文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Global optimisation Evolutionary neural networks Evolutionary algorithms Genetic algorithms
侵权投诉

Global optimisation Evolutionary neural networks Evolutionary algorithms Genetic algorithms

developments and future directions of bio-inspired

ECOLOGICALINFORMATICS1(2006)9–22

11

Unreasonableinitialvaluescouldcausethefittingprogramtotrapinalocalminimum.

2.2.Twostageevolutionaryapproach(Lietal.,2004)

Theevolutionaryapproachconsistsoftwomajorsteps:(i)thefirststepaimstoseekpromisingmathematicalfunctionsusinggeneticprogramming(GP).(ii)Thesecondstepisaimedatfittingparametersinthefunctionformulathusfound,usingevolutionaryprogramming(EP)techniques.

FirstweintroduceGP(KozaandAndre,1999)usedinthemodelling.InthecaseofGPhere,wetakeafunctionset,F={+, ,*,/,exp,log,sin,cos}andaterminalset,T={r,R}whereristhevariableradiusandRisarandomfloat-pointvaluebetween 10.0and10.0.Theoretically,allprofilemodels(1)–(3)couldbepotentiallygeneratedbyGP.

Thefitnessfunctiondetermineshowwellanindividualexpressionfitsanobservationalgalaxyprofile.TheoverallfitnessfunctionusedhereforGPconsistsoftwoitems.Wetakehitsasthefirstpartofthefitnessfunction.Thehitsmeasureherecountsthenumberoffitnessdatapointsforwhichthenumericalvaluereturnedbythemodelexpressionlieswithinasmalltolerance(thehitscriterion)oftheobservedintensityvalue.Forexample,thehitscriteriontakeninthispaperis0.005forGP.Selectinganappropriatevalueforthehitcriterionseemsabittrickyhere.TheaimofusingGPistofindpromisingfunctionalformsthatpotentiallyhavecapabilityfordescribingasmanyprofilesaspossible.Therefore,inourapproach,takingahigherhitcriterionforGPispreferableaslongasthemaximumnumberofhits(i.e.,50inthisstudy)isachievableintermsoftheresultson18galaxiesprofiles.

Inaddition,weaddaseconditemintothefitnessfunctiontopenalizeanycomplexmathematicalexpressionbecauseasimplefunctionalformisalmostalwayspreferable.Thecom-plexityofanexpressionismeasuredbythelengthoftheexpression,i.e.,thenumberofnodeswithintheexpression.Theextenttowhichanexpressionshouldbepenalizedischangeablebyaweightw.Insummary,thefitnessfunctionthatweuseisgivenby

f¼HitsÀwÃthelengthoftheexpression:

ð5Þ

ThecriteriaforterminatingaGPrunareeitherthemaxi-mumruntimeallowedorthemaximumofgenerationthatwe

set,whicheverreachedfirst.

Inthesecondstep,EPisusedtotuneparametersintheform.Inthisstudy,theEPalgorithmissimilartotheoneusedinYaoetal.(1999),exceptthatself-adaptivemutationisnotused.CauchymutationisusedintheEP.Thefitnessfunctionisdefinedbythehits,thefirstitemofthefitnessfunctionforGP.Theterminationcriteriontakenhereisthatthemaximumgenerationhasbeenreached.

2.3.Experimentsandresults(Lietal.,2004)

Thedataareasetof18ellipticalgalaxiesintheComaclusterobservedinnear-infraredband(centralwavelength2.2mm;forfurtherdescriptionoftheobservationsseeKhosroshahietal.,2000).Theyarealsochosenfromthesameclusterofgal-axiestoenableustoeliminatethesignatureofdifferentenvir-

onmentsonthesurfacebrightnessofthesamplegalaxies.EllipticalgalaxieshistoricallyhavebeencharacterizedbythedeVaucouleurs(r1/4)profile,whichisanempiricalfit,notyetshowntoemergefromtheoreticalorsemi-analyticscenariosofgalaxyformation.Carefulobservationsrevealafargreatervarietyofprofiles,leading,amongothers,Khosroshahietal.(2000)toadvocatetheuseofgeneralizedmodelsthatallowforshallowerorsteeperprofilestoellipticalgalaxiesandbulgesofdiskgalaxies.

One-dimensionalradialprofileshavebeenextractedfromfittingellipticalcontoursofthesamebrightness(isophotes)totheimagesofthegalaxies,andcomputingtheaveragebright-nessasafunctionofradiusalongthemajoraxisoftheellipseatvariousvaluesoftheradialdistancer.Eachprofilehas50datapoints.Theskybackgroundishighlyvariableatnear-infraredwavelengths.Thereforethelocalbackgroundishigh-lyvariable.

Theexperimentswerecarriedoutasfollows.WefirstrantheGPsystemoneachof18galaxyprofiledatasets.ThedetailsofparametersettingsforrunningGParegiveninthepaper(Lietal.,2004).

OneprofilemodeliscreatedaftereachGPrun.18differentprofilemodelswithdifferentmathematicalformsaregener-atedintotal.Theyseemdifferentintheirshapesoftreestruc-ture.Therefore,findingsimilaritybetweenthoseformsgeneratedisneededinordertofindamoregenericmodelform.

Wesimplifyatree-structuredfunctionforminthewayofcuttingeachpossiblesub-treeandreplacingitbyonesinglenumericalnodeiftheultimatereturnvalueofthesub-treeisanumericalvalue.Afterthesimplification,theformhastobegeneralizedinthewayofreplacingafewnumericalvaluesbyvariables(e.g.,a,b,andc),respectively.

Afterthegeneralizationprocess,thereemergeafewformswithparametersthatarerelativelysimple.Amongtheseformsweonlyselectthefollowingtwosimplestfunctionformsfg1andfg2forfurtherinvestigation.fg1¼aþb=ðcþrÞ;ð6Þfg2¼aVþbV=ðbVþcVr2Þ:

ð7Þ

Eachofthetwoformsinvolvesthreeparametersi.e.,a,b,andcforfg1;a′,b′andc′forfg2,andoneradiusr.Now,weuseEPtoseekpropervaluesforthreeparameterswithinthebothfunctionsrespectivelytofiteachgalaxyprofileinthelogarithmicregime.TheparametersettingsforrunningEParegiveninthepaper(Lietal.,2004).

Forbrevity,resultsoffittingparametersfoundbyEPforfg1andforfg2arebothlistedinTable1.

Severalpointsareworthindicatinghere.Firstofall,EPalgorithmsalmostalwaysachievethemaximumhits:50(thenumberofdatapointsforeachprofile),basedoneitherfunctionformsforeachgalaxyprofile.Thisispartlyattrib-utedtoalargepopulationofsolutions.Secondly,unlikenon-linearfittingalgorithms,EPisnotsensitivetotheinitialvaluesprovidedatall.Throughoutourexperimentswekeeptheinitialvaluesforthoseparameters(i.e.,a,b,andc;a′,b′andc′),whichdonotneedtobetunedinordertoavoidfailureofarun.ThisisattributedtothemutationinEP.

上一页第3页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email