Research ArticleA Novel User Classification Method for FemtocellNetwork by Using Affinity Propagation Algorithm andArtificial Neural Network

Afaz Uddin Ahmed1 Mohammad Tariqul Islam2 Mahamod Ismail2

Salehin Kibria1 and Haslina Arshad3

1 Space Science Centre (ANGKASA) Universiti Kebangsaan Malaysia (UKM) 43600 Bangi Selangor Malaysia2 Department of Electrical Electronic and Systems Engineering Universiti Kebangsaan Malaysia (UKM) 43600 BangiSelangor Malaysia

3 Centre of Artificial Intelligence Technology Faculty of Information Science and Technology Universiti Kebangsaan Malaysia (UKM)43600 Bangi Selangor Malaysia

Correspondence should be addressed to Mohammad Tariqul Islam titareqgmailcom

Received 5 May 2014 Accepted 12 June 2014 Published 16 July 2014

Academic Editor Su Fong Chien

Copyright copy 2014 Afaz Uddin Ahmed et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited

An artificial neural network (ANN) and affinity propagation (AP) algorithm based user categorization technique is presented Theproposed algorithm is designed for closed access femtocell network ANN is used for user classification process and AP algorithmis used to optimize the ANN training process AP selects the best possible training samples for faster ANN training cycleThe usersare distinguished by using the difference of received signal strength in a multielement femtocell device A previously developeddirective microstrip antenna is used to configure the femtocell device Simulation results show that for a particular house patternthe categorization technique without AP algorithm takes 5 indoor users and 10 outdoor users to attain an error-free operationWhile integrating AP algorithm with ANN the system takes 60 less training samples reducing the training time up to 50 Thisprocedure makes the femtocell more effective for closed access operation

1 Introduction

The number of cellular users has increased significantly overthe last few decades To some extent the cellular operatorswere able to provide the increasing demand of voice anddata services However due to massive growth of multimediaapplications the demand has reached beyond the limitwhere the existing macrocell cannot support such high nodedensity The increasing number of indoor data traffic (morethan 70) has made it quite difficult for the operators toprovide quality coverage using the existing macrocell [1]An alternative to this is that femtocell opens up a cost-effective solution by offloading excess voice and data trafficIt provides high quality indoor coverage which connects to

the core network throughwired backhaul [2]Without properplanning vast deployment of femtocell causes interferenceproblem in dense heterogeneous network Overlapping ofcoverage zones of both macrocell and femtocell is mostlysubjected to unwanted handover cell overshooting and highmobility events On the contrary the users buy the femtocellto enjoy the service of high quality indoor coverage [3ndash5]Access control algorithm is introduced to macrofemtocellnetwork to minimize the interference caused by excessmobility event Among the three access control mechanismsclosed access allows only particular users (mainly indoorusers) to get access to the network [6ndash9] In such techniquethe outdoor users cannot get access to the femtocell and themobility event reduces However under supreme coverage

Hindawi Publishing Corporatione Scientific World JournalVolume 2014 Article ID 253787 14 pageshttpdxdoiorg1011552014253787

2 The Scientific World Journal

of femtocell the outdoor user attempts to change basestation that creates signaling congestions in the networkTherefore multielement antenna configuration for femtocellapplication has been proposed in various articles It utilizesbeam-forming technique to control the coverage of femtocell[10ndash13] The antennas are usually mounted on the verticalsurface of the device with an individual scanning angle andseparation distance It creates null coverage in the interfer-ence regions and optimizes the coverage to avoid supremeoutdoor coverage Thus far all the efforts aimed to reducethe interference by making null coverage in the affectedregion

Smart antenna concept is an add-on to wireless networkin recent years Direction of arrival (DOA) estimation andbeam steering are considered the fundamental function ofthe smart antenna [14ndash16] In addition new features like userlocalization based on distinctive characters of usersrsquo signalsare also under consideration Array antennas give flexibilityto identify the users in an adaptive spatially sensitive mannerIt represents leading-edge smart antenna approach by usingdiverse signal-processing algorithm adjusted to real timeIn this paper a novel technique for user classification isproposed formultielement femtocell device by using artificialneural network (ANN) Clustering algorithmof affinity prop-agation (AP) is also introduced to make the process fasterand effective In multielement femtocell each of the antennashas different receiving gain in different angle that gives a setof received power pattern for every user Based on this thefemtocell is trained to identify the indoor and outdoor usersTo model the nonlinear relationship between the indoor andoutdoor user ANN is trained using randomly generated usersamples The trained ANN allows the femtocell to selectthe indoor and outdoor users from the antenna end Inaddition the training process is upgraded usingAP clusteringalgorithm This paper focuses on unwanted user admissioncontrol in femtocell to decrease the unwanted handover andsignaling congestion As femtocell distinguishes between theusers after a certain time it does not accept users outsidethe house which results in a less number of handoverrequests The performance of the proposed technique isshown as percentage of error rate in identifying the correctusers The remainder of the paper is described as followsuser categorization technique is explained in Section 2 anddetailed structure of the ANN and AP clustering algorithmis described in Sections 21 and 22 respectively Resultsand Discussions are in Section 3 and Conclusion is inSection 4

2 User Categorization in ClosedAccess Femtocell

Closed access mechanism in femtocell network avoidsunwanted handover and mobility events in dense macro-femtonetwork The users are predefined and femtocell onlyallows access to particular group of users In case of superiorcoverage which is beyond the threshold limit of the receivedsignal level outdoor users want to switch serving cell Asa result the femtocell gets continuous handover request

on SDCCH (stand-alone dedicated control channel) fromthe outdoor user This induces signalling congestion thatencompasses the core network for each request Most of thetime this event occurs due to overshooting of the femtocell inunwanted direction Use of multielement antenna instead ofomnidirectional antenna optimizes the coverage of femtocelland minimizes the overshooting effect However in initialstage femtocell does not have any prior knowledge ofhousersquos dimension and its own position In such conditionmultielement antenna also creates the overshooting problemIn multielement femtocell device the antennas are facedin different direction which allows forming of directionalbeam for particular user to avoid interference Since allthe proposed multiantenna concepts used planner antennaslike PIFA (planner inverted-F antenna) and patch antennapreviously designed microstrip antenna has been used inthis paper to simulate the femtocell device The antenna wasdesigned for LTE band 7 [17] It has a directional gain patternthat gives different receiving gain for different position ofthe user A 6-element antenna structure is considered for thefemtocell device with a scanning angle of 60∘ degree eachFor a particular user in the uplink the femtocell will have 6different received power patterns The relation between thereceived power and antenna gain which was shown in Friistransmission equation is given below [18]

119875119903= 119875119905times 119866119905(120579119905 120593119905) times 119866119903(120579119903 120593119903) times (

120582

4120587119877)

2

(1)

119875119903(dBm) = 119875

119905(dBm) + 119866

119905(120579119905 120593119905) (dB) + 119866

119903(120579119903 120593119903) (dB)

+ 20 log10

(120582

4120587119877)

⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟

free space pathloss

(dB)

(2)

where 119875119903and 119875

119905are the receive and transmit power respec-

tively 119866119905(120579119905 120593119905) and 119866

119903(120579119903 120593119903) are the transmit and receive

antenna gain at the receiver and transmitter direction respec-tively

The transmitting antenna of the userrsquos equipment isassumed to be omnidirectional Even if the antennas aredirectional the received signal strength on the antennapatch will change scantly as the mutual distance among theantennas is very small compared to the distance from thefemtocell to the user equipment In (2) the receiving gain andfree space path-loss for every user are different Comparingwith the distance between the users and femtocell the size ofthe femtocell is quite small As a result the free space path-loss is almost the same for each antenna element Figure 1visualizes the scenario of the above discussion

Femtocell antennas respond to an incoming wave froma given direction according to the pattern value in thatdirection Each of 6 antenna elements holds different gainpattern in each directionTherefore the received power variesdue to the prospective antenna gainThe variation of receivedpower is used to differentiate between the outdoor and indoorusers Femtocell performsmapping from incident wave to thereceived power pattern The neural network is trained to do

The Scientific World Journal 3

Outdoor-user

Femtocell

Indoor-user

Figure 1 User pattern in closed access femtocell network

the inverse mapping It uses the vectors comprised of energyE from all antennas over multiple instances of n

119864119894119899

= int

119879(119899+1)

119899times119879

119875119894(119905) 119889119905 119894 = 1 2 3 4 5 6

119864 =[[

[

1198641119899

1198646119899

]]

]

(3)

where T is the sampling period and 119899 = 0 1 2 3 In the training stage the ANN learns the behaviour of

indoor and outdoor users using the value of 119864 The networkcategorizes the user based on the previous learning For thetask a simulated environment is developed in MATLABIndoor and outdoor users are randomly generated usinguniformly distributed pseudorandomnumber A 2D layout ofa house is also designed considering the indoor and outdoorwalls Moreover AP clustering algorithm is used to filterout the best possible samples from randomly generated datapoints It allows the ANN to learn faster with the same levelof accuracy but a less number of iterations After the traininganother set of random samples are generated to evaluate theperformance of the network Standard path-loss model andadditive white Gaussian noise are considered in free spacepath-loss calculation

Pathloss119891(db) = 3846 + 20 log

10

119863 + 071198892119863indoor

+ 183119899((119899minus2)(119899+1) minus046)

+ 119908119871119894119908

(4)

where 119863 119908 119899 071198892119863indoor and Pathloss

119891are distance

number of walls number of floors penetration loss inside thehouse and path-loss of the users respectively [19]

Generating random indoor and outdoor users position

Clustering the generated samples using AP algorithm

Neural network trainingInput clustered samples of the received power of the antenna

pattern

Training process

Trained system

Output categorizes the indoor and outdoor users

Calculating free space path-loss andreceived powers in multi-element

Figure 2 Proposed femtocell user selection technique using ANNand AP algorithm

Using AP algorithm and ANN femtocell determines theusersrsquo category to allow access For random values of 119864 theneural network determines the usersrsquo category by giving anoutput of ldquo+1rdquo or ldquominus1rdquo The details of process is projected in aflow chart in Figure 2

21 Artificial Neural Network for User Categorization Arti-ficial neural network (ANN) is a machine-learning processthat is modelled after the brain architecture Like the brainrsquossmallest cell neuron it contains hundreds of processing unitswired together as a complex network It is trained using thesample data to predict the behaviour of the future data [20]User categorizing is a supervised learning process A modelis prepared through a training process where it is requiredto make predictions and is corrected when those predictionsare wrong The training process continues until the modelachieves a desired level of accuracy on the training data Ingeneral algorithms are presented in groups by similaritiesin terms of their operation process and function Thereare algorithms that could easily fit into multiple categorieslike learning vector quantization It is both an instance-based method and a neural network inspired method Thereare categories that have the same name that describes theproblem and the class of algorithm such as regression andclustering The popular machine leaning algorithms areregression instance-based methods regularization methodsdecision tree learning Bayesian kernel methods clustering

4 The Scientific World Journal

sgn(x) =

Sigmoid function

Input 1

Input 2

Input 3

Input N

Input layer Hidden layer(s) Output layer

Neuron 1

Neuron 2

Neuron Nminus 1

Neuron N

W11

W12

W21

W31

W32

W22

WN(Nminus1)

W3(Nminus1)

W1(Nminus1)

WN1WN2

b1

b2

bnminus1

bn

120575(1)

120575(1)

120575(N minus 1)

W1N

W2N

W(Nminus1)N

sgn(middot)OutputIndoor(+1) orOutdoor(minus1)

+1 if x ge 0

minus1 if x lt 0

Figure 3 Structure of MLPFFBP in the proposed technique

methods association rule learning deep learning dimen-sionality reduction ensemble methods and artificial neuralnetwork [21] However in machine learning algorithmsthemselves there is no perfect model just a good enoughmodel depending on how the application layout is designedANN has many attractive theoretic properties specificallythe ability to detect nonpredefined relations such as nonlineareffects andor interactions These theoretic advantages comeat the cost of reduced interpretability of the model outputMany authors have analysed the same data set based on thesefactors with both standard statisticalmethods andANN [22ndash24]

In the proposed technique multilayer perceptron feedforward backpropagation (MLPFFBP) neural network is usedto categorize the usersMLPFFBPuses error backpropagationto adjust the weights of the neurons There are two passes inthe layers of the network forward pass and backward passThe network consists of three layers input layer output layerand the hidden layer The input layer is fed with initial dataThe output layer gives the desired solution In between thereexists a series of hidden layersThe primary layer is connectedwith the input layer and the last layer is connected to theoutput layer Each subsequent layer is connected with theprevious layer Based on the network design each hiddenlayer consists of multiple numbers of neurons The neuronsuse differentiable transfer function to generate the outputDuring the training period the input and output values ofthe network are specified and based on these values and thehidden layer builds up a set of weights for the neurons [25]

The differentiable transfer function (tansig) used hereis a sigmoid function In multilayer sigmoid function ifthe input vector is very large the weight becomes so smallto prevent the transfer function being saturated Thus thegradient will be very small and the neural network will

be very slow On the contrary higher number of trainingsamples with higher number of neurons makes the networkmore accurate but such a process makes the network bulkyand time-consuming For this preprocessing steps are addedin-between the input layers and the hidden layers Theperformance of the neural network is made more effective byusing a preprocessing step in training sample selection In thiscase AP clustering algorithm is used to select the best-suitedsamples for the network training

In Figure 3 1198871 1198872 119887

119899minus1 119887119899

and 11990811 11990812sdot sdot sdot 11990821

11990822sdot sdot sdot 119908119873(119873minus1)

119908119873119873

are the biases and the weights of thenetwork nodes respectively Biases are also consideredthe primary weights that are initially put as 1 Moreoverldquosignumrdquo function is used to compute the actual response ofthe perceptron The final output from the last neuron passesthrough the ldquosignumrdquo function that gives the binary output

The transfer function is

120593 (V) =1

1 + exp (minusV) (5)

The signum function is

sgn (119909) = +1 if 119909 ge 0

minus1 if 119909 lt 0(6)

The weights are calculated as

119908 (119899 + 1) = 119908 (119899) + 120572 lowast 119908 (119899 minus 1) + 120578 lowast 120575 (119899) lowast 119910

120575 (119899) = 1205931015840

(V) lowast (119889 minus 119910)

(7)

where 120572 120578 119889 119910 and 120575 are the mobility factor the trainingparameter the desired output the real output and the localgradient for the nodes of the network respectively [26 27]

The Scientific World Journal 5

After the training process of the network the femtocell takesthe 6-element antennas received power as input and gives thecategory of the users in the output

22 Affinity Propagation Algorithm for Selecting the Best Sam-ples AP algorithm is a recent clustering algorithm proposedby Frey and Dueck [28] It is widely accepted because ofits high quality set of cluster samples The proposed userclassification in neural network is a supervised techniqueTheperformance of the network is subjected to the nature andquantity of the training samples Higher number of trainingsamples led to precise values of the neuronsrsquo weight butit makes the training process slower Clustering of data setbased on similarities is a vital step in data analysis problemA common practice in both supervised and unsupervisedlearning is to cluster the data based on the similarities [29 30]Affinity propagation (AP) is the latest clustering algorithmthat reduces the redundancy of the training data set Itaccelerates the computing process of ANN by reducing thesample numbers

Traditional clustering algorithms follow random selec-tion of initial data subset as exemplars and refine it iterativelyAP takes an input set of pairwise similarities between thedata points and finds the clusters based on maximum totalsimilarities between the exemplars and the data points [31]The real messages are exchanged between the data pointsuntil the finest set of exemplars and corresponding clustersprogressively emerges It has a better clustering performancethan K-means K-medians fuzzy c-means Hill combining(HC) and self-organizing map (SOM) algorithms [32 33]It is computationally efficient and simple to implement andcustomize In AP algorithm all the sample data points areconsidered a possible candidate to be the desired exemplarsEach step exchanges real-valued messages between themuntil a superior set of exemplar shows up Messages areupdated based on simple formulae that reflect on the sum-product or max-product It updates the rules until themagnitude of the messages reflects on the current affinityfor choosing another data point as its exemplar Each datapoint is considered a note in the network The process of thealgorithm is described briefly below

Input is a set of pairwise similarities as

119904 (119894 119896) = minus1003817100381710038171003817119909119894 minus 119909

119896

1003817100381710038171003817

2

119894 = 119896 (squared Euclidean distance)

where (119894 119896) isin 1 1198732

119904 (119894 119896) isin R

(8)

Here 119904(119894 119896) isin R indicates how well suited the data point 119896 isas an exemplar for data point i

For each data point 119896 a real number 119904(119896 119896) represents thepreference that is to be considered as an exemplar

119904 (119896 119896) = 120588 forall119896 isin 1 119873 (9)

Initialization set availabilities to zero for all 119894 119896

119886(119894 119896) = 0

Table 1 System parameters

System parameters ValuerangeFrequency 253GHzNumber of training indoor users 10Number of training outdoor users 15Number of randomly placed users (after training) 20Femtocell antenna height 1mUser equipment height 1mFrequency 26GHzUE transmit power (fixed) 13 dBmIndoor wall loss 5 dBOutdoor wall loss 10 dBShadow fading std 6 dBWhite noise power density minus174 dBmHzNumber of neurons in hidden layer 10

Repeat responsibility and availability updates until con-vergence

forall119894 119896 119903 (119894 119896) = 119904 (119894 119896) minusmax [119904 (119894 1198961015840) + 119886 (119894 1198961015840

)]

forall119894 119896 119886 (119894 119896)

=

summax [0 119903 (1198941015840 119896)] for 119896 = 119894

min [0 119903 (119896 119896)]

+sum11989410158401198941015840notin119894119896

max [0 119903 (1198941015840 119896)] for 119896 = 119894

(10)

Output is assignments 119888 = (1198881 119888

119873) where 119888

119894=

arg max119896[119886(119894 119896) + 119903(119894 119896)] Here 119888

119894indexes the clusterrsquos exem-

plar at which point 119894 is assigned If point 119894 is a cluster withpoint 119896 as the exemplar then 119888

119894= 119896 and 119888

119896= 119896 [34]

3 Results and Discussions

A layout of functioning area is modelled with a femto-cell in the middle of the house Six-microstrip antennasare operating with 60∘ separation angle on the same axisinside the femtocell device A previously designed microstripantenna is used here to configure the directive gain patternof each antenna element [17] The house has indoor andoutdoor walls that decrease the strength of the signal basedon their thickness Initially random indoor and outdoor usersare generated and the received powers are measured Thedimension of the house is set to 7m times 6m In Figure 4(a)the users and the house are plotted in a 20m times 20m windowThe radiation pattern of the microstrip antenna is shown inFigure 4(b)

To demonstrate the performance of the technique ANNis initially trained without using AP clustering Randomsamples are generated by varying the numbers of indoorand outdoor users In the performance analysis stage againrandom samples are generated to categorize users using theprevious experiences The system parameters that have beenused in the simulation are given in Table 1 In the model theoutdoor wall loss is considered higher than the indoor wall

6 The Scientific World Journal

00

2

4

6

8

10

12

14

16

18

20

2 4 6 8 10 12 14 16 18 20

FemtocellIndoor userOutdoor user

WallIndoor connectingOutdoor connecting line

(a)

Farfield gain Abs (120593 = 0)

Farfield (f = 253) [1]Farfield (f = 266) [1]

Frequency = 266

Thetadeg versus (dB)

Main lobe magnitude = 65dBMain lobe direction = minus50deg

Side lobe level = minus103 dB

0

30

60

90

120

150

180

minus30

minus60

minus90

minus120

minus150

minus15minus40 10

Frequency = 253

Main lobe magnitude = 61dB

Main lobe direction = minus40degAngular width (3dB) = 919deg

Angular width (3dB) = 821deg

(b)

Figure 4 (a) Layout of the simulation environment (b) Radiation pattern of the microstrip antenna at 253GHz and 266GHz

One of the reasons is that usually the outdoorwalls are thickerthan the indoor wall with more concrete and steel materialsfor the foundation or shape This increases the loss exponentof the outdoor walls Another reason is that outdoor walls aremore subjected to rust and moist from the environment thatweakens the incoming signal [35]

31 Femtocell Network Performance with ANN Figure 5(a)shows the training stage of the femtocell device The reddots are the outdoor users and green dots are the indoorusers In Figure 5(b) random users are generated for the

femtocell to classify the indoor and outdoor users by usingthe learning experience Femtocell only allows connectionto the indoor users to be connected The green connectinglines between the femtocell and the indoor users confirm theproper recognition of the users

Figures 6(a) and 6(b) show the training state and perfor-mance validation state for a simulation with 10 indoor and15 outdoor training samples The minimum gradient of theANN is set to 1 times 10minus6 In this particular iteration ANNtakes 38 epochs to train up and adjust the values of biasesand weights to achieve the minimum gradient value Theldquovalidation graphrdquo shows a downward curve It confirms that

The Scientific World Journal 7

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 2 4 6 8 10 12 14 16 18 200

2

4

6

8

10

12

14

16

18

20

(b)

Figure 5 (a) Training and (b) testing of the femtocell device

Table 2

Number of samples ANN training performance AP + ANN training performance Performance comparison after AP

Indoor Outdoor Training time(sec)

Numberof

epochslowastAP clusteringtime (sec)

Total trainingtime (sec)

Number ofepochslowast

Training timedecreases ()

Number ofepochs decreases

()5 10 15165 21 02885 07647 13 4957 38096 11 15655 21 03633 08257 14 4725 33337 12 16221 22 03921 09071 14 4407 36368 13 16416 23 04172 0934 15 4310 34789 14 16443 24 04226 09504 16 4220 333310 15 16504 26 04244 09611 16 4176 384611 16 16541 26 0439 09803 17 4073 346112 17 16669 27 04461 09848 18 4092 333313 18 16709 27 04556 09951 19 4044 296214 19 16801 28 04671 09951 19 4077 321415 20 16819 28 04702 09964 19 4075 321416 21 1688 29 04704 09982 19 4086 344817 22 17322 30 04997 10017 21 4217 3018 23 17767 31 05073 10397 22 4148 290319 24 17966 32 05234 10695 22 4047 312520 25 17996 34 05484 10511 22 4159 3529lowastThe fraction values of the epochs are expressed by the nearest integer values

after every epoch the latest values of the weights and biasesvalidate the previous training samples

Figure 6(c) shows the performance of the femtocell inpercentage of error for different number of outdoor andindoor training user samples In every iteration the networkis tested using 20 random users to verify the performanceIn both types of users the error rate is quite high at thebeginning Due to lack of knowledge of the usersrsquo behaviourthe system cannot categorize the nature of the randomlycreated users For the same number of indoor users outdoorusersrsquo percentage of error rate is higher This is becauseof the unpredictable nature of wireless signal propagation

from the outdoor users end The outdoor walls their shapesand constructing materials also add more variations in theoutdoor users signal strength due to absorption losses anddiffraction loss As a result the ANN requires higher numberof outdoor users training samples for categorizing the usersHowever after 5 indoor and 10 outdoor user samples thenetwork reaches the perfectionwith error-free user detectionIt shows that the performance of the indoor sample isbetter than the outdoor sample In the indoor situation thevariation of the signal strength is limited to a certain boundThe effects of indoor free-space loss refraction diffractionreflection aperture-medium coupling loss and absorption

8 The Scientific World Journal

100

Gra

dien

t

10minus10

10minus5

Val f

ail

1

05

00 5 10 15 20 25 30 35

Gradient = 90011e minus 007 at epoch 38

Validation checks = 0 at epoch 38

38 epochs

(a)

Best validation performance is 76703e minus 008 at epoch 38

100

10minus2

10minus6

10minus4

10minus8

Mea

n sq

uare

d er

ror (

mse

)

0 5 10 15 20 25 30 35

38 Epochs

TrainValidation

TestBest

(b)

100

80

60

40

20

0

Erro

rs (

)

24

68

10

Training indoor users

5

10

15

Training outdoor users

(c)

Figure 6 (a) Training state and (b) performance of best validation (c) Performance of femtocell for different number of samples

are comparatively smaller which allows the system to verifyany random users signal strength within a certain variationof received power strength Nevertheless the number of thesample users always depends on the geographical shape of thehousesThe system requires higher number of indoor sampleswhen the variation bounds overlap with the outdoor usersvariation bound Such a case is studied below

The proposed method is now tested in a more complexscenario A ldquoUrdquo shaped house layout is designed to test theperformance of the system In this layout indoor wall isignored Figures 7(a) and 7(b) show the training and testingprocess of the femtocell network The challenging shape of

the house makes the user pattern more improvised thanthe previous one In this case the system requires highernumber of indoor and outdoor training user samples to reachan error-free performance Figure 7(c) shows the requirednumber of indoor and outdoor users against the percentageof error occurrences in detecting the usersrsquo category Herethe required number of users for both categories is above 25users The rest of the performance analysis of the process isdone using the previous layout of the house

32 Femtocell Network Performance with ANN and APClustering Algorithm AP algorithm clusters the users into

The Scientific World Journal 9

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

010

2030

010

2030

0

20

40

60

80

100

Training outdoor usersTraining indoor users

Err

ors (

)

1020

30

1020

30 utdoor useining indoor us

(c)

Figure 7 (a) Training and (b) testing process of the femtocell with ldquoUrdquo shaped house (c) Performance of femtocell for different number ofsamples

Indoor user

Outdoor user

0

minus200

minus400

minus600

minus800

minus1200

minus1000

minus1400

minus1600

minus1800

Fitn

ess (

net s

imila

rity)

of q

uant

ized

inte

rmed

iate

solu

tion

0 5 10 15 20 25 30

Number of iterations

Figure 8 Fitness of AP clustering algorithm for indoor and outdoorusers

subgroups based on their power pattern and selects arepresentative from each subgroup Unlike other clustering

methods AP algorithm selects the clusterssubgroups basedon the samples nature If the nature of the sample variesimmensely the number of clusters gets higherThe clusteringperformance of the AP algorithm is presented in Figure 8 as aform of achieved fitness (net similarities) with respect to theiteration number Both the outdoor and indoor users reachtheir best fitness before 8 iterations However a safe marginof 25 is kept to ensure the best fitness for both types of users

Figure 9(a) shows the general ANN training processDuring the training the ANN adjusts the values of theweights and the biases of the network In Figure 9(b) theAP algorithm clusters the users based on their similaritiespower pattern A representative has been chosen among thedata points of a subgroup which has most of similaritieswith the other data points of the subgroup There might alsoexist subgroups with only one data point Figure shows thatinstead of training ANN with 15 outdoor users and 10 indoorusers the AP selects 3 outdoor users and 3 indoor usersFigures 9(c) and 9(d) show the performance of the networkwith and without AP algorithm For a random simulationboth processes show the same accuracy

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

2 The Scientific World Journal

of femtocell the outdoor user attempts to change basestation that creates signaling congestions in the networkTherefore multielement antenna configuration for femtocellapplication has been proposed in various articles It utilizesbeam-forming technique to control the coverage of femtocell[10ndash13] The antennas are usually mounted on the verticalsurface of the device with an individual scanning angle andseparation distance It creates null coverage in the interfer-ence regions and optimizes the coverage to avoid supremeoutdoor coverage Thus far all the efforts aimed to reducethe interference by making null coverage in the affectedregion

Smart antenna concept is an add-on to wireless networkin recent years Direction of arrival (DOA) estimation andbeam steering are considered the fundamental function ofthe smart antenna [14ndash16] In addition new features like userlocalization based on distinctive characters of usersrsquo signalsare also under consideration Array antennas give flexibilityto identify the users in an adaptive spatially sensitive mannerIt represents leading-edge smart antenna approach by usingdiverse signal-processing algorithm adjusted to real timeIn this paper a novel technique for user classification isproposed formultielement femtocell device by using artificialneural network (ANN) Clustering algorithmof affinity prop-agation (AP) is also introduced to make the process fasterand effective In multielement femtocell each of the antennashas different receiving gain in different angle that gives a setof received power pattern for every user Based on this thefemtocell is trained to identify the indoor and outdoor usersTo model the nonlinear relationship between the indoor andoutdoor user ANN is trained using randomly generated usersamples The trained ANN allows the femtocell to selectthe indoor and outdoor users from the antenna end Inaddition the training process is upgraded usingAP clusteringalgorithm This paper focuses on unwanted user admissioncontrol in femtocell to decrease the unwanted handover andsignaling congestion As femtocell distinguishes between theusers after a certain time it does not accept users outsidethe house which results in a less number of handoverrequests The performance of the proposed technique isshown as percentage of error rate in identifying the correctusers The remainder of the paper is described as followsuser categorization technique is explained in Section 2 anddetailed structure of the ANN and AP clustering algorithmis described in Sections 21 and 22 respectively Resultsand Discussions are in Section 3 and Conclusion is inSection 4

2 User Categorization in ClosedAccess Femtocell

Closed access mechanism in femtocell network avoidsunwanted handover and mobility events in dense macro-femtonetwork The users are predefined and femtocell onlyallows access to particular group of users In case of superiorcoverage which is beyond the threshold limit of the receivedsignal level outdoor users want to switch serving cell Asa result the femtocell gets continuous handover request

on SDCCH (stand-alone dedicated control channel) fromthe outdoor user This induces signalling congestion thatencompasses the core network for each request Most of thetime this event occurs due to overshooting of the femtocell inunwanted direction Use of multielement antenna instead ofomnidirectional antenna optimizes the coverage of femtocelland minimizes the overshooting effect However in initialstage femtocell does not have any prior knowledge ofhousersquos dimension and its own position In such conditionmultielement antenna also creates the overshooting problemIn multielement femtocell device the antennas are facedin different direction which allows forming of directionalbeam for particular user to avoid interference Since allthe proposed multiantenna concepts used planner antennaslike PIFA (planner inverted-F antenna) and patch antennapreviously designed microstrip antenna has been used inthis paper to simulate the femtocell device The antenna wasdesigned for LTE band 7 [17] It has a directional gain patternthat gives different receiving gain for different position ofthe user A 6-element antenna structure is considered for thefemtocell device with a scanning angle of 60∘ degree eachFor a particular user in the uplink the femtocell will have 6different received power patterns The relation between thereceived power and antenna gain which was shown in Friistransmission equation is given below [18]

119875119903= 119875119905times 119866119905(120579119905 120593119905) times 119866119903(120579119903 120593119903) times (

120582

4120587119877)

2

(1)

119875119903(dBm) = 119875

119905(dBm) + 119866

119905(120579119905 120593119905) (dB) + 119866

119903(120579119903 120593119903) (dB)

+ 20 log10

(120582

4120587119877)

⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟

free space pathloss

(dB)

(2)

where 119875119903and 119875

119905are the receive and transmit power respec-

tively 119866119905(120579119905 120593119905) and 119866

119903(120579119903 120593119903) are the transmit and receive

antenna gain at the receiver and transmitter direction respec-tively

The transmitting antenna of the userrsquos equipment isassumed to be omnidirectional Even if the antennas aredirectional the received signal strength on the antennapatch will change scantly as the mutual distance among theantennas is very small compared to the distance from thefemtocell to the user equipment In (2) the receiving gain andfree space path-loss for every user are different Comparingwith the distance between the users and femtocell the size ofthe femtocell is quite small As a result the free space path-loss is almost the same for each antenna element Figure 1visualizes the scenario of the above discussion

Femtocell antennas respond to an incoming wave froma given direction according to the pattern value in thatdirection Each of 6 antenna elements holds different gainpattern in each directionTherefore the received power variesdue to the prospective antenna gainThe variation of receivedpower is used to differentiate between the outdoor and indoorusers Femtocell performsmapping from incident wave to thereceived power pattern The neural network is trained to do

The Scientific World Journal 3

Outdoor-user

Femtocell

Indoor-user

Figure 1 User pattern in closed access femtocell network

the inverse mapping It uses the vectors comprised of energyE from all antennas over multiple instances of n

119864119894119899

= int

119879(119899+1)

119899times119879

119875119894(119905) 119889119905 119894 = 1 2 3 4 5 6

119864 =[[

[

1198641119899

1198646119899

]]

]

(3)

where T is the sampling period and 119899 = 0 1 2 3 In the training stage the ANN learns the behaviour of

indoor and outdoor users using the value of 119864 The networkcategorizes the user based on the previous learning For thetask a simulated environment is developed in MATLABIndoor and outdoor users are randomly generated usinguniformly distributed pseudorandomnumber A 2D layout ofa house is also designed considering the indoor and outdoorwalls Moreover AP clustering algorithm is used to filterout the best possible samples from randomly generated datapoints It allows the ANN to learn faster with the same levelof accuracy but a less number of iterations After the traininganother set of random samples are generated to evaluate theperformance of the network Standard path-loss model andadditive white Gaussian noise are considered in free spacepath-loss calculation

Pathloss119891(db) = 3846 + 20 log

10

119863 + 071198892119863indoor

+ 183119899((119899minus2)(119899+1) minus046)

+ 119908119871119894119908

(4)

where 119863 119908 119899 071198892119863indoor and Pathloss

119891are distance

number of walls number of floors penetration loss inside thehouse and path-loss of the users respectively [19]

Generating random indoor and outdoor users position

Clustering the generated samples using AP algorithm

Neural network trainingInput clustered samples of the received power of the antenna

pattern

Training process

Trained system

Output categorizes the indoor and outdoor users

Calculating free space path-loss andreceived powers in multi-element

Figure 2 Proposed femtocell user selection technique using ANNand AP algorithm

Using AP algorithm and ANN femtocell determines theusersrsquo category to allow access For random values of 119864 theneural network determines the usersrsquo category by giving anoutput of ldquo+1rdquo or ldquominus1rdquo The details of process is projected in aflow chart in Figure 2

21 Artificial Neural Network for User Categorization Arti-ficial neural network (ANN) is a machine-learning processthat is modelled after the brain architecture Like the brainrsquossmallest cell neuron it contains hundreds of processing unitswired together as a complex network It is trained using thesample data to predict the behaviour of the future data [20]User categorizing is a supervised learning process A modelis prepared through a training process where it is requiredto make predictions and is corrected when those predictionsare wrong The training process continues until the modelachieves a desired level of accuracy on the training data Ingeneral algorithms are presented in groups by similaritiesin terms of their operation process and function Thereare algorithms that could easily fit into multiple categorieslike learning vector quantization It is both an instance-based method and a neural network inspired method Thereare categories that have the same name that describes theproblem and the class of algorithm such as regression andclustering The popular machine leaning algorithms areregression instance-based methods regularization methodsdecision tree learning Bayesian kernel methods clustering

4 The Scientific World Journal

sgn(x) =

Sigmoid function

Input 1

Input 2

Input 3

Input N

Input layer Hidden layer(s) Output layer

Neuron 1

Neuron 2

Neuron Nminus 1

Neuron N

W11

W12

W21

W31

W32

W22

WN(Nminus1)

W3(Nminus1)

W1(Nminus1)

WN1WN2

b1

b2

bnminus1

bn

120575(1)

120575(1)

120575(N minus 1)

W1N

W2N

W(Nminus1)N

sgn(middot)OutputIndoor(+1) orOutdoor(minus1)

+1 if x ge 0

minus1 if x lt 0

Figure 3 Structure of MLPFFBP in the proposed technique

methods association rule learning deep learning dimen-sionality reduction ensemble methods and artificial neuralnetwork [21] However in machine learning algorithmsthemselves there is no perfect model just a good enoughmodel depending on how the application layout is designedANN has many attractive theoretic properties specificallythe ability to detect nonpredefined relations such as nonlineareffects andor interactions These theoretic advantages comeat the cost of reduced interpretability of the model outputMany authors have analysed the same data set based on thesefactors with both standard statisticalmethods andANN [22ndash24]

In the proposed technique multilayer perceptron feedforward backpropagation (MLPFFBP) neural network is usedto categorize the usersMLPFFBPuses error backpropagationto adjust the weights of the neurons There are two passes inthe layers of the network forward pass and backward passThe network consists of three layers input layer output layerand the hidden layer The input layer is fed with initial dataThe output layer gives the desired solution In between thereexists a series of hidden layersThe primary layer is connectedwith the input layer and the last layer is connected to theoutput layer Each subsequent layer is connected with theprevious layer Based on the network design each hiddenlayer consists of multiple numbers of neurons The neuronsuse differentiable transfer function to generate the outputDuring the training period the input and output values ofthe network are specified and based on these values and thehidden layer builds up a set of weights for the neurons [25]

The differentiable transfer function (tansig) used hereis a sigmoid function In multilayer sigmoid function ifthe input vector is very large the weight becomes so smallto prevent the transfer function being saturated Thus thegradient will be very small and the neural network will

be very slow On the contrary higher number of trainingsamples with higher number of neurons makes the networkmore accurate but such a process makes the network bulkyand time-consuming For this preprocessing steps are addedin-between the input layers and the hidden layers Theperformance of the neural network is made more effective byusing a preprocessing step in training sample selection In thiscase AP clustering algorithm is used to select the best-suitedsamples for the network training

In Figure 3 1198871 1198872 119887

119899minus1 119887119899

and 11990811 11990812sdot sdot sdot 11990821

11990822sdot sdot sdot 119908119873(119873minus1)

119908119873119873

are the biases and the weights of thenetwork nodes respectively Biases are also consideredthe primary weights that are initially put as 1 Moreoverldquosignumrdquo function is used to compute the actual response ofthe perceptron The final output from the last neuron passesthrough the ldquosignumrdquo function that gives the binary output

The transfer function is

120593 (V) =1

1 + exp (minusV) (5)

The signum function is

sgn (119909) = +1 if 119909 ge 0

minus1 if 119909 lt 0(6)

The weights are calculated as

119908 (119899 + 1) = 119908 (119899) + 120572 lowast 119908 (119899 minus 1) + 120578 lowast 120575 (119899) lowast 119910

120575 (119899) = 1205931015840

(V) lowast (119889 minus 119910)

(7)

where 120572 120578 119889 119910 and 120575 are the mobility factor the trainingparameter the desired output the real output and the localgradient for the nodes of the network respectively [26 27]

The Scientific World Journal 5

After the training process of the network the femtocell takesthe 6-element antennas received power as input and gives thecategory of the users in the output

22 Affinity Propagation Algorithm for Selecting the Best Sam-ples AP algorithm is a recent clustering algorithm proposedby Frey and Dueck [28] It is widely accepted because ofits high quality set of cluster samples The proposed userclassification in neural network is a supervised techniqueTheperformance of the network is subjected to the nature andquantity of the training samples Higher number of trainingsamples led to precise values of the neuronsrsquo weight butit makes the training process slower Clustering of data setbased on similarities is a vital step in data analysis problemA common practice in both supervised and unsupervisedlearning is to cluster the data based on the similarities [29 30]Affinity propagation (AP) is the latest clustering algorithmthat reduces the redundancy of the training data set Itaccelerates the computing process of ANN by reducing thesample numbers

Traditional clustering algorithms follow random selec-tion of initial data subset as exemplars and refine it iterativelyAP takes an input set of pairwise similarities between thedata points and finds the clusters based on maximum totalsimilarities between the exemplars and the data points [31]The real messages are exchanged between the data pointsuntil the finest set of exemplars and corresponding clustersprogressively emerges It has a better clustering performancethan K-means K-medians fuzzy c-means Hill combining(HC) and self-organizing map (SOM) algorithms [32 33]It is computationally efficient and simple to implement andcustomize In AP algorithm all the sample data points areconsidered a possible candidate to be the desired exemplarsEach step exchanges real-valued messages between themuntil a superior set of exemplar shows up Messages areupdated based on simple formulae that reflect on the sum-product or max-product It updates the rules until themagnitude of the messages reflects on the current affinityfor choosing another data point as its exemplar Each datapoint is considered a note in the network The process of thealgorithm is described briefly below

Input is a set of pairwise similarities as

119904 (119894 119896) = minus1003817100381710038171003817119909119894 minus 119909

119896

1003817100381710038171003817

2

119894 = 119896 (squared Euclidean distance)

where (119894 119896) isin 1 1198732

119904 (119894 119896) isin R

(8)

Here 119904(119894 119896) isin R indicates how well suited the data point 119896 isas an exemplar for data point i

For each data point 119896 a real number 119904(119896 119896) represents thepreference that is to be considered as an exemplar

119904 (119896 119896) = 120588 forall119896 isin 1 119873 (9)

Initialization set availabilities to zero for all 119894 119896

119886(119894 119896) = 0

Table 1 System parameters

System parameters ValuerangeFrequency 253GHzNumber of training indoor users 10Number of training outdoor users 15Number of randomly placed users (after training) 20Femtocell antenna height 1mUser equipment height 1mFrequency 26GHzUE transmit power (fixed) 13 dBmIndoor wall loss 5 dBOutdoor wall loss 10 dBShadow fading std 6 dBWhite noise power density minus174 dBmHzNumber of neurons in hidden layer 10

Repeat responsibility and availability updates until con-vergence

forall119894 119896 119903 (119894 119896) = 119904 (119894 119896) minusmax [119904 (119894 1198961015840) + 119886 (119894 1198961015840

)]

forall119894 119896 119886 (119894 119896)

=

summax [0 119903 (1198941015840 119896)] for 119896 = 119894

min [0 119903 (119896 119896)]

+sum11989410158401198941015840notin119894119896

max [0 119903 (1198941015840 119896)] for 119896 = 119894

(10)

Output is assignments 119888 = (1198881 119888

119873) where 119888

119894=

arg max119896[119886(119894 119896) + 119903(119894 119896)] Here 119888

119894indexes the clusterrsquos exem-

plar at which point 119894 is assigned If point 119894 is a cluster withpoint 119896 as the exemplar then 119888

119894= 119896 and 119888

119896= 119896 [34]

3 Results and Discussions

A layout of functioning area is modelled with a femto-cell in the middle of the house Six-microstrip antennasare operating with 60∘ separation angle on the same axisinside the femtocell device A previously designed microstripantenna is used here to configure the directive gain patternof each antenna element [17] The house has indoor andoutdoor walls that decrease the strength of the signal basedon their thickness Initially random indoor and outdoor usersare generated and the received powers are measured Thedimension of the house is set to 7m times 6m In Figure 4(a)the users and the house are plotted in a 20m times 20m windowThe radiation pattern of the microstrip antenna is shown inFigure 4(b)

To demonstrate the performance of the technique ANNis initially trained without using AP clustering Randomsamples are generated by varying the numbers of indoorand outdoor users In the performance analysis stage againrandom samples are generated to categorize users using theprevious experiences The system parameters that have beenused in the simulation are given in Table 1 In the model theoutdoor wall loss is considered higher than the indoor wall

6 The Scientific World Journal

00

2

4

6

8

10

12

14

16

18

20

2 4 6 8 10 12 14 16 18 20

FemtocellIndoor userOutdoor user

WallIndoor connectingOutdoor connecting line

(a)

Farfield gain Abs (120593 = 0)

Farfield (f = 253) [1]Farfield (f = 266) [1]

Frequency = 266

Thetadeg versus (dB)

Main lobe magnitude = 65dBMain lobe direction = minus50deg

Side lobe level = minus103 dB

0

30

60

90

120

150

180

minus30

minus60

minus90

minus120

minus150

minus15minus40 10

Frequency = 253

Main lobe magnitude = 61dB

Main lobe direction = minus40degAngular width (3dB) = 919deg

Angular width (3dB) = 821deg

(b)

Figure 4 (a) Layout of the simulation environment (b) Radiation pattern of the microstrip antenna at 253GHz and 266GHz

One of the reasons is that usually the outdoorwalls are thickerthan the indoor wall with more concrete and steel materialsfor the foundation or shape This increases the loss exponentof the outdoor walls Another reason is that outdoor walls aremore subjected to rust and moist from the environment thatweakens the incoming signal [35]

31 Femtocell Network Performance with ANN Figure 5(a)shows the training stage of the femtocell device The reddots are the outdoor users and green dots are the indoorusers In Figure 5(b) random users are generated for the

femtocell to classify the indoor and outdoor users by usingthe learning experience Femtocell only allows connectionto the indoor users to be connected The green connectinglines between the femtocell and the indoor users confirm theproper recognition of the users

Figures 6(a) and 6(b) show the training state and perfor-mance validation state for a simulation with 10 indoor and15 outdoor training samples The minimum gradient of theANN is set to 1 times 10minus6 In this particular iteration ANNtakes 38 epochs to train up and adjust the values of biasesand weights to achieve the minimum gradient value Theldquovalidation graphrdquo shows a downward curve It confirms that

The Scientific World Journal 7

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 2 4 6 8 10 12 14 16 18 200

2

4

6

8

10

12

14

16

18

20

(b)

Figure 5 (a) Training and (b) testing of the femtocell device

Table 2

Number of samples ANN training performance AP + ANN training performance Performance comparison after AP

Indoor Outdoor Training time(sec)

Numberof

epochslowastAP clusteringtime (sec)

Total trainingtime (sec)

Number ofepochslowast

Training timedecreases ()

Number ofepochs decreases

()5 10 15165 21 02885 07647 13 4957 38096 11 15655 21 03633 08257 14 4725 33337 12 16221 22 03921 09071 14 4407 36368 13 16416 23 04172 0934 15 4310 34789 14 16443 24 04226 09504 16 4220 333310 15 16504 26 04244 09611 16 4176 384611 16 16541 26 0439 09803 17 4073 346112 17 16669 27 04461 09848 18 4092 333313 18 16709 27 04556 09951 19 4044 296214 19 16801 28 04671 09951 19 4077 321415 20 16819 28 04702 09964 19 4075 321416 21 1688 29 04704 09982 19 4086 344817 22 17322 30 04997 10017 21 4217 3018 23 17767 31 05073 10397 22 4148 290319 24 17966 32 05234 10695 22 4047 312520 25 17996 34 05484 10511 22 4159 3529lowastThe fraction values of the epochs are expressed by the nearest integer values

after every epoch the latest values of the weights and biasesvalidate the previous training samples

Figure 6(c) shows the performance of the femtocell inpercentage of error for different number of outdoor andindoor training user samples In every iteration the networkis tested using 20 random users to verify the performanceIn both types of users the error rate is quite high at thebeginning Due to lack of knowledge of the usersrsquo behaviourthe system cannot categorize the nature of the randomlycreated users For the same number of indoor users outdoorusersrsquo percentage of error rate is higher This is becauseof the unpredictable nature of wireless signal propagation

from the outdoor users end The outdoor walls their shapesand constructing materials also add more variations in theoutdoor users signal strength due to absorption losses anddiffraction loss As a result the ANN requires higher numberof outdoor users training samples for categorizing the usersHowever after 5 indoor and 10 outdoor user samples thenetwork reaches the perfectionwith error-free user detectionIt shows that the performance of the indoor sample isbetter than the outdoor sample In the indoor situation thevariation of the signal strength is limited to a certain boundThe effects of indoor free-space loss refraction diffractionreflection aperture-medium coupling loss and absorption

8 The Scientific World Journal

100

Gra

dien

t

10minus10

10minus5

Val f

ail

1

05

00 5 10 15 20 25 30 35

Gradient = 90011e minus 007 at epoch 38

Validation checks = 0 at epoch 38

38 epochs

(a)

Best validation performance is 76703e minus 008 at epoch 38

100

10minus2

10minus6

10minus4

10minus8

Mea

n sq

uare

d er

ror (

mse

)

0 5 10 15 20 25 30 35

38 Epochs

TrainValidation

TestBest

(b)

100

80

60

40

20

0

Erro

rs (

)

24

68

10

Training indoor users

5

10

15

Training outdoor users

(c)

Figure 6 (a) Training state and (b) performance of best validation (c) Performance of femtocell for different number of samples

are comparatively smaller which allows the system to verifyany random users signal strength within a certain variationof received power strength Nevertheless the number of thesample users always depends on the geographical shape of thehousesThe system requires higher number of indoor sampleswhen the variation bounds overlap with the outdoor usersvariation bound Such a case is studied below

The proposed method is now tested in a more complexscenario A ldquoUrdquo shaped house layout is designed to test theperformance of the system In this layout indoor wall isignored Figures 7(a) and 7(b) show the training and testingprocess of the femtocell network The challenging shape of

the house makes the user pattern more improvised thanthe previous one In this case the system requires highernumber of indoor and outdoor training user samples to reachan error-free performance Figure 7(c) shows the requirednumber of indoor and outdoor users against the percentageof error occurrences in detecting the usersrsquo category Herethe required number of users for both categories is above 25users The rest of the performance analysis of the process isdone using the previous layout of the house

32 Femtocell Network Performance with ANN and APClustering Algorithm AP algorithm clusters the users into

The Scientific World Journal 9

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

010

2030

010

2030

0

20

40

60

80

100

Training outdoor usersTraining indoor users

Err

ors (

)

1020

30

1020

30 utdoor useining indoor us

(c)

Figure 7 (a) Training and (b) testing process of the femtocell with ldquoUrdquo shaped house (c) Performance of femtocell for different number ofsamples

Indoor user

Outdoor user

0

minus200

minus400

minus600

minus800

minus1200

minus1000

minus1400

minus1600

minus1800

Fitn

ess (

net s

imila

rity)

of q

uant

ized

inte

rmed

iate

solu

tion

0 5 10 15 20 25 30

Number of iterations

Figure 8 Fitness of AP clustering algorithm for indoor and outdoorusers

subgroups based on their power pattern and selects arepresentative from each subgroup Unlike other clustering

methods AP algorithm selects the clusterssubgroups basedon the samples nature If the nature of the sample variesimmensely the number of clusters gets higherThe clusteringperformance of the AP algorithm is presented in Figure 8 as aform of achieved fitness (net similarities) with respect to theiteration number Both the outdoor and indoor users reachtheir best fitness before 8 iterations However a safe marginof 25 is kept to ensure the best fitness for both types of users

Figure 9(a) shows the general ANN training processDuring the training the ANN adjusts the values of theweights and the biases of the network In Figure 9(b) theAP algorithm clusters the users based on their similaritiespower pattern A representative has been chosen among thedata points of a subgroup which has most of similaritieswith the other data points of the subgroup There might alsoexist subgroups with only one data point Figure shows thatinstead of training ANN with 15 outdoor users and 10 indoorusers the AP selects 3 outdoor users and 3 indoor usersFigures 9(c) and 9(d) show the performance of the networkwith and without AP algorithm For a random simulationboth processes show the same accuracy

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World Journal 3

Outdoor-user

Femtocell

Indoor-user

Figure 1 User pattern in closed access femtocell network

the inverse mapping It uses the vectors comprised of energyE from all antennas over multiple instances of n

119864119894119899

= int

119879(119899+1)

119899times119879

119875119894(119905) 119889119905 119894 = 1 2 3 4 5 6

119864 =[[

[

1198641119899

1198646119899

]]

]

(3)

where T is the sampling period and 119899 = 0 1 2 3 In the training stage the ANN learns the behaviour of

indoor and outdoor users using the value of 119864 The networkcategorizes the user based on the previous learning For thetask a simulated environment is developed in MATLABIndoor and outdoor users are randomly generated usinguniformly distributed pseudorandomnumber A 2D layout ofa house is also designed considering the indoor and outdoorwalls Moreover AP clustering algorithm is used to filterout the best possible samples from randomly generated datapoints It allows the ANN to learn faster with the same levelof accuracy but a less number of iterations After the traininganother set of random samples are generated to evaluate theperformance of the network Standard path-loss model andadditive white Gaussian noise are considered in free spacepath-loss calculation

Pathloss119891(db) = 3846 + 20 log

10

119863 + 071198892119863indoor

+ 183119899((119899minus2)(119899+1) minus046)

+ 119908119871119894119908

(4)

where 119863 119908 119899 071198892119863indoor and Pathloss

119891are distance

number of walls number of floors penetration loss inside thehouse and path-loss of the users respectively [19]

Generating random indoor and outdoor users position

Clustering the generated samples using AP algorithm

Neural network trainingInput clustered samples of the received power of the antenna

pattern

Training process

Trained system

Output categorizes the indoor and outdoor users

Calculating free space path-loss andreceived powers in multi-element

Figure 2 Proposed femtocell user selection technique using ANNand AP algorithm

Using AP algorithm and ANN femtocell determines theusersrsquo category to allow access For random values of 119864 theneural network determines the usersrsquo category by giving anoutput of ldquo+1rdquo or ldquominus1rdquo The details of process is projected in aflow chart in Figure 2

21 Artificial Neural Network for User Categorization Arti-ficial neural network (ANN) is a machine-learning processthat is modelled after the brain architecture Like the brainrsquossmallest cell neuron it contains hundreds of processing unitswired together as a complex network It is trained using thesample data to predict the behaviour of the future data [20]User categorizing is a supervised learning process A modelis prepared through a training process where it is requiredto make predictions and is corrected when those predictionsare wrong The training process continues until the modelachieves a desired level of accuracy on the training data Ingeneral algorithms are presented in groups by similaritiesin terms of their operation process and function Thereare algorithms that could easily fit into multiple categorieslike learning vector quantization It is both an instance-based method and a neural network inspired method Thereare categories that have the same name that describes theproblem and the class of algorithm such as regression andclustering The popular machine leaning algorithms areregression instance-based methods regularization methodsdecision tree learning Bayesian kernel methods clustering

4 The Scientific World Journal

sgn(x) =

Sigmoid function

Input 1

Input 2

Input 3

Input N

Input layer Hidden layer(s) Output layer

Neuron 1

Neuron 2

Neuron Nminus 1

Neuron N

W11

W12

W21

W31

W32

W22

WN(Nminus1)

W3(Nminus1)

W1(Nminus1)

WN1WN2

b1

b2

bnminus1

bn

120575(1)

120575(1)

120575(N minus 1)

W1N

W2N

W(Nminus1)N

sgn(middot)OutputIndoor(+1) orOutdoor(minus1)

+1 if x ge 0

minus1 if x lt 0

Figure 3 Structure of MLPFFBP in the proposed technique

methods association rule learning deep learning dimen-sionality reduction ensemble methods and artificial neuralnetwork [21] However in machine learning algorithmsthemselves there is no perfect model just a good enoughmodel depending on how the application layout is designedANN has many attractive theoretic properties specificallythe ability to detect nonpredefined relations such as nonlineareffects andor interactions These theoretic advantages comeat the cost of reduced interpretability of the model outputMany authors have analysed the same data set based on thesefactors with both standard statisticalmethods andANN [22ndash24]

In the proposed technique multilayer perceptron feedforward backpropagation (MLPFFBP) neural network is usedto categorize the usersMLPFFBPuses error backpropagationto adjust the weights of the neurons There are two passes inthe layers of the network forward pass and backward passThe network consists of three layers input layer output layerand the hidden layer The input layer is fed with initial dataThe output layer gives the desired solution In between thereexists a series of hidden layersThe primary layer is connectedwith the input layer and the last layer is connected to theoutput layer Each subsequent layer is connected with theprevious layer Based on the network design each hiddenlayer consists of multiple numbers of neurons The neuronsuse differentiable transfer function to generate the outputDuring the training period the input and output values ofthe network are specified and based on these values and thehidden layer builds up a set of weights for the neurons [25]

The differentiable transfer function (tansig) used hereis a sigmoid function In multilayer sigmoid function ifthe input vector is very large the weight becomes so smallto prevent the transfer function being saturated Thus thegradient will be very small and the neural network will

be very slow On the contrary higher number of trainingsamples with higher number of neurons makes the networkmore accurate but such a process makes the network bulkyand time-consuming For this preprocessing steps are addedin-between the input layers and the hidden layers Theperformance of the neural network is made more effective byusing a preprocessing step in training sample selection In thiscase AP clustering algorithm is used to select the best-suitedsamples for the network training

In Figure 3 1198871 1198872 119887

119899minus1 119887119899

and 11990811 11990812sdot sdot sdot 11990821

11990822sdot sdot sdot 119908119873(119873minus1)

119908119873119873

are the biases and the weights of thenetwork nodes respectively Biases are also consideredthe primary weights that are initially put as 1 Moreoverldquosignumrdquo function is used to compute the actual response ofthe perceptron The final output from the last neuron passesthrough the ldquosignumrdquo function that gives the binary output

The transfer function is

120593 (V) =1

1 + exp (minusV) (5)

The signum function is

sgn (119909) = +1 if 119909 ge 0

minus1 if 119909 lt 0(6)

The weights are calculated as

119908 (119899 + 1) = 119908 (119899) + 120572 lowast 119908 (119899 minus 1) + 120578 lowast 120575 (119899) lowast 119910

120575 (119899) = 1205931015840

(V) lowast (119889 minus 119910)

(7)

where 120572 120578 119889 119910 and 120575 are the mobility factor the trainingparameter the desired output the real output and the localgradient for the nodes of the network respectively [26 27]

The Scientific World Journal 5

After the training process of the network the femtocell takesthe 6-element antennas received power as input and gives thecategory of the users in the output

22 Affinity Propagation Algorithm for Selecting the Best Sam-ples AP algorithm is a recent clustering algorithm proposedby Frey and Dueck [28] It is widely accepted because ofits high quality set of cluster samples The proposed userclassification in neural network is a supervised techniqueTheperformance of the network is subjected to the nature andquantity of the training samples Higher number of trainingsamples led to precise values of the neuronsrsquo weight butit makes the training process slower Clustering of data setbased on similarities is a vital step in data analysis problemA common practice in both supervised and unsupervisedlearning is to cluster the data based on the similarities [29 30]Affinity propagation (AP) is the latest clustering algorithmthat reduces the redundancy of the training data set Itaccelerates the computing process of ANN by reducing thesample numbers

Traditional clustering algorithms follow random selec-tion of initial data subset as exemplars and refine it iterativelyAP takes an input set of pairwise similarities between thedata points and finds the clusters based on maximum totalsimilarities between the exemplars and the data points [31]The real messages are exchanged between the data pointsuntil the finest set of exemplars and corresponding clustersprogressively emerges It has a better clustering performancethan K-means K-medians fuzzy c-means Hill combining(HC) and self-organizing map (SOM) algorithms [32 33]It is computationally efficient and simple to implement andcustomize In AP algorithm all the sample data points areconsidered a possible candidate to be the desired exemplarsEach step exchanges real-valued messages between themuntil a superior set of exemplar shows up Messages areupdated based on simple formulae that reflect on the sum-product or max-product It updates the rules until themagnitude of the messages reflects on the current affinityfor choosing another data point as its exemplar Each datapoint is considered a note in the network The process of thealgorithm is described briefly below

Input is a set of pairwise similarities as

119904 (119894 119896) = minus1003817100381710038171003817119909119894 minus 119909

119896

1003817100381710038171003817

2

119894 = 119896 (squared Euclidean distance)

where (119894 119896) isin 1 1198732

119904 (119894 119896) isin R

(8)

Here 119904(119894 119896) isin R indicates how well suited the data point 119896 isas an exemplar for data point i

For each data point 119896 a real number 119904(119896 119896) represents thepreference that is to be considered as an exemplar

119904 (119896 119896) = 120588 forall119896 isin 1 119873 (9)

Initialization set availabilities to zero for all 119894 119896

119886(119894 119896) = 0

Table 1 System parameters

System parameters ValuerangeFrequency 253GHzNumber of training indoor users 10Number of training outdoor users 15Number of randomly placed users (after training) 20Femtocell antenna height 1mUser equipment height 1mFrequency 26GHzUE transmit power (fixed) 13 dBmIndoor wall loss 5 dBOutdoor wall loss 10 dBShadow fading std 6 dBWhite noise power density minus174 dBmHzNumber of neurons in hidden layer 10

Repeat responsibility and availability updates until con-vergence

forall119894 119896 119903 (119894 119896) = 119904 (119894 119896) minusmax [119904 (119894 1198961015840) + 119886 (119894 1198961015840

)]

forall119894 119896 119886 (119894 119896)

=

summax [0 119903 (1198941015840 119896)] for 119896 = 119894

min [0 119903 (119896 119896)]

+sum11989410158401198941015840notin119894119896

max [0 119903 (1198941015840 119896)] for 119896 = 119894

(10)

Output is assignments 119888 = (1198881 119888

119873) where 119888

119894=

arg max119896[119886(119894 119896) + 119903(119894 119896)] Here 119888

119894indexes the clusterrsquos exem-

plar at which point 119894 is assigned If point 119894 is a cluster withpoint 119896 as the exemplar then 119888

119894= 119896 and 119888

119896= 119896 [34]

3 Results and Discussions

A layout of functioning area is modelled with a femto-cell in the middle of the house Six-microstrip antennasare operating with 60∘ separation angle on the same axisinside the femtocell device A previously designed microstripantenna is used here to configure the directive gain patternof each antenna element [17] The house has indoor andoutdoor walls that decrease the strength of the signal basedon their thickness Initially random indoor and outdoor usersare generated and the received powers are measured Thedimension of the house is set to 7m times 6m In Figure 4(a)the users and the house are plotted in a 20m times 20m windowThe radiation pattern of the microstrip antenna is shown inFigure 4(b)

To demonstrate the performance of the technique ANNis initially trained without using AP clustering Randomsamples are generated by varying the numbers of indoorand outdoor users In the performance analysis stage againrandom samples are generated to categorize users using theprevious experiences The system parameters that have beenused in the simulation are given in Table 1 In the model theoutdoor wall loss is considered higher than the indoor wall

6 The Scientific World Journal

00

2

4

6

8

10

12

14

16

18

20

2 4 6 8 10 12 14 16 18 20

FemtocellIndoor userOutdoor user

WallIndoor connectingOutdoor connecting line

(a)

Farfield gain Abs (120593 = 0)

Farfield (f = 253) [1]Farfield (f = 266) [1]

Frequency = 266

Thetadeg versus (dB)

Main lobe magnitude = 65dBMain lobe direction = minus50deg

Side lobe level = minus103 dB

0

30

60

90

120

150

180

minus30

minus60

minus90

minus120

minus150

minus15minus40 10

Frequency = 253

Main lobe magnitude = 61dB

Main lobe direction = minus40degAngular width (3dB) = 919deg

Angular width (3dB) = 821deg

(b)

Figure 4 (a) Layout of the simulation environment (b) Radiation pattern of the microstrip antenna at 253GHz and 266GHz

One of the reasons is that usually the outdoorwalls are thickerthan the indoor wall with more concrete and steel materialsfor the foundation or shape This increases the loss exponentof the outdoor walls Another reason is that outdoor walls aremore subjected to rust and moist from the environment thatweakens the incoming signal [35]

31 Femtocell Network Performance with ANN Figure 5(a)shows the training stage of the femtocell device The reddots are the outdoor users and green dots are the indoorusers In Figure 5(b) random users are generated for the

femtocell to classify the indoor and outdoor users by usingthe learning experience Femtocell only allows connectionto the indoor users to be connected The green connectinglines between the femtocell and the indoor users confirm theproper recognition of the users

Figures 6(a) and 6(b) show the training state and perfor-mance validation state for a simulation with 10 indoor and15 outdoor training samples The minimum gradient of theANN is set to 1 times 10minus6 In this particular iteration ANNtakes 38 epochs to train up and adjust the values of biasesand weights to achieve the minimum gradient value Theldquovalidation graphrdquo shows a downward curve It confirms that

The Scientific World Journal 7

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 2 4 6 8 10 12 14 16 18 200

2

4

6

8

10

12

14

16

18

20

(b)

Figure 5 (a) Training and (b) testing of the femtocell device

Table 2

Number of samples ANN training performance AP + ANN training performance Performance comparison after AP

Indoor Outdoor Training time(sec)

Numberof

epochslowastAP clusteringtime (sec)

Total trainingtime (sec)

Number ofepochslowast

Training timedecreases ()

Number ofepochs decreases

()5 10 15165 21 02885 07647 13 4957 38096 11 15655 21 03633 08257 14 4725 33337 12 16221 22 03921 09071 14 4407 36368 13 16416 23 04172 0934 15 4310 34789 14 16443 24 04226 09504 16 4220 333310 15 16504 26 04244 09611 16 4176 384611 16 16541 26 0439 09803 17 4073 346112 17 16669 27 04461 09848 18 4092 333313 18 16709 27 04556 09951 19 4044 296214 19 16801 28 04671 09951 19 4077 321415 20 16819 28 04702 09964 19 4075 321416 21 1688 29 04704 09982 19 4086 344817 22 17322 30 04997 10017 21 4217 3018 23 17767 31 05073 10397 22 4148 290319 24 17966 32 05234 10695 22 4047 312520 25 17996 34 05484 10511 22 4159 3529lowastThe fraction values of the epochs are expressed by the nearest integer values

after every epoch the latest values of the weights and biasesvalidate the previous training samples

Figure 6(c) shows the performance of the femtocell inpercentage of error for different number of outdoor andindoor training user samples In every iteration the networkis tested using 20 random users to verify the performanceIn both types of users the error rate is quite high at thebeginning Due to lack of knowledge of the usersrsquo behaviourthe system cannot categorize the nature of the randomlycreated users For the same number of indoor users outdoorusersrsquo percentage of error rate is higher This is becauseof the unpredictable nature of wireless signal propagation

from the outdoor users end The outdoor walls their shapesand constructing materials also add more variations in theoutdoor users signal strength due to absorption losses anddiffraction loss As a result the ANN requires higher numberof outdoor users training samples for categorizing the usersHowever after 5 indoor and 10 outdoor user samples thenetwork reaches the perfectionwith error-free user detectionIt shows that the performance of the indoor sample isbetter than the outdoor sample In the indoor situation thevariation of the signal strength is limited to a certain boundThe effects of indoor free-space loss refraction diffractionreflection aperture-medium coupling loss and absorption

8 The Scientific World Journal

100

Gra

dien

t

10minus10

10minus5

Val f

ail

1

05

00 5 10 15 20 25 30 35

Gradient = 90011e minus 007 at epoch 38

Validation checks = 0 at epoch 38

38 epochs

(a)

Best validation performance is 76703e minus 008 at epoch 38

100

10minus2

10minus6

10minus4

10minus8

Mea

n sq

uare

d er

ror (

mse

)

0 5 10 15 20 25 30 35

38 Epochs

TrainValidation

TestBest

(b)

100

80

60

40

20

0

Erro

rs (

)

24

68

10

Training indoor users

5

10

15

Training outdoor users

(c)

Figure 6 (a) Training state and (b) performance of best validation (c) Performance of femtocell for different number of samples

are comparatively smaller which allows the system to verifyany random users signal strength within a certain variationof received power strength Nevertheless the number of thesample users always depends on the geographical shape of thehousesThe system requires higher number of indoor sampleswhen the variation bounds overlap with the outdoor usersvariation bound Such a case is studied below

The proposed method is now tested in a more complexscenario A ldquoUrdquo shaped house layout is designed to test theperformance of the system In this layout indoor wall isignored Figures 7(a) and 7(b) show the training and testingprocess of the femtocell network The challenging shape of

the house makes the user pattern more improvised thanthe previous one In this case the system requires highernumber of indoor and outdoor training user samples to reachan error-free performance Figure 7(c) shows the requirednumber of indoor and outdoor users against the percentageof error occurrences in detecting the usersrsquo category Herethe required number of users for both categories is above 25users The rest of the performance analysis of the process isdone using the previous layout of the house

32 Femtocell Network Performance with ANN and APClustering Algorithm AP algorithm clusters the users into

The Scientific World Journal 9

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

010

2030

010

2030

0

20

40

60

80

100

Training outdoor usersTraining indoor users

Err

ors (

)

1020

30

1020

30 utdoor useining indoor us

(c)

Figure 7 (a) Training and (b) testing process of the femtocell with ldquoUrdquo shaped house (c) Performance of femtocell for different number ofsamples

Indoor user

Outdoor user

0

minus200

minus400

minus600

minus800

minus1200

minus1000

minus1400

minus1600

minus1800

Fitn

ess (

net s

imila

rity)

of q

uant

ized

inte

rmed

iate

solu

tion

0 5 10 15 20 25 30

Number of iterations

Figure 8 Fitness of AP clustering algorithm for indoor and outdoorusers

subgroups based on their power pattern and selects arepresentative from each subgroup Unlike other clustering

methods AP algorithm selects the clusterssubgroups basedon the samples nature If the nature of the sample variesimmensely the number of clusters gets higherThe clusteringperformance of the AP algorithm is presented in Figure 8 as aform of achieved fitness (net similarities) with respect to theiteration number Both the outdoor and indoor users reachtheir best fitness before 8 iterations However a safe marginof 25 is kept to ensure the best fitness for both types of users

Figure 9(a) shows the general ANN training processDuring the training the ANN adjusts the values of theweights and the biases of the network In Figure 9(b) theAP algorithm clusters the users based on their similaritiespower pattern A representative has been chosen among thedata points of a subgroup which has most of similaritieswith the other data points of the subgroup There might alsoexist subgroups with only one data point Figure shows thatinstead of training ANN with 15 outdoor users and 10 indoorusers the AP selects 3 outdoor users and 3 indoor usersFigures 9(c) and 9(d) show the performance of the networkwith and without AP algorithm For a random simulationboth processes show the same accuracy

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

4 The Scientific World Journal

sgn(x) =

Sigmoid function

Input 1

Input 2

Input 3

Input N

Input layer Hidden layer(s) Output layer

Neuron 1

Neuron 2

Neuron Nminus 1

Neuron N

W11

W12

W21

W31

W32

W22

WN(Nminus1)

W3(Nminus1)

W1(Nminus1)

WN1WN2

b1

b2

bnminus1

bn

120575(1)

120575(1)

120575(N minus 1)

W1N

W2N

W(Nminus1)N

sgn(middot)OutputIndoor(+1) orOutdoor(minus1)

+1 if x ge 0

minus1 if x lt 0

Figure 3 Structure of MLPFFBP in the proposed technique

methods association rule learning deep learning dimen-sionality reduction ensemble methods and artificial neuralnetwork [21] However in machine learning algorithmsthemselves there is no perfect model just a good enoughmodel depending on how the application layout is designedANN has many attractive theoretic properties specificallythe ability to detect nonpredefined relations such as nonlineareffects andor interactions These theoretic advantages comeat the cost of reduced interpretability of the model outputMany authors have analysed the same data set based on thesefactors with both standard statisticalmethods andANN [22ndash24]

In the proposed technique multilayer perceptron feedforward backpropagation (MLPFFBP) neural network is usedto categorize the usersMLPFFBPuses error backpropagationto adjust the weights of the neurons There are two passes inthe layers of the network forward pass and backward passThe network consists of three layers input layer output layerand the hidden layer The input layer is fed with initial dataThe output layer gives the desired solution In between thereexists a series of hidden layersThe primary layer is connectedwith the input layer and the last layer is connected to theoutput layer Each subsequent layer is connected with theprevious layer Based on the network design each hiddenlayer consists of multiple numbers of neurons The neuronsuse differentiable transfer function to generate the outputDuring the training period the input and output values ofthe network are specified and based on these values and thehidden layer builds up a set of weights for the neurons [25]

The differentiable transfer function (tansig) used hereis a sigmoid function In multilayer sigmoid function ifthe input vector is very large the weight becomes so smallto prevent the transfer function being saturated Thus thegradient will be very small and the neural network will

be very slow On the contrary higher number of trainingsamples with higher number of neurons makes the networkmore accurate but such a process makes the network bulkyand time-consuming For this preprocessing steps are addedin-between the input layers and the hidden layers Theperformance of the neural network is made more effective byusing a preprocessing step in training sample selection In thiscase AP clustering algorithm is used to select the best-suitedsamples for the network training

In Figure 3 1198871 1198872 119887

119899minus1 119887119899

and 11990811 11990812sdot sdot sdot 11990821

11990822sdot sdot sdot 119908119873(119873minus1)

119908119873119873

are the biases and the weights of thenetwork nodes respectively Biases are also consideredthe primary weights that are initially put as 1 Moreoverldquosignumrdquo function is used to compute the actual response ofthe perceptron The final output from the last neuron passesthrough the ldquosignumrdquo function that gives the binary output

The transfer function is

120593 (V) =1

1 + exp (minusV) (5)

The signum function is

sgn (119909) = +1 if 119909 ge 0

minus1 if 119909 lt 0(6)

The weights are calculated as

119908 (119899 + 1) = 119908 (119899) + 120572 lowast 119908 (119899 minus 1) + 120578 lowast 120575 (119899) lowast 119910

120575 (119899) = 1205931015840

(V) lowast (119889 minus 119910)

(7)

where 120572 120578 119889 119910 and 120575 are the mobility factor the trainingparameter the desired output the real output and the localgradient for the nodes of the network respectively [26 27]

The Scientific World Journal 5

After the training process of the network the femtocell takesthe 6-element antennas received power as input and gives thecategory of the users in the output

22 Affinity Propagation Algorithm for Selecting the Best Sam-ples AP algorithm is a recent clustering algorithm proposedby Frey and Dueck [28] It is widely accepted because ofits high quality set of cluster samples The proposed userclassification in neural network is a supervised techniqueTheperformance of the network is subjected to the nature andquantity of the training samples Higher number of trainingsamples led to precise values of the neuronsrsquo weight butit makes the training process slower Clustering of data setbased on similarities is a vital step in data analysis problemA common practice in both supervised and unsupervisedlearning is to cluster the data based on the similarities [29 30]Affinity propagation (AP) is the latest clustering algorithmthat reduces the redundancy of the training data set Itaccelerates the computing process of ANN by reducing thesample numbers

Traditional clustering algorithms follow random selec-tion of initial data subset as exemplars and refine it iterativelyAP takes an input set of pairwise similarities between thedata points and finds the clusters based on maximum totalsimilarities between the exemplars and the data points [31]The real messages are exchanged between the data pointsuntil the finest set of exemplars and corresponding clustersprogressively emerges It has a better clustering performancethan K-means K-medians fuzzy c-means Hill combining(HC) and self-organizing map (SOM) algorithms [32 33]It is computationally efficient and simple to implement andcustomize In AP algorithm all the sample data points areconsidered a possible candidate to be the desired exemplarsEach step exchanges real-valued messages between themuntil a superior set of exemplar shows up Messages areupdated based on simple formulae that reflect on the sum-product or max-product It updates the rules until themagnitude of the messages reflects on the current affinityfor choosing another data point as its exemplar Each datapoint is considered a note in the network The process of thealgorithm is described briefly below

Input is a set of pairwise similarities as

119904 (119894 119896) = minus1003817100381710038171003817119909119894 minus 119909

119896

1003817100381710038171003817

2

119894 = 119896 (squared Euclidean distance)

where (119894 119896) isin 1 1198732

119904 (119894 119896) isin R

(8)

Here 119904(119894 119896) isin R indicates how well suited the data point 119896 isas an exemplar for data point i

For each data point 119896 a real number 119904(119896 119896) represents thepreference that is to be considered as an exemplar

119904 (119896 119896) = 120588 forall119896 isin 1 119873 (9)

Initialization set availabilities to zero for all 119894 119896

119886(119894 119896) = 0

Table 1 System parameters

System parameters ValuerangeFrequency 253GHzNumber of training indoor users 10Number of training outdoor users 15Number of randomly placed users (after training) 20Femtocell antenna height 1mUser equipment height 1mFrequency 26GHzUE transmit power (fixed) 13 dBmIndoor wall loss 5 dBOutdoor wall loss 10 dBShadow fading std 6 dBWhite noise power density minus174 dBmHzNumber of neurons in hidden layer 10

Repeat responsibility and availability updates until con-vergence

forall119894 119896 119903 (119894 119896) = 119904 (119894 119896) minusmax [119904 (119894 1198961015840) + 119886 (119894 1198961015840

)]

forall119894 119896 119886 (119894 119896)

=

summax [0 119903 (1198941015840 119896)] for 119896 = 119894

min [0 119903 (119896 119896)]

+sum11989410158401198941015840notin119894119896

max [0 119903 (1198941015840 119896)] for 119896 = 119894

(10)

Output is assignments 119888 = (1198881 119888

119873) where 119888

119894=

arg max119896[119886(119894 119896) + 119903(119894 119896)] Here 119888

119894indexes the clusterrsquos exem-

plar at which point 119894 is assigned If point 119894 is a cluster withpoint 119896 as the exemplar then 119888

119894= 119896 and 119888

119896= 119896 [34]

3 Results and Discussions

A layout of functioning area is modelled with a femto-cell in the middle of the house Six-microstrip antennasare operating with 60∘ separation angle on the same axisinside the femtocell device A previously designed microstripantenna is used here to configure the directive gain patternof each antenna element [17] The house has indoor andoutdoor walls that decrease the strength of the signal basedon their thickness Initially random indoor and outdoor usersare generated and the received powers are measured Thedimension of the house is set to 7m times 6m In Figure 4(a)the users and the house are plotted in a 20m times 20m windowThe radiation pattern of the microstrip antenna is shown inFigure 4(b)

To demonstrate the performance of the technique ANNis initially trained without using AP clustering Randomsamples are generated by varying the numbers of indoorand outdoor users In the performance analysis stage againrandom samples are generated to categorize users using theprevious experiences The system parameters that have beenused in the simulation are given in Table 1 In the model theoutdoor wall loss is considered higher than the indoor wall

6 The Scientific World Journal

00

2

4

6

8

10

12

14

16

18

20

2 4 6 8 10 12 14 16 18 20

FemtocellIndoor userOutdoor user

WallIndoor connectingOutdoor connecting line

(a)

Farfield gain Abs (120593 = 0)

Farfield (f = 253) [1]Farfield (f = 266) [1]

Frequency = 266

Thetadeg versus (dB)

Main lobe magnitude = 65dBMain lobe direction = minus50deg

Side lobe level = minus103 dB

0

30

60

90

120

150

180

minus30

minus60

minus90

minus120

minus150

minus15minus40 10

Frequency = 253

Main lobe magnitude = 61dB

Main lobe direction = minus40degAngular width (3dB) = 919deg

Angular width (3dB) = 821deg

(b)

Figure 4 (a) Layout of the simulation environment (b) Radiation pattern of the microstrip antenna at 253GHz and 266GHz

One of the reasons is that usually the outdoorwalls are thickerthan the indoor wall with more concrete and steel materialsfor the foundation or shape This increases the loss exponentof the outdoor walls Another reason is that outdoor walls aremore subjected to rust and moist from the environment thatweakens the incoming signal [35]

31 Femtocell Network Performance with ANN Figure 5(a)shows the training stage of the femtocell device The reddots are the outdoor users and green dots are the indoorusers In Figure 5(b) random users are generated for the

femtocell to classify the indoor and outdoor users by usingthe learning experience Femtocell only allows connectionto the indoor users to be connected The green connectinglines between the femtocell and the indoor users confirm theproper recognition of the users

Figures 6(a) and 6(b) show the training state and perfor-mance validation state for a simulation with 10 indoor and15 outdoor training samples The minimum gradient of theANN is set to 1 times 10minus6 In this particular iteration ANNtakes 38 epochs to train up and adjust the values of biasesand weights to achieve the minimum gradient value Theldquovalidation graphrdquo shows a downward curve It confirms that

The Scientific World Journal 7

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 2 4 6 8 10 12 14 16 18 200

2

4

6

8

10

12

14

16

18

20

(b)

Figure 5 (a) Training and (b) testing of the femtocell device

Table 2

Number of samples ANN training performance AP + ANN training performance Performance comparison after AP

Indoor Outdoor Training time(sec)

Numberof

epochslowastAP clusteringtime (sec)

Total trainingtime (sec)

Number ofepochslowast

Training timedecreases ()

Number ofepochs decreases

()5 10 15165 21 02885 07647 13 4957 38096 11 15655 21 03633 08257 14 4725 33337 12 16221 22 03921 09071 14 4407 36368 13 16416 23 04172 0934 15 4310 34789 14 16443 24 04226 09504 16 4220 333310 15 16504 26 04244 09611 16 4176 384611 16 16541 26 0439 09803 17 4073 346112 17 16669 27 04461 09848 18 4092 333313 18 16709 27 04556 09951 19 4044 296214 19 16801 28 04671 09951 19 4077 321415 20 16819 28 04702 09964 19 4075 321416 21 1688 29 04704 09982 19 4086 344817 22 17322 30 04997 10017 21 4217 3018 23 17767 31 05073 10397 22 4148 290319 24 17966 32 05234 10695 22 4047 312520 25 17996 34 05484 10511 22 4159 3529lowastThe fraction values of the epochs are expressed by the nearest integer values

after every epoch the latest values of the weights and biasesvalidate the previous training samples

Figure 6(c) shows the performance of the femtocell inpercentage of error for different number of outdoor andindoor training user samples In every iteration the networkis tested using 20 random users to verify the performanceIn both types of users the error rate is quite high at thebeginning Due to lack of knowledge of the usersrsquo behaviourthe system cannot categorize the nature of the randomlycreated users For the same number of indoor users outdoorusersrsquo percentage of error rate is higher This is becauseof the unpredictable nature of wireless signal propagation

from the outdoor users end The outdoor walls their shapesand constructing materials also add more variations in theoutdoor users signal strength due to absorption losses anddiffraction loss As a result the ANN requires higher numberof outdoor users training samples for categorizing the usersHowever after 5 indoor and 10 outdoor user samples thenetwork reaches the perfectionwith error-free user detectionIt shows that the performance of the indoor sample isbetter than the outdoor sample In the indoor situation thevariation of the signal strength is limited to a certain boundThe effects of indoor free-space loss refraction diffractionreflection aperture-medium coupling loss and absorption

8 The Scientific World Journal

100

Gra

dien

t

10minus10

10minus5

Val f

ail

1

05

00 5 10 15 20 25 30 35

Gradient = 90011e minus 007 at epoch 38

Validation checks = 0 at epoch 38

38 epochs

(a)

Best validation performance is 76703e minus 008 at epoch 38

100

10minus2

10minus6

10minus4

10minus8

Mea

n sq

uare

d er

ror (

mse

)

0 5 10 15 20 25 30 35

38 Epochs

TrainValidation

TestBest

(b)

100

80

60

40

20

0

Erro

rs (

)

24

68

10

Training indoor users

5

10

15

Training outdoor users

(c)

Figure 6 (a) Training state and (b) performance of best validation (c) Performance of femtocell for different number of samples

are comparatively smaller which allows the system to verifyany random users signal strength within a certain variationof received power strength Nevertheless the number of thesample users always depends on the geographical shape of thehousesThe system requires higher number of indoor sampleswhen the variation bounds overlap with the outdoor usersvariation bound Such a case is studied below

The proposed method is now tested in a more complexscenario A ldquoUrdquo shaped house layout is designed to test theperformance of the system In this layout indoor wall isignored Figures 7(a) and 7(b) show the training and testingprocess of the femtocell network The challenging shape of

the house makes the user pattern more improvised thanthe previous one In this case the system requires highernumber of indoor and outdoor training user samples to reachan error-free performance Figure 7(c) shows the requirednumber of indoor and outdoor users against the percentageof error occurrences in detecting the usersrsquo category Herethe required number of users for both categories is above 25users The rest of the performance analysis of the process isdone using the previous layout of the house

32 Femtocell Network Performance with ANN and APClustering Algorithm AP algorithm clusters the users into

The Scientific World Journal 9

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

010

2030

010

2030

0

20

40

60

80

100

Training outdoor usersTraining indoor users

Err

ors (

)

1020

30

1020

30 utdoor useining indoor us

(c)

Figure 7 (a) Training and (b) testing process of the femtocell with ldquoUrdquo shaped house (c) Performance of femtocell for different number ofsamples

Indoor user

Outdoor user

0

minus200

minus400

minus600

minus800

minus1200

minus1000

minus1400

minus1600

minus1800

Fitn

ess (

net s

imila

rity)

of q

uant

ized

inte

rmed

iate

solu

tion

0 5 10 15 20 25 30

Number of iterations

Figure 8 Fitness of AP clustering algorithm for indoor and outdoorusers

subgroups based on their power pattern and selects arepresentative from each subgroup Unlike other clustering

methods AP algorithm selects the clusterssubgroups basedon the samples nature If the nature of the sample variesimmensely the number of clusters gets higherThe clusteringperformance of the AP algorithm is presented in Figure 8 as aform of achieved fitness (net similarities) with respect to theiteration number Both the outdoor and indoor users reachtheir best fitness before 8 iterations However a safe marginof 25 is kept to ensure the best fitness for both types of users

Figure 9(a) shows the general ANN training processDuring the training the ANN adjusts the values of theweights and the biases of the network In Figure 9(b) theAP algorithm clusters the users based on their similaritiespower pattern A representative has been chosen among thedata points of a subgroup which has most of similaritieswith the other data points of the subgroup There might alsoexist subgroups with only one data point Figure shows thatinstead of training ANN with 15 outdoor users and 10 indoorusers the AP selects 3 outdoor users and 3 indoor usersFigures 9(c) and 9(d) show the performance of the networkwith and without AP algorithm For a random simulationboth processes show the same accuracy

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World Journal 5

After the training process of the network the femtocell takesthe 6-element antennas received power as input and gives thecategory of the users in the output

22 Affinity Propagation Algorithm for Selecting the Best Sam-ples AP algorithm is a recent clustering algorithm proposedby Frey and Dueck [28] It is widely accepted because ofits high quality set of cluster samples The proposed userclassification in neural network is a supervised techniqueTheperformance of the network is subjected to the nature andquantity of the training samples Higher number of trainingsamples led to precise values of the neuronsrsquo weight butit makes the training process slower Clustering of data setbased on similarities is a vital step in data analysis problemA common practice in both supervised and unsupervisedlearning is to cluster the data based on the similarities [29 30]Affinity propagation (AP) is the latest clustering algorithmthat reduces the redundancy of the training data set Itaccelerates the computing process of ANN by reducing thesample numbers

Traditional clustering algorithms follow random selec-tion of initial data subset as exemplars and refine it iterativelyAP takes an input set of pairwise similarities between thedata points and finds the clusters based on maximum totalsimilarities between the exemplars and the data points [31]The real messages are exchanged between the data pointsuntil the finest set of exemplars and corresponding clustersprogressively emerges It has a better clustering performancethan K-means K-medians fuzzy c-means Hill combining(HC) and self-organizing map (SOM) algorithms [32 33]It is computationally efficient and simple to implement andcustomize In AP algorithm all the sample data points areconsidered a possible candidate to be the desired exemplarsEach step exchanges real-valued messages between themuntil a superior set of exemplar shows up Messages areupdated based on simple formulae that reflect on the sum-product or max-product It updates the rules until themagnitude of the messages reflects on the current affinityfor choosing another data point as its exemplar Each datapoint is considered a note in the network The process of thealgorithm is described briefly below

Input is a set of pairwise similarities as

119904 (119894 119896) = minus1003817100381710038171003817119909119894 minus 119909

119896

1003817100381710038171003817

2

119894 = 119896 (squared Euclidean distance)

where (119894 119896) isin 1 1198732

119904 (119894 119896) isin R

(8)

Here 119904(119894 119896) isin R indicates how well suited the data point 119896 isas an exemplar for data point i

For each data point 119896 a real number 119904(119896 119896) represents thepreference that is to be considered as an exemplar

119904 (119896 119896) = 120588 forall119896 isin 1 119873 (9)

Initialization set availabilities to zero for all 119894 119896

119886(119894 119896) = 0

Table 1 System parameters

System parameters ValuerangeFrequency 253GHzNumber of training indoor users 10Number of training outdoor users 15Number of randomly placed users (after training) 20Femtocell antenna height 1mUser equipment height 1mFrequency 26GHzUE transmit power (fixed) 13 dBmIndoor wall loss 5 dBOutdoor wall loss 10 dBShadow fading std 6 dBWhite noise power density minus174 dBmHzNumber of neurons in hidden layer 10

Repeat responsibility and availability updates until con-vergence

forall119894 119896 119903 (119894 119896) = 119904 (119894 119896) minusmax [119904 (119894 1198961015840) + 119886 (119894 1198961015840

)]

forall119894 119896 119886 (119894 119896)

=

summax [0 119903 (1198941015840 119896)] for 119896 = 119894

min [0 119903 (119896 119896)]

+sum11989410158401198941015840notin119894119896

max [0 119903 (1198941015840 119896)] for 119896 = 119894

(10)

Output is assignments 119888 = (1198881 119888

119873) where 119888

119894=

arg max119896[119886(119894 119896) + 119903(119894 119896)] Here 119888

119894indexes the clusterrsquos exem-

plar at which point 119894 is assigned If point 119894 is a cluster withpoint 119896 as the exemplar then 119888

119894= 119896 and 119888

119896= 119896 [34]

3 Results and Discussions

A layout of functioning area is modelled with a femto-cell in the middle of the house Six-microstrip antennasare operating with 60∘ separation angle on the same axisinside the femtocell device A previously designed microstripantenna is used here to configure the directive gain patternof each antenna element [17] The house has indoor andoutdoor walls that decrease the strength of the signal basedon their thickness Initially random indoor and outdoor usersare generated and the received powers are measured Thedimension of the house is set to 7m times 6m In Figure 4(a)the users and the house are plotted in a 20m times 20m windowThe radiation pattern of the microstrip antenna is shown inFigure 4(b)

To demonstrate the performance of the technique ANNis initially trained without using AP clustering Randomsamples are generated by varying the numbers of indoorand outdoor users In the performance analysis stage againrandom samples are generated to categorize users using theprevious experiences The system parameters that have beenused in the simulation are given in Table 1 In the model theoutdoor wall loss is considered higher than the indoor wall

6 The Scientific World Journal

00

2

4

6

8

10

12

14

16

18

20

2 4 6 8 10 12 14 16 18 20

FemtocellIndoor userOutdoor user

WallIndoor connectingOutdoor connecting line

(a)

Farfield gain Abs (120593 = 0)

Farfield (f = 253) [1]Farfield (f = 266) [1]

Frequency = 266

Thetadeg versus (dB)

Main lobe magnitude = 65dBMain lobe direction = minus50deg

Side lobe level = minus103 dB

0

30

60

90

120

150

180

minus30

minus60

minus90

minus120

minus150

minus15minus40 10

Frequency = 253

Main lobe magnitude = 61dB

Main lobe direction = minus40degAngular width (3dB) = 919deg

Angular width (3dB) = 821deg

(b)

Figure 4 (a) Layout of the simulation environment (b) Radiation pattern of the microstrip antenna at 253GHz and 266GHz

One of the reasons is that usually the outdoorwalls are thickerthan the indoor wall with more concrete and steel materialsfor the foundation or shape This increases the loss exponentof the outdoor walls Another reason is that outdoor walls aremore subjected to rust and moist from the environment thatweakens the incoming signal [35]

31 Femtocell Network Performance with ANN Figure 5(a)shows the training stage of the femtocell device The reddots are the outdoor users and green dots are the indoorusers In Figure 5(b) random users are generated for the

femtocell to classify the indoor and outdoor users by usingthe learning experience Femtocell only allows connectionto the indoor users to be connected The green connectinglines between the femtocell and the indoor users confirm theproper recognition of the users

Figures 6(a) and 6(b) show the training state and perfor-mance validation state for a simulation with 10 indoor and15 outdoor training samples The minimum gradient of theANN is set to 1 times 10minus6 In this particular iteration ANNtakes 38 epochs to train up and adjust the values of biasesand weights to achieve the minimum gradient value Theldquovalidation graphrdquo shows a downward curve It confirms that

The Scientific World Journal 7

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 2 4 6 8 10 12 14 16 18 200

2

4

6

8

10

12

14

16

18

20

(b)

Figure 5 (a) Training and (b) testing of the femtocell device

Table 2

Number of samples ANN training performance AP + ANN training performance Performance comparison after AP

Indoor Outdoor Training time(sec)

Numberof

epochslowastAP clusteringtime (sec)

Total trainingtime (sec)

Number ofepochslowast

Training timedecreases ()

Number ofepochs decreases

()5 10 15165 21 02885 07647 13 4957 38096 11 15655 21 03633 08257 14 4725 33337 12 16221 22 03921 09071 14 4407 36368 13 16416 23 04172 0934 15 4310 34789 14 16443 24 04226 09504 16 4220 333310 15 16504 26 04244 09611 16 4176 384611 16 16541 26 0439 09803 17 4073 346112 17 16669 27 04461 09848 18 4092 333313 18 16709 27 04556 09951 19 4044 296214 19 16801 28 04671 09951 19 4077 321415 20 16819 28 04702 09964 19 4075 321416 21 1688 29 04704 09982 19 4086 344817 22 17322 30 04997 10017 21 4217 3018 23 17767 31 05073 10397 22 4148 290319 24 17966 32 05234 10695 22 4047 312520 25 17996 34 05484 10511 22 4159 3529lowastThe fraction values of the epochs are expressed by the nearest integer values

after every epoch the latest values of the weights and biasesvalidate the previous training samples

Figure 6(c) shows the performance of the femtocell inpercentage of error for different number of outdoor andindoor training user samples In every iteration the networkis tested using 20 random users to verify the performanceIn both types of users the error rate is quite high at thebeginning Due to lack of knowledge of the usersrsquo behaviourthe system cannot categorize the nature of the randomlycreated users For the same number of indoor users outdoorusersrsquo percentage of error rate is higher This is becauseof the unpredictable nature of wireless signal propagation

from the outdoor users end The outdoor walls their shapesand constructing materials also add more variations in theoutdoor users signal strength due to absorption losses anddiffraction loss As a result the ANN requires higher numberof outdoor users training samples for categorizing the usersHowever after 5 indoor and 10 outdoor user samples thenetwork reaches the perfectionwith error-free user detectionIt shows that the performance of the indoor sample isbetter than the outdoor sample In the indoor situation thevariation of the signal strength is limited to a certain boundThe effects of indoor free-space loss refraction diffractionreflection aperture-medium coupling loss and absorption

8 The Scientific World Journal

100

Gra

dien

t

10minus10

10minus5

Val f

ail

1

05

00 5 10 15 20 25 30 35

Gradient = 90011e minus 007 at epoch 38

Validation checks = 0 at epoch 38

38 epochs

(a)

Best validation performance is 76703e minus 008 at epoch 38

100

10minus2

10minus6

10minus4

10minus8

Mea

n sq

uare

d er

ror (

mse

)

0 5 10 15 20 25 30 35

38 Epochs

TrainValidation

TestBest

(b)

100

80

60

40

20

0

Erro

rs (

)

24

68

10

Training indoor users

5

10

15

Training outdoor users

(c)

Figure 6 (a) Training state and (b) performance of best validation (c) Performance of femtocell for different number of samples

are comparatively smaller which allows the system to verifyany random users signal strength within a certain variationof received power strength Nevertheless the number of thesample users always depends on the geographical shape of thehousesThe system requires higher number of indoor sampleswhen the variation bounds overlap with the outdoor usersvariation bound Such a case is studied below

The proposed method is now tested in a more complexscenario A ldquoUrdquo shaped house layout is designed to test theperformance of the system In this layout indoor wall isignored Figures 7(a) and 7(b) show the training and testingprocess of the femtocell network The challenging shape of

the house makes the user pattern more improvised thanthe previous one In this case the system requires highernumber of indoor and outdoor training user samples to reachan error-free performance Figure 7(c) shows the requirednumber of indoor and outdoor users against the percentageof error occurrences in detecting the usersrsquo category Herethe required number of users for both categories is above 25users The rest of the performance analysis of the process isdone using the previous layout of the house

32 Femtocell Network Performance with ANN and APClustering Algorithm AP algorithm clusters the users into

The Scientific World Journal 9

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

010

2030

010

2030

0

20

40

60

80

100

Training outdoor usersTraining indoor users

Err

ors (

)

1020

30

1020

30 utdoor useining indoor us

(c)

Figure 7 (a) Training and (b) testing process of the femtocell with ldquoUrdquo shaped house (c) Performance of femtocell for different number ofsamples

Indoor user

Outdoor user

0

minus200

minus400

minus600

minus800

minus1200

minus1000

minus1400

minus1600

minus1800

Fitn

ess (

net s

imila

rity)

of q

uant

ized

inte

rmed

iate

solu

tion

0 5 10 15 20 25 30

Number of iterations

Figure 8 Fitness of AP clustering algorithm for indoor and outdoorusers

subgroups based on their power pattern and selects arepresentative from each subgroup Unlike other clustering

methods AP algorithm selects the clusterssubgroups basedon the samples nature If the nature of the sample variesimmensely the number of clusters gets higherThe clusteringperformance of the AP algorithm is presented in Figure 8 as aform of achieved fitness (net similarities) with respect to theiteration number Both the outdoor and indoor users reachtheir best fitness before 8 iterations However a safe marginof 25 is kept to ensure the best fitness for both types of users

Figure 9(a) shows the general ANN training processDuring the training the ANN adjusts the values of theweights and the biases of the network In Figure 9(b) theAP algorithm clusters the users based on their similaritiespower pattern A representative has been chosen among thedata points of a subgroup which has most of similaritieswith the other data points of the subgroup There might alsoexist subgroups with only one data point Figure shows thatinstead of training ANN with 15 outdoor users and 10 indoorusers the AP selects 3 outdoor users and 3 indoor usersFigures 9(c) and 9(d) show the performance of the networkwith and without AP algorithm For a random simulationboth processes show the same accuracy

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

6 The Scientific World Journal

00

2

4

6

8

10

12

14

16

18

20

2 4 6 8 10 12 14 16 18 20

FemtocellIndoor userOutdoor user

WallIndoor connectingOutdoor connecting line

(a)

Farfield gain Abs (120593 = 0)

Farfield (f = 253) [1]Farfield (f = 266) [1]

Frequency = 266

Thetadeg versus (dB)

Main lobe magnitude = 65dBMain lobe direction = minus50deg

Side lobe level = minus103 dB

0

30

60

90

120

150

180

minus30

minus60

minus90

minus120

minus150

minus15minus40 10

Frequency = 253

Main lobe magnitude = 61dB

Main lobe direction = minus40degAngular width (3dB) = 919deg

Angular width (3dB) = 821deg

(b)

Figure 4 (a) Layout of the simulation environment (b) Radiation pattern of the microstrip antenna at 253GHz and 266GHz

One of the reasons is that usually the outdoorwalls are thickerthan the indoor wall with more concrete and steel materialsfor the foundation or shape This increases the loss exponentof the outdoor walls Another reason is that outdoor walls aremore subjected to rust and moist from the environment thatweakens the incoming signal [35]

31 Femtocell Network Performance with ANN Figure 5(a)shows the training stage of the femtocell device The reddots are the outdoor users and green dots are the indoorusers In Figure 5(b) random users are generated for the

femtocell to classify the indoor and outdoor users by usingthe learning experience Femtocell only allows connectionto the indoor users to be connected The green connectinglines between the femtocell and the indoor users confirm theproper recognition of the users

Figures 6(a) and 6(b) show the training state and perfor-mance validation state for a simulation with 10 indoor and15 outdoor training samples The minimum gradient of theANN is set to 1 times 10minus6 In this particular iteration ANNtakes 38 epochs to train up and adjust the values of biasesand weights to achieve the minimum gradient value Theldquovalidation graphrdquo shows a downward curve It confirms that

The Scientific World Journal 7

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 2 4 6 8 10 12 14 16 18 200

2

4

6

8

10

12

14

16

18

20

(b)

Figure 5 (a) Training and (b) testing of the femtocell device

Table 2

Number of samples ANN training performance AP + ANN training performance Performance comparison after AP

Indoor Outdoor Training time(sec)

Numberof

epochslowastAP clusteringtime (sec)

Total trainingtime (sec)

Number ofepochslowast

Training timedecreases ()

Number ofepochs decreases

()5 10 15165 21 02885 07647 13 4957 38096 11 15655 21 03633 08257 14 4725 33337 12 16221 22 03921 09071 14 4407 36368 13 16416 23 04172 0934 15 4310 34789 14 16443 24 04226 09504 16 4220 333310 15 16504 26 04244 09611 16 4176 384611 16 16541 26 0439 09803 17 4073 346112 17 16669 27 04461 09848 18 4092 333313 18 16709 27 04556 09951 19 4044 296214 19 16801 28 04671 09951 19 4077 321415 20 16819 28 04702 09964 19 4075 321416 21 1688 29 04704 09982 19 4086 344817 22 17322 30 04997 10017 21 4217 3018 23 17767 31 05073 10397 22 4148 290319 24 17966 32 05234 10695 22 4047 312520 25 17996 34 05484 10511 22 4159 3529lowastThe fraction values of the epochs are expressed by the nearest integer values

after every epoch the latest values of the weights and biasesvalidate the previous training samples

Figure 6(c) shows the performance of the femtocell inpercentage of error for different number of outdoor andindoor training user samples In every iteration the networkis tested using 20 random users to verify the performanceIn both types of users the error rate is quite high at thebeginning Due to lack of knowledge of the usersrsquo behaviourthe system cannot categorize the nature of the randomlycreated users For the same number of indoor users outdoorusersrsquo percentage of error rate is higher This is becauseof the unpredictable nature of wireless signal propagation

from the outdoor users end The outdoor walls their shapesand constructing materials also add more variations in theoutdoor users signal strength due to absorption losses anddiffraction loss As a result the ANN requires higher numberof outdoor users training samples for categorizing the usersHowever after 5 indoor and 10 outdoor user samples thenetwork reaches the perfectionwith error-free user detectionIt shows that the performance of the indoor sample isbetter than the outdoor sample In the indoor situation thevariation of the signal strength is limited to a certain boundThe effects of indoor free-space loss refraction diffractionreflection aperture-medium coupling loss and absorption

8 The Scientific World Journal

100

Gra

dien

t

10minus10

10minus5

Val f

ail

1

05

00 5 10 15 20 25 30 35

Gradient = 90011e minus 007 at epoch 38

Validation checks = 0 at epoch 38

38 epochs

(a)

Best validation performance is 76703e minus 008 at epoch 38

100

10minus2

10minus6

10minus4

10minus8

Mea

n sq

uare

d er

ror (

mse

)

0 5 10 15 20 25 30 35

38 Epochs

TrainValidation

TestBest

(b)

100

80

60

40

20

0

Erro

rs (

)

24

68

10

Training indoor users

5

10

15

Training outdoor users

(c)

Figure 6 (a) Training state and (b) performance of best validation (c) Performance of femtocell for different number of samples

are comparatively smaller which allows the system to verifyany random users signal strength within a certain variationof received power strength Nevertheless the number of thesample users always depends on the geographical shape of thehousesThe system requires higher number of indoor sampleswhen the variation bounds overlap with the outdoor usersvariation bound Such a case is studied below

The proposed method is now tested in a more complexscenario A ldquoUrdquo shaped house layout is designed to test theperformance of the system In this layout indoor wall isignored Figures 7(a) and 7(b) show the training and testingprocess of the femtocell network The challenging shape of

the house makes the user pattern more improvised thanthe previous one In this case the system requires highernumber of indoor and outdoor training user samples to reachan error-free performance Figure 7(c) shows the requirednumber of indoor and outdoor users against the percentageof error occurrences in detecting the usersrsquo category Herethe required number of users for both categories is above 25users The rest of the performance analysis of the process isdone using the previous layout of the house

32 Femtocell Network Performance with ANN and APClustering Algorithm AP algorithm clusters the users into

The Scientific World Journal 9

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

010

2030

010

2030

0

20

40

60

80

100

Training outdoor usersTraining indoor users

Err

ors (

)

1020

30

1020

30 utdoor useining indoor us

(c)

Figure 7 (a) Training and (b) testing process of the femtocell with ldquoUrdquo shaped house (c) Performance of femtocell for different number ofsamples

Indoor user

Outdoor user

0

minus200

minus400

minus600

minus800

minus1200

minus1000

minus1400

minus1600

minus1800

Fitn

ess (

net s

imila

rity)

of q

uant

ized

inte

rmed

iate

solu

tion

0 5 10 15 20 25 30

Number of iterations

Figure 8 Fitness of AP clustering algorithm for indoor and outdoorusers

subgroups based on their power pattern and selects arepresentative from each subgroup Unlike other clustering

methods AP algorithm selects the clusterssubgroups basedon the samples nature If the nature of the sample variesimmensely the number of clusters gets higherThe clusteringperformance of the AP algorithm is presented in Figure 8 as aform of achieved fitness (net similarities) with respect to theiteration number Both the outdoor and indoor users reachtheir best fitness before 8 iterations However a safe marginof 25 is kept to ensure the best fitness for both types of users

Figure 9(a) shows the general ANN training processDuring the training the ANN adjusts the values of theweights and the biases of the network In Figure 9(b) theAP algorithm clusters the users based on their similaritiespower pattern A representative has been chosen among thedata points of a subgroup which has most of similaritieswith the other data points of the subgroup There might alsoexist subgroups with only one data point Figure shows thatinstead of training ANN with 15 outdoor users and 10 indoorusers the AP selects 3 outdoor users and 3 indoor usersFigures 9(c) and 9(d) show the performance of the networkwith and without AP algorithm For a random simulationboth processes show the same accuracy

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World Journal 7

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 2 4 6 8 10 12 14 16 18 200

2

4

6

8

10

12

14

16

18

20

(b)

Figure 5 (a) Training and (b) testing of the femtocell device

Table 2

Number of samples ANN training performance AP + ANN training performance Performance comparison after AP

Indoor Outdoor Training time(sec)

Numberof

epochslowastAP clusteringtime (sec)

Total trainingtime (sec)

Number ofepochslowast

Training timedecreases ()

Number ofepochs decreases

()5 10 15165 21 02885 07647 13 4957 38096 11 15655 21 03633 08257 14 4725 33337 12 16221 22 03921 09071 14 4407 36368 13 16416 23 04172 0934 15 4310 34789 14 16443 24 04226 09504 16 4220 333310 15 16504 26 04244 09611 16 4176 384611 16 16541 26 0439 09803 17 4073 346112 17 16669 27 04461 09848 18 4092 333313 18 16709 27 04556 09951 19 4044 296214 19 16801 28 04671 09951 19 4077 321415 20 16819 28 04702 09964 19 4075 321416 21 1688 29 04704 09982 19 4086 344817 22 17322 30 04997 10017 21 4217 3018 23 17767 31 05073 10397 22 4148 290319 24 17966 32 05234 10695 22 4047 312520 25 17996 34 05484 10511 22 4159 3529lowastThe fraction values of the epochs are expressed by the nearest integer values

after every epoch the latest values of the weights and biasesvalidate the previous training samples

Figure 6(c) shows the performance of the femtocell inpercentage of error for different number of outdoor andindoor training user samples In every iteration the networkis tested using 20 random users to verify the performanceIn both types of users the error rate is quite high at thebeginning Due to lack of knowledge of the usersrsquo behaviourthe system cannot categorize the nature of the randomlycreated users For the same number of indoor users outdoorusersrsquo percentage of error rate is higher This is becauseof the unpredictable nature of wireless signal propagation

from the outdoor users end The outdoor walls their shapesand constructing materials also add more variations in theoutdoor users signal strength due to absorption losses anddiffraction loss As a result the ANN requires higher numberof outdoor users training samples for categorizing the usersHowever after 5 indoor and 10 outdoor user samples thenetwork reaches the perfectionwith error-free user detectionIt shows that the performance of the indoor sample isbetter than the outdoor sample In the indoor situation thevariation of the signal strength is limited to a certain boundThe effects of indoor free-space loss refraction diffractionreflection aperture-medium coupling loss and absorption

8 The Scientific World Journal

100

Gra

dien

t

10minus10

10minus5

Val f

ail

1

05

00 5 10 15 20 25 30 35

Gradient = 90011e minus 007 at epoch 38

Validation checks = 0 at epoch 38

38 epochs

(a)

Best validation performance is 76703e minus 008 at epoch 38

100

10minus2

10minus6

10minus4

10minus8

Mea

n sq

uare

d er

ror (

mse

)

0 5 10 15 20 25 30 35

38 Epochs

TrainValidation

TestBest

(b)

100

80

60

40

20

0

Erro

rs (

)

24

68

10

Training indoor users

5

10

15

Training outdoor users

(c)

Figure 6 (a) Training state and (b) performance of best validation (c) Performance of femtocell for different number of samples

are comparatively smaller which allows the system to verifyany random users signal strength within a certain variationof received power strength Nevertheless the number of thesample users always depends on the geographical shape of thehousesThe system requires higher number of indoor sampleswhen the variation bounds overlap with the outdoor usersvariation bound Such a case is studied below

The proposed method is now tested in a more complexscenario A ldquoUrdquo shaped house layout is designed to test theperformance of the system In this layout indoor wall isignored Figures 7(a) and 7(b) show the training and testingprocess of the femtocell network The challenging shape of

the house makes the user pattern more improvised thanthe previous one In this case the system requires highernumber of indoor and outdoor training user samples to reachan error-free performance Figure 7(c) shows the requirednumber of indoor and outdoor users against the percentageof error occurrences in detecting the usersrsquo category Herethe required number of users for both categories is above 25users The rest of the performance analysis of the process isdone using the previous layout of the house

32 Femtocell Network Performance with ANN and APClustering Algorithm AP algorithm clusters the users into

The Scientific World Journal 9

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

010

2030

010

2030

0

20

40

60

80

100

Training outdoor usersTraining indoor users

Err

ors (

)

1020

30

1020

30 utdoor useining indoor us

(c)

Figure 7 (a) Training and (b) testing process of the femtocell with ldquoUrdquo shaped house (c) Performance of femtocell for different number ofsamples

Indoor user

Outdoor user

0

minus200

minus400

minus600

minus800

minus1200

minus1000

minus1400

minus1600

minus1800

Fitn

ess (

net s

imila

rity)

of q

uant

ized

inte

rmed

iate

solu

tion

0 5 10 15 20 25 30

Number of iterations

Figure 8 Fitness of AP clustering algorithm for indoor and outdoorusers

subgroups based on their power pattern and selects arepresentative from each subgroup Unlike other clustering

methods AP algorithm selects the clusterssubgroups basedon the samples nature If the nature of the sample variesimmensely the number of clusters gets higherThe clusteringperformance of the AP algorithm is presented in Figure 8 as aform of achieved fitness (net similarities) with respect to theiteration number Both the outdoor and indoor users reachtheir best fitness before 8 iterations However a safe marginof 25 is kept to ensure the best fitness for both types of users

Figure 9(a) shows the general ANN training processDuring the training the ANN adjusts the values of theweights and the biases of the network In Figure 9(b) theAP algorithm clusters the users based on their similaritiespower pattern A representative has been chosen among thedata points of a subgroup which has most of similaritieswith the other data points of the subgroup There might alsoexist subgroups with only one data point Figure shows thatinstead of training ANN with 15 outdoor users and 10 indoorusers the AP selects 3 outdoor users and 3 indoor usersFigures 9(c) and 9(d) show the performance of the networkwith and without AP algorithm For a random simulationboth processes show the same accuracy

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

8 The Scientific World Journal

100

Gra

dien

t

10minus10

10minus5

Val f

ail

1

05

00 5 10 15 20 25 30 35

Gradient = 90011e minus 007 at epoch 38

Validation checks = 0 at epoch 38

38 epochs

(a)

Best validation performance is 76703e minus 008 at epoch 38

100

10minus2

10minus6

10minus4

10minus8

Mea

n sq

uare

d er

ror (

mse

)

0 5 10 15 20 25 30 35

38 Epochs

TrainValidation

TestBest

(b)

100

80

60

40

20

0

Erro

rs (

)

24

68

10

Training indoor users

5

10

15

Training outdoor users

(c)

Figure 6 (a) Training state and (b) performance of best validation (c) Performance of femtocell for different number of samples

are comparatively smaller which allows the system to verifyany random users signal strength within a certain variationof received power strength Nevertheless the number of thesample users always depends on the geographical shape of thehousesThe system requires higher number of indoor sampleswhen the variation bounds overlap with the outdoor usersvariation bound Such a case is studied below

The proposed method is now tested in a more complexscenario A ldquoUrdquo shaped house layout is designed to test theperformance of the system In this layout indoor wall isignored Figures 7(a) and 7(b) show the training and testingprocess of the femtocell network The challenging shape of

the house makes the user pattern more improvised thanthe previous one In this case the system requires highernumber of indoor and outdoor training user samples to reachan error-free performance Figure 7(c) shows the requirednumber of indoor and outdoor users against the percentageof error occurrences in detecting the usersrsquo category Herethe required number of users for both categories is above 25users The rest of the performance analysis of the process isdone using the previous layout of the house

32 Femtocell Network Performance with ANN and APClustering Algorithm AP algorithm clusters the users into

The Scientific World Journal 9

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

010

2030

010

2030

0

20

40

60

80

100

Training outdoor usersTraining indoor users

Err

ors (

)

1020

30

1020

30 utdoor useining indoor us

(c)

Figure 7 (a) Training and (b) testing process of the femtocell with ldquoUrdquo shaped house (c) Performance of femtocell for different number ofsamples

Indoor user

Outdoor user

0

minus200

minus400

minus600

minus800

minus1200

minus1000

minus1400

minus1600

minus1800

Fitn

ess (

net s

imila

rity)

of q

uant

ized

inte

rmed

iate

solu

tion

0 5 10 15 20 25 30

Number of iterations

Figure 8 Fitness of AP clustering algorithm for indoor and outdoorusers

subgroups based on their power pattern and selects arepresentative from each subgroup Unlike other clustering

methods AP algorithm selects the clusterssubgroups basedon the samples nature If the nature of the sample variesimmensely the number of clusters gets higherThe clusteringperformance of the AP algorithm is presented in Figure 8 as aform of achieved fitness (net similarities) with respect to theiteration number Both the outdoor and indoor users reachtheir best fitness before 8 iterations However a safe marginof 25 is kept to ensure the best fitness for both types of users

Figure 9(a) shows the general ANN training processDuring the training the ANN adjusts the values of theweights and the biases of the network In Figure 9(b) theAP algorithm clusters the users based on their similaritiespower pattern A representative has been chosen among thedata points of a subgroup which has most of similaritieswith the other data points of the subgroup There might alsoexist subgroups with only one data point Figure shows thatinstead of training ANN with 15 outdoor users and 10 indoorusers the AP selects 3 outdoor users and 3 indoor usersFigures 9(c) and 9(d) show the performance of the networkwith and without AP algorithm For a random simulationboth processes show the same accuracy

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World Journal 9

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

010

2030

010

2030

0

20

40

60

80

100

Training outdoor usersTraining indoor users

Err

ors (

)

1020

30

1020

30 utdoor useining indoor us

(c)

Figure 7 (a) Training and (b) testing process of the femtocell with ldquoUrdquo shaped house (c) Performance of femtocell for different number ofsamples

Indoor user

Outdoor user

0

minus200

minus400

minus600

minus800

minus1200

minus1000

minus1400

minus1600

minus1800

Fitn

ess (

net s

imila

rity)

of q

uant

ized

inte

rmed

iate

solu

tion

0 5 10 15 20 25 30

Number of iterations

Figure 8 Fitness of AP clustering algorithm for indoor and outdoorusers

subgroups based on their power pattern and selects arepresentative from each subgroup Unlike other clustering

methods AP algorithm selects the clusterssubgroups basedon the samples nature If the nature of the sample variesimmensely the number of clusters gets higherThe clusteringperformance of the AP algorithm is presented in Figure 8 as aform of achieved fitness (net similarities) with respect to theiteration number Both the outdoor and indoor users reachtheir best fitness before 8 iterations However a safe marginof 25 is kept to ensure the best fitness for both types of users

Figure 9(a) shows the general ANN training processDuring the training the ANN adjusts the values of theweights and the biases of the network In Figure 9(b) theAP algorithm clusters the users based on their similaritiespower pattern A representative has been chosen among thedata points of a subgroup which has most of similaritieswith the other data points of the subgroup There might alsoexist subgroups with only one data point Figure shows thatinstead of training ANN with 15 outdoor users and 10 indoorusers the AP selects 3 outdoor users and 3 indoor usersFigures 9(c) and 9(d) show the performance of the networkwith and without AP algorithm For a random simulationboth processes show the same accuracy

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

10 The Scientific World Journal

0 5 10 15 2002468

101214161820

(a)

0 2 4 6 8 10 12 14 16 18 2002468

101214161820

Representativeoutdoor sample

Representative indoor sample

(b)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 9 (a) Training with ANN (b) Training with ANN+AP (c) Performance of the network with ANN training (d) Performance of thenetwork with ANN+AP training

Results show that training the ANN in corporation withAP clustering requires less number of training samples Theprocess takes less number of epochs to reach the gradientrsquosthreshold value For the above simulation the ANN took 25epochs while it took 12 epochs using AP clustered samplesThe representative of the data points helps the ANN toexplore all the possible variations of the characters of theusersrsquo power pattern and guide the network to balance thevalues of weights and the biases with a faster time intervalFigures 10(a) and 10(b) show the mean square rate (MSE) ofthe training process Due to higher number of sample datapoints the accuracy of the regular ANN training is moreprecise However in the training process with clustered datasamples the mean square error decreases drastically and getsto the desired value with less number of epochs In Figures10(c) and 10(d) the validation check shows a good fitnesssince the number of indoor and outdoor users is chosen fromthe error-free region achieved in the result in Figure 6(c)

The performance analysis of both processes is shown inTable 2 Randomly 20 users have been generated every timeto test the performance of the network Each resultant data is

an average value of 1000 simulations The AP+ANN trainingprocess takes around 75ndash85 less time than the regularANN training process meanwhile AP clustering processtakes some additional time which makes the total AP+ANNtime around 50ndash60 less than ANN regular training timeAfter AP algorithm implementation the number of epochsalso decreases down to 40The fraction values of the epochsin Table 2 are expressed by the nearest integer value

33 AP Clustering Algorithm versus K-Means Clustering Algo-rithm and Fuzzy c-Means Clustering To justify the selectionof AP clustering algorithm over the traditional clusteringalgorithm two popular algorithms K-means and fuzzy c-means clustering are compared with AP clustering in theANN training process

K-Means K-means is one of the simplest unsupervisedlearning algorithms that solves the well-known clusteringproblems It partitions the data set into 119896 mutually exclusiveclusters and returns the index of the cluster to which it

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World Journal 11

10minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

Best validation performance is 26883e minus 009 at epoch 25

0 5 10 15 20 25

25 epochs

TrainValidation

TestBest

(a)

0 2 4 6 8 10 1210minus8

10minus2

10minus4

10minus6

100

Mea

n sq

uare

d er

ror (

mse

)

12 epochs

TrainValidation

TestBest

Best validation performance is 2489e minus 007 at epoch 12

(b)

10minus10

10minus5

100

Gra

dien

t

0

0

5

1

10 15 20 25minus1

minus05

05

Val f

ail

25 epochs

Gradient = 55626e minus 007 at epoch 25

Validation checks = 0 at epoch 25

(c)

10minus10

10minus5

100G

radi

ent

0

1

minus1

minus05

05

Val f

ail

0 2 64 8 10 1212 epochs

Validation checks = 0 at epoch 12

Gradient = 62198e minus 007 at epoch 12

(d)

Figure 10 (a) Best validation performance of ANN (b) Best validation performance of ANN+AP (c) Training state of ANN (d) Trainingstate of ANN+AP femtocell network

has assigned each observation Unlike AP K-means createsa single level of clusters and needs the number of clustersassigned before the execution The algorithm breaks the dataset into 119896 different clusters If it is unable to find 119896 clustersit breaks the data set into 119896 minus 1 clusters Initially it takes 119896number of random observation data set which is consideredthe seeds of the algorithm Then it assigns all the otherobservations to 119896 seeds based on their proximity to the seedsIn general sense the algorithm takes a set of objects 119878 andan integer 119896 and gives a partition of 119878 into subsets 119878

1 119878

119896

defined by 119896 cluster centroid locations or centres [36]

Fuzzy c-Means The central idea in fuzzy clustering isthe nonunique partitioning of the data in a collection ofclusters LikeK-means fuzzy c-means creates a single level ofclusters and needs the number of clusters assigned before the

execution Cluster centres are randomly initialized and datapoint (119909

119894) assigned into clusters (119862

119895 119895 = 1 to 119896) Distance

metric (Euclidean distance are widely used) calculate how faraway a point is from a cluster centre When all data pointshave been assigned to clusters new cluster centres (centroids)are calculated The process of calculating cluster member-ships and recalculating cluster centres continues until thecluster centres no longer change from one cycle to the next[37 38]

Figures 11(b) 11(c) and 11(d) illustrate the representativeselection process of AP K-means and fuzzy c-means cluster-ing algorithm in the functioning area The green dots showthe indoor representative points of the data set while the reddots represent the outdoor In both K-means and fuzzy c-means the centroid points are not user data sample it is apoint of each cluster that has a minimum value distance from

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

12 The Scientific World Journal

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(a)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(b)

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

(c)

0 5 10 15 200

2

4

6

8

10

12

14

16

18

20

(d)

Figure 11 (a) Position of the sample indoor and outdoor users (b) AP clustering of indoor and outdoor users based on signal strength (c)K-means clustering with 9 clusters (d) Fuzzy c-means clustering with 9 clusters

each of themembers of the clusters In the case ofK-means itjust executes the distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting To obtain theerror-free performance in the ANN K-means and fuzzy c-means require different number of clusters each time A littlecomparison of the performance is shown in Table 3

K-means minimizes the sum of distances from each datapoints to its cluster centroid The process repeats until thesum of distances cannot be decreased further This processtakes more time than AP On the other hand K-meansneeds to do a distance calculation whereas fuzzy c-meansneeds to do a full inverse-distance weighting Fuzzy c-meansthus performs slower than both clustering algorithms inthis particular case However for higher number of datasamples the time increment is a little less than the APclustering algorithm Although the overall clustering time ofAP algorithm is always less by a fair distance the number ofclusters has to be determined maintaining the same accuracyof the ANNoutput Except AP algorithm the challenge in theother clustering processes mostly lies in selecting the number

of clusters to perform an error-free training On this note APalgorithm is the best candidate in this process as it selects thenumber of clusters by itself analysing the samples in everysimulation

4 Conclusion

This paper proposed a novel technique to classify the usersin closed access femtocell network by using ANN and APclustering algorithm The technique is developed using amultielement antenna femtocell device The power patternof each user is used to distinguish different level of usersA machine learning process is adopted by using ANN toinaugurate the user recognition feature in the femtocellAfter using a certain number of user samples the femtocellsuccessfully recognizes the indoor and outdoor users In thelater part AP clustering algorithm is included along withANN to speed up the training process Performance analysisshows that the femtocell takes less time to recognize user

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World Journal 13

Table 3

Number of samples ANN + AP performance ANN + 119870-means performance ANN + fuzzy 119888-means clustering

Indoor OutdoorNumber of samples

for error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)

Number of samplesfor error-freeoperationlowast

Clustering +training time

(sec)5 10 6 07647 8 12516 8 1371210 15 6 09611 9 13354 8 1415715 20 7 09964 9 13847 9 1486920 25 7 10511 9 14964 9 15738lowastThe fraction values of the epochs are expressed by the nearest integer values

without compromising the accuracy Finally a comparisonof AP clustering K-means clustering and fuzzy c-meansis showed in the user classification process to justify theselection of AP clustering methodThe result shows for samesimulation that both K-means and fuzzy c-means consumemore time and give less efficiency

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

The authors would like to thank the Ministry of HigherEducation Research Grant FRGS12014TK03UKM011 forsponsoring this work

References

[1] V Chandrasekhar J G Andrews and A Gatherer ldquoFemtocellnetworks a surveyrdquo IEEE Communications Magazine vol 46no 9 pp 59ndash67 2008

[2] A Rath S Hua and S S Panwar ldquoFemtoHaul using femtocellswith relays to increase macrocell backhaul bandwidthrdquo in Pro-ceedings of the IEEE Conference on Computer CommunicationsWorkshops (INFOCOM rsquo10) March 2010

[3] D Lopez-Perez A Valcarce G de La Roche and J ZhangldquoOFDMA femtocells a roadmap on interference avoidancerdquoIEEE Communications Magazine vol 47 no 9 pp 41ndash48 2009

[4] H Widiarti S Pyun and D Cho ldquoInterference mitigationbased on femtocells grouping in low duty operationrdquo in Pro-ceedings of the IEEE 72nd Vehicular Technology Conference Fall(VTC-Fall rsquo10) pp 1ndash5 September 2010

[5] A U Ahmed M T Islam M Ismail and M GhanbarisabaghldquoDynamic resource allocation in hybrid access femtocell net-workrdquoThe ScientificWorld Journal vol 2014 Article ID 539720p 7 2014

[6] G de La Roche A Valcarce D Lopez-Perez and J ZhangldquoAccess control mechanisms for femtocellsrdquo IEEE Communica-tions Magazine vol 48 no 1 pp 33ndash39 2010

[7] H A Mahmoud I Guvenc and F Watanabe ldquoPerformanceof open access femtocell networks with different cell-selectionmethodsrdquo in Proceedings of the IEEE 71st Vehicular TechnologyConference (VTC rsquo10-Spring) pp 1ndash5 Taipei Taiwan May 2010

[8] P Xia V Chandrasekhar and J G Andrews ldquoOpen vs closedaccess femtocells in the uplinkrdquo IEEE Transactions on WirelessCommunications vol 9 no 12 pp 3798ndash3809 2010

[9] W Zheng H Zhang X Chu and X Wen ldquoMobility robust-ness optimization in self-organizing LTE femtocell networksrdquoEURASIP Journal onWireless Communications and Networkingvol 2013 article 27 no 1 2013

[10] H Claussen and F Pivit ldquoFemtocell coverage optimizationusing switched multi-element antennasrdquo in Proceedings of theIEEE International Conference on Communications (ICC 09)pp 1ndash6 Dresden Germany June 2009

[11] A Cabedo J Anguera C Picher M Ribo and C PuenteldquoMultiband handset antenna combining a PIFA slots andground plane modesrdquo IEEE Transactions on Antennas andPropagation vol 57 no 9 pp 2526ndash2533 2009

[12] A-H Tsai L-C Wang J-H Huang and R-B Hwang ldquoHigh-capacity OFDMA femtocells by directional antennas and loca-tion awarenessrdquo IEEE Systems Journal vol 6 no 2 pp 329ndash3402012

[13] S Al-Rubaye A Al-Dulaimi and J Cosmas ldquoCognitive fem-tocellrdquo IEEE Vehicular Technology Magazine vol 6 no 1 pp44ndash51 2011

[14] M Agatonovic Z Stankovic N Doncova L Sit B Milo-vanovic and T Zwick ldquoApplication of artificial neural networksfor efficient high-resolution 2D DOA estimationrdquo Radioengi-neering vol 21 p 1179 2012

[15] D Inserra and A M Tonello ldquoA multiple antenna wirelesstestbed for the validation of DoA estimation algorithmsrdquoAEUmdashInternational Journal of Electronics andCommunicationsvol 68 no 1 pp 10ndash18 2014

[16] T S G Basha M N G Prasad and P V Sridevi ldquoHybrid tech-nique for beam forming in smart antenna with spatial diversityrdquoInternational Journal of Wireless and Mobile Computing vol 5no 2 pp 126ndash136 2012

[17] AUAhmedM T Islam R AzimM Ismail andM FMansorldquoMicrostrip antenna design for femtocell coverage optimiza-tionrdquo International Journal of Antennas and Propagation vol2014 Article ID 480140 8 pages 2014

[18] S Promwong and J-I Takada ldquoFree space link budget estima-tion scheme for ultra wideband impulse radio with imperfectantennasrdquo IEICE Electronic Express vol 1 pp 188ndash192 2004

[19] A U Ahmed M T Islam and M Ismail ldquoA review onfemtocell and its diverse interference mitigation techniquesin heterogeneous networkrdquoWireless Personal Communicationspp 1ndash22 2014

[20] D F Specht ldquoProbabilistic neural networksrdquo Neural Networksvol 3 no 1 pp 109ndash118 1990

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

14 The Scientific World Journal

[21] J Wang P Urriza Y Han and D Cabric ldquoWeighted cen-troid localization algorithm theoretical analysis and distributedimplementationrdquo IEEE Transactions on Wireless Communica-tions vol 10 no 10 pp 3403ndash3413 2011

[22] J Benedicto S Dinwiddy G Gatti R Lucas and M LugertGALILEO Satellite System Design European Space Agency2000

[23] WG Griswold R Boyer SW Brown et alActiveCampus Sus-taining Educational Communities through Mobile TechnologyDepartment of Computer Science and Engineering Universityof California San Diego Calif USA 2002

[24] TOgawa S Yoshino andM Shimizu ldquoLocation determinationmethod for wireless systems based on learning vector quantiza-tionrdquo NTT Technical Review vol 1 no 9 pp 27ndash36 2003

[25] P S Roy and S Chakraborty ldquoDesign of C-slotted microstripantenna using artificial neural network modelrdquo InternationalJournal for Research in Science amp Advanced Technologies vol 22012

[26] L Fausett Fundamentals of Neural Networks ArchitecturesAlgorithms and Applications Prentice-Hall New York NYUSA 1994

[27] M T Hagan H B Demuth and M H Beale Neural NetworkDesign Pws Boston Mass USA 1996

[28] B J Frey and D Dueck ldquoClustering by passing messagesbetween data pointsrdquo Science vol 315 no 5814 pp 972ndash9762007

[29] H Liu H Darabi P Banerjee and J Liu ldquoSurvey of wirelessindoor positioning techniques and systemsrdquo IEEE Transactionson Systems Man and Cybernetics C Applications and Reviewsvol 37 no 6 pp 1067ndash1080 2007

[30] J Zhao Y Zhang and M Ye ldquoResearch on the received signalstrength indication location algorithm for RFID systemrdquo inProceedings of the International Symposium on Communica-tions and Information Technologies (ISCIT rsquo06) pp 881ndash885Bangkok Thailand October 2006

[31] W-S Lai M-E Chiang S-C Lee and T-S Lee ldquoGame theo-retic distributed dynamic resource allocation with interferenceavoidance in cognitive femtocell networksrdquo in Proceedings ofthe IEEE Wireless Communications and Networking Conference(WCNC 13) pp 3364ndash3369 Shanghai China April 2013

[32] D Dembele and P Kastner ldquoFuzzy C-means method forclustering microarray datardquo Bioinformatics vol 19 no 8 pp973ndash980 2003

[33] T Kohonen E Oja O Simula A Visa and J Kangas ldquoEngi-neering applications of the self-organizing maprdquo Proceedings ofthe IEEE vol 84 no 10 pp 1358ndash1384 1996

[34] J Meinila P Kyosti T Jamsa and L Hentila ldquoWINNER IIchannel modelsrdquo in Radio Technologies and Concepts for IMT-Advanced pp 39ndash92 2009

[35] Y Miura Y Oda and T Taga ldquoOutdoor-to-indoor propagationmodelling with the identification of path passing through wallopeningsrdquo in Proceedings of the 13th IEEE International Sym-posium on Personal Indoor and Mobile Radio Communications(PIMRC rsquo02) pp 130ndash134 September 2002

[36] A Ghosh R Ratasuk W Xiao et al ldquoUplink control channeldesign for 3GPP LTErdquo in Proceedings of the 18th AnnualIEEE International Symposium on Personal Indoor and MobileRadio Communications (PIMRC rsquo07) pp 1ndash5 Athens Ga USASeptember 2007

[37] N R Pal K Pal J M Keller and J C Bezdek ldquoA possibilisticfuzzy c-means clustering algorithmrdquo IEEE Transactions onFuzzy Systems vol 13 no 4 pp 517ndash530 2005

[38] R L Cannon J V Dave and J C Bezdek ldquoEfficient imple-mentation of the fuzzy c-means clustering algorithmsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol8 no 2 pp 248ndash255 1986

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Distributed Sensor Networks

International Journal of

Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014

International Journal of

ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

ArtificialNeural Systems

Advances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Computational Intelligence and Neuroscience

Industrial EngineeringJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014