Performance Enhancement in Machine Learning System using Hybrid Bee Colony based Neural Network

Transcription

1 Performance Enhancement n Machne Learnng System usng Hybrd Bee Colony based Neural Network S. Karthck 1* 1 Team Manager, Sea Sense Softwares (P) Ltd., Marthandam, Taml Nadu, nda *** Abstract - Metrc learnng and data predcton s a notable ssue n many data mnng and machne learnng applcatons. n recent advances many researches have been proposed for the data predcton. Stll there s a problem such as reduced predcton accuracy, computatonal tme cost etc. Hence n ths paper a novel Hybrd Bee Colony based Neural Network (HBCNN) s proposed for the data predcton n data mnng. The proposed technque s segmented nto three phases, n the frst phase the raw nput data s pre-processed to make t sutable for computaton. The second phase s tranng of the proposed HBCNN technque s whch the neural network s traned usng bee colony algorthm. The last phase s dynamc testng n whch the neural network s undergone rapd testng wth dynamc nput for the valdaton and testng of traned ANN. The proposed technque s tested usng cancer dataset and the performance s compared wth KNN technque. Key Words: Metrc Learnng, neural network, bee colony, data predcton 1.NTRODUCTON Artfcal Neural Network (ANN) s often a new ndustry connected wth computatonal scence whch ntegrates the varous strateges for dffculty resoluton whch cannot be consequently effortlessly descrbed wthout the help of an algorthmc conventonal concentraton. The partcular ANNs stand for bg as well as dverse classes connected wth computatonal types [1]. Together wth bologcal template modules these people are created bascally by more or less comprehensve examples. The partcular general approxmates as well as computatonal types havng specfc trats such as the ablty to learn or even adapt, to prepare or to generalze data are recognzed as the partcular Artfcal Neural Networks. Contemporary development of a (near) optmal network archtecture s carred out by means of people sklled together wth use of a monotonous learnng from tral and error process. Neural networks are generally algorthms regardng optmzaton as well as fndng out, and are dependent freely wth prncples prompted merely by researchng on the characterstcs from the bran. For optmzaton as well as fndng out neural networks n tandem wth genetc algorthms, there are generally two technques, each usng ts own advantages coupled wth weaknesses. By means of dfferent paths, the two processes are commonly evolved [2, 3]. Optmzaton algorthms, n addton to beng called Learnng algorthms, are, accordng to a number of features of scentfc advancement, usually known as Genetc algorthms. An easy method nvolvng encodng answers to the ssue n chromosomes, an assessment purpose whch on return starts to attan Chromosome drected at the dea, nvolvng ntalzng populaton of chromosomes, workers that could be put on parents whenever they reproduce to correct the genetc composton are the fve expected crtera. ntegraton could be of mutaton, crossover n addton to ste-specfc workers. Parameter confguratons for the partcular crteron are the workers or can be anythng else [4]. The crtera can develop populatons nvolvng much better ones n addton to the much better ndvduals, convergng eventually n effect close to an nternatonal deal, every tme a genetc algorthm s actually run havng a portrayal whch usefully encodes a soluton to a problem n addton to workers that can generate much better young ones through good parents [5]. n numerous stuatons, wth regard to performng the partcular optmzaton on the standard workers, mutaton n addton to crossover, s usually adequate [6]. Genetc algorthms can assst lke a black box functon optmzer certanly not requestng any knd of know-how about computers of the partcular doman n these cases. However, understandng of the partcular doman s frequently exploted to rase the partcular genetc algorthms performance wth the ncorporaton nvolvng new workers hghlghted n ths partcular paper [7]. n neural network, functonalty enhancement partton space can be a space whch s used to classfy data sample rght after the test s actually mapped by neural network [8]. Centrod can be a space wth partton space and also denotes the partcular mddle of the class. n tradtonal neural network functonalty, locaton of centrods and the partnershp nvolvng centrods and also classes are generally fxed personally [9]. Furthermore, wth reference to the quantty of classes varety of centrods s actually set. To locate optmum neural network ths partcular set centrod restrcton mnmzes the rsk [9]. Genetc algorthms (GAs) have emerged to be practcal software for the heurstc alternatve of sophstcated dscrete optmzaton ssues. Partcularly, there has been large fascnaton wth ther use n the most effectve arrangng and tmetablng ssues [10]. However, nowadays there have been several attempts made to become lsted on both systems. Neural networks could be rewrtten snce a type of genetc algorthm s termed as some sort of classfer program and also vce-versa [11]. Genetc algorthm s meant for tranng feed forward networks. t does not merely work wth ts task but t does perform rear dstrbuton, the normal tranng 2018, RJET mpact Factor value: SO 9001:2008 Certfed Journal Page 931

2 Cancer Data Attrbute crtera [12, 13]. Ths accomplshment comes from talorng the partcular genetc algorthm to the doman of tranng neural networks. Although the ANN features enjoy several advantages t has got several complcatons smlar to convergence n addton to recevng throughout neghborhood mnma durng tranng by means of back propagaton [14]. That s why the analysts are engaged to produce a new fnest protocol to tran the ANN. On ths routne optmzaton protocol, algorthms smlar to PSO, CS etc have been utlzed to enhance the overall performance however t certanly does not fulfll the overall performance prerequstes. On ths mpresson our system s also thought out to be able to get over a real ssue durng tranng nvolvng ANN. Thus, we proudly propose the novel AGCS technque for tranng the ANN. 2. PROPOSED HBCNN TECHNQUE FOR MACHNE LEARNNG SYSTEM Metrc learnng and Data predcton s a notable ssue n many data mnng and machne learnng applcatons. n recent advances many researches have been proposed for the data predcton. Stll there s a problem such as reduced predcton accuracy, computatonal tme cost etc. Hence a novel hybrd HBCNN s proposed for the data predcton n data mnng. The archtecture of the proposed system s gven n fg Dsease Database Pre-processng Feature extracton Classfcaton Output Fg -1: Archtecture of the proposed system The proposed system ncludes four phases they are; 1. Data Collecton 2. Data Preprocessng 3. ANN-ABC Predctor Modellng 4. Dynamc Testng 2.1 Artfcal Neural Network An ANN s a programmed computatonal model that ams to duplcate the neural structure and the functonng of the human bran. t s made up of an nterconnected structure of artfcally produced neurons that functon as pathways for data transfer. Artfcal neural networks are flexble and adaptve, learnng and adjustng wth each dfferent nternal or external stmulus. Artfcal neural networks are used n sequence and pattern recognton systems, data processng, robotcs and modelng. The ANN conssts of a sngle nput layer, and a sngle output layer n addton to one or more hdden layers. All nodes are composed of neurons except the nput layer. The number of nodes n each layer vares dependng on the problem. The complexty of the archtecture of the network s dependent upon the number of hdden layers and nodes. Tranng an ANN s to fnd a set of weghts that would gve desred values at the output when presented wth dfferent patterns at ts nput. The two man process of an ANN s tranng and testng. The total of eght attrbutes or nput s gven as the nput for the ANN and the correspondng class or dsease can be obtaned as the output of the ANN. Thus the proposed ANN archtecture contans eght nputs (eght attrbutes) and correspondng dsease output. The man two process of a classfcaton algorthm s tranng and testng. n tranng the nput as well as the output wll be defned and the approprate weght s fxed so that the classfer (ANN) can able to predct the apt object (dsease) n the testng phase. Hence the tranng phase s the major part n a common classfer algorthm. n the proposed system the artfcal bee colony (ABC) algorthm s used for the tranng. The process nvolved n the tranng s gven as follows and the archtecture of the back propagaton neural network (BPNN) s gven n the fg , RJET mpact Factor value: SO 9001:2008 Certfed Journal Page 932

3 F1 w 11 w 12 w 13 F2 w 21 w 15 w 22 w 23 w 14 F3 w 24 w 25 w 31 w 32 w 33 w 34 w 35 w 41 w 42 w O 2 w O 1 F4 w 43 w 44 w 45 w 51 w 52 w O 3 C w 53 F5 w 54 w 55 w O 4 w 61 w 62 F6 w 65 w 63 w 64 w O 5 w 71 w 72 w 73 F7 w 75 w 74 w 81 w 82 w 83 w 84 F8 w 85 Fg -2: Proposed back propagaton neural network The proposed ANN consst of eght nput unts, one output unts, and M hdden unts (M=5). Frst, the nput data s transmtted to the hdden layer and then, to the output layer. Ths s called the forward pass of the back propagaton algorthm. Each node n the hdden layer gets nput from the nput layer, whch are multplexed wth approprate weghts and summed. The output of the neural network s obtaned by the Eqn. (1) gven below. C M j1 w 1 exp( O j N 1 F w j ) (1) n Eqn. (1), F s the th nput value and w s the weghts assgned between hdden and output layer, w s the weght assgned between nput and hdden layer and M s the number of hdden neurons. The output of the hdden node s the non-lnear transformaton of the resultng sum. Same process s followed n the output layer. The output values from the output layer are compared wth target values and the learnng error rate for the neural network s calculated, whch s gven n Eqn. (2). n Eqn. (2), s the k 1 ft Y C 2 k (2) 2 th k O j j learnng error of the ANN, Y s the desred output and C s the actual output. The error between the nodes s transmtted back towards the hdden layer. Ths s called the backward pass of the back propagaton algorthm. Then the tranng s repeated for some other tranng dataset by changng the weghts of the neural network. The learnng error s consdered as the ftness functon n ABC for the error mnmzaton. B. Artfcal Bee Colony Algorthm Artfcal Bee Colony (ABC) algorthm s a comparatvely new technque proposed by Karaboga. ABC s encouraged by the foragng behavour of honey bee swarms. n the ABC algorthm, the colony conssts of three knds of bees namely: Employee bee Onlooker bee Scout bee Among these three knds of bees, the employee bee and the onlooker bee are the employed bees, whereas, the scout bee s an unemployed bee. The number of employee bees and onlooker bees are sad to be equal n ABC algorthm. The food sources are consdered as the possble solutons for a gven problem and the nectar amount of the food source s the relatve ftness of that partcular soluton. The populaton of the colony s twce the sze of the food sources. The number of food sources represents the poston of the possble solutons of the optmzaton problem and the nectar amount of a food source represents the qualty (ftness) of the assocated soluton. The pseudo code of the ABC algorthm s gven below. 2018, RJET mpact Factor value: SO 9001:2008 Certfed Journal Page 933

4 ntalze food sources Repeat Employed Bee Phase: Send the employed bees to the food sources assgned to them. Determne the amount of nectar (ftness value) n the food source. Calculate the probablty value of the food sources. Onlooker Bee Phase: Select the food sources dscovered by the employee bee based on the probablty value. Fnd the neghborng food source and estmate ts nectar amount. Compare both the food sources and select the food source wth better ftness. Scout Bee Phase: Select the food sources randomly and replace the abandoned food source wth the new food source. Memorze the best food source. Untl (Requrements are met) n the ntal step of ABC algorthm, a set of food sources are selected randomly by the employed bee and ther correspondng nectar amounts (.e. ftness value) are computed. The employee bee shares these detals to another set of employed bees called as onlooker bees. After sharng ths nformaton, the employee bee vsts the same food source agan and then, fnds a new food source nearby usng the Eqn. (3). v, j, j, j (, j k, j x x x ) (3) Where, k and j are random selected ndex that represent the partcular soluton from the populaton. s a random number between [-1,1]. When new neghbourng food sources are generated, ther ftness values are calculated and the employee bee apples greedy selecton to make a decson on whether to replace the exstng food source n memory usng new food source or not. Then, the probablty value P s calculated for each food source based on ts ftness amount usng the followng Eqn. (4). ft P (4) CS / 2 ft 1 Where, CS s the colony sze. 3. Results and Dscusson The proposed system s tested usng cancer dataset and the results are compared wth conventonal technques. The ALL / AML dataset for ths expermental analyss s collected from onlne. After the normalzaton the randomly chosen sample s dvded nto three categores such as tranng, cross valdaton and testng data sets. The tranng data set s used for learnng the network. Cross valdaton s used to measure the tranng performance durng the tranng as well as to stop the tranng f necessary. The Leukaema cancer contans four types. n our system we consder only two major types namely ALL (Acute Lymphoblastc Leukaema) and AML (Acute Myelod Leukaema). For ths tranng purpose we consder 38 patents who are dvded nto two clusters wth 25 and 13 patents for ALL and AML respectvely wth overall 7129 genes. The performance of the proposed system s compared base on the executon tme and accuracy of the predcton. Executon tme s the tme n whch a sngle nstructon s executed. t makes up the last half of the nstructon cycle. Accuracy s used to descrbe the closeness of a measurement to the true value. When the term s appled to sets of measurements of the same measured, t nvolves a component of random error and a component of systematc error. n ths case trueness s the closeness of the mean of a set of measurement results to the actual (true) value and precson s the closeness of agreement among a set of results. K=1 Data Set Table 1: performance analyss table for k=1 accuracy rs 96.00% glass 71.03% vowel 96.77% cancer 67.14% letter 93.95% DNA 73.38% Technques Proposed Exstng System Cosne Based System executon tme accuracy 96.00% 67.29% 96.36% 63.59% 94.55% 73.38% executon tme accuracy 94.67% 67.29% 91.72% 65.48% 94.08% 71.25% executon tme seconds For k=1 performance analyss s compared wth the exstng technque. For the cosne based system executon tme s evaluated wth the proposed and the exstng strateges. Comparson analyss s evaluated for the exstng and the cosne based system. Ths comparson s carred out for the rs, glass, vowel, cancer, letter and DNA. For k=2 performance analyss s compared wth the exstng technque. For the cosne based system executon tme s evaluated wth the proposed and the exstng strateges. 2018, RJET mpact Factor value: SO 9001:2008 Certfed Journal Page 934

5 Fg -3: Performance analyss chart for k=2 Fgure 3 shows the performance analyss of the chart for k=2.comparson analyss s evaluated for the exstng and the cosne based system. Ths comparson s carred out for the rs, glass, vowel, cancer, letter and DNA. Fgure 4 shows the performance analyss of the chart for k=1 for the executon tme. Comparson analyss s evaluated for the exstng and the cosne based system. Ths comparson s carred out for the rs, glass, vowel, cancer, letter and DNA. For k=3 performance analyss s compared wth the exstng technque. For the cosne based system executon tme s evaluated wth the proposed and the exstng strateges. Fgure 5 shows the performance analyss of the chart for k=3. Comparson analyss s evaluated for the exstng and the cosne based system. Ths comparson s carred out for the rs, glass, vowel, cancer, letter and DNA. Fgure 6 shows the performance analyss of the chart for k=1 for the executon tme. Comparson analyss s evaluated for the exstng and the cosne based system. Ths comparson s carred out for the rs, glass, vowel, cancer, letter and DNA. Fg -4: Performance analyss chart the executon tme for K=2 2018, RJET mpact Factor value: SO 9001:2008 Certfed Journal Page 935

6 Fg -5: Performance analyss chart for k=3 Fg -6: Performance analyss chart the executon tme for K=2 4. Concluson Machne learnng s the motvated research topc n the recent decade. Metrc learnng become a challengng task n machne learnng. Hence a novel procedure for the metrc learnng s proposed n ths paper. The proposed system uses two stages of executon, n the frst stage sem supervsed clusterng and second stage predcton s performed. The herarchy forest clusterng technque s utlzed for the sem supervsed clusterng and KNN technque for the predcton. The performance of the system s verfed based on the accuracy and executon tme. The performance analyss outperforms the exstng technques and proves ts effectveness for the metrc learnng n machne learnng. REFERENCES [1] Mantas Lukosevcus, and Herbert Jaeger, Reservor Computng Approaches to Recurrent Neural Network Tranng, Computer Scence Revew, Vol. 3, 2009, pp [2] Davd E. Morarty, Alan C. Schultz, and John J. Grefenstette, Evolutonary Algorthms for Renforcement Learnng, Journal of Arteral ntellgence Research, Vol. 11, 1999, pp [3] Lucas Negr, AdemrNed, Hypolto Kalnowsk, and Alexander Paterno, Benchmark for Peak Detecton Algorthms n Fbre Bragg Gratng nterrogaton and a New Neural Network for ts Performance mprovement, Sensors, Vol. 11, 2011, pp , [4] Davd J. Montana and Lawrence Davs, Tranng Feed forward Neural Networks Usng Genetc Algorthms, n Proceedngs of the 11th nternatonal jont conference on Artfcal ntellgence, Vol. 1, 1989, pp , [5] Wllam B. Langdon, Robert. McKay and Lee Spector, Genetc Programmng, nternatonal Seres n 2018, RJET mpact Factor value: SO 9001:2008 Certfed Journal Page 936

7 Operatons Research & Management Scence, Vol. 146, 2010, pp , [6] Swagatam Das, Ajth Abraham, Uday K. Chakraborty, and Amt Konar, Dfferental Evoluton Usng a Neghborhood-Based Mutaton Operator, EEE Transactons on Evolutonary Computaton, Vol. 13, 2009, pp , [7] Adam M. Smth and Mchael Mateas, Answer Set Programmng for Procedural Content Generaton: A Desgn Space Approach, EEE Transactons on Computatonal ntellgence and A n Games, Vol. 3, 2011, pp , [8] Ln Wang, Bo Yang, Yuehu Chen, Ajth Abraham, Hongwe Sun, Zhenxang Chen and Hayang Wang, mprovement of neural network classfer usng floatng centrods, Vol. 31, 2012, pp , [9] [9] Le Zhang, Ln Wang, Xujewen Wang, Keke Lu, and Ajth Abraham, Research of Neural Network Classfer Based on FCM and PSO for Breast Cancer Classfcaton, Hybrd Artfcal ntellgent Systems, Vol. 7208, 2012, pp , [10] Tanomaru J., Staff Schedulng by a Genetc Algorthm wth Heurstc Operators, n Proceedngs of the EEE Conference on Evolutonary Computaton, 1995, pp [11] Patrycja Vaslyev Mssuro, Predctng Genetc nteractons n Caenorhabdts Elegans usng Machne Learnng, Massachusetts nsttute of Technology, 2010, pp [12] Potr Mrowsk, Deepak Madhavan, Yann LeCun, and Ruben Kuznecky, Classfcaton of patterns of EEG synchronzaton for sezure predcton, Clncal Neurophysology, Vol. 120, 2009, pp [13] XS Yang and S. Deb, Cuckoo search va Lévy flghts, n Proceedngs of World Congress on Nature & Bologcally nspred Computng, 2009, pp [14] L. Mngguang and L. Gaoyang, Artfcal Neural Network Co-optmzaton Algorthm based on Dfferental Evoluton, n Proceedngs of Second nternatonal Symposum on Computatonal ntellgence and Desgn, , RJET mpact Factor value: SO 9001:2008 Certfed Journal Page 937