You are on page 1of 3

Prediction of Information Propagation in a Drone

Network by using Machine Learning


Jinsoo Park, Yoojoong Kim, and Junhee Seok*
The School of Electrical Engineering
Korea University
Seoul, Korea
{waterlemona, sunbisunbi, jseok14}@korea.ac.kr

Abstract Drones cooperate with each other by transmitting


and receiving packets. Therefore, it is important to conjecture II. PROPOSED METHOD
the packet transmission rates within the network. However, the Communications between UAVs are carried out in a three-
conventional methods are not suitable to describe the dimensional environment, and each UAV has its own spatial
transmission patterns with satisfactory computing speed and movement. The probability of successful packet transmission
accuracy. In this paper, we demonstrated that machine learning
decreases as the distance between UAVs increases. Therefore,
can successfully predict the transmission patterns in drone
network. The packet transmission rates of a communication
relayed packet transmission via spatially-neighboring UAVs
network with twenty drones were simulated, of which results would be necessary in order to deliver the packet to the specific
were used to train the linear regression and Support Vector UAV. As a result, the communication network between UAVs
Machine with Quadratic Kernel (SVM-QK). We found out SVM- become more complex and time-variant. As shown in Fig. 1,
QK can precisely predict the communication between drones. two assumptions were introduced in order to construct a model
for complex communication network structure.
Keywords Drone; Communication; Network; Supported
1. Transmission channel between UAVs exists only when the
Vector Machine; linear regression; Monte-Carlo method
successful packet transmission rate is greater than .
I. INTRODUCTION 2. If the channels are connected, every channel has the same
packet transmission rate ().
With the increasing number of Unmanned Aerial Vehicles
(UAV), packet communications between them have become
more complex and difficult [1]. Moreover, various industries
are trying to deploy drones for commercial purposes. Given
this trend, the number of UAVs will naturally increase even
further. This implies that packet transmissions between UAVs
will become more complex. And the chance of transmission
failures will be higher concurrently. Transmission failure can
hinder the effective movement between UAVs [2]. Thus, it is
critically important to determine the possibility of transmission
failure in packet transmission networks between UAVs.
There are several ways to model and predict a packet
transmission network [3-4]. This paper focuses on successful
communication spread between UAVs. The purpose of this
research is to determine whether the success or failure of Fig. 1. Two assumptions in order to construct a model of the drone
packet transmissions between UAVs can be predicted by the communication network.
machine learning methods. The success or failure probabilities
of packet transmission is computed using Monte-Carlo As shown in Fig. 1, the communication network between
Simulation (MCS) [4] after simple modeling for channel design UAVs is composed of both UAVs and transmission channels
of the transmission during communication between UAVs. To between them. According to above mentioned two assumptions,
be specific, the packet transmission process in the UAV each UAV can concurrently transmit or receive the packets
network is simulated using Susceptible-Infected-Recovery to/from its neighbors, respectively. The transmission status of
(SIR) model [5] mainly used in the disease spread model [6]. previous network affects the packet propagation of current
Thereafter, the simulation results of the process would be network. Probability that one UAV receives the packet at a
trained using Linear Regression using nth order (LR(nth)) and specific time can be mathematically represented as follows;
Support Vector Machine using Quadratic Kernel (SVM-QK) (1)
methods in order to assess the prediction speed and accuracy
resulted by two methods. (2)

This work was supported by grants from the National Research


Foundation of Korea (NRF-2014R1A1A2A16050527, NRF-
2015R1A2A1A16074932). Correspondence should be addressed to
jseok14@korea.ac.kr

978-1-5090-1325-8/16/$31.00 2016 IEEE 147 ICTC 2016


pi (t) is the probability that UAV node i successfully
receives packets at/before the time t. qi (t) is the probability that
UAV node i fails to receive the packets from its neighboring
nodes. Thus, eq. (1) represents the probability that specific
UAV has received the packet at t+1. This value can be
represented by the sum of both probability that it already
received the packet and probability that it receives the packet
from other UAVs at time t. Above mentioned two equations
have been commonly used in many propagation models (e.g.
SIR model). Therefore, these equations can also be applied to
the drone communication network following previously-
mentioned two assumptions. Applying Eq. (1) and (2) will
make it possible to simulate the packet transmissions in the
communication network by using MCS. In this paper,
simulation data derived from MCS were used as the training
data for machine learning methods such as LR and SVM-QK
to predict the transmission probability.
In the training of LR, various equations that possess up to
7th order were used to represent the complex data as precise as
possible. In the training of SVM-QK, the probability data
(floating point data) obtained from MCS were quantized into Fig. 2. UAV packet transmission network used in the simulation; Each circle
the integers by multiplying specific constants such as 1000, represents the UAV nodes. Red bar represents the available transmission
channel between two UAVs. Blue node is the UAV used in the simulation in
10000, and 100000 (Quantization Factor: QF). This order to show the prediction accuracies and times by two methods.
quantization process changes the numerical numbers into labels.
The training of SVM-QK is carried out using these quantized
data as the class labels. QF can affect the accuracy of trained
results from SVM-QK. However, SVM-QK can predict the
probability data with tremendous accuracy regardless of QF
factor. Root mean square error (RMS) is used to measure the
accuracy of results estimated from both LR and SVM-QK.

III. SIMULATION RESULT


As shown in Fig. 2, a scenario, in which twenty UAVs
work as the routers to propagate the packets transmitted from
the UAV node 1 in the drone communication network, were
used in the simulation. Each UAV obeys the assumptions
introduced in Fig. 1, and transmits the packets to the
neighboring UAVs. As a result, the packets can be spread out (a) Comparisons of the prediction accuracies resulted by both LR(4th and 7th)
according to various routes. Therefore, the conventional and SVM-QK against MCS data at UAV node 9
methods may not be possible to predict the packet transmission
probability of the scenario precisely. In the scenario, it is
assumed that each UAV has a communication channel with
0.05 successful packet transmission probability. This value is
determined to collect the sufficient amount of data in the small
network. If the transmission probability is determined less than
0.05, it would be unrealistic in actual communication, and if
the probability is decided greater than 0.05, collection of
sufficient data to carry out machine learning would be difficult
in the suggested network, since the packets propagate too fast.
In the communication network used in the scenario, the
probability data of individual UAV have calculated by using
discrete-time MCS. While processing discrete-time MCS, one
million iterations were performed to calculate the probability of (b) Comparisons of the prediction accuracies resulted by both LR(4th and 7th)
packet transmission to get accurate data as much as possible. and SVM-QK against MCS data at UAV node 20
Also all simulations were carried out until 150 packet
transmissions were finished. The simulation data calculated by Fig. 3. Comparison results of the prediction accuracies between LR(4th or 7th)
MCS had been used to train LR(nth; n=1-7) and SVM-QK and SVM-QK using 1,000 as a QF for the MCS data.
using various QFs.

148
As an example of simulation results, the packet
transmission probabilities for 150 packets were calculated
when UAV nodes 9 and 20 received the packets sent from
UAV node 1. As shown in Fig. 2, two UAV nodes (9 and 20)
can be representative for all remainder nodes because they
possess the complicated data characteristics fairly enough. Fig.
3 shows the comparison results of prediction accuracies
obtained by both LR and SVM-QK against the results of MCS,
at UAV nodes 9 and 20. The transmission environments for the
other UAV nodes can be trained and predicted by using
identical method.

TABLE I. PREDICTION ERROR OF EACH MACHINE LEARNING


METHOD
Fig. 4. Representation of computational time required in the simulation of
Linear Regression SVM-QK packet transmission probability: computation time of MCS, LR(nth; n = 1-7),
Error QF and SVM-QK in log-scale showed an learning time and two prediction times,
LR(4th) LR(7th) QF 1000 QF 10000 respectively.
100000
RMS 1.71E-02 2.28E-03 5.97E-04 5.84E-05 5.31E-06
On the contrary, machine learning methods such as LR(nth)
and SVM-QK require reasonable time. However, it can be
easily found that the results of SVM-QK are superior because it
In the training LR method, the features such as transmission requires very short time to process in comparison with that LR
probability data, node locations in the network, transmission only can obtain negligible error while applying far more than
probability within channel and times, were used. All of such 4th order equations.
features were obtained through MCS. Various orders of
equation from 1st to nth orders can be applied in the LR
IV. CONCLUSION
training process (LR(nth; n = 1-N)). Therefore, a LR system to
predict transmission probability can be constructed with any This paper shows that the nonlinear model of the packet
nth order term. In our simulation, n is used from 1 to 7. It is transmissions network between twenty UAVs can be precisely
found that the prediction accuracies are tremendously increased trained by the machine learning methods such as LR and SVM-
according to the order increased while measuring by RMS, QK. Simulation results show that the packet transmission
after training LR(1th-7th). However, as shown in Fig. 3, probability can be exactly predicted by 0.0000531% in RMS
inadequate prediction can be possible in the borders of error measurement, especially when the SVM-QK is used.
transmissions (starting and terminating parts) due to non-linear Even the prediction speed and accuracy by SVM-QK are far
characteristics of the packet transmission. As shown in TABLE faster and more precise than those of LR(higher-order nth). If
I, simulation results show that the average prediction errors of several drone networks could be sufficiently trained in advance,
1.71% and 0.228% were measured in LR(4th) and LR(7th), another complicated network can be quickly predicted while
respectively. maintaining high accuracy. This approach is helpful in
analyzing a time-varying UAV transmission network.
In the training SVM-QK, the features such as the quantized
values by QFs on the results of MCS, transmission
probabilities and time, and locations of nodes in the network, REFERENCES
were used. As the same as LR, SVM-QK also trained the [1] P. R. Chandler, M. Pacher, D. Swaroop, J. M. Fowler, J. K. Howlett, S.
features resulted from the individual UAV nodes. Thereafter, Rasmussen, C.Schumacher, and K.Nygard, Complexity in UAV
coorperative control, IEEE, vol.3, pp 1831-1836, May 2002.
the trained SVM-QK are combined to predict the probabilities
[2] Burdoin, Robert B., Nicolaas J. Moolenijzer, and Fred M. Strohacker.
of packet transmissions of whole UAV network. As shown in "Airborne drone formation control system." U.S. Patent 5521817, May
TABLE I, simulation results show that the average prediction 1996.
errors of 5.97E-04 (0.00597%), 5.84E-05 (0.000584%), 5.31E- [3] Chang, Fi-John, and Yen-Chang Chen. "A counterpropagation fuzzy-
06 (0.0000531%) measured by RMS were obtained according neural network modeling approach to real time streamflow prediction."
to the QF variations, respectively. Consequently, it can be Journal of hydrology 245.1, pp. 153-164, May 2001.
regarded that the predicted values of SVM-QK are precisely [4] S. Gomez, A. Arenas, J. Borge-Holthoefer, S. Meloni, and Y.Moreno
matched to the outcome data of MCS, even if the packet Discrete-time Markov chain approach to contact-based disease
transmission characteristics are non-linear. spreading in complex networks, EPL, vol. 89, pp. 38009, Jan 2010.
[5] N. Madar, T. Kalisky, R. Cohen, D. ben-Avraham, and S. Havlin.
Fig. 4 shows the computation times required in the "Immunization and epidemic dynamics in complex networks." The
simulations of three methods, such as MCS, LRs, and SVM- European Physical Journal B-Condensed Matter and Complex Systems,
vol. 38, pp. 269-276, March 2004.
QK. Log-scale is used to represent the time required because
the variations are too wide. As shown in Fig. 4, it can be easily [6] Grassly, Nicholas C., and Christophe Fraser. "Mathematical models of
infectious disease transmission." Nature Reviews Microbiology, vol. 6,
found that the MCS cannot be used in the prediction of real- pp. 477-487, Jun 2008.
world network characteristics because it requires unreasonably
long time.

149

You might also like