Multi-objective Genetic Algorithm Based Selective Neural Networks Ensemble for Concentration Estimation of Indoor Air Pollutants Using Electronic Nose

Transcription

1 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 10, Issue 3, No, May 013 ISSN (Prnt): ISSN (Onlne): Mult-objectve Genetc Algorthm Based Selectve Neural Networks Ensemble for Concentraton Estmaton of Indoor Ar Pollutants Usng Electronc Nose Chabou Kadr 1*, Fengchun Tan 1, Le Zhang 1, Xongwe Peng 1, and Xn Yn 1 1 College of Communcaton Engneerng, Chongqng Unversty ShaZheng street 174, ShaPngBa dstrct, Chongqng , Chna Abstract Neural networks emble or commttee of neural networks s a learnng approach where many neural networks are combned to solve a gven problem. Ths approach has been proved to mprove the generalzaton performance of ndvdual networks (base networks), provded these networks are accurate enough whle beng error-ndependent (dverse). In ths paper, varance nflaton factor (VIF) s defned as dversty measure. A mult-objectve genetc algorthm (MOGA) wth two objectves (emble error and the new dversty metrc) s used to select approprate members of the emble from a pool of traned neural networks. The proposed method heren called MOGASEN(Mult Objectve Genetc Algorthm based Selectve emble) and other popular emble approaches were evaluated on data from an electronc nose (E-Nose) for concentraton estmaton of four ndoor ar pollutants (formaldehyde, benzene, toluene, and carbon monoxde). Emprcal results show that the proposed method, whle havng hgher capablty n reducng the sze of the emble, was, n most cases, able to outperform other methods. Keywords: Neural network emble, Electronc nose, varance nflaton factor, Mult-objectve genetc algorthm, ar qualty montorng 1. Introducton Neural network emble (NNE) s a learnng method where multple neural networks are traned to fulfll a gven task, and ther predctons are combned to form the emble s output [1]. Snce ts ncepton, ths approach to learnng has been successfully appled n many domans, ncludng medcal dagnoss [], electronc nose systems [3, 4], optcal character recognton [5], and so forth. There are generally two steps n the course of constructon of neural network emble: tranng of the base networks, and combnaton of the traned networks. Durng the tranng phase, the man goal s to obtan networks wth acceptable accuracy whle commttng ther errors dfferently. The last crteron s commonly known as dversty (ether mplct or explct dversty). Baggng and boostng, whch operate by changng the tranng data, are the most wdely used methods to generate dverse base networks. Baggng s a name derved from bootstrap aggregaton; t s an effectve method of emble learnng ntroduced by Breman [6]. The method uses bootstrap samplng to generate multple data sets from the orgnal tranng data, and then each of these data sets s used to tran a specfc model. The output of the emble s obtaned by averagng the outputs of all the models (for regresson) or through votng (for classfcaton). It s worth notng that baggng s more effectve on unstable (.e. a small change n the tranng set can cause a sgnfcant change n the model) models [6] such as neural networks, regresson trees, etc... Moreover, Optz and Macln [7] compared baggng and two boostng methods: they concluded that, as a general method, baggng s the most approprate. As a result of that, baggng s consdered n ths paper. It s worth mentonng that n the lterature, other methods whch operate dfferently from baggng and boostng are also reported, wth some as follows. Krogh and Vedelsby [8] use cross-valdaton technque to generate several base networks. Optz and Shavlk [9] use genetc algorthm wth accuracy and ambguty as search crteron (ftness) to generate dverse and accurate base networks for classfcaton. In [10], Yao and Lu evolve a populaton of neural networks and consder the ndvduals n the last generaton as base networks. In [11], Zh-Hua Zhou et al. employ an approach named GASEN whch frst trans the base networks usng bootstrap replcates of the orgnal tranng data, after that t assgns random weghts to those base networks and uses GA to evolve the weghts. At the end, base networks wth weghts above a desgned threshold are selected to form the emble. Emprcal results show that ths method compares favorably wth other popular emble methods. Havng a gven number of base networks at hand, the next step s to combne them. The most wdely used methods for ths task are smple averagng or weghted sum [1, 13] for regresson problems, and votng for classfcaton problems [1]. Other methods for combnng base networks are also reported n the lterature [14, 15, 16]. In ths paper, as the Copyrght (c) 013 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

2 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 10, Issue 3, No, May 013 ISSN (Prnt): ISSN (Onlne): problem at hand s a regresson task and the focus s not on combnaton method, the output of the emble s obtaned by averagng (smple) the predctons of the selected base networks. It s worth mentonng that, although the general practce n most emble methods s to consder all base networks as component networks of the emble, Zh-Hua Zhou et al. [11] have demonstrated the beneft of consderng several base networks nstead of all. In ths paper, a smlar approach s adopted. However, to select the most effectve base networks, a new method s proposed. The method uses a mult-objectve genetc algorthm (GA) wth two objectves: emble error and the dversty metrc (VIF), to select approprate members of the emble from a pool of traned neural networks. Classcal method of combnng multple objectves generally requres normalzaton of the ndvdual objectves to get the fnal objectve [9].Ths encourages us to use MOGA, although, at the last generaton, t requres selectng one soluton from the Pareto-optmal front. Indeed all the solutons are optmal. The proposed method and some other approaches were evaluated on data from an electronc nose (E-Nose) for concentraton estmaton of four ndoor ar pollutants (formaldehyde, benzene, toluene, and carbon monoxde). An electronc nose s an artfcal olfacton system whch uses a fnte number of partally selectve sors along wth assocated crcutry and a sutable sgnal processng system. Electronc nose systems fnd applcaton n many felds whch nclude ndustral hazards montorng, homeland securty, food qualty, publc health, and envronmental polluton. Owng to ther versatlty and ease of use, these systems can be a better alternatve to conventonal methods (gas chromatography, mass spectrometry) for contnuous real-tme montorng and control of ndoor ar qualty. However, ther performance depends on the calbraton model generally bult usng some pror measurements. Ths s the ratonale behnd usng data from such an mportant nstrument to evaluate the methods consdered n ths paper.. Expermental Detals.1. Data sets generaton Our E-nose conssts of eght sors: two auxlary sors (temperature and humdty module), and sx gas sors (GSBT11, TGS60, TGS60, dual sor TGS01 wth two outputs named TGS01A and TGS01B, and one O sor). These sors are mounted on a self-made prnted crcut board (PCB), along wth assocated crcutry. An analog-dgtal converter (AD) s used as nterface between the FPGA processor and the sors. Also, an addtonal flash memory s used for real-tme data storage. The samplng rate durng data acquston was one pont every three seconds. For further processng, the saved data can be transferred to a personal computer (PC) usng Nos II IDE and the Jont Test Acton Group (JTAG) cable. Fgure 1 shows our electronc nose system. Fg. 1. Photography of the mplemented E-nose (left), and nsde components (rght), All experments were carred out n an atmospherecontrolled chamber by exposng our E-nose to four gas analytes each at dfferent concentratons. Detaled descrpton of the expermental setup and procedure can be found n our prevous publcatons [17,18]. However, to make the paper self-contaned, we reproduce the expermental setup (see Fgure ). As for the expermental procedure t s worth notng that durng all the experments, the respectve ranges of the temperature and humdty were and 5-80%. Also, a sngle experment conssts of three phases: exposure to clean ar for 10s to stablze the sors, exposure to gas analyte for 480s, and another exposure to clean ar for 10s to allow the sors recover. Between any two consecutve experments, the chamber s cleaned for about 10mns to avod (mnmze) nterference from any chemcal remnant. It s worth mentonng that the real concentraton of benzene was determned by gas chromatography method; whle that of formaldehyde was determned usng two dfferent methods: acetylacetone spectrophotometrc method for concentratons greater than 0.5ppm, and the 3-methyl--benzothazolnone hydrochlorde (MBTH) method for concentratons less than 0.5ppm. Ths s the am of usng organc gas sampler. For the other gases, standard measurement equpments were placed nsde the chamber and dsplayed concentratons were recorded. The number of measurements for formaldehyde, benzene, toluene, and carbon monoxde s 16, 7, 66, and 58, at concentraton ranges of ppm, ppm, ppm, and 4-55ppm, respectvely. Copyrght (c) 013 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

3 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 10, Issue 3, No, May 013 ISSN (Prnt): ISSN (Onlne): Thus, for each gas analyte, an orgnal data set s obtaned, whch contans raw measurements from the sor array TGS Feature Sor Response TGS01B Temperature Sor 1500 Humdty Sor Sample Index Fg. 3. Responses of two gas sors when exposed to 5.30ppm of formaldehyde, at 45, 50% RH (humdty and temperature sors responses are shown for nterpretaton) Fg.. Expermental setup Pror to feature extracton, raw measurements are fltered to remove measurement nose. For gas concentraton estmaton, Szczurek et al. [19] demonstrated that features from the steady-state porton of a gas sor response are more nformatve. Takng ths nto account, we selected one feature from that porton (see Fgure 3). For the auxlary sors (temperature, humdty) we selected features at the same tme postons wth other sors. The extracted features are normalzed to have values n the nterval [0, 1]. Havng an array of eght sors, an 8 m (m s the number of measurements or samples) feature data matrx s formed for each data set. Then we used Kennard and Stone (K-S) algorthm [0] to dvde each data set nto three sub data sets: 50% for tranng, 5% for valdaton, and 5% for test.. Mult-objectve genetc algorthm based selectve emble Evolutonary technques can be set to optmze sngle objectves or multple objectves. The goal of multobjectve optmzaton (MOO) s to fnd solutons that are optmal, or at least acceptable, accordng to all crtera smultaneously. The most prmtve form of MOO s to combne multple objectves nto a scalar ftness functon. And the smplest form of ths combnaton s a (scaled) lnear combnaton of the dfferent objectves. Copyrght (c) 013 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

4 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 10, Issue 3, No, May 013 ISSN (Prnt): ISSN (Onlne): Another sound alternatve to the approach mentoned above s to keep the objectves apart. In fact, the man motvaton for keepng the objectves apart s to encourage dversty among solutons, whch encourages us to adopt the alternatve. It s worth mentonng that a key dea n MOO s the noton of Pareto domnance. For nstance, gven a set of solutons S, a soluton a s non-domnated f and only f there s no other alternatve a j S, j so that a j s better than a on all crtera. Or, expressng the opposte relaton less formally, a soluton s sad to Pareto domnate another soluton f t s as good as that soluton on all objectves and better on at least one objectve. Ths results n a partal orderng, where several solutons can be nondomnated, and thus consttute the set of best solutons for the partcular set of objectves. The set of all nondomnated solutons n the search space are called the Pareto front, or the Pareto optmal set. In [6], t has been ponted out that mplct dversty can be acheved through baggng. However, n practcal applcatons where lmted number of samples s avalable, ths dversty s not guaranteed. We therefore used multobjectve genetc algorthm (MOGA) to further optmze the derved emble. More specfcally, the man dea behnd our approach s to consder many networks traned usng baggng algorthm and keep a subset of the networks that are both accurate and dverse. Genetc algorthms are effectve n ther use of global nformaton [1]; they allow us to consder a wde varety of networks durng our search, so they are sutable for our search method. The Mult- Objectve Genetc Algorthm functon gamultobj n MATLAB was used. Each gene n the GA s a bt strng of length L (the number of ANN generated usng baggng), where a 1 n any locaton ndcates that the ANN wth correspondng ndex should be ncluded n the emble...1 Varance nflaton factor as dversty metrc In the course of emble constructon dversty s one of the most mportant crtera. However, most of the dversty metrcs are drectly applcable to classfcaton embles rather than regresson embles. In ths paper, we explore the possblty of usng varance nflaton factor (VIF) as dversty metrc n selectng approprate members of neural network embles. For comparson purpose, we also evaluate an exstng method named GASEN. In regresson analyss, multcollnearty between ndependent varables can affect the varance of the estmated regresson coeffcents severely. Par-wse correlatons between predctors, t-tests and F-test, are some of the common methods used for detectng multcollnearty. Another method whch s favored by many regresson analysts s the use of varance nflaton factors (VIF). Varance nflaton factor s a statstcal measure that quantfes the severty of multcollnearty of the th ndependent varable wth the other ndependent varables, n regresson analyss. More specfcally, t quantfes how much the varances of the estmated regresson coeffcents are nflated. In [], Greene derved the varancecovarance matrx of the regresson coeffcents as: T 1 σ( bb j) = σε ( XX ) (1) where X s an n by k +1 matrx wth the frst column consstng of ones and the next k columns consstng of the σ T ε values of k ndependent varables, X s the transpose of X, and σ s the populaton varance of the resduals. ε Based on Eq. (1), Robert [3] derved an equaton that provdes the unbased estmate of the varance of the th regresson coeffcent as: ˆ ( b ) σ (1 R ) ( Y Y) ( n k 1) = (1 R ) x y where n s the sample sze, k the number of ndependent varables n the regresson analyss, x are the ndependent varables (the last k-1 columns of X n Eq. (1)), Y s the dependent varable (generally one dependent varable s used), R s the proporton of the varance n the th ndependent varable that s assocated wth the other ndependent varables n the analyss. R s the squared multple correlaton of the dependent varable regressed on all other ndependent varables n the analyss. Rearrangng Eq. (), we get, ˆ ( b ) σ y (1 Ry ) ( Y Y) ( n k 1) 1 = x (1 R ) The second term on the rght hand sde of Eq. (3) represents the VIF of the th ndependent varable, and t ndcates the multplcatve ncrease n the varance of the regresson coeffcent of ths varable [3]. It s worth mentonng that the varance nflaton factor VIF of the th ndependent () (3) Copyrght (c) 013 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

5 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 10, Issue 3, No, May 013 ISSN (Prnt): ISSN (Onlne): varable can be found by regressng t on the other ndependent varables. In our case, the ndependent varables are the valdaton errors of the emble members (.e. the base networks). Thus, for the n th base network VIF n s defned as follows. VIFn = 1 (1 R n ) Where VIF n s the varance nflaton factor for the n th network, and R n s the R value obtaned by regressng the n th ndependent varable(the valdaton errors of the n th network) on the remanng ndependent varables (.e. the valdaton errors of the others networks). There are many rules of thumb assocated wth VIF that are regarded as a sgn of severe multcollnearty. The most commonly used s the rule of 10, that s f VIF n >10, the n th network has serous multcollnearty wth the other emble members, otherwse there s less or even no multcollnearty [3]. In ths paper, nstead of consderng any threshold value, we try to mnmze the sum of all VIF n. The value of ths sum s consdered as our second objectve n secton.3. For detaled dscusson on VIF we refer nterested readers to [3, 4]... Tranng Component Networks For each data set, bootstrap samplng was used on the orgnal tranng data to generate 50 new tranng data. Each of these new tranng data s then used to tran a component network (or base network), usng back-propagaton algorthm wth dfferent ntal weghts and wth earlystoppng opton (on the valdaton data). Early stoppng s a method to mprove the generalzaton capablty of ANN n case of small sze tranng data; the tranng s stopped when the error on the valdaton set has reached a certan threshold. Ffty sngle-hdden-layered base networks wth smlar structure (8:5:1, that s 8 nput neurons, 5 hdden neurons, and one output neuron) were traned. These networks consttute the pool of base networks on whch we appled two selecton based methods: GASEN, and MOGASEN. For the standard baggng method, all the base networks are consdered as component networks of the emble, and the output of the emble s obtaned by averagng the outputs from these base networks; whereas only outputs from the selected networks are consdered n other methods. (4)..3 Best Ensemble Selecton Two selecton-based methods are used to select the best emble: GASEN, and MOGASEN. For GASEN method, default settngs specfed by the authors [5] were used, except for the number of generatons and the populaton sze. For MOGASEN method, MATLAB mplementaton of MOGA was used. Table 1 shows settngs of some mportant parameters n MOGA; default settngs were used for other parameters. Table 1: MOGA parameters settng Parameter Value/Scheme Populaton Sze 50 Populaton type Bt Strng Number of varables 50 Generatons 50 Mutaton probablty 0. Crossover probablty 0.8 Selecton Tournament MOGASEN based selecton process s performed through four steps as descrbed below. Indvdual encodng: To solve our optmzaton problem, a soluton s frst encoded to chromosome form, the sze of the search space s the same as the number of prmary base networks, K. Bnary encodng scheme s used, wheren 0 means the base network s excluded and 1 means the base network s selected. For nstance, f chromosome C = (when K = 8) means that the base learners #1, #3, #5, #7, and #8 are selected as members of the emble (h = 5). Intal populaton: The ntal populaton s randomly generated. Objectve functons: Two objectve functons were used, the emble error and the dversty metrc (VIF). More specfcally, let s call these objectve functons f 1, and f, respectvely. Then we can defne them as follows. f 1 N 1 ( y y ) = 100 (5) N y = 1 where N s the number of valdaton samples, y s the actual value of the th sample and y s ts predcted value by the emble. It s worth mentonng that s the average value of the outputs from all the selected base y Copyrght (c) 013 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

6 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 10, Issue 3, No, May 013 ISSN (Prnt): ISSN (Onlne): networks. For nstance, f we have K base networks obtaned usng Eq. (6). y s y 1 K K k = 1 k = y (6) k where y s the output from the k th network for the th sample. f K = VIF (7) n= 1 n where K s the number of base networks n a potental soluton, and VIF n s the VIF of the n th base network, as defned n Eq. (4). Genetc operatons: In standard GA, three operatons are generally nvolved: selecton, crossover, and mutaton. Durng the selecton step, chromosomes wth hghest ftness values are chosen as parents. From these parents, canddates (chldren, offsprng) are generated usng crossover and mutaton operatons. The algorthm calculates a ftness score for each canddate and replaces chromosomes wth low scores by new canddates wth hgh scores. Ths process s repeated untl stoppng condtons (maxmum number of generatons, a certan value of the ftness functon, etc ) are satsfed. The proposed approach s summarzed n Table, where D Tr s the orgnal tranng data; D V s the valdaton data used durng MOGA based base networks selecton, D Te s the test data whch s only used after emble constructon, and B s the number of bootstrap replcates whch s also equal to the number of ntal base networks (n our case 50 base networks). Base networks generaton: Table : MOGASEN method 1) Use bootstrap samplng on D Tr to generate B new tranng sets ) Use each of the new tranng sets obtaned n 1) to tran a base network usng BP wth early-stoppng (aganst the valdaton data) Best emble selecton: 3) Randomly generate an ntal populaton 4) Use MOGA wth two objectve functons as defned n Eqs. (5) & (7) to evolve the ntal populaton. At termnaton, trace out the soluton wth the smallest f on D V and report t as the selected emble. 5) Evaluate the emble selected n 4) on D Te, compute the test error usng Eq. (5) wth N as the number of test samples nstead. 3. Results and Dscusson All computatons were carred out usng MATLAB R010a (MathWorks Inc.) software on a desktop computer wth Intel(R) Core(TM) 3 T GHz CPU, GB RAM and Wndows XP professonal operatng system. It s worth mentonng the followng notatons: BNN for the best base network, Baggng for the standard baggng method, GASEN for GASEN method, and MOGASEN for our method. To avod based comparson, for each method and each data set we perform ten runs and recorded the averages of mean absolute percentage errors(mape) for the selected embles as well as for the best base networks. Expermental results are reported n Tables 3 and 4. Also, for comparson purpose, results of standard baggng are shown. When constructng emble of predctors, the capablty of a method n reducng the sze of an emble whle mantanng or even mprovng ts performance s also of great mportance. Havng ths n mnd, we computed the average number of selected base networks over ten runs, for GASEN and MOGASEN. Results are reported n Table 5. Copyrght (c) 013 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

7 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 10, Issue 3, No, May 013 ISSN (Prnt): ISSN (Onlne): Table 3: Averaged valdaton errors over ten runs Data sets Methods BNN Baggng GASEN MOGASEN Formaldehyde (31.45)(40.073)(35.955) a Benzene (8.3587)(8.4771)(8.593) Toluene (1.55)(1.55)( ) Carbon monoxde (15.314)( )( ) a Numbers n parentheses are errors of best component networks for baggng, GASEN, MOGASEN, n order Table 4: Averaged test errors over ten runs Data sets Methods BNN Baggng GASEN MOGASEN Formaldehyde ( )( )( ) a Benzene (7.9458)(14.033)( ) Toluene (13.54)(16.536)( ) Carbon monoxde ( )(18.136)( ) a Numbers n parentheses are errors of best component networks for baggng, GASEN, MOGASEN, n order Table 5: Average number of selected nets over ten runs Data sets Methods GASEN MOGASEN Formaldehyde 5 5 Benzene 7 5 Toluene 5 5 Carbon monoxde 5 5 From Tables 3 and 4 one can notce that the theory of many could be better than all was verfed. In most cases, GASEN and MOGASEN outperformed the standard baggng and ther correspondng best base networks, wth unque excepton on benzene. The only case where GASEN performs better than MOGASEN on both valdaton and test data sets s wth toluene. An ntutve remark from these results s that an emble less effectve than the best component network on a gven data set (here, valdaton data set) may perform well on a novel data set (e.g. GASEN on formaldehyde data set). Also, results from carbon monoxde data set nfer that there were too many redundant networks n the ntally generated pool of base networks. Ths resulted n poor performance of standard baggng method whch s normally known to be effectve on component networks that are suffcently accurate and dverse. Indeed, beng the smallest data set wth almost 60 samples, usng smple bootstrappng and dfferent ntal weghts on such small data set to tran base networks was nsgnfcant for generatng dverse networks. Ths s n perfect agreement wth results obtaned n our recent work [6], where base networks wth dfferent topologes were even used. GASEN tends more to selectng best networks than dverse ones. Ths can be seen wth benzene and toluene data sets. A possble reason for ths s the value of the threshold λ. Settng the threshold to hgh values wll cause the algorthm to only emphasze on accurate base networks, whle settng t to small values wll result n selecton of naccurate networks. Results from Table 5 show that GASEN and MOGASEN selected almost the same number of component networks over ten runs, except wth benzene data set where the average number of selected networks by GASEN s superor to that selected by MOGASEN. By selectng more component networks n the case of benzene, GASEN mght have overftted the valdaton data, thereby resultng n lower performance on test data compared to MOGASEN. Another mportant remark s that, as both GASEN and MOGASEN are off-lne selecton methods, one may suspect that component networks selected by these methods are n fact smlar. However, results from Tables 3 and 4 evdenced that ths rarely happened n practce. 4. Concluson Dversty metrcs (par-wse as well as non-parwse) play an mportant role n emble learnng. In ths paper, a new non-parwse dversty metrc based on varance nflaton factor s proposed. A mult-objectve GA wth two objectves (emble error and the new dversty metrc) s Copyrght (c) 013 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

8 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 10, Issue 3, No, May 013 ISSN (Prnt): ISSN (Onlne): used to select best neural network embles for concentraton estmaton of some ndoor ar pollutants. Emprcal results show that varance nflaton factor can effectvely be used as dversty metrc. Although the proposed method can outperform GASEN (another selecton based method), standard baggng, and the best component network; more study on VIF for multcollnearty measurement s requred to further mprove the method. Ths method s not restrcted to electronc nose data; t can be appled n other felds. Also, t can be extended to classfcaton problem. In ths paper small sze data sets were used, we therefore need to evaluate ths method on large-scale data sets (for both classfcaton and regresson); ths wll be consdered n our future work. Acknowledgments The authors would lke to acknowledge fnancal supports from the Key Scence and Technology Research Program (CSTC010AB00, CSTC009BA01), the Chongqng Unversty Postgraduate Scence and Innovaton Fund (CDJXS ), and the Fundamental Research Funds for the Central Unverstes of Chna (CDJXS ). Specal thanks to the anonymous revewers for ther valuable comments and suggestons. References [1] L.K. Hansen, P. Salamon, Neural network embles, IEEE Trans. Pattern Analyss and Machne Intellgence, Vol. 1, No.10,1990,pp [] Z.-H. Zhou, Y. Jang, Y.-B. Yang, S.-F. Chen, Lung cancer cell dentfcaton based on artfcal neural network embles, Artfcal Intellgence n Medcne, Vol. 4, No.1, 00, pp [3] G. Daq, C. We, Smultaneous estmaton of odor classes and concentratons usng an electronc nose wth functon approxmaton model embles, S. Actuators, B Chem. Vol.10, 007, pp [4] Evandro Bona, et al., Optmzed Neural Network for Instant Coffee Classfcaton through an Electronc Nose, Internatonal Journal of Food Engneerng: Vol. 7, Issue 6, 011. [5] L.K. Hansen, L. Lsberg, P. Salamon, Ensemble methods for handwrtten dgt recognton, n Proceedngs IEEE Workshop on Neural Networks for Sgnal Processng, 199, pp [6] L. Breman, Baggng predctors, Machne Learnng, Vol. 4, No., 1996, pp [7] Optz, D. and Macln, R., Popular emble methods: An emprcal study, Journal of Artfcal Intellgence Research, Vol. 11, 1999, pp [8] A. Krogh, J. Vedelsby, Neural network embles, cross valdaton, and actve learnng, Advances n Neural Informaton Processng Systems, Vol. 7, 1995, pp [9] D.W. Optz, J.W. Shavlk, Generatng accurate and dverse members of a neural network emble, Advances n Neural Informaton Processng Systems, Vol. 8, 1996, pp [10] X. Yao, Y. Lu, Makng use of populaton nformaton n evolutonary artfcal neural networks, IEEE Transactons on Systems, Man and Cybernetcs - Part B: Cybernetcs Vol.8, No. 3, 1998, pp [11] Zh-Hua Zhou et al., Ensemblng Neural Networks: Many Could Be Better Than All, Artfcal Intellgence, Vol.137, No.1-, 00, pp [1] R. J. Mammone, Artfcal Neural Networks for Speech and Vson, London: Chapman & Hall, pp.16-14, [13] D.W. Optz, J.W. Shavlk, Actvely searchng for an effectve neural network emble, Connecton Scence Vol.8, No.3-4, 1996, pp [14] C.J. Merz, M.J. Pazzan, Combnng neural network regresson estmates wth regularzed lnear weghts, Advances n Neural Informaton Processng Systems, Vol. 9, 1997, pp [15] D. Jmenez, Dynamcally weghted emble neural networks for classfcaton, n Proceedngs IJCNN-98, vol. 1, 1998, pp [16] N. Ueda, Optmal lnear combnaton of neural networks for mprovng classfcaton performance, IEEE Transacton Pattern Analyss and Machne Intellgence, Vol., No., 000,pp [17] L. Zhang et al., Classfcaton of multple ndoor ar contamnants by an electronc nose and a hybrd support vector machne, Sors and Actuators B, Vol.174, 01, pp [18] L.Zhang et al., Chaos based neural network optmzaton for Concentraton Estmaton of Indoor Ar Contamnants, S. and Actuat. A, Vol.189, 013, pp [19] A. Szczurek et al., The stop-flow mode of operaton appled to a sngle chemresstor, Sors and Actuators B, Vol. 148, 010, pp [0] R.W. Kennard, L.A. Stone, Computer aded desgn of experments, Technometrcs, Vol. 11, 1969, pp [1] Goldberg, D. E. Genetc Algorthms n Search, Optmzaton, and Machne Learnng, Addson- Wesley Pub. Co., [] Greene, W. H., Econometrc Analyss, nd ed. Macmllan, New York, [3] Robert, M. O Bren, A Cauton Regardng Rules of Thumb for Varance Inflaton Factors. Qual. Quant. 41, 007, pp [4] Kutner, Nachtshem, Neter, Appled Lnear Regresson Models, fourth ed. McGraw-Hll, Irwn, 004. [5] Zh-Hua Zhou, et al., Genetc Algorthm based Selectve Neural Network Ensemble, n Proceedngs of the 17th Internatonal Jont Conference on Artfcal Intellgence, 001, vol., pp [6] Chabou Kadr, Fengchun Tan, Le Zhang, Guojun L, Ljun Dang and Guoru L, Neural Network Ensembles for Onlne Gas Concentraton Estmaton Usng an Electronc Nose, Internatonal Journal of Computer Scence Issues, Vol. 10, Issue, No. 1, 013, pp Copyrght (c) 013 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.