.Yang,andZ.Zhang,“Resourceallocation inProc.IEEEWirelessCommun.Netw.Conf.(WCNC),2024,pp.1–6.
forgreenprobabilisticsemanticcommunicationwithratesplitting,”in [90] B.Hu,J.Ma,Z.Sun,J.Liu,R.Li,andL.Wang,“DRL-basedintel-
Proc. IEEE Int. Conf. Commun. Workshops (ICC Workshops), 2024, ligent resource allocation for physical layer semantic communication
pp.2017–2022. withIRS,”Phys.Commun.,vol.63,Apr.2024,Art.no.102270.
[70] Z. Zhao et al., “Spectral efficiency Maximization for probabilistic [91] J. Dai, H. Fan, Z. Zhao, Y. Sun, and Z. Yang, “Secure resource
semanticcommunicationwithratesplitting,”inProc.IEEE99thVeh. allocationforintegratedsensingandsemanticcommunicationsystem,”
Technol.Conf.(VTC-Spring),2024,pp.1–5. inProc.IEEEInt.Conf.Commun.Workshops(ICCWorkshops),2024,
[71] Y.Chengetal.,“Resourceallocationandcommonmessageselection pp.1225–1230.
for task-oriented semantic information transmission with RSMA,” [92] J.X.Dai,H.Fan,Z.X.Zhaoetal.,“Jointcommunicationandcompu-
IEEE Trans. Wireless Commun., vol. 23, no. 6, pp.5557–5570, Jun. tationdesignforsecureintegratedsensingandsemanticcommunication
2024. system,”Sci.ChinaInf.Sci.,vol.68,no.3,2025,Art.no.132301.
[72] M. Lu, J. Huang, T. Yang, Y. Wang, J. Jiao, and Q. Zhang, [93] Y.Yang,M.Shikh-Bahaei,Z.Yang,C.Huang,W.Xu,andZ.Zhang,
“Utilitylossofinformation-optimalforsemanticempoweredRSMAin “Jointsemanticcommunicationandtargetsensingfor6Gcommunica-
satellite-integratedInternet,”IEEEInternetThingsJ.,vol.11,no.24, tionsystem,”2024,arXiv:2401.17108.
pp.40572–40587,Dec.2024.
[94] Y.Zhang,J.Li,G.Mu,andX.Chen,“ADRL-basedresourceallocation
[73] N. G. Evgenidis et al., “Delay minimization for hybrid semantic- for IRS-enhanced semantic spectrum sharing networks,” EURASIP J.
Shannon communications,” in Proc. IEEE Wireless Commun. Netw. Adv.SignalProcess.,vol.2024,no.1,pp.1–17,2024.
Conf.(WCNC),2024,pp.1–6.
[95] Z.Shao,Q.Wu,P.Fan,N.Cheng,Q.Fan,andJ.Wang,“Semantic-
[74] J.Zhao,M.Chen,Z.Yang,C.You,andM.Chen,“Resourceallocation
aware resource allocation based on deep reinforcement learning
forsemanticrelayaidedwirelessnetworkswithprobabilitygraph,”in
for 5G-V2X HetNets,” IEEE Commun. Lett., vol. 28, no. 10,
Proc.IEEEInt.Conf.Commun.,2024,pp.5317–5322.
pp.2452–2456,Oct.2024.
[75] M. Zhang, R. Zhong, X. Mu, Y. Chen, and Y. Liu, “Resource man-
[96] Z.Shaoetal.,“Semantic-awarespectrumsharinginInternetofVehicles
agementforheterogeneoussemanticandbitcommunicationsystems,”
basedondeepreinforcementlearning,”IEEEInternetThingsJ.,vol.11,
inProc.IEEEInt.Conf.Commun.Workshops(ICCWorkshops),2023,
no.23,pp.38521–38536,Dec.2024.
pp.1629–1634.
[97] L.Xia,Y.Sun,X.Li,G.Feng,andM.A.Imran,“Wirelessresource
[76] H.Noh,S.Park,andH.J.Yang,“Deepreinforcementlearning-based
managementinintelligentsemanticcommunicationnetworks,”inProc.
resourceallocationandmodeselectionforsemanticcommunication,”
IEEE Conf. Comput. Commun. Workshops (INFOCOM WKSHPS),
inProc.22ndInt.Symp.Model.Optim.Mobile,AdHoc,WirelessNetw.
2022,pp.1–6.
(WiOpt),2024,pp.1–6.
[98] L.Xia,Y.Sun,D.Niyato,X.Li,andM.A.Imran,“Jointuserassoci-
[77] P. Li, Y. Wang, M. Liu, and H. Liu, “Matching game based resource
ationandbandwidthallocationinsemanticcommunicationnetworks,”
allocation scheme for adaptive semantic and bit communication
IEEETrans.Veh.Technol.,vol.73,no.2,pp.2699–2711,Feb.2024.
networks,”inProc.IEEE99thVeh.Technol.Conf.(VTC-Spring),2024,
[99] X. Jia, X. Wang, Y. Zhang, M. Sheng, and G. Cheng, “Resource
pp.1–7.
allocation for multi-cell semantic communication systems based on
[78] L. Xia, Y. Sun, D. Niyato, L. Zhang, and M. A. Imran,
DRL,”inProc.12thInt.Conf.Inf.Syst.Comput.Technol.(ISCTech),
“Wirelessresourceoptimizationinhybridsemantic/bitcommunication
2024,pp.1–6.
networks,” IEEE Trans. Commun., vol. 73, no. 5, pp. 3318–3332,
[100] L. Li, J. Dai, Z. Yang, Q. Yang, C. Huang, and Z. Zhang, “Joint
May2025.
compression ratio and user association for multi-cell probabilistic
[79] J. Li, H. Gao, T. Lv, and Y. Lu, “Deep reinforcement learning based
semantic communication,” in Proc. IEEE/CIC Int. Conf. Commun.
computation offloading and resource allocation for MEC,” in Proc.
China(ICCCWorkshops),2024,pp.645–650.
IEEEWirelessCommun.Netw.Conf.(WCNC),2018,pp.1–6.
[101] X. Pu, T. Lei, W. Wen, and Q. Chen, “Enhancing communication
[80] Y. Liu, H. Yu, S. Xie, and Y. Zhang, “Deep reinforcement learning
efficiency of semantic transmission via joint processing technique,”
for offloading and resource allocation in vehicle edge comput-
IEEECommun.Lett.,vol.28,no.3,pp.657–661,Mar.2024.
ing and networks,” IEEE Trans. Veh. Technol., vol. 68, no. 11,
pp.11158–11168,Nov.2019. [102] J.Suetal.,“Semanticcommunication-baseddynamicresourcealloca-
[81] S. Wang, T. Lv, W. Ni, N. C. Beaulieu, and Y. J. Guo, “Joint tion in d2d vehicular networks,” IEEE Trans. Veh. Technol., vol. 72,
resource management for MC-NOMA: A deep reinforcement learn- no.8,pp.10784–10796,Aug.2023.
ing approach,” IEEE Trans. Wireless Commun., vol. 20, no. 9, [103] L.Wang,W.Wu,F.Zhou,Z.Qin,andQ.Wu,“IRS-enhancedsecure
pp.5672–5688,Sep.2021. semantic communication networks: Cross-layer and context-awared
[82] R. Lin, C. Guo, J. Chen, and Y. Wang, “Multidimensional resource resource allocation,” IEEE Trans. Wireless Commun., vol. 24, no. 1,
jointallocationalgorithmbasedondeepdoubleQnetworkinsemantic pp.494–508,Jan.2025.
communication, (in Chinese),” Mobile Commun., vol. 47, no. 4, [104] X. Xu, C. He, X. Li, and J. Xu, “Joint optimization trajectory and
pp.45–53,2023. resourceallocationforUAV-assistedsemanticcommunications,”Phys.
[83] H. Wang, L. Wang, and W. Wu, “Resource allocation and intelli- Commun.,vol.68,Feb.2025,Art.no.102555.
genttrajectoryoptimizationforUAV-assistedsemanticcommunication [105] Y. Li, X. Zhou, and J. Zhao, “Resource allocation for the training
system,” in Proc. IEEE 23rd Int. Conf. Commun. Technol. (ICCT), of image semantic communication networks,” IEEE Trans. Wireless
2023,pp.1370–1374. Commun.,vol.24,no.4,pp.2968–2984,Apr.2025.
[84] Z. Zhao, Z. Yang, Q.-V. Pham, Q. Yang, and Z. Zhang, “Semantic [106] C. Liu, C. Guo, and Y. Yang, “Performance optimization for task-
communication with probability graph: A joint communication and oriented communications,” in Proc. IEEE Int. Conf. Commun., 2024,
computation design,” in Proc. IEEE 98th Veh. Technol. Conf. (VTC- pp.968–973.
Fall),2023,pp.1–5. [107] J. Liu, Y. Lu, H. Wu, and Y. Dai, “Efficient resource allocation
[85] Z.Q.Liew,Y.Cheng,W.Y.B.Lim,D.Niyato,C.Miao,andS.Sun, and semantic extraction for federated learning empowered vehicular
“Economics of semantic communication system in wireless powered semantic communication,” in Proc. IEEE 98th Veh. Technol. Conf.
Internet of Things,” in Proc. IEEE Int. Conf. Acoust., Speech Signal (VTC-Fall),2023,pp.1–5.
Process.(ICASSP),2022,pp.8637–8641. [108] X.Xiang,X.Li,Q.Cui,X.Zhang,andX.Tao,“EoSI-awareresource
[86] Q. Cai et al., “Query-aware semantic encoder-based resource alloca- allocationforsemanticcommunication-enabledindustrialIoTsystem,”
tion in task-oriented communications,” IEEE Trans. Mobile Comput., inProc.Int.Conf.WirelessCommun.SignalProcess.(WCSP),2023,
vol.24,no.7,pp.6413–6429,Jul.2025. pp.477–483.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

3000 IEEECOMMUNICATIONSSURVEYS&TUTORIALS,VOLUME28,2026
[109] Y. Wang et al., “Feature importance-aware task-oriented semantic [131] C. Feng, K. Zheng, Y. Wang, K. Huang, and Q. Chen, “Goal-
transmission and optimization,” IEEE Trans. Cogn. Commun. Netw., orientedwirelesscommunicationresourceallocationforcyber-physical
vol.10,no.4,pp.1175-1189,Aug.2024. systems,” IEEE Trans. Wireless Commun., vol. 23, no. 11,
[110] G. Liu, H. Du, D. Niyato, J. Kang, Z. Xiong, and B. H. Soong, pp.15768–15783,Nov.2024.
“Vision-based semantic communications for metaverse services: A [132] Z.Zhao,Z.Yang,Q.Yang,C.Huang,M.Shikh-Bahaei,andZ.Zhang,
contest theoretic approach,” in Proc. IEEE Global Commun. Conf., “Sum rate maximization for distributed riss assisted probabilistic
2023,pp.2426–2432. semantic communication,” in Proc. IEEE 34th Int. Workshop Mach.
[111] C.Liu,C.Guo,Y.Yang,andN.Jiang,“Adaptablesemanticcompres- Learn.SignalProcess.(MLSP),2024,pp.1–6.
sionandresourceallocationfortask-orientedcommunications,”IEEE [133] K.Brunnströmetal.,“QualiNetwhitepaperondefinitionsofquality
Trans.Cogn.Commun.Netw.,vol.10,no.3,pp.769–782,Jun.2024. of experience,” presented at Eur. Netw. Qual. Exp. Multimedia Syst.
[112] H.Zhang,H.Wang,Y.Li,K.Long,andA.Nallanathan,“DRL-driven Services(COSTActionIC1003),2013.
dynamic resource allocation for task-oriented semantic communica- [134] N. Banovic´-C´urguz and D. Iliševic´, “Mapping of QoS/QoE in 5G
tion,”IEEETrans.Commun.,vol.71,no.7,pp.3992–4004,Jul.2023. networks,” in Proc. 42nd Int. Conv. Inf. Commun. Technol., Electron.
[113] B. Du et al., “YOLO-based semantic communication with generative Microelectron.(MIPRO),2019,pp.404–408.
AI-aided resource allocation for digital twins construction,” IEEE [135] A. Takahashi, “Framework and standardization of quality of experi-
InternetThingsJ.,vol.11,no.5,pp.7664–7678,Mar.2024. ence (QoE) design and management for audiovisual communication
[114] J.Zheng,B.Du,H.Du,J.Kang,D.Niyato,andH.Zhang,“Energy- services,”NTTTech.Rev.,vol.7,no.4,pp.1–5,2009.
efficient resource allocation in generative AI-aided secure semantic
[136] M.I.Belghazietal.,“Mutualinformationneuralestimation,”inProc.
mobile networks,” IEEE Trans. Mobile Comput., vol. 23, no. 12,
Int.Conf.Mach.Learn.,Stockholm,Sweden,Jul.2018,pp.531–540.
pp.11422–11435,Dec.2024.
[137] R. Kneser and H. Ney, “Improved backing-off for m-gram language
[115] W. C. Ng, H. Du, W. Y. B. Lim, Z. Xiong, D. Niyato, and C. Miao,
modeling,” in Proc. IEEE Int. Conf. Acoust. Speech, Signal Process.,
“Stochasticresourceallocationforsemanticcommunication-aidedvir-
1995,pp.181–184.
tualtransportationnetworksinthemetaverse,”inProc.IEEEWireless
[138] M. E. Peters et al., “Deep contextualized word representations,” in
Commun.Netw.Conf.(WCNC),2024,pp.1–6.
Proc.NorthAmer.ChapterAssoc.Comput.Linguist.Hum.Lang.Tech.,
[116] H. Saadat, A. Albaseer, M. Abdallah, A. Mohamed, and A. Erbad,
Jun.2018,pp.2227–2237.
“Energy-aware service offloading for semantic communications in
wireless networks,” in Proc. IEEE Int. Conf. Commun., 2024, [139] S. Banerjee and A. Lavie, “METEOR: An automatic metric for MT
pp.5467–5472. evaluationwithimprovedcorrelationwithhumanjudgments,”inProc.
[117] X.Sun,J.Chen,andC.Guo,“Semantic-drivencomputationoffloading ACL Workshop Intrinsic Extrinsic Eval. Meas. Mach. Transl. /Or
andresourceallocationforUAV-assistedmonitoringsysteminvehic- Summarization,2005,pp.65–72.
ular networks,” in Proc. 48th Annu. Conf. IEEE Ind. Electron. Soc., [140] A. Radford et al., “Learning transferable visual models from natural
2022,pp.1–6. languagesupervision,”2021,arXiv:2103.00020.
[118] Y. Zheng, T. Zhang, and J. Loo, “Dynamic multi-time scale user [141] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image
admission and resource allocation for semantic extraction in MEC qualityassessment:Fromerrorvisibilitytostructuralsimilarity,”IEEE
systems,”IEEETrans.Veh.Technol.,vol.72,no.12,pp.16441–16453, Trans.ImageProcess.,vol.13,pp.600–612,2004.
Dec.2023. [142] Y.Ao,Y.Li,S.He,D.Chen,Z.Qin,andX.Tao,“Researchonresource
[119] X. Han, B. Feng, Y. Shi, Y. Wu, and W. Zhang, “Semantic- allocationincellularsemanticcommunicationsystems,(inChinese),”
aware resource allocation for wireless image transmission,” in Proc. MobileCommun.,vol.48,no.2,pp.104–110,2024.
IEEE/CICInt.Conf.Commun.China(ICCC),2024,pp.2071–2076. [143] K. Niu et al., “A paradigm shift toward semantic communications,”
[120] X. Yang, H. Yang, Y. Jiang, A. Alphones, and L. Xiao, “Game- IEEECommun.Mag.,vol.60,no.11,pp.113–119,Nov.2022.
guidedmatchingtheory-basedresourceallocationforsecuresemantic [144] S. Kadam and D. I. Kim, “Knowledge-aware semantic communi-
communications,” IEEE Trans. Veh. Technol., vol. 74, no. 5, cation system design,” in Proc. IEEE Int. Conf. Commun., 2023,
pp.8357–8362,May2025. pp.6102–6107.
[121] L. Yan, Z. Qin, C. Li, R. Zhang, Y. Li, and X. Tao, “QoE-based [145] H. Xie, Z. Qin, X. Tao, and K. B. Letaief, “Task-oriented multi-user
semantic-aware resource allocation for multi-task networks,” IEEE semanticcommunications,”IEEEJ.Sel.AreasCommun.,vol.40,no.9,
Trans.WirelessCommun.,vol.23,no.9,pp.11958–11971,Sep.2024. pp.2584–2597,Sep.2022.
[122] C. Huang, X. Chen, G. Chen, P. Xiao, G. Y. Li, and W. Huang, [146] Y. Wang et al., “A survey on metaverse: Fundamentals, security, and
“Deepreinforcementlearning-basedresourceallocationforhybridbit privacy,” IEEE Commun. Surveys Tuts., vol. 25, no. 1, pp.319–352,
andgenerativesemanticcommunicationsinspace-air-groundintegrated Jan.2023.
networks,”2024,arXiv:2412.05647. [147] H.Duetal.,“Attention-awareresourceallocationandQoEanalysisfor
[123] Z. Shao et al., “Semantic-aware resource management for C- metaverse xURLLC services,” IEEE J. Sel. Areas Commun., vol. 41,
V2X platooning via multi-agent reinforcement learning,” 2024, no.7,pp.2158–2175,Jul.2023.
arXiv:2411.04672.
[148] H. Du et al., “Exploring attention-aware network resource allocation
[124] J. Chen, C. Feng, C. Guo, Y. Yang, Q. Sun, and M. Zhu, “Video
for customized metaverse services,” IEEE Netw., vol. 37, no. 6,
semantics-drivenresourceallocationalgorithminInternetofVehicles,
pp.166–175,Nov.2023.
(inChinese),”J.Commun.,vol.42,no.7,pp.1–11,2021.
[149] A. Kosta, N. Pappas, and V. Angelakis, “Age of information: A new
[125] R. Lin, C. Guo, B. Zhang, J. Chen, and H. Li, “Tasks-oriented concept, metric, and tool,” Found. Trends(cid:2)R Netw., vol. 12, no. 3,
channeloptimizationandresourceallocationinvehicularnetworks:A
pp.162–259,2017.
hierarchicalreinforcementlearningbasedapproach,”IEEETrans.Veh.
[150] S. Guo, Y. Wang, S. Li, and N. Saeed, “Semantic importance-aware
Technol.,vol.74,no.5,pp.7624–7636,May2025.
communications using pre-trained language models,” IEEE Commun.
[126] F. Zhao, G. Bagwe, E. Mohammed, L. Feng, L. Zhang, and Y. Sun,
Lett.,vol.27,no.9,pp.2328–2332,Sep.2023.
“Joint computing resource and bandwidth allocation for semantic
communication networks,” in Proc. IEEE 98th Veh. Technol. Conf. [151] S.KadamandD.I.Kim,“Knowledge-awaresemanticcommunication
(VTC-Fall),2023,pp.1–5. systemdesignanddataallocation,”IEEETrans.Veh.Technol.,vol.73,
[127] Y. Zhu, X. Yuan, Y. Hu, and A. Schmeink, “Semantic reliability no.4,pp.5755–5769,Apr.2024.
Maximization:Acooperativeperspectiveinintegratedsensing,commu- [152] M. Noor-A-Rahim, Z. Liu, H. Lee, G. G. M. N. Ali, D. Pesch, and
nication and computation networks,” in Proc. IEEE Global Commun. P.Xiao,“Asurveyonresourceallocationinvehicularnetworks,”IEEE
Conf.,2023,pp.5073–5079. Trans.Intell.Transp.Syst.,vol.23,no.2,pp.701–721,Feb.2022.
[128] S. Liang et al., “Fair resource allocation for probabilistic semantic [153] H. W. Kuhn, “The Hungarian method for the assignment problem,”
communicationinIloT,”inProc.IEEE/CICInt.Conf.Commun.China Nav.Res.LogisticsQuart.,vol.2,nos.1–2,pp.83–97,1955.
(ICCCWorkshops),2024,pp.242–247. [154] Z. Zhou, Y. Guo, Y. He, X. Zhao, and W. M. Bazzi, “Access
[129] H.Chen,F.Fang,andX.Wang,“Semanticextractionmodelselection controlandresourceallocationforM2Mcommunicationsinindustrial
for IoT devices in edge-assisted semantic communications,” IEEE automation,”IEEETrans.Ind.Informat.,vol.15,no.5,pp.3093–3103,
Commun.Lett.,vol.28,no.7,pp.1733–1737,Jul.2024. May2019.
[130] S. Hua et al., “Optimizing spectral efficiency through bandwidth [155] Z. Yu, Y. Gong, S. Gong, and Y. Guo, “Joint task offloading and
managementinsemanticcommunicationsystems,”inProc.IEEEInt. resource allocation in UAV-enabled mobile edge computing,” IEEE
Conf.Commun.Workshops(ICCWorkshops),2024,pp.1635–1640. InternetThingsJ.,vol.7,no.4,pp.3147–3159,Apr.2020.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.

---PAGE BREAK---

ZHANGetal.:RESOURCEALLOCATIONINWIRELESSSem-Com:ACOMPREHENSIVESURVEY 3001
[156] H.Zhang,N.Yang,W.Huangfu,K.Long,andV.C.M.Leung,“Power [176] P.Dütting,Z.Feng,H.Narasimhan,D.Parkes,andS.S.Ravindranath,
control based on deep reinforcement learning for spectrum sharing,” “Optimal auctions through deep learning,” in Proc. Int. Conf. Mach.
IEEE Trans. Wireless Commun., vol. 19, no. 6, pp.4209–4219, Learn.,2019,pp.1706–1715.
Jun.2020. [177] R. Carnap and Y. Bar-Hillel, “An outline of a theory of seman-
[157] Z.Ding,R.Schober,andH.V.Poor,“No-painno-gain:DRLassisted tic information,” Res. Lab. Electron., Massachusetts Inst. Technol.,
optimizationinenergy-constrainedCR-NOMAnetworks,”IEEETrans. Cambridge,MA,USA,Rep.247,Oct.1952.
Commun.,vol.69,no.9,pp.5917–5932,Sep.2021. [178] B. Güler, A. Yener, and A. Swami, “The semantic communication
[158] C.J.C.H.WatkinsandP.Dayan,“Q-learning,”Mach.Learn.,vol.8, game,”IEEETrans.Cogn.Commun.Netw.,vol.4,no.4,pp.787–802,
pp.279–292,May1992. Dec.2018.
[159] V. Mnih et al., “Human-level control through deep reinforcement [179] J. Konecˇny`, B. McMahan, and D. Ramage, “Federated
learning,”Nature,vol.518,no.7540,pp.529–533,2015. optimization: Distributed optimization beyond the datacenter,” 2015,
[160] L.-J. Lin, “Self-improving reactive agents based on reinforcement arXiv:1511.03575.
learning, planning and teaching,” Mach. Learn., vol. 8, nos. 3–4,
pp.293–321,1992.
[161] Y. Li, “Deep reinforcement learning: An overview,” 2017,
arXiv:1701.07274.
[162] H.vanHasselt,A.Guez,andD.Silver,“Deepreinforcementlearning ChujunZhangreceivedtheB.E.degreeincommu-
with double Q-learning,” in Proc. AAAI Conf. Artif. Intell. (AAAI), nication engineering from the School of Electrical
2016,pp.2094–2100. Engineering and Information, Southwest Petroleum
[163] H.V.Hasselt,“Doublelearning,”inProc.24thAnnu.Conf.NeuralInf. University, Chengdu, China, in 2023. He is cur-
Process.Syst.,2010,pp.2613–2616. rently pursuing the M.Sc. degree in information
[164] Z.Wang,T.Schaul,M.Hessel,H.Hasselt,M.Lanctot,andN.Freitas, andcommunicationengineeringwiththeCollegeof
“Dueling network architectures for deep reinforcement learning,” in Electronics and Information Engineering, Sichuan
Proc.Int.Conf.Mach.Learn.,2016,pp.1995–2003. University,Chengdu.
[165] R.S.Sutton,D.McAllester,S.Singh,andY.Mansour,“Policygradient His current research interests include wireless
methods for reinforcement learning with function approximation,” in communications, semantic communications, and
Proc.Adv.NeuralInf.Process.Syst.,vol.12,1999,pp.1–7. resourceallocation.
[166] D. Silver, G. Lever, N. Heess, T. Degris, D. Wierstra, and
M.Riedmiller,“Deterministicpolicygradientalgorithms,”inProc.Int.
Conf.Mach.Learn.(ICML),2014,pp.1–9.
[167] T. P. Lillicrap et al., “Continuous control with deep reinforcement Linyu Huang (Member, IEEE) received the B.E.
learning,”inProc.Int.Conf.Learn.Represent.(ICLR),2016,pp.1–10. degree in electronic information engineering from
[168] S.Fujimoto,H.Hoof,andD.Meger,“Addressingfunctionapproxima- theUniversityofElectronicScienceandTechnology
tion error in actor-critic methods,” in Proc. Int. Conf. Mach. Learn., of China, Chengdu, China, in 2008, and the Ph.D.
2018,pp.1587–1596. degree in electronic engineering from the City
[169] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, UniversityofHongKongin2014.Hejoinedthefac-
“Proximalpolicyoptimizationalgorithms,”2017,arXiv:1707.06347. ultywiththeCollegeofElectronicsandInformation
[170] T.Haarnoja,A.Zhou,P.Abbeel,andS.Levine,“Softactor-critic:Off- Engineering,SichuanUniversity,Chengdu,in2014.
policymaximumentropydeepreinforcementlearningwithastochastic Hiscurrentresearchinterestsincludewirelesscom-
actor,”inProc.Int.Conf.Mach.Learn.,2018,pp.1861–1870. munication,signalprocessing,andmachinelearning.
[171] S. Mirjalili and A. Lewis, “The whale optimization algorithm,” Adv.
Eng.Softw.,vol.95,pp.51–67,May2016.
[172] Y.Gu,W.Saad,M.Bennis,M.Debbah,andZ.Han,“Matchingtheory
for future wireless networks: Fundamentals and applications,” IEEE
Commun.Mag.,vol.53,no.5,pp.52–59,May2015. Qian Ning received the bachelor’s degree from
[173] C.Yuetal.,“ThesurprisingeffectivenessofPPOincooperativemulti- XidianUniversityin1990,themaster’sdegreefrom
agentgames,”inProc.Adv.NeuralInf.Process.Syst.,vol.35,2022, theUniversityofElectronicScienceandTechnology
pp.24611–24624. ofChinain1997,andthePh.D.degreefromSichuan
[174] S.Bayat,Y.Li,L.Song,andZ.Han,“Matchingtheory:Applicationsin University, Chengdu, China, in 2006, where she is
wirelesscommunications,”IEEESignalProcess.Mag.,vol.33,no.6, currently an Associate Professor with the College
pp.103–122,Nov.2016. of Electronics and Information Engineering. Her
[175] Y. Zhang, C. Lee, D. Niyato, and P. Wang, “Auction approaches for currentresearchinterestsincludeintelligentsystems
resource allocation in wireless systems: A survey,” IEEE Commun. andwirelessadhocnetworks.
SurveysTuts.,vol.15,no.3,pp.1020–1041,3rdQuart.,2013.
Authorized licensed use limited to: WUHAN UNIVERSITY OF TECHNOLOGY. Downloaded on February 09,2026 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.