Artificial Neural Networks for Cognitive Radio Network: A Survey

Transcription

1 Internatonal Journal of Electroncs and Communcaton Engneerng Artfcal Neural Networks for Cogntve Rado Network: A Survey Vshnu Pratap Sngh Krar Abstract The man am of a communcaton system s to acheve maxmum performance. In Cogntve Rado any user or transcever has ablty to sense best sutable channel, whle channel s not n use. It means an unlcensed user can share the spectrum of a lcensed user wthout any nterference. Though, the spectrum sensng consumes a large amount of energy and t can reduce by applyng varous artfcal ntellgent methods for determnng proper spectrum holes. It also ncreases the effcency of Cogntve Rado Network (CRN). In ths survey paper we dscuss the use of dfferent learnng models and mplementaton of Artfcal Neural Network (ANN) to ncrease the learnng and decson makng capacty of CRN wthout affectng bandwdth, cost and sgnal rate. Keywords Artfcal Neural Network, Cogntve Rado, Cogntve Rado Networks, Back Propagaton, Spectrum Sensng. I. INTRODUCTION N evoluton of communcaton system the necessty of I hgher data rate s major concern because at present tme user not only use voce servces but also use vdeo and data servces. The electromagnetc rado frequency spectrum has ts own lmtatons and t s tghtly regulated and allocated wthn all countres of the world by Internatonal Telecommuncaton Unon (ITU). In any country, local government can provde spectrum lcense for servce provders. Rado spectrum allocaton s categorzed as lcensed and unlcensed band. In Lcensed band frequency can used or transmt only n allocated band that they purchased, whle unlcensed band can use any frequency. Thus to use optmum frequency, communcaton system can use dfferent technques lke modulaton, attenuaton, codng. Many research shows that n fxed spectrum allocaton some frequences are used heavly whle some frequences are not used or partally used. Unused frequency s also known as spectrum hole. The spectrum holes are belongs to lcensed user but for some nstants these holes are not used by user. The cogntve rado s a devce that senses these spectrum holes and make avalable for unlcensed user. In CR lcensed user also known as prmary user and unlcensed user as secondary user. The major characterstcs of CR s to ablty to sense, learn, measure, be aware about communcaton channel and ts avalablty.e. spectrum avalablty and power. The concept of Cogntve Rado (CR) s frst ntroduced by Mtola n hs PhD work [1]. He proposed an dea to enhance effectveness of wreless communcaton by make aware to ts Vshnu Pratap Sngh Krar s wth the Computer Scence Department, Unversty of Bedfordshre, Luton, Unted Kngdom. (Phone: ; ; e-mal: Vshnu.krar@study.beds.ac.uk, vshnupskrar@lve.com). rado unts to utlze ts surroundng sources.e. communcaton channel. Haykn [2] ntroduce sgnal processng and communcatons realzaton of CR technology. Cabrc, Mshra and Brodersen and also proposed some fundamental ssues about CR [3]. All these early contrbutons ntroduce spectrum sensng to detect vacant spectrum band and utlze these t. Artfcal Intellgent play an mportant role n wreless communcaton specally to sense the surroundng envronment. It has ablty to learn thngs and adapt tself accordng to nput and provde output. Cogntve Rado Networks fulfll these requrements. Thus, f we apply ANN on CRN then we acheve maxmum performance and maxmum utlzaton of wreless communcaton. Human bran learns new thngs every day, by ths behavor he gan knowledge and become more and more ntellgent and smarter. ANN adopted the property of human bran and provdes soluton for non-lner and probablstc problems. Smlarly, f we want CR to work more ntellgently then we should enable CR to learn. Varous learnng technques of ANN enable CR to learn. Meanwhle all ntellgent algorthms are not useful for CR. Some learnng algorthms can be used to predct communcaton performance but genetc algorthms are sutable for transcever s parameters. Thus, combnaton of dfferent knd of ntellgent algorthms s better for CR. Some of them are ANN, ANFIS, renforcement learnng genetc algorthms, and hdden Markov models renforcement algorthms. In ths survey paper we dscuss varous learnng technques of ANN that mplemented on CRN. Ths survey paper s organzed as follows: sectons II descrbes and compare dfferent learnng models for cogntve rado network. Secton III explans ANN and ts learnng algorthms. We manly dscuss the Back propagaton (BP) algorthm and Feed Forward Neural Network (FFNN). Secton IV dscusses the varous ANN that mplemented on CRN and provde a comparson between them. Fnally we conclude the paper n Secton V. II. DEFERENT LEARNING MODELS FOR COGNITIVE RADIO NETWORKS Cogntve rado has specal ablty of learnng about ts surroundng communcaton system and remembers the nformaton lke knowledge. Accordng to stuaton f CR needs help from prevous knowledge then t can retreve nformaton. It s also useful to make an accurate decson. In CRN for spectrum sensng, spectrum behavor, spectrum selecton, performance and other dfferent parameters, dfferent learnng models are mplemented. Some of them are Internatonal Scholarly and Scentfc Research & Innovaton 9(1)

2 Internatonal Journal of Electroncs and Communcaton Engneerng Machne Learnng, Collaboratve flterng and self-learnng model. Neural Network, Genetc Algorthm, Markov Model and Game Theory are used for dynamc parameters lke spectrum and channel selecton. Neural Network s best sutable soluton for pattern recognton and probablstc problems. Thus, ANN s applyng for transmsson rate, sgnal predcton, decson makng n CRN [4]. Self-organzng learnng technques are mplemented for detecton of surroundng spectrums. Dfferent learnng technques are explaned n Table I. Learnng Model Markov Model Q-Learnng Game Theory Fuzzy Logc Genetc Algorthms Neural Networks TABLE I LEARNING MODELS FOR COGNITIVE RADIO NETWORK Technques used for Dynamc Spectrum Access Advantages Improved throughput Lmtatons Provdes some undesrable solutonss Modellng, Cogntve cycle Channel Selecton Hgher performance rate Hgher utlzaton rate Only for local parameters Performance depends upon partcular parameter selecton Transmsson rate predcton Less complexty Need some global nformaton n addton Optmzaton Excellent for parameter Performance depends upon optmzaton partcular parameter selecton Dynamc channel Selecton, Learn n absence of prevous Performance depends upon Cogntve Engne, nformatonn partcular parameter selecton III. ARTIFICIAL NEURAL NETWORK ANN s dentcal to the bologcal cells of the human bran, t consst of a number of nterconnected processors also known as neurons. Its neuron model, archtecture and learnng algorthms can explan ANN. Archtecture refers to a number of neurons and the lnks connected to these neurons n dfferent layers. The lnk between neurons s known as weght and t s adjusted durng the tranng phase. Neuron model processes the nformaton that they receve as nput and provde an output. Thus, there s a fxed output for a partcular nput nformaton, attrbute or data. Learnng algorthms are essental and fundamental characterstc of ANN and used for ts tranng. Durng the tranng weghts are automatcally updated usng the negatve gradent of Mean Square Error (MSE). If network fnd an error than ths error sgnal s agan feed nto the lower layer of ANN. ANN conssts of number of neurons and they are nterconnected by weghted lnks. Generally ANN structure made by three layers: ncomng neurons layer whch receve the ncomng sgnals, hdden neuron layer and output neuron layer. The ncomng sgnals (x ) are multpled by the correspondng weghts (w ) of the lnks and a bas term (b ) s added. All these terms now multpled to form an nput to next layers neuron, whch s subjected to a nonlnear functon.e. actvaton functon lke sgmodal or hyperbolc. A sngle neuron classfer model s shown n the Fg. 1 [5]. Fg. 1 Sngle Neuron as a classfer Fnally, the standard outpu y(t) defne as: n y ( w x b) 1 A. Feed forward Neural Network (FFNN) Feed Forward (FF) neural network have most powerful mappng technques. Applyng on multplcatve network t gves faster learnng tme and excellent approxmaton capactes. Its results are better than mult-layer perceptons because t can process hgher order nformaton. Multplcatve Neuron Model (MNM) s mplemented n hgher order neural networks. Fg. 2 P Neuron Based MNM A P neuron based ANN s shown here n Fg. 2. In ths MNM, at each neuron adds all weghts ncomng to t along wth bas, now these summatons are multpled and generate an output for ths neuron. Now ths output wll become an nput for next layers neuron. A bpolar actvaton functon s appled before the fnal output. These MNM neurons look very complex at the frst but t requred a less number of parameters as compare to other neuron models [6]. Observaton and scope Also mplemented n cogntve engne Realzaton of Cogntve engne Ablty to enhance the capacty for next generaton servces Can Predct other parameters of CRN Can be update nput weght automatcally Can also mplemented n all aspectss of CRN (1) (2) (3) Internatonal Scholarly and Scentfc Research & Innovaton 9(1)

3 Internatonal Journal of Electroncs and Communcaton Engneerng B. Back Propagaton (BP) Algorthm The Back Propagaton (BP) algorthm was frst ntroduced by Rumelhart, Hnton and Wllams n 1986 [7]. BP provdes effectve learnng for many practcal applcatons. In BP weght change s calculated by usng two term algorthms, leanng rate and momentum factor. BP algorthm faces problem of local mnma and slow converson speed. Zwer [8] propose proportonal factor of two-term cost of ANN. The BP method BP to reduce complexty and computatonal calculates the frst dervatvee for estmatng the gradent. Thus, t s the most popular among other methods. On BP varous approaches are proposed to avod local mnma and they are based on selecton of momentum and dynamc varaton of machne learnng along wth sutable cost functon and actvaton functon. Momentum coeffcent and learnng rate selected accordng to the prevous weght update of neuron and coeffcent of downhlll gradent [9]. For fast mnmum search Drago et al. [10] proposed an adaptve momentum BP. Chen et al. proposed sequence of weghtng vector at the learnng phase. Vshnu et al. [11] mplemented three terms BP on XOR problem to solve the local mnma problem. The BP learnng algorthm wth multplcatve neural model explaned n Fg. 3. Fg. 3 Multplcatve Neural Network Model The standardd algorthm can modfy by applyng momentum term and proportonal factor term. The momentum term obtan by quantzatonn of weght change. Varatons n slop of error suppress the oscllaton and gradent due to anomales. It prevents ANN to fall n to local mnma problem. Saturaton behavor of actvaton functon keeps thee convergencee speed relatvely slow or almost constant. The problem of convergence speed s resolved by addngg proportonal factor between the tranng targets and outputs. The mproved term BP weght and bas can be calculated as ww bb mproved mproved w b w old b old ( y d) ( y d) s the proportonal term. The error functon optmzaton depends on these ndependent quanttes [11]. IV. VARIOUS V ANNN IMPLEMENTED ON CRN CRN s a new emergng technology that attracts researchers. Many of them apply varous ANN methods on t (4) (5) and acheve excellent results because the some characterstcs and propertes of ANN and CNR are smlar especally the ntellgence toward the sensng or trackng. In ths secton we dscuss some proposals that presented by researchers. Performance of ANN n CRN s shown n Table II. A. Proposal 1 Baldo and Zorz [12] proposed a Multlayered Feed forward Neural Network (MNFF) for performance of real-tme communcaton for Cogntve Rado System. They use the functon approxmaton of MFNN to obtan envronment measurement and performance measurement. Now by usng the sub set of ths nformaton MFNN s tran wth Back Propagaton (BP) algorthm. The CR s now tran and able to perform n varous dfferent envronments. NS-Mracle smulator used to obtan the sub set of nformaton. Ths performance s compared wth Banch s model [13]. Banch proposed kalman flter for performancee calculaton. After comparng the results Baldo and Zorzm [12] found that Banch s model have much more complexty and have less accuracy of throughput performance. MFNN provdes good accuracy and t s very flexble. It ncreases the performance and optmzng the confguraton of CRN. B. Proposal 2 Zhang and Xe [14] desgn a neural network for decson makng of CR engne, whch s based on evaluaton, and learnng. CR engne s based on Genetc Algorthm (GA). GA works on varous parameters of system chromosomes. These chromosomes are based on knowledge and help to make a decson. For ths both changeable nformaton (sgnal rate, automatc repeat request (ARQ), FCC, bandwdth, modulaton and encrypton) and unchangeable nformaton (lcensed user or owner and cost) are collected and processed. These nformaton become nput for neural network and t gves best decson as the output. In ths study Levenberg-Marquardt (LM) algorthm s used for tranng and performance ndex s Mean Square Error (MSE). Ths neural network has great ablty of non-lnear.e. nformaton and t also has a less complex structure. It can easly smulate the nformaton at nput and reflecton, t need very less prevous knowledge the output for complex network. Zhang and Xe [14] compare ther model wth Reser s cogntve rado engne. Reser s model has lmtatons because ts decson s based only on unchangeable nformaton [ 15]. On the other hand, neural network model proposed by Zhang and Xe [14] s able to make decson for CR engne, whch s based on both changeable and unchangeablee nformaton. C. Proposal 3 Zhu et al. [ 16] propose an Adaptve Resonance Theory (ART-2) Neural Network for channel sensng. It also satsfes the cogntve Wreless Mesh Network (WMN) structure, whch also combnes wth sgnal broadcast system. Zhu et al. also consder that WMN coexst wth Wreless Regonal Area Network (WRAN). WRAN dstrbutes the sgnal spectrum nto separate sub-bands. These sub-bands perform the channel sensng. WMN consst of clusters and mult channels, so Mesh Internatonal Scholarly and Scentfc Research & Innovaton 9(1)

4 Internatonal Journal of Electroncs and Communcaton Engneerng Pont sense the nformaton and nform to cluster head. The data fuson n channel sensng s smlar to pattern recognton problem thus ART-2 s more sutable for CRN. ART-2 has ablty to self-organze the nput and creates resonance state and assocate wth categores. These categores follow specfc prototype patterns. Neural network accept all these patterns of same category as an nput and tran the system. After the tranng we select the most sgnfcant node. If network not provde any node then t s assumed that MP provde wrong nformaton. Zhu et al. [16] compare ther smulaton results wth Bayesan Draft Protocol and found that ther network provdes better accuracy. D. Proposal 4 Tumuluru et al. [17] proposed a spectrum predctor for cogntve rado usng Multlayer Perceptron (MLP) Neural Network. Ths MLP has specal characterstc that t does not requre prevous knowledge or data of traffc characterstcs of lcensed user. Neural network creates mappng functon between nput data and output data. Ths data s n the form of bnary whch s obtaned by channel sensng durng dfferent tme. When lcensed user s actve then channel status s busy and when user s absent then channel s dle. In ANN we represent ths stuaton as a two-class problem. Bnary representaton for busy and dle condton s 1 and -1 respectvely. MLP predctor uses the BP algorthm. For tranng channel sensng data s provded as nput. Neural network maps nput data wth output data. In ths problem output data s 1/-1. We have desred output and neural network provdes desred output. Dfference between desred and estmated output provdes error. The less error provdes the better results and accuracy.e. predcton. Tumuluru et al. compare ther results wth HMM based spectrum predcton scheme [18], [19], whch does not provde detals about length of observaton sequence and number of states. Spectrum TABLE II PERFORMANCE OF ARTIFICIAL NETWORK IN COGNITIVE RADIO NETWORK Authors Uses for CRN Input Attrbutes Output Actvaton Baldo and Zorz [12] Performance charactersaton of component Decson Makng Receved Frames Idle Tme SNR ARQ, FCC, Sgnal Rate, Bandwdth, Modulaton, Encrypton, Cost, Owner predcton CRN save the sensng energy and mprove spectrum utlzaton of communcaton channel. E. Proposal 5 Ca et al. [20] proposed an Incremental Self-Organzng Map ntegrated wth neural Network (ISOM-HNN) for sgnal classfcaton n CRN. Ths approach detects unknown rado sgnals n wde communcaton network or channel. ISOM mproves real tme learnng performance and HNN mproves learnng along wth predcton accuracy. ISOM provde ncremental learnng to SOM. ISOM update the weght of neurons by calculatng the total number of nputs n neurons. As number of nput s ncreased the magntude of weght s also ncreases. By ths method ISOM grows dynamcally and detect the unknown sgnals contnuously. For learnng, predcton and assocaton of HNN, the modfed Hebban learnng algorthm s proposed. ISOM-HNN dscards the dependency of data dmensonalty and t enhances capacty of CRN to dentfy authorzed and unauthorzed rado sgnal n communcaton spectrum. F. Proposal 6 Tang et al. [21] propose an Artfcal Neural Network for spectrum sensng of CRN under low Sgnal-to-Nose Rato (SNR). Prmary user has Ampltude Modulaton (AM) sgnals. Secondary user perform ANN based detecton method to sense whether the prmary user occupy the channel or not. The attrbutes of four nput neurons are energy and three cyclostatonary values. At the tranng phase, weghts and threshold of each neuron are updated at each-teraton. Tranng followed the feature abstracton. Addtve Whte Gaussan Nose (AWGN) s added to AM sgnals to ntroduce SNR n network. Proposed ANN has advantages of cyclostatonary values detecton and energy detecton. Ths ANN has less computatonal complexty and reduces the nterference n CRN. Attrbutes Throughput RelabltyDelay Layers Functon Sgmod functon Multlayer forward NN Zhang and Xe [14] Mean Square Error (MSE) Sgmod Functon Multlayer (ML) BP NN Zhu et al. Channel Sensng Prototype Patterns Mean Square Poson ART-2 NN [16] Node Dstrbuton Tumuluru et Spectrum Predcton Traffc Characterstcs Two classes (1/-1) Sgmod functon MLBP NN al. [17] MSE Ca et al. [20] Sgnal Classfcaton Channel bandwdth, Mohalanobs Incremental ISOM-HNN Dwellng tme dstance Functon Tang et al. Spectrum Sensng Energy Cyclostatonary values Cyclc Spectrum Threshold BPNN [21] functon Shams et al. Predctve Modellng Traffc Dstrbuton MSE Hyperbolc Feedforword NN [22] Mult secondary user functon BPNN Tan et al. [23] Frequency Allocaton Frequency MSE Sgmod BPNN Functon Zhang et al. Cooperatve spectrum Probablty forecast of Fuson Centre MSE Threshold BPNN [24] sensng functon Gatla et al. Lnk qualty MSE Sgmod [25] Sgnal Strength Functon Performance (Throughput, Data Rate) Focused tme delay Neural Network Consderato ns Number of users n CRN System Chromosomes Data Fuson Pror Knowledge Data dmensonalty SNR AWGN Delay Lne Weght at dfferent tme AWGN SNR FTDNN Internatonal Scholarly and Scentfc Research & Innovaton 9(1)

5 Internatonal Journal of Electroncs and Communcaton Engneerng G. Proposal 7 Shams et al. [22] desgn Tme Delay Neural Network (TDNN) and Recurrent Neural Network (RNN) for predctve modelng and mult secondary user scenaro n CRN. It mproves the spectrum utlzaton. It helps the secondary user to choose best possble and avalable communcaton channel. The error for predcton s almost zero n these methods. Secondary user dvde lcensed channel n to small tme slots. For each tme slot secondary user sense the spectrum holes. In deal condton secondary user sense the vacant channel correctly. TDNN s a Feed Forward network and a delay lne s appled to ts nput. RNN s back propagaton network and t has feedback connecton from output node to nput node. Ths feedback generates a pattern for each tme nstance. Prmary user traffc dstrbuton as a bnary sequence work as an Input for tranng of TDNN and RNN. Two data sets are generated by these networks by usng feedback pattern and bnary sequence. For learnng of TDNN and RNN, BP algorthm s used. After successful tranng, MSE s calculated as performance of CRN. Secondary user uses the channel status predctor wth maxmum prorty. Shams et al. [22] also explan the spectrum resource securty. These networks help to secondary user to sense the spectrum holes and make them to accessble. It also provdes accurate actvty of prmary user. Thus t also reduces the nterference. TDNN and RNN both have hghest predcton probablty. H. Proposal 8 Tan et al. [23] propose an ANN to solve a frequency allocaton problem n CR. In CR prmary user has lcense or rght to use frequency any tme. On the other hand, secondary user only use at partcular tme. User has dfferent weght n CR. Thus, two hypotheses are adopted. Frstly, mult user wth dfferent weght at same tme and secondly, sngle user wth dfferent tme. BPNN tran the weght of each user for same tme nstant and dfferent tme. The output provdes decrease dstance between actual output and expected output. In CRN, demand of frequency s changes over the tme. Thus, user has to keep n touch wth the envronment. The network desgned by Tan et al. provdes a faster and accuracy toward frequency allocaton due to less complexty n computaton. I. Proposal 9 Zhang et al. [24] proposed an ANN for cooperatve spectrum sensng of CRN. Fuson centre s used to fnd the probablty of weghts. Secondary User/Unt (SU) sense for prmary user/unt (PU) and send the nformaton to fuson centre. Fuson centre s used to fnd the probablty of weghts. And t s work as an nput of ANN. Spectrum sensng s dvded nto three hypotheses n ths model. These are spectrum sensng of ndvdual SU, communcaton between SU and fuson centre, and fuson scheme. For tranng phase SU provde the sensng nformaton as nput and after the tranng SU stop to work. Now fuson centre also stop to send reference sgnal to PU. Thus PU gets knowledge about probablty of SU weght. As SU and PU are both nvolve n spectrum sensng of CRN thus t s known as cooperatve spectrum sensng. Ths model provdes the excellent performance of probablty and detecton probablty. J. Proposal 10 Gatla et al. [25] proposed a learnng model usng neural network to calculate performance of CRN and ts parameters lke throughput and data rate. Ths network uses the non-lnear transfer functon to map lnear as well as non-lnear nput and output. The lnear output has two classes and generally represented by 1 and -1. Preprocessng s appled to normalze the data. Focused Tme-Delay Neural Network (FTDNN) provdes delay lnes n the nput. To update the weght and the bas, LM algorthm s used. To measure the data rate, bt rate and sgnal strength works as an nput of NN. Ths model explans the relaton between sgnal strength and data rate of CRN. V. CONCLUSION In present scenaro, wreless network spectrum resources are backbone of communcaton across the world and t has potental to rapd ncrease. There are a lot of possbltes n the research feld of CRN especally n aeronautcal and satellte communcaton systems. In ths survey paper we dscuss varous mplementaton presented by dfferent authors. From the gven proposals we conclude that BP algorthm s best sutable algorthm for ANN. Sgmod functon provdes best result n ANN. And most popular method to descrbe output parameter for ANN s Mean Square Error (MSE). The performance of proposal presented by Shams et al. [22], gve the most sgnfcant and accurate results. The accuracy of the network s hgher than other proposal that descrbe above. The MSE for network s almost zero. In general the proposal of Shams et al. [22] s best among the other proposals. The man mportance of CRN s to sense the spectrum or predcton. If CRN has effectve sensng power than then t can use all the resources of communcaton channel. ANN has very good ablty for recognton and predcton of physcal and logcal attrbutes. These abltes of ANN are mplemented on CRN to acheve maxmum performance of CRN. It also ncreases the accuracy and effectveness of CRN. REFERENCES [1] J. Mtola, Cogntve Rado: An Integrated Agent Archtecture for Software defne Rado, Ph.D. dssertaton, Royal Insttute of Technology (KTH), Sweden, [2] S. Haykn, "Cogntve Rado: Bran-Empowered Wreless Communcatons," IEEE Journal on Selected Areas n Communcatons, vol. 23, no. 2, pp , February [3] D. Cabrc and R. W. Brodersen, "Physcal Layer Desgn Issues Unque to Cogntve Rado Systems," n Proc. of PIMRC-2005, pp , September [4] M. Venkatesan, A. V. Kulkarn, Soft Computng based Learnng for Cogntve Rado, nternatonal journal on Recent Trends n Engneerng and Technology, vol. 10, ssue 1, pp , January [5] K. Burse, A. Mshra, A. Somkuwar, Convergence Analyss of Complex Valued Multplcatve Neural Network for varous Actvaton Functons, IEEE Internatonal Conference on Computatonal Intellgence and Communcaton System (CICN 2011), pp , October [6] V. P. S. Krar, K. Burse, R. N. Yadav, S. C. Srvastav, A Compact P Network for Reducng Bt Error Rate n Dspersve FIR Channel Nose Internatonal Scholarly and Scentfc Research & Innovaton 9(1)

6 Internatonal Journal of Electroncs and Communcaton Engneerng Model, Proceedngs of World Academy of Scence, Engneerng and Technology, vol. 38, pp , Ferbuary [7] D.E. Rumelhart, G.E. Hnton and R.J. Wllams, "Learnng representatons by back -propagatng errors," Nature (London), 323, , [8] Yahya H. Zwer, Lakmal D. Senevratne, and Kaspar Althoefer Stablty analyss of a three-term backpropagaton algorthm. Neural Netw. 18, 10 (December 2005), [9] Y. F. Yam and T.W.S. Chow, Extended Back Propagaton Algorthm, Electroncs Letters, vol. 29(19), pp , [10] G. P. Drago, M. Morando and S. Rdella, An Adaptve Momentum Back Propagaton, Neural Computng and Applcaton, vol. 3, pp , [11] V. P. S. Krar, K. Burse, M. Manora, Improved Back Propagaton Algorthm for Complex Multplcatve Neuron Model, Proceedngs of Sprnger conference, Informaton Technology and Moble Communcaton, Communcaton n Computer and Informaton Scence (CCIS), vol. 147, pp , Aprl [12] N. Baldo and M. Zorz, Learnng and Adaptaton n Cogntve Rados usng Neural Networks, 5th IEEE Consumer Communcatons and Networkng Conference (CCNC 2008), pp , january [13] G. Banch, Performance Analyss of the IEEE Dstrbuted Coordnaton Functon, IEEE Journal on Selected Areas n Communcatons, vol. 18, no. 3, pp , March [14] Z. Zhang and X. Xe, Intellgent Cogntve Rado: Research on Learnng and Evaluaton of CR Based on Neural Network, Proceedngs ITI 5th Internatonal Conference on Informaton and Communcatons Technology (ICICT 2007), pp , December [15] C. J. Reser, T. W. Rondeau, C. W. Bostan, and T. M. Gallagher. Cogntve RadoTest bed: Further Detals and Testng of a Dstrbuted Genetc Algorthm Based Cogntve Engne for Programmable Rados, IEEE MILCOM, October [16] X. Zhu, Y. Lu, W. Weng, and D. Yuan, Channel Sensng Algorthm based on Neural Network for Cogntve Wreless Mesh Network, n Proceedngs of IEEE Internatonal Conference on Wreless Communcatons (WCom), pp. 1-4, [17] V. K. Tumuluru, P. Wang, and D. Nyato, A Neural Network Based Spectrum Predcton Scheme for Cogntve Rado, In IEEE Internatonal Conference on Communcaton (ICC), Cape Town, South Afrca, pp. 1-5, [18] A. Akbar and W. H. Tranter, Dynamc Spectrum Allocaton n Cogntve Rado usng Hdden Markov Models: Posson Dstrbuted Case, n Proceedngs of IEEE SoutheastCon, pp , March [19] C. H. Park, S. W. Km, S. M. Lm and M. S. Song, HMM based Channel Status Predctor for Cogntve Rado, n Proceedngs of Asa- Pacfc Mcrowave Conference (APMC), pp. 1-4, December [20] Q. Ca, S. Chen, X. L, N. Hu, H. He, Y.-D. Yao, and J. Mtola, An Integrated Incremental Self-Organzng Map and Herarchcal Neural Network Approach for Cogntve Rado Learnng, The 2010 Internatonal Jont Conference on n Neural Networks (IJCNN), pp. 1-6, July [21] Yu-Je Tang, Qn-Yu Zhang, We Ln, Artfcal Neural Network based Spectrum Sensng Method for Cogntve Rado, IEEE conference on wreless communcatons and moble computng, pp. 1-4, September [22] N. Shams, A. Mousavna, H. Amrpour, A Channel State Predcton for Mult-Secondary users n a Cogntve Rado based on Neural Network, Internatonal Conference on Electroncs, Computer and Computaton (ICECCO)2013, pp , November [23] X. Tan, H. Huang, L. Ma, Frequency Allocaton wth Artfcal Neural Networks n Cogntve Rado System, IEEE TENCON Sprng Conference 2013, pp , Aprl [24] T. Zhang, M. Wu. C. Lu, Cooperatve Spectrum Sensng based on Artfcal Neural Network for Cogntve Rado System, 8 th Internatonal Conference on Wreless Communcaton, Networkng and Moble Computng (WCOM) 2012, pp. 1-5, September [25] V. Gatla, M. Venkatesan, A. V. Kulkarn, Feed Forward Neural Network based learnng scheme for cogntve rado systems, Thrd Internatonal Conference on Computatonal Intellgance and Informaton technology, CIIT 2013, pp , October Internatonal Scholarly and Scentfc Research & Innovaton 9(1)