Enhanced Routing Algorithm for Opportunistic Networking

On the Improvement of the Basic Opportunistic Networking Routing Algorithm

by the Application of Machine Learning

Ladislava Smítková Janků and Kateřina Hyniová

Institute of Department of Digital Design, Faculty of Information Technology, Czech Technical University in Prague,

Thákurova 9, 160 00 Praha 6, Czech Republic

Keywords: Opportunistic Networking, Information Retrieval, Probabilistic Models, HMM, Mobility Models,

Autonomous Mobile Robots, Machine Learning based Routing Method.

Abstract: The opportunistic communication networks are special communication networks where no assumption is

made on the existence of a complete path between two nodes wishing to communicate; the source and

destination nodes needn't be connected to the same network at the same time. This assumption makes the

routing in these networks extremely difficult. We proposed the novel opportunistic networking routing

algorithm, which improves the basic opportunistic networking routing algorithm by application of machine

learning. The HMM Autonomous Robot Mobility Models and Node Reachability Model are constructed

from the observed data and used in a proposed routing scheme in order to compute the combined

probabilities of message delivery to the destination node. In the proposed routing scheme, the messages are

coppied between two nodes only if the combined probability of the message delivery to the destination node

is higher than the preliminary defined limit value. The routing scheme was developed for the networks of

autonomous mobile robots. The improvement about 70% in a network load is reported.

1 INTRODUCTION

The opportunistic communication networks are

special communication networks where no

assumption is made on the existence of a complete

path between two nodes wishing to communicate

(Pelusi, 2006); the source and destination nodes

needn't be connected to the same network at the

same time. This assumption makes the routing in

these networks extremely difficult. In contrast to the

common ad-hoc networks, source and destination

nodes needn't be connected to the same network at

the same time, but they are allowed to exchange

messages between them due to techniques of the

opportunistic networking. These techniques allow

nodes to connect and disconnect anytime. In

opportunistic networking, no limitations are also set

on the nodes to keep their positions; the nodes can

move. This opportunity networking paradigm opens

a space for a number of novel application scenarios.

This paper describes the improvement of the

basic opportunistic network routing algorithm by

application of machine learning. Machine learning is

applied to construct robot mobility models and node

reachability model. These models are furthermore

used in routing to make decisions about messages.

1.1 Basic Opportunistic Networking

Routing Schema

The basic opportunistic networking schema is based

on a flooding algorithm extended by sending of

message delivery acknowledgement. There are

several variants of flooding algorithm. Usually each

node tries to forward every message to every one of

its neighbours except the source node. Every

message is delivered to all reachable parts of the

network. One of the most important disadvantages

of this routing schema is the fact, that messages can

become duplicated in the network. It increases the

network load and can cause the network

overloading.

1.2 Previous Work

There are two main categories of mobility models:

macroscopic models and microscopic models

(Helbing, 2001); (Qiang, 2009). The macroscopic

771

Smítková Jank˚u L. and Hyniová K..

Enhanced Routing Algorithm for Opportunistic Networking - On the Improvement of the Basic Opportunistic Networking Routing Algorithm by the

Application of Machine Learning.

DOI: 10.5220/0004923507710776

In Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), pages 771-776

ISBN: 978-989-758-018-5

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

models operate with the global parameters, e.g. an

average node velocity or a traffic density. The

microscopic models model each node of the network

particularly. There are many subcategories of these

models.

The Stochastic Models are based on the

assumption of random movements of the nodes

(Timo, 2007). Random Walk or Random Walk or

Random Direction models belong to this category of

mobility models. Traffic Stream Models explore the

mathematical and physical techniques developed in

hydrodynamics. In Car Following Models, the

behavior of each driver is modeled according to

vehicles.

Statistical models, also called Survey-based

Models (Harri, 2009), constitute a second important

group of mobility models. These models are based

on probability functions describing properties of the

node behavior.

Both the synthetic and statistical models

described above are the dynamic models; all of them

can generate mobility of patterns, but they can't

operate on pattern data only. A different approach

can be followed. Instead of developing complex

models and then calibrating them using mobility

traces or surveys, it is possible to extract generic

mobility patterns from movement traces. This

approach is called Trace-based Modeling (Harri,

2009).

2 METHOD DESCRIPTION

This section describes the whole process of the

creation of the new routing method.

2.1 Mobility Model

The mobility model is constructed for each of the

autonomous mobile robots from the finite set of its

positions in time. Mobility model has to cover both

the changes in a space and the changes in time.

Hidden Markov Model (HMM) is defined by the

set of the observation symbols, initial state

probablity vector and the state transition probability

matrix. Since the HMMs were successfully applied

in speech recognition, they become a standard tool

for recognition and modeling spatio-temporal data.

2.1.1 HMM Background

HMM is the spatio-temporal model defined as

(Rabiner, 1989):

Θ=

(

A,B,π

)

(1)

where A is a state transition probability matrix, B is

a observation probability matrix and π is an initial

state probability vector.





ij t j t i

a=Pq+1=s|q=s,1 i,j n

(2)

j=1

a=1



(3)









jt t t t j

v=Po=v|q=s

(4)

ijn,ikm







ij t

i=1



(5)





i1i

π =P q =s ,1 i n

(6)

i=1

π =1



(7)

2.1.2 The Set of Mobile Robot Positions

The set of mobile robot positions consists of triples

containing three elements: x-coordinate, y-

coordinate and time. The set of mobile robot

positions is finite.

In simulations, the mobile robot positions are

generated by the simulator. Otherwise, the data set

can be collected by any device, which makes

possible to use GPS for position localization.

The finite data set of mobile robot positions is

collected independently fo each mobile robot.

2.1.3 Observation Symbols in a Mobility

Model

The set of mobile robot positions is too large to be

used directly as the set of observation symbols in

HMM based robot mobility model. We need to find

a smaller set of symbols.

We decided to use cluster analysis algorithm,

which was was proposed in order to construct

Human Mobility Models (Hyunuk, 2012). The

clustering vonsists from the following operations:

cluster initialization

2.1.4 Autonomous Robot Mobility Model

Identification

The well-known algorithms for the estimation of the

Hidden Markov Model are EM algorithm and

Viterbi training and their derivations. We used EM

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

772

algorithm.

2.2 Network Scanning

The data on how the messages are boradcasted in the

network are collected during the scanning. Each

mobile robot generates a scanning message. The

mobile robot moves and copies the scanning

messages to the memory of each robot he meets

during the scanning. The time of message copying is

preserved as a part of the message in a form of a

time stamp.

When a mobile robot meets another mobile

robot, it copies its message memory to it. Message

memory contains the messages, which the mobile

robot received from the other robots and which are

not expired. Expired messages are removed from the

message memory. Network scanning algorithm is

described below.

There are two important parameters of network

scanning: maximum delivery time tmax and the

maximum number of message copying cmax. The

output of the network scanning is a log file, which

contains a) messages, which were delivered to the

destination node in time less or equal to tmax and

simultaneously were not copied more times than

cmax times b) messages, which were copied cmax

times and cannot be copied anymore.

2.2.1 Maximum Delivery Time

Maximum delivery time is defined as a maximum

time between the scanning message transmission by

the source mobile robot (source node) and its

reception by the destination mobile robot

(destination node).

During the scanning process, only the scanning

messages delivered to the destination before the

maximum delivery time is reached, are logged.

In our experiments, mobile robots are

synchronised using the same internal clock in the

simulator, e.g. we do not need solve the problems

which occur when each mobile robot has its own

clock.

2.2.2 Maximum Number of Message

Copying

The maximum number of message copying cmax

defines the depth of the network scanning. If the

number of copying of a scanning message is equal to

the cmax, the scanning message is removed from the

node memory and it is logged.

2.2.3 Scanning Message Structure

The scanning message has the following structure:

Number of the Source Node … unique number

of the source mobile robot (source node),

Number of the Destination Node … unique

number of the destination mobile robot (destination

node),

Message Id – a number of scanning message sent

by the the source mobile robot (source node),

Source Node Cluster Id … a unique number of

the mobility model cluster, in which the source

mobile robot starts its movement,

Copying Counter … it identifies how many times

the message was copyied,

Copy Info Section: the copy info section is

repated in the message cmax times to preserve

information of each copying of the message.

i-th Copy Info Node Number …. number of the

node the message is copyied into, i=1 to cmax.

i-th Time Stamp ... time, when the message copy

was done, i=1 to cmax

i-th “To” Node Cluster … unique number of the

mobile robot (node), which received a copy of a

message, i=1 to cmax.

2.2.4 Network Scanning Algorithm

This section describes the network scanning

algorithm. The Network Scanning Process has the

following phases:

1. Initialization. The message buffer of each mobile

robot is cleaned in this step. The mobile robots

are in their initial positions.

2. Scanning. The mobile robots move in some

meaningful ways, e.g. solve some tasks. Every

time the mobile robot meets another mobile

robot, the followng operations are performed:

1. Each mobile robot generates a scanning

message; the another mobile robot makes a

copy of this message with the time-stamp.

2. Each mobile robot makes a copy of the

message buffer of another mobile robot.

During the copying process, the mobile robot

reads each scanning message. If the mobile

robot is the destination node for some

message, it put the time stamp to this

message and wrote the message copy in the

log file; the message is cleaned. Otherweise,

it modified the messages by adding the

information on copying to each message.

This information contains time and mobile

robot ID. Also, the Copying Counter in each

message is incremented.

EnhancedRoutingAlgorithmforOpportunisticNetworking-OntheImprovementoftheBasicOpportunisticNetworking

RoutingAlgorithmbytheApplicationofMachineLearning

773

3. Each mobile robot scans its message buffer

for the “garbage” messages. Garbage

Messages are:

a) the messages, which are too old, e.g.

messages which are not valid, because they

exist longer than the Maximum Delivery

Time parameter dwfines,

b) the messages which cannot be copied

anymore; the message cannot be copied

anymore if the Copy Counter in the message

is equal to the Maximum Number of Message

Copying parameter.

All the garbage messages are moved to the log file

and removed from the mobile robot message buffer.

3. End of Logging. The network scanning is ended

by the human user on demand.

2.2.5 Node Reachability Model

Node Reachability Model is constructed from the

data collected during the network scanning. Node

Reachability Model is defined on the MxMxCl,

where M is the set of the autonomous network nodes

(autonomous robots) and Cl is the union of all the

observation symbols used in HMM Robot Mobility

Models. The elements of the Node Reachability

Model are values of the probability function. Each

value represents the probability than the message

sent from the i-th source node in k-th cluster reaches

the j-th destination node in time less than tmax and

the number of message copying will not be higher

than cmax.

2.3 Routing Algorithm Alpha 09

We proposed the novel opportunistic networking

routing algorithm called ALPHA09. ALPHA09

improves the basic opportunistic networking routing

algorithm by application of HMM Autonomous

Robot Mobility Models described in section 2.2. and

Node Reachability Model described in section 2.3.5.

The detailed description of the algorithm needs more

space than is available in this paper. The next

section presents the most important parts of the

ALPHA09.

2.3.1 How the ALPHA09 Works

Let A be the source node and Z be the destination

node and m be a message generated by the node A.

Let tmax be a maximum delivery time. Let cmax be

a maximal number of a message copying, e.g.

message m sent from the source node A can be most

highly cmax times copyied on its way through the

network until it reaches the destination node Z.

If the source node A meets an unknown node X,

the following steps are done:

1. Node A asks unknown node X for its Id Number;

1a. if IdX IS IdZ, the message m is copied

directly from the node A to the node Z. The

message m was delivered. The acceptation

message is copied from Z to A and m is

removed from the node A memory.

1b. if IdX IS NOT IdZ, then

2. Node A asks node X if it already carries the

message m.

2a if node X carries a message m, the Copy

Count of message m in node X is set to 1.

2b if node X does not carry a message, node A

asks node X for its HMM mobility model.

Node A computes a combined probability of

message m delivery on assumption the

message is copied to the node X. This

combined probability PAX is computed using

the Node Reachability Model and the HMM

Mobility Model of the node X as a sum of the

all probabilities over the set of output

symbols of the HMM Mobility Model of the

node X.

3. Node A compares PAX to Pe;

3a. if PAX >= Pe, the message is copied from

node A to node X.

3b. if PAX < Pe, the message is not copied from

node A to node X. The continue with the step

4. Node A stops the active communication.

Let M be a node, which carries a k-th copy of the

message m generated by the source node A.

If the node M meets an unknown node X, the

following steps are done:

1. Node M asks unknown node X for its Id

Number;

1a if IdX IS IdZ, the message m is copied

directly from the node M to the node Z. The

message m is delivered. The acceptation

message is copied from Z to A and m is

removed from the node A memory.

1b if IdX IS NOT IdZ,

2 Node M compares k and the cmax.

2a If k=cmax-1, the message is not copied and

M stops the active communication.

2b k<cmax-1; the node M asks node X for its

HMM mobility model. Node M computes a

combined probability of message m delivery

on assumption the message is to the node X.

This combined probability PMX is computed

using the Node Reachability Model and the

HMM Mobility Model of the Node X as a

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

774

weightened sum of the all probabilities of

reaching the destination node computed over

the all output symbols of the HMM Robot

Mobility System. The weightening is done by

the paramter 1/k.

3. Node M compares PMX to Pe;

3a. if PMX >= Pe, the message is copied from

node M to node X.

3b. if PMX < Pe, the message is not copied from

node M to node X. The continue with the step

4. Node M stops the active communication. Pe is

the threshold probability. Pe is defined by the

human user. The value of Pe describes the

netowrk senstivity. The efficiency of the

proposed method is strongly influenced by the

value of Pe.

3 DATA COLLECTION

The machine learning enhanced routing algorithm

for opportunistic networking ALPHA9 was tested on

artificially created data generated by the simulating

software NESCUAR 1.0. (Natural Environment

Simulator for Cooperative Unmanned Aerial Robots,

version 1.0).

3.1 The Simulation Survey

NESCUAR 1.0. is a simulation environment

designed and developed at the Faculty of

Information Technology at Czech Technical

University for the purposes of research, development

and testing of a communication platform for mobile

robots.

The random 2D model of a homogeneous

landscape was generated in the NESCUAR 1.0.

Then, the traffic of mobile robots was modeled. The

trajectory of each mobile robot was simulated

separately using probabilistic models and task-

oriented models, e.g. it is supposed the robots do

some meaningful movements. Mobile robots are

simulated as autonomous devices. Some random

noise was added to the data during the data

collection process to simulate measurement

inaccuracies.

The set of autonomous mobile robot positions

consists of triples containing three elements: x-

coordinate, y-coordinate and time. The finite data set

of mobile robot positions was collected

independently fo each mobile robot.

The simulation environment NESCUAR 1.0.

enables integration of external routing algorithmes.

Both the routing algorithmes, the were tested in this

simulator. The number of mobile robots was set to

25.

4 RESULTS

The positions of each of the 25 mobile robots was

clustered separately. Each cluster was described by

the following parameters: center latitude, center

longitude, maximal distance, mean distance, mean

speed, number of data elements.

The number of clusters obtained from the mobile

robot position data depends on the complexity of the

mobile robot movements. The numbers of clusters

generated for the different mobile robots changes

from 5 (the least complex traces) to 21 (the most

complex traces).

The mobility model based on HMM of each

mobile robot was identified. We received 25 HMM

based mobility models Ω

MOBBIE01

, Ω

MOBBIE02

, …

Ω

MOBBIE25

. The numbers of internal states of HMMs

generated for the different mobile robots changes

from 2 to 7.

The Node Reachability Model was constructed

from the data collected during the network scanning.

It was implemented as a multi-dimensional matrix.

Both the estimated Node Reachability Model and the

HMM Autonomous Robot Mobility Models were

used in the testing the ALPHA09 routing algorithm.

The number of mobile robots was set to 25.

Maximum delivery time was set to 120 seconds. The

maximal number of a message copying cmax was

set to 10, e.g. each message sent from any source

node could be most highly 10 times copyied on its

way through the network until it reached the

destination node. The pairs of communicating

mobile robots were randomly generated before the

simulation and were used in both simulations. The

Pe delivery probability limit value was selected by a

human operator; the automatic selection of this

parameter is a challenge for the further research.

5 CONCLUSIONS

Thus paper deals with the proposal of the novel

opportunistic networking routing algorithm, which

improves the flooding routing algorithm by

application of machine learning. The HMM

Autonomous Robot Mobility Models and Node

Reachability Model are constructed from the

observed data and used in a proposed routing

scheme in order to compute the combined

EnhancedRoutingAlgorithmforOpportunisticNetworking-OntheImprovementoftheBasicOpportunisticNetworking

RoutingAlgorithmbytheApplicationofMachineLearning

775

probabilities of message delivery to the destination

node. In comparision to the basic opportunistic

networking routing algorithm, the significant

improvement was observed in network load

reduction about 70%. The future work will be

focused on application of data preprocessing

techniques, exploration of the algirithmes for the

combined probability computation. We also intend

to redesign the proposed method to enable fully

distributed routing.

ACKNOWLEDGEMENTS

This work has been supported from the research

grant SGS12/094/OHK3/1T/18 "Metody a postupy

číslicového návrhu a jejich experimentální ověření"

hold by Assoc. Prof. Hana Kubátová, CSc. at the

Dept. of Digital Design at the Faculty of Information

Technology of the Czech Technical University in

Prague.

REFERENCES

Pelusi, L., Passarella, A., Conti, M., 2006.. Opportunistic

networking: data forwarding in disconnected mobile

ad hoc networks. In IEEE Communications Magazine,

Vol. 44, No. 11, pp. 134-141. IEEE.

Ravikiran, G.; Singh, S., 2004. Influence of mobility

models on the performance of routing protocols in ad-

hoc wireless networks. In Vehicular Technology

Conference, 2004. VTC 2004, vol. 4, p. 2185-2189.

IEEE.

Helbing, D., 2001. Traffic and Related Self-driven Many-

particles Systems. In Rev. Modern Physics, Vol. 73,

2001, pp. 1067-1141.

Qiang, Z. et al. 2009. Formal taxonomy research on

opportunistic networks. In Proceedings of 2nd IEEE

International Conference on Broadband Network &

Multimedia Technology 2009, pp. 854–857.

Ariyakhajorn, J. et al. 2006. A Comparative Study of

Random Waypoint and Gauss-Markov Mobility

Models in the Performance Evaluation of MANET. In

Proceedings of the International Symposium on

Communications and Information Technologies 2006,

pp. 894-899.

Timo, R., Blackmore, K., Hanlen, L., 2007. Strong

Stochastic Stability for MANET Mobility Models. In

Proceedings of the 15th IEEE International

Conference on Networks 2007, pp. 13-18. IEEE.

Harri, J., Filali, F., Bonnet, C., 2009. Mobility models for

vehicular ad hoc networks: a survey and taxonomy. In

IEEE Communications Surveys & Tutorials. Vol. 11,

No. 4, pp. 19-41. IEEE.

Rabiner, L. R., 1989. A tutorial on hidden Markov models

and selected applications in speech recognition, In

Proceedings of the IEEE, Vol. 77, No. 2, 1989, pp.

257-286. IEEE.

Hyunuk Kim, Ha Yoon Song, 2012 Formulating Human

Mobility Model in a Form of Continuous Time Markov

Chain, Procedia CS, Vol. 10, 2012.

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

776