Computation Rate Maximization in UAV-Enabled

Wireless Powered Mobile-Edge Computing Systems

Fuhui Zhou, Member, IEEE, Yongpeng Wu, Senior Member, IEEE,
Rose Qingyang Hu, Senior Member, IEEE, and Yi Qian, Senior Member, IEEE

Abstract—Mobile edge computing (MEC) and wireless power computation offloading, partial computation offloading.
transfer (WPT) are two promising techniques to enhance the
computation capability and to prolong the operational time I. I NTRODUCTION
of low-power wireless devices that are ubiquitous in Internet
arXiv:1806.04589v2 [eess.SP] 17 Jul 2018

of Things. However, the computation performance and the

harvested energy are significantly impacted by the severe prop-
agation loss. In order to address this issue, an unmanned aerial
T HE Internet of Things (IoT) has been widely developed
with the unprecedented proliferation of mobile devices,
such as smart phones, cloud-based mobile sensors, tablet
vehicle (UAV)-enabled MEC wireless powered system is studied computers and wearable devices, which facilitates the real-
in this paper. The computation rate maximization problems in ization of smart environment (e.g. smart city, smart home,
a UAV-enabled MEC wireless powered system are investigated
under both partial and binary computation offloading modes, smart transportation, etc.) [1]. IoT enables mobile users to
subject to the energy harvesting causal constraint and the UAV’s experience intelligent applications (e.g., automatic navigation,
speed constraint. These problems are non-convex and challenging face recognition, unmanned driving, etc.) and to enjoy diverse
to solve. A two-stage algorithm and a three-stage alternative services with high quality of service (QoS) such as mobile
algorithm are respectively proposed for solving the formulated online gaming, augmented reality, etc. These services normally
problems. The closed-form expressions for the optimal central
processing unit frequencies, user offloading time, and user require a massive number of size-constrained and low-power
transmit power are derived. The optimal selection scheme on mobile devices to perform computation-intensive and latency-
whether users choose to locally compute or offload computation sensitive tasks [2]. However, it is challenging for mobile
tasks is proposed for the binary computation offloading mode. devices to perform these services due to their low computing
Simulation results show that our proposed resource allocation capability and finite battery lifetime.
schemes outperforms other benchmark schemes. The results also
demonstrate that the proposed schemes converge fast and have Mobile edge computing (MEC) and wireless power transfer
low computational complexity. (WPT) have been deemed two promising technologies to
Index Terms—Mobile-edge computing, wireless power transfer, tackle the above mentioned challenges [2]-[4]. Recently, MEC
unmanned aerial vehicle-enabled, resource allocation, binary has received an ever-increasing level of attention from industry
and academia since it can significantly improve the compu-
Manuscript received January 4, 2018; revised May 1, 2018 and accepted tation capability of mobile devices in a cost-effective and
June 4, 2018. Date of publication ****; date of current version ****. The energy-saving manner [2]. It enables mobile devices to offload
research of F. Zhou was supported in part by the Natural Science Foundation
of China under Grant 61701214, in part by the Young Natural Science partial or all of their computation-intensive tasks to MEC
Foundation of Jiangxi Province under Grant 20171BAB212002, in part by servers that locate at the edge of the wireless network, such as
The Open Foundation of The State Key Laboratory of Integrated Services cellular base stations (BSs) and access points (APs). Different
Networks under Grant ISN19-08, and in part by The Postdoctoral Science
Foundation of Jiangxi Province under Grant 2017M610400, Grant 2017KY04 from the conventional cloud computing, MEC servers are
and Grant 2017RC17. The research of Y. Wu was supported by the Natural deployed in a close proximity to end users. Thus, MEC has
Science Foundation of China under Grant 61701301 and in part by Young Elite the potential to provide low-latency services, to save energy
Scientist Sponsorship Program by CAST. The research of Prof. R. Q. Hu was
supported in part by the National Science Foundation under Grants EECS- for mobile users, and to achieve high security [2]. Up to now,
1308006, NeTS-1423348, EARS-1547312 and the Natural Science Foundation there are a number of leading companies (e.g., IBM, Intel, and
of China under Grant 61728104. The research of Prof. Y. Qian was supported Huawei) that have identified MEC as a promising technique
by the National Science Foundation under Grants EECS-1307580, NeTS-
1423408 and EARS-1547330. The corresponding author is Yongpeng Wu. for the future wireless communication networks. In general,
F. Zhou is with the Department of Electrical and Computer Engineering MEC has two operation modes, namely, partial and binary
as a Research Fellow at Utah State University, U.S.A. F. Zhou is also with computation offloading. In the first mode, the computation
the School of Information Engineering, Nanchang University, P. R. China,
330031. He is also with State Key Laboratory of Integrated Services Networks, task can be partitioned into two parts, and one part is locally
Xidian University, Xian, 710071, P. R. China (e-mail: [email protected]). executed while the other part is offloaded to the MEC servers
Y. Wu is with Shanghai Key Laboratory of Navigation and Location for computing [5]-[9]. For the second mode, computation tasks
Based Services, Shanghai Jiao Tong University, Minhang, 200240, China
(Email:[email protected]). cannot be partitioned. Thus they can be either executed locally
R. Q. Hu is with the Department of Electrical and Computer Engineering, or completely offloaded [10].
Utah State University, USA. (e-mail: [email protected]). On the other hand, WPT can provide low-power mobile
Y. Qian is with the Department of Electrical and Computer Engineer-
ing, University of Nebraska-Lincoln, Omaha, NE 68182, USA. (E-mail: devices with sustainable and cost-effective energy supply by
[email protected]). using radio-frequency (RF) signals [3]. It facilitates a perpetual
operation and enables users to have high QoE, especially in optimizing the communication and computation resources is
the case that mobile devices do not have sufficient battery superior to that obtained by optimizing these resources sep-
energy for offloading task or taking the services when the arately. The authors in [6] extended the energy minimization
battery energy is exhausted. Compared to the conventional problem into the multi-user MEC systems with time-division
energy harvesting techniques, such as solar or wind charging, multiple access (TDMA) and orthogonal frequency-division
WPT is more attractive since it can provide a controllable and multiple access (OFDMA), respectively. It was proved that
stable power supply [4]. It is envisioned that the computation the optimal offloading policy has a threshold-based structure,
performance can be significantly improved by integrating which is related to the channel state information (CSI) [6].
WPT into MEC networks [11]-[16]. However, the harvested Particularly, mobile users offload their computation tasks when
power level can be significantly degraded by the severe prop- the channel condition is strong; otherwise, they can locally
agation loss. Recently, an unmanned aerial vehicle (UAV)- execute the computation tasks. In [7], the revenue of the
enabled WPT architecture has been proposed to improve the wireless cellular networks with MEC was maximized by
energy transfer efficiency [17]-[20]. It utilizes an unmanned jointly designing the computation offloading decision, resource
aerial vehicle (UAV) as an energy transmitter for powering allocation, and content caching strategy. The works in [5]-
the ground mobile users. It was shown that the harvested [7] focused on optimizing a single objective, which over-
power level can be greatly improved due to the fact that emphasizes the importance of one metric and may not achieve
there is a high possibility that short-distance line-of-sight a good tradeoff among multiple metrics. Recently, the authors
(LoS) energy transmit links exist [17]-[20]. Moreover, the in [8] and [9] studied the fairness and multi-objective opti-
computation performance can also be improved by using the mization problem in MEC networks. It was shown that there
UAV-assisted MEC architecture [21]-[25]. Furthermore, UAV- exist multiple tradeoffs in MEC systems, such as the tradeoff
assisted architectures can provide flexible deployment and low between the total computation rate and the fairness among
operational costs, and are particularly helpful in the situations users. Different from the works in [5]-[9], MEC systems
that the conventional communication systems are destroyed by with the binary computation offloading mode were considered
natural disasters [26]-[32]. and the optimal resource allocation strategy was designed to
Motivated by the above mentioned reasons, a UAV-enabled minimize the consumption energy in [10].
and wireless powered MEC network is studied in this paper. Energy harvesting was not considered in the MEC systems
In order to maximize the achievable computation rate, the [5]-[10]. Recently, the authors in [11]-[16] have studied the
communication and computation resources and the trajectory resource allocation problem in various MEC systems relying
of the UAV are jointly optimized under both partial and binary on energy harvesting. In [11] and [12], The reinforcement
computation offloading modes. To the authors’ best knowl- learning and Lyapunov optimization theory were used to
edge, this is the first work that considers the UAV-enabled design resource allocation schemes in MEC systems relying
wireless powered MEC network and studies the computation on the conventional energy harvesting techniques. Different
rate maximization problems in this type of network. from [11] and [12], the resource allocation problems were
studied in wireless powered MEC systems [13]-[16]. Specif-
A. Related Work and Motivation ically, the authors in [13] proposed an energy-efficient com-
In wireless powered MEC systems, it is of great importance puting framework in which the energy consumed for local
to design resource allocation schemes so as to efficiently computing and task offloading is from the harvested energy.
exploit energy, communication, and computation resources and The consumed energy was minimized by jointly optimizing
improve the computation performance. Resource allocation the CPU frequency and the mode selection. In [14], the
problems have been extensively investigated in the conven- energy minimization problem was extended into a multi-input
tional MEC networks [5]-[10] and also in MEC networks rely- single-out wireless powered MEC system, and the offloading
ing on energy harvesting [11]-[16]. Recently, efforts have also time, the offloading bits, the CPU frequency and the energy
been dedicated to designing resource allocation and trajectory beamforming were jointly optimized. Unlike [14], energy
schemes in UAV-enabled wireless powered communications efficiency was defined and maximized in a full-duplex wireless
network [17]-[20] and UAV-assisted MEC networks [21]-[25]. powered MEC system by jointly optimizing the transmission
These contributions are summarized as follows. power, offloaded bits, computation energy consumption, time
In MEC networks, the communication and computation slots for computation offloading and energy transfer [15]. In
resources and the selection of the offloading mode were jointly contrast to the work in [13]-[15], the computation bits were
optimized to achieve the objective of the system design, maximized in a wireless powered MEC system under the
e.g., the users’ consumption energy minimization [5], [6], the binary computation offloading mode [16]. Two sub-optimal
revenue maximization [7], the maximum cost minimization algorithms based on the alternating direction method were
[8], etc. Specifically, in [5], the total energy of all users in a proposed to solve the combinatorial programming problem.
multi-cell MEC network was minimized by jointly optimizing The proposed algorithms actually did not provide the optimal
the user transmit precoding matrices and the central processing selection scheme for the user operation mode.
unit (CPU) frequencies of the MEC server allocated to each Although WPT has been exploited to improve the com-
user. It was shown that the performance achieved by jointly putation performance of MEC systems [13]-[16], the energy
harvested by using WPT can be significantly degraded by in UAV-enabled wireless powered MEC systems have binary
the severe propagation loss. The energy conversion efficiency variables related to the selection of either local computation
is low when the distance between the energy transmitter or offloading tasks. It makes the problem a mixed integer non-
and the harvesting users is large. In order to tackle this convex optimization problem.
challenge, the authors in [17]-[20] proposed a UAV-enabled
B. Contributions and Organization
wireless powered architecture where a UAV transmits energy
to the harvesting users. Due to the high possibility of having In contrast to [5]-[16], this paper studies the resource
line-of-sight (LoS) air-to-ground energy harvesting links, the allocation problem in UAV-enabled wireless powered MEC
harvesting energy can be significantly improved by using this systems, where a UAV transmits energy signals to charge
architecture. Moreover, it was shown that the harvesting energy multiple mobile users and provides computation services
can be further improved by optimizing the trajectory of the for them. Although the computation performance is limited
UAV [18]-[20]. Thus, it is envisioned that the application of the by the flight time of the UAV, it is worth studying UAV-
UAV-enabled architecture into wireless powered MEC systems enabled wireless powered MEC systems since these systems
is promising and valuable to be studied [26]. However, to the are promising in environments such as mountains and desert
authors’ best knowledge, few investigations have focused on areas, where no terrestrial wireless infrastructures exist, and
this area. in environments where the terrestrial wireless infrastructures
Recently, the UAV-enabled MEC systems have been studied are destroyed due to the natural disasters [33], [34]. Thus, in
and their resource allocation schemes have been proposed this paper, the weighted sum computation bits of all users
[21]-[25]. In [21], the UAV-enabled MEC architecture was are maximized under both partial and binary computation
first proposed and the computation performance was improved offloading modes. The main contributions of this work are
by using UAV. The authors in [22] proposed a new caching summarized as follows:
UAV framework to help small cells to offload traffic. It was 1) It is the first time that the resource allocation framework
shown that the throughput can be greatly improved while the is formulated in UAV-enabled MEC wireless powered
overload of wireless backhaul can be significantly reduced. systems under both partial and binary computation of-
In order to further improve the computation performance, floading modes. The weighted sum computation bits are
the authors in [23] and [24] designed a resource allocation maximized by jointly optimizing the CPU frequencies,
scheme that jointly optimizes the CPU frequency and the the offloading times and the transmit powers of users as
trajectory of the UAV. In [25], a theoretical game method was well as the UAV trajectory. Under the partial computa-
applied to design a resource allocation scheme for the UAV- tion offloading mode, a two-stage alternative algorithm
enabled MEC system and the existence of Nash Equilibrium is proposed to solve the non-convex and challenging
was demonstrated. computation bits maximization problem. The closed-
Although resource allocation problems have been well stud- form expressions for the optimal CPU frequencies, the
ied in MEC systems [5]-[10], MEC systems relying on energy offloading times and the transmit powers of users are
harvesting [11]-[16] and UAV-enabled MEC systems [21]- derived for any given trajectories.
[25], few investigations have been conducted for designing 2) Under the binary computation offloading mode, the
resource allocation schemes in the UAV-enabled wireless pow- weighted sum computation bits maximization problem
ered MEC systems. Moreover, resource allocation schemes is a mixed integer non-convex optimization problem, for
proposed in the above-mentioned works are inappropriate to which a three-stage alternative algorithm is proposed.
UAV-enabled MEC wireless powered systems since the com- The optimal selection scheme on whether users choose
putation performance not only depends on the optimization to locally compute or offload tasks is derived in a closed-
of energy, communication and computation resources, but form expression for a given trajectory. The structure for
also relies on the design of the UAV trajectory. Furthermore, the optimal selection scheme shows that whether users
the application of UAV into wireless powered MEC systems choose to locally compute or offload their tasks to the
has the potential to enhance the user computation capability UAV for computing depends on the tradeoff between
since it can improve the energy conversion efficiency and task the achievable computation rate and the operation cost.
offloading efficiency [33], [34]. Thus, in order to improve the Moreover, the trajectory of the UAV is optimized by us-
computation performance and provide mobile users with high ing the successive convex approximation (SCA) method
QoE, it is of great importance and worthiness to study resource under both partial and binary computation offloading
allocation problems in UAV-enabled wireless powered MEC modes.
systems. However, these problems are indeed challenging to 3) The simulation results show that the computation per-
tackle. The reasons are from two aspects. On one hand, there formance obtained by using the proposed resource allo-
exists dependence among different variables (e.g., the CPU cation scheme is better than these achieved by using the
frequency, the task offloading time and the variables related disjoint optimization schemes. Moreover, it only takes
to the trajectory of the UAV), which makes the problems several iterations for the proposed alternative algorithms
non-convex. On the other hand, when the binary computation to converge. Furthermore, simulation results verify that
offloading mode is applied, the resource allocation problems the priority and fairness of users can be improved by
 trajectory [18]-[20]. A finite time horizon with duration T is
considered. During T , the UAV flies at the same altitude level
denoted by H (H > 0). In practice, the fixed altitude is the
minimum altitude that is appropriate to the work terrain and
can avoid building without the requirement of frequent aircraft
descending and ascending. A block fading channel model is
z
applied, i.e., during each T , the channel remains static.
For the ease of exposition, the finite time T is discretized
Energy
harvesting into N equal time slots, denoted by n = 1, 2, · · · , N . At the
User1
UseM nth slot, it is assumed that the horizontal plane coordinate of
y Local
computing the UAV is qu [n] = [xu [n], yu [n]]. Similar to [27]-[32], it
User2 Usem is assumed that the wireless channel between the UAV and
Usem
each user is dominated by LOS. Thus, the channel power gain
x
Wireless powered link
between the UAV and the mth user, denoted by hm [n], can
Computation offloading link
be given as
β0
hm [n] = β0 d−2
m,n = 2,m ∈ M, n ∈ N ,
Fig. 1: The system model. H2 + kqu [n] − qm k
(1)
where β0 is the channel power gain at a reference distance
using the weight vector. Additionally, it is shown that
d0 = 1 m; dm,n is the horizontal plane distance between
the total computation bits increase with the number of
the UAV and the mth user at the nth slot, n ∈ N , N =
users.
{1, 2, · · · , N }; k·k denotes its Euclidean norm. The details for
The remainder of this paper is organized as follows. Section the UAV-enabled wireless powered MEC system are presented
II gives the system model. The resource allocation problem under partial and binary computation offloading modes in the
is formulated under the partial computation offloading mode following, respectively.
in Section III. Section IV formulates the resource allocation
problem under the binary computation offloading mode. Sim- A. Partial Computation Offloading Mode
ulation results are presented in Section V. Finally, our paper Under the partial computation offloading mode, the com-
is concluded in Section VI. putation task of each user can be partitioned into two parts,
one for local computing and one for offloading to the UAV.
II. S YSTEM M ODEL The energy consumed for local computing and task offloading
A UAV-enabled wireless powered MEC system is consid- comes from the harvested energy. In this paper, in order to
ered in Fig. 1, where an RF energy transmitter and an MEC shed meaningful insights into the design of a UAV-enabled
server are implemented in UAV. The UAV transmits energy to wireless powered MEC system, similar to [4], [13]-[16], the
M users and provides MEC services for these users. Each user linear energy harvesting model is applied. Thus, the harvested
has an energy harvesting circuit and can store energy for its energy Em [n] at the mth user during n time slots is given as
operation. The UAV has an on-board communication circuit n
X T η0 hm [i] P0
and an on-board computing processor. So does each user. Em [n] = , m ∈ M, n ∈ N , (2)
N
The computing processor of each user is an on-chip micro- i=1
processor that has low computing capability and can locally where η0 denotes the energy conservation efficiency, 0 < η0 ≤
execute simple tasks. The UAV has a powerful processor that 1 and P0 is the transmit power of the UAV. In this paper,
can perform computation-intensive tasks [21]-[25]. Similar the UAV employs a constant power transmission [18]-[20].
to [13]-[16], each user can simultaneously perform energy The details for the operation of each user under the partial
harvesting, local computing and computation offloading while computation offloading mode are presented as follows.
the UAV can simultaneously transmit energy and perform 1) Local Computation: Similar to [14]-[16], the energy
computation. In this paper, all devices are equipped with a harvesting circuit, the communication circuit, and the compu-
single antenna. tation unit are all separate. Thus, each user can simultaneously
Without loss of generality, a three-dimensional (3D) Eu- perform energy harvesting, local computing, and computation
clidean coordinate is adopted. Each user’s location is fixed on offloading. Let C denote the number of CPU cycles required
the ground. The location of the mth ground user is denoted by for computing one bit of raw data at each user. In order
qm , where qm = [xm , ym ], m ∈ M and M = {1, 2, · · · , M }. to efficiently use the harvested energy, each user adopts a
Boldface lower case letters represent vectors and boldface dynamic voltage and frequency scaling technique and then can
upper case letters represent matrices. xm and ym are the hori- adaptively control the energy consumed for performing local
zontal plane coordinates of the mth ground user. It is assumed computation by adjusting the CPU frequency during each time
that user positions are known to the UAV for designing the slot [14]-[16]. The CPU frequency of the mth user during
 T one has
M
The first slot The second slot .. . The n th slot .. . The N th slot
X
tm [n] ≤ 1, n ∈ N . (4)
T m=1
N
User1 UAV .. . UserM UAV . . . UAV User1 .. . UAV UserM
Since the energy consumed for local computing and task
Offloading Offloading Download Download offloading comes from the harvested energy, the following
t1 1 .. .
tM 1 0 0 energy harvesting causal constraint should be satisfied.
n n
Fig. 2: The TDMA protocol for multiuser computation offload- T X 3
 η0 T X
γc fm [k] + tm [k] Pm [k] ≤ hm [k] P0 ,
ing. N N
k=1 k=1
n ∈ N , m ∈ M.
(5)
the nth slot is denoted by fm [n] with a unit of cycles per
second. Thus, the total computation bits executed at the mth Under the partial computation offloading mode, the total
user during n slots and the total consumed energy at the mth computation bits Rm of the mth user is given as
n
P T fm [k]
user during n slots are respectively given as NC and N  
k=1
X T fm [n] BT tm [n] hm [n] Pm [n]
n Rm = + log2 1 + ,
P 3
γc fm [k] [14]-[16], where γc is the effective capacitance n=1
NC νm N σ02
k=1
coefficient of the processor’s chip at the mth user, n ∈ N , m ∈ M. (6)
m ∈ M. Note that γc is dependent of the chip architecture of
B. Binary Computation Offloading Mode
the mth user.
Under the binary computation offloading mode, the compu-
2) Computation Offloading: In order to avoid interference
tation task cannot be partitioned. All the users need to choose
among users during the offloading process, a TDMA protocol
to either locally compute the task completely or offload the
shown in Fig. 2 is applied. Specifically, each time slot consists
entire task. This case can be widely experienced in practice.
of three stages, namely, the offloading stage, the computation
For example, in order to improve the estimation accuracy,
stage, and the downloading stage. In the offloading stage, M
the raw data samples that are correlated need to be jointly
users offload their respective computation task one by one
computed altogether [10], [16]. Let M0 and M1 denote the set
during each slot. Let tm [n] × T /N (0 ≤ tm [n] ≤ 1) denote
of users that choose to perform local computation and the set
the duration in which the mth user offloads its computation
of users that choose to perform task offloading, respectively.
task to the UAV at the nth slot, n ∈ N , m ∈ M. Similar
Thus, M = M0 ∪ M1 and M0 ∩ M1 = Θ, where Θ denotes
to [16], the computation task of the mth user to be offload
the null set.
is composed of raw data and communication overhead, such
as the encryption and packer header. Let νm Rm [n] denote 1) Users Choosing to Perform Local Computing: In this
the total number of bits that the mth user offloads to the UAV case, a user in M0 exploits all the harvested energy to perform
during the nth slot, where Rm [n] is the number of raw data to local computing. Thus, the total computation rate of the ith
be computed at the UAV and νm indicates the communication user denoted by RiL can be given as
overhead included in the offloading task. Thus, one has N
X T fi [n]
RiL = , i ∈ M0 . (7)
  NC
BT tm [n] hm [n] Pm [n] n=1
Rm [n] ≤ log2 1 + ,
νm N σ02 And the energy harvesting causal constraint for a user in M0
n ∈ N , m ∈ M, (3) can be given as
n n
T X 3 η0 T X
where B is the communication bandwidth; Pm [n] is the γc fi [k] ≤ hi [k] P0 , n ∈ N , i ∈ Mi . (8)
N N
transmit power of the mth user at the nth slot and σ02 denotes k=1 k=1

the noise power at the mth user. 2) Users Choosing to Perform Task Offloading: Each user
After all users offload their computation tasks at the nth in M1 exploits all the harvested energy to perform task of-
slot, the UAV performs computing task and sends the com- floading. The TDMA protocol is applied to avoid interference
puting results back to all the users. Similar to [14]-[16], the among these users during the offloading process. Since the
computation time and the downloading time of the UAV are total offloading time of all users in M1 at the nth slot cannot
neglected since the UAV has a much stronger computation exceed the duration of a time slot, one has
capability than the users and the number of the bits related to X
the computation result is very small. Since the total offloading tj [n] ≤ 1, n ∈ N . (9)
time of all users does not exceed the duration of one time slot, j∈M1
 Let RjO denote the total computation rate of the jth user in constraint imposed on each user; C2 represents the energy
the set M1 . Then, one has harvesting causal constraint; C3 is the time constraint that the
N   total time of all users offloading the computation bits cannot
O
X BT tj [n] hj [n] Pj [n] exceed the duration of each time slot; C4 and C5 are the
Rj = log2 1 + , j ∈ M1 .
n=1
νj N σ02 speed constraint and the initial and final horizontal location
(10) constraint of the UAV, respectively. P1 is non-convex since
The energy harvesting causal constraint for a user in M1 can there exist non-linear couplings among the variables, fm [n],
be given as Pm [n],qu [n], tm [n] and the objective function is non-concave
with respect to the trajectory of the UAV. In order to solve it, a
n n
T X η0 T X two-stage alternative optimization algorithm is proposed. The
tj [k] Pj [k] ≤ hj [k] P0 , n ∈ N , j ∈ M1 .
N N details for the algorithm are presented as follows.
k=1 k=1
(11) B. Two-Stage Alternative Optimization Algorithm
Sections III and IV will respectively formulate the compu- Let zm [n] = tm [n] Pm [n] , n ∈ N . For a given trajectory,
tation rate maximization problem for the partial and binary P1 can be transformed into P2 .
computation offloading modes. M
X
III. R ESOURCE A LLOCATION U NDER T HE PARTIAL P2 : max wm
fm [n],zm [n],tm [n]
m=1
C OMPUTATION O FFLOADING M ODE " #
N  
In this section, the resource allocation problem is studied
X T fm [n] BT tm [n] hm [n] zm [n]
× + log2 1 +
under the partial computation offloading mode. The weighted n=1
NC νm N tm [n] σ02
sum computation bits are maximized by jointly optimizing the (13a)
CPU frequencies, the offloading times and the transmit powers s.t. C1, C3, (13b)
of users as well as the trajectory of the UAV. In order to tackle n n
this non-convex problem, a two-stage alternative algorithm is T X  3
 η0 T X
C5 : γc fm [k] + zm [k] ≤ hm [k] P0 ,
proposed. N N
k=1 k=1

A. Resource Allocation Problem Formulation m ∈ M, n ∈ N . (13c)

Under the partial computation offloading mode, the It is easy to prove that P2 is convex and can be solved by using
weighted sum computation bits maximization problem in the the Lagrange duality method [35], based on which the optimal
UAV-enabled wireless powered MEC system is formulated as solutions for the CPU frequency and the transmit power can
opt opt
P1 , be derived. Let fm [n] and Pm [n] denote the optimal CPU
M frequency and transmit power of the mth user at the nth time
slot, respectively, where m ∈ M and n ∈ N . By solving P2 ,
X
P1 : max wm ×
fm [n],Pm [n],qu [n],tm [n]
m=1 Theorem 1 can be stated as follows.
" N  # Theorem 1: For a given trajectory qu [n], the optimal CPU
X T fm [n] BT tm [n] hm [n] Pm [n]
+ log2 1 + frequency and transmit power of users can be respectively
n=1
NC νm N σ02 expressed as
(12a) v
opt
u wm
s.t. C1 : fm [n] ≥ 0, Pm [n] ≥ 0, m ∈ M, n ∈ N , (12b) fm [n] = u u N
, (14a)
n n
P
3Cγc λm,k
t
T X 3
 η0 T X
C2 : γc fm [k] + tm [k] Pm [k] ≤ hm [k] P0 k=n
N N 
k=1 k=1
 0, if tm [n] = 0,
m ∈ M, n ∈ N , (12c) +

 
opt
M Pm [n] = wm B σ02
X   N
− hm [n]
 , otherwise,
C3 : tm [n] ≤ 1, n ∈ N , (12d)
 P
 νm ln 2 λm,k
k=n
m=1
(14b)
2 T
C4 : kqu [n + 1] − qu [n]k ≤ Vmax ,n ∈ N, (12e) where λm,n ≥ 0 is the dual variable associated with the
N +
C5 : qu [1] = q0 , qu [N + 1] = qF , (12f) constraint C2; [a] = max (a, 0) and max (a, 0) denotes the
bigger value of a and 0.
where Vmax denotes the maximum speed of the UAV in the Proof: See Appendix A.
unit of meter per second; q0 and qF are the initial and final Remark 1: It can be seen from Theorem 1 that users choose
horizontal locations of the UAV, respectively. In (12), wm to offload their computation tasks only when the channel state
denotes the weight of the mth user, which takes the priority information between users and the UAVNis stronger than a
and the fairness among users into consideration. C1 is the CPU
threshold, namely, hm [n] ≥ σ02 νm ln 2
P
frequency constraint and the computation offloading power λm,k / (wm B).
k=n
 l,opt
This indicates that the user chooses to perform local com- where fm [n], zm l,opt [n], and tl,opt
m [n] denote the optimal
putation when the horizontal distance between the user and solutions at the lth iterations. According to [35], the subgra-
β0 wm B
the UAV is larger than N
− H 2 . Moreover, dient guarantees to converge to the optimal value with a very
σ02 νm ln 2
P
λm,k small error range.
k=n
it can be seen that the larger the weight is, the higher the
chance for the user to chooses to offload its computation task.
Furthermore, users prefers to offload their computation task C. Trajectory Optimization
when the local
q computation frequency is very large, namely,
σ02 νm ln 2 For any given CPU frequency, transmit power, and offload-
opt
fm [n] ≥ 3Cγc Bhm [n] . ing time of users, the trajectory optimization problem can be
opt
Theorem 2: If there exists a time slot that fm [n] = 0, the formulated as P3 .
opt
equation fm [k] = 0 must hold, 0 ≤ k ≤ n.
Proof: Since λm,n is the dual variable and λm,n ≥ 0, M
X
opt P3 : max wm
from Theorem 1 fm [n] increases with n. Thus, if there exists
opt opt qu [n]
a time slot n so that fm [n] = 0, one must have fm [k] = 0,  m=1
 
for 0 ≤ k ≤ n. Theorem 2 is proved. N
BT t [n] β P [n]
m 0 m
X
Remark 2: Theorem 2 indicates that the user CPU frequency × log2 1 +   
ν m N 2 2
σ0 H + kqu [n] − qm k
2
increases with the time slot index. This means that the number n=1
of computation bits obtained by local computing increases (18a)
with the time slot index. Moreover, the user CPU frequency T X
n
3

increases with the weight assigned to that user since more s.t. C2 : γc fm [k] + tm [k] Pm [k]
N
resources are allocated to the user with a higher weight. k=1
n
Theorem 3: For a given trajectory qu [n], the optimal user η0 T X β 0 P0
offloading time can be obtained by solving the following ≤ 2 , m ∈ M, n ∈ N (18b)
N H 2 + kqu [k] − qm k
k=1
equation.
  C4 and C5. (18c)
hm [n] zm [n] hm [n] zm [n]
log2 1 + −
σ02 tm [n] ln 2 {σ02 tm [n] + hm [n] zm [n]} Since C2 is non-convex and the objective function is non-
νm N αn concave with respect to qu [n], P3 is non-convex and we use
− = 0. (15)
BT the SCA technique to solve the optimization problem. The
Remark 3: Theorem 3 can be readily proved based on the obtained solutions can be guaranteed to satisfy the Karush-
proof for Theorem 1. Thus this proof is omitted for the sake Kuhn-Tucker (KKT) conditions of P3 [27]. By using the SCA
of saving space. Moreover, (15) can be solved by using the technique, Theorem 4 is given as follows.
bisection method [35]. Theorem 4: For any local trajectory qu, [n] , n ∈ N at the
The values of the dual variables are needed in order to obtain th iteration, one has
the optimal CPU frequency, the optimal transmit power and the
n
optimal offloading time for all users. The subgradient method X P0 β 0
in Lemma 1 can be used to tackle this problem [36]. 2 ≥ P0 β0 hm [n] , (19a)
i=1 H2
+ kqu [i] − qm k
Lemma 1: The subgradient method for obtaining the dual  
variables is given as n  2 2
X H + 2kqu, [i] − qm k − kqu [i] − qm k 
 2 
+
hm [n] =  2
λm,n (l + 1) = [λm,n (l) − θ (l) ∆λm,n (l)] , m ∈ M, n ∈ N 
i=1  H 2 + kqu, [i] − qm k
2 

(16a)
+
(19b)
αn (l + 1) = [αn (l) − ϑ (l) ∆αn (l)] , n ∈ N , (16b)

where l denotes the iteration index; θ (l) and ϑ (l) represent where the equality holds when qu [n] = qu, [n].
a
the iterative steps at the lth iteration. In (16), ∆λm,n (l) and Proof: Let f (z) = b+z , where a and b are positive
∆αn (l) are the corresponding subgradients, given as constants, and z ≥ 0. Since f (z) is convex with respect to z,
the following inequality can be obtained:
n
η0 T X
∆λm,n (l) = hm [k] P0 a a a
N ≥ −
k=1
b+z b + z0 2 (z − z0 ) , (20)
n (b + z0 )
T Xh l,opt

3 i
− γc fm [k] + zm l,opt [k] , (17a)
N where z0 is a given local point. By using (20), Theorem 4 is
k=1
M
X proved.
∆αn (l) = 1 − tl,opt
m [n], n ∈ N , (17b) In order to tackle the objective function of P3 , Lemma 2 is
m=1 given as follows.
 Lemma 2: [27] Using the SCA method, the following optimization algorithm denoted by Algorithm 1 is further
inequality can be obtained, developed to solve P1 . The details for Algorithm 1 can be
  found in Table I. In Table I, Ri denotes the value of the
β0 Pm [n] objective function of P1 at the ith iteration.
log2 1 +    ≥ ym, ({qu [n]}) ,
2 2 2
σ0 H + kqu [n] − qm k IV. R ESOURCE A LLOCATION IN B INARY C OMPUTATION
(21a) O FFLOADING M ODE
 
β0 Pm [n] In this section, the weighted sum computation bits max-
ym, ({qu [n]}) = log2 1 +   imization problem is studied in the UAV-enabled wireless
2 2 2
σ0 H + kqu, [n] − qm k powered MEC system under the binary computation offloading
β0 Pm [n] log2 e mode. The CPU frequencies of the users that choose to
−   perform local computation, the offloading times, the transmit
2 2
σ02 H 2 + β0 Pm [n] + σ02 kqu, [n]k H 2 + kqu, [n]k powers of users that choose to perform task offloading, the
trajectory of the UAV, and the mode selection are jointly
 
2 2
× kqu [n]k − kqu, [n]k , (21b)
optimized to maximize the weighted sum computation bits
where the equality holds when qu [n] = qu, [n]. of all users. The formulated problem is a mixed integer
non-convex optimization problem, for which a three-stage
TABLE I: Two-stage alternative optimization algorithm alternative optimization problem is proposed.
A. Resource Allocation Problem Formulation
Algorithm 1: The two-stage alternative optimization algorithm
1: Setting:
Under the binary computation offloading mode, the
P0 , T , N , Vmax , q0 , qF , and the tolerance errors ξ, ξ1 ; weighted sum computation bit maximization problem subject
2: Initialization: to the energy harvesting causal constraints, the UAV speed and
The iterative number i = 1, λim,n , αin and qiu [n];
position constraints is formulated as P5 ,
3: Repeat 1:
opt,i opt,i N
calculate fm [n] and Pm [n] using Theorem 1 X X fi [n] T
for given qiu [n]; P5 : max wi
use the bisection method to solve (20) and obtain ti,optm [n]; fi [n],Pj [n],q[n], CN
tj [n],M0 ,M1 i∈M0 n=1
update λim,n and αin using the subgradient algorithm;
initialize the iterative number j = 1; X wj BT XN  
Repeat 2:
hj [n] Pj [n]
+ tj [n] log2 1 + (23a)
solve P4 by using CVX for the given fm opt,i opt,i
[n], Pm [n] νj N n=1 σ02
i,opt
j∈M1
and tm [n]; n n
update j = j + 1, and qju [n]; T X η0 T X
N
s.t. γc fi3
[k] ≤ hi [k] P0 , n ∈ N , i ∈ M0 ,
if
P
qju [n] − qj−1 [n] ≤ ξ
N N
u k=1 k=1
n=1
(23b)
qiu [n] = qju [n] ;
n n
break; T X η0 T X
end tj [k] Pj [k] ≤ hj [k] P0 , n ∈ N , j ∈ M1 ,
end Repeat 2 N N
k=1 k=1
update the iterative number i = i + 1; (23c)
if Ri − Ri−1 ≤ ξ1 X
break; tj [n] ≤ 1, n ∈ N , (23d)
end
j∈M1
end Repeat 1
4: Obtain solutions: M = M0 ∪ M1 , M0 ∩ M1 = Θ, (23e)
opt opt
fm [n], Pm [n] and topt opt
m [n] and qu [n].
fi [n] ≥ 0, Pj [n] ≥ 0, i ∈ M0 , j ∈ M1 , (23f)
C4 and C5. (23g)
Using Theorem 4 and Lemma 2, P3 can be solved by
iteratively solving the approximate problem P4 , given as (23b) and (23c) are the energy harvesting causal constraints
"N # imposed on these users who choose to perform local computa-
M X BT tm [n] ym, ({qu [n]})
X tion and on these users who choose to perform task offloading,
P4 : max wm (22a)
qu [n] νm N respectively; (23d) is the offloading time constraint during
m=1 n=1
each slot and (23e) is the user operation selection constraint.
s.t. C4 and C5, (22b)
In P5 there exist close couplings among different optimiza-
n
X  3
 tion variables. Furthermore, the binary user operation mode
γc fm [k] + tm [k] Pm [k] ≤ η0 P0 β0 hm [n],
selection makes P5 a mixed integer programming problem.
k=1
The exhaustive search method leads to a prohibitively high
m ∈ M, n ∈ N . (22c)
computational complexity, especially when there exist a large
It can be seen that P4 is convex and can be readily solved by number of users. Motivated by how we solve P1 , P5 has a
using CVX [4]. By solving P2 and P4 , a two-stage alternative similar structure as P1 when the operation modes of users
are determined. Thus, the optimal CPU frequency, transmit where υm,n ≥ 0 and εn ≥ 0 are the dual variables associated
power, and offloading time of users can be obtained by using with the constraints given by (24b) and (24c), respectively.
the same method as the one used for P1 and the trajectory Proof: See Appendix B.
optimization for the UAV can also be achieved by using the Remark 4: Theorem 5 indicates that the user operation
SCA method. As such, a three-stage alternative optimization selection scheme depends on the tradeoff between the achiev-
algorithm is proposed based on the two-stage Algorithm 1. able computation rate and the operation cost. If the tradeoff
The details for the algorithm are presented as follows. of the user achieved by local computing is better than that
B. Three-Stage Alternative Optimization Algorithm obtained by task offloading, the user chooses to perform
local computing; otherwise, the user chooses to offload its
In order to efficiently solve P5 , a binary variable denoted computation tasks to the UAV for computing.
by ρm is introduced, where ρm ∈ {0, 1} and m ∈ M.
Finally, the trajectory optimization for any given ρm , fm [n],
ρm = 0 indicates that the mth user performs local computation
Pm [n] and tm [n] can be obtained by solving P7 , given as
mode while ρm = 1 means that the mth user performs task
offloading. Moreover, the user operation selection indicator M
"N
#
X X BT tm [n] ym, ({qu [n]})
variable ρm is relaxed as a sharing factor ρm ∈ [0, 1]. Thus, P7 : max wm ρm
P5 can be rewritten as qu [n]
m=1 n=1
νm N
M X N  (26a)
X fm [n] T
P6 : max wm (1 − ρm ) s.t. C4 and C5, (26b)
fm [n],Pn [n],q[n],
m=1 n=1
CN n n
tm [n],ρm X X
3
  (1 − ρm ) γc fm [k] + ρm tm [k] Pm [k]
BT tm [n] ρm hm [n] Pm [n]
+ log2 1 + (24a) k=1 k=1
νm N σ02
n n
≤ η0 P0 β0 hm [n], m ∈ M, n ∈ N , (26c)
T X 3 T X
s.t. (1 − ρm ) γc fm [k] + ρm tm [k] Pm [k]
N N where hm [n] and yj ({qu [n]}) are given by (19b) and (21b),
k=1 k=1
n respectively. P7 is convex and can be efficiently solved by
η0 T X
≤ hm [k] P0 , m ∈ M, (24b) using CVX [4]. Based on Theorem 1, Theorem 5 and the so-
N lutions of P7 , a three-stage alternative optimization algorithm
k=1
M
X denoted by Algorithm 2 is proposed to solve P5 . The details
ρm tm [n] ≤ 1, n ∈ N , (24c) for Algorithm 2 are presented in Table 2. In Table 2, Rl and
m=1 Ri denote the value of the objective function of P5 at the lth
fm [n] ≥ 0, Pm [n] ≥ 0, n ∈ N , m ∈ M, (24d) and i iteration, respectively.
C4 and C5. (24e)
C. Complexity Analysis
Even by relaxing the binary variable ρm , P6 is still difficult
to solve as there exist couplings among different variables. The complexity of Algorithm 1 comes from four aspects.
For any given ρm and the trajectory of the UAV, P6 has The first aspect is from the computation of the CPU frequency
a similar structure as P1 . Thus, using the same techniques and the offloading power. The second aspect is from the
applied to P1 , the optimal CPU frequency, transmit power bisection method for obtaining the offloading time. The third
and offloading time of users for a given ρm and the UAV aspect is from the subgradient method for computing the
trajectory can be obtained. It is easy to verify that the optimal dual variables. The fourth aspect comes from the application
CPU frequency, transmit power and offloading time of users of CVX for solving P4 . Let L1 and L2 denote the num-
for a given trajectory have the same forms given by Theorem ber of iterations required for the outer loop and the inner
1 and Theorem 3. loop of Algorithm 1, respectively. Let `1 and `2 denote the
Theorem 5: For any given fm [n], Pm [n], tm [n] and qu [n], tolerance error for the bisection method and the subgradi-
the user operation selection scheme can be obtained by ent method, respectively. Thus, according to the works in

0 if G1 ≥ G2 , [35],
 [38] and [39], the total complexity of Algorithm 1 is
opt
ρm = (25a)


1 otherwise; O L1 2M N + M log2 (`1 /T ) + 1/`22 + L2 N 3 and O (·)
N
( n
) is the big-O notation [35].
X wm fm [n] X
3 The complexity of Algorithm 2 comes from five aspects.
G1 = − υm,n γc fm [k] , (25b)
n=1
C Four aspects are the same as these of Algorithm 1. The fifth
k=1
N    aspect is from the computation of the operation selection
X Btm [n] hm [n] Pm [n] indicator variable ρm . Let L1 , L2 and L3 denote the number
G2 = log2 1 +
n=1
νm σ02 of iterations required for the first, second and third loop of
Algorithm 2, respectively. Similar to the complexity analysis
n
)
X N
−υm,n tm [k] Pm [k] − εn tm [n] , (25c) for Algorithm 1, the total complexity of Algorithm
 2 is
T
k=1 O L1 L2 2M N + M + M log2 (`1 /T ) + 1/`22 + L3 N 3 .
 TABLE II: Three-stage alternative optimization algorithm TABLE III: Simulation Parameters
Parameters Notation Typical Values
Algorithm 2: The three-stage alternative optimization algorithm
Numbers of Users M 4
1: Setting:
The height of the UAV H 10 m
P0 , T , N , Vmax , q0 , qF , and the tolerance errors ξ, ξ1 and ξ2 ;
The time length of the UAV flying T 2 sec
2: Initialization:
i Numbers of CPU cycles C 103 cycles/bit
The iterative number i = 1, υm,n and εin , and qiu [n];
Energy conversation efficiency η0 0.8
3: Repeat 1: Communication bandwidth B 40 MHz
initialize the iterative number l = 1 and ρlm ; The receiver noise power σ02 10−9 W
Repeat 2: The number of time slots N 50
opt,i opt,i
calculate fm [n] and Pm [n] using Theorem 1 The effective switched capacitance γc 10−28
for given qiu [n] and ρopt,l
m ; The channel power gain β0 −50 dB
use the bisection method to solve (20) and obtain ti,opt m [n]; The tolerance error ξ, ξ1 10−4
i
update υm,n and εin using the subgradient algorithm; The initial position of the UAV q0 [0, 0]
calculate ρopt,l using Theorem 5 and update l = l + 1; The final position of the UAV qF [10, 0]
m
The maximum speed of the UAV Vmax 20 m/s
if Rl − Rl−1 ≤ ξ
break;
end
initialize the iterative number j = 1; 10

Repeat 3:
opt,i opt,i 9
solve P7 by using CVX for the given fm [n], Pm [n], User 2 (0,10)
i,opt opt,l
tm [n] and ρm ; 8 User 3 (10,10)
update j = j + 1, and qju [n];
N 7
qju [n] − qj−1
P
if u [n] ≤ ξ
n=1 6 A semi−circle
qiu [n] = qju [n] ; The offloading mode

y(m)
break; 5
end
end Repeat 3 4

update the iterative number i = i + 1; The binary mode

if Ri − Ri−1 ≤ ξ1 3

break;
2
end User 1 (0,0) User 4 (10,0)
A constant speed
end Repeat 2 1
end Repeat 1
4: Obtain solutions: 0
opt opt 0 1 2 3 4 5 6 7 8 9 10
fm [n], Pm [n] and topt opt opt
m [n], ρm and qu [n]. x (m)

Fig. 3: The trajectory of the UAV under different schemes with

T = 2 seconds.
V. S IMULATION R ESULTS

In this section, simulation results are presented to compare the operation modes. The reason is that the weights of user 2
the performance of our proposed designs with that of other and user 3 are larger than these of user 1 and user 4. Thus,
benchmark schemes. The convergence performance of the the UAV needs to fly close to user 2 and user 3 so as to
proposed algorithms is also evaluated. The simulation settings provide more energy to them. This indicates that the priority
are based on the works in [7], [14], [16] and [23]. The and the fairness among users can be obtained by using the
positions of users are set as: q1 = [0, 0], q2 = [0, 10], weight vector.
q3 = [10, 10], q4 = [10, 0]. The detailed settings are given Fig. 4 shows the weighted sum computation bits of all users
in Table III. The weight vector of each user [w1 w2 w3 w4 ] versus the transmit power of the UAV under different schemes.
is set as [0.1 0.4 0.3 0.2]. The optimal local computing is the mode that all users
Fig. 3 shows the UAV trajectory under different schemes only perform local computing while the optimal offloading
with T = 2 seconds. The UAV transmit power is set as mode is that all users only perform task offloading. And the
P0 = 0.1 W. In the constant speed scenario, the UAV flies trajectory of the UAV is jointly optimized under these two
straight with a constant speed from the initial position to the benchmark schemes. The results under the binary mode and
final position. In the semi-circle scenario, the UAV flies along the partial offloading mode are obtained by using Algorithm
the trajectory that is a semi-circle with its diameter being 2 and Algorithm 1, respectively. In Fig. 4 the weighted sum
kqF − q0 k. The trajectory of the offloading mode is obtained computation bits achieved under the partial offloading mode is
by using Algorithm 1 for the partial computation offloading the largest among these obtained by other schemes. The reason
mode and the trajectory of the binary mode is obtained by is that all the users can dynamically select the operation mode
using Algorithm 2 for the binary computation offloading mode. based on the quality of the channel state information under the
It can be seen from the trajectories of our proposed schemes partial computation offloading mode. Moreover, the optimal
the UAV is always close to user 2 and user 3, irrespective of offloading mode outperforms the optimal local computing.
 7 7
x 10 x 10
4.5 4.5

The weighted sum computation bits of all users (bits)

The weighted sum computation bits of all users (bits)

The partial offloading mode
4 Optimal local computing 4 The partial offloading mode with the semi−circle trajectory

Optimal offloading The partial offloading mode with a constant speed

3.5 The binary mode

3.5
The partial offloading mode

3
3

2.5

2.5

2
1.5

1.5
1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
The transmit power of the UAV (W)
1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 4: The weighted sum computation bits of all users versus The transmit power of the UAV (W)
the transmit power of the UAV under different schemes. x 10
7

4.5

The weighted sum computation bits of all users (bits)

The binary mode with joint optimization
This result is consistent with the results obtained in [13]. 4
The binary mode with the semi−circle trajectory
Furthermore, the weighted sum computation bits of all users
The binary mode with a constant speed
increase with the UAV transmit power. It can be explained by 3.5

the fact that the harvesting energy increases with the transmit
power of the UAV. Thus, users have more energy to perform 3
local commutating or task offloading.
Fig. 5 shows the weighted sum computation bits of all the
2.5
users versus the transmit power of the UAV under different
trajectories with the partial computation offloading mode and
2
the binary computation offloading mode. As shown in Fig. 5,
the weighted sum computation bits of all the users achieved
by using our proposed schemes are larger than that obtained 1.5

by using the trajectory with a constant speed and than that

obtained by using the semi-circle trajectory, irrespective of the 1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
operation modes. This indicates that the optimization of the The transmit power of the UAV (W)
trajectory of the UAV can improve the weighted sum computa-
(a) (b)
tion bits. It also verifies that our proposed resource allocation
scheme outperforms the disjoint optimization schemes. Fig. 5: (a) The weighted sum computation bits of all users
Fig. 6 shows the total computation bits of each user under versus the transmit power of the UAV under different trajec-
different operation modes. The transmit power of the UAV is tories with the partial computation offloading mode; (b) The
set as P0 = 0.1 W. The total computation bits of user 2 and weighted sum computation bits of all users versus the transmit
user 3 are higher than those of user 1 and user 4. The reason power of the UAV under different trajectories with the binary
is that the weights of user 2 and user 3 are larger than those computation offloading mode.
of user 1 and user 4. Thus, the resource allocation scheme
should consider the priority of user 2 and user 3. This further
verifies that the application of the weight vector can improve computation offloading mode are larger than those obtained
the priority and also the fairness of users. under the binary computation offloading mode. The reason is
Fig. 7 is given to verify the efficiency of our proposed that users can simultaneously perform local computing and
Algorithm 1 and Algorithm 2. The transmit power of the task offloading when the channel state information is strong
UAV is given as 0.1 W or 0.2 W. The results show that under the partial computation offloading mode. However, users
Algorithm 1 and Algorithm 2 only need several iterations to can only perform either local computing or task offloading in
converge. This indicates that the proposed Algorithm 1 and the binary offloading mode even when the channel state infor-
Algorithm 2 are computationally effective and have a fast mation is strong. The computation performance is improved
convergence rate. It can also be seen that the weighted sum by the flexible selection of the operation mode based on the
computation bits of all the users achieved under the partial channel state information.
 6 7
x 10 x 10
7 4

The weighted sum computation bits of all users (bits)

The binary mode
The total compuation bits of each user (Bits)
6
3.5
The partial offloading mode

5
3

2.5

2 The partial offloading mode, P0=0.4 W

2
The binary mode, P0=0.4 W

1.5 The partial offloading mode, P0=0.2 W

1
The binary mode, P0=0.2 W

0 1
User 1 User 2 User 3 User 4 2 3 4 5 6 7 8 9 10 11 12
Users The number of users

Fig. 6: The total computation bits of each user under different Fig. 8: The weighted sum computation bits of all users versus
operation modes with P0 = 0.1 W. the number of users under different transmit powers of the
UAV and different operation modes.
7
x 10
2.8
The weighted sum computation bits of all users (bits)

Core(TM) i7-4790 CPU, 8 GB RAM. From Table IV we

2.6
can see that the required run time of Algorithm 1 is smaller
than that of Algorithm 2. This indicates that the complexity
2.4 P =0.2 W, Algorithm 1
0
of Algorithm 1 is lower than that of Algorithm 2. It can be
P0=0.2 W, Algorithm 2
verified by the complexity analysis presented in Subsection C
2.2 P =0.1 W, Algorithm 1
0 of Section IV. Moreover, the effect of the number of time slots
P =0.1 W, Algorithm 2
0 on the run time is larger than that of the number of users. The
2
reason is that the complexity of these two algorithms mainly
depends on the number of time slots. This can also be verified
1.8
by the complexity analysis.
1.6 VI. C ONCLUSIONS
The resource allocation problems were studied for UAV-
1.4
2 4 6 8 10 12 14 16 18 20 enabled wireless powered MEC systems under both the par-
The number of iterations
tial and binary computation offloading modes. The weighted
Fig. 7: The weighted sum computation bits of all users versus sum computation rates of users were maximized by jointly
the number of iterations required by using Algorithms 1 and optimizing the CPU frequencies, the user offloading times,
2 under different transmit powers of the UAV and different the user transmit powers, and the UAV trajectory Two alter-
operation modes. native algorithms were proposed to solve these challenging
problems. The closed-form expressions for the optimal CPU
frequencies, user offloading times, and user transmit power
Fig. 8 shows the weighted sum computation bits of all users were derived. Moreover, the optimal selection scheme whether
versus the number of users under different operation modes. users choose to locally compute or offload tasks was proposed
The transmit power of the UAV is set as P0 = 0.2 W or for the binary computation offloading mode. It was shown
P0 = 0.4 W. In Fig. 8 the weighted sum computation bits of that the performance achieved by using our proposed resource
all users increase with the number of users. The reason is that allocation scheme is superior to these obtained by using the
more users can exploit the harvesting energy to perform local disjoint optimization schemes. Simulation results also verified
computing and computation offloading. It is also observed that the efficiency of our proposed alternative algorithms and our
the growth rate decreases with the increase of the number of theoretical analysis.
users. The reason is that the offloading time allocated for each The exploitation of UAV to improve the energy conversation
user decreases with the increase of the number of users since efficiency and the computation performance was studied in this
the total offloading time is limited by T . paper. However, the computation performance is also limited
Table IV is given to evaluate the run times of Algorithm 1 by the flight time of the UAV. It is interesting to exploit
and Algorithm 2 shown in the top of the next page. The run multiple antennas techniques to tackle this challenge. This will
times are obtained by using a computer with 64-bit Intel(R) be investigated in our future work.
 `TABLE IV: Comparison of the required run time of Algorithm 1 with that of Algorithm 2 (s)
` (N, M ) (50, 2) (50, 4) (50, 8) (60, 2) (60, 4) (60, 8) (70, 2) (70, 4) (70, 8)
```
Algorithms ````
Algorithm 1 43.72 104.54 186.38 154.74 198.65 235.85 224.74 291.53 352.72
Algorithm 2 89.35 167.17 265.46 223.19 275.42 321.87 308.56 388.92 468.39

N
A PPENDIX A $m,n =
P
υm,k . Thus, according to [37], the optimal ρm
P ROOF OF T HEOREM 1 k=n
denoted by ρopt be obtained by (35) at the tope of the next
m can
Let λm,n and αn denote the dual variables associated with
page. Based on (35), since zm [n] = tm [n] Pm [n], Theorem
the constraint C2 and C3, respectively, where λm,n ≥ 0 and
5 is proved.
αn ≥ 0. Then, the Lagrangian of P2 can be given by (27) at
the tope of this page, where Ξ denotes a collection of all the
N
P R EFERENCES
primal and dual variables related to P2 . Let µm,n = λm,k
k=n [1] F. Zhou, Y. Wu, R. Q. Hu, Y. Wang, and K. K. Wong, “Energy-
3
and gm [k] = η0 hm [k] P0 − γc fm [k] − zm [k]. Then, the efficient NOMA enabled heterogeneous cloud radio access networks,”
Lagrangian function L (Ξ) can be rewritten by (28) at the IEEE Network, vol. 32, no. 2, pp.152-160, 2018.
[2] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey
tope of this page. And the Lagrangian dual function of P2 on mobile edge computing: The communication perspective,” IEEE
can be presented as Commun. Surveys Tuts., vol. 19, no. 4, pp. 2322-2358, Fourth Quarter,
2017.
g (λm,n , αn ) = max L (Ξ). (29) [3] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks
0≤fm [n]
with RF energy harvesting: A contemporary survey,” IEEE Commun.
Surveys Tuts., vol. 17, pp. 757-789, Second Quarter, 2015.
Based on (29), the optimal solutions of P2 can be obtained
[4] F. Zhou, Z. Li, J. Cheng, Q. Li, and J. Si, “Robust AN-aided beamform-
by solving its dual problem, given as ing and power splitting design for secure MISO cognitive radio with
SWIPT,” IEEE Trans. Wireless Commun., vol. 16, no. 4, pp. 2450-2464,
min g (λm,n , αn ) . (30) April 2017.
λm,n ,αn
[5] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of radio
and computational resources for multicell mobile-edge computing,”
It can be seen from (30) that the dual problem can be IEEE Trans. Signal Inf. Process. Over Netw., vol. 1, no. 2, pp. 89-103,
decoupled into M independent optimization problems, given Jun. 2015.
by (31) at the tope of the next page. Thus, let the derivation [6] C. You, K. Huang, H. Chae, and B. Kim, “Energy-efficient resource al-
of (31b) with respect to fm [n] andzm [n] be zero, one has location for mobile-edge computation offloading,” IEEE Trans. Wireless
Commun., vol. 16, no. 3, pp. 1397-1411, Mar. 2017.
2 N [7] C. Wang, C. Liang, F. R. Yu, Q. Chen, L. Tang, “Computation offloading
T wm 3T γc fm [k] X and resource allocation in wireless cellular networks with moible edge
− λm,k = 0, (32a)
NC N computing,” IEEE Trans. Wireless Commun., vol. 16, no. 8, pp. 4924-
k=n 4938, Aug. 2017.
N [8] J. Du, L. Zhao, J. Feng, and X. Chu, “Computation offloading and
wm BT tm [n] hm [n] T X
− λm,k =0. resource allocation in mixed fog/cloud computing systems with min-
νm N ln 2 σ02 tm [n] + hm [n] zm [n] N max fairness guarantee,” IEEE Trans. Commun., vol. 66, no. 4, pp.
k=n
1594-1608, April 2018.
(32b) [9] L. Liu, Z. Chang, X. Guo, S. Mao, and T. Ristaniemi, “Multi-objective
optimization for computation offloading in fog computing,” IEEE Inter-
Note that zm [k] = tm [k] Pm [k] and Pm [k] ≥ 0. Moreover, net Things J., vol. 5, no. 1, pp. 283-294, Jan. 2018.
the case that tm [n] = 0 can be identified as Pm [n] = 0. Thus, [10] W. Zhang, Y. Wen, K. Guan, D. Kilper, H. Luo, and D. O. Wu, “Energy-
based on (32), Theorem 1 is proved. The proof for Theorem optimal mobile cloud computing under stochastic wireless channel,”
IEEE Trans. Wireless Commun., vol. 12, no. 9, pp. 4569-4581, Sep.
1 is complete. 2013.
[11] J. Xu, L. Chen, and S. Ren, “Online learning for offloading and
A PPENDIX B autoscaling in energy harvesting mobile edge computing,” IEEE Trans.
P ROOF OF T HEOREM 5 Cogn. Netw., vol. 3, no. 3, pp. 361-373, Sep., 2017.
[12] Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloading
Let υm,n and εn denote the dual variables with respect to for mobile-edge computing with energy harvesting devices,” IEEE J.
the constraints given by (24b) and (24c), respectively, where Sel. Areas Commun., vol. 34, no. 12, pp. 3590-3605, Dec., 2016.
υm,n ≥ 0 and εn ≥ 0. Then, for any given fm [n], Pm [n], [13] C. You, K. Huang, and H. Chae, “Energy efficient mobile cloud
computing powered by wireless energy transfer,” IEEE J. Sel. Areas
tm [n] and qu [n], the Lagrangian of P6 can be expressed Commun., vol. 34, no. 5, pp. 1757-1771, May, 2016.
by (33) at the tope of the next page, where Ξ1 denotes a [14] F. Wang, J. Xu, X. Wang, and S. Cui, “Joint offloading and computing
collection of all the primal and dual variables related to P6 . optimization in wireless powered mobile-edge computing systems,”
Ξ2 denotes a collection of υm,n , αn , fm [n] , zm [n], tm [n] and IEEE Trans. Wireless Commun., vol. 17, no. 3, pp. 1784-1797, March
2018.
ρm . Using the same techniques that are used for the proof of [15] S. Mao, S. Leng, K. Yang, X. Huang, and Q. Zhao, “Fair energy-efficient
Theorem 1, for any given fm [n], zm [n], tm [n] and qu [n], scheduling in wireless powered full-duplex mobile-egde computing
P6 can be solved by solving M independent optimization systems,” in Proc. IEEE Global Commun. Conf., Singapore, 2017.
[16] S. Bi and Y. Zhang, “Computation rate maximization for wireless
problems, given by (34) at the tope of the next page, where powered mobile egde computing with binary computation offloading,”
3
`m [n] = η0 hm [n] P0 − (1 − ρm ) γc fm [n] − ρm zm [n] and IEEE Trans. Wireless Commun., to be published, 2018.
 M N
"  #
X X fm [n] T BT tm [n] hm [n] zm [n]
L (Ξ) = wm + log2 1 +
m=1 n=1
C N νm N σ02 tm [n]
M X N
( n n
) N
( M
)
X η0 T X T X 3
 X X
+ λm,n hm [k] P0 − γc fm [k] + zm [k] + αn 1 − tm [n] , (27)
m=1 n=1
N N n=1 m=1
k=1 k=1

M X
N   
X T fm [n] BT tm [n] hm [n] zm [n] αn
L (Ξ) = wm + log2 1 + +µm,n gm [k] + − αn tm [n]. (28)
m=1 n=1
N C νm N σ02 tm [n] M

max Lm (λm,n , αn , fm [n] , zm [n] , tm [n]) (31a)

λm,n ,αn ,fm [n]≥0

Lm (λm,n , αn , fm [n] , zm [n] , tm [n]) (31b)

N     
X T fm [n] BT tm [n] hm [n] zm [n] αn
= wm + log2 1 + +µm,n gm [n] + − αn tm [n] . (31c)
n=1
N C νm N σ02 tm [n] M

M
" N  #
X X fm [n] T BT ρm tm [n] hm [n] zm [n]
L1 (Ξ1 ) = wm (1 − ρm ) + log2 1 +
m=1 n=1
C N νm N tm [n] σ02
M X N
( n n
)
X η0 T X T X 3

+ υm,n hm [k] P0 − (1 − ρm ) γc fm [k] + ρm zm [k]
m=1 n=1
N N
k=1 k=1
N
( M
)
X X
+ εn 1 − ρm tm [n] , (33)
n=1 m=1

max L1m (Ξ2 ) (34a)

υm,n ,εn ,fm [n]≥0
N   
X T (1 − ρm ) fm [n] BT ρm tm [n] hm [n] zm [n]
L1m (Ξ2 ) = wm + log2 1 +
n=1
NC νm N tm [n] σ02
N
X εn
+ $m,n `m [n] + − εn tm [n], (34b)
n=1
M


< 0, ρopt
m = 0,
∂L1m(Ξ2 ) 
= 0, 0 < ρopt
m < 1, m∈M (35a)
∂ρopt
m 
> 0, ρopt = 1;
m
(N )
1

∂Lm (Ξ2 ) X wm fm [n] T BT tm [n] hm [n] zm [n]
= − + log2 1 +
∂ρopt
m n=1
C N νm N tm [n] σ02
N
( n
) N
X T X 3
 X
+ υm,n − −γc fm [k] + zm [k] − εn tm [n]. (35b)
n=1
N n=1
k=1
[17] H. Wang, J. Wang, G. Ding, L. Wang, T. A. Tsiftsis, P. K. Sharma, Fuhui Zhou received the Ph. D. degree from Xidian
“Resource allocation for energy harvesting-powered D2D communica- University, Xian, China, in 2016. He is an associate
tion underlaying UAV-assisted networks,” IEEE Trans. Cogn. Netw., vol. Professor with School of Information Engineering,
2, no. 1, pp. 14-24, Jan. 2018. Nanchang University. He is now a Research Fellow
[18] S. Yin, J. Tan, and L. Li, “UAV-assisted cooperative communications at Utah State University. He has worked as an
with wireless information and power transfer,” submitted to IEEE Trans. international visiting Ph. D student of the University
Wireless Commun., https://arxiv.org/abs/1710.00174v1. of British Columbia from 2015 to 2016. His research
[19] J. Xu, Y. Zeng, and R. Zhang, “UAV-enabled wireless power transfer: interests focus on cognitive radio, green communi-
Trajectory design and energy region charaterization,” in Proc. IEEE cations, edge computing, machine learning, NOMA,
Global Commun. Conf. Singapore, 2017, physical layer security, and resource allocation. He
[20] J. Xu, Y. Zeng, and R. Zhang, “UAV-enabled wireless power transfer: has published more than 40 papers, including IEEE
Trajectory design and energy optimization,” IEEE Trans. Wireless Com- Journal of Selected Areas in Communications, IEEE Transactions on Wireless
mun., to be published, 2018. Communications, IEEE Wireless Communications, IEEE Network, IEEE
[21] N. H. Motlagh, M. Bagaa, and T. Taleb, “UAV-based IoT platform: A GLOBECOM, etc. He has served as Technical Program Committee (TPC)
crowd surveillance use case,” IEEE Commun. Mag., vol. 55, no. 2, pp. member for many International conferences, such as IEEE GLOBECOM,
128-134, Feb. 2017. IEEE ICC, etc. He serves as an Associate Editor of IEEE Access.
[22] N. Zhao, F. Cheng, F. R. Yu, J. Tang, Y. Chen, G. Gui, and H. Sari,
“Caching UAV assisted secure transmission in hyper-dense networks
based on interference alignment,” IEEE Trans. Commun., vol. 66, no. 5,
pp. 2281-2294, May 2018.
[23] S. Jeong, O. Simeone, and J. Kang, “Mobile edge computing via a UAV- Yongpeng Wu (S’08–M’13–SM’17) received the
mounted cloudlet: Optimization of bit allocation and path planning,” B.S. degree in telecommunication engineering from
IEEE Trans. Vehicular Technol., vol. 67, no. 3, pp. 2049-2063, Mar. Wuhan University, Wuhan, China, in July 2007,
2018. the Ph.D. degree in communication and signal pro-
[24] S. Jeong, O. Simeone, and J. Kang, “Mobile edge computing with a cessing with the National Mobile Communications
UAV-mounted cloudlet: Optimal bit allocation for communication and Research Laboratory, Southeast University, Nanjing,
computation,” IET Commmn., vol. 11, no. 7, pp. 969-974, Nov. 2017. China, in November 2013.
[25] M. A. Messous, H. Sedjelmaci, N. Houari, and S. M. Senouci, “Compu- Dr. Wu is currently a Tenure-Track Associate
tation offloading game for an UAV network in mobile egde computing,” Professor with the Department of Electronic Engi-
in Proc. IEEE Int. Conf. Commuun., France, May, 2017. neering, Shanghai Jiao Tong University, China. Pre-
[26] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications with un- viously, he was senior research fellow with Institute
manned aerial vehicles: Opportunities and challenges,” IEEE Commun. for Communications Engineering, Technical University of Munich, Germany
Mag., vol. 54, no. 5, pp. 36-42, May 2016. and the Humboldt research fellow and the senior research fellow with Institute
[27] Y. Zeng and R. Zhang, “Energy-efficient UAV communication with for Digital Communications, University Erlangen-Nürnberg, Germany. During
trajectory optimization,” IEEE Trans. Wireless Commun., vol. 16, no. his doctoral studies, he conducted cooperative research at the Department
6, pp. 3747-3760, June 2017. of Electrical Engineering, Missouri University of Science and Technology,
[28] P. Yang, X. Cao, C. Yin, Z. Xiao, X. Xi, and D. Wu, “Proactive drone- USA. His research interests include massive MIMO/MIMO systems, physical
cell deployment: Overload relief for a cellular network under flash crowd layer security, signal processing for wireless communications, and multivariate
traffic,” IEEE Trans. Intell. Transportation Sys., vol. 18, no. 10, pp. statistical theory.
2877-2892, Oct, 2017. Dr. Wu was awarded the IEEE Student Travel Grants for IEEE Inter-
[29] E. Kalantari, H. Yanikomeroglu, and A. Yongacoglu, “On the number national Conference on Communications (ICC) 2010, the Alexander von
and 3D placement of drone base stations in wireless cellular networks,”in Humboldt Fellowship in 2014, the Travel Grants for IEEE Communication
Proc. IEEE VTC fall, 2016, Montreal, Canada, Sept. 2016. Theory Workshop 2016, and the Excellent Doctoral Thesis Awards of China
[30] E. Kalantari, M. Z. Shakir, H. Yanikomeroglu, and A. Yongacoglu, Communications Society 2016. He was an Exemplary Reviewer of the IEEE
“Backhaul-aware robust 3D drone placement in 5G+ wireless net- Transactions on Communications in 2015, 2016. He is the lead guest editor
works,”in Proc. ICC Workshops, 2017, Paris, France, May 2017. for the upcoming special issue “Physical Layer Security for 5G Wireless
[31] L. Zeng, X. Cheng, C. X. Wang, and X. Yin, “A 3D geometry-based Networks” of the IEEE Journal on Selected Areas in Communications. He is
stochastic channel model for UAV-MIMO channels,” in Proc. IEEE currently an editor of the IEEE Access and IEEE Communications Letters. He
WCNC 2017, San Francisco, USA, Mar. 2017. has been a TPC member of various conferences, including Globecom, ICC,
[32] C. X. Wang, A. Ghazal, B. Ai, P. Fan, and Y. Liu, “Channel mea- VTC, and PIMRC, etc.
surements and models for high-speed train communication systems: a
survey,” IEEE Commun. Surveys Tuts., vol. 18, no. 2, pp. 974-987, 2nd
Quart., 2016.
[33] N. Cheng, W. Xu, W. Shi, Y. Zhou, N. Lu, H. Zhou, and X. Shen,
“Aire-ground integrated mobile edge networks: Architecture, challenges Rose Qingyang Hu is a Professor of Electri-
and opportunities,” IEEE Commun. Mag., to be published, 2018. cal and Computer Engineering Department at Utah
[34] M. Mozaffari, W. Saad, M. Bennis, Y. H. Nam, and M. Debbah, State University. She received her B.S. degree from
“A tutorial on UAVs for wireless networks: Applications, challenges, University of Science and Technology of China,
and open problems,” IEEE Commun. Surveys Tuts., submitted, 2018. her M.S. degree from New York University, and
https://arxiv.org/abs/1803.00680 her Ph.D. degree from the University of Kansas.
[35] S. P. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, She has more than 10 years of R&D experience
U.K.: Cambridge Univ. Press, 2004. with Nortel, Blackberry and Intel as a technical
[36] F. Zhou, N. C. Beaulieu, Z. Li, J. Si, and P. Qi, “Energy-efficient manager, a senior wireless system architect, and
optimal power allocation for fading cognitive radio channels: Ergodic a senior research scientist, actively participating in
capacity, outage capacity and minimum-rate capacity,” IEEE Trans. industrial 3G/4G technology development, standard-
Wireless Commun., vol. 15, no. 4, pp. 2741-2755, Apr. 2016. ization, system level simulation and performance evaluation. Her current
[37] F. Zhou, Z. Li, J. Cheng, Q. Li, and J. Si, “Robust max-min fairness research interests include next-generation wireless communications, wireless
resource allocation in sensing-based wideband cognitive radio with system design and optimization, green radios, Internet of Things, Cloud
SWIPT: Imperfect channel sensing,” IEEE Syst. J., to be published, 2017. computing/fog computing, multimedia QoS/QoE, wireless system modeling
[38] S. Bubeck, “Convex optimization: Algorithms and complexity,” In and performance analysis. She has published over 180 papers in top IEEE
Foundations and Trends in Machine Learning, vol. 8, no. 3, pp. 231-357, journals and conferences and holds numerous patents in her research areas.
2015. https://arxiv.org/abs/1405.4980 Prof. Hu is an IEEE Communications Society Distinguished Lecturer Class
[39] C. Gutierrez, F. Gutierrez, M.C. Rivara, “Complexity on the bisection 2015-2018 and the recipient of Best Paper Awards from IEEE Globecom
method,” Theoretical Computer Science, vol. 382, pp. 131-138, 2007. 2012, IEEE ICC 2015, IEEE VTC Spring 2016, and IEEE ICC 2016.
 Yi Qian received a Ph.D. degree in electrical engi-
neering from Clemson University. He is a professor
in the Department of Electrical and Computer En-
gineering, University of Nebraska-Lincoln (UNL).
Prior to joining UNL, he worked in the telecommu-
nications industry, academia, and the government.
Some of his previous professional positions include
serving as a senior member of scientific staff and a
technical advisor at Nortel Networks, a senior sys-
tems engineer and a technical advisor at several start-
up companies, an assistant professor at University
of Puerto Rico at Mayaguez, and a senior researcher at National Institute of
Standards and Technology. His research interests include information assur-
ance and network security, network design, network modeling, simulation and
performance analysis for next generation wireless networks, wireless ad-hoc
and sensor networks, vehicular networks, smart grid communication networks,
broadband satellite networks, optical networks, high-speed networks and the
Internet. Prof. Yi Qian is a member of ACM and a senior member of IEEE.
He was the Chair of IEEE Communications Society Technical Committee for
Communications and Information Security from January 1, 2014 to December
31, 2015. He is a Distinguished Lecturer for IEEE Vehicular Technology
Society and IEEE Communications Society. He is serving on the editorial
boards for several international journals and magazines, including serving as
the Associate Editor-in-Chief for IEEE Wireless Communications Magazine.
He is the Technical Program Chair for IEEE International Conference on
Communications (ICC) 2018.