1

On Greedy Routing in Dynamic UAV Networks

Mehrdad Khaledi, Arnau Rovira-Sugranes, Fatemeh Afghah, and Abolfazl Razi
School of Informatics, Computing and Cyber Systems, Northern Arizona University,
Email:{[Link],ar2832,[Link],[Link]}@[Link]

Abstract—Unmanned aerial vehicles (UAVs), commonly known have been proposed recently. For instance, in vehicular Ad-Hoc
as drones, are becoming increasingly popular for various appli- networks (VANETs), a routing protocol based on the Dijkstra’s
cations. Freely flying drones create highly dynamic environments, algorithm [8] and an optimized multicast routing protocol [9]
where conventional routing algorithms which rely on stationary have been developed. In [10], a stable routing protocol is
network contact graphs fail to perform efficiently. Also, link es-
presented which finds the most stable path by considering
arXiv:1806.04587v1 [[Link]] 4 Jun 2018

tablishment through exploring optimal paths using hello messages

velocity, direction and the link expiration time using fuzzy
(as is used in AODV algorithm) deems extremely inefficient and
costly for rapidly changing network topologies.
logic. In addition, learning-based routing protocols (e.g. Q-
In this paper, we present a distance-based greedy routing routing) have been proposed to learn the link states based on
algorithm for UAV networks solely based on UAVs’ local ob- the current transmission experience [11]. Such algorithms rely
servations of their surrounding subnetwork. Thereby, neither a on the assumption of low speed vehicles moving in a confined
central decision maker nor a time consuming route setup and two dimensional space with limited and predefined movement
maintenance mechanism is required. To evaluate the proposed patterns dictated by obstacles and roads. These assumptions,
method, we derive an analytical bound for the expected number however, are not realistic in FANETs with UAVs moving freely
of hops that a packet traverses. Also, we find the expected in space with potentially high speeds.
end-to-end distance traveled by each packet as well as the
For dynamic UAV networks, however, there have been few
probability of successful delivery. The simulation results verify
efforts on routing algorithm design. For instance, the authors
the accuracy of the developed analytical expressions and show
considerable improvement compared to centralized shortest path
in [12], presented the RARP protocol which utilizes GPS
routing algorithms. information to estimate the duration of a path. This algorithm
is based on AODV [13] and requires a route setup phase before
I. I NTRODUCTION transmission. To avoid repeated route setups the algorithm
Unmanned Aerial Vehicles (UAV) have recently attracted assumes that nodes keep their current movement patterns for
significant interest in many civilian, commercial and military some period of time which is a typical assumption in mobile
applications. The popularity of UAVs emanate from their low Ad-Hoc networks (MANETs). This assumption is not valid for
cost, rapid deployment, and ability to fly above obstacles. network of autonomous UAVs. This protocol is also evaluated
It is anticipated that the global market revenue of UAVs using a mobility model primarily developed for low-speed
reach $11.2 billion by 2020 [1]. Some applications require terrestrial nodes, hence not suited for freely flying drones.
deployment of a large number of UAVs to complete designated Another recently proposed routing protocol for UAV networks
tasks [2]–[4]. A network of autonomous UAVs form a flying is the Predictive-OLSR [14] which utilizes GPS coordinates to
Ad-hoc network (FANET), which provides more sensing and estimate the quality of a link. However, they adopt a restrictive
actuation capabilities. However, it also poses many challenges assumption that only the source node is mobile.
in designing networking protocols. In this paper, we present a distance greedy routing algorithm
One of the main challenges in FANETs is the high degree for dynamic UAV networks. This algorithm relies on local
of mobility which causes frequent changes in network topol- forwarding decisions and does not require a route setup phase.
ogy. As such, conventional communication protocols face a This low-complexity algorithm imposes no additional signaling
considerable performance degradation [5]–[7]. In routing, for overhead to the system, hence well suited to dynamic UAV
instance, link breakage and frequent topology changes can networks. In order to analyze the performance of the greedy
lead to packet loss, excessive re-transmissions and eventually routing, we derive analytical lower and upper bounds for the
increased delay. expected number of hops that a packet traverses. Using this
For environments with low degrees of mobility or with result, we find the expected end-to-end distance that a packets
geographically confined movements, several routing protocols travels using the greedy algorithm, which can be a ground to
 2

optimize distance-based performance metrics such as end-to-

end delay and transmission power consumption. We also derive
the probability of successful delivery which is the probability
that all intermediate nodes along the path can forward the
packet to a node closer to the destination. Finally, simulation
results verify the accuracy of the derived analytical results
and also indicate superior performance of the greedy routing
compared to the conventional shortest path routing based on Fig. 1. The intermediate node, ni , forwards the packet to a neighbor which
Dijkstra’s algorithm [15]. is closer to the destination, t, than its other neighbors (ni+1 = a here). The
shaded area called progress area which is the valid locations for neighbors
II. S YSTEM M ODEL that have lower distances to the destination than the current distance, Di .
We consider a FANET with N UAV nodes (n1 , n2 , . . . , nN ), a neighbor does not exist the current session fails and the
distributed uniformly in a L×L rectangular area. Let R denote packet journey is re-initiated. This algorithm continues until
the radius of the circular communication range of a UAV node, the packet is delivered to the destination. This constraint is
then the set of neighbors for node ni is defined as: required to ensure a loop-free path towards the destination. In
n q o this way, we guarantee a progress at each step provided that
Si = nj : dij = (xi − xj )2 + (yi − yj )2 ≤ R there is at least one neighbor in the progress area.
where dij denotes the Euclidean distance between nodes ni Fig. 1, represents one iteration of this algorithm. The shaded
and nj . Also, we use the popular mobility model for UAV area represent the valid locations for the next node. This
nodes which integrates linear and circular motions [16]. Transi- progress area is the intersection of two circles centered at
tion between two mobility modes occur based on an underlying ni and t with radii R and Di , respectively. Here, Di is the
Markov process with adjustable transition probabilities. The remaining distance to the destination, once the packet reaches
mobility parameters include speed and direction of linear node ni . The algorithm chooses a node which makes the
motion, and initial phase, angular speed, and radius of circular highest progress towards the destination (here ni+i = a).
motion. Speed and radii are drawn from exponential distribu-
tions and the rest of parameters are uniformly distributed. The IV. A NALYSIS OF T HE G REEDY ROUTING A LGORITHM
parameters are initialized when transition occurs between the In this section, we evaluate the performance of the greedy
two states and remain constant until the next transition. routing algorithm through several steps. We first bound the
UAV nodes do not keep track of entire network topology, expected number of hops a packet travels from source to
but each node needs to know the location of its neighbors. destination. Next, using the results we find the expected end-to-
We assume the anticipated locations of neighbors can be end distance traversed by the greedy algorithm. Finally, we find
predicted by a node, either through a model-based motion the probability of successful delivery in terms of transmission
trajectory prediction method [17] or exploiting online path- range, number of nodes and the size of the area through finding
planning information by UAVs [18]. Also, since the algorithm the probability of having at least one node in the progress area
works based on the remaining distance to the destination, the throughout the selected path.
source node needs to know the location of the destination.
This information can be embedded into the packet to inform A. Analysis of The Number of Hops
the intermediate nodes about the destination’s location. In order to bound the expected number of hops for sending a
III. T HE D ISTANCE G REEDY ROUTING A LGORITHM packet from source to destination, we first find the probability
distribution of the progress made at each hop.
Consider a source node s who wants to send a packet to a
Let ADi (R, Di ) denote the area of the shaded region in
destination node t which is located at distance D. The distance
Fig. 1 which is the intersection of two circles with centers at
greedy algorithm works based on a simple forwarding rule. At
ni , t and with radii R and Di , respectively. Also, let Xi be
each step, i, the packet is passed by the current node ni to
a random variable representing the remaining distance to the
a neighbor node nj which is closest to the destination, i.e.
destination at the ith hop. In fact, Xi represents the distance
ni+1 = argmin dj,t . The packet is passed only if the next
nj ∈Si from destination t to its closest node in the shaded area. The
node makes a progress, (i.e. if the next node is closer to probability of Xi being at least x equals the probability that
the destination than the current node: di+1,t < di,t ). If such there are no nodes in the area ADi (R, x) which is the overlap
 3

of two circles with centers at ni and t, which are at distance t, which are at distance Di of each other, and with radii R
Di of each other, and with radii R and x, respectively. Since, and Di − R, respectively.
nodes are distributed uniformly, the number of nodes in any Now that we have the average progress at each hop, we
region with area A follows a Binomial distribution with N find the number of hops that a packet traverses to reach a
trials and success probability of LA2 . Thus, we find destination located at distance D of the source node. The
 N number of hops is of this form; n = m + 1 where m is the
ADi (R, x)
P [Xi ≥ x] = 1 − , (1) number of hops needed to reach the communication range of
L2 the destination. That means, the first m hops takes the packet
where Di − R ≤ x < Di and we can use geometric analysis to the destination’s communication range where there is only
to find the area ADi (R, x) as one hop left to the destination node. We have:
 D 2 + R 2 − x2  m−1
X m
X
ADi (R, x) = R2 cos−1 i
Yi < D − R ≤ Yi . (6)
2Di R i=1 i=1
 2 2 2
2 −1 Di + x − R
+ x cos It should be noted that m is a stopping time step with respect
2Di x
1p to the sequence Yi . That means, at time m we have enough
− (R − Di + x)(Di − R + x)(Di + R − x)(Di + R + x). information to stop and we do not need any future information
2
(2) to decide. For a special case of stopping times when the
sequence of random variables are independent and identically
Now, we can find the probability distribution of the progress
distributed (i.i.d.), we can utilize the Wald’s equation [19] to
made at each hop. Let Yi = Di − Xi denote the progress the
find the sum of random variables up to time m.
ith hop, we have:
Lemma 1 (Wald’s Equation [19]): If τ is a stopping time


 0 y<0 with respect to an i.i.d. sequence {Xi : i ≥ 1}, and if E[τ ] <
ADi (R,D−y) N ∞ and E[X] < ∞, then
 
P [Yi ≤ y] = 1− L2 0 ≤ y ≤ R (3)

 X τ 
1 y>R

E Xi = E[τ ] E[X].
i=1
The probability density function (PDF) of Yi can be computed
by taking the derivative of its distribution function in (3). Note In our case, however, the sequence Yi is not i.i.d., therefore,
that there is a discontinuity point at Yi = 0, therefore we can we cannot directly use the Wald’s equation. For this reason,
write the PDF as: we first find the number of hops using some i.i.d. random
variables Zi . Next, we replace Zi with i.i.d. random variables
fYi (y) = P [Yi 6= 0]fYci (y|Yi 6= 0) + P [Yi = 0]δ(y), (4) that upper bound and lower bound Yi . Thereby, we conclude
about the bounds on the number of hops a packet travels. Using
where fYci (y|Yi 6= 0) denotes the continuous part conditioned
Lemma 1 for i.i.d. random variables Zi , we have
on progress, which is the derivative of the distribution function
for Yi between 0 and R. Knowing the PDF, we can find the m
X 
expected progress at the ith hop as follows: E Zi = E[m] E[Z]. (7)
i=1
Z R
E[Yi ] = y fYi (y|Yi 6= 0) dy. from the inequality in (6), we have:
0 Xm 
Using integration by parts, we have: E Zi ≥ D − R. (8)
R i=1
 ADi (R, Di − y) N
E[Yi ] = y 1− Combining (7) and (8) we get:
L2 0
Z R D−R
ADi (R, Di − y) N E[m] ≥ . (9)
− 1− ) dy E[Z]
0 L2
Z R
ADi (R, Di − y) N To find an upper bound for E[m] we use the left inequality
= R− 1− ) dy, (5) in (6) and the fact that the progress at each hop is at most R,
0 L2
m m−1
where the first term equals R since ADi (R, Di − R) = 0 as
X X
Zi ≤ Zi + R < D.
there is no intersection between two circles with centers at ni , i=1 i=1
 4

taking expectation we have

Xm 
E Zi < D.
i=1

using (7), we get:

D
E[m] < (10)
E[Z]
Fig. 2. The length of line ab¯ can range between Yi , the progress made at
Now, utilizing (10) and (9) the expected number of hops the ith hop, and the transmission range R.
E[n] = E[m] + 1 can be bounded as follows
D−R D B. Analysis of The End-to-End Delay
+ 1 ≤ E[n] < + 1. (11)
E[Z] E[Z] In this section, we analyze the total distance a packet travels
As mentioned earlier the random variables Yi are not i.i.d. from source to destination using the distance greedy routing.
and we need to bound them using i.i.d. random variables to Considering the delay is proportional to the distance, this will
be able to use the result in (11). give us the end-to-end delay metric.
For this purpose, let us define the progress at each hop as a Consider the forwarding scenario at the ith hop, depicted in
function of the remaining distance to the destination, as Y (D). Fig. 2, where node b has been chosen by the intermediate node
It is worth noting that the Y (D) is a non-decreasing function a and we make a progress of Yi towards the destination. We
of the remaining distance D. More precisely, for D ≥ D0 we want to find the distance traveled at the ith hop (i.e. the length
¯ in Fig. 2), which we denote it by Wi . Given the
of the line ab
have P [Y (D) > y] ≥ P [Y (D0 ) > y]. Intuitively, if there is
more distance to the destination it is more likely that we make progress Yi , we know that node b should be on the arc cc0 .
more progress than the case of having less distance to the Considering the fact that nodes are uniformly distributed, node
destination. This can be shown formally using (3), as follows b can be anywhere on the arc cc0 with equal probability. Also,
the distance of any node on the arc cc0 from the transmitting
node a ranges between Yi and R. Therefore, we can roughly
 AD (R, D − y) N estimate that the distance from node a to node b is uniformly
P [Y (D) > y] = 1 − 1 − ≥ (12)
L2 distributed between [Yi , R]. That is Wi ∼ U [Yi , R] for the ith
 AD0 (R, D0 − y) N

1− 1− = P [Y (D0 ) > y], hop.
L2
Now, we can find the expected end-to-end distance traveled
where the inequality follows from the fact that if D ≥ D0 we by a packet as
have AD (R, D − y) ≥ AD0 (R, D0 − y). X n  X n n
Now, observe that at any of the m hops, the distance between
X Yi + R 1
E Wi = E[Wi ] = = (D + E[n]R),
the intermediate node and the destination is between R and D. i=1 i=1 i=1
2 2
Using (12), we get (15)
where n is the number of hops whose expectation is charac-
P [Y (R) > y] ≤ P [Yi > y] ≤ P [Y (D) > y]. (13) terized in (14). Also, we have used the fact that the progresses
for i = 1, · · · , m at each hop sum up to the distance between source and
Pn
destination, i=1 Yi = D.
Thus, in (11) we can replace Zi with i.i.d. random variables
Y (D) to find a lower bound for the number of hops. Similarly,
C. Analysis of Network Density for Successful Delivery
we can use i.i.d. random variables Y (R) to find an upper
bound. In the greedy routing algorithm, we assume that there is al-
ways a neighbor to forward the packet towards the destination
D−R D
(1{D≥R} ) · + 1 ≤ E[n] < + 1, (14) and thereby we ignore the possibility of the packet reaching
E[Y (D)] E[Y (R)]
an isolated node (more precisely, a node without any progress-
where E[Y (D)] can be computed using (5) and 1{D≥R} is an making neighbors). In fact, if the network is dense enough or
indicator to account for the case where source and destination if the transmission range of nodes are large enough, such a
are immediate neighbors. situation can be avoided.
 5

In this section, we find the probability of successful deliv-

ery1 . First, we find the probability of a node being isolated
which equals the probability of having no node in its progress
area. To simplify the analysis, we estimate the progress area
of a node by the half of its transmission region which faces
the destination. Then, we can write the probability of a node’s
isolation, Piso , as
 πR2 N −1
Piso ≈ 1 − 22 (16)
L
we can now solve (16) for R and find the minimum transmis-
sion range such that the probability of node isolation is less
than  r q Fig. 3. Total travel distance from source to destination per packet vs the
2 1 number of nodes in the network (N ): simulation results are compared against
R> L2 + L2 ( N −1 ) (17) analytical lower and upper bounds.
π
The probability of success equals the probability of no node
along the path being isolated which is (1 − Piso )n , where n
denotes the number of hopes that is bounded in expectation
by (14)
!n
 πR2 N −1
Psuccess ≈ 1 − 1 − (18)
2L2

V. S IMULATION R ESULTS
To test simulations results and prove the efficacy of the
analysis work, random networks are generated using uniform
distributions for the initial real positions in a L × L grid. We Fig. 4. Probability of delivery success versus the number of nodes:
use the mobility model explained in Section II to generate comparison between the simulation results and analytical lower and upper
motion trajectories for N nodes. We use the actual positions bounds.

for all nodes when quantifying the performance metric, but on 15, where the bounds on the expected number of hops, E[n]
use the predicted positions when finding the optimal path. is obtained from (14). We note that the lower bound is tighter,
The predicted locations are the actual locations mixed with which is due to the tightness of the lower bound in (14). The
2
Normally distributed prediction noise of variance σN = 10. fluctuation in the results is due to the average distance between
We use dynamic contact graph by making connections the source and destination (D), which is a probabilistic value.
between nodes with pairwise distances below R. The rest of Another important performance indicator of the proposed
simulation parameters include number of nodes: N = 10, the algorithm is the probability of success, which means possibility
grid size: L = 10 km, communication range: R = 5 km, of progress at all intermediate nodes (having at least one node
average node velocities: v̄ = 50m/sec, and average waiting in the current node’s progress area), as characterized in (18)
time: w̄ = 20, unless specified otherwise. Also the transition based on the average number of hops per packet in (14). Fig 4
probability between the circular and linear motions is 20%. suggests that as we increase the number of nodes, the network
Finally, we note that for all figures, we take average over 100 density increases and therefore the probability of getting stuck
runs of the algorithm with different initializations. in an intermediate node with empty progress area diminishes.
We first, verify the accuracy of the derived upper and lower Similar to Fig. 3, the obtained lower bound is more accurate.
bounds for two important performance metrics, namely the Now, we compare the performance of the proposed greedy
end-to-end delay per packet and the probability of success. algorithm with the conventional Dijkstra’s shortest path al-
In Fig.3, we present the expected end-to-end distance per gorithm in Fig. 5. We also evaluate the proposed algorithm
packet vs N . The upper and lower bounds are presented based with and without including predictive location information
1 It
is worth noting that by successful delivery we mean the packet travels under different average node velocities. The results show that
from source to destination without facing an isolated node. the probability of delivery success for the greedy method is
 6

VI. C ONCLUSION

In this paper, we studied the routing problem in dynamic

UAV networks. Prior approaches fail to perform well in such
dynamic environments due to the requirement of maintaining
information about the network topology or using frequent
route-establishment phases. We studied an agile distance-
greedy routing algorithm which is low-complexity and the
intermediate nodes take decisions solely based on the predicted
locations of their neighbors. This algorithm is fully distributed
and incorporates predicted locations into the algorithm, hence
outperforms the centralized shortest algorithm. We character-
Fig. 5. Probability of delivery success for predictive greedy algorithm, static
ized the number of hops, the probability of success delivery
greedy algorithm and conventional Dijkstra’s algorithm.
and the expected distance a packet travels based on system
parameters. We plan to characterize the impact of the predic-
tion uncertainty on the optimality of the selected path as an
extension of this work.

R EFERENCES

[1] “Gartner says almost 3 million personal and commercial drones

will be shipped in 2017,” Feb 2017. [Online]. Available: http:
//[Link]/newsroom/id/3602317
[2] A. Razi, F. Afghah, and J. Chakareski, “Optimal measurement policy
for predicting uav network topology,” in 51th Asilomar Conference on
Signals, Systems and Computers (Asilomar’17), November 2017.

Fig. 6. End-to-end power comparison between the conventional Dijkstra’s [3] S. Mousavi, F. Afghah, J. Ashdown, and K. Truck, “Leader-follower
algorithm and the predictive greedy algorithm. based coalition formation in large-scale uav networks, a quantum
evolutionary approach,” in IEEE INFOCOM, Workshop on Wireless
Sensor, Robot, and UAV Networks, April 2018.
[4] F. Afghah, M. Zaeri-Amirani, A. Razi, J. Chakareski, and E. Bentley,
higher than that of the conventional shortest path algorithm “A Coalition Formation Approach to Coordinated Task Allocation in
consistently for all average node velocities. Also, the predictive Heterogeneous UAV Networks,” ArXiv e-prints, Nov. 2017.
greedy method outperforms the static greedy algorithm, which [5] A. Rovira-Sugranes and A. Razi, “Predictive routing for dynamic uav
shows including predictive location information decreases the networks,” in 2017 IEEE International Conference on Wireless for
probability of selecting nodes with empty progress area, as Space and Extreme Environments (WiSEE), Oct 2017, pp. 43–47.

was expected. Finally, when the network is more dynamic, [6] A. Razi, C. Wang, F. Almaraghi, Q. Huang, Y. Zhang, H. Lu,
and A. Rovira-Sugranes, “Predictive routing for wireless networks:
more nodes are subject to getting out of the communication
Robotics-based test and evaluation platform,” in 2018 IEEE 8th Annual
ranges of their neighbors, and hence the probability of success Computing and Communication Workshop and Conference (CCWC),
declines. Jan 2018, pp. 993–999.
[7] F. Afghah, A. Razi, and A. Abedi, “Stochastic game theoretical model
Lastly, in Fig. 6 we present the average power consumption for packet forwarding in relay networks,” Springer Telecommunication
per packet to complete the path for the proposed predictive Systems journal, Special Issue on Mobile Computing and Networking
greedy algorithm and the standard Dijkstra’s algorithm without Technologies, vol. 54, no. 2, pp. 1877–1893, 2013.
including predictive information in order to show the practical [8] J. d. Zhang, Y. j. Feng, F. f. Shi, G. Wang, B. Ma, R. s. Li, and X. y. Jia,
utility of the proposed method. Since the power consumption “Vehicle routing in urban areas based on the oil consumption weight
-dijkstra algorithm,” IET Intelligent Transport Systems, vol. 10, no. 7,
is proportional to the sum of all link distances squared,
pp. 495–502, 2016.
considering predictive information provides a significant gain
[9] W. Farooq, M. A. Khan, and S. Rehman, “Amvr: A multicast routing
for our suboptimal algorithm. This gain is higher for more protocol for autonomous military vehicles communication in vanet,” in
dynamic networks, since the inclusion of predictive locations 2017 14th International Bhurban Conference on Applied Sciences and
is more beneficial for higher average node velocities. Technology (IBCAST), Jan 2017, pp. 699–706.
 7

[10] H. C. Premkumar, V. R. Budyal, and M. S. Kakkasageri, “Cognitive

agent based stable routing protocol for vehicle-to-vehicle communica-
tion,” in 2016 IEEE Annual India Conference (INDICON), Dec 2016,
pp. 1–5.
[11] S. P. Choi and D.-Y. Yeung, “Predictive q-routing: A memory-based
reinforcement learning approach to adaptive traffic control,” in Advances
in Neural Information Processing Systems, 1996, pp. 945–951.
[12] G. Gankhuyag, A. P. Shrestha, and S. J. Yoo, “Robust and reliable
predictive routing strategy for flying ad-hoc networks,” IEEE Access,
vol. 5, pp. 643–654, 2017.
[13] C. Perkins and E. Royer, “Ad-hoc on-demand distance vector routing,”
in Mobile Computing Systems and Applications, 1999. Proceedings.
WMCSA’99. Second IEEE Workshop on. IEEE, 1999, pp. 90–100.
[14] S. Rosati, K. Kruelecki, G. Heitz, D. Floreano, and B. Rimoldi,
“Dynamic routing for flying ad hoc networks,” IEEE Transactions on
Vehicular Technology, vol. 65, no. 3, pp. 1690–1700, March 2016.
[15] E. W. Dijkstra, “A note on two problems in connexion with graphs,”
Numer. Math., vol. 1, no. 1, pp. 269–271, Dec. 1959.
[16] O. Bouachir, A. Abrassart, F. Garcia, and N. Larrieu, “A mobility
model for uav ad hoc network,” in 2014 International Conference on
Unmanned Aircraft Systems (ICUAS), May 2014, pp. 383–388.
[17] G. Aoude, J. Joseph, N. Roy, and J. How, “Mobile agent trajectory
prediction using bayesian nonparametric reachability trees,” Proc. of
AIAA Infotech@ Aerospace, pp. 1587–1593, 2011.
[18] G. D. Goez, R. A. V. Velez, and J. S. B. Valencia, “Uav route planning
optimization using pso implemented on microcontrollers,” IEEE Latin
America Transactions, vol. 14, no. 4, pp. 1705–1710, April 2016.
[19] A. Wald, “Sequential tests of statistical hypotheses,” Ann. Math. Statist.,
vol. 16, no. 2, pp. 117–186, 06 1945.