You are on page 1of 6

2022 IEEE Wireless Communications and Networking Conference (WCNC)

Joint Coverage and Resource Allocation for


Federated Learning in UAV-Enabled Networks
Mariam Yahya∗ and Setareh Maghsudi∗†
∗ Department of Computer Science, University of Tübingen, Germany
† Fraunhofer Heinrich Hertz Institute, Berlin, Germany
2022 IEEE Wireless Communications and Networking Conference (WCNC) | 978-1-6654-4266-4/22/$31.00 ©2022 IEEE | DOI: 10.1109/WCNC51071.2022.9771880

Abstract—Thanks to its communication efficiency and low the FL procedure. Moreover, the transmission delay is affected
latency, federated learning (FL) has emerged as a promising by the UAVs’ placement and the radio resource allocation.
learning paradigm in the unmanned aerial vehicle (UAV)-enabled There are several approaches to optimizing the FL conver-
networks; nevertheless, the great potential of FL in UAV networks
is realizable only upon optimizing crucial factors such as coverage gence time in resource-limited wireless networks. Some aim
and transmission delay. In this paper, we study the problem of at reducing the data rate by implementing communication-
joint coverage optimization and efficient radio resource alloca- efficient compression algorithms [3]. Others leverage a client
tion. The objective is to minimize the convergence time of FL in selection method where only a subset of clients receive the
a UAV-enabled network, where UAVs perform learning over an network’s scarce resources for data transmission [4]. The work
inhomogeneous sensor network. To this end, we develop a method
that minimizes the FL computation and communication time in [5] exploits the signal superposition property of multi-access
in each global iteration: First, the algorithm adjusts the UAVs’ channels to allocate the limited bandwidth to maximize the
locations to control the average number of sensors associated number of participants in the model update to accelerate the FL
with each UAV to maximize the coverage and to reduce the convergence. Some recent researches consider the convergence
overall computation time. The UAVs’ locations also affect their of UAV-enabled networks. Zeng et al., for example, propose a
transmission delay. Thus, in the second step, the method uses a
fair resource allocation scheme for channel allocation and power novel framework for implementing FL within a UAV swarm
control to minimize the FL communication time while retaining [6]. They propose a joint power allocation and UAV scheduling
the efficiency of resource expenditure. approach to optimize the convergence performance.
Index Terms—Convergence time, coverage, federated learning, In this paper, we develop a novel method for joint UAV
unmanned aerial vehicles, resource allocation. placement and resource allocation. We aim to maximize the
I. I NTRODUCTION network coverage while satisfying the constraints on the
convergence delay of FL and resource consumption. The
Recently, unmanned aerial vehicle (UAV)-enabled network- challenge arises since the UAV’s placement affects both sensor
ing has emerged as a new paradigm for managing the com- coverage and transmission delay, especially in inhomogenous
plexity of processing the excessive amount of data. To this sensor networks. The core idea of our method is to equalize
end, UAVs cooperate in learning by taking advantage of the computation and communication times of UAVs through
their implemented machine learning algorithms [1]. While optimal UAV placement, spectrum allocation, and power con-
centralized machine learning algorithms might be impractical trol. We provide theoretical and numerical analysis to evaluate
due to the UAVs’ mobility and the scarcity of radio resources, the performance of our proposed scheme. The results show
distributed learning is envisioned as a potential alternative. the superior performance of our approach compared to the
Federated learning (FL) is a privacy preserving distributed state-of-the-art literature in terms of FL convergence delay,
learning algorithm. In each iteration, upon receiving the global coverage, and power consumption.
model from the server, every UAV updates the model using
the data collected by the sensors in its coverage area, then II. S YSTEM M ODEL
transmits the locally updated models to the server. The server We consider a wireless network consisting of a federated
aggregates the local parameters and, based on them, generates learning ground server located at (xs , ys , 0) and a set of U
a new global model. This process continues until the global UAVs, where i ∈ {1, ..., U } with homogeneous processing
model converges with a predetermined accuracy [2]. capabilities. Each UAV i is located at (xi , yi , hi ), and has a
FL preserves privacy and reduces the communication cost; maximum transmission power Pi . The UAVs cover a field of
nevertheless, the iterative updates and transmissions raise sensors that are randomly distributed according to an inhomo-
some challenges concerning convergence time and resource geneous Poisson point process (IHPPP) with a known density.
expenditure. The FL convergence time depends on the UAVs’ The exact locations of the sensors are however unknown. The
computation times and the transmission delays between the sensors are homogeneous in the sense that they collect and
UAVs and the server. These parameters, in turn, depend on transmit the same number of samples. The communication
the network design. Therefore, deploying UAVs to merely between the UAVs and the server is OFDM-based. There is a
maximize the network coverage does not suffice to optimize set N of N subchannels, n ∈ {1, ..., N }.
This research was supported by the German Federal Ministry of Education Each UAV i ∈ U has a set of Di data samples. By {xil }D i
l=1 ,
d
and Research under Grant 16KIS1134 and Grant 16KIS1165. xil ∈ R , we represent the set of input data samples that

978-1-6654-4266-4/22/$31.00 ©2022
Authorized licensed use limited 2476
IEEE STATE UNIV. Downloaded
to: OREGON on June 25,2022 at 05:53:56 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE Wireless Communications and Networking Conference (WCNC)

UAV i receives from the sensors in its coverage area. Besides, III. T HE E FFECT OF C OVERAGE AND R ESOURCE
{yil }D
l=1 , yil ∈ R, is the corresponding set of outputs. Let ω
i
A LLOCATION ON THE FL C ONVERGENCE T IME
be the parameter vector relating the two vectors at UAV i, and As described previously, we aim to maximize the network
fi (ω, xil , yil ) be the loss function for the lth data sample. The coverage while ensuring minimizing the FL convergence time
total loss function of UAV i, Fi (ω, xi1 , yi1 , . . . , xiDi , yiDi ), and ensuring the efficiency of resource expenditure. Before
or simply Fi (ω), is proceeding to the formal problem statement, we analyze
Di the relationship between the computation time and network
1 X coverage and the relationship between the communication time
Fi (ω) := fi (ω, xil , yil ). (1)
Di and resource allocation.
l=1

The global loss function is defined as [2] A. FL Computation Time and Network Coverage
U U Di The computation time of UAV i in each global iteration m,
Di 1 XX
Ticp , is the time required to update the model using the data
X
F (ω) := Fi (ω) = fi (ω, xil , yil ), (2)
i=1
D D i=1 received from the sensors in the UAV’s coverage region. Let
l=1
PU C be the number of CPU cycles required for computing one
where D = i=1 Di . The objective of FL is to find a shared data sample. Besides, fCPU is the UAVs’ CPU frequency. Then
model ω ∗ that minimizes the global loss function we have
CDi
U Di
1 XX Ticp = I , ∀i ∈ U, (8)
ω ∗ = arg min F (ω) := fi (ω, xil , yil ). (3) fCPU
ω∈Rd D i=1
l=1 where Di is the number of data samples at UAV i and I is
We adopt the FEDL algorithm [2] to solve the objective the number of local iterations as given in (7).
function in (3). A summary of the procedure is provided in As a consequence of the inhomogeneous sensor distribution,
Algorithm 1. the number of sensors covered by each UAV can differ signif-
icantly depending on its location. Such a difference results in
Algorithm 1 The federated learning procedure [2] long computation times for the UAVs covering dense areas and
1: In the first global iteration m = 0, the server generates and vice versa. In a synchronous FL scheme, the server requires
broadcasts an initial global model ω 0 to all UAVs. It also the input of all UAVs for global aggregation. As such, the
determines the local accuracy parameter κ ∈ (0, 1). model update time is determined by the UAV with the highest
2: In the following global iterations, each UAV i receives ω m−1 Ticp . Since the UAVs are homogeneous, that UAV is the one
and ∇F̄ m−1 from the server and solves ω m m
i = arg min Ji (ω),
ω∈Rd
that covers the highest number of sensors and, consequently,
has the highest number of data samples
Jim (ω) := Fi (ω) + ⟨η∇F̄ m−1
− ∇Fi (ω m−1
), ω⟩, (4)
T cp = max Ticp . (9)
where ⟨., .⟩ is the inner product of two vectors, η is the learning i∈U
rate. The resulting ω mi satisfies the accuracy κ ∈ (0, 1), i.e., B. FL Communication Time and Resource Allocation
||∇Jim (ω m m
i )|| ≤ κ||∇Ji (ω
m−1
)||, ∀i ∈ U , (5) At iteration m, the uplink data rate of UAV i is given by
κ = 0 solves the local problem optimally, whereas κ = 1 means ri,m (βin,m , pin,m , xi , yi , hi ) =
that the local model doesn’t change.
pin,m |gin,m |2
 
3: The UAVs send their local models ω m and local gradients X
∇F m
i
aggregates them to calculate ω m =
B U L βin,m log 1 + , (10)
PUi (ωDi i ) to the server. It P βin,m Li σ 2
m
i=1 D ω i and ∇F̄
m
= U Di m
i=1 D ∇F̄i (ωi ). The server then
n∈N
transmits these values to all UAVs. where B U L is the uplink subchannel bandwidth and gin,m
4: The global iterations, i.e., the collection of updating and transmis-
sion processes are repeated until convergence, i.e., until achieving
is the fading channel gain between UAV i and the server
a certain difference between the current and the minimal loss on subchannel n. Besides, βin,m represents the fraction of
function. Formally, F (ω m ) − F (ω ∗ ) ≤ ϵ. subchannel
P n ∈ {1, ..., N } allocated to UAV i such that
n∈N β in,m ≤ 1. We allow only for full subchannel al-
The time required for one global iteration yields location so that βin,m ∈ {0, 1}. Furthermore, pin,m is the
transmission power of UAV i on subchannel n such that
Tglobal = T cm + T cp (κ). (6)
P
n∈N pin,m ≤ Pi . Pi is the maximum power constraint for
In (6), T cm is the communication time between the UAVs UAV i. The path loss Li between UAV i and the server is
and the server. Besides, T cp is the computation time required determined by the air-to-ground channel model [7]
to achieve the accuracy κ. The number of iterations for the A
LdB
i (xi , yi , hi ) =    
convergence of the local model is 1 + a exp −b[ 180 −1 hi
− a]
π tan di
α2
I(κ) = α1 log . (7) + 10 log d2i + h2i + Y,

κ (11)
The constants α1 and α2 depend on the condition number of where A = ηLoS − ηNLoS is the mean loss incurred in addition
the Hessian matrix of Fi (.) [2]. to free space path loss due to shadowing and scattering for the

2477
Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:53:56 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE Wireless Communications and Networking Conference (WCNC)

LoS and NLos paths. Besides, a and p b are constants that de- (i) Computation time minimization through optimal UAV
pend on the environment [8]. di = (xi − xs )2 + (yi − ys )2 placement while ensuring the maximum coverage, and (ii)
is the horizontal distance between
 the ground server and UAV Communication time minimization through fair resource al-
i. Furthermore, Y = 20 log 4πf c
c
+ ηNLoS . Finally, fc is the location while ensuring the optimal resource expenditure. A
carrier frequency and c is the speed of light. detailed description follows in Section IV and Section V,
Different UAV-server links have different lengths and qual- correspondingly.
ities. Therefore, they can be significantly distinct concerning IV. C OMPUTATION T IME M INIMIZATION T HROUGH UAV
data rates and transmission time, as reflected by (10). Here C OVERAGE
UL
we consider the transmission times in both the uplink, Ti,m ,
DL
Based on (9), one can minimize the computation time of
and downlink, Ti,m [6]. Let SL,i be the total size of the local FL by shrinking the coverage area of UAVs in dense areas
model and the gradient of UAV i and let SG be the total size to reduce the number of covered sensors. While UAVs in
of the global model and the gradients in the server, then low-density areas must increase their coverage area to max-
UL SL,i DL SG imize coverage; nevertheless, this solution is challenging to
Ti,m = , Ti,m = . (12)
ri,m ri,m implement due to the inhomogeneous sensor distribution and
the unknown sensors’ locations. In this section, we propose
Due to the channel reciprocity and without loss of generality,
a novel heuristic algorithm to solve this problem. Before
we assume that the uplink and downlink data rates are equal.
proceeding to describe this algorithm, we explain the physical
We define the FL communication time as the time required
parameters that determine UAV coverage.
to communicate the model parameters and the gradients be- Every UAV i ∈ U located at (xi , yi , hi ) has a disk-shaped
tween the UAVs and the server. Similar to the computation coverage region with radius Ri . Let LdB th be the maximum
time, the total time required to complete an iteration is allowable path loss for any sensor. Then LdB th determines the
determined by the longest transmission delay; that is, maximum coverage radius of a UAV, denoted by Rmax . More
cm UL DL
 precisely, any sensor located at a horizontal distance below
Tm = max Ti,m + Ti,m . (13) Rmax to a UAV i ∈ U experiences a path loss less than LdB
i∈U th
when communicating to that UAV; In this case, we say that
C. Problem Formulation UAV i covers that sensor. The relation between the threshold
Our objective is to minimize the time needed to complete path loss LdB
th and the maximum coverage radius Rmax is
 
one global iteration of FL in UAV-enabled networks. A Rmax
LdB
th =  +20 log +Y. (22)
1 + a exp −b[ 180 θ − a]
π opt
cos(θopt )
minimize (T cp + Tm
cm
) (14)
βin,m ,pin,m ,xi ,yi ,hi In (22), θopt is the optimal elevation angel, which is an
subject to environment-dependent factor that determines the optimum
altitude to maximize the coverage radius [8]. For example,
pin,m ≥ 0, i ∈ U, n ∈ N , (15) the optimal elevation angle for an urban environment θopt =
42.44◦ . The optimal altitude for Rmax yields
X
pin,m ≤ Pi , ∀i ∈ U, (16)
n∈N
hopt = Rmax tan(θopt ). (23)
βin,m ∈ {0, 1}, i ∈ U, n ∈ N , (17)
X A. Problem Statement
βin,m ≤ 1, n ∈ N, (18)
The average number of sensors covered by UAV i ∈ U is a
i∈U
function of the UAV’s location. Formally,
xmin ≤ xi ≤ xmax , i ∈ U, (19) Z Z
ymin ≤ yi ≤ ymax , i ∈ U, (20) n̄i = λ(x, y)dx dy, ∀i ∈ U, (24)
hmin ≤ hi ≤ hmax , i ∈ U, (21) Ai

where T cp and Tm cm
are given by (9) and (13), respectively. where λ(x, y) is the IHPPP density function and Ai is
Constraints (19) and (20) ensure that the UAVs are within the UAV i’s coverage area. Thus, even for UAVs with the same
field boundaries [xmin , xmax ] and [ymin , ymax ]. The constraint coverage radius, the value of n̄i , thereby the computation
in (21) ensures that the UAV’s altitude remains within the time, can differ drastically. Our objective is to maximize the
technically feasible interval [hmin , hmax ]. total number of sensor covered by all UAVs while keeping
Solving optimization problem (14) is challenging for several n̄i bounded between [Nmin , Nmax ]. The values of Nmin and
reasons. First, both T cp and Tm cm
depend on the UAVs’ Nmax depend on the sensor density, number of UAVs, and
locations. That implies that placing the UAVs to minimize the UAVs’ computation capabilities. The upper bound prevents
the computation time and improve coverage can change the long computation times, whereas the lower bound determines
UAV-server channel status which affects the transmission time. the coverage; hence the difference Nmax − Nmin balances
Besides, the full (binary) allocation of OFDM subchannels the trade-off between maximizing coverage and equalizing the
renders subchannel scheduling a hard integer programming UAVs’ computation delays. The problem is formally stated as
problem. Furthermore, the UAVs’ power constraints can vary.
X
maximize n̄i (xi , yi , hi ) , (25)
Hence, we solve (14) by decomposing it into two parts: xi ,yi ,hi
i∈U

2478
Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:53:56 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE Wireless Communications and Networking Conference (WCNC)

4 4 RD(j) reduces by ∆R . Then it shifts UAV D(j) in proportion


to ∆R in the direction of higher node density. The procedure
continues until Nmin ≤ n̄D(j) ≤ Nmax . Afterwards, the
y (km)

y (km)
2 2 algorithm shifts the remaining UAVs in the direction of
RCP RCP increasing sensor density without overlapping.
Rmax Rmax,shifted
• If n̄D (j) < Nmin , the algorithm increases hD (j) by a constant
0 0
0 1 2 3 4 0 1 2 3 4 ∆h , with hD(j) +∆h ≤ hmax . Thus, RD(j) increases by ∆R ,
x (km) x (km)
subject to the constraint RD(j) + ∆R ≤ Rmax . Then it shifts
(a) Before shifting the UAVs (b) After shifting the UAVs UAV D(j) in proportion to ∆R in the direction of decreasing
N̄total = 193.23 nodes. N̄total = 275.97 nodes.
sensor density. This continues until Nmin ≤ n̄D(j) ≤ Nmax .
Fig. 1: Illustration of the effect of shifting the UAVs. Finally, the algorithm shifts the remaining UAVs in the di-
rection of the decreasing sensor density without overlapping.
Step 5: The algorithm recalculates n̄i , ∀i ∈ U −
subject to {D(1) , . . . , D(j) }. The same procedure is repeated for the
(19), (20), (21), (26) remaining UAVs D(j+1) , D(j+2) , . . . , D(U ) . We summarize
Nmin ≤ n̄i ≤ Nmax , ∀i ∈ U , (27) the procedure in Algorithm 2.
p
(xi − xj )2 + (yi − yj )2 ≥ Ri + Rj ∀i, j ∈ U , i ̸= j. (28) TABLE I: Equal CP in a square of edge length E
Constraint (28) is to avoid an overlap between the UAVs’ Number of UAVs U Coverage radius RCP Total coverage (%)
1√
coverage regions. 2
2+ 2
E 53.90
3 √2 √ E 60.96
B. Solution 4+ 2+ 6
1
4 E 78.54
To solve (25), we propose the algorithm described below. 1
√4
5 2
( 2 − 1)E 67.38
Step 1: The algorithm determines the UAVs’ initial placement
(xi,CP , yi,CP ) and coverage radius RCP using the well-known C. Complexity Analysis
circle packing (CP) method [9]. The CP method finds the The CP problem is intractable, as there is no general method
locations and the maximum radius of U equal circles that one for packing U circles in a square. However, the solution for CP
can pack in a surface to maximize the packing density without is listed in tables for a large number of circles, it is optimal
overlapping. Table I shows the maximum radius as a function in a few cases and approximate in the rest [10]. Therefore,
of the square length E for different number of UAVs. the CP part renders a simple value assignment operation. The
Step 2: Sensors located outside Rmax experience high path remaining part of the algorithm requires solving U (U2−1) non-
loss and are not covered by any UAV. Therefore, the UAVs’ linear problems with non-convex constraints. An approximate
radii are adjusted to R0 = min{RCP , Rmax }. solution is found by dividing the area into convex sub-regions.
Step 3: The locations of the UAVs depend on R0 as follows:
• If R0 = RCP , the UAVs remain at the locations determined
V. C OMMUNICATION T IME M INIMIZATION T HROUGH
by CP, i.e., (xi,CP , yi,CP ). Besides, h0 = R0 tan(θopt ). FAIR R ESOURCE A LLOCATION
• If R0 = Rmax , some gaps will result from reducing the ra- The lengths and quality of the UAV-to-server communica-
dius from RCP to Rmax . Therefore, we shift the UAVs in the tion links can vary significantly between the UAVs; thus, the
direction of increasing density in proportion to (RCP −Rmax ) communication times of the UAVs become crucially different.
to increase coverage. Besides, h0 = R0 tan(θopt ). This case To reduce this difference, we implement a joint subchannel
is illustrated in Fig. 1 for a scenario with four UAVs. scheduling and power control algorithm to maintain fairness
Here, we select λ(x, y) = 3(x2 + y 2 ), RCP = 1 km and among the UAVs that participate in the learning process while
Rmax = 0.707 km. If located at the position determined by retaining the efficiency of resource expenditure. At each round
CP, the average number of sensors that the UAVs cover is m and for each UAV i ∈ U, we define a utility function ui as
N̄total = 193.23, as shown in Fig. 1a. Shifting the UAVs
(
1
(Wi,m )α α ≤ 1, α ̸= 0
in the direction of higher density, as shown in Fig. 1b, ui (Wi,m ) = α . (29)
improves the coverage so that N̄total = 275.97. log(Wi,m ) α = 0
Step 4: Up to this point, all UAVs have the same radius R0 The utility function in (29) depends on the average throughput
and altitude h0 but different horizontal positions (xi , yi ), ∀i ∈ so far, Wi,m , and α is the fairness parameter. Namely, α = 0
U. In this step, the UAVs’ radii Ri are adjusted according gives the proportionality fair (PF) rule whereas α = 1 results
in a maximum throughput scheduling.
to the number of sensors covered by each UAV. First, the To guarantee fairness, we use gradient-based scheduling to find
algorithm determines the UAV with the highest n̄i using (24), the uplink data rate r ∗m = {r1,m
∗ ∗
, ..., rU,m } that maximizes the
and denotes this UAV by D(1) , where D(j) is the UAV with weighted sum of the UAVs’ data rates, where the weights are
the jth highest average number of sensors. Then, the following the gradients of the UAVs’ utility functions. Formally,
two cases are considered: X ∂ui (Wi,m ) X
maximize ri,m = maximize (Wi,m )α−1 ri,m .
• If n̄D (j) > Nmax , the algorithm reduces the altitude hD (j) βin,m ,pin,m ∂W i,m βin,m ,pin,m
i∈U i∈U
by a constant ∆h , with hD(j) − ∆h ≥ hmin . Consequently, (30)

2479
Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:53:56 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE Wireless Communications and Networking Conference (WCNC)

Algorithm 2 Joint coverage maximization and computation time sin , which ensures the max-min fairness. In other words, the
minimization UAV with the weakest channel uses its maximum available
1: Input: U , E, λ(x, y), Lth , θopt , Nmax , Nmin . power, whereas other UAVs reduce their power consumption
2: Output: x, y, h.
3: Find the maximum coverage radius Rmax using (22).
to save energy. Formally, it aims at
 
4: Find the CP radius RCP using Table I.
5: ∀i ∈ U find the initial placement (xi , yi ) using the circle packing maximize minimum ri,m (33)
pin,m i∈U
theory.
6: if Rmax ≥ RCP then subject to constraints (15) and (16).
7: Ri = RCP . The UAVs’ locations do not change.
8: else A. Complexity Analysis
9: Ri = Rmax . The computational complexity of subchannel allocation is
10: Shift the UAVs in the direction of the higher node density
O(N U + N log N ) and the complexity of the water filling
without overlapping.
11: end if algorithm is O(N U ) [11], resulting in a total complexity of
12: hi = Ri tan(θopt ). O(N U + N log N ) for resource allocation.
13: for j = 1 : U do
14: Find the UAV with the jth highest average number of sensors, VI. N UMERICAL R ESULTS
D(j) . In this section we evaluate the proposed algorithm’s FL
15: if n̄D(j) > Nmax then convergence time and coverage probability. We consider a 2
16: while n̄D(j) > Nmax do
km × 2 km square region with IHPPP distributed sensors with
17: Reduce RD(j) by ∆R by reducing hD(j) by ∆h .
18: Shift UAV D(j) towards higher node density in propor- density function λ(x, y) = 15(x2 + y 2 ) sensor/km2 and four
tion to ∆R . UAVs. Besides, we select ηLoS = 1 dB, ηN LoS = 20 dB,
19: Update the location of UAV D(j) , (xD(j) , yD(j) ). a = 9.61, b = 0.61, fc = 2 GHz, and c = 3 × 108 m/s.
20: Find n̄D(j) using (24). For each UAV, the CPU frequency is fCPU = 2 GHz and
21: end while C = 20 cycles/bit. The size of a data sample collected from
22: Shift the remaining UAVs in the direction of higher node
density without overlapping. each sensor is 200 Kb. The training size Di at UAV i in bits
23: else if n̄D(j) < Nmin then is the product of n̄i and the data size of each sensor. There
24: while n̄D(j) < Nmin do are 20 OFDM subchannels, each with a bandwidth B UL = 1
25: Increase RD(j) by ∆R by increasing hD(j) by ∆h such MHz. The maximum power constraint is 1 W for all UAVs.
that RD(j) ≤ Rmax . The channel gains follow the Ricean distribution. The path
26: Shift UAV D(j) towards its initial position in proportion loss is obtained by substituting the UAVs’ locations into (11).
to ∆R .
27: Update the location of UAV D(j) , (xD(j) , yD(j) ). We select α = 0 for proportionally fair resource allocation.
28: Find n̄D(j) using (24). First, we evaluate the FL computation time for our algorithm
29: end while and compare it to the circle packing and also the k-means
30: Shift the remaining UAVs in towards their initial positions clustering [12] methods. The k-means clustering method aims
without overlapping. at maximizing the sensor coverage by dividing the sensors into
31: end if
32: end for U clusters. It then finds the maximum UAV coverage radius
and the placement that fits into these subareas to maximize
coverage. According to Table I, we have RCP = 0.5 km. By
Substituting the value of ri,m from (10) gives: (22), the maximum coverage radius that corresponds to Lth =
100 dB is equal to Rmax = 0.707 km. Since RCP ≤ Rmax ,
pin,m |gin,m |2
 
maximize
X
(Wi,m ) α−1
X UL
B βin,m log 1 + ,
the initial radius and position of the UAVs in our proposed
β 2
βin,m ,pin,m in,m Li σ algorithm are equivalent to the CP method.
i∈U n∈N
(31) Fig. 3a shows the computation time for one local iteration
subject to (15), (16), (17), and (18). at each of the four UAVs, for the three coverage methods.
We adopt the approach in [11] to solve problem (31) for In CP, the four UAVs have equal radius RCP , but n̄i varies
the uplink. The method starts with a subchannel allocation drastically between them. This is due to the inhomogeneous
phase, where each subchannel is allocated to one UAV at most. nodes distribution that causes a significant difference in the
Then, each UAV uses the water-filling concept to allocate the computation time, thereby slowing down the FL process. The
power to the subchannels assigned to it, subject to its power same holds for the k-means method. In contrast, our proposed
constraint. As a result, scheme guarantees that n̄i ∈ [18, 25]; therefore, the computa-
 ∗ + ! tion times of the UAVs differ only slightly. Most importantly,
βin,m 1 sin

pin,m = min − , , (32) the maximum delay, which affects the FL, decreases. Fig. 3b
νi,m cin,m ein,m shows the elapsed time in updating the local model in each
iteration. We show the UAV with the highest computation
|g |2
where cin,m = Lin,m iσ
2 . Besides, νi,m , ∀i ∈ U satisfies the delay as it determines the update time in synchronous FL.
UAVs’ power constraint n∈N p∗in,m = Pi .
P
Fig. 4 shows the coverage probability for the three UAV
The power allocation scheme given by (32) corresponds to a placement methods. Naturally, the proposed algorithm mini-
constraint on the maximum SNR on subchannel n for UAV i, mizes the computation delay at the expense of reducing the

2480
Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:53:56 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE Wireless Communications and Networking Conference (WCNC)

10
2 2 2 PF with sin = 0 dB 20 PF with sin = 0 dB
1.5 1.5 1.5 8 PF with sin = ∞ PF with sin = ∞
16

Data rate (Mbps)


y (km)
y (km)

Tcm (ms)
y (km)
1 1 1
6 12
0.5 0.5 0.5

0 0 0
4 8
0 0.5 1 1.5 2 0 0.5 1 1.5 2 0 0.5 1 1.5 2
x (km) x (km) x (km)
2 4
(a) Circle packing (b) K-means clustering (c) Proposed algorithm
0 0
1 2 3 4 1 2 3 4
Fig. 2: Three UAV placement methods. UAV number UAV number
(a) Data rates (b) Communication time
Computation time (ms)
120 60 1500
Number of covered sensors

Proposed algorithm Proposed algorithm


K-means clustering K-means clustering PF with sin = 0 dB
Computation time (ms)

100 50

Total power transmitted (W)


Circle packing Circle packing
PF with sin= ∞
80 40 1000
1
60 30

40 20 500
20 10 0.5
0 0
0
1 2 3 4 2 4 6 8 10
UAV Number Iteration number
0
(a) The computation time for one (b) Cumulative computation time 1 2 3 4
local iteration for the four UAVs. in one global iteration. UAV number

Fig. 3: FL computation time (c) Transmission power


Fig. 5: Resource allocation.

coverage probability. Its performance can be improved by


increasing Nmax ; nevertheless, this increases T cp . It should be R EFERENCES
noted that the k-means method, despite the highest coverage [1] B. Brik, A. Ksentini, and M. Bouaziz, “Federated learning for UAVs-
probability, results in high computation delay. Besides, it enabled wireless networks: Use cases, challenges, and open problems,”
is extremely costly in terms of power consumption, as the IEEE Access, vol. 8, pp. 53841–53849, 2020.
[2] C. T. Dinh, N. H. Tran, M. N. H. Nguyen, C. S. Hong, W. Bao, A. Y.
UAVs’ radii are maximized. Indeed, the UAVs are located to Zomaya, and V. Gramoli, “Federated learning over wireless networks:
maximize the coverage even if a significant part of the UAVs’ Convergence analysis and resource allocation,” IEEE/ACM Transactions
coverage region lies outside the square field. on Networking, vol. 29, no. 1, pp. 398–409, 2021.
[3] F. Sattler, S. Wiedemann, K.-R. Müller, and W. Samek, “Robust and
After the UAV placement as shwon in Fig. 2c, the resource communication-efficient federated learning from non-i.i.d. data,” IEEE
allocation procedure starts. Fig. 5a shows the data rate between Transactions on Neural Networks and Learning Systems, vol. 31, no. 9,
pp. 3400–3413, 2020.
the UAVs and the server in one global iteration, averaged [4] W. Zhang, X. Wang, P. Zhou, W. Wu, and X. Zhang, “Client selection for
over 1000 rounds. From Fig. 5b, the proposed method results federated learning with non-iid data in mobile edge computing,” IEEE
in homogeneous data rates among the UAVs, and thereby Access, vol. 9, pp. 24462–24474, 2021.
[5] K. Yang, T. Jiang, Y. Shi, and Z. Ding, “Federated learning via over-
homogeneous communication times. The latter figure shows the-air computation,” IEEE Transactions on Wireless Communications,
the sum of the communication times of the local data of vol. 19, no. 3, pp. 2022–2035, 2020.
size SL,i = 36 kbit, ∀i ∈ U and the global data of size [6] T. Zeng, O. Semiari, M. Mozaffari, M. Chen, W. Saad, and M. Bennis,
“Federated learning in the sky: Joint power allocation and scheduling
SG = 36 kbit. The homogeneous data rates and communi- with UAV swarms,” in ICC 2020 - 2020 IEEE International Conference
cation times are achieved by maximizing the power allocated on Communications (ICC), pp. 1–6, 2020.
to the UAV with the minimum data rate as shown in Fig. 5c. [7] M. Alzenad, A. El-Keyi, F. Lagum, and H. Yanikomeroglu, “3-D
placement of an unmanned aerial vehicle base station (UAV-BS) for
energy-efficient maximal coverage,” IEEE Wireless Communications
Letters, vol. 6, no. 4, pp. 434–437, 2017.
[8] A. Al-Hourani, S. Kandeepan, and S. Lardner, “Optimal LAP altitude
1 Proposed algorithm for maximum coverage,” IEEE Wireless Communications Letters, vol. 3,
Coverage Probability

K-means clustering no. 6, pp. 569–572, 2014.


0.8 Circle packing [9] P. Szabó, M. Markót, and T. Csendes, Global Optimization in Geometry
— Circle Packing into the Square, pp. 233–265. Boston, MA: Springer
0.6 US, 2005.
[10] E. Specht, “The best known packings of equal circles in a square (up
0.4 to n = 10000).” http://hydra.nat.uni-magdeburg.de/packing/csq/csq.html.
(Accessed: 08.09.2021).
0.2 [11] J. Huang, V. G. Subramanian, R. Agrawal, and R. Berry, “Joint schedul-
ing and resource allocation in uplink OFDM systems for broadband
0 wireless access networks,” IEEE Journal on Selected Areas in Commu-
2 3 4 nications, vol. 27, no. 2, pp. 226–234, 2009.
Number of UAVs [12] J. Sun and C. Masouros, “Drone positioning for user coverage maxi-
mization,” in International Symposium on Personal, Indoor and Mobile
Fig. 4: Coverage probability for different number of UAVs. Radio Communications, pp. 318–322, 2018.

2481
Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:53:56 UTC from IEEE Xplore. Restrictions apply.

You might also like