1484 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 41, NO.

5, MAY 2023

Energy Efficient Semantic Communication Over

Wireless Networks With Rate Splitting
Zhaohui Yang , Member, IEEE, Mingzhe Chen , Member, IEEE, Zhaoyang Zhang , Senior Member, IEEE,
and Chongwen Huang , Member, IEEE

Abstract— In this paper, the problem of wireless resource Index Terms— Rate splitting multiple access, semantic commu-
allocation and semantic information extraction for energy effi- nication, energy efficient design.
cient semantic communications over wireless networks with rate
splitting is investigated. In the considered model, a base station I. I NTRODUCTION
(BS) first extracts semantic information from its large-scale data,
and then transmits the small-sized semantic information to each
user which recovers the original data based on its local common
knowledge. At the BS side, the probability graph is used to extract
T HE rapid development of emerging applications such
as digital twin, edge learning, and metaverse requires
wireless networks to support high transmission data rate,
multi-level semantic information. In the downlink transmission,
a rate splitting scheme is adopted, while the private small-sized ultra low latency, and seamless connectivity [1], [2], [3], [4].
semantic information is transmitted through private message However, due to limited wireless resources such as fre-
and the common knowledge is transmitted through common quency and time, conventional orthogonal multiple access
message. Due to limited wireless resource, both computation schemes cannot support massive connectivity concern for next-
energy and transmission energy are considered. This joint
computation and communication problem is formulated as an generation wireless communication networks [5]. Through
optimization problem aiming to minimize the total communica- using the same time or frequency resource, multiple users
tion and computation energy consumption of the network under can be served in non-orthogonal multiple access (NOMA)
computation, latency, and transmit power constraints. To solve [6], [7], [8], [9], [10], [11], where users can be split in the
this problem, an alternating algorithm is proposed where the power or code domain. Since additional users can be served
closed-form solutions for semantic information extraction ratio
and computation frequency are obtained at each step. Numerical with superposition coding at the transmitter and successive
results verify the effectiveness of the proposed algorithm. interference cancellation (SIC) at the receiver, the spectral
efficiency of NOMA is generally higher than conventional
Manuscript received 25 July 2022; revised 20 October 2022; accepted orthogonal multiple access schemes.
1 December 2022. Date of publication 30 January 2023; date of current version In downlink NOMA transmission, the receiver side decodes
18 April 2023. This work was supported in part by the National Natural
Science Foundation of China under Grant 61725104 and Grant U20A20158 the interference for all received strong messages [11]. Thus,
and in part by the National Key Research and Development Program of the computation capacity of NOMA decoding is generally
China under Grant 2018YFB1801104 and Grant 2020YFB1807101. The work high. To balance the decoding tradeoff of intended signal
of Chongwen Huang was supported in part by the China National Key
Research and Development Program under Grant 2021YFA1000500, in part and interference signal, the concept of rate splitting multiple
by the National Natural Science Foundation of China under Grant 62101492, access (RSMA) was introduced in [12], [13], [14] and [15].
in part by the Zhejiang Provincial Natural Science Foundation of China For downlink RSMA transmission, the transmission message
under Grant LR22F010002, in part by the National Natural Science Fund for
Excellent Young Scientists Fund Program (Overseas), in part by the Zhejiang intended for each user is divided into both common and
University Education Foundation Qizhen Scholar Foundation, and in part by private parts. All users intend to receive the common part
the Fundamental Research Funds for the Central Universities under Grant of the message, i.e., common message, while only part of
2021FZZX001-21. (Corresponding author: Zhaoyang Zhang.)
Zhaohui Yang is with the Zhejiang Laboratory, Hangzhou, Zhejiang the users wish to receive and decode the specific private
311121, China, also with the College of Information Science and Electronic part of the message, i.e., private message. At the user side,
Engineering, Zhejiang University, Hangzhou, Zhejiang 310027, China, and the common message is decoded first with regarding all
also with the Zhejiang Provincial Key Laboratory of Information Processing,
Communication and Networking (IPCAN), Hangzhou, Zhejiang 310007, private messages as interference, while the intended private
China (e-mail: [email protected]). message is decoded with only considering the private messages
Mingzhe Chen is with the Department of Electrical and Computer Engineer- intended for other users as interference. Through dynamically
ing and the Institute for Data Science and Computing, University of Miami,
Coral Gables, FL 33146 USA (e-mail: [email protected]). controlling the split of private and common messages, the
Zhaoyang Zhang and Chongwen Huang are with the College of Information computation complexity of RSMA can be adjusted to achieve
Science and Electronic Engineering, Zhejiang University, Hangzhou, Zhejiang the specific spectral efficiency requirements. To implement
310027, China, and also with the Zhejiang Provincial Key Laboratory
of Information Processing, Communication and Networking (IPCAN), RSMA for wireless communication systems, there are still
Hangzhou, Zhejiang 310007, China (e-mail: [email protected]; many challenges, which include the resource allocation for
[email protected]). private and common messages, decoding order optimization,
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/JSAC.2023.3240713. system design in imperfect channel and hardware mismatch
Digital Object Identifier 10.1109/JSAC.2023.3240713 cases.
0733-8716 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: ENERGY EFFICIENT SEMANTIC COMMUNICATION OVER WIRELESS NETWORKS WITH RATE SPLITTING 1485

There are many contributions investigating the problems to meet the multimodal quality-of-experience (QoE) require-
of RSMA in wireless communication systems. The general ments with massive data, which makes the traditional Shan-
challenges of RSMA were pointed out in [14] for multi- non capacity characterized transmission infeasible. Especially
ple input multiple output (MIMO) communication systems. in human-computer interaction scenarios, humans can con-
To maximize the sum rate of all users, a distributed rate trol multiple IoE devices simultaneously through voice and
splitting technique was proposed in [16]. For a two-receiver augmented/virtual reality commands, making communication
multiple input single output (MISO) communication system ubiquitous in small-range wireless networks, which poses
with limited rate feedback, the rate analysis was investigated severe challenges to traditional bit-oriented communication
in [17]. Compared with NOMA and space-division multi- challenge. Supporting real-time human-machine interaction
ple access (SDMA), it was shown in [18] that RSMA can and machine-to-machine interaction through the use of text,
achieve the best performance in terms of spectral and energy speech, images, and augmented/virtual reality is important
efficiency [19]. In particular, the energy efficiency optimiza- for future wireless communications. In order to support this
tion for RSMA and NOMA transmissions in a unmanned interaction, the important information finally received depends
aerial vehicle assisted wireless communication system was mainly on the intent, rather than the bit information depen-
investigated in [20]. Considering wireless energy transfer and dence of common sense. These applications use advanced
information transmission, the linear precoding method for signal processing to facilitate the development of task-oriented
RSMA was investigated in [21]. For the case with imperfect semantic communication [3], [46], [47]. In semantic commu-
channel state information, the sum rate maximization with nication, both transmitter and receiver share common knowl-
partial channel state information for RSMA was studied in edge, which can be used to extract small-size information
[22], while a downlink MISO RSMA system with bounded at the transmitter and recover the original information at
channel errors was investigated in [23]. the receiver [48]. Similarly, in downlink RSMA, all users
The interplay between rate splitting with emerging tech- also need to receive both common information and pri-
nologies has been investigated. With the help of reconfigurable vate information. Due to the inherent similarity of common
intelligent surface, the energy efficient resource allocation for knowledge and common message, RSMA can be utilized
reconfigurable intelligent surface assisted RSMA was investi- to enhance the system performance of downlink semantic
gated in [24], where the phase shift, rate allocation, and trasn- communication. To our best knowledge, there is no prior works
mit beamforming were jointly scheduled. The learning based that consider the integration of semantic communication and
traffic prediction method was studied in [25] for unmanned RSMA.
aerial vehicle enabled wireless communication system with The main contributions of this paper include:
rate splitting. The neural network was proposed in [26] to • The problem of wireless resource allocation and semantic
solve the user clustering problem in hierarchical rate splitting information extraction for energy efficient semantic com-
communication systems. Due to coupled rate and power allo- munications over wireless networks with rate splitting is
cation relationship, the resource allocation of RSMA usually investigated. In the considered model, the BS first extracts
leads to the nonconvex problem, which can be solved by the semantic information from its large-scale data, and
utilizing the learning techniques such as deep reinforcement then transmits the small-sized semantic information to
learning. Several deep learning algorithms were designed each user which recovers the original data based on the
to solve various complex resource allocation problems for local common knowledge.
RSMA, which include total power minimization problem [27], • In the downlink transmission, the rate splitting scheme is
joint power control, beamforming design, and splitting opti- adopted, while the private small-sized semantic informa-
mization problem [28], [29], power allocation problem with tion is transmitted through private message and the com-
limited channel state information knowledge [30], [31], joint mon knowledge is transmitted through common message.
transmit power, user clustering, and resource block allocation Due to limited wireless resource, both computational
problem [32], joint passive precoding at the reconfigurable energy and transmission energy must be considered.
intelligent surface and active precoding at the transmitter [33]. This joint computation and communication problem is
In the federated learning frameworks [34], the authors in [35] formulated as an optimization problem whose goal is to
utilized RSMA for uplink model transmission to minimize the minimize the total energy consumption of the network
total delay of the whole system. A model-based deep learning under a latency constraint.
algorithm was developed to solve the receiver design problem • To solve this problem, an iterative algorithm is proposed
of RSMA in [36]. where, at every step, closed-form solutions for semantic
Recently, semantic communication has attracted a lot of information extraction ratio and computation frequency
attention [37], [38], [39], [40], [41], [42], [43], [44], [45]. For are derived. Numerical results show the effectiveness of
the wireless communication system characterized by Shannon the proposed algorithm.
capacity, the receiver side needs to recover the information that The rest of this paper is organized as follows. The system
is exactly the same as the transmitted information. However, model and problem formulation are described in Section II.
in the emerging wireless applications such as virtual reality, The algorithm design is presented in Section III. Simulation
personalized healthcare, autonomous driving, and the Internet- results are analyzed in Section IV. Conclusions are drawn in
of-Everything (IoE), the wireless communication systems aim Section V.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 1486 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 41, NO. 5, MAY 2023

Fig. 2. An example of the multi-level semantic information extraction.

Fig. 1. Illustration of the considered SWC network with rate splitting.

II. S YSTEM M ODEL AND P ROBLEM F ORMULATION

Consider a downlink semantic wireless communication
(SWC) network with one multiple-antenna BS and K single-
antenna users, as shown in Fig. 1. The BS is equipped with
N antennas and the set of users is denoted by K. Each user k
has a large-sized data Dk to receive. Due to limited wireless
resource, the BS needs to extract the small-sized semantic
information from the original data Dk . In the considered
model, the BS first extracts the semantic information based on
directional probability graph and then transmits the semantic
information via rate splitting technique.

A. Semantic Communication Model Fig. 3. Illustration of the SWC model.

In this part, we utilize the directional probability graph
to characterize the inherent information of the transmitted semantic information and the output is denoted by G(Dk ).
information. In the directional probability graph, each vertex To efficiently transmit information, in the second part, a subset
represents the semantic entity with different semantic levels. Sk out of G(Dk ) is selected at user k, which is used for data
The higher level the semantic level is, the more complicated transmission.
the semantic information is. The link between any two vertexes At the user side, each user utilizes the shared common direc-
represents the probability. tional probability graph to recover the original data and the
To construct the directional probability graph, we use the recovered data is denoted by R(Sk ). The semantic accuracy
deep neural network to train the stored dataset, which includes of the recovered data
P|R(Sk )|
three main steps. In the first step, the semantic entity is min{σ(R(Sk ), s′ki ), σ(Dk , s′ki )}
uk (Dk , Sk ) = i=1 P|R(S )| ,
recognized from the dataset, where the semantic entity means k
σ(R(Sk ), s′ki )
i=1
the names in text, including person names, place names, etc. (1)
The name of semantic entity is highly open (various types,
flexible lengths, unregistered words), contains rich knowledge where |R(Sk )| is the number of bits in R(Sk ), s′ki denotes
and highlights individuality. Three common methods, i.e., rule the i-th word in text or frame in video in R(Sk ), and
method, taxonomy method, and sequence labeling [49] can be σ(R(Sk ), s′ki ) is the number of occurrences of s′ki in R(Sk ).
used to identify the semantic entity. The semantic entity is
presented as a vertex in the directional probability graph. In the B. RSMA Model
second step, the link between any two vertexes means the In RSMA, the message intended for each user can be split
probability that one vertex can be linked with the other vertex. into two parts, i.e., common part and private part [50]. The
Through training the dataset, the probability between two common parts from all users are collected and combined into
vertexes can be calculated via convolutional neural networks. a common message. Through sharing the same codebook for
In the third step, the semantic information fusion is conducted. all users, the common message is encoded into the common
For two vertexes, if the link probabilities between these two message s0 , which all users need to decode. The private part
vertexes are higher than a predefined threshold. As a result, the of each user k is encoded into the private stream sk , which is
final directional probability graph becomes a multi-tier graph, intended for the specific user k. As a result, the transmitted
as shown in Fig. 2. signal x of the BS can be written as:
To obtain the small-size semantic communication, the K
extraction process includes two parts, as shown in Fig. 3. In the √ X√
x= p0 w0 s0 + pk wk sk , (2)
first part, the directional probability graph is used to extract k=1

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: ENERGY EFFICIENT SEMANTIC COMMUNICATION OVER WIRELESS NETWORKS WITH RATE SPLITTING 1487

where w0 is the transmit beamforming of the common mes-

sage s0 , wk is the transmit beamforming of the private
message sk intended for user k, p0 is the transmit power of
the common message s0 , and pk is the transmit power of the
private message sk .
For user k, the received message can be represented by:
K
H √ X √
hk x + nk = hH
k p0 w0 s0 + pk hH
k w k sj + nk , (3)
j=1

where hk stands for the channel between user k and the BS.
To decode the common message s0 , the rate of user k can be
Fig. 4. Illustration of the computation and communication time.
given by:
!
p0 |hHk w0 |
2
With private rate (7) and allocated common rate ak , the
ck = B log2 1 + PK H
. (4)
2
j=1 pj |hk w j | + σ
2 downlink transmission time for transmitting Sk is given by
where B is the bandwidth of the BS. Note that all users need Z(Sk )
t2k1 = , (10)
to decode the same common message. To ensure that all users rk + ak
can successfully decode the common message, the rate of the where Z(Sk ) is the data size of set Sk . To transmit the
common message can be set as [18] renewed information about the knowledge base, i.e., updated
c0 = min ck . (5) information of directional probability graph, the transmission
k∈K time of all users can be formulated as
In our considered SWC with rate splitting, the common K0
knowledge is shared by all users. Thus, the common knowl- t0 = , (11)
a0
edge required for semantic communication can be encoded
in the common message. Besides, the common message also where K0 is the size of updated information of directional
includes the parts that are allocated for different users, i.e., the probability graph. Combining (10) and (11), the downlink
rate in the common message allocated to user k is denoted by transmission time for user k is
ak . As a result, the rate constraint for the common message t2k = max{t2k1 , t0 }. (12)
can be given by
K
The transmission energy for sending Sk is
X
a0 + ak ≤ ck , ∀k ∈ K, (6) E2k = t2k1 pk , (13)
k=1
where a0 is the rate allocated to updated common knowledge and the transmission energy for broadcasting updated infor-
that all users need to receive. In SWC, a0 represents the rate of mation of directional probability graph is
transmitting the information of updated directional probability E20 = t0 p0 . (14)
graph.
For each user, the common message is decoded first, and At user k, to recover the original data, the user needs to
then the common message can be subtracted for decoding the compute the semantic information Sk . The computation time
private message. As a result, the rate for decoding the private of user k
message for user k can be calculated as y2k (Dk , Sk )
t3k = , (15)
gk
 
pk |hH
k wk |
2
rk = B log2 1 + K P
p |hH w |2 +σ 2
. (7)
j=1,j̸=k j k j where y2k (Dk , Sk ) is the number of computation cycles of
recovering Dk from Sk and gk is the computation capac-
C. Transmission and Computation Model ity at user k. The total complete time for user k includes
For each user k, the computation time for extracting seman- computation time at the BS, downlink transmission time, and
tic information from data Dk is computation time at user k, as shown in Fig. 4. The overall
complete time of user k including both computation and
y1k (Dk , Sk )
t1k = , (8) computation is
fk
where y1k (Dk , Sk ) is the required amount of CPU cycles for tk = t1k + t2k + t3k  
calculating Sk out of Dk , and fk is the computing capacity of y1k (Dk , Sk ) Z(Sk ) K0 y2k (Dk , Sk )
= + max , + .
user k. The local computation energy can be given by: fk rk + ak a0 gk
(16)
E1k = κy1k (Dk , Sk )fk2 , (9)
The energy consumption at user k is
where κ is a constant coefficient to measure the effective
switched capacitance. E3k = κy2k (Dk , Sk )gk2 . (17)

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 1488 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 41, NO. 5, MAY 2023

With the above considered model, the total communication

and computation energy consumption of the system is
K
X
E= (E1k + E2k + E3k ) + E0
k=1
K  
X Z(Sk )pk
= κy1k (Dk , Sk )fk2 + + κy2k (Dk , Sk )gk2
rk + ak
k=1
K0 p0 Fig. 5. Illustration of the accuracy and computation functions versus the
+ . (18) extraction rate.
a0
We aim to minimize the total energy consumption of the A. Semantic Information Extraction
whole system with considering the completion time, transmit
With given computation capacity, power control, rate alloca-
information accuracy, computation capacity, rate allocation,
tion, and beamforming design, problem (19) can be simplified
and power allocation constraints. Mathematically, the formu-
as
lated total energy minimization problem can be given by:
K  
X Z(Sk )pk
min E, (19) min κy1k (Dk , Sk )fk2 + + κy2k (Dk , Sk )gk2
S,f ,g,p,a,w S rk + ak
  k=1
y1k (Dk , Sk ) Z(Sk ) K0 K0 p0
s.t. + max , + (20)
fk rk + ak a0 a0  
y2k (Dk , Sk ) y1k (Dk , Sk ) Z(Sk ) K0
+ ≤ T, ∀k ∈ K, (19a) s.t. + max ,
gk fk rk + ak a0
uk (Dk , Sk ) ≥ Ak , ∀k ∈ K, (19b) y2k (Dk , Sk )
+ ≤ T, ∀k ∈ K, (20a)
Sk ⊆ G(Dk ) ∀k ∈ K, (19c) gk
K
X uk (Dk , Sk ) ≥ Ak , ∀k ∈ K, (20b)
a0 + ak ≤ ck , ∀k ∈ K, (19d) Sk ⊆ G(Dk ) ∀k ∈ K. (20c)
k=1
K
X Problem (20) is hard to solve because of two general diffi-
max culties. The first difficulty lies in the discrete value space of
fk ≤ F (19e)
k=1 variable Sk , which leads to the discrete optimization problem
XK and the complexity to find the optimal solution is usually
p0 ≤ P max (19f) extremely too high. The second difficulty is the implicit
k=0 expressions of accuracy function uk (Dk , Sk ) and computation
ak , fk , pk ≥ 0, ∀k, (19g) functions f1k (Dk , Sk ) and f2k (Dk , Sk ).
∥wk ∥ = 1, ∀k ∈ K ∪ {0}, (19h) To handle the first difficulty, we introduce the new variable,
extraction rate ρk , which is defined as
0 ≤ gk ≤ gkmax , ∀k ∈ K, (19i)
Z(Sk )
ρk = . (21)
where S = {S1 , · · · , SK }, f = [f0 , f1 , · · · , fK ]T , g = Z(G(Dk ))
[g1 , · · · , gK ]T , p = [p1 , · · · , pK ]T , a = [a0 , · · · , pK ]T ,
Absolutely, the value of ρk lies in (0,1]. In the following,
w = [w0 ; w1 ; · · · ; wK ], T is the maximum communication
we use variable to replace Sk for the purpose of obtaining the
delay of the system, Ak is the minimum semantic accuracy
insights about extraction rate.
for user k, F max is the maximum computation capacity at
To handle the second difficulty, we first analyze the trend
the BS, P max is the transmission power of the BS, and
of accuracy and computation functions. For the accuracy
gkmax is the maximum local computation capacity of user
function, the accuracy always increases with the extraction
k. Since both objective function and constraints (19a)-(19c)
rate since more information can be used to recover the original
are nonconvex, it is generally hard to solve this problem.
data, as shown in Fig. 5. As a result, the minimum accuracy
To solve this problem, we propose an iterative algorithm using
constraint (20c) can be equivalent to
the alternating method and successive convex approximation
(SCA) approach. ρk ≥ Γk , (22)
where Γk is the minimum extraction rate satisfying
III. A LGORITHM D ESIGN
uk (Dk , Γk ) = Ak . For computation function y1k (Dk , ρk )
In this section, an alternating algorithm is proposed to is the number of required CPU cycles for computing the
iteratively solve problem (19) through optimizing three sub- information with extraction rate ρk out of Dk , y1k (Dk , Sk )
problems, i.e., semantic information extraction subproblem, includes two parts. The first part is computing the directional
computation capacity subproblem, joint power control, rate probability graph, which can be modeled as a function only
allocation and beamforming design subproblem. related to the size of Dk , i.e., y3k (Dk ). The second part is

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: ENERGY EFFICIENT SEMANTIC COMMUNICATION OVER WIRELESS NETWORKS WITH RATE SPLITTING 1489

selecting the information with extraction rate ρk out from the Ck4 (ρ∗k2 (λ1k2 )|1Γk )−Ck5
+ = T, (28)
directional probability graph. Since ρk = 0 or ρk = 1, the gk
selection scheme is straightforward, which leads to the lowest with a|cb = min{max{a, b}, c}.
computation cycles. Hence, the computation of the second Proof: Denoting λ11 , · · · , λ1K > 0 as the Lagrange
part first increases and then decreases with the extraction rate multiplier variables associated with constraint (25a), we obtain
ρk . As an example, computation function y1k (Dk , ρk ) can be the Lagrange function of problem (25) as
expressed as
K 
y1k (Dk , ρk ) = y3k (Dk ) + Ck1 (ρk − Ck2 )Ck3 ,
X
(23) L1 (ρ, λ1 ) = κfk2 (y3k (Dk ) + Ck1 (ρk − Ck2 )Ck3 )
where Ck1 > 0, Ck2 ∈ (0, 1), and Ck3 > 0 are constant k=1

parameters and theses parameters can be obtained through sim- Z(G(Dk ))pk ρk  K p
0 0
+ + κCk4 ρ−C
k
k5 2
gk +
ulations. For computation function y2k (Dk , ρk ), the number rk + ak a0
of computation cycles decreases with ρk since more semantic K  y (D ) + C (ρ − C )Ck3
3k k k1 k k2
X
information can be helpful in recovering the original infor- + λ1k
fk
mation. As an example, the computation function y2k (Dk , ρk ) k=1
 
can be expressed as Z(G(Dk ))ρk K0
+ max ,
rk + ak a0
y2k (Dk , ρk ) = Ck4 ρ−C
k
k5
, (24) −Ck5
Ck4 ρk 
where Ck4 > 0 and Ck5 > 0 are constant parameters through + −T , (29)
gk
simulations.
With the above variable substitution (21) and expressions where λ1 = [λ11 , · · · , λ1K ]T . The first derivative of (29)
(22)-(24), problem (20) can be reformulated as: becomes
K  ∂L1 (ρ, λ1 )
X = κfk2 Ck1 Ck3 (ρk − Ck2 )Ck3 −1
min κfk2 (y3k (Dk ) + Ck1 (ρk − Ck2 )Ck3 ) ∂ρk
ρ
k=1 Z(G(Dk ))pk
Z(G(Dk ))pk ρk  + − κCk4 Ck5 ρ−C
k
k5 −1 2
gk
+ + κCk4 ρ−C k5 2
gk rk + ak
k
rk + ak  C C (ρ − C )Ck3−1
k1 k3 k k2 Z(G(Dk ))
K0 p0 +λ1k +
+ (25) fk rk + ak
a0 −Ck5 −1 
Ck4 Ck5 ρk
y3k (Dk ) + Ck1 (ρk − Ck2 )Ck3 − (30)
s.t. gk
fk

Z(G(Dk ))ρk K0

Ck4 ρ−Ck5 for
k
+ max , + ≤ T, K0 (ak + rk )
rk + ak a0 gk ρk ≥ , (31)
∀k ∈ K, (25a) a0 Z(G(Dk ))
Γk ≤ ρk ≤ 1, ∀k ∈ K, (25b) and
where ρ = [ρ1 , · · · , ρK ]T . Since both objective function and ∂L1 (ρ, λ1 )
feasible set are convex, problem (25) is a convex problem. ∂ρk
Thus, we can apply the dual method to obtain the Karush- = κfk2 Ck1 Ck3 (ρk − Ck2 )Ck3 −1
Kuhn-Tucker (KKT) point. To calculate the solution of prob- Z(G(Dk ))pk
lem (25), we can obtain the following theorem. + − κCk4 Ck5 ρ−C
k
k5 −1 2
gk
rk + ak
Theorem 1: The optimal solution of problem (25) is  C C (ρ − C )Ck3−1 Ck4 Ck5 ρ−Ck5 −1 
k1 k3 k k2 k

K0 (ak + rk ) +λ1k −
 ρ∗k1 (λ1k1 ) if ρ∗k1 (λ11 ) ≥

 fk gk
∗ a0 Z(G(Dk )) (32)
ρk = K , (26)
∗ ∗ 0 (ak + rk )
 ρk2 (λ1k2 ) if ρk2 (λ12 ) <

for

a0 Z(G(Dk ))
where ρ∗k1 (λ1k ) and ρ∗k2 (λ1k ) are respectively the solutions to K0 (ak + rk )
ρk < , (33)
∂L1 (ρ,λ1k ) a0 Z(G(Dk ))
∂ρk = 0 in (30) and (32), λ1k1 and λ1k2 respectively
satisfy Denote the solution of ∂L1∂ρ (ρ,λ1 )
k
= 0 to equations (30) and
y3k (Dk ) + Ck1 (ρ∗k1 (λ1k1 )|1Γk − Ck2 )Ck3 Z(G(Dk )) (32) by ρ∗k1 (λ1 ) and ρ∗k2 (λ1k ), respectively. Note that the left
+ hand sides of (30) and (32) are monotonically increasing with
fk rk + ak
respect to ρk , solutions ρ∗k1 (λ1k ) and ρ∗k2 (λ1k ) can be obtained
Ck4 (ρ∗k1 (λ1k1 )|1Γk )−Ck5 via the bisection method. Considering constraints (25b), (31),
+ = T, (27)
gk and (33), the Lagrange multiplier should meet the KKT con-
y3k (Dk ) + Ck1 (ρ∗k2 (λ1k2 )|1Γk − Ck2 )Ck3 K0 dition, i.e., the optimal solution of problem can be presented
+ in (26).
fk a0

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 1490 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 41, NO. 5, MAY 2023

B. Optimal Computation Capacity the gradient method. In the t-th iteration, the value of λ2k and
With given semantic information extraction, power control, λ3 are updated by
"
rate allocation, and beamforming design, problem (19) can be  y (D , S )
1k k k
simplified as λ2k (t) = λ2k (t − 1) − υ(t)
fk
K   +
   #
X Z(Sk )pk Z(Sk ) K0 y2k (Dk , Sk )
min κy1k (Dk , Sk )fk2 + + κy2k (Dk , Sk )gk2 + max , + −T ,
f ,g rk + ak rk + ak a0 gk
k=1
K0 p0 (40)
+ (34)
a0   and
y1k (Dk , Sk ) Z(Sk ) K0 " K
!#+
s.t. + max ,
fk rk + ak a0
X
max
λ3 (t) = λ3 (t − 1) − υ(t) fk −F , (41)
y2k (Dk , Sk ) k=1
+ ≤ T, ∀k ∈ K, (34a)
gk where [a]+ = max a, 0 and υ(t) > 0 is the dynamic step size.
K
X Through iteratively updating (fk , gk ) and (λ2k , λ3 ), the overall
fk ≤ F max (34b) procedure yields the global optimal solution of problem (34).
k=1
fk ≥ 0, ∀k, (34c) C. Joint Power Control, Rate Allocation, and Beamforming
Design
0 ≤ gk ≤ gkmax , ∀k ∈ K. (34d)
With given semantic information extraction and computa-
The Language function of problem (34) can be given by tion capacity, problem (19) can be simplified as
K  
L2 (f , g, λ2 , λ3 )
X Z(Sk )pk
min κy1k (Dk , Sk )fk2 + + κy2k (Dk , Sk )gk2
K  p,a,w rk + ak
X Z(Sk )pk k=1
= κy1k (Dk , Sk )fk2 + K0 p 0
rk + ak + , (42)
k=1 a0
 K p
0 0
 
+κy2k (Dk , Sk )gk2 + y1k (Dk , Sk ) Z(Sk ) K0
a0 s.t. + max ,
fk rk + ak a0
K  y (D , S )  
X 1k k k Z(Sk ) K0 y2k (Dk , Sk )
+ λ2k + max , + ≤ T, ∀k ∈ K, (42a)
fk rk + ak a0 gk
k=1
K
! K
y2k (Dk , Sk )  X
ak ≤ ck , ∀k ∈ K,
X
+ − T + λ3 fk −F max , (35) a0 + (42b)
gk k=1
k=1
K
X
T
where λ2 = [λ21 , · · · , λ2K ] is the Language multiplier p0 ≤ P max (42c)
associated with constraint (34a) and λ3 > 0 is the Language k=0
multiplier associated with constraint (34b). The first derivative a0 , ak , pk ≥ 0, ∀k, (42d)
of (35) becomes ∥wk ∥ = 1, ∀k ∈ K ∪ {0}, (42e)
∂L2 (f , g, λ2 , λ3 ) Problem (42) is nonconvex owing to the nonconvex objec-
∂fk tive function and constraints (42a), (42b) and (42e). To handle
λ2k y1k (Dk , Sk ) the nonconvexity of the objective function, we introduce new
= 2κy1k (Dk , Sk )gk − + λ3 (36) variable rk and use variable p2k to replace power pk . Thus,
gk2
∂L2 (f , g, λ2 , λ3 ) problem (42) can be equivalently transformed to
K 
∂gk Z(Sk )p2k
X 
λ2k y2k (Dk , Sk ) min κy1k (Dk , Sk )fk2 + + κy2k (Dk , Sk )gk2
= 2κy2k (Dk , Sk )fk − (37) p,a,r,w rk + ak
k=1
fk2
K0 p20
∂L2 (f ,g,λ2 ,λ3 ) ∂L2 (f ,g,λ2 ,λ3 )
+ , (43)
Setting = 0 and = 0 yields a0
∂fk ∂gk  
y1k (Dk , Sk ) Z(Sk ) K0
s.t. + max ,
2κy1k (Dk , Sk )fk3 + λ3 fk2 − λ2k y1k (Dk , Sk ) = 0, (38) fk rk + ak a0
 1 y2k (Dk , Sk )
λ2k y2k (Dk , Sk ) 3 + ≤ T, ∀k ∈ K, (43a)
gk = . (39) gk
2κy2k (Dk , Sk ) !
K
X p20 |hH
k w0 |
2
The value of fk can be obtained via solving the cubic function a0 + ak ≤ B log2 1+ PK ,
2 H 2 2
in (38). Having obtained the value of computation capacity fk k=1 j=1 pj |hk w j | + σ
and gk , the value of Language multiplier can be updated via ∀k ∈ K, (43b)

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: ENERGY EFFICIENT SEMANTIC COMMUNICATION OVER WIRELESS NETWORKS WITH RATE SPLITTING 1491

!
p2k |hH
k wk |
2
constraint (47) can be approximated by
rk ≤ B log2 1 + PK ,
2 H 2 σ2
j=1,j̸=k pj |hk w j | +
R(hH
kqwk )
∀k ∈ K, (43c) (n) (n)
q
(n)
γk αk γk
a0 , ak , pk ≥ 0, ∀k, (43d) ≥ +
(n)
(αk − αk )
(n)
q
∥wk ∥ ≤ 1, ∀k ∈ K ∪ {0}, (43e) pk (n)
2pk αk
(n)
q q
(n) (n) (n)
where r = [r0 , r1 , · · · , rK ]T , the objective function is convex, αk (n)
γk α k (n)
and constraint (43e) is replaced by the inequality without + q (γk − γk ) − (n) 2
(pk − pk ),
(n) (n) (pk )
loss of generality. In problem (43), we only need to deal 2pk γk
with the nonconvexity of constraints (43b) and (43c). Through (48)
introducing slacking variables γk and ηk , problem (43) can be
reformulated as: where the superscript (n) means the value of the variable
in the n-th iteration. Moreover, (46) can be reformulated
K 
X Z(Sk )p2k as
min κy1k (Dk , Sk )fk2 +
p,a,r,w,γ,η rk + ak K
k=1 X 1 2
 K p2 ((p + |hH 2 2 2 H 2 2
k w j | ) − (pj − |hk w j | ) )
0 0 4 j
+ κy2k (Dk , Sk )gk2 + , (44) j=1,j̸=k
a0 K
 
y1k (Dk , Sk ) Z(Sk ) K0
X
s.t. + max , = p2j |hH 2 2
k w j | + σ ≤ αk . (49)
fk rk + ak a0 j=1,j̸=k
y2k (Dk , Sk )
+ ≤ T, ∀k ∈ K, (44a) Through replacing the left hand side of (49) with its first-order
gk
Taylor approximation, we can obtain
XK
K
a0 + ak ≤ B log2 (1 + ηk ) , ∀k ∈ K,

X 1 (n) (n) 2 2 (n) 2
k=1 ((pj )2 + |hHk w j | ) + 4((pj )
4
(44b) j=1,j̸=k
(n) (n) (n)
rk ≤ B log2 (1 + γk ) , ∀k ∈ K, (44c) +|hH 2 2 H 2 2
k w j | )pj (pj − pj ) − (pj − |hk w j | )

a0 , ak , pk ≥ 0, ∀k, (44d) (n) (n) 2 H (n) H H (n) 2
+4((pj )2 +|hH
k w j | )(R(hk w j hk w j )−|hk w j | )
∥wk ∥ ≤ 1, ∀k ∈ K ∪ {0}, (44e)
p2k |hH 2 +σ 2 ≤ αk . (50)
k wk |
PK ≥ γk , ∀k ∈ K,
2 H 2 + σ2
j=1,j̸=k pj |hk w j | Similarly, we can introduce slack variable βk and constraint
(44f) (44g) can be rewritten as:
p20 |hH
k w0 |
2 1 2
≥ ηk , ∀k ∈ K, (44g) ((p + |hH 2 2 2 H 2 2
k w 0 | ) − (p0 − |hk w 0 | ) )
PK 2 H 2 σ2 4 0
j=1 pj |hk w j | +
1
= p20 |hH 2 2 2
k w 0 | ≥ βk ηk = ((βk + ηk ) − (βk − ηk ) ),
where γ = [γ1 , · · · , γK ]T and η = [η1 , · · · , ηK ]T . 4
In problem (44), the objective function is transformed into (51)
K
convex. Because of nonconex constraints (44f) and (44g), X
p2j |hH 2 2
k w j | + σ ≤ βk . (52)
problem (44) is nonconvex. In the following, we uti-
j=1
lize the SCA method to handle these two nonconvex
constraints. Note that we cannot make hH k w 0 as real values for all k
For constraint (44f), it can be equivalent to through changing the phase of w0 . To handle the nonconvexity
of (52), we use first-order Taylor approximation on both
p2k |hH 2
k w k | ≥ γk αk , (45) sides of (52), which is different from the method in [51].
K
X Considering the first-order Taylor approximation on both sides,
p2j |hH 2 2
k w j | + σ ≤ αk , (46)
(52) can be transformed to
j=1,j̸=k
(n) (n) (n)
where αk is a nonnegative slack variable. In (45), we can ((p0 )2 + |hH 2 2
k w 0 | ) + 4((p0 )
2

always choose the term hH

k w k as a real value through chang- +|hH 2 (n) (n) 2 (n)
H 2 2
k w 0 | )p0 (p0 − p0 ) − (p0 − |hk w 0 | )
ing the phase of beamforming wk . Thus, constraint (45) can (n) (n) (n) (n)
+4((p0 )2 +|hH 2 H H H 2
k w 0 | )(R(hk w 0 hk w 0 )−|hk w 0 | )
be rewritten as (n) (n) (n) (n)
√ ≥ (βk + ηk )2 − (βk − ηk )(βk − ηk ) + (βk − ηk )2 ,
H γk αk
R(hk wk ) ≥ , (47) (53)
pk
where the left hand side is convex now. Through using the For constraint (52), we can use the similar method to handle
first-order Taylor series to replace the right hand side of (47), the nonconvexity of (46). Thus, (52) can be approximated

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 1492 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 41, NO. 5, MAY 2023

by Algorithm 1 Joint Communication and Computation Resource

Allocation for SWC with RSMA
K 
X 1 (n) (n) 2 2 (n) 2 (0)
1: Initialize S (0) , f , g (0) , p(0) , a(0) , w (0) . Set iteration
((pj )2 + |hH
k w j | ) + 4((pj )
j=1
4 number n = 1.
(n) (n) (n) 2: repeat
+|hH 2 2 H 2 2
k w j | )pj (pj − pj ) − (pj − |hk w j | ) 3: With given f (n−1) , g (n−1) , p(n−1) , a(n−1) , w(n−1) ,
(n) (n) (n)
+4((pj )2 + |hH 2 H H
k w j | )(R(hk w j hk w j ) solve the semantic information extraction subproblem
 and obtain the solution S (n) .
(n) 2
−|hH 2
k w j | ) + σ ≤ βk . (54) 4: With given S (n) , p(n−1) , a(n−1) , w(n−1) , solve the
computation capacity subproblem and obtain the solu-
With the above approximations, we can approximate the tion f (n) , g (n) .
nonconvex constraints (44f) and (44g) with the corresponding 5: With given S (n) , f (n) , g (n) , solve the joint power con-
convex approximation terms. Thus, the original problem (44) trol, rate allocation, and beamforming design subprob-
can be approximated by the following convex one: lem, of which the solution is p(n) , a(n) , w(n) .
K  6: Set n = n + 1.
X Z(Sk )p2k
min κy1k (Dk , Sk )fk2 + 7: until the objective value (19) converges.
p,a,r,w,γ,α,η,β rk + ak
k=1
K0 p20

+ κy2k (Dk , Sk )gk2 + , (55) TABLE I
a0 M AIN S YSTEM PARAMETERS
s.t. (43a) − (43e), (48), (50), (53), (54), (55a)
αk ≥ 0, βk ≥ 0, ∀k ∈ K, (55b)
where α = [α, · · · , α]T and β = [β1 , · · · , βK ]T . The
convex problem (55) can be solved by the existing convex
optimization toolbox.

power spectral density of the noise power is −174 dBm/Hz.

D. Algorithm Analysis Unless specified otherwise, we set maximum transmit power
The overall joint communication and computation resource P max = 30 dBm, the effective switched capacitance in local
allocation for SWC with RSMA is presented in Algorithm 1. computation is κ = 10−28 , maximum local computation
According to Algorithm 1, the complexity of solving problem capacity g1max = · · · = gK max
= 2 GHz. For the considered
(19) lies in solving three subproblems at each iteration. For semantic information task, we consider the same parameters
the semantic information extraction subproblem, the optimal as in [54]. The main system parameters are summarized in
solution is calculated by (26) in Theorem 1 with complexity Table I.
O(K log2 (1/ϵ1 )), where O(log2 (1/ϵ1 )) is the complexity of The proposed joint communication and computation
solving (27) and (28) with the bisection method of accuracy resource allocation for SWC with RSMA is labeled as
ϵ1 . For the computation capacity subproblem, the complexity ‘RSMA’. To compare the results of the proposed scheme,
is O(N1 K), where N1 denotes the number of iterations of we consider the conventional orthogonal multiple access,
using the dual method for solving the computation capacity frequency division multiple access (FDMA) [55], which is
subproblem. For the joint power control, rate allocation, and labeled as ‘FDMA’, the total energy minimization problem
beamforming design subproblem, the complexity lies in solv- for NOMA [56], which is labeled as ‘NOMA’. To better show
ing the approximated convex problem (55). The complexity the performance of multiple antenna scheme, we consider the
of obtaining the solution of problem (55) is O(M12 M2 ) [52], SDMA system as in [18].
where M1 = (N +7)K +N +2 is the total number of variables Fig. 6 illustrates that the total communication and com-
and M2 = 13K +1 is the total number of constraints. The total putation energy changes as the maximum transmit power of
complexity of solving the joint power control, rate allocation, each user varies. According to this figure, the EXH-RSMA
and beamforming design subproblem is O(N2 N 2 K 3 ), where scheme stands for the exhaustive search method, which can
N2 is the number of iterations for the SCA method. As a yield a near globally optimal solution through running the
result, the total complexity of the proposed Algorithm 1 is proposed algorithm with 1000 initial solutions. It can be
O(N3 K log2 (1/ϵ1 ) + N1 N3 K + N2 N3 N 2 K 3 ), where N3 is shown from this figure that the total energy decreases with
the number of outer iterations of Algorithm 1. the maximum transmit power of the BS. This is due to the
fact that large transmit power can lead to low transmit time,
IV. S IMULATION R ESULTS which allows more time for computation and yields low total
In the simulations, there are K = 5 users in the considered energy consumption. It is observed that the proposed RSMA
area. For the pathloss model between each user and the BS, outperforms FDMA, NOMA, since RSMA can achieve higher
we set 128.1 + 37.6 log10 d (d is in km) [53] and the standard spectral efficiency than FDMA and NOMA. Compared to
deviation of shadow fading is 4 dB [51]. Furthermore, the SDMA, RSMA can still achieve better energy consumption,
total bandwidth of the system is B = 20 MHz and the in particular the maximum transmit power is high. The reason

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: ENERGY EFFICIENT SEMANTIC COMMUNICATION OVER WIRELESS NETWORKS WITH RATE SPLITTING 1493

Fig. 6. Total communication and computation energy vs. maximum transmit

power. Fig. 8. Total communication and computation energy vs. transmit data size
of each user.

Fig. 7. Total communication and computation energy vs. bandwidth of the

system.
Fig. 9. Total communication and computation energy vs. maximum compu-
tation capacity of each user.
is that SDMA is more likely to serve the users with higher
channel gains, while the users with poor channel gains tend
to have long transmit time and high computation power is computation energy versus the maximum computation capac-
needed for task computation, thus leading to higher total ity of each user. According to this figure, the total energy first
energy consumption than RSMA. It can be also found that decreases rapidly and then the total energy tends to approach
the proposed RSMA achieves near performance as the EXH- a fixed value. The reason lies in that for small computa-
RSMA, which indicates the effectiveness of the proposed tion capacity region, the increase of maximum computation
scheme. capacity can greatly decrease the computation time and more
Fig. 7 shows the total energy versus bandwidth of the time can be used for transmission, thus reducing the transmit
system. Based on this figure, the total communication and power and total energy. For high computation capacity region,
computation energy decreases as the bandwidth of the system each user has chosen its optimal computation capacity and the
increases for all schemes. This is because high bandwidth increase of maximum computation capacity does not affect the
decreases the transmit time between users and the BS, which computation capacity allocation result, thus leading to stable
allows long computation time and consequently reduces the energy consumption.
local computation energy consumption.
Fig. 8 illustrates the trend of total communication and V. C ONCLUSION
computation energy with the transmit data size of each user. In this paper, the problem of wireless resource alloca-
It is observed that the total energy increases as the data size for tion and semantic information extraction for energy efficient
all schemes. This is due to the fact that more information needs semantic communications over wireless networks with rate
to be transmitted, thus increasing the transmit and computation splitting is investigated. In the considered model, the BS
power. It can be found that the growing speed of total energy first extracts the semantic information from its large-scale
od the proposed RSMA is slower than that of NOMA and data, and then transmits the small-sized semantic information
FDMA, which shows the robustness of the RSMA. to each user which recovers the original data based on the
To show how the computation capacity affects the system local common knowledge. In the downlink transmission, the
performance, Fig. 9 presents the total communication and rate splitting scheme is adopted, while the private small-sized

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 1494 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 41, NO. 5, MAY 2023

semantic information is transmitted through private message [18] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting multiple access for
and the common knowledge is transmitted through common downlink communication systems: Bridging, generalizing, and outper-
forming SDMA and NOMA,” EURASIP J. Wireless Commun. Network.,
message. Due to limited wireless resource, both computational vol. 2018, no. 1, pp. 1–54, May 2018.
energy and transmission energy need to be considered. This [19] X. Li et al., “Rate-splitting multiple access-enabled security analysis
joint computation and communication problem is considered in cognitive satellite terrestrial networks,” IEEE Trans. Veh. Technol.,
vol. 71, no. 11, pp. 11756–11771, Nov. 2022.
as an optimization problem whose goal is to minimize the
[20] A. Rahmati, Y. Yapıcı, N. Rupasinghe, I. Guvenc, H. Dai, and
total energy consumption of the network under both task A. Bhuyany, “Energy efficiency of RSMA and NOMA in cellular-
completion and semantic accuracy constraints. An iterative connected mmWave UAV networks,” 2019, arXiv:1902.04721.
algorithm is presented to solve this problem, where at each [21] Y. Mao, B. Clerckx, and V. O. K. Li, “Rate-splitting for multi-
user multi-antenna wireless information and power transfer,” 2019,
step, the optimal solutions for semantic information extraction arXiv:1902.07851.
ratio and computation frequency are derived. Numerical results [22] H. Joudeh and B. Clerckx, “Sum-rate maximization for linearly precoded
show the effectiveness of the proposed algorithm. downlink multiuser MISO systems with partial CSIT: A rate-splitting
approach,” IEEE Trans. Commun., vol. 64, no. 11, pp. 4847–4861,
Nov. 2016.
R EFERENCES [23] H. Joudeh and B. Clerckx, “Robust transmission in downlink multiuser
MISO systems: A rate-splitting approach,” IEEE Trans. Signal Process.,
[1] Y. Mao, O. Dizdar, B. Clerckx, R. Schober, P. Popovski, and vol. 64, no. 23, pp. 6227–6242, Dec. 2016.
H. V. Poor, “Rate-splitting multiple access: Fundamentals, survey, and [24] Z. Yang, J. Shi, Z. Li, M. Chen, W. Xu, and M. Shikh-Bahaei, “Energy
future research trends,” IEEE Commun. Surveys Tuts., vol. 24, no. 4, efficient rate splitting multiple access (RSMA) with reconfigurable
pp. 2073–2126, 4th Quart., 2022. intelligent surface,” in Proc. IEEE Int. Conf. Commun. Workshops (ICC
[2] B. Clerckx et al., “A primer on rate-splitting multiple access: Tutorial, Workshops), Jun. 2020, pp. 1–6.
myths, and frequently asked questions,” 2022, arXiv:2209.00491. [25] L. Lu, Y. Hu, Y. Zhang, G. Jia, J. Nie, and M. Shikh-Bahaei, “Machine
[3] W. Xu, Z. Yang, D. W. K. Ng, M. Levorato, Y. C. Eldar, and M. Debbah, learning for predictive deployment of UAVs with rate splitting multiple
“Edge learning for B5G networks with distributed signal processing: access,” in Proc. IEEE Globecom Workshops (GC Wkshps), Dec. 2020,
Semantic communication, edge computing, and wireless sensing,” 2022, pp. 1–6.
arXiv:2206.00422. [26] R. Pereira et al., “User clustering for rate splitting using machine
[4] M. Chen et al., “Distributed learning in wireless networks: Recent learning,” 2022, arXiv:2205.11373.
progress and future challenges,” IEEE J. Sel. Areas Commun., vol. 39, [27] M. R. Camana, C. E. Garcia, and I. Koo, “Deep learning-assisted power
no. 12, pp. 3579–3605, Dec. 2021. minimization in underlay MISO-SWIPT systems based on rate-splitting
[5] W. Saad, M. Bennis, and M. Chen, “A vision of 6G wireless systems: multiple access,” IEEE Access, vol. 10, pp. 62137–62156, 2022.
Applications, trends, technologies, and open research problems,” IEEE [28] J. Liu, C.-H.-R. Lin, Y.-C. Hu, and P. K. Donta, “Joint beamforming,
Netw., vol. 34, no. 3, pp. 134–142, May 2019. power allocation, and splitting control for SWIPT-enabled IoT networks
[6] Y. Liu, Z. Qin, M. Elkashlan, Z. Ding, A. Nallanathan, and L. Hanzo, with deep reinforcement learning and game theory,” Sensors, vol. 22,
“Nonorthogonal multiple access for 5G and beyond,” Proc. IEEE, no. 6, p. 2328, Mar. 2022.
vol. 105, no. 12, pp. 2347–2381, Dec. 2017. [29] Z. Yang, M. Chen, W. Saad, W. Xu, and M. Shikh-Bahaei, “Sum-rate
[7] M. Vaezi, Z. Ding, and H. V. Poor, Multiple Access Techniques maximization of uplink rate splitting multiple access (RSMA) commu-
for 5G Wireless Networks and Beyond. Cham, Switzerland: Springer, nication,” IEEE Trans. Mobile Comput., vol. 21, no. 7, pp. 2596–2609,
2018. Jul. 2020.
[8] L. Dai, B. Wang, Y. Yuan, S. Han, and Z. Wang, “Non-orthogonal [30] N. Q. Hieu, D. T. Hoang, D. Niyato, and D. I. Kim, “Optimal power
multiple access for 5G: Solutions, challenges, opportunities, and future allocation for rate splitting communications with deep reinforcement
research trends,” IEEE Commun. Mag., vol. 53, no. 9, pp. 74–81, learning,” IEEE Wireless Commun. Lett., vol. 10, no. 12, pp. 2820–2823,
Sep. 2015. Dec. 2021.
[9] Z. Ding, X. Lei, G. K. Karagiannidis, R. Schober, J. Yuan, and [31] J. Huang, Y. Yang, L. Yin, D. He, and Q. Yan, “Deep reinforcement
V. Bhargava, “A survey on non-orthogonal multiple access for 5G learning-based power allocation for rate-splitting multiple access in 6G
networks: Research challenges and future trends,” IEEE J. Sel. Areas LEO satellite communication system,” IEEE Wireless Commun. Lett.,
Commun., vol. 35, no. 10, pp. 2181–2195, Oct. 2017. vol. 11, no. 10, pp. 2185–2189, Oct. 2022.
[10] K. Wang, Y. Liu, A. Nallanathan, and Z. Ding, “User association in non- [32] M. Z. Hassan, M. J. Hossain, J. Cheng, and V. C. M. Leung, “Joint
orthogonal multiple access networks,” in Proc. IEEE ICC, Kansas City, throughput-power optimization of fog-RAN using rate-splitting multiple
MO, USA, May 2018, pp. 1–6. access and reinforcement-learning based user clustering,” IEEE Trans.
[11] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and Veh. Technol., vol. 70, no. 8, pp. 8019–8036, Aug. 2021.
K. Higuchi, “Non-orthogonal multiple access (NOMA) for cellular [33] M. Wu et al., “Deep learning-based rate-splitting multiple access
future radio access,” in Proc. 77th IEEE VTC-Spring, Dresden, Germany, for reconfigurable intelligent surface-aided tera-hertz massive MIMO,”
Jun. 2013, pp. 1–5. 2022, arXiv:2209.08456.
[12] B. Rimoldi and R. Urbanke, “A rate-splitting approach to the Gaussian [34] M. Chen, Z. Yang, W. Saad, C. Yin, H. V. Poor, and S. Cui, “A joint
multiple-access channel,” IEEE Trans. Inf. Theory, vol. 42, no. 2, learning and communications framework for federated learning over
pp. 364–375, Mar. 1996. wireless networks,” IEEE Trans. Wireless Commun., vol. 20, no. 1,
[13] T. Han and K. Kobayashi, “A new achievable rate region for the inter- pp. 269–283, Jan. 2021.
ference channel,” IEEE Trans. Inf. Theory, vol. IT-27, no. 1, pp. 49–60, [35] S.-H. Park and H. Lee, “Completion time minimization of fog-RAN-
Jan. 1981. assisted federated learning with rate-splitting transmission,” IEEE Trans.
[14] B. Clerckx, H. Joudeh, C. Hao, M. Dai, and B. Rassouli, “Rate splitting Veh. Technol., vol. 71, no. 9, pp. 10209–10214, Sep. 2022.
for MIMO wireless networks: A promising PHY-layer strategy for LTE [36] R. Cerna Loli, O. Dizdar, B. Clerckx, and C. Ling, “Model-based
evolution,” IEEE Commun. Mag., vol. 54, no. 5, pp. 98–105, May 2016. deep learning receiver design for rate-splitting multiple access,” 2022,
[15] H. Liu, T. A. Tsiftsis, K. J. Kim, K. S. Kwak, and H. V. Poor, “Rate arXiv:2205.00849.
splitting for uplink NOMA with enhanced fairness and outage perfor- [37] Z. Weng and Z. Qin, “Semantic communication systems for speech
mance,” IEEE Trans. Wireless Commun., vol. 19, no. 7, pp. 4657–4670, transmission,” IEEE J. Sel. Areas Commun., vol. 39, no. 8,
Jul. 2020. pp. 2434–2444, Aug. 2021.
[16] J. Cao and E. M. Yeh, “Asymptotically optimal multiple-access commu- [38] Y. Wang et al., “Performance optimization for semantic communications:
nication via distributed rate splitting,” IEEE Trans. Inf. Theory, vol. 53, An attention-based reinforcement learning approach,” IEEE J. Sel. Areas
no. 1, pp. 304–319, Jan. 2007. Commun., vol. 40, no. 9, pp. 2598–2613, Sep. 2022.
[17] C. Hao, Y. Wu, and B. Clerckx, “Rate analysis of two-receiver MISO [39] H. Xie, Z. Qin, G. Y. Li, and B.-H. Juang, “Deep learning enabled
broadcast channel with finite rate feedback: A rate-splitting approach,” semantic communication systems,” IEEE Trans. Signal Process., vol. 69,
IEEE Trans. Commun., vol. 63, no. 9, pp. 3232–3246, Sep. 2015. pp. 2663–2675, 2021.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: ENERGY EFFICIENT SEMANTIC COMMUNICATION OVER WIRELESS NETWORKS WITH RATE SPLITTING 1495

[40] M. Chen, Y. Wang, and H. V. Poor, “Performance optimization for Mingzhe Chen (Member, IEEE) received the
wireless semantic communications over energy harvesting networks,” Ph.D. degree from the Beijing University of Posts
in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), and Telecommunications, Beijing, China, in 2019.
May 2022, pp. 8647–8651. From 2016 to 2019, he was a Visiting Researcher
[41] J. Kang et al., “Personalized saliency in task-oriented semantic com- at the Department of Electrical and Computer
munications: Image transmission and performance analysis,” 2022, Engineering, Virginia Tech. From 2019 to 2021,
arXiv:2209.12274. he was a Post-Doctoral Research Associate with the
[42] R. Kaewpuang et al., “Cooperative resource management in quantum Department of Electrical and Computer Engineering,
key distribution (QKD) networks for semantic communication,” 2022, Princeton University. In 2022, he worked as an AI
arXiv:2209.11957. Researcher at Ericsson Research, USA. Currently,
[43] S. Xie, Y. Wu, S. Ma, M. Ding, Y. Shi, and M. Tang, “Robust information he is an Assistant Professor with the Department
bottleneck for task-oriented communication with digital modulation,” of Electrical and Computer Engineering and Institute of Data Science and
2022, arXiv:2209.10382. Computing, University of Miami. His research interests include federated
[44] S. Wang et al., “Wireless deep video semantic transmission,” 2022, learning, reinforcement learning, virtual reality, unmanned aerial vehicles, and
arXiv:2205.13129. the Internet of Things. He has received from the IEEE Communication Society
[45] H. Zou, C. Zhang, S. Lasaulce, L. Saludjian, and V. Poor, “Goal-oriented two paper awards, including the 2021 Young Author Best Paper Award and
quantization: Analysis, design, and application to resource allocation,” the 2022 Fred W. Ellersick Prize Award, and three Conference Best Paper
2022, arXiv:2209.15347. Awards at IEEE ICC in 2020, IEEE GLOBECOM in 2020, and IEEE WCNC
[46] D. Gunduz et al., “Beyond transmitting bits: Context, semantics, and in 2021. He currently serves as an Associate Editor for IEEE W IRELESS
task-oriented communications,” 2022, arXiv:2207.09353. C OMMUNICATIONS L ETTERS, IEEE T RANSACTIONS ON G REEN C OMMU -
[47] Z. Qin, X. Tao, J. Lu, W. Tong, and G. Ye Li, “Semantic communica- NICATIONS AND N ETWORKING , and IEEE T RANSACTIONS ON M ACHINE
tions: Principles and challenges,” 2021, arXiv:2201.01389. L EARNING IN C OMMUNICATIONS AND N ETWORKING. Previously, he guest
[48] H. Tong, Z. Yang, S. Wang, Y. Hu, W. Saad, and C. Yin, “Federated edited a Special Issue on Distributed Learning over Wireless Edge Networks
learning based audio semantic communication over wireless networks,” for IEEE J OURNAL ON S ELECTED A REAS IN C OMMUNICATIONS.
in Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2021,
pp. 1–6.
[49] L. Deng and Y. Liu, Deep Learning in Natural Language Processing.
Cham, Switzerland: Springer, 2018.
[50] B. Clerckx et al., “Rate-splitting unifying SDMA, OMA, NOMA, Zhaoyang Zhang (Senior Member, IEEE) received
and Multicasting in MISO broadcast channel: A simple two-user rate the Ph.D. degree from Zhejiang University,
analysis,” IEEE Wireless Commun. Lett., vol. 9, no. 3, pp. 349–353, Hangzhou, China, in 1998.
Mar. 2020. He is currently a Qiushi Distinguished Professor
[51] Z. Yang, M. Chen, W. Saad, and M. Shikh-Bahaei, “Optimization of rate with Zhejiang University. His research interests
allocation and power control for rate splitting multiple access (RSMA),” include the fundamental aspects of wireless
IEEE Trans. Commun., vol. 69, no. 9, pp. 5988–6002, Sep. 2021. communications and networking, such as
information theory and coding theory, network
[52] M. S. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret, “Applications
signal processing and distributed learning, AI-
of second-order cone programming,” Linear Algebra Appl., vol. 284,
empowered communications and networking, and
nos. 1–3, pp. 193–228, Nov. 1998.
synergetic sensing, computing, and communication.
[53] M. S. Ali, E. Hossain, A. Al-Dweik, and D. I. Kim, “Downlink
He has coauthored more than 300 peer-reviewed international journals and
power allocation for CoMP-NOMA in multi-cell networks,” IEEE Trans.
conference papers, including eight conference best papers awarded by IEEE
Commun., vol. 66, no. 9, pp. 3982–3998, Sep. 2018.
ICC 2019 and IEEE GlobeCom 2020. He was awarded the National Natural
[54] W. Ammar et al., “Construction of the literature graph in semantic Science Fund for Distinguished Young Scholars by NSFC in 2017. He is
scholar,” 2018, arXiv:1805.02262. serving or has served as an Editor for IEEE T RANSACTIONS ON W IRELESS
[55] K. Seong, M. Mohseni, and J. M. Cioffi, “Optimal resource allocation C OMMUNICATIONS, IEEE T RANSACTIONS ON C OMMUNICATIONS, and
for OFDMA downlink systems,” in Proc. IEEE Int. Symp. Inf. Theory, IET Communications, and the General Chair, the TPC Co-Chair or the
Seattle, WA, USA, Jul. 2006, pp. 1394–1398. Symposium Co-Chair for PIMRC 2021 Workshop on Native AI Empowered
[56] Z. Yang, C. Pan, J. Hou, and M. Shikh-Bahaei, “Efficient resource allo- Wireless Networks, VTC-Spring 2017 Workshop on HMWC, WCSP
cation for mobile-edge computing networks with NOMA: Completion 2013/2018, and Globecom 2014 Wireless Communications Symposium.
time and energy minimization,” IEEE Trans. Commun., vol. 67, no. 11, He was also a Keynote Speaker for APCC 2018 and VTC-Fall 2017
pp. 7771–7784, Nov. 2019. Workshop NOMA.

Zhaohui Yang (Member, IEEE) received the Ph.D.

degree from Southeast University, Nanjing, China,
in 2018. From 2018 to 2020, he was a Post-Doctoral
Research Associate with the Center for Telecom- Chongwen Huang (Member, IEEE) received the
munications Research, Department of Informatics, B.Sc. degree from Nankai University in 2010, the
King’s College London, U.K. From 2020 to 2022, M.Sc. degree from the University of Electronic
he was a Research Fellow with the Department Science and Technology of China in 2013, and the
of Electronic and Electrical Engineering, University Ph.D. degree from the Singapore University of Tech-
College London, U.K. He is currently a ZJU Young nology and Design (SUTD) in 2019. From October
Professor with the College of Information Science 2019 to September 2020, he was a Post-Doctoral
and Electronic Engineering and the Zhejiang Key Researcher with SUTD. Since September 2020,
Laboratory of Information Processing Communication and Networking, Zhe- he has been with Zhejiang University as a Tenure-
jiang University, and also a Research Scientist with the Zhejiang Laboratory. Track Young Professor. His main research interests
His research interests include joint communication, sensing, and computation, include holographic MIMO surface/reconfigurable
federated learning, and semantic communication. He was the Co-Chair for intelligent surface, B5G/6G wireless communication, mmWave/THz com-
international workshops with more than ten times, including ICC, GLOBE- munications, and deep learning technologies for wireless communications.
COM, WCNC, PIMRC, and INFOCOM. He is an Associate Editor for the He was a recipient of IEEE Marconi Prize Paper Award in Wireless Communi-
IEEE C OMMUNICATIONS L ETTERS, IET Communications, and EURASIP cations and IEEE ComSoc Asia-Pacific Outstanding Young Researcher Award
Journal on Wireless Communications and Networking. He has served as a in 2021. He has been serving as an Editor for IEEE C OMMUNICATIONS
Guest Editor for several journals, including IEEE J OURNAL ON S ELECTED L ETTER, Signal Processing (Elsevier), and EURASIP Journal on Wireless
A REAS IN C OMMUNICATIONS. Communications and Networking and Physical Communication since 2021.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY GUWAHATI. Downloaded on May 14,2023 at 14:21:10 UTC from IEEE Xplore. Restrictions apply.