IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO.

4, APRIL 2023 4981

Computation Offloading and Resource Allocation in

Unmanned Aerial Vehicle Networks
Binghong Liu, Chenxi Liu , Senior Member, IEEE, and Mugen Peng , Fellow, IEEE

Abstract—Deploying unmanned aerial vehicles (UAVs) as mobile promising paradigm. Specifically, by bringing computing and
edge computing (MEC) servers has been attracting significant at- cache resources closer to terminals, MEC can extend the cloud
tention, since the UAVs’ inherent maneuverability and mobility can computing capability to the edge and as such reduce the trans-
further reduce the distance between the users and computational
functionalities. In the UAV-assisted MEC systems, the caching and mission latency. Nevertheless, there are also many problems of
offloading decision optimization, subject to the computation and the current edge computing system. For example, devices in
energy constraints of the UAVs, is one of the key design issues. internet-of-things (IoT) applications such as smart agriculture
Against this backdrop, in this paper, we propose a novel cloud-edge and forest fire monitoring are often far away from the core
framework to facilitate MEC in the UAV networks. Specifically, in network [5], thus the rigidity of ground base stations (BSs) and
our framework, the edge UAVs (EUAVs), together with the cloud,
provide caching and computing services for the terrestrial users. their susceptibility to infrastructure failure may prevent them
In order to minimize the weighted sum cost of latency and energy from satisfying the quality-of-service (QoS) requirements.
consumption, we jointly optimize the caching and offloading deci- Recently, the UAVs have been increasingly applied in the edge
sions, the EUAV deployment, the radio and computation resource computing system to open new opportunities for the next genera-
allocation, while simultaneously satisfying the UAVs’ cache and tion mobile network [6], [7], [8], [9], and the UAV-assisted MEC
computation capacity constraints, as well as the users’ latency and
energy consumption constraints. To solve this NP-hard problem, system has been attracting widespread research attention [10],
we propose a sequential convex programming (SCP) and sequential [11], [12], [13], [14], due to the flexible and maneuverable
quadratic programming (SQP) based deep Q-learning (SS-DQN) potential of UAVs to deliver prompt services [15], [16], [17].
algorithm. The proposed algorithm allows the system to adaptively Compared with conventional MEC systems, the UAV-assisted
adjust caching and offloading decisions with the EUAV deployment MEC system can offer more flexible and on-demand services,
scheme and the resource allocation scheme obtained by SCP and
SQP algorithms, respectively. Simulation results verify the supe- where the UAVs act as the MEC servers to supply faster re-
riority of our proposed algorithm compared to two benchmark sponse to the user requests via high-quality line of sight (LoS)
schemes, and demonstrate how the combination of DRL and convex communication links [18]. Specifically, in [10], a UAV was
optimization can improve the system performance significantly. deployed to facilitate the MEC service for the IoT devices, and
Index Terms—Unmanned aerial vehicles (UAVs), wireless the latency and energy consumption were minimized by jointly
caching, computation offloading. optimizing the task split ratio, UAV deployment and resource
allocation, where the task split ratio is defined as the ratio of the
I. INTRODUCTION offloaded bits to the total bits of the computation task. In [11],
HE demands for higher data rate and better network capa- the scenario was expanded into a swarm of the UAVs, in which
T bility have experienced an exponential growth over the last
few years. As forecasted in the Cisco report, by 2023, the number
each UAV had a time-varying task to be completed. With the
same goal as in [10], the transmission mode and the offloading
of networked devices will reach 29.3 billion, generating almost decisions were jointly optimized. In [12], the throughput of the
one zettabyte each year [1]. In order to meet users’ needs and UAV-assisted MEC systems was maximized by jointly planning
support a wide variety of fast-growing applications, which are the trajectory and scheduling the tasks under constraints of
generally delay-sensitive and computation-intensive, the mobile transmission rate and task atomicity. In [13], the ultra-reliable
edge computing (MEC) [2], [3], [4] has been presented as a low-latency computation offloading problem was studied in a
UAV-assisted IoT network to maximize the transmission rate.
Manuscript received 15 October 2022; accepted 14 November 2022. Date In [14], the 5G network slice extension problem was investigated
of publication 17 November 2022; date of current version 18 April 2023. This in a UAV-assisted edge computing system, where the system
work was supported in part by the National Key Research and Development controller was able to turn on and off the computing functions
Program of China under Grant 2021YFB2900200, in part by the National Natural
Science Foundation of China under Grants 62001047, 61925101, 61921003, of UAVs to serve users flexibly and save energy.
and 61831002, in part by the Young Elite Scientist Sponsorship Program by The aforementioned works [10], [11], [12], [13], [14] focused
China Institute of Communications, and in part by Beijing Municipal Science on single-tier UAV-assisted MEC networks. As such, the limited
and Technology under Grant Z211100004421017. The review of this article was
coordinated by Prof. Alfredo Luigi Grieco. (Corresponding author: Chenxi Liu.) computation capacity and energy storage of the UAVs [17]
The authors are with the State Key Laboratory of Networking and Switch- becomes a vital issue that hinder the real implementations of
ing Technology, Beijing University of Posts and Telecommunications, Beijing the UAV-assisted MEC networks. One possible solution to this
100876, China (e-mail: [email protected]; [email protected];
[email protected]). problem is to incorporate the cloud computing as a supplement
Digital Object Identifier 10.1109/TVT.2022.3222907 to the edge UAVs (EUAVs). In such multi-tier cloud-edge net-

0018-9545 © 2022 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: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
4982 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

works, the cloud center has powerful computing resource and to the constraints on the computation and caching capacity
huge cache capacity, to which the EUAVs can connect via the of the EUAVs, and the latency and energy consumption of
fronthaul links, thus making up the deficiency of UAV-assisted the users.
MEC to achieve cost-effective computation with latency satis- r We show how the formulated problem can be decomposed
faction. Despite its potential, little efforts has so far been devoted to three subproblems, namley, the EUAV deployment sub-
to the design and deployment of the UAV-assited cloud-edge problem, transmit power and computation capacity alloca-
computing systems [19], [20], [34]. The authors in [19] con- tion subproblem, and caching and offloading decision sub-
sidered a typical scenario, where one UAV was dispatched as problem. Sequential convex programming (SCP), sequen-
the MEC server to handle the computation tasks together with tial quadratic programming (SQP), and deep Q-learning
the cloud. The resource allocation and three-dimensional (3D) (DQN) algorithms are applied to solve these subproblems,
trajectory design were jointly optimized to minimize the energy respectively. Note that the proposed SCP and SQP-based
consumption. Similar to [19], the authors in [34] studied the DQN (SS-DQN) algorithm is a hybrid method that com-
energy consumption minimization problem by jointly optimiz- bines the benefits of both the convex optimization method
ing the computation resource scheduling, bandwidth allocation and the DRL method. Specifically, the convex optimization
and UAV trajectory. In [20], a UAV-assisted cloud-edge system is utilized to accelerate the convergence rate of the DRL
for virtual reality (VR) applications was considered, where the by reducing the large state-action space.
UAVs were deployed as aerial base stations (BSs) to cache, r We exam the impacts of key parameters, i.e. cache and com-
process and deliver VR contents from the cloud to users. The putation capacities, task sizes, on the cost performance,
user association, caching policy and locations of UAVs were demonstrating the superiority of our proposed algorithm
jointly optimized to minimize the latency. over various benchmark solutions. We further show that our
proposed algorithm works well under the stringent latency
constraints of the users, indicating the applicability of our
A. Motivations and Contributions proposed algorithm in practical UAV-assisted cloud-edge
Based on the above observations, although the UAV-assisted systems.
computing and caching have been separately discussed in previ-
ous studies [19], [20] and [34], few researches have comprehen- B. Organization
sively considered both of them. By avoiding the duplicate trans-
The rest of this paper is organized as follows. Section II
mission of popular file contents and handling the burst traffic
presents the system model of the considered UAV-assisted cloud-
timely, the introduction of caching mechanism into computing
edge system, analyzes the latency and energy consumption for
can effectively improve the resource reuse, reduce the latency,
executing tasks, and formulates the weighted sum cost mini-
and enhance the user experience. As such, how to exploit the
mization problem. Section III presents the proposed algorithm
relationship between caching and computing to schedule tasks
that solve the formulated problem. Numerical results and related
between the edge and the cloud, and further, to balance the
discussions are provided in Section IV. Finally, Section V draws
latency and energy consumption of the UAV-assisted cloud-edge
the conclusion. Unless otherwise specified, the notations used
system is worth to be studied. The reason is that the uneven user
throughout the paper are summarized in Table I.
requests and heterogeneity of the cloud-edge architecture pose a
challenge to the selection of cached data and offloading location.
II. SYSTEM MODEL AND PROBLEM FORMULATION
In addition, the aforementioned works mainly considered the use
of either convex optimization method or the deep reinforcement In this paper, we consider a UAV-assisted cloud-edge network,
learning (DRL) method. For example, works [10], [12], [13], which consists of a cloud layer, an edge layer, and a terminal
[19], [20], and [34] adopted the traditional convex optimization layer. In the edge layer, the UAVs with limited cache and
method, while works [11] and [14] utilized the DRL method. computing capacities are deployed as the MEC servers in the
However, the combination of these two methods can contribute vicinity of users. In the cloud layer, a tethered UAV (TUAV)
to a more efficient system. Therefore, in this paper, we exploit physically wired to a ground control center (GCC) of power
the caching and computing functions of the edge server and caching and computing capabilities is deployed. Both the energy
consider a UAV-assisted cloud-edge system, in which the EU- and data transmission between the TUAV and the GCC are
AVs, together with the cloud, provide caching and computation aggregated in the same tether [35], which not only enables
offloading services for the terrestrial users. We formulate the cost the TUAV to maintain longer flight time, but also ensures the
minimization problem and reveal its dependence on caching and reliable communication at higher data rate. While due to the
offloading decisions, resource allocation, as well as UAV deploy- weight of the fiber, the tether length is usually limited to a range
ment strategies. The contributions of this paper are summarized between 80 m and 150 m [36]. As such, we assume that the
as follows. transmission latency between the TUAV and GCC is negligible,
r We formulate a cost minimization problem for the con- due to the use of high-speed fiber link and the limited tether
sidered multi-tier UAV-assited cloud-edge framwork, in length, and for the sake of simplicity, use the TUAV to represent
which the balance between the latency and energy con- the GCC [37]. The set of EUAVs and ground users are denoted by
sumption can be achieved by jointly optimizing the caching N = {1, 2, . . ., N } and M = {1, 2, . . ., M }, respectively. The
and offloading decisions, the power and computation re- two-dimensional coordinate of EUAV n, TUAV and user m are
source allocations, as well as the UAVs deployment, subject un = [xn , yn ], u0 = [x0 , y0 ], and um = [xm , ym ], respectively.
Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: COMPUTATION OFFLOADING AND RESOURCE ALLOCATION IN UNMANNED AERIAL VEHICLE NETWORKS 4983

TABLE I
LIST OF NOTATIONS

the interference, the users and UAVs occupy orthogonal band-

width for transmission. The air-to-ground communication link
between user m and EUAV n is modeled as follows
m,n = ĥm,n Lm,n ,
hAtG (1)
where ĥm,n and Lm,n represent the small-scale and large-scale
fading, respectively. Specifically, ĥm,n is assumed to follow the
Rician fading [21], given by
  
Qm,n 1
ĥm,n = h̄m,n + h̃m,n , (2)
Qm,n + 1 Qm,n + 1
 
h̄m,n is the deterministic LoS channel component, and h̄m,n =1
h̃m,n is the random scattering component, which is a circularly
symmetric complex Gaussian (CSCG) random variable with
zero mean and unit variance, namely, h̃m,n ∼ CN (0, 1). Here
CSCG means h̃m,n has independently and identically distributed
real and imaginary parts, each follows N (0, 1/2) [38] [39].
Fig. 1. Illustration of the UAV-assisted cloud-edge system.
Qm,n denotes the Rician factor, Qm,n = A1 exp(A2 θm,n ),
where A1 and A2 are environment-dependent constants, and
The height of EUAV n and TUAV are fixed as hn and h0 , θm,n represents the elevation angle, θm,n = arcsin( dhn,m n
). In
respectively. (1), Lm,n can be expressed as
As shown in Fig. 1, the transmission can be divided into two 
phases, namely, the caching content delivery phase and the com- Lm,n = α0 (dm,n )−a1 , (3)
putation offloading phase. In the caching content delivery phase, where α 0 is the power gain at reference distance d0 = 1 m,
the users obtain the requested files from the the cache servers
dm,n = un − um 2 + h2n is the distance between user m
(i.e., the EUAVs or the TUAV). In the computation offloading
and EUAV n, and a1 is the pathloss exponent of the air-to-ground
phase, the users decide where to complete the computation tasks
link.
based on the obtained files in the caching content delivery phase.
Based on (1)–(3), the downlink transmission rate of EUAV n
Note that we assume that the computation can only be performed
for transmitting the requested cache file of user m is given by
after the files are obtained. This assumptions is practical in   AtG 2 
many scenarios. For the real-time gaming applications, the game q n,m
hm,n 
E,Down
Rn,m = Bm,n log2 1 + , (4)
files with higher request rate need to be cached closer to users 2
σm,n
to reduce the computational burden of the cloud and lower
the latency [24], and the user equipments perform subsequent
2
where σm,n = N0 Bm,n , Bm,n is the transmission bandwidth
computations based on these files. between user m and EUAV n, N0 is the noise power spectrum
density. qn,m is the transmit power of EUAV n to user m.
Similarly, after receiving the requested data, the uplink trans-
A. The Communication Model mission rate of user m for offloading its computation task to
We assume that the EUAVs can communicate with both the EUAV n is given by
  2 
TUAV and the terrestrial users, while the terrestrial users can pm,n hAtG
m,n

only communicate with the EUAVs, due to the long distance Rm,n = Bm,n log2 1 +
E,U p
2
, (5)
σm,n
between the TUAV and the users. In order to efficiently control

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
4984 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

where pm,n is the transmit power of user m to EUAV n. Therefore, the downlink transmission rate of TUAV delivering
Based on (4) and (5), the downlink latency for transmitting user m’s requested cache file to EUAV n is given by
  2 
requested cache file and the uplink latency for offloading com- pcloud hAtA
n

puting task between EUAV n and user m can be expressed C,Down
Rn,m = Bn log2 1 + , (11)
as L̄n,m = RE,Down
Om
and L̄m,n = RO σn2
E,U p , respectively. Due to
m
n,m m,n
the dynamic feature of the air-to-ground links, there exists a where σn2 = N0 Bn , Bn is the transmission bandwidth between
timeout probability that user m cannot successfully receive the TUAV and EUAV n, pcloud is the downlink transmit power of
requested cache files from the EUAV n. Mathematically, the TUAV.
timeout probability is given by Similarly, the uplink transmission rate of EUAV n offloading
  user m’s computation task to TUAV is given by
O m    
out
Pn,m = P Ln,m = E,Down < L̄n,m AV  AtA 2
pU
n hn
rn,m Rm,n = Bn log2 1 +
C,U p
, (12)
⎛ ⎞ σn2
 2
E,Down

  σ 2
m,n 2 rn,m /Bm,n
− 1 where pU AV
is the transmit power of EUAV n for offloading
= P ⎝ĥm,n  < ⎠ n
α0 d−a
m,n qn,m
1 user m’s computing task to TUAV. For the sake of simplicity,
we assume that the transmit power of all the EUAVs is fixed as
⎛ ⎞
2
σm,n 2rn,m
E,Down
/Bm,n
−1 puav .
= Γn,m ⎝ ⎠, (6)
α0 d−a
m,n qn,m
1
B. Caching and Computing Model
where Γn,m (·) is the cumulative distribution function (CDF) of The task of user m can be represented as Sm =
the random variable |ĥm,n |2 . (Cm , Om , Fm ), in which Cm and Om are the size of the re-
Denote δ as the maximum tolerable outage probability of the quested cache file and the related computation task, respectively,
air-to-ground links. In order to ensure the reliable communica- and Fm is the required CPU cycles for computing the task. Note
tion between the EUAVs and the users, rn,m E,Down
is chosen to that each EUAV is able to provide its cached data for users
out
satisfy the condition: Pn,m ≤ δ and we can obtain the outage- who request to compute the data. The un-cached files need to
aware achievable rate as be firstly obtained from the cloud through backhaul, and then
  transmitted to users. We denote αn,m as the caching decision
down
γn,m χn,m variable that indicates whether the file requested by user m is
E,Down
rn,m = Bm,n log2 1 + , (7)
dam,n
1
cached in EUAV n. Specifically, αn,m = 1 means that the file is
cached, otherwise, αn,m = 0.
where χn,m is the unique solution of Γn,m (·) = δ, and γn,m
down
=
qn,m α0 Since the computation capabilities of the terrestrial users are
σ 2 . To facilitate the analysis, we approximate χ n,m by the
m,n in general limited, we consider that the users can offload their
E,Down
log function [22], as such, rn,m can be further expressed as tasks to EUAVs or even the cloud via the EUAVs. Assuming
   
down
γn,m that the computation task has to be offloaded as a whole, we
E,Down . C2
rn,m = Bm,n log2 1+ C1+   , introduce the offloading decision variable βm,n E
∈ {0, 1} and
− C3+C4 dH n dam,n
1
1+e m,n
βm,n ∈ {0, 1} to indicate where to offload the tasks. Specifi-
C
(8) E
cally, βm,n = 1 means that user m choose to offload its com-
where C1 > 0, C2 > 0 and C1 + C2 = 1, C3 < 0 represents the putation task to EUAV n, and βm,n C
= 1 means that the task
positive logistic mid-point, and C4 > 0 represents the logistic is offloaded to the cloud. Note that in this paper, the binary
growth rate. offloading can be transformed to the partial offloading when
Similarly, the outage-aware achievable rate in the uplink from the packet segmentation is fine enough. Specifically, when we
user m to EUAV n can be expressed as divide the task, which is previously offloaded as a whole in
    the binary offloading mode, into L equal-size segments, the
up
E,U p . C2 γm,n
rm,n = Bm,n log2 1+ C1+   . offloading becomes the partial offloading in another sense, and
− C3+C4 dH n dam,n
1
1+e m,n L represents the segmentation index.
(9) In the following, we analyze the latency and energy consump-
up
where γm,n
p
= σm,n 0α
. tion of our considered network in detail.
2
m,n 1) Latency: Note that the users firstly need to obtain the
The air-to-air communication link from the EUAV to the
requested cache files from the EUAVs or the cloud. Based on the
TUAV is modeled as the quasi-static block fading LoS link,
obtained cache data, the users start the computation. There are
which is given by
 three modes that the users that can choose from for computing
hAtA
n = α0 d−a
n ,
2
(10) the tasks, namely, the local computing mode, the edge computing
 mode, and the cloud computing mode. We assume that the
where dn = un − u0 2 + (hn − h0 )2 is the distance be- transmission of cached files and the offloading of computation
tween EUAV n and TUAV, a2 is the pathloss exponent of the tasks are carried out by sequence and independently with each
air-to-air link. other. Then, the latency for transmitting the requested file to user

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: COMPUTATION OFFLOADING AND RESOURCE ALLOCATION IN UNMANNED AERIAL VEHICLE NETWORKS 4985

m from EUAV n is given by The energy consumption of the cloud computing mode is
Cm Cm given by
m,n = (1 − αn,m )
Lcache C,Down
+ E,Down . (13)
Om Om
Rn,m rn,m cloud
Em,n = Em,n
cache
+ E,U p · pm,n + C,U p · puav + Ec,m com
.
In the local computing mode, the latency consists of the trans- rm,n Rm,n
mission latency for delivering the cache file and the computation (21)
latency for processing tasks locally, which is given by com
Similarly, Ec,m is the energy consumption of the cloud cen-
Fm ter for computing the offloaded task of user m, Ec,m com
=
m,n = Lm,n +
Llocal cache
, (14) v
fc,m (e3 (fc,m ) + e4 ).
εFm
fm
where fm represents the local processing ability of user m. Then, the total energy consumption for executing tasks of user
In the edge computing mode, the latency consists of the trans- m is given by
mission latency for delivering the cache file, the transmission   local
total
Em,n = 1−βm,n E
−βm,n
C
Em,n +βm,n E edge
Em,n +βm,n
C cloud
Em,n .
latency for offloading user m’s task to EUAV n, and the edge
(22)
computing latency, which is given by
Note that the energy consumption of each EUAV n consists
Om Fm
m,n = Lm,n + E,U p +
Ledge cache
, (15) of the energy consumed by data transmission, task computation
rm,n f n,m and hovering. As referred in [40], [41], the energy consumed by
where fn,m represents the allocated edge computing resources hovering is much smaller than that consumed by propulsion [42].
of EUAV n to process the task offloaded by user m. Moreover, the energy consumed by data transmission is much
In the cloud computing mode, the latency consists of the trans- smaller than that consumed by computing and hovering [43].
mission latency for delivering the cache file, the transmission In this paper, we only consider the UAVs’ hovering situation in
latency for offloading user m’s task to EUAV n, and then to a short-time period. As such, the energy consumption of each
TUAV, and the cloud computing latency, which is given by EUAV is within its energy capacity.
Om Om Fm Based on the above analysis on latency and energy consump-
m,n = Lm,n + E,U p +
Lcloud +
cache
C,U p
, (16)
rm,n Rm,n fc,m tion, respectively, the entire cost of the UAV-assisted cloud-edge
where fc,m represents the allocated computing resources of the system is defined as follows
cloud center to process the task offloaded by user m. 
N 
M

In this paper, we mainly focus on the applications in which the W total = total
Wm,n
n=1 m=1
amount of uplink data is significantly larger than that of downlink
data. Therefore, as in the previous studies [10], [11], [13], [19], 
N 
M
 
[20], the latency for receiving the outcome of computation tasks = m,n + ωm,2 Em,n ,
ωm,1 Ltotal total
(23)
is ignored. As such, the total latency for executing tasks of user n=1 m=1
m is given by where ωm,1 and ωm,2 are the weight coefficients for balancing
  local the latency and the energy consumption. Note that the users with
m,n = 1−βm,n −βm,n Lm,n +βm,n Lm,n +βm,n Lm,n .
Ltotal E C E edge C cloud
different service types have various preference for latency (rep-
(17)
resented as QoS index 1) and energy consumption (represented
2) Energy Consumption: Next, we focus on the energy con- as QoS index 2). Assuming that each QoS index can be further
sumption of our considered framework. The energy consump- divided into K preference levels, in which a higher level means
tion for transmitting the requested cache file from EUAV n to that this index is more favored [25] and has more significant
user m, is given by effect on the users. The preference level of user m on QoS
Cm Cm index i, (i = 1, 2), can be represented as km i
∈ {1, 2, . . ., K}.
cache
Em,n = (1 − αm,n ) C,Down · pcloud + E,Down · qn,m .
Rn,m rn,m In addition, we introduce am = km /km to measure user m’s
i,j i j

(18) relative preference degree on the QoS index 1 compared with

Therefore, the energy consumption of the local computing the QoS index 2. Therefore, the weight coefficient can be further
mode is given by represented as follows
1
2
local
Em,n = Em,n
cache
+ εfm
2
Fm . (19) ai,j
ωm,i = m
, ∀i = {1, 2} . (24)
where ε is the effective switched capacitance determined by the 2 j=1 a1,j
m + am
2,j

specific chip architecture.

The energy consumption of the edge computing mode is given C. Problem Formulation
by
In this paper, the goal of our work is to minimize the
Om
edge
Em,n = Em,n
cache
+ E,U p · pm,n + En,mcom
. (20) total cost of the UAV-assisted cloud-edge system by jointly
rm,n optimizing the caching and offloading decision schemes α =
com
Note that En,m is the energy consumption of EUAV n {αn,m |∀m, ∀n} and β = {βm,n
E C
, βm,n |∀m, ∀n}, the EUAV de-
for computing the offloaded task of user m, En,m com
= ployment scheme u = {un |∀n}, the power allocation scheme
1
v
fn,m (e1 (fn,m ) + e2 ), where e1 > 0 and e2 > 0 are constants
εFm

obtained by the offline power fitting, and 2.5 < v < 3 according 1 In this paper, we assume that the TUAV is located right above the GCC, due
to [26]. to the limited mobility introduced by the tether.
Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
4986 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

P = {pm,n , qn,m |∀m, ∀n}, and the computing capacity alloca-

Algorithm 1: SCP-based EUAV deployment algorithm.
tion scheme F = {fn,m , fc,m |∀m, ∀n}, while simultaneously
satisfying the cache and computing capacity constraints of
UAVs, as well as the latency and energy consumption constrains
of users’ tasks. Mathematically, the optimization problem can
be expressed as
min W total
α,β,u,P,F


N
 E 
s.t. βm,n + βm,n
C
= 1, (a)
n=1
E
βm,n C
, βm,n , αn,m ∈ {0, 1} (b)

M
E
βm,n fn,m ≤ ψn , (c)
m=1


M subproblem, the resource allocation subproblem, as well as the
αn,m Cm ≤ ϕn , (d) caching and offloading decision subproblem. Then, the SCP-,
m=1 SQP-, and DQN-based algorithms are proposed to solve these
three subproblems, respectively. The proposed SCP- and SQP-
m,n ≤ Tmax , (e)
Ltotal
based algorithms optimize the deployment and the resource
total
Em,n ≤ Emax , (f ) allocation of the EUAVs, respectively. Based on which, the
proposed DQN-based algorithm allows for adaptively adjusting
0 ≤ pm,n ≤ pmax
m , (g) the caching and offloading decisions of the system such that the
0 ≤ qn,m ≤ qnmax , (h) system cost can be minimized.
xmin ≤ xn ≤ xmax , ymin ≤ yn ≤ ymax , (i) (25)
A. EUAV Deployment Subproblem
where constraint (25a) means that user m can only choose to
We first focus on the EUAV deployment subproblem. With the
compute the task Om locally or offload it to the edge or the
given caching and offloading decision scheme and the resource
cloud for further computation, constraint (25b) regulates that the
allocation scheme, the problem in (25) can be expressed as
caching and offloading decision variables are binary variables,
constraint (25c) means that the computation tasks offloaded to min W total
u (26)
EUAV n should not exceed its computing capacity ψn , constraint s.t. (25e)(25f )(25i).
(25d) means that the content files cached in EUAV n should not
Note that the objective function is non-convex. To address this
exceed its cache capacity ϕn , constraints (25e) and (25f) mean
issue, we transform W total to a convex form following the steps.
that the time and energy consumption for completing tasks of total
According to (8)–(9) and (13)–(23), we first express Wm,n as
user m shouldn’t exceed the corresponding threshold, which are
application-specific and changes with user needs, constraints Om  E 
total
Wm,n = δm,n + E,U p
(ωm,1 + ωm,2 pm,n ) βm,n + βm,n
C
(25g) and (25h) are the transmit power constraints for the users rm,n
and EUAVs, respectively, and constraint (25i) restricts the scope C
of the EUAVs’ deployment area. Om βm,n
+ C,U p
[(ωm,1 + ωm,2 puav )]
Note that (25) is a mixed integer non-linear programming Rm,n
(MINLP) problem, due to the discrete and continuous variables, Cm
as well as the non-convex objective function and constraints. As + E,Down
(ωm,1 + ωm,2 qn,m )
such, (25) is in general difficult to solve. To address this issue, rn,m
in the next section, taking the combinatorial characteristic and Cm (1 − αn,m )
inherent non-convexity of (25) into consideration, we propose a + C,Down
(ωm,1 + ωm,2 pcloud ) , (27)
Rn,m
hybrid problem-solving framework that minimizes the weighted
sum cost of latency and energy consumption of our considered where δm,n is expressed as
  
networks, while inducing relatively low computational complex-   Fm
δm,n = 1 − βm,n − βm,n ωm,1
E C
+ ωm,2 εfm Fm
2
ity. fm
 
Fm
III. PROPOSED HYBRID PROBLEM-SOLVING FRAMEWORK: THE + βm,n (ωm,1 + ωm,2 e2 )
E
+ ωm,2 e1 fn,m Fm
v−1
fn,m
SS-DQN ALGORITHM  
Fm
In this section, we propose a novel SS-DQN algorithm to solve +βm,n
C
(ωm,1 +ωm,2 e4 ) +ωm,2 e3 fc,m
v−1
Fm .
the optimization problem in (25). Specifically, we decompose fc,m
(25) into three subproblems, namely, the EUAV deployment (28)
Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: COMPUTATION OFFLOADING AND RESOURCE ALLOCATION IN UNMANNED AERIAL VEHICLE NETWORKS 4987

By denoting up,(l) up,(l)

with bm,n and cm,n respectively given by
 E  up
c1 = (ωm,1 + ωm,2 pm,n ) βm,n + βm,n
C
, γm,n Bm,n C2
bup,(l)
m,n =  a1 ,
(l) (l) (l) 2 up (l)
c2 = βm,n
C
(ωm,1 + ωm,2 puav ) , ūm,n ln 2 ūm,n v̄m,n +γm,n ūm,n C1 +C2
c3 = (ωm,1 + ωm,2 qn,m ) , (l)
up
a1 γm,n Bm,n ūm,n C1 +C2
c4 = (1 − αn,m ) (ωm,1 + ωm,2 pcloud ) , (29) m,n =
cup,(l)  a1 .
(l) (l) (l) 2 up (l)
we can re-express (26) as v̄m,n ln 4 ūm,n v̄m,n +γm,n ūm,n C1 +C2
  (34)
N M
c 1 Om c 2 Om c3 Cm c4 Cm
min δm,n + E,U p + C,U p + E,Down + C,Down Similarly, the constraint (30d) can be upper bounded by
u,e,g em,n em,n en,m en,m
n=1m=1
eE,Down
n,m ≤ r̄n,m
E,Down
(gm,n , un ) , (35)
s.t. (25i)
where
 
Ltotal
m,n eE,U
m,n , em,n , en,m ≤ Tmax , (a)
p C,U p E,Down C,Down
, en,m E,Down
r̄n,m (gm,n , un )
 
total
Em,n eE,U p C,U p E,Down C,Down
m,n , em,n , en,m , en,m ≤ Emax , (b) = rn,m
E,Down (l)
gm,n , u(l)
n

m,n ≤ rm,n (gm,n , un ) , (c)

eE,U p E,U p  
(l)

−(C3 +C4 gm,n ) − C3 +C4 gm,n
− bdown,(l) e − e
eE,Down
n,m ≤ rn,m
E,Down
(gm,n , un ) , (d) n,m

  2 
hn  (l)
2 
gm,n ≤ , (e) − cdown,(l) un − um  − un − um  , (36)
dm,n (un ) n,m

down,(l) down,(l)
m,n ≤ Rm,n (un ) , (f )
eC,U p C,U p
with bn,m and cn,m respectively given by

eC,Down ≤ Rn,m
C,Down
(un ) .(g) (30) bdown,(l)
n,m
n,m
down
γn,m Bm,n C2
where eE,U p C,U p E,Down C,Down
m,n , em,n , en,m , en,m and gm,n are the intro- =  a1 ,
duced slack variables. As such, the objective function is now (l) (l) (l) 2 (l)
ūm,n ln 2 ūm,n v̄m,n + γn,m
down ū
m,n C1 + C2
convex.
However, the constraints (30c)–(30g) are non-convex. We
cdown,(l)
define the function f (u, v) = Blog2 (1 + (C1 + Cu2 ) va/2
γ
). Note n,m

that f (u, v) is a convex function of the variables u and v, which down

a1 γn,m
(l)
Bm,n ūm,n C1 + C2
means f (u, v) can be approximated by its lower bound at any =  a1 .
points that satisfy ū > 0 and v̄ > 0. Therefore, we obtain the (l) (l) (l) 2 (l)
v̄m,n ln 4 ūm,n v̄m,n + γn,m
down ū
m,n C1 + C2
following inequality [20]
(37)
γBC2 (u − ū)
f (u, v) ≥ f (ū, v̄) −     Now the constraints (30c) and (30d) have been success-
ū ln 2 ū v̄ a/2 + γ (ūC1 + C2 )
fully converted into convex constraints (32) and (35), respec-
βγB (ūC1 + C2 ) (v − v̄) tively. Next, we deal with the non-convex constraints (30e)–
−    . (31)
v̄ ln 4 ū v̄ a/2 + γ (ūC1 + C2 ) (30f). We firstly define two functions f (x) = xh1/2 and f (y) =
Blog2 (1 + ya/2 γ
). Note that f (x) and f (y) are convex functions
Based on (31), we make u = 1 + e−(C3 +C4 gm,n ) , of the variable x and y, respectively, which means they can
(l) (l) (l)
v = d2m,n , ū = ūm,n = 1 + e−(C3 +C4 gm,n ) , and v̄ = v̄m,n = separately be approximated by their lower bound at any points
(l) that satisfy x̄ > 0 and ȳ > 0. Therefore, we can get the following
un − um 2 + h2n . As such, the constraint (30c) can be
converted to the following convex form inequalities
h (x̄)
m,n ≤ r̄m,n (gm,n , un ) ,
eE,U p E,U p
(32) f (x) ≥ f (x̄) − , (38)
2x̄3/2
where βγB (y − ȳ)
f (y) ≥ f (ȳ) −  . (39)
ln 4 · ȳ ȳ β/2 + γ
E,U p
r̄m,n (gm,n , un )
 
(l)
 Based on (38) and (39), we substitute the functions at right side
(l) − C +C g
= rm,n
E,U p
gm,n , u(l)
n −bup,(l)
m,n e−(C3+C4 gm,n ) − e 3 4 m,n of the constraints (30e)–(30g) with their lower bounds. Specifi-
(l)
cally, in constraint (30e), we set f (x) = dhm,n
n
and x̄ = (dm,n )2 .
  2
2  (l)  In constraints (30f) and (30g), we set f1 (y) = Rm,nC,U p
(un ) and
− cup,(l)
m,n un − um  − un − um  , (33) (l)
f2 (y) = Rn,m
C,Down
(un ) with the same ȳ = (dn )2 , respectively.

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
4988 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

which can be solved by the CVX solver [29]. The deployment

Algorithm 2: Resource allocation algorithm based on SQP.
algorithm of EUAVs is detailed in Algorithm 1, in which the
optimization problem P1 is iteratively solved until the relevant
cost converges to a predefined threshold.

B. Resource Allocation Subproblem

Next, we focus on the resource allocation subproblem. With
the given caching and offloading decision scheme and the EUAV
deployment scheme, the problem in (25) can be expressed as
min W total
P,F

s.t. (25e)–(25h). (44)

which is non-convex and difficult to be directly solved. To tackle
this problem, we decompose it into a computing resource allo-
cation subproblem and a transmit power allocation subproblem,
respectively, given by
min W total
F

s.t. (25e), (25f). (45)

min W total
P

s.t. (25e)–(25h). (46)

Note that, if the power allocation is determined, (45) is a
convex optimization problem with regard to the variables fn,m
and fc,m , which can be solved by the CVX solver. However,
for a given computation resource allocation, (46) is still non-
convex. Therefore, we first use the variable transformation
Therefore, we can get the following convex constraints method by introducing the variables νm,n 1
= 1/R(pm,n ) and
  2  νm,n = 1/R(qn,m ) . According to the equations (4) and (5), the
2
2  (l) 
hn un − um  − un − um  relationship between the transmit power and transmission rate
hn can be expressed as
gm,n ≤ (l) − 32
,
dm,n (l) R(pm,n ) /Bm,n
2 dm,n pm,n = 2 − 1 κ,
(40)
R(qn,m ) /Bm,n
qn,m = 2 − 1 κ, (47)
m,n ≤ Rm,n
eC,U p C,U p
u(l)
n
where κ = κ1 /κ2 with κ1 = and κ2 = (C1 +
2
σm,n dam,n
1
  2 
 (l)  C2
)α0 . As such, the optimization problem
a2 γnup un − u0 2 − un − u0  1+e−(C3 +(C4 Hn /dm,n ))
(46) can be reformulated as
− 2  , (41) M    
(l) (l)
a2 N  1
ln 4 dn dn + γnup min K1 νm,n ωm,1 + ωm,2 κ 2
1 1
Bm,n νm,n
−1
p
n=1 m=1
eC,Down
n,m ≤ C,Down
Rn,m u(l)
n   1

+ Cm νm,n ωm,1 + ωm,2 κ 2 Bm,n νm,n − 1
2 2
  2
2  (l) 
a2 γn un −u0  − un −u0 
dw

− 2 a2  , (42) +K2 + K3
(l) (l)
ln 4 dn dn + γndw
 1 
m,n νm,n , νm,n ≤ Tmax , (a)
s.t. Ltotal 2
where γnup= puav α0
and = γndw Pcloud α0
.
 1 
2
σn 2
σn
Based on the above analysis, the EUAV deployment problem total
Em,n 2
νm,n , νm,n ≤ Emax , (b)
can be converted into a convex optimization problem as
  1
M
c 1 Om c 2 Om c3 Cm c4 Cm
1
νm,n ≥ , (c)
P 1 : min δm,n + E,U p + C,U p + E,Down + C,Down R(pmax )
m,n
u,e,g em,n em,n en,m en,m
m=1
1
2
νm,n ≥ , (d) (48)
s.t. (25i), (30a), (30b), (32), (35), (40)–(42), (43) R(qmax )
n,m

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: COMPUTATION OFFLOADING AND RESOURCE ALLOCATION IN UNMANNED AERIAL VEHICLE NETWORKS 4989

where the expressions of the parameters K1 -K5 , Ltotal m,n and Ck = ∇fv(k) v(k) ,
total
Em,n in (49), shown at the bottom of the page.
 T
Lemma 1: (48) is a convex optimization problem. Ak = ∇v(k) g1 v(k) , . . ., ∇v(k) g4 v(k) ,
1
Proof: We define the function f (v) = v · 2 Bm,n ·v , and its
 T
first-order derivative with respect to v is given by
  Bk = g1 v(k) , . . ., g4 v(k) , (54)
1 ln 2
f  (v) = 2 Bm,n ·v 1 − − 1. (50) (k) (k) (k)
Bm,n · v where v(k) = [vm,1 , vm,2 ], λi is the Lagrange multiplier of the
The second-order derivative with respect to v is given by i-th constraint at the k-th iteration. The update principle of the
1 (ln 2)2 Lagrange multiplier is given by
f  (v) = 2 Bm,n ·v . (51)  
2
Bm,n v3 T −1
(k+1) (k) (k)
λi = Aa Aa A(k)
a (Hk s + Ck ) , (55)
When v > 0, the second-order derivative of f (v) satisfies
f  (v) > 0, which means f (v) is a convex function with respect (k)
where Aa represents all the rows in A(k) at the k-th iteration
to the variable v. As such, the objective function and constraints
that satisfies the condition Ak s + Bk = 0. At each iteration of
(48a), (48b) in (48) are the combinations of convex and linear
the SQP algorithm, the descent direction searching problem is
functions with respect to the variables vm,1 and vm,2 . Therefore,
converted into a quadratic optimal problem, and the obtained
(48) is a convex optimization problem. The proof is completed.
solution is the Newton direction of the Lagrange function with
Note that both the objective function and constraints in
respect to variable v.
(48) contain non-linear function, which makes it a non-linear
The resource allocation algorithm is demonstrated in Algo-
programming problem and can be solved by the SQP algo-
rithm 2. In this algorithm, we solve the optimization problem
rithm. The core idea of the SQP algorithm is to convert the
(45) and (46) iteratively until converging to a prescribed accu-
complex nonlinearly constrained optimization problem into a
racy to obtain the optimal computation and power allocation
relatively simple quadratic programming (QP) problem. Specif-
results F and P, respectively.
ically, we set the objective function in (48) as f (νm,n
1 2
, νm,n ), and
represent the constraints (48a)–(48d) as g1 (νm,n , νm,n ) ≥ 0,
1 2
C. Caching and Offloading Decision Subproblem
g2 (νm,n
1 2
, νm,n ) ≥ 0, g3 (νm,n
1 2
, νm,n ) ≥ 0 and g4 (νm,n
1 2
, νm,n )≥
(k) (k) (k) In this subsection, we solve the caching and offloading de-
0, respectively. Then, the Lagrange multipliers λ1 , λ2 , λ3 ,
(k) cision subproblem. To this end, we re-express the problem in
and λ4 are introduced to obtain the Lagrange function as
(25) for given EUAV deployment and resource allocation, as
 1   1  4
 1  follows
L νm,n 2
, νm,n , λ = f νm,n 2
, νm,n − λi gi νm,n 2
, νm,n .
i=1 min W total
α,β
(52)
Based on (52) and Karush-Kuhn-Tucker (KKT) conditions, s.t. (25a)–(25d). (56)
we can get the following quadratic programming problem Note that the problem in (56) is a NP-hard problem, due
N  M
1 T to the computational complexity is O(M N +2 ) for computation
P 2 : min s Hk s + C k T s offloading decision determination only, where M and N denote
n=1 m=1
2
the number of users and EUAVs, respectively.
s.t. Ak s ≤ −Bk , (53) To address this issue, we propose a modified DQN-based
where algorithm to perform the caching and offloading decision, where
the central controller manages and coordinates the users and

4
(k)
Hk = ∇2v(k) f v(k) − λi ∇2v(k) gi v(k) , UAVs in a unified manner. Note that the DQN method is
i=1 based on the neural network and Q-learning. In the DQN,

 E  C
Om βm,n (ωm,1 + ωm,2 puav ) Cm (ωm,1 + ωm,2 pcloud )
K1 = Om βm,n + βm,n
C
, K2 = δm,n + C,U p
, K3 = (1 − αn,m ) C,Down
,
Rm,n Rn,m
 
  F m F m O m F m Cm
K4 = 1 − βm,n E
− βm,n
C
+ βm,nE
+ βm,n
C
C,U p
+ + (1 − αm ) C,Down ,
fm fn,m Rm,n f c,m Rn,m
 
  2 Om puav Cm pcloud
K5 = 1 − βm,n E
− βm,n
C
εfm Fm + βm,n E com
En,m + βm,n
C
C,U p
+ Ec,m
com
+ (1 − αm ) C,Down ,
Rm,n Rn,m
 1
  1

= + + = − + − + K5 .
1 2
Ltotal
m,n K ν 1
1 m,n C ν 2
m m,n K 4 , E total
m,n K ν 1
1 m,n κ 2 Bm,n νm,n
1 C ν 2
m m,n κ 2 Bm,n νm,n
1 (49)

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
4990 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

we apply the experience replay and target network technolo-

Algorithm 3: SS-DQN Algorithm for Caching and Offload-
gies to enhance the learning reliability and efficiency. After
ing Decisions.
each iteration, the experience tuple (st , at , rt , st+1 ) is stored
in the experience replay buffer E, where in this paper, the 1: Input: episode length Lep , discount factor γ, replay
state st , action at and reward rt are respectively defined as buffer size E, mini-batch size Nb , number of steps to
follows update the target network Ls ;
r State st : The state obtained from observing the envi- 2: Output: Well-trained neural network parameters;
ronment can be represented as st = (Wttotal , ut , Ψt , Φt ), 1: Randomly initialize the prediction network
which contains the system cost, deployment of the EU- Q(st , at ; θ) and target network Q (st , at ; θtar ) with
AVs, the remaining caching and computation resources weights θ and θtar ;
of the EUAVs at current time slot t, respectively. Note 2: l = 0;
that we have Φt = (Φt1 , Φt2 , . . ., ΦtN ) with Φtn = φn − 3: for episode l = 0:Lep do
M 4: Randomly initialize the EUAV deployment scheme
φn,m Cm , and Ψt = (Ψt1 , Ψt2 , . . ., ΨtN ) with Ψtn =
m=1 u0 and the resource allocation scheme F0 and P0 .
ψn − M βm,n
t E,t
fn,m .
r Action am=1 : The action includes the caching and offloading
Then observe the environment to obtain the state s0 ;
t 5: for time step t = 0:T do
decision, at = (αtm , β E,t m , β m |∀m ∈ M), where αm =
C,t t
E,t E,t E,t
6:  Select action: at =
(αm,1
t t
, αm,2 t
, . . ., αm,N ), β E,t
m = (βm,1 , βm,2 , . . ., βm,N ), arg maxat Q(st , at ; θ) with probability 1−
C,t C,t C,t ;
and β m = (βm,1 , βm,2 , . . ., βm,N
C,t
). random action with probability 
r Reward rt : After executing the action at , the system tran- 7: Execute the corresponding action at ;
sits from the state st to the state st+1 and generates the 8: Compute the EUAV deployment scheme ut and
corresponding reward rt . We define the utility function the resource allocation scheme Ft and Pt based
U = −W total to specify the optimization objective, where on the Algorithm 1 and 2, respectively;
a higher cost leads to a smaller utility value. However, 9: Compute the reward rt based on obtained
U cannot be directly used as the reward function since it solutions at , ut , Ft and Pt . Then the system
cannot evaluate the desirability of the action taken under transits to the next state st+1 ;
a particular state. To this end, we consider adopting the 10: Store the experience tuple (st , at , rt , st+1 ) into
difference of two consecutive utility values and as such, replay buffer E;
define the reward function as 11: if E is full then
⎧
⎨r1 , if Ut+1 − Ut > ζ 12: Randomly sample Nb tuples from E to form the
rt = r2 , if Ut+1 − Ut < ζ (57) mini batch;
⎩ for i = 1:Nb do
0, otherwise 13:
14: Obtain the target value yt ;
where r1 > 0, r2 < 0, and ζ is a tunable parameter that
15: end for
adjust the agent’s sensitivity to changes of utility values.
16: Update the weight of evaluation network θ by
Moreover, we also introduce the penalties to rt when
minimizing L(θ) based on (58);
constraints (25c) and (25d) are violated. Based on the above
17: end if
reward function, the system is able to learn to judge the
18: if Remainder ( Lls )==0 then
quality of the action, and thus, obtain the optimal caching
19: Update the weight of target network θtar by
and offloading decision policies.
θtar = θ;
When E is full, the newest samples will replace the oldest
20: end if
samples to be stored in the buffer. θ is the weight of the neural
21: t = t + 1, l = l + 1;
network and is updated based on the minimum mean square
22: end for
error principle, which is given by
23: end for
1 
Nb
L (θ) = (yt − Q (st , at ; θ))2 , (58)
Nb t=1
ci is the cost of sample i. After calculating all the weights, the
where yt is the target value, yt = rt + sumtree is updated by adding the sample i and its corresponding
γ maxat+1 Q(st+1 , at+1 ; θ), γ ∈ (0, 1] represents the discount weight ηi as a new leaf. As such, the samples with the top Nb
factor for weighting future rewards, and Nb is the size of the highest weights will be chosen to train the network.
mini-batch sampled from replay buffer E. The parameter of the Finally, based on the proposed SCP-, SQP-, and DQN- based
target network is θtar . We adopt the soft updating principle to algorithms, we propose the SS-DQN algorithm shown in Algo-
update θtar = θ every few episodes. rithm 3. In this algorithm, we adopt a reward incentive mecha-
In order to enhance the performance of sampling and training, nism, in which with the EUAV deployment scheme u and the
we propose a cost sensitive replay memory method, in which resource allocation scheme F and P obtained by Algorithm 1
the sumtree [30] is introduced to record the training samples and Algorithm 2, the relevant reward can be calculated as an
based on their weights. The weight of the sample i is given by incentive to motivate the system to take actions toward the direc-
ηi = tanh Z/ci , where Z is the cost influence coefficient and tion of low-cost. In particular, the SS-DQN algorithm includes

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: COMPUTATION OFFLOADING AND RESOURCE ALLOCATION IN UNMANNED AERIAL VEHICLE NETWORKS 4991

TABLE II
SIMULATION PARAMETERS

Fig. 2. The framework of the proposed SS-DQN.

two phases, namely, the training phase and the test phase. In
the training phase, at each time slot t, the optimal caching and
offloading decision policies αt and β t are computed via DQN
with Ft−1 and Pt−1 obtained by Algorithm 1 and Algorithm 2
at the last time slot t − 1. The actions αt and β t are then used to
compute ut , Ft and Pt based on Algorithm 1 and Algorithm 2,
respectively. As such, the reward rt is computed and the system latency and energy consumption can have equal importance on
transit to the next state st+1 . The above processes are performed the cost evaluation. Other parameters utilized throughout the
iteratively in each time slot until the training phase is over. In whole simulation are summarized in Table II, unless otherwise
the test phase, the well-trained neural network is employed to specified.
perceive the environment and react in real-time. The framework In order to evaluate the performance of our proposed al-
of the proposed SS-DQN algorithm is illustrated in Fig. 2. gorithm, we introduce two benchmark strategies, namely, 1)
Note that with the help of SCP and SQP algorithms, SS-DQN the UAV-enabled local-edge combined (ULEC) scheme, which
algorithm only needs to train the caching and offloading decision reserves the local and edge computing parts, and have no
variables instead of all the involved variables in the optimization cloud center. In this scheme, three EUAVs work as the edge
problem (25), which drastically reduces the corresponding state- servers as in our proposed algorithm; 2) the UAV and access
action space and effectively enhance the executive efficiency of point (AP)-assisted local-edge combined (UAPLEC) scheme
the DQN algorithm. Since the computational complexity of the proposed in [32], in which a novel MEC system is considered
SCP and SQP algorithms is polynomial, the overall computa- with two ground MEC servers at the APs and one MEC server
tional complexity of the proposed SS-DQN algorithm is within carried by the UAV. The locations of the APs are fixed at the
an acceptable range. edge of the network, while the UAV can fly around to provide
faster response to the user requests. Note that the flight energy
IV. SIMULATION RESULTS consumption in the UAPLEC scheme is much larger than the
communication and computation energy consumption. There-
In this section, we provide numerical results to validate the
fore, in this section, we only focus on the energy consumption
performance of the proposed algorithm. Specifically, 10 users
produced by the communication and computation. In addition,
are uniformly and randomly distributed in a square area of
since the users start the computation only after obtaining their
1000 m × 1000 m, where three EUAVs and one TUAV are de-
required cache files, and the cache capacity of edge servers in
ployed. The 2-D coordinate2 of the TUAV is fixed as [1000,1000]
the above two comparison schemes is relatively limited. Thus,
m. The heights of UAVs and TUAV are set as hn = 100 m
the computing tasks cannot be executed if their requested files
and h0 = 150 m, respectively. Following [23], the parameters
have not been cached in EUAVs or APs, and an extra cost is
in (8) are separately set as C1 = 0.01, C2 = 0.99, C3 = −4.7
introduced as a punishment.
and C4 = 8.9, which guarantees the maximum possible time-
In Fig. 3(a), we plot the system cost versus the number
out probability δ = 0.0025. α0 = 10−5 . The parameters in the
of users for different schemes. It can be seen that our pro-
energy consumption model are separately set as ε = 10−28 [27],
posed scheme induces the smallest cost compared to the two
e1 = 0.5, e2 = 0.4, e3 = 0.6, e4 = 0.6 [28]. We set the default
comparison schemes. The powerful processing capabilities of
values of weights in (23) as ωm,1 = ωm,2 = 0.5. As such, the
the cloud bring a significant reduction to cost, showing the
superiority of the cloud-edge architecture and how a proper
2 Note that the 2-D coordinate of the TUAV can be any point within the offloading decision plays a crucial role in enhancing the system
simulation area, and the trend observed from the figures will not be changed. performance. Moreover, with an increasing number of users, the
In our work, considering the limited length of the tether [36] and the complex
topographic conditions [37], we choose to place the TUAV at the edge of the total system cost increases as well. In addition, the performance
simulation area. gap between the ULEC scheme and the UAPLEC scheme is

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
4992 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

Fig. 3. System cost performance of the considered three schemes at varying number of users (M ), cache capacity of each EUAV (φn ) and computation capacity
of each EUAV (ψn ). (a) Cost versus number of users. (b) Cost versus cache capacity of EUAV. (c) Cost versus computation capacity of EUAV.

Fig. 4. The impact of parameter settings (including the size of computation tasks, cached files and workloads) on the system cost under the scenario of M = 10,
N = {1, 3, 5}. (a) System cost versus size of the computation task. (b) System cost versus size of the cached file. (c) System cost versus the workload.

relatively small, verifying the effectiveness of the reasonable by higher computation frequency is offset by the increase of
EUAV deployment. energy consumption. In addition, the performance gain of the
We further examine the effect of cache capacity φn on the cost ULEC and UAPLEC schemes is smaller than that of our pro-
performance in Fig. 3(b). It is again observed that our proposed posed scheme, leading to that the performance gap between
scheme outperforms the other two benchmark solutions. We can our scheme and the two comparison schemes becomes increas-
also see that the system costs of the three considered schemes ingly larger with ψn . This observation is different from that in
decreases with the increase of φn , due to the enhanced capabil- Fig. 3(b). Note that the computing tasks of the users are executed
ities of the edge to cache more files and provide faster response only when the required cache files are obtained, which explains
to the user requests. It is worth noting that the performance why the limited cache capacity of the two comparison schemes
gap between our solution and the two comparison schemes is restricts the effect of increasing ψn .
becoming smaller with higher φn . The reason lies in φn , when φn The impacts of the parameter settings (i.e., size of the com-
is small, most of users’ tasks in the ULEC and UAPLEC schemes putation task, size of the cached file and the workload) on the
cannot be accomplished, resulting in a heavy penalty on the system cost are contemplated in Fig. 4. We consider M = 10
system cost. However, when φn is relatively large, all users’ tasks and N = {1, 3, 5}. It can be observed from Fig. 4(a) that, with
can be executed normally, and further increasing the capacity is the fixed M and N , the increasing size of computation tasks
no longer available for enhancing the cost performance, and thus Om degrades the cost performance of the system, due to the
the cost approaches to a certain value. higher transmission latency of communication links between
The effect of computation capacity ψn on the cost perfor- the users and the EUAVs, the EUAVs and the TUAV. As such,
mance is investigated in Fig. 3(c). Similar to the observations in the gain of processing latency produced by powerful computing
Fig. 3(b), the increase of ψn can bring reduction to the system capabilities of the edge and the cloud will finally be offset by the
cost. This is because the enhanced capabilities of the edge make larger transmission latency, leading to the scenario that all the
the task computation faster. Moreover, we can observe how the users choose the local computing mode. In the same setting, we
costs of the three considered schemes gradually approach to then consider the effect of cached file size Cm and workload Fm
certain values. This is because the reduction of latency brought in Fig. 4(b) and (c), respectively. It can be seen from Fig. 4(b)

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: COMPUTATION OFFLOADING AND RESOURCE ALLOCATION IN UNMANNED AERIAL VEHICLE NETWORKS 4993

Fig. 6. CDF of users’ latency for completing the related computation tasks.
Fig. 5. The selection of computing mode for different user densities.

that the cost increases with Cm , which is because that a small

Cm can be handled by EUAVs easily, while a relatively large Cm
requires the involvement of the cloud, leading to the increase of
the system cost. Similarly, it can be seen that in Fig. 4(c), with
an increasing size of workload Fm , the system cost increases as
well. The reason is that when Fm is small, the QoS requirements
of users can be met through local computing, which saves the
transmission latency to the edge and the cloud. However, when
Fm is relatively large, the local computing mode is no longer
applicable for satisfying the QoS requirements of the users,
which requires the assistance of the more powerful EUAVs and
cloud center. In addition, from Fig. 4(a) to (c), we can commonly
see that the increase of N simultaneously raises the capabilities
of data transmitting, caching and computing, thus, lowering the
system cost.
Fig. 5 shows the users’ computing mode selection for different
Fig. 7. System cost versus the training episode with various learning rate lr =
user densities. It can be observed that the number of scheduled 0.001 and lr = 0.0001.
computation tasks of all three modes increases with the user
density. Moreover, as the user density increases, an increased
number of users tend to choose the local computing mode. The also introduce a complete DQN (CDQN) algorithm [33] as a
reason is that the computation resources at the edge and cloud comparison. Specifically, the CDQN scheme is a pure DQN
gradually become insufficient for users’ demands. We can also algorithm, where both the caching and offloading decisions and
observe that the number of users choosing edge computing is the resource allocation strategies are jointly optimized by the
always larger than that choosing cloud computing, due to the centralized DQN controller. Note that the action space for UAV
closer distance between users and EUAVs. deployment is too large to be directly solved by DQN, therefore,
To further analyze the variability of user performance, in we assume that the location of the UAVs are determined in
Fig. 6, we plot the cumulative distribution function (CDF) of advance. It can be observed that when lr = 0.0001, our proposed
user latency for completing the computation tasks. It can be scheme is stabilized at around 100 episodes. However, not only
observed that our proposed scheme can provide the lowest does the CDQN scheme converges slower, but it also suffers a
latency on average. Moreover, our proposed scheme is more large performance loss compared to our scheme, even after hav-
compact compared with the other two schemes, which means ing reached its convergence. The reason is that the quantization
a smaller variance, namely, most of the users in our scheme of the transmit power and the computing capacity reduces the
can complete their computation tasks with a close-to-average chances of obtaining the optimal solutions, and thus, deteriorates
time. Such a small variance is extremely significant toward its performance. Such an observation shows that our proposed
implementing the real-time applications, which require low and DQN-based scheme that combining advantages of both the DQN
constant completion time. and convex optimization methods is an efficient and valuable
Then, we investigate the convergence performance in Fig. 7 solution to enhance the system performance. Moreover, we can
with different learning rate lr = 0.001 and lr = 0.0001. We see that a smaller learning rate contributes to a higher converge

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
4994 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 4, APRIL 2023

optimization to improve the cost performance and the training

efficiency.

REFERENCES
[1] Cisco, “Cisco visual networking index: Global mobile data traffic forecast
update, 2018-2023,” White Paper, Feb. 2020.
[2] F. Spinelli and V. Mancuso, “Toward enabled industrial verticals in 5G: A
survey on MEC-based approaches to provisioning and flexibility,” IEEE
Commun. Surv. Tut., vol. 23, no. 1, pp. 596–630, Jan.–Mar. 2021.
[3] B. Liu, C. Liu, and M. Peng, “Resource allocation for energy-efficient
MEC in NOMA-enabled massive IoT networks,” IEEE J. Sel. Areas
Commun., vol. 39, no. 4, pp. 1015–1027, Apr. 2021.
[4] F. Jiang, K. Wang, L. Dong, C. Pan, W. Xu, and K. Yang, “AI driven
heterogeneous MEC system with UAV assistance for dynamic environ-
ment: Challenges and solutions,” IEEE Netw., vol. 35, no. 1, pp. 400–408,
Jan./Feb. 2021.
[5] S. Gangakhedkar, H. Cao, A. R. Ali, K. Ganesan, M. Gharba, and J.
Eichinger, “Use cases, requirements and challenges of 5G communication
Fig. 8. The performance comparison between the binary and partial offloading. for industrial automation,” in Proc. IEEE Int. Conf. Commun. Workshops,
Kansas City, MO, USA, 2018, pp. 1–6.
[6] J. Wang, C. Jin, Q. Tang, N. N. Xiong, and G. Srivastava, “Intelligent
ubiquitous network accessibility for wireless-powered MEC in UAV-
speed, a smaller jitter, and a better cost performance, which assisted B5G,” IEEE Trans. Netw. Sci. Eng., vol. 8, no. 4, pp. 2801–2813,
motivate our choice to set the default value of lr as 0.0001. Oct.–Dec. 2021.
Fig. 8 shows how the partial offloading mode can enhance [7] L. Wang and M. Hua, “Optimal bit allocation for UAV-enabled mo-
bile communication,” in Proc. 3rd IEEE Int. Conf. Comput. Commun.,
the cost performance. To this end, we compare the performance Chengdu, China, 2017, pp. 474–478.
of the binary offloading mode to that of the partial offloading [8] G. Lee, W. Saad, and M. Bennis, “Online optimization for UAV-assisted
mode with L = 10, where each computation task is assumed distributed fog computing in smart factories of industry 4.0,” in Proc. IEEE
Glob. Commun. Conf., Abu Dhabi, UAE, 2018, pp. 1–6.
to be further divided into 10 equal-size segments, as such, the [9] L. Pacheco et al., “Towards the future of edge computing in the sky:
original 10 computation tasks with size [0.5,4] Mbits in the Outlook and future directions,” in Proc. 2021 17th Int. Conf. Distrib.
binary offloading scenario become 100 tasks with size [50,400] Comput. Sens. Syst., 2021, pp. 220–227.
[10] Z. Yu, Y. Gong, S. Gong, and Y. Guo, “Joint task offloading and resource
Kbits. It can be observed in Fig. 7 that the cost performance allocation in UAV-enabled mobile edge computing,” IEEE Internet Things
of the partial offloading mode is better than that of the binary J., vol. 7, no. 4, pp. 3147–3159, Apr. 2020.
offloading, due to the more sufficient utilization of the local, [11] A. Sacco, F. Esposito, G. Marchetto, and P. Montuschi, “Sustainable task
offloading in UAV networks via multi-agent reinforcement learning,” IEEE
edge and cloud resources. Moreover, the convergence speed of Trans. Veh. Technol., vol. 70, no. 5, pp. 5003–5015, May 2021.
the partial offloading is slower than that of the binary offloading, [12] Z. Ning et al., “5G-enabled UAV-to-community offloading: Joint trajectory
due to the enlarged action space. design and task scheduling,” IEEE J. Sel. Areas Commun., vol. 39, no. 11,
pp. 3306–3320, Nov. 2021.
[13] E. E. Haber, H. A. Alameddine, C. Assi, and S. Sharafeddine, “UAV-aided
ultra-reliable low-latency computation offloading in future IoT networks,”
V. CONCLUSION IEEE Trans. Commun., vol. 69, no. 10, pp. 6838–6851, Oct. 2021.
[14] G. Faraci, C. Grasso, and G. Schembra, “Design of a 5G network slice
In this paper, we considered a UAV-assisted cloud-edge sys- extension with MEC UAVs managed with reinforcement learning,” IEEE
tem, where the EUAVs serve as MEC servers to cooperate with J. Sel. Areas Commun., vol. 38, no. 10, pp. 2356–2371, Oct. 2020.
the cloud center to provide cache and computing services for [15] Z. Yuan, J. Jin, L. Sun, K. -W. Chin, and G. -M. Muntean, “Ultra-reliable
IoT communications with UAVs: A swarm use case,” IEEE Commun.
the ground users. The task scheduling between the users, the Mag., vol. 56, no. 12, pp. 90–96, Dec. 2018.
EUAVs and the cloud was investigated. In order to minimize [16] Y. Huang, Q. Wu, R. Lu, X. Peng, and R. Zhang, “Massive MIMO
the weighted sum cost of latency and energy consumption, the for cellular-connected UAV: Challenges and promising solutions,” IEEE
Commun. Mag., vol. 59, no. 2, pp. 84–90, Feb. 2021.
caching and offloading decisions, EUAV deployment, as well [17] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications with un-
as transmit power and computation capacity allocation schemes manned aerial vehicles: Opportunities and challenges,” IEEE Commun.
were jointly optimized under the constraints of the EUAV capac- Mag., vol. 54, no. 5, pp. 36–42, May 2016.
[18] A. Al-Hourani, S. Kandeepan, and S. Lardner, “Optimal LAP altitude
ities and user-specific latency and energy consumption limits. for maximum coverage,” IEEE Wireless Commun. Lett., vol. 3, no. 6,
In order to deal with the NP-hard problem, an novel SS-DQN pp. 569–572, Dec. 2014.
algorithm was proposed, which combines the advantages of SCP, [19] H. Mei, K. Yang, Q. Liu, and K. Wang, “Joint trajectory-resource optimiza-
tion in UAV-enabled edge-cloud system with virtualized mobile clone,”
SQP and DQN methods to deal with the decomposed EUAV IEEE Internet Things J., vol. 7, no. 7, pp. 5906–5921, Jul. 2020.
deployment problem, the resource allocation problem, as well [20] A. A. Nasir, “Latency optimization of UAV-enabled MEC system for
as the caching and offloading decision problem, respectively. virtual reality applications under rician fading channels,” IEEE Commun.
Lett., vol. 10, no. 8, pp. 1633–1637, Aug. 2021.
The proposed algorithm allowed the system to learn the optimal [21] Y. Liu, K. Xiong, Y. Lu, Q. Ni, P. Fan, and K. B. Letaief, “UAV-aided
parameter tuple that minimize the system cost in a more effi- wireless power transfer and data collection in rician fading,” IEEE J. Sel.
cient way. Numerical results showed significant performance Areas Commun., vol. 39, no. 10, pp. 3097–3113, Oct. 2021.
[22] C. You and R. Zhang, “3D trajectory optimization in rician fading for
gain compared to the existing edge computing only schemes UAV-enabled data harvesting,” IEEE Trans. Wireless Commun., vol. 18,
and illustrated the advantages of combining DRL and convex no. 6, pp. 3192–3207, Jun. 2019.

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: COMPUTATION OFFLOADING AND RESOURCE ALLOCATION IN UNMANNED AERIAL VEHICLE NETWORKS 4995

[23] S. Sun, G. Zhang, H. Mei, K. Wang, and K. Yang, “Optimizing multi- Binghong Liu received the B.E. degree in informa-
UAV deployment in 3-D space to minimize task completion time in UAV- tion and communication engineering from the Beijing
enabled mobile edge computing systems,” IEEE Commun. Lett., vol. 25, University of Posts and Telecommunications, Bei-
no. 2, pp. 579–583, Feb. 2021. jing, China, in 2017. She is currently working to-
[24] H. Wei, H. Luo, Y. Sun, and M. S. Obaidat, “Cache-aware computation ward the Ph.D. degree with the State Key Laboratory
offloading in IoT systems,” IEEE Syst. J., vol. 14, no. 1, pp. 61–72, of Networking and Switching Technology, Beijing
Mar. 2020. University of Posts and Telecommunications. Her
[25] X. Huang, X. Yang, Q. Chen, and J. Zhang, “Task offloading optimization research interests include resource allocation in radio
for UAV-assisted fog-enabled Internet of Things networks,” IEEE Internet access networks, UAV, NOMA, convex optimization,
Things J., vol. 9, no. 2, pp. 1082–1094, Jan. 2022. and deep reinforcement learning.
[26] S. Guo, B. Xiao, Y. Yang, and Y. Yang, “Energy-efficient dynamic offload-
ing and resource scheduling in mobile cloud computing,” in IEEE Conf.
Comput. Commun., 2016, pp. 1–9.
[27] X. Cheng et al., “Space/aerial-assisted computing offloading for IoT
applications: A learning-based approach,” IEEE J. Sel. Areas Commun.,
vol. 37, no. 5, pp. 1117–1129, May 2019.
[28] L. Rao, X. Liu, M. D. Ilic, and J. Liu, “Distributed coordination of internet
data centers under multiregional electricity markets,” Proc. IEEE, vol. 100,
no. 1, pp. 269–282, Jan. 2012.
[29] S. Boyd, S. P. Boyd, and V. Lieven. Convex Optimization. Cambridge,
U.K.: Cambridge Univ. Press, 2004. Chenxi Liu (Senior Member, IEEE) received the B.E.
[30] T. Schaul, J. Quan, I. Antonoglou, and D. Silver, “Prioritized experience degree from Central South University, Changsha,
replay,” in Proc. Int. Conf. Learn. Represent., 2016, pp. 1–21. China, in 2010, and the Ph.D. degree from the Uni-
[31] C.-B. Park, B.-S. Park, H.-J. Uhm, H. Choi, and H.-S. Kim, “IEEE 802.15.4 versity of New South Wales, Sydney, NSW, Australia,
based service configuration mechanism for smartphone,” IEEE Trans. in 2016. From 2017 to 2019, he was a Postdoctoral
Consum. Electr., vol. 56, no. 3, pp. 2004–2010, Aug. 2010. Research Fellow with the Singapore University of
[32] Y. Xu, T. Zhang, Y. Liu, D. Yang, L. Xiao, and M. Tao, “UAV-assisted Technology and Design, Singapore. Since 2019, he
MEC networks with aerial and ground cooperation,” IEEE Trans. Wireless has been with the Beijing University of Posts and
Commun., vol. 20, no. 12, pp. 7712–7727, Dec. 2021. Telecommunications, Beijing, China, where he is cur-
[33] Y. -C. Wu, T. Q. Dinh, Y. Fu, C. Lin, and T. Q. S. Quek, “A hybrid DQN rently an Associate Professor. His research interests
and optimization approach for strategy and resource allocation in MEC include wireless security, unmanned aerial vehicle-
networks,” IEEE Trans. Wireless Commun., vol. 20, no. 7, pp. 4282–4295, enabled wireless networks, and network intelligence. He was the recipient the
Jul. 2021. Best Paper Award from the IEEE ICC 2022. He is currently the Editor of the
[34] X. Hu, K. Wong, K. Yang, and Z. Zheng, “UAV-assisted relaying and edge IEEE WIRELESS COMMUNICATIONS LETTERS.
computing: Scheduling and trajectory optimization,” IEEE Trans. Wireless
Commun., vol. 18, no. 10, pp. 4738–4752, Oct. 2019.
[35] M. Kishk, A. Bader, and M. -S. Alouini, “Aerial base station deployment
in 6G cellular networks using tethered drones: The mobility and endurance
tradeoff,” IEEE Veh. Technol. Mag., vol. 15, no. 4, pp. 103–111, Dec. 2020.
[36] M. A. Kishk, A. Bader, and M.-S. Alouini, “On the 3-D placement
of airborne base stations using tethered UAVs,” IEEE Trans. Commun.,
vol. 68, no. 8, pp. 5202–5215, Aug. 2020.
[37] S. Zhang, W. Liu, and N. Ansari, “On tethered UAV-assisted heteroge-
neous network,” IEEE Trans. Veh. Technol., vol. 71, no. 1, pp. 975–983,
Jan. 2022. Mugen Peng (Fellow, IEEE) received the Ph.D. de-
[38] A. Lapidoth, A Foundation in Digital Communication. Cambridge, U.K.: gree in communication and information systems from
Cambridge Univ. Press, 2017, pp. 500–501. the Beijing University of Posts and Telecommunica-
[39] R. G. Gallager, “Circularly-symmetric Gaussian random vectors,” tions (BUPT), Beijing, China, in 2005. Afterward,
Preprint, pp. 1–9, 2008. he joined BUPT, where he has been a Full Professor
[40] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Performance opti- with the School of Information and Communication
mization for UAV-enabled wireless communications under flight time con- Engineering since 2012. In 2014, he was an Academic
straints,” in Proc. IEEE Glob. Commun. Conf., Singapore, 2017, pp. 1–6. Visiting Fellow with Princeton University, Princeton,
[41] C. Zhan, H. Hu, X. Sui, Z. Liu, and D. Niyato, “Completion time and NJ, USA. He leads a Research Group focusing on
energy optimization in the UAV-enabled mobile-edge computing system,” wireless transmission and networking technologies
IEEE Internet Things J., vol. 7, no. 8, pp. 7808–7822, Aug. 2020. with the State Key Laboratory of Networking and
[42] Y. Zeng, J. Xu, and R. Zhang, “Energy minimization for wireless commu- Switching Technology, BUPT. He has authored or coauthored more than 100
nication with rotary-wing UAV,” IEEE Trans. Wireless Commun, vol. 18, refereed IEEE journal papers and more than 300 conference proceeding papers.
no. 4, pp. 2329–2345, Apr. 2019. He was the recipient of the 2018 Heinrich Hertz Prize Paper Award, the 2014
[43] H. -T. Ye, X. Kang, J. Joung, and Y. -C. Liang, “Optimization for IEEE ComSoc AP Outstanding Young Researcher Award, and the Best Paper
wireless-powered IoT networks enabled by an energy-limited UAV under Award in the JCN 2016 and IEEE WCNC 2015. He is on the Editorial/Associate
practical energy consumption model,” IEEE Commun. Lett., vol. 10, no. 3, Editorial Board of the IEEE Communications Magazine, the IEEE INTERNET OF
pp. 567–571, Mar. 2021. THINGS JOURNAL, and IEEE ACCESS.

Authorized licensed use limited to: KTH Royal Institute of Technology. Downloaded on August 30,2024 at 10:51:52 UTC from IEEE Xplore. Restrictions apply.