Mobility Prediction in Mobile Ad Hoc Networks Using Deep Learning Approach

Abstract
Mobile ad-hoc network (MANET) becomes the Random Waypoint Mobility (RWM) model was
common self-configuring wireless networks with used to create location pattern. From collected
comprising heterogeneous devices without datasets, we used 75% of the data for training the
depending on device directions. In mobile ad-hoc proposed system, and the remaining part for testing
networks where users are potentially highly mobile, purpose. The simulation result shows that 86.23%
knowledge of future location and movement can be can successfully estimate the predictions.
of great value to routing protocols. When the node Therefore, the proposed system can successfully
moves from one place to another place, the data predict the Mobile user’s mobility network
obtained by mobile is changed due to protocol topology through RSS values in dynamic
changes because future distance between environments.
neighboring nodes is changes. Therefore, to
overcome this problem and balancing dataflow Keywords: Back propagation through time
through protocol changes, mobility prediction (BPTT), deep recurrent neural network (DRNN),
application is very important. Therefore, mobile mobility prediction, mobile ad-hoc networks
user’s mobility prediction is very important to (MANETs), random waypoint mobility
know the future location of user nodes network
topology. It also uses to allow for estimating the 1. Introduction
stability of paths in mobile wireless ad hoc Mmobile networks can be categorized mobile IP:
networks. Moreover, the proposed system helps for networks with infrastructures that is networks with
location prediction to allocate the next suitable base stations, gateway and routing support andad
access point before the user nodes’ leaves its hoc networks: networks without any a pre-
current one, in order to reduce the interruption time established of infrastructure [1]. Ad hoc networks
in communication between user nodes through can be static or mobile. In static ad hoc networks,
complex environments. In this paper, we used a the topology of the mobile node may not change
deep learning recurrent neural network (DRNN) for once it has become portion of the network.
predict the future topology.
Nowadays, a Mobile ad hoc network (MANET)
To evaluate the proposed system, we collect RSS becomes popular because of self-configuring
2
values from 1000X1000 m from a conference wireless networks of mobile nodes; May
room. From the specified room we seat four access compriseheterogeneous devises (mobile phones,
points (Aps) in specific topology and stored in the PDAs, laptops, etc.) connected by a wireless link to
server for training and testing from a conference form an arbitrary topology without a fixed
room. The prediction is done based on the node infrastructure and central administration in a
movement history that is previous user movement dynamic environment.Every mobile node(MN) has
patterns based on collected data obtained. Each a maximum transmission power which determines
mobile node is assumed to know its current the maximum transmission range of the MN to
mobility information such as position, speed and communicate with other MNs. Due to the
movement direction angle. In the proposed system,
availability of small and inexpensive wireless mobility prediction in different approaches briefly
communicating devices, mobile ad-hoc networks explained. In the 3rd section we discussed the newly
play a very important role in different applications proposed scheme for mobility prediction.
in a day to day activities such as battlefield Performance of the proposed mobility prediction is
communication, disaster relief applications, sensor discussed in section 4. The final section shows the
networks and emergency medical situations [10]. conclusion of this work.
2. Related Works
Every mobile node in MANET can move
Mobility prediction is a method to deal with the
autonomously in any direction, therefore the
problems emerging from the nodes’ mobility by
communication link between other devices can
predicting the future network topology changes and
break due to protocol changes and/or the previous
estimating trajectory of the future position of the
established topology changes. Two mobile wireless
MNs in a dynamic environment to protect link
nodes out of the communication range can use
failure due to mobility [15]. Due to the importance
other MNs within their communication range to
of mobility prediction in ad hoc networks, a lot of
reply packets by discovering and maintaining
scholars have been studied in cellular networks and
routes to other MNs.To establish more robust and
wireless ad hoc networks [1] - [4], [7] - [15]. In
reliable communications, it is important to predict
[12] the Authors proposed the prediction model to
the next suitable access point of the MNs before the
predict the topology changes in order to provide the
user nodes’ leaves its current one, to reduce the
seamless connection service for MANETs.
interruption time in communication between user
Actually, GPS information is used to estimate the
nodes through a dynamic environments [2] [14].
expiration time of the link between two adjacent

MANETs are typically applied in emergency mobile nodes. Based on this prediction, routes are

operations and military background, where future reconstructed before the links expire. The paper

node actions cannot be based on a record of showed the route delivery ratio maintained above

previous movements because of the dissimilar 90% for all mobility speeds.

requirements of each circumstance [16]. Mobility

Methods are proposed in [3] an Autoregressive
prediction scheme for MANETs are executed on
Hello protocol (ARH) for neighborhood discovery
the mobile nodes, they must be more light weight
in mobile ad hoc networks. Each node predicts its
than the methods for fixed wireless system,
own position by an ever-update autoregressive
typically perform on the base station and in
based mobility model. Each mobile node updates
MANET no need of any infrastructure. By taking
its location by transmit a “hello” message. [7]
these significance we are motivate to contribute
Investigated to predict the future movement of
towards the achievement of mobility prediction in
MNs position in ad hoc networks using
MANET. In this study, we proposed to build a
Feedforward neural network based method.
Deep Recurrent Neural Network (DRNN)
Random Waypoint Mobility (RWM) model was
prototype for mobility prediction in mobile ad hoc
used to create location patterns and backpropagaion
networks.
algorithm (BPA) for training the network to

The remainder of this paper is organized as minimize the mean square error by adjusting the

follows: In the 2nd section the previous work in weights of each node. The prototype evaluated by
presenting the new input patterns to this network Positioning System (GPS) receiver. In this study,
and works well for prediction the next position of we develop a technique for mobility prediction in
MNs in mobile ad hoc network. MANET. This technique is not dependent of the
mobility model used by the MNs. We assume that
Model for MANET using Recurrent Neural
each MN in the ad hoc network is aware of its
Network (RNN) and Extended Kalman Filter
location. The node will be able to learn its location
(EKF) was proposed in [2] to predict the behaviour
using a Global Positioning System (GPS) receiver,
of mobile ad hoc network. The Kalman Filtering
so it can periodically record its geographical
addresses the estimation of a state vector in a linear
location.
model of a dynamic system. To apply EKF for
estimating optimal weights of RNN, weights of
3. Proposed Model
network were interpreted as transition matrix for
Our proposed model is used to predict the future
the system state. For simulation Reference Point
location of the mobile node in a dynamic
Group Mobility Model (RPGM) used as a mobility
environment. To build the location prediction
model and the model trained with RNN using
network model, four static Aps and 4 MNs moving
Levenberg-Marquardt and the performance result
in a grid of 1000x1000 m2 were used to collect the
of the model was satisfactory.
data from real environment as shown in figure 1.

In [1] the Authors proposed to estimate the The data were collected from conference room

availability of communication link of the mobility located at first floor of complex building in Addis

of the MNs using neural network based method. In Ababa science and technology university (ASTU).

this paper a recurrent neural network has been used To create location patterns we used RWM model

for long time series prediction. Backpropagation since it is the most popular mobility models used to

through time (BPTT) algorithm was used for evaluate the mobility prediction in MANET and

training the network to minimize the error function. well-suited in our case. During data collection, our

Random Waypoint Mobility (RWP) model were selected mobility model works as follow to collect

used to examine the efficiency of the prediction RSS values from various locations.

prototype. • A MN stays in one location for a certain

period of time (also known as pause time).
Until now, there are many methods proposed for • A node chooses a random destination
mobility prediction in ad hoc mobile networks as anywhere in the specified area
we described on the above, the researchers’ were • The node moves towards that destination
done mobility prediction by using different with a velocity chosen randomly from [0,
mobility models to collect the previous MNs maxspeed]
movement patterns(current geographic location, • After reaching the destination, the node stops
speed and moving direction of a node ) using for a duration defined by the “pause time”
different simulation software’s. To the best of our (Pause) parameter.
knowledge, there is no work done on mobility • This procedure is repeated until we get
prediction in mobile ad hoc network using DRNN enough dataset for better prediction
by capturingprevious and current movement prototype.
patterns of real mobile nodes via Global
All of the training and testing datasets was obtained Radio transmission range two-way radio
range
from the real mobile nodes.We collect a total of
1000 patterns. Among these patterns 700 patterns
5. Conclusion
are used for training network and the rest are used
Regarding mobile ad-hoc network lots of work
for testing the performance of predicting model.
done like security and routing for better
The number of hidden nodes was selected by trial efficiency, mobility is main concern because of
and error. the nature of devices moves of nodes is
random so that communication between nodes
going to be easy and faster in our future work
we apply RWM model.

References
Figure 1: Area of data collection: conference room.
[1]. A. Ali, L. A. Latiff, N. Fisal, “GPS-Free Indoor
In our work we have used deep recurrent neural
Geometry-Based Location Tracking In Mobile Ad
network. We used Matlab to implement our
Hoc Network”, RF and Microwave Conference
proposed method and learning automation. BPTT is
(IEEE Cat. No.04EX924), pp. 251-255, 2004.
the training algorithm used to update weights in
[2] P. Sharma, S. Kohli, and A. K. Sinha, “Model
recurrent neural networks.
for MANET Using Recurrent Neural Network
and Extended Kalman Filter”, International
4. Results and Discussion Journal of Applied Engineering Research.
To implement the proposed system that described 11(1) p. 8-10, 2016
in section III, we used MATLAB 2017b tool. The [3] X. Li, N. Mitton, and D. Simplot-ryl,
entire experimentation was done on a laptop “Mobility Prediction Based Neighborhood
computer with processor of Intel(R) Core(TM) i5- Discovery in Mobile Ad Hoc Networks”, vol.
4210M CPU @2.60GHz, 4 GB RAM, Microsoft 2011, pp. 147–159, 2011.
Windows 10, 64- operating system and 500GB [4] L. Ghouti., “Mobility prediction using fully
hard disk. complex extreme learning machines”,

Table 1: Data collection parameters European Symposium on Artificial Neural

Networks, Computational Intelligence and
Parameters
Mobility model RWM model machine learning , p. 607-612, 2014.
Maximum speed(m/s) 0-1m/s
[5]
Pause time(s) 0,5,10,15,20,25,30,40
sec [6] Y. Zhang, J.Hu, J. Dong, Y. Yuan, J. Zhou, J.
Simulation area(m) 1000mX1000m Shi “Location Prediction Model Based on
Number of MNs 100,200,300,350
 Bayesian Network Theory,” IEEE [16] Garud V. S, More S.N. and Panpatil A.Y.,
GLOBECOM (Global Telecommunications “Mobility Location Prediction: A review of
Conference), 2009, pp.1-6. Techniques used in Mobile Ad-Hoc Network”,
[7] N. Makhlouf, “Exploiting Neural Networks Journal of Electronics and Communication
for Mobility Prediction in Mobile Ad Hoc Engineering (IOSR-JECE) e-ISSN: 2278-
Networks”, journal of electrorevue, vol.18, 2834, p-ISSN: 2278-8735 PP 59-68, 2015.
no.3, pp. 66-67, 2016 [17] H. Kaaniche and F. Kamoun, “Mobility
[8] Prediction in Wireless Ad hoc Networks using
[9] F. Bai and A. Helmy, “A Survey of Mobility Neural Networks,” Journal of
Modelsin Wireless Adhoc Networks,” Wirel. Telecommunications, vol. 2, no. 1, pp. 95 -
Ad Hoc Sens.Networks, pp. 1–28, 2001. 101, 2010.
[10] L. Ghouti, T. R. Sheltami, and K. S.
Alutaibi,“Mobility Prediction in Mobile Ad
Hoc Networks Using Extreme Learning
Machines,” ProcediaComput. Sci., vol. 19, pp.
305–312, 2013.
[11]
[12] W. Su, S. J. Lee, and M. Gerla, “Mobility
Prediction and Routing in Ad Hoc Wireless
Networks,” International Journal of Network
management, vol. 11, no. 1, pp. 3 - 30, 2001.
[13] Cadger F, Curran K, Santos J, and Moffett
S., “MANET location prediction using
machine learning algorithms,” Proceedings of
the International Conference on
Wired/Wireless Internet Communication,
Springer-Verlag Berlin, Heidelberg, pp. 174-
185, 2012.
[14] S. M. Mousavi, H. R. Rabiee, M. Moshref,
A. Dabirmoghaddam, “Model Based Adaptive
Mobility Prediction in Mobile Ad-Hoc
Networks”, International Conference on
Wireless Communications, Networking and
Mobile Computing, pp. 1713-1716, 2007
[15] H.Si, Y.Wang, J. Yuan and X. Shan,
“Mobility Prediction in Cellular Network
Using Hidden Markov Model”, 7th IEEE
Consumer Communications and Networking
Conference, 2010.