Charging Task Scheduling For Directional Wireless Charger Networks

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.2997602, IEEE
Transactions on Mobile Computing
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1
Charging Task Scheduling for

Directional Wireless Charger Networks
Haipeng Dai, Member, IEEE, Ke Sun, Alex X. Liu, Fellow, IEEE, Lijun Zhang, Member, IEEE, Jiaqi Zheng,
and Guihai Chen, Member, IEEE
Abstract—This paper studies the problem of cHarging tAsk Scheduling for direcTional wireless chargEr networks (HASTE), i.e., given
a set of rotatable directional wireless chargers on a 2D area and a series of offline (online) charging tasks, scheduling the orientations
of all the chargers with time in a centralized offline (distributed online) fashion to maximize the overall charging utility for all the tasks.
We prove that HASTE is NP-hard. Then, we prove that a relaxed version of HASTE falls within the realm of maximizing a submodular
function subject to a partition matroid constraint, and propose a centralized offline algorithm that achieves (1 − ρ)(1 − 1e )
approximation ratio to address HASTE where ρ is the switching delay of chargers. Further, we propose a distributed online algorithm
and prove it achieves 12 (1 − ρ)(1 − 1e ) competitive ratio. We conduct simulations and field experiments on a testbed consisting of 8
off-the-shelf power transmitters and 8 rechargeable sensor nodes. The results show that our distributed online algorithm achieves
92.97% of the optimal charging utility, and outperforms the comparison algorithms by up to 15.28% in terms of charging utility.
Index Terms—Charging task, scheduling, directional wireless chargers.
1 I NTRODUCTION beamforming), and thus enhances the power intensity in the

intended directions [8]. For this reason, directional antennas
1.1 Motivation and Problem Statement
for WPT are widely adopted in applications such as mil-
The last decade has witnessed the rapid development of limeter wave cellular networks [9]–[12], wireless recharge-
Wireless Power Transfer (WPT) technology, which enjoys able sensor networks [13], and wireless charging systems
huge advantages such as no contact, reliable power supply, adopting the simultaneous wireless information and power
and ease of maintenance compared to traditional wired transfer technology [14], [15], and are also studied in [16]–
power supply technologies. WPT technology has numerous [18]. Further, static chargers are more preferable than mobile
applications, including wireless identification and sensing chargers in some scenarios. First, using static chargers is a
platform (WISP) [1], wireless rechargeable sensor networks more robust and timely way to handle unexpected arrived
[2], electric vehicles [3], solar power satellites [4], and wire- charging tasks in an online manner, such as urgent charging
less powered drone aircraft [5], etc. As per the record pro- requests caused by accidental energy depletion of existing
vided by Wireless Power Consortium, an organization ded- sensor nodes or new nodes join, than using mobile chargers,
icated to promote standardization of WPT, the number of because mobile chargers may need to travel a long distance
registered WPT products from its 214 member companies, for handling tasks. Second, static chargers can also serve
including IT leaders Samsung, Philips, LG, and Huawei, has as data collectors, which allows fast and efficient data col-
surged to 848 [6]. By a recent report, 35% of consumers in lection than using mobile chargers. Third, it is more cost-
the United States have used WPT products [7]. efficient for some applications where, for example, sensor
Directional wireless charger network, which consists of nodes form multiple clusters with long distance between
static directional wireless chargers, is one of the critical them. Moreover, from a long term view, purchasing wireless
topics for WPT technology. To begin with, it is well-known chargers is a one-time investment and can be amortized over
that directional charging is more energy efficient than time, while using mobile chargers usually require much
omnidirectional charging. Unlike omnidirectional charging higher energy cost and human cost than maintaining static
which broadcasts the electromagnetic waves equally in all chargers, and such cost constantly accumulates over time.
directions regardless of the locations of the rechargeable Fourth, there have emerged a lot of on-the-shelf products
devices, directional charging concentrates radiated energy based on wireless power transfer technologies [19]–[21],
in the directions of the rechargeable devices (i.e., energy and they offer solutions for popular applications such as
charging at coffee shops, security systems, smart home, and
• H. Dai, K. Sun, A. X. Liu, L. Zhang, J. Zheng, and G. Chen are with the
Department of Computer Science and Technology, Nanjing University,
in-vehicle charging. These applications require dedicated
China. static chargers.
E-mail: {haipengdai, zlj, gchen}@nju.edu.cn, kesun@smail.nju.edu.cn, In this paper, we consider the problem of cHarging
alexliu360@gmail.com, jiaqi369@gmail.com tAsk Scheduling for direcTional wireless chargEr networks
• Alex X. Liu is a no-pay Adjunct Professor of the Department of Computer (HASTE) aiming for maximizing the overall charging utility
Science and Technology at Nanjing University. of offline/online charging tasks. We adopt the directional
Corresponding author: H. Dai (E-mail: haipengdai@nju.edu.cn). Manuscript charging model for wireless chargers and rechargeable de-
received August 29, 2018. vices, which captures the characteristics of power transmit-
Corresponding Author: Haipeng Dai

1536-1233 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Exeter. Downloaded on June 18,2020 at 15:33:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2020.2997602, IEEE
ters and receivers equipped with directional antennas. In wireless charging task scheduling issue for omnidirectional
this model, the power charging area for a charger and the wireless chargers in offline scenarios, which are fundamen-
power receiving area for a device are modeled as sectors. tally different from our paper. In the conference version of
A rechargeable device can be charged via wireless by a this paper [51], we initiated the first study on scheduling
charger with non-zero power if and only if they are located wireless charging tasks for directional wireless chargers and
in each other’s covered sector. All wireless chargers can designing online algorithms.
freely adjust its orientation in [0, 2π) while rechargeable
devices cannot. Moreover, a charging task initiated by a 1.3 Key Technical Challenges
rechargeable device consists of five elements: the position
We are faced with three major challenges to address HASTE.
and orientation of its associated device, the release time and
The first challenge is that HASTE is non-linear and is NP-
end time of the task, and its required charging energy. To
hard. HASTE is nonlinear because that the orientation of
evaluate the effectiveness of wireless charging for a task, we
chargers can be freely scheduled; a task can be either cov-
define the task’s charging utility as a linear and bounded
ered by a charger and have a certain constant power incre-
function with its harvested energy from its release time to
ment or not with no power increment, which has the flavor
its end time.
of 0-1 integer programming; the charging utility function
With these models, we consider two scenarios for charg-
is linear but bounded, let alone that we extend our results
ing task scheduling, i.e., offline and online. In the offline sce-
to the case where the utility function is a general concave
nario, information for all charging tasks is known a priori,
function. In addition, by reducing from the classical NP-
and thereby the scheduling policies for all chargers at any
hard separate assignment problem, we prove that HASTE is
moment can be determined beforehand. To accommodate
NP-hard.
practical concerns, we assume that each charger needs an
The second challenge is how to design an efficient cen-
amount of time for switching its orientation, which we
tralized offline algorithm for HASTE in the offline scenario
call switching delay. In the online scenario, charging tasks
while considering the switching delay of chargers. The
stochastically arrive, and chargers reschedule their orien-
switching delay happens if and only if a charger’s next
tations in realtime. Nevertheless, in addition to switching
intended orientation is different from its current orientation,
delay, each charger needs an additional amount of time for
which implies that the switching delay as well as its caused
recomputing the scheduling policies with negotiating with
performance loss is history-dependent. Moreover, the per-
neighboring chargers, which we call rescheduling delay. To
formance loss is difficult to evaluate as there are potentially
avoid global management effort and reduce update cost,
multiple tasks are affected by a charger’s switching delay,
we desire a distributed and local algorithm which is scal-
and the charging utility function for tasks is non-linear.
able with network size. For both scenarios, we want to
The third challenge is how to design an efficient dis-
dynamically schedule the orientations of chargers as time
tributed online algorithm for HASTE in the online scenario
goes on such that the overall weighted charging utility for
where all chargers are asynchronous and the rescheduling
all charging tasks is maximized. Moreover, we stress that
delay needs to be considered. To the best of our knowledge,
chargers can be either in the working mode for the offline
there are neither existing distributed online algorithms di-
scenario or in that for the online scenario, but cannot switch
rectly applicable to our problem even when the reschedul-
between these two different statuses. To sum up, we state
ing delay is omitted, nor existing online algorithms that deal
our problem HASTE as follows. Given a set of rotatable
with the case in our considered scenario with rescheduling
directional wireless chargers on a 2D area and a series of
delay being concerned for which the response is delayed
offline (online) charging tasks, scheduling the orientations of
and the algorithm is not truly “online”.
all the chargers with time in a centralized offline (distributed
online) fashion to maximize the overall charging utility for
all the tasks. 1.4 Proposed Approach
To address the first challenge, we propose that rather than
considering all possible orientations in [0, 2π) for chargers,
1.2 Prior Art we can safely consider a limited number of orientations for
On one hand, there exist numerous literatures [22]–[37] them without causing performance loss, and therefore, ex-
studying on the mobile charging case where one single tract the so-called “dominant task sets” as the corresponding
or multiple chargers travel in a field to charge wireless sets of covered tasks. Then, we neglect the switching delay
rechargeable devices to guarantee their normal working. for wireless chargers, and thus reformulate the original con-
They are fundamentally different from ours as we consider tinuous optimization problem into a discrete optimization
static chargers. problem HASTE-R. Further, we prove that the reformulated
On the other hand, the other works consider wireless problem is exactly a problem of maximizing a submodular
charger networks consisted of static wireless chargers, but function subject to a partition matroid constraint, which
nearly none of them investigate charging task scheduling. greatly facilitates approximation algorithm design.
In particular, most of them focus on scheduling issues in To address the second challenge, based on the theoret-
coarse granularity rather than task levels, such as those ical results obtained by addressing the first challenge, we
overlooking the harmful effect of high electromagnetic ra- can either use a simple greedy algorithm that achieves 12
diation (EMR) [16], [17], [38]–[41] and those taking the approximation ratio [52] or a randomized algorithm with
EMR safety into consideration [42]–[48]. To the best of our the optimal approximation guarantee, namely, 1− 1e approx-
knowledge, there is only one work [49], [50] investigate imation ratio [53]. Nevertheless, as the former is not good
enough and the latter is too computationally demanding, we working time in a charging time period [30], [31]. We refer
tailor the TABULARGREEDY algorithm proposed in [54], readers to survey [56] for more related works.
[55] to address HASTE-R as it can achieve an approximation Second, the other works are dedicated to wireless
ratio between 12 and 1 − 1e (1 − 1e as default in our setting) charger networks consisted of static wireless chargers, but
depending on the value of a control parameter and resulting nearly none of them consider charging task scheduling.
in different time complexity. Further, to bound the perfor- First, most of them study scheduling issues in coarse gran-
mance loss of switching delay, we exploit the concavity of ularity rather than task levels. On one hand, some works
the utility function and consider all the caused performance (e.g., [16], [17], [38]–[41]) overlook the detrimental effect of
loss for all impacted tasks in the worst case, and prove that the electromagnetic radiation (EMR) to human health. For
the switching delay introduces a constant factor of 1 − ρ in instance, He et al. considered the triangular deployment
the ultimate achieved approximation ratio, i.e., (1−ρ)(1− 1e ), problem of wireless chargers [38]. They attempted to min-
of the proposed algorithm, where ρ is the switching delay. imize the number of chargers while rechargeable tags can
To address the third challenge, we propose a distributed receive sufficient power. In addition, we first proposed the
online algorithm based on the proposed centralized offline directional charging problem based on empirical experimen-
algorithm to HASTE. We prove that if the rescheduling tal results, and investigated the ominidirectional charging
delay is neglected, as long as the local executions of a problem using directional chargers in [16], the wireless
charger and its neighbors are in order and repeat regularly charger placement problem for directional charging in [17],
with time, the achieved global charging utility is the same [39]–[41]. On the other hand, other literatures [42]–[48] take
as that of the centralized offline algorithm. Further, by the EMR safety into consideration, and guarantee that the
leveraging the concavity of the utility function and the sub- EMR intensity at any point in the area does not exceed a
modularity of the objective function in HASTE, we bound predefined EMR threshold. For instance, we presented and
the performance loss of scheduling delay, and prove that studied how to schedule non-adjustable chargers [42], [43]
our distributed online algorithm achieves 21 (1 − ρ)(1 − 1e ) and adjustable chargers [44], [45] to maximize the charging
competitive ratio. utility for chargers under the EMR safety constraint. Niko-
letseas et al. [46] considered more practical constraints such
as the energy limitations of chargers and devices, the non-
1.5 Evaluation Results
linear constraints in the time domain, and their goal is to
We conducted simulations and field experiments to evalu- optimize the amount of energy transferred from chargers to
ate our proposed algorithms. Our simulation results show devices and truly utilized. Moreover, we reported a wireless
that our proposed distributed online algorithm can achieve charger placement scheme that ensures EMR safety in [47].
92.97% of the optimal charging utility which corroborates Second, to the best of our knowledge, there is only one work
our theoretical findings, outperform the other two compari- [49], [50] that study the wireless charging task scheduling.
son algorithms by 10.96%. We implemented our algorithms Nevertheless, [49], [50] consider omnidirectional wireless
on a testbed consisting of 8 off-the-shelf TX91501 power chargers whose charging power is adjustable and focus
transmitters produced by [19] and 8 rechargeable sensor on offline scenarios, which are fundamentally different
nodes associated with 8 charging tasks. Our experimental from our paper. Moreover, we launched the first study on
results show that our distributed online algorithm out- scheduling wireless charging tasks for directional wireless
performs the comparison algorithms by up to 15.28% on chargers and designing online algorithms in the conference
average, and 29.63% at most. version of this paper [51].
2 R ELATED W ORK 3 P ROBLEM F ORMULATION

In this section, we briefly review related works regarding 3.1 Preliminaries
wireless charging. Suppose there is a set of directional wireless chargers
First, there exist some literatures focus on mobile charg- S = {s1 , ..., sn } located in a 2D plane Ω , which can contin-
ing scenarios where one single or multiple chargers travel uously rotate with orientation angle within [0 2π). Suppose
in a field to charge rechargeable devices deployed there there are also some rechargeable devices located in Ω , which
to make them work perpetually, which are fundamentally either keep static or dynamically join or leave the wireless
different from ours. [22]–[26] study the charging efficiency charger network. These rechargeable devices launch (wire-
issues of wireless chargers, e.g., Zhang et al. presented an less) charging tasks and sending them to wireless chargers
optimal scheme for multiple mobile chargers to charge a now and then, and the chargers accordingly schedule their
linear WSN while the ratio of truly charged energy to orientations to serve the tasks. Formally, charging tasks are
wasted energy is maximized. [27]–[29] concentrate on re- defined by a five-tuple Tj =< oj , φj , tjr , tje , Ej > where oj
ducing the service delay of mobile chargers, e.g., Fu et al. denotes the position of the rechargeable device that raises
considered the problem of minimizing the overall charging the task, φj is the orientation of the device, tjr and tje are the
delay of a single mobile charger by planning its charging release time and end time of the task, and Ej is required
route and charging strategy [27]. [30]–[37] pay attention charging energy. We adopt a discrete time model for which
to the overall network performance such as data routing, the time is divided into multiple slots with uniform duration
event monitoring, data collection, and task assignment. For Ts . For simplicity, we assume that tjr is exactly at the
instance, Shi et al. proposed to use a single mobile charger beginning of a time slot while tje is at the end of a time slot.
to improve data collection performance and the charger’s We summarize the notations used in this paper in Table 1.
TABLE 1 where α and β are two known constants determined by

Notations and symbols used in this paper hardware parameters of chargers as well as surrounding
environment [16]–[18], ||si oj || is the distance between si
Symbol Description and oj , As and Ao are respectively the charging angle of
si The ith directional wireless charger, or its position chargers and the receiving angle of devices, − r→ −→
θi and rφj
n Number of directional wireless chargers are respectively the unit vectors denoting the orientations
θi (θi (t)) Orientation of charger si (its function with time t) of the charger and the device. Further, if a device oj is
θi,k The value of θi (t) at the kth time slot if charger si is
not switching
covered by more than one directional wireless chargers, its
Tj The jth charging task received power is the sum of the received power from all
oj Position of the rechargeable device that raises charg- chargers [16], [17]. Note that there is another directional
ing task Tj , or the jth rechargeable device charging model proposed in [57], which is more practical
φj Orientation of the rechargeable device that raises
charging task Tj , or the orientation of device oj
as it considers the anisotropic energy receiving property of
tjr (tje ) Release time (end time) of charging task Tj
rechargeable sensors. We plane to consider it in our future
Ej Required charging energy of charging task Tj work.
m Number of charging tasks A charger can either keep its orientation unchanged
As Charging angle of chargers during the same time slot, or switch its orientation in the
Ao Receiving angle of devices
Ts Duration of a time slot starting ρ (0 < ρ < 1) portion of a time slot, which we call
Pr (.) Charging power function switching delay, and keep static in the rest 1 − ρ portion of
α, β Constants in the charging model the time slot. We argue that this assumption makes sense
D Radius of charging/receiving area because typically a charging task can last up to tens of
ρ Switching delay
τ Rescheduling delay
minutes or even more than an hour, the duration of time
U (.) Charging utility function slots can be set to a few minutes, and the switching time
wj Weight of charging task Tj for commercial rotatable heads or cradles [58] on which the
Ti Set of charging tasks that cover charger si chargers are mounted or soft switching of smart antennas
Γi (Γip ) Set of dominant task sets for charger si (the pth of chargers [59], [60] is commonly a few seconds or even
dominant task set in Γi )
p
Γi,k (Γi,k ) Set of dominant task sets for charger si at the kth shorter. We assume that a charger stops emitting power
time slot (the pth dominant task set in Γi,k ) during its switching. For convenience of exposition, we
K Number of considered time slots for all tasks define θi = Ø for a charger during its switching process,
C Number of colors and further define Pr (si , Ø, oj , φj ) = 0. In the offline
N (si ) Neighbors of charger si (two chargers are neighbors
to each other if and only if they cover at least one case, we assume the information for all charging tasks are
charging task in common) known a prior, then the scheduling policies for all time
Ki Number of considered time slots for all tasks ob- slots for each charger are determined beforehand. In the
served by charger si
online case, we assume the charging tasks stochastically
arrive, and chargers recompute their scheduling policies in

an on-the-fly fashion. Especially, we assume each charger
rk needs τ (τ ∈ Z+ ) number of time slots, which we name as
rescheduling delay, for negotiation with neighboring chargers
oj and computation to update its future scheduling policies,
As 
 si ri and then, if necessary, starts switching with a delay of ρ

Ao
r j ok time slot. Typically, the rescheduling delay is expected to be
much less than the duration of charging tasks. In this paper,
we assume the latter is at least two times that of the former,
Fig. 1. Directional charging model (oj can receive power from si while i.e., tje − tjr ≥ 2τ Ts for any task Tj , where Ts is the duration
ok cannot) of a time slot.
We adopt the general and practical directional charging 3.2 Charging Utility Model
model proposed in [16]–[18]. As Figure 1 shows, a charger
We adopt a linear and bounded charging utility model for
si with working orientation denoted by vector −r→
θi can only
harvested energy for a task, which is similar to the charging
charge devices in a charging area in the shape of a sector
utility model for received power proposed in [17]. That is,
with charging angle As and radius D. A rechargeable device
the charging utility for a task is first proportional to the
oj with orientation denoted by vector −r→
φj can only receive
harvested energy of its associated device, and then reaches
non-zero power in a receiving area in the shape of a sector
a constant if the harvested energy exceeds a predetermined
with receiving angle Ao and radius D. The charging power
threshold, i.e.,
from si to oj is given by
( 1
 α
, 0 ≤ ||si oj || ≤ D, Ej
· x, x ≤ Ej

 (||si oj ||+β)2 U(x) = (1)
−−→ −→ 1, x > Ej

si oj · rθ − ksi oj kcos(As /2) ≥ 0,


i
Pr (si , θi , oj , φj ) =


 and −
o−→ −→
j si · rφj − koj si kcos(Ao /2) ≥ 0.
where Ej is the required charging energy of charging task

0, otherwise Tj .

3.3 Problem Formulation and Hardness Analysis Algorithm 1: Dominant Task Sets Extraction
Let θi (t) (θi : R≥0 7→ {[0 2π) ∪ Ø}) be the func- Input: The wireless charger si , all charging tasks
tion of orientation for charger si with time t. Suppose {Tj }mj=1
the value of θi (t) at the kth time slot is θi,k if charger Output: All dominant task sets
m
si is not switching; otherwise, θi (t) is set to Ø and 1 Find the subset of charging tasks in {Tj }j=1 that cover
the charging power of si is zero. Then, for a charg- si , say T i ;

2 Initialize the orientation of the charger to 0;
Pn task Tj , its harvested power at time t is given by
ing
3 Rotate the charger anticlockwise to cover the tasks in
i=1 Pr (si , θi (t), oj , φj ), and jits aggregate harvested en-
R t Pn T i one by one until there is some covered task is
ergy during its whole life is je i=1 Pr (si , θi (t), oj , φj )dt.
tr Pm going to be uncovered. During the rotating process, if
And the overall (weighted) charging utility is j=1 wj · the rotated angle is larger than 2π , then terminate;
R tje Pn
U( tj i=1 Pr (si , θi (t), oj , φj )dt) where wj is the weight of 4 Add the current covered set of tasks to the collection
r
charging task Tj . of dominant task sets;
Our task is to determine the decision variables θi,k 5 Rotate the charger anticlockwise until a new task in
defined in θi (t) for all the chargers so that the overall T i is included in the covered set. During the rotating
charging utility is maximized. With all above, we define process, if the rotated angle is larger than 2π , then
the problem of cHarging tAsk Scheduling for direcTional terminate. If not, goto Line 3.
wireless chargEr networks (HASTE) as follows.
m
X Z tje n
X
(P1) max U= wj · U ( Pr (si , θi (t), oj , φj )dt)
θi,k
j=1
j
tr i=1
4 P ROBLEM R EFORMULATION
In this section, considering the complexity of the formu-
(
 Ø, kTs < t ≤ (k + ρ)Ts

, θi,k 6= θi,k−1
s.t. θi (t) =
lation P1 of HASTE, we reformulate HASTE to make it
θi,k , (k + ρ)Ts < t ≤ (k + 1)Ts
 tractable. In particular, we first propose a dominant task sets
θi,k , kTs < t(k + 1)Ts , otherwise

extraction algorithm for chargers to reduce the continuous
where k ∈ Z+
0 , and θi (0) = Ø solution space for orientations of chargers to a discrete one
0 ≤ θi,k < 2π. with limited choices. Then, we consider a relaxed version
The following theorem shows the complexity of HASTE. of HASTE, i.e., HASTE-R, and prove it falls into the realm
of maximizing a submodular function subject to a partition
Theorem 3.1. HASTE is NP-hard. matroid constraint, which assists the further algorithm de-
sign.
Proof: Due to space limit, we only sketch the proof
here. Suppose ρ → 0, tjr = 0 and tje = Ts for all charging
tasks, which means each task simply occupies the first time 4.1 Extraction of Dominant Task Sets
slot and we only need to consider one round scheduling Though each charger can continuously rotate within [0 2π),
in this time slot. Moreover, suppose the required charging we do NOT need to consider all possible orientations.
energy for each task Ej is so small that as long as a task Instead, we only need to care about the possible sets of
is covered by a charger, it certainly obtains an amount of covered tasks, whose number is obviously limited for any
energy greater then Ej and therefore achieves a charging given charger. Further, among these sets we only need to
utility of wj in the overall charging utility. Besides, though consider the following specific ones.
the orientation of chargers can be freely chosen in [0 2π), its Definition 4.1. (dominant task set) Given a set of tasks T 1i
covered sets of charging tasks can be enumerated in a fixed covered by a charger si with some orientation, if there doesn’t
number of steps and are limited, as we will see in Algo- exist another set of tasks T 2i covered by si with some other
rithm 1. Consequently, with the above settings, our problem orientation such that T 1i ⊂ T 2i , then T 1i is a dominant
changes to choosing the orientation for each charger among task set.
its candidate choices such that the overall charging utility of
We describe our algorithm for extracting dominant task
all tasks is maximized. We can prove this simplified problem
sets in Algorithm 1. Basically, the considered charger rotates
is NP-hard by reducing from the classical NP-hard separate
for 2π and extracts the dominant task sets one by one. We
assignment problem [61], which is defined as follows: given
use a toy example for illustration in our conference version
a set of bins and a set of items to pack in each bin, a
of this paper [51].
value for assigning item j to bin i, and a separate packing
constraint for each bin, i.e., for bin i, a family Ii of subsets of
items that fit in bin i, packing items into bins to maximize 4.2 Problem Relaxation and Reformulation
the aggregate value. Here we can regard each charger as a As the switching delay is hard to be analyzed for optimiza-
bin, each task as an item, each set of covered tasks for a tion, we first consider a relaxed version of HASTE, HASTE-
candidate orientation of charger si as a subset in the family R for short, by neglecting the switching delay of all chargers,
Ii for bin i, the achieved utility of a task as the value for and then analyze HASTE. We will bound the performance
assigning this item to a bin, and therefore, we can reduce loss for the relaxation in our proposed algorithms.
any instance of the separate assignment problem to the Suppose the obtained set of dominant task sets for
considered simplified problem. As the separate assignment p p
charger si is Γi , the pth dominant task set in Γi is Γi . Let xi,k
problem is NP-hard [61], HASTE is also NP-hard. be a binary indicator denoting whether the pth dominant
task set in Γi in the kth time slot is selected or not. For Besides, it can be easily proved that M = {S, I} is a matroid
convenience of expression, we define by verifying the three properties proposed in Definition 4.3.
Thus we have the following lemma.
( α
(||si oj ||+β)2
, 0 ≤ ||si oj || ≤ D,
Pr (si , oj ) =
0, otherwise. Lemma 4.1. The constraint in the scheduling problem RP1 can
p
be written as a partition matroid on the ground set S .
Moreover, we abuse the notation slightly by defining Γi 3
oj as ∃Tj 0 ∈ Γip |Tj 0 .oj 0 = oj . That is, there exists a charing Accordingly, problem RP1 can be rewritten as
p (RP2)
task Tj 0 in Γi and its associated position of rechargeable
device is oj . Then, the problem HASTE-R can be formulated m tje /Ts
X X X
as max f (X) = wj · U ( Pr (si , oj )Ts )
X
j=1 j p
(RP1) k=tr /Ts +1 Γi,k 3oj , i∈[n],
p
p∈{p|Θi,k =X∩Θi,k }
m tje /Ts
X∈I
X X X
max UR = wj · U ( xpi,k Pr (si , oj )Ts ) s.t.
p
xi,k p
j=1 j
k=tr /Ts +1 Γi 3oj ,
i∈[n], p∈[|Γi |]
where X is the decision variable, and f (X) (f : 2S → R≥0 )
|Γi |
is the objective function. Note that we abuse the notation
p p
slightly, and here Γi,k 3 oj means ∃Tj 0 ∈ Γi,k |Tj 0 .oj 0 = oj .
X
s.t. xpi,k = 1, (xpi,k ∈ {0, 1})
p=1 For RP2, we have the following critical lemma.
where xpi,k s
are the decision variables, Γi is the pth domi-
p Lemma 4.2. The objective function f (X) in RP2 is a monotone
nant task set in Γi . submodular set function.
Clearly, RP1 is a combinatorial optimization problem. Proof: By Definition 4.2, we need to check whether
To facilitate further analysis, we first give the following f (X) satisfies the three listed conditions.
definitions. First, when there are no active scheduling policies, i.e.,
Definition 4.2. [62] (submodular set function) Let S be a X = ∅, the received energy for any task is zero, then we
finite ground set. A real-valued set function f : 2S → R have f (X) = 0.
is normalized, monotonic and submodular if and only if it Second, let A be a set of scheduling strategies in S and
satisfies the following conditions, respectively: e ∈ S\A. For simplicity, define
tje /Ts
1) f (∅) = 0; X X
2) f (A ∪ {e}) − f (A) ≥ 0 for any A ⊆ S and e ∈ S\A; g(X, j) = U( Pr (si , oj )Ts )
j p
(3)
Γi,k 3oj , i∈[n],
3) f (A ∪ {e}) − f (A) ≥ f (B ∪ {e}) − f (B) for any k=tr /Ts +1
p
A ⊆ B ⊆ S and e ∈ S\B .
as the achieved utility for task Tj . It is easy to see that
Definition 4.3. [62] (matroid) A matroid M is a strategy g(A ∪ {e}, j) − g(A, j) ≥ 0 because there are possibly more
M = (S, L) where S is a finite ground set, L ⊆ 2S is a chargers cover task Tj as all possible dominant task sets that
collection of independent sets, such that: p p
cover Tj , i.e., Γi,k (i ∈ [n], p ∈ {p|Θi,k = A ∩ Θi,k }) would
1)∅ ∈ L; be enlarged as A becomes A ∪ {e}, and the utility function
2)if X ⊆ Y ∈ L, then X ∈ L; U(.) is non-decreasing. Hence we have
3)if X, Y ∈ L, and |X| < |Y |, then ∃y ∈ Y \X , X ∪ m
X
{y} ∈ L. f (A ∪ {e}) − f (A) = wj · [g(A ∪ {e}, j) − g(A, j)] ≥ 0. (4)
j=1
Sk
Definition 4.4. [62] (partition matroid) Given S = i=1 Si0 is Third, let A and B be two sets such that A ⊆ B ⊆ S and
the disjoint union of k sets, l1 , l2 , . . . , lk are positive integers, element e ∈ S\B . On one hand, it is easy to see that
a partition matroid M = (S, I) is a matroid where I = tj tj
e e
{X ⊂ S : |X ∩ Si0 | ≤ li for i ∈ [k]}. Ts
X X Ts
X X
Pr (si , oj )Ts − Pr (si , oj )Ts
We will show that the problem RP1 fits perfectly in t
j p
Γi,k 3oj , t
j p
Γi,k 3oj ,
k= Tr +1 k= Tr +1
the realm of maximizing a monotone submodular function s i∈[n], p∈P1 s i∈[n], p∈P2
subject to a partition matroid. First, we define Γi,k = Γi tj
e tj
e
Ts Ts
(k ∈ [K]) as the set of dominant task sets for charger si at the X X X X
= Pr (si , oj )Ts − Pr (si , oj )Ts
kth time slot, where K is the total number of time slots and j p
Γi,k 3oj , j p
Γi,k 3oj ,
p t
k= Tr
t
k= Tr
the notation [n] = {1, 2, . . . , n}. Then, we define Θi,k as the s
+1
i∈[n], p∈P3 s
+1
i∈[n], p∈P4
p (5)
corresponding scheduling policy for Γi,k , i.e., the orientation
p p p p
that covers Γi,k = Γi , for charger si at the kth time where P1 = {p|Θi,k = {A ∪ e} ∩ Θi,k }, P2 = {p|Θi,k =
p p
slot, define Θi,k = {Θi,k }p∈[|Γi,k |] as the set of scheduling A ∩ Θi,k }, P3 = {p|Θi,k = {B ∪ e} ∩ Θi,k }, and P4 =
p
policies for si at the kth time slot, and define a ground set {p|Θi,k = B ∩ Θi,k }. On the other hand, it is clear that
of all scheduling policies S = {Θi,k }i∈[n],k∈[K] . Further, we (U(x1 + ∆x) − U (x1 )) − (U(x2 + ∆x) − U (x2 )) ≥ 0, (6)
define the scheduling policies for all chargers at all K time
slots as X , which is subject to |X ∩ Θi,k | ≤ 1. Therefore, as for any x2 ≥ x1 ≥ 0 and ∆x ≥ 0 due to the concavity of the
Θi,k s are disjoint sets, we write the independent sets as charging utility function U(.).
Consequently, we have [g(A∪{e}, j)−g(A, i, q)]−[g(B∪
I = {X ⊆ S : |X ∩ Θi,k | ≤ 1 for i ∈ [n], k ∈ [K]}. (2) {e}, j) − g(B, j)] ≥ 0, and therefore,
Algorithm 2: Centralized Offline Algorithm to HASTE

[f (A ∪ {e}) − f (A)] − [f (B ∪ {e}) − f (B)] Input: Integer C , set of scheduling policies Θi,k for
Xm charger si (i ∈ [n], k ∈ [K]), objective function
= wj · {[g(A ∪ {e}, j) − g(A, j)] − [g(B ∪ {e}, j) − g(B, j)]} f (.)
j=1 Output: Scheduling policies for all chargers X
≥0. (7) 1 Q ← ∅;
2 for all c ∈ [C] do
In summary, we conclude that f (X) is a monotone
3 for all i ∈ [n], k ∈ [K] do
submodular set function. This completes the proof.
4 ei,k,c ← arg maxx∈Θi,k ×{c} F(Q + x);
5 Q ← Q ∪ ei,k,c ;
5 C ENTRALIZED O FFLINE A LGORITHM
6 for all i ∈ [n], k ∈ [K] do
In this section, we propose a centralized offline algorithm to 7 Choose ci,k uniformly at random from [C];
address HASTE in the offline scenario. We note that in this
case, the information for all charging tasks is known before- 8 X ← sample~c (Q), where
hand, and thereby the scheduling policies for all chargers at ~c = (c1,1 , . . . , cn,1 , . . . , c1,K , . . . , cn,K ).
any time can be determined a priori. 9 return X
5.1 Algorithm Description

color ci,k in ~c is selected uniformly at random from
After proved that HASTE-R is a problem of maximizing a [C].
submodular function under a partition matroid constraint,
we can resort to existing schemes to address HASTE-R. For We present our centralized offline algorithm in Algo-
example, we can use a simple greedy algorithm to find a rithm 2. We can see that at each step in the two-level loop,
solution that achieves 21 approximation ratio according to Algorithm 2 greedily optimizes F(Q).
the classical results presented in [52]. Moreover, [53] pro-
poses a randomized algorithm with optimal approximation 5.2 Theoretical Analysis
guarantees, namely, 1 − 1e approximation ratio. Neverthe- Following Theorem 2 in [55], we have the following lemma.
less, it is too computationally demanding to practically
Lemma 5.1. Algorithm 2 achieves 1 − (1 − C1 )C − nK
−1
implement. In this paper, we tailor the TABULARGREEDY 2 C
approximation ratio for HASTE-R.
algorithm proposed in [54], [55] to address HASTE-R as it
can achieve an approximation ratio between 12 and 1 − 1e Obviously, when C → +∞, the approximation ratio
depending on the value of a control parameter and resulting approaches 1 − 1e . Further, when C = 1, there is only
in different time complexity, which provides flexibility in one possible choice for ~c, and TABULARGREEDY is simply
practical applications. the locally greedy algorithm that achieves 12 approximation
We first propose some useful concepts in our context, ratio [52]. For simplicity, we assume C → +∞ and say
which also capture the essential elements in the TABU- Algorithm 2 achieves 1 − 1e approximation ratio for HASTE-
LARGREEDY algorithm, to facilitate understanding our R.
algorithm. Theorem 5.1. Algorithm 2 achieves (1−ρ)(1− 1e ) approximation
ratio for HASTE, and its time complexity is O(C(nmK)2 )
• S-C tuple. An S-C tuple is a tuple of a scheduling
where ρ is the switching delay, C , n, and m are the color
policy for a charger at a time slot and a color from
number, charger number, and task number, respectively, K is
a palette [C] of C colors (note that here color and
the number of considered time slots for all tasks.
palette have no concrete meaning, and they are only
used to assist sampling). A set Q ⊆ S × [C] consists Proof: Suppose the optimal charging utility for
∗ ∗
of S-C tuples which can be regarded as labeling each HASTE is U , and that for HASTE-R is U R . Apparently, we
scheduling policy for a charger with one or more have
colors. ∗ ∗
UR ≥ U . (9)
• S-C tuple sampling function. We associate with each
partition Θi,k a color ci,k . For any set Q ⊆ S × [C] Further, suppose the output X of Algorithm 2 achieves
and vector ~c = (c1,1 , . . . , cn,1 , . . . , c1,K , . . . , cn,K ), overall charging utility U R for HASTE-R, i.e.,
we define S-C tuple sampling function as m tje /Ts
X X X
[ UR = wj · U ( Pr (si , oj )Ts ),
sample~c (Q) = {x ∈ Θi,k : (x, ci,k ) ∈ Q}. j=1 j
k=tr /Ts +1
p
i∈[n],k∈[K] p
(8)
In other words, sample~c (Q) returns a set containing and achieves U (U ≤ U R ) for HASTE by taking the switch-
each item x that is exactly labeled with the color ci,k ing delay into consideration. Consider the worst case, i.e.,
assigned by ~c to the partition Θi,k that contains x. every charger needs to rotate at the beginning of each time
• Expected charging utility function after S-C tuple sam- slot and lead to switching delay, which results in a time
pling. It is defined as F(Q) = E(f (sample~c (Q))) duration of (1 − ρ)Ts for effective charging in each time slot
as the expected value of f (sample~c (Q)) when each for all tasks, then we have
6.1 Algorithm Description

m tje /Ts
We design our distributed online algorithm for HASTE
U≥
X
wj · U (
X X
Pr (si , oj )(1 − ρ)Ts ) based on our proposed centralized offline algorithm, and
j=1 j p
partially borrow the idea of the distributed algorithm in [63].
k=tr /Ts +1
p
p∈{p|Θi,k =X∩Θi,k } First, we present some concepts to assist analysis.
m
X tje /Ts
X X • Neighbors of a charger. We say two chargers are
≥ (1 − ρ) wj · U ( Pr (si , oj )Ts ) neighbors to each other if and only if they cover
j=1 j
k=tr /Ts +1
p
Γi,k 3oj , i∈[n], at least one charging task in common. We assume
p
p∈{p|Θi,k =X∩Θi,k } that the communication range of wireless chargers is
= (1 − ρ)U R . (10) at least twice of their charging range, and therefore,
the neighboring wireless chargers can communicate
Note that the second inequality in the above formula with each other. The set of neighbors of charger si is
is due to the concavity of the charging utility function. denoted as N (si ).
Following Lemma 5.1 and letting C → +∞, we have • Local charging utility function. The local charging util-
ity function for charger si is defined as the aggre-
1 ∗ gated charging utility of all charging tasks that can
U R ≥ (1 − )U R . (11)
e be charged by si , i.e., T i . Denote by Xi as the set of
Combining Equs. (9), (10), and (11) we obtain scheduling policies of si , and Xi the set of scheduling
polices of si and its neighbors N (si ), we can for-
1 ∗ mally express the local charging utility function for
U ≥ (1 − ρ)(1 − )U , (12)
e HASTE-R as fi : ∪si0 ∈{si }∪N (si ),k∈[Ki ] Θi0 ,k 7→ R≥0
as
which indicates that Algorithm 2 achieves (1 − ρ)(1 − 1e )
tje /Ts
approximation ratio. X X X
For time complexity, it is clear that the computation for fi (Xi ) = wj U( Pr (si0 , oj )Ts )
Tj ∈T i j p
Q is the dominating part. The computation inside the two- k=tr /Ts +1 Γi,k 3oj , si0 ∈{si }∪N (si ),
p
p∈{p|Θi,k =Xi ∩Θi,k }
level loop involves testing all possible scheduling policies,
which is O(m) in the worst case when the considered where Ki is the number of considered time slots for
charger covers all m tasks. Moreover, computing F(Q + x) all tasks T i observed by charger si .
needs O(nmK) computational cost. Thus, considering all • Local expected charging utility function after S-C tuple
CnK loops, the overall time complexity for computing Q is sampling. Similar to the expected charging utility
O(C(nmK)2 ), so does the time complexity of Algorithm 2. function after S-C tuple sampling defined in Section
This completes the proof. 5.1, we define Fi (Qi ) = E(fi (sample~c (Qi ))) as the
expected value of fi (sample~c (Qi )) when each color
ci,k in ~c is selected uniformly at random from [C].
6 D ISTRIBUTED O NLINE A LGORITHM • Control message. The control message exchanged
between wireless chargers is expressed as
In this section, we propose a distributed online algorithm to msg(ID, T IM, COL, CM D, ∆F∗i (Qi ), ek∗ i ). The
address HASTE in the online scenario. Note that in this case, field ID is the charger ID; T IM is the index of the
charging tasks stochastically arrive, and chargers reschedule time slots; COL is an integer between 1 and C , which
their orientations in realtime. Moreover, chargers can be stands for the parameter c in the centralized offline
either in the working mode for the offline scenario or in that algorithm; CM D can be U P D which indicates an
for the online scenario, and cannot switch between these update command; and ∆Fk∗ i (Qi ) is the “maximum”
two different statuses. marginal increment for the local expected charging
We face two main challenges. First, we need to adapt utility function after S-C tuple sampling for charger
the centralized offline algorithm to HASTE, whose relaxed si for all possible scheduling policies at the kth time
version HASTE-R is a submodular function maximization slot, and ek∗ i is the corresponding scheduling policy.
problem, to cater to the distributed online scenario where We show our distributed online algorithm in Algorithm
all chargers are asynchronous and charging tasks randomly 3, which is invoked at charger si upon arrival of new charg-
arrive. Nevertheless, to the best of our knowledge, there are ing tasks that can be charged by si . Each charger accordingly
no distributed online schemes for maximizing a submodular updates the set of charging tasks T i , all possible scheduling
function with or without constraints. Second, the response policies in all Ki time slots Θi,k , and the local charging
of each charger has a delay of up to τ + ρ time slots, that utility function fi (.). Then, each charger si enumerates all
is, τ number of time slots for computation and negotiation C colors in all Ki time slots. For each color c at the kth
with neighboring chargers and, possibly, plus ρ time slot for time slot, si computes ∆Fk∗ i (Qi ) and the corresponding
switching delay. This setting is fundamentally different from scheduling policy ek∗ , and broadcasts them to its neighbors.
i
existing ones of online scheduling problems and invalidates Note that ∆Fk∗ (Q ) for charger s is obtained by greedily
i i i
traditional online algorithms. We address these challenges choosing the scheduling policies that yield the maximum
by proposing a distributed online algorithm that achieves additional local expected charging utility in all Ki time
1 1
2 (1 − ρ)(1 − e ) competitive ratio. slots in an increasing order, and therefore, ek∗ i is a set of
Algorithm 3: Distributed Online Algorithm to HASTE 6.2 Theoretical Analysis

(at each wireless charger si ) Theorem 6.1. Algorithm 3 achieves 21 (1 − ρ)(1 − 1e )
Input: Neighbor set N (si ) competitive ratio for HASTE, and its time complex-
Output: Scheduling policy Xi ity is O(C(|N (si )||T i |Ki )2 ), its communication cost is
1 Update the set of charging tasks that can cover O(CKi (|N (si )|)2 ) where ρ is the switching delay, C is the
charger si , i.e., T i to include the new arrived tasks; number of colors, N (si ) is the set of neighbors of charger si ,
2 Compute the dominant task sets and determine all T i is the set of tasks that can cover si , Ki is the number of
possible scheduling policies Θi,k ; considered time slots for all tasks in T i .
3 Exchange the information of dominant task sets and Proof: First of all, we refer the readers to the confer-
scheduling policies with the neighbors, and thus ence of this paper [51] for the proof of the competitive ratio.
derive the local charging utility function fi (.); The time complexity analysis is similar to that in the
4 Qi ← ∅; proof to Algorithm 5.1, we omit it to save space. For
5 for k from 1 to Ki do communication cost, it is clear that there are in total CKi
6 for c from 1 to C do loops, and in each loop, there are O(|N (si )|) rounds to
7 Calculate ∆Fk∗ k∗
i (Qi ) and obtain ei ; determine a local maximum marginal increment for the local
8 Broadcast msg(i, k, c, N U LL, ∆Fk∗ k∗
i (Qi ), ei ); expected charging utility function after S-C tuple sampling
k∗
9 while ∆Fi (Qi ) > 0 do for a charger and its neighbors. Each round in turn needs
10 if ∆Fk∗ j (Qj ) of all neighbors sj ∈ N (si ) are O(|N (si )|) times of message sending and receiving. To sum
collected and all their colors are equal to c, and up, the total communication cost is O(CKi (|N (si )|)2 ). This
∆Fk∗ i (Qi ) is larger than any of them then completes the proof.
11 Qi ← Qi ∪ (ek∗ i , c);
12 Broadcasts 7 S IMULATION R ESULTS
msg(i, k, c, U P D, ∆Fk∗ k∗
i (Qi ), ei ); In this section, we perform simulations to evaluate the per-
13 break; formance of the proposed centralized offline and distributed
if msg(j, k, c, U P D, ∆Fk∗ k∗ online algorithms to HASTE.
14 j (Qj ), ej ) is
received then
7.1 Evaluation Setup
15 Update the stored scheduling policy of
its neighbor sj at the kth time slots to Unless otherwise stated, we use the following setup in our
ek∗ simulations. The considered field is a 50 m × 50 m square
j ;
Calculate ∆Fk∗ k∗ area, and wireless chargers and charging tasks are uniformly
16 i (Qi ) and obtain ei ;
17 Broadcast distributed in this filed. We set α = 10000, β = 40,
1
msg(i, k, c, N U LL, ∆Fk∗ k∗ D = 20 m, n = 50, m = 200, wj = 200 , Ts = 1 min,
i (Qi ), ei ); 1
18 continue; ρ = 12 , τ = 1, As = π/3, Ao = π/3, respectively. The
required charging energy and duration of charging tasks
19 if msg(j, k, c, N U LL, ∆Fk∗ k∗
j (Qj ), ej ) is are randomly selected in [5kJ 20kJ] and [10min 120min],
received then respectively. If we choose 3.7 Volts for the voltage, the
20 Update ∆Fk∗ k∗
j (Qj ) and ej for the required charging energy is selected in [375mAh 1500mAh].
neighbor sj ; Therefore, the simulation setup is reasonable for the re-
21 continue; quired battery of wireless sensor and mobile devices. To
cover the area of 50 m × 50 m, we chose n = 50 wireless
22 for c from 1 to C do chargers. Note that if the number of wireless chargers is too
23 Choose cik uniformly at random from [C]; small, some charging tasks area will not be covered, which
will make HASTE meaningless. Conversely, if the number of
24 Xi ← sample~c (Qi ), where ~c = (ci1 , . . . , ciKi ). wireless chargers is too large, the charging utility will close
25 return Xi to 1.0. Besides, each data point in the figures in this section
stands for an averaging result for 100 random topologies.
Ki scheduling policies. Meanwhile, si receives the control
messages sent from its neighbors. If it collects the messages 7.2 Baseline Setup
from all its neighbors and finds that it has the maximum As there are no existing schemes for scheduling charging
value of ∆Fk∗ i (Qi ) (if there are two or more chargers have tasks in directional wireless charger networks, we propose
the same value of ∆Fk∗ i (Qi ) which leads to a tie, we break two algorithms named GreedyUtility and GreedyCover for
it based on the IDs of these chargers), si adds the S-C tuple comparison. For GreedyUtility, each charger greedily picks
(ek∗
i , c) to its global S-C tuple set Qi , and broadcasts the the orientation that leads to maximum charging utility
update command to its surrounding neighbors. Otherwise, while ignoring the scheduling policies of its neighboring
if it receives an update command from one of its neighbors, chargers. For GreedyCover, the only difference compared
si updates the stored scheduling policy for the neighbor, with GreedyUtility is that each charger greedily selects the
recomputes ∆Fk∗ k∗
i (Qi ) and ei , and repeats the above ne- orientation that covers the maximum number of charging
gotiation procedure. After traversing all C colors for all Ki tasks. Apparently, both of these algorithms can be easily
time slots, Algorithm 3 obtains a set of S-C tuples Qi , and implemented in a distributed way by letting each charger
applies a sampling function on Qi to get a solution Xi . execute them locally.
0.8 1 0.75 0.85
0.76 0.9 0.72 0.8

Charging Utility
Charging Utility
Charging Utility
Charging Utility
0.72 0.8 0.69 0.75
0.68 0.7 0.66 0.7

HASTE(C=4) HASTE(C=4) HASTE(C=4)
0.64 HASTE(C=1) 0.6 HASTE(C=1) 0.63 HASTE(C=1) 0.65
GreedyUtility GreedyUtility GreedyUtility
GreedyCover GreedyCover GreedyCover
0.6 0.5 0.6 0.6
30 60 90 120 150 180 210 240 270 300 330 360 30 60 90 120 150 180 210 240 270 300 330 360 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 2 3 4 5 6 7 8
As (°) Ao (°) ρ C
Fig. 2. As vs. charging utility (cen- Fig. 3. Ao vs. charging utility (cen- Fig. 4. ρ vs. charging utility (central- Fig. 5. C vs. charging utility (cen-
tralized offline algorithm) tralized offline algorithm) ized offline algorithm) tralized offline algorithm)
7.3 Centralized Offline Algorithm Evaluation 0.7

0.7
Charging Utility
Charging Utility
1
7.3.1 Impact of Charging Angle As 0.8
0.65 1
0.8
0.65
0.6
0.6
0.6
Our simulation results show that on average HASTE outperforms 0.4
0.55
0.6
0.4
0.55
0.5
GreedyUtility and GreedyCover by 2.67% and 3.40% (at most 0.2

0.5
0.45
0.2 0.45
0 0
4.34% and 6.03%), respectively, in terms of As . Figure 2 shows 70
60 10
0.4 70
60 10
0.4
0.35
20 20
that the charging utilities of HASTE, GreedyUtility, and Δt (min)
50
40 40
30 0.35
Δt (min)
50
40 40
30 0.3
30 50 Ej (kJ) 30 50 Ej (kJ)
GreedyCover steadily increase with the charging angle of
chargers As , and achieve the same maximum overall charg- Fig. 6. Required charging energy Fig. 7. Required charging energy
ing utility when As = 360◦ . Note that for simplicity, we & task duration vs. charging utility & task duration vs. charging utility
(centralized offline algorithm) (distributed online algorithm)
still use HASTE to denote our proposed centralized offline
algorithm or distributed online algorithm to HASTE in all 7.3.4 Impact of Color Number C
simulation figures if no confusion arises. This observation
Our simulation results show that on average the achieved charg-
is consistent with our intuition as the larger the charging
ing utility of HASTE steadily increases with color number C .
angle, the larger the chance that a charger can cover more
Figure 5 shows the box plot of the charging utilities of
charging tasks even with the same orientation, and all the
HASTE. It can be seen that the average charging utility
chargers cover the same set of tasks regardless of their
of HASTE increases by 3.29% when the color number C
orientations when As = 360◦ and thus make no difference
increases from 1 to 8. The maximum and minimum charging
in the performance for the three algorithms. Moreover, the
utilities of HASTE also smoothly increase with C . The
solution for HASTE with the color number C = 4 always
variance of charging utility for the eight colors is at most
outperforms that with C = 1, and has a performance gain
of 0.39% on average (at most 2.59%).
8.56 × 10−3 .
7.3.5 Impact of Required Charging Energy and Task Dura-

7.3.2 Impact of Receiving Angle Ao
tion
Our simulation results show that on average HASTE outperforms Our simulation results show that the achieved charging utility
GreedyUtility and GreedyCover by 5.63% and 8.81% (at most of HASTE steadily increases with a decreasing charging energy
7.36% and 11.27%), respectively, in terms of Ao . Figure 3 or an increasing task duration. We set the required charging
shows that the charging utilities of the three algorithms energy being randomly selected from [0.5Ej 1.5Ej ], and
increase monotonically with the receiving angle of devices task duration from [0.5∆t 1.5∆t]. Figure 6 shows that when
Ao . This is because tasks with larger receiving angles can be Ej decreases from 50 kJ to 10 kJ and ∆t increases from
charged with more potential chargers. Clearly, the increas- 30 min to 70 min, the overall charging utility increases by
ing speeds of charging utilities for these algorithms are first 44.28%. Moreover, the increasing speed for charging utility
fast and then become slow as Ao increases from 30◦ to 360◦ . slows down when Ej is large or ∆t is small, which indicates
On average, HASTE with C = 4 outperforms HASTE with a marginal diminishing gain property.
C = 1 by 1.04% on average (at most 1.45%).
7.3.3 Impact of Switching Delay ρ 7.4 Distributed Online Algorithm Evaluation

Our simulation results show that on average HASTE outperforms 7.4.1 Impact of Charging Angle As
GreedyUtility and GreedyCover by 3.20% and 6.30% (at most Our simulation results show that on average HASTE outperforms
3.25% and 6.34%), respectively, in terms of ρ. Not surprisingly, GreedyUtility and GreedyCover by 3.33% and 4.47% (at most
we observe in Figure 4 that the charging utilities for all the 5.59% and 7.59%), respectively, in terms of As . We denote
algorithms smoothly decrease with an increasing switching by HASTE-DO the distributed online algorithm for HASTE
delay ρ. HASTE with C = 4 outperforms HASTE with C = in the following figures. Figure 8 demonstrates that the
1 by 0.99% (at most 1.00%). Note that even when ρ = 1, charging utilities of HASTE, GreedyUtility, and Greedy-
which means the switching delay is up to one time slot, the Cover smoothly increase with the charging angle of chargers
achieved charging utilities for all the algorithms just slightly As , and reach the same maximum overall charging utility
degrade. The reason is that each charger keeps still most of when As = 360◦ . This is a natural result because the larger
the time and the orientation switching seldom happens, and the charging angle, the larger the chance that a charger
therefore, the performance loss caused by switching is little. can cover more charging tasks with the same orientation.
0.8 1 0.73 0.85
0.9
0.76 0.7 0.8
Charging Utility
Charging Utility
Charging Utility
Charging Utility
0.8
0.72 0.67 0.75
0.7
0.68 0.64 0.7
0.6
HASTE-DO(C=4) HASTE-DO(C=4) HASTE-DO(C=4)
0.64 HASTE-DO(C=1) HASTE-DO(C=1) 0.61 HASTE-DO(C=1) 0.65
0.5
GreedyUtility-DO GreedyUtility-DO GreedyUtility-DO
GreedyCover-DO GreedyCover-DO GreedyCover-DO
0.6 0.4 0.58 0.6
30 60 90 120 150 180 210 240 270 300 330 360 30 60 90 120 150 180 210 240 270 300 330 360 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 2 3 4 5 6 7 8
As (°) Ao (°) ρ C
Fig. 8. As vs. charging utility (dis- Fig. 9. Ao vs. charging utility (dis- Fig. 10. ρ vs. charging utility (dis- Fig. 11. C vs. charging utility (dis-
tributed online algorithm) tributed online algorithm) tributed online algorithm) tributed online algorithm)
Moreover, if As = 360◦ , each charger covers the same set 25

Number of Messages
of tasks regardless of its orientations, and therefore, the 20
Number of Rounds
three algorithms have the same performance. The solution

15
for HASTE with C = 4 always outperforms that with
C = 1 with a gain of 0.77% on average (at most 2.59%). 10
Besides, we can see that the charging utility for each of the 5
three distributed online algorithms is less than that of its

0
corresponding centralized offline algorithm. 10 20 30 40 50 60 70 80 90 100
Number of Chargers
7.4.2 Impact of Receiving Angle Ao Fig. 12. Communication cost
Our simulation results show that on average HASTE outperforms 7.4.5 Impact of Required Charging Energy and Task Dura-
GreedyUtility and GreedyCover by 6.83% and 8.95% (at most tion
8.68% and 10.96%), respectively, in terms of Ao . Figure 9
Our simulation results show that the achieved charging utility
illustrates that the charging utilities of the three algorithms
of HASTE steadily increases with a decreasing charging energy
monotonically increase with the receiving angle of devices
or an increasing task duration. Similar to the setting for the
Ao . The reason is that tasks with larger receiving angles can
centralized offline algorithm, we set the required charging
potentially be charged by more chargers, and thus receive
energy being randomly selected from [0.5Ej 1.5Ej ], and
more energy on average. Moreover, it is clear that the
task duration from [0.5∆t 1.5∆t]. Figure 7 shows that when
increasing trends of charging utilities for these algorithms
Ej downgrades from 50 kJ to 10 kJ and ∆t rises from
are first fast and then become slow as Ao increases from 30◦
30 min to 70 min, the achieved charging utility increases by
to 360◦ . Besides, HASTE with C = 4 outperforms HASTE
45.47%. The increasing speed for charging utility decreases
with C = 1 by 1.42% on average (at most 2.23%). Again,
when Ej increases or ∆t decreases, which demonstrates a
the charging utilities for the distributed online algorithms
marginal diminishing gain property.
are less than their corresponding centralized offline version.
7.4.6 Communication Cost
7.4.3 Impact of Switching Delay ρ
Our simulation results show that the number of messages and
Our simulation results show that on average HASTE outperforms
the number of rounds for a time slot increase quadratically and
GreedyUtility and GreedyCover by 5.20% and 7.3% (at most
linearly, respectively, with the number of chargers. We set the
5.20% and 7.31%), respectively, in terms of ρ. Figure 10 shows
number of color C to 1, and plot the average numbers of
that the charging utilities for all the algorithms steadily
messages and rounds in Algorithm 3 in Figure 12. We can
decrease with switching delay ρ. Especially, HASTE with
see that when the number of chargers increases from 10
C = 4 outperforms HASTE with C = 1 by 1.98%. When
to 100, the numbers of messages and rounds increase by
the switching delay is even up to one time slot, i.e., ρ = 1,
223.77% and 952.29%, respectively. The number of rounds
the achieved charging utilities for all the algorithms only
linearly increases because the number of neighboring charg-
slightly degrade compared with ρ = 0. This is because most
ers linearly increases. Further, as the number of messages
chargers keep still most of the time, and thus the caused
in each round also grows proportionally to the number of
performance loss is little.
neighboring chargers, it thus grows quadratically with the
7.4.4 Impact of Color Number C number of neighboring chargers, or the number of chargers.
Our simulation results show that on average the achieved charg- This finding supports Theorem 6.1.
ing utility of HASTE steadily increases with color number C .
Figure 11 demonstrates the box plot of the charging utilities 7.5 Insights
of HASTE when the color number C increases from 1 to 8. First, we investigate the impact of distribution of positions
We can see that both of the maximum and minimum charg- of charging tasks on the overall charging utility. Suppose
ing utilities of HASTE steadily increase with C . Moreover, there are 50 tasks distributed in a 50 m × 50 m area, and
on average the average charging utility of HASTE increases Ao = As = π/3. The required charging energy and charging
by 3.08% when the color number C increases by 1. Besides, duration for all tasks are randomly chosen from [5 kJ 20 kJ]
the variance of charging utility for all the eight colors is at and [10 min 120 min], respectively. The positions of tasks
most 8.42 × 10−3 , which indicates the stable performance of are randomly generated following a 2D Gaussian distribu-
our algorithm. tion with both x- and y - coordinates obeying a Gaussian
1
Charging Utility 0.35
Charging Utility
0.4 0.8
0.3
0.3
0.6
1
0.2
0.25
U∝ Ej o
0.2
0.4 87 [1 3]
o
0.1
4 114 [2 6]
348
o
[1 2]
7
0.2
0 0.15 12 8
25 o
39 [4 8]
20 o
25 342 [0 3]
15 20 0.1 0 o 20
σy 10 15 0 1 2 3 4 5 6 7 8 9 10 54 [2 5] o
5 5
10
σx Ej (J) ×104 2 298 [1 5]
0 0 3 9
o o 6
228 [1 5] 296 [3 5]
16
Fig. 13. Overall charging utility vs. Fig. 14. Individual charging utility o
5 20
o
[3 4] 4
o
[3 4]
o
0 [2 5] 25 [0 3]
Guassian distribution variance vs. required charging energy 15 18 138
o
[3 4]
1
151
o
[3 6]
2 10
14
17
o
305 [2 3]
o
112
o 130
o
[0 4]
11 245 [0 3]
[3 5]
3
o
13 4 20 [3 5]
1 19
0 Task
Charger
0 1 2 3 4 5
Fig. 16. Topology 2
Fig. 15. Testbed 0.8
Charging Utility
0.6
distribution with µ = 25. Figure 13 shows that generally the 0.4
charging utility increases with either σx or σy , which indi- 0.2 HASTE(C=4)

GreedyUtility
cates that the uniformness of tasks’ distribution contributes GreedyCover
0
to the overall charging utility. This is because with a higher 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
degree of uniformness of positions, the phenomenon that Task
some tasks are over-charged while the others are starved out Fig. 17. Charging utility of 20 tasks for the centralized offline algorithms
can be effectively avoided, and according to the concavity
of the charging utility function, the overall charging utility
will be enhanced. Second, we study the impact of Ej on
9 C ONCLUSION
the individual charging utility of each charger. Compared The key novelty of this paper is on proposing the first
with the above setting, we uniformly distribute 50 chargers scheduling algorithm for charging tasks in directional wire-
and 200 tasks. The required charging energy is a random less charging networks. The key contributions of this paper
number in [5 kJ 100 kJ]. Figure 14 shows that generally are proposing a centralized offline algorithm that achieves
the charging utility first can achieve 1 for a small Ej , and (1 − ρ)(1 − 1e ) approximation ratio where ρ denotes the
then rapidly decreases when Ej continues growing. The switching delay, and a distributed online algorithm that
maximum individual charging utility is approximately in- achieves 12 (1 − ρ)(1 − 1e ) competitive ratio, and conducting
versely proportional to Ej , as shown by the curve in Figure both simulations and field experiments for evaluation. The
14. The reason is that to achieve the same charging utility, key technical depth of this paper is in transforming the
a task with a higher required Ej needs a higher average problem into maximizing a submodular function subject to
charging power from its surrounding chargers, which is not a partition matroid constraint, bounding the performance
cost efficient. Thus, higher Ej leads to lower charging utility. loss caused by the switching delay and proving the approx-
imation ratio for the centralized offline algorithm, making
the centralized offline algorithm distributed and bounding
8 F IELD E XPERIMENTS the performance loss caused by the rescheduling delay and
We have conducted field experiments to evaluate our proving the competitive ratio for the distributed online algo-
scheme. First of all, we implemented our proposed schemes rithm. Our simulation and field experimental results show
on a small textbed. Due to space limit, we refer readers to that our proposed distributed online algorithm can achieve
the conference version of this paper [51] for details. 92.97% of the optimal charging utility, outperform the other
Next, we implemented our proposed schemes on a large two comparison algorithms, and its communication cost
testbed which consists of 16 TX91501 power transmitters moderately increases as the charger number scales up.
and 20 rechargeable sensor nodes. Figure 16 shows the
topology of this large testbed, which is much more irreg-
ular than the small testbed as it is randomly generated. ACKNOWLEDGEMENT
Similarly, Figures 17 and 18 show that HASTE respectively This work was supported in part by the National Natural
outperforms GreedyUtility and GreedyCover by 4.38% and Science Foundation of China under Grant 61872178, in part
10.12% on average, and by 13.27% and 23.60% at most in by the Natural Science Foundation of Jiangsu Province
the centralized offline scenario; and by 6.04% and 15.28% under Grant No. BK20181251, in part by Supported by the
on average, and by 22.58% and 29.63% at most for in the Fundamental Research Funds for the Central Universities
distributed online scenario. under Grant 14380062, in part by the open research fund
1
[18] X. Xu et al., “Optimizing charging efficiency and maintaining sen-
0.8 sor network perpetually in mobile directional charging,” Sensors,
Charging Utility
vol. 19, no. 12, p. 2657, 2019.
0.6
[19] “www.powercastco.com,” 2018.
0.4 [20] “https://blog.ossia.com/news/ossia-and-t-mobile-pilot-wireless-
0.2 HASTE-DO(C=4) charging-asset-trackers-in-walmart-warehouses,” 2019.
GreedyUtility-DO [21] “https://wi-charge.com/applications,” 2019.
GreedyCover-DO
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 [22] C. Wang et al., “Multi-vehicle coordination for wireless energy
Task replenishment in sensor networks,” in IEEE IPDPS, 2013, pp.
1101–1111.
Fig. 18. Charging utility of 20 tasks for the distributed online algorithms [23] ——, “Recharging schedules for wireless sensor networks with ve-
hicle movement costs and capacity constraints,” in IEEE SECON,
2014, pp. 468–476.
of Key Lab of Broadband Wireless Communication and [24] W. Xu et al., “Towards perpetual sensor networks via deploying
multiple mobile wireless chargers,” in IEEE ICPP, 2014, pp. 80–89.
Sensor Network Technology (Nanjing University of Posts
[25] L. Chen et al., “Charge me if you can: Charging path optimization
and Telecommunications), Ministry of Education, in part by and scheduling in mobile networks,” in ACM MobiHoc, 2016, pp.
the National Natural Science Foundation of China under 101–110.
Grant 61802172, Grant 61672276, Grant 61832005, Grant [26] S. Nikoletseas et al., “Interactive wireless charging for energy
61472184, Grant 61872173, and Grant 61872082, in part balance,” in IEEE ICDCS, 2016, pp. 262–270.
[27] L. Fu et al., “Minimizing charging delay in wireless rechargeable
by the Key Research and Development Project of Jiangsu sensor networks,” in IEEE INFOCOM, 2013, pp. 2922–2930.
Province under Grant BE2015154 and Grant BE2016120, [28] L. He et al., “Esync: An energy synchronized charging protocol for
and the Collaborative Innovation Center of Novel Software rechargeable wireless sensor networks,” in ACM MobiHoc, 2014,
Technology and Industrialization, Nanjing University, and pp. 247–256.
[29] ——, “Mobile-to-mobile energy replenishment in mission-critical
in part by the Jiangsu Innovation and Entrepreneurship robotic sensor networks,” in IEEE INFOCOM, 2014, pp. 1195–1203.
(Shuangchuang) Program, the Guangdong Leading Talent [30] Y. Shi et al., “On renewable sensor networks with wireless energy
Program No. 2016LJ06D658. transfer,” in IEEE INFOCOM, 2011, pp. 1350–1358.
[31] L. Xie et al., “On renewable sensor networks with wireless energy
transfer: the multi-node case,” in IEEE SECON, 2012, pp. 10–18.
[32] S. Guo et al., “Mobile data gathering with wireless energy replen-
R EFERENCES ishment in rechargeable sensor networks,” in IEEE INFOCOM,
2013, pp. 1932–1940.
[1] A. P. Sample et al., “Design of an rfid-based battery-free pro- [33] L. Xie et al., “Bundling mobile base station and wireless energy
grammable sensing platform,” IEEE Transactions on Instrumentation transfer: Modeling and optimization,” in IEEE INFOCOM, 2013,
and Measurement, vol. 57, no. 11, pp. 2608–2615, 2008. pp. 1636–1644.
[2] L. Xie et al., “Wireless power transfer and applications to sensor [34] ——, “On traveling path and related problems for a mobile station
networks,” IEEE Wireless Communications, vol. 20, no. 4, pp. 140– in a rechargeable sensor network,” in ACM MobiHoc, 2013, pp.
145, 2013. 109–118.
[3] S. Li et al., “Wireless power transfer for electric vehicle appli- [35] C. Wang et al., “Improve charging capability for wireless recharge-
cations,” IEEE Journal of Emerging and Selected Topics in Power able sensor networks using resonant repeaters,” in IEEE ICDCS,
Electronics, vol. 3, no. 1, pp. 4–17, 2015. 2015, pp. 133–142.
[4] P. E. Glaser, “An overview of the solar power satellite option,” [36] ——, “A hybrid framework combining solar energy harvesting
IEEE Transactions on Microwave Theory and Techniques, vol. 40, no. 6, and wireless charging for wireless sensor networks,” in IEEE
pp. 1230–1238, Jun 1992. INFOCOM, 2016, pp. 1–9.
[5] S. F. Bush, Smart grid: Communication-enabled intelligence for the [37] Z. Dong et al., “Energy synchronized task assignment in recharge-
electric power grid. John wiley & sons, 2014. able sensor networks,” in IEEE SECON, 2016, pp. 1–9.
[6] “https://www.wirelesspowerconsortium.com/,” 2018. [38] S. He et al., “Energy provisioning in wireless rechargeable sensor
[7] “https://www.wirelesspowerconsortium.com/blog/273/wireless- networks,” IEEE Transactions on Mobile Computing, vol. 12, no. 10,
power-market-surges-as-usage-leaps-forward,” 2018. pp. 1931–1942, 2013.
[8] Z. Wang et al., “Adaptively directional wireless power transfer [39] H. Dai et al., “Wireless charger placement for directional charg-
for large-scale sensor networks,” IEEE Journal on Selected Areas in ing,” IEEE/ACM Transactions on Networking, vol. 26, no. 4, pp.
Communications, vol. 34, no. 5, pp. 1785–1800, 2016. 1865–1878, 2018.
[9] L. Wang et al., “Millimeter wave power transfer and information
[40] N. Yu et al., “Placement of connected wireless chargers,” in IEEE
transmission,” IEEE Globecom, pp. 1–6, 2015.
INFOCOM, 2018. To appear.
[10] N. Deng and M. Haenggi, “A novel approximate antenna pattern
[41] X. Wang et al., “Heterogeneous wireless charger placement with
for directional antenna arrays,” IEEE Wireless Communications Let-
obstacles,” in ACM ICPP, 2018, p. 16.
ters, vol. 7, no. 5, pp. 832–835, 2018.
[11] T. A. Khan and R. W. Heath, “Wireless power transfer in millimeter [42] H. Dai et al., “Safe charging for wireless power transfer,” in IEEE
wave tactical networks,” IEEE Signal Processing Letters, vol. 24, INFOCOM, 2014, pp. 1105–1113.
no. 9, pp. 1284–1287, 2017. [43] ——, “Safe charging for wireless power transfer,” IEEE/ACM
[12] T. A. Khan and R. W. Heath Jr, “Wireless power transfer in Transactions on Networking, vol. 25, no. 6, pp. 3531–3544, 2017.
millimeter wave,” Wireless Information and Power Transfer: Theory [44] ——, “SCAPE: Safe charging with adjustable power,” in IEEE
and Practice, pp. 139–156, 2018. ICDCS, 2014, pp. 439–448.
[13] “http://www.powercastco.com/pdf/p2110-evb-reva.pdf,” 2018. [45] ——, “SCAPE: Safe Charging With Adjustable Power,” IEEE/ACM
[14] O. Georgiou, “Simultaneous wireless information and power Transactions on Networking, vol. 26, no. 1, pp. 520–533, Feb 2018.
transfer in cellular networks with directional antennas,” IEEE [46] S. Nikoletseas et al., “Low radiation efficient wireless energy
Communications Letters, vol. 21, no. 4, pp. 885–888, 2017. transfer in wireless distributed systems,” in IEEE ICDCS, 2015,
[15] L. Fan et al., “Secure wireless information and power transfer pp. 196–204.
based on tilt adaptation in 3-d massive mimo systems,” IEEE [47] H. Dai et al., “Radiation constrained wireless charger placement,”
Access, vol. 7, pp. 5531–5540, 2019. in IEEE INFOCOM, 2016, pp. 1–9.
[16] H. Dai et al., “Omnidirectional chargability with directional anten- [48] ——, “Robustly safe charging for wireless power transfer,” in IEEE
nas,” in IEEE ICNP, 2016, pp. 1–10. INFOCOM, 2018. To appear.
[17] ——, “Optimizing wireless charger placement for directional [49] ——, “Radiation constrained scheduling of wireless charging
charging,” in IEEE INFOCOM, 2017, pp. 2922–2930. tasks,” in ACM MobiHoc, 2017, pp. 17–26.
[50] ——, “Radiation constrained scheduling of wireless charging Ke Sun received his B.S. degree in Computer
tasks,” IEEE/ACM Transactions on Networking, vol. 26, no. 1, pp. Science from Nanjing University of Aeronautics
314–327, Feb 2018. and Astronautics, Jiangsu, China, in 2016. He is
[51] ——, “Charging task scheduling for directional wireless charger currently a Master student in Nanjing University.
networks,” in ICPP, 2018. To appear. His research interests are in the area of of mo-
[52] G. L. Nemhauser et al., “An analysis of approximations for max- bile computing.
imizing submodular set functions-I,” Mathematical Programming,
vol. 14, no. 1, pp. 265–294, 1978.
[53] G. Calinescu et al., “Maximizing a monotone submodular function
subject to a matroid constraint,” SIAM J. Comput., vol. 40, no. 6,
pp. 1740–1766, 2011.
[54] M. Streeter et al., “Online learning of assignments,” in Advances in
Neural Information Processing Systems 22, Y. Bengio, D. Schuurmans,
J. D. Lafferty, C. K. I. Williams, and A. Culotta, Eds. Curran Alex X. Liu received his Ph.D. degree in Com-
Associates, Inc., 2009, pp. 1794–1802. [Online]. Available: http:// puter Science from The University of Texas at
papers.nips.cc/paper/3719-online-learning-of-assignments.pdf Austin in 2006, and is a professor at the De-
[55] D. Golovin et al., “Online submodular maximization under a ma- partment of Computer Science and Engineer-
troid constraint with application to learning assignments,” arXiv ing, Michigan State University. He received the
preprint arXiv:1407.1082, 2014. IEEE & IFIP William C. Carter Award in 2004, a
[56] X. Lu et al., “Wireless charging technologies: Fundamentals, stan- National Science Foundation CAREER award in
dards, and network applications,” IEEE Communications Surveys & 2009, and the Michigan State University Withrow
Tutorials, vol. 18, no. 2, pp. 1413–1452, 2016. Distinguished Scholar Award in 2011. He has
[57] C. Lin et al., “Minimizing charging delay for directional charging served as an Editor for IEEE/ACM Transactions
in wireless rechargeable sensor networks,” in IEEE INFOCOM, on Networking, and he is currently an Associate
2019, pp. 1819–1827. Editor for IEEE Transactions on Dependable and Secure Computing
[58] “http://www.sevenoak.biz/pantiltheads/sk-ebh01.html,” 2018. and IEEE Transactions on Mobile Computing, and an Area Editor for
[59] R. Zhang et al., “MIMO broadcasting for simultaneous wireless Computer Communications. He received Best Paper Awards from ICNP-
information and power transfer,” IEEE Transactions on Wireless 2012, SRDS-2012, and LISA-2010. His research interests focus on
Communications, vol. 12, no. 5, pp. 1989–2001, 2013. networking and security.
[60] Z. Ding et al., “Application of smart antenna technologies in
simultaneous wireless information and power transfer,” IEEE
Communications Magazine, vol. 53, no. 4, pp. 86–93, 2015.
[61] L. Fleischer et al., “Tight approximation algorithms for maximum Lijun Zhang received the B.S. and Ph.D. de-
general assignment problems,” in ACM SODA, 2006, pp. 611–620. grees in Software Engineering and Computer
[62] S. Fujishige, Submodular functions and optimization. Elsevier, 2005, Science from Zhejiang University, China, in 2007
vol. 58. and 2012, respectively. He is currently an as-
[63] H. Dai et al., “Practical scheduling for stochastic event capture in sociate professor of the Department of Com-
energy harvesting sensor networks,” International Journal of Sensor puter Science and Technology, Nanjing Univer-
Networks, vol. 18, no. 1-2, pp. 85–100, 2015. sity, China. Prior to joining Nanjing University,
[64] “https://en.wikipedia.org/wiki/online algorithm,” 2019. he was a postdoctoral researcher at the De-
partment of Computer Science and Engineering,
Michigan State University, USA. His research
interests include machine learning, optimization,
information retrieval and data mining.
Jiaqi Zheng is currently an assistant researcher

from Department of Computer Science and
Technology, Nanjing University, China. His re-
search area is computer networking, particularly
data center networks, SDN, and NFV. He re-
Haipeng Dai received the B.S. degree in the De- ceived Ph.D. degree from Nanjing University in
partment of Electronic Engineering from Shang- 2017. He was an assistant researcher in the
hai Jiao Tong University, Shanghai, China, in City University of Hong Kong in 2015, and a
2010, and the Ph.D. degree in the Department visiting scholar in Temple University in 2016.
of Computer Science and Technology in Nanjing He received the best paper award from IEEE
University, Nanjing, China, in 2014. His research ICNP 2015 and Doctorial Dissertation Award
interests are mainly in the areas of wireless from ACM SIGCOMM China 2018. He is a member of ACM and IEEE.
charging, mobile computing, and data mining.
He is an associate professor in the Department
of Computer Science and Technology in Nan-
jing University. His research papers have been Guihai Chen received B.S. degree in computer
published in many prestigious conferences and journals such as ACM software from Nanjing University in 1984, M.E.
MobiSys, ACM MobiHoc, ACM VLDB, ACM SIGMETRICS, ACM Ubi- degree in computer applications from Southeast
Comp, IEEE INFOCOM, IEEE ICDCS, IEEE ICNP, IEEE SECON, IEEE University in 1987, and Ph.D. degree in com-
IPSN, IEEE JSAC, IEEE/ACM TON, IEEE TMC, IEEE TPDS, and IEEE puter science from the University of Hong Kong
TOSN. He is an IEEE and ACM member. He serves/ed as Poster Chair in 1997. He is a professor and deputy chair of the
of the IEEE ICNP’14, Track Chair of the ICCCN’19, TPC member of the Department of Computer Science, Nanjing Uni-
ACM MobiHoc’20, IEEE INFOCOM’20, IEEE ICDCS’20, IEEE ICNP’14, versity, China. He had been invited as a visiting
IEEE IWQoS’19, IEEE IPDPS’20, IEEE MASS’18-19, IEEE ICC’14-18, professor by many foreign universities including
IEEE ICCCN’15-18 and IEEE Globecom’14-18. He received Best Paper Kyushu Institute of Technology, Japan in 1998,
Award from IEEE ICNP’15, Best Paper Award Runner-up from IEEE University of Queensland, Australia in 2000, and
SECON’18, and Best Paper Award Candidate from IEEE INFOCOM’17. Wayne State University, USA during Sept. 2001 to Aug. 2003. He has
a wide range of research interests with focus on sensor networks,
peer-to-peer computing, high-performance computer architecture and
combinatorics.

Charging Task Scheduling For Directional Wireless Charger Networks

Загружено:

Сведения о документе

Исходное описание:

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Charging Task Scheduling For Directional Wireless Charger Networks

Загружено:

Авторское право:

Доступные форматы

This article has been accepted for publication in a future issue of this journal, but has not been

Charging Task Scheduling for

Index Terms—Charging task, scheduling, directional wireless chargers.

1 I NTRODUCTION beamforming), and thus enhances the power intensity in the

Corresponding Author: Haipeng Dai

2 R ELATED W ORK 3 P ROBLEM F ORMULATION

TABLE 1 where α and β are two known constants determined by

 si ri and then, if necessary, starts switching with a delay of ρ

the charging power of si is zero. Then, for a charg- si , say T i ;

Algorithm 2: Centralized Offline Algorithm to HASTE

5.1 Algorithm Description

6.1 Algorithm Description

Algorithm 3: Distributed Online Algorithm to HASTE 6.2 Theoretical Analysis

0.8 1 0.75 0.85

0.76 0.9 0.72 0.8

0.68 0.7 0.66 0.7

7.3 Centralized Offline Algorithm Evaluation 0.7

GreedyUtility and GreedyCover by 2.67% and 3.40% (at most 0.2

7.3.5 Impact of Required Charging Energy and Task Dura-

7.3.3 Impact of Switching Delay ρ 7.4 Distributed Online Algorithm Evaluation

0.8 1 0.73 0.85

Moreover, if As = 360◦ , each charger covers the same set 25

three algorithms have the same performance. The solution

three distributed online algorithms is less than that of its

7.4.2 Impact of Receiving Angle Ao Fig. 12. Communication cost

Fig. 16. Topology 2

Fig. 15. Testbed 0.8

distribution with µ = 25. Figure 13 shows that generally the 0.4

charging utility increases with either σx or σy , which indi- 0.2 HASTE(C=4)

degree of uniformness of positions, the phenomenon that Task

Jiaqi Zheng is currently an assistant researcher

Вам также может понравиться