A clustering guided ant colony optimization algorithmservice selection problemWord下载.docx

A clustering guided ant colony optimization algorithmservice selection problemWord下载.docx

《A clustering guided ant colony optimization algorithmservice selection problemWord下载.docx》由会员分享，可在线阅读，更多相关《A clustering guided ant colony optimization algorithmservice selection problemWord下载.docx（14页珍藏版）》请在冰豆网上搜索。

Abstract

TotacklethelargescaleQoS-basedserviceselectionproblem,anovelefficientclusteringguidedantcolonyserviceselectionalgorithmcalledCASSisproposedinthispaper.Inthisalgorithm,askylinequeryprocessisusedtofilteringthecandidatesrelatedeachserviceclassandaclusteringbasedshrinkingprocessisusedtoguidetheantsearchdirections.Weevaluateourapproachexperimentallyusingstandardrealdatasetsandsyntheticallygenerateddatasets,andcomparedwiththerecentlyproposedrelatedserviceselectionalgorithms.Itrevealsveryencouragingresultsintermsofthequalityofsolution,andtheprocessingtimerequired.

Keywords:

antcolonyoptimization;

serviceselection;

clustering.

1.Introduction

WiththeproliferationoftheCloudComputingandSoftwareasaService（SaaS）concepts,moreandmorewebserviceswillbeofferedonthewebatdifferentlevelsofquality[2].Theremaybemultipleserviceproviderscompetingtoofferthesamefunctionalitywithdifferentqualityofservice.QualityofService（QoS）hasbecomeacentralcriterionfordifferentiatingthesecompetingserviceprovidersandplaysamajorroleindeterminingthesuccessorfailureofthecomposedapplication.Therefore,aServiceLevelAgreement（SLA）isoftenusedasacontractualbasisbetweenserviceconsumersandserviceprovidersontheexpectedQoSlevel.TheQoS-basedserviceselectionproblemaimsatfindingthebestcombinationofwebservicesthatsatisfyasetofend-to-endQoSconstraintsinordertofulfillagivenSLA,whichisaNP-hardproblem[1].

Thisproblembecomesespeciallyimportantandchallengingasthenumberoffunctionally-equivalentservicesofferedonthewebatdifferentQoSlevelsincreasesexponentially[8].Asthenumberofpossiblecombinationscanbeveryhuge,basedonthenumberofsubtaskscomprisingthecompositeprocessandthenumberofalternativeservicesforeachsubtask,usingtheproposedexactsearchalgorithms[6][17]toperformanexhaustivesearchtofindthebestcombinationthatsatisfiesacertaincompositionlevelSLAisimpractical.So,themostresearchesareconcentratedonheuristic-basedalgorithmsespeciallythemeta-heuristicapproachesaimingtofindnear-optimalcompositions.In[17],theauthorsproposeheuristicalgorithmsthatcanbeusedtofindanearoptimalsolutionmoreefficientlythanexactsolutions.TheauthorsproposetwomodelsfortheQoS-basedservicecompositionproblemandintroduceaheuristicforeachmodel.In[5],theauthorspresentageneticalgorithmforthisproblem,includingthedesignofaspecialrelationmatrixcodingschemeofchromosomes,evolutionfunctionofpopulationandpopulationdiversityhandlingwithsimulatedannealing.In[18],anewcooperativeevolution（Co-evolution）algorithmconsistsofstochasticparticleswarmoptimization（SPSO）andsimulatedannealing（SA）ispresentedtosolvethisproblem.

Asameta-heuristicapproach,theACOalgorithmisdefinedbyM.Dorigo[12],motivatedbytheintelligentbehaviorofantsystem.Ithasbeenappliedtosolvemanyproblemsandobtainedsatisfyingresults[3][14].Inthispaper,theACOalgorithmisextendedforsolvingtheQoSbasedserviceselectionproblem.Inthisalgorithm,anunsupervisedclusteringprocessisusedforconstructingadirectedclusteringgraphtoguidetheantsmakingexploration,andadynamicexpandingprocessisusedtoenlargethispathforantsmakingexploitationbasedontheobtainedglobalinformation.Furthermore,theMulti-criteriaDominanceRelationships[7]isintroducedtoreducethesearchspaceaswellasant-basedclustering[11]tofurtherimprovetheserviceselectionefficiency.WehavecomparedourapproachwiththerecentlyproposedserviceselectionalgorithmDiGA[5]andSPSO[18].Theperformanceofthesealgorithmshasbeentestedonavarietyofdatasetsprovidedfromseveralreal-lifesituationsandsyntheticallygenerateddatasets.Thecomputationalresultsdemonstratetheeffectivenessofourapproachincomparisontotheseapproaches.Thispaperisorganizedasfollows.InSection2,wegivethedefinitionoftheQoS-basedserviceselectionproblemandthebasicantcolonyalgorithm.TheCASSalgorithmincludingitsmodelandconcretealgorithmdescriptionisprovidedinsection3.Section4presentexperimentalstudiesandcomparedtheCASSwithsomeotherrecentlyproposedalgorithms.Finally,Section5summarizesthecontributionofthispaperalongwithsomefutureresearchdirections.

2.Problemdefinitionandantcolonyalgorithm

2.1TheQoS-basedserviceselectionproblem

Foracompositeapplicationthatisspecifiedasabstractprocessescomposedofasetofabstractservices

eachabstractservicei,i[0,||

||-1]correspondstoaserviceclassSi={si1,si2,…,sin},andSiconsistsofallservicesthatdeliverthesamefunctionalitybutpotentiallydifferintermsQoSvalues.SincetheQoSattributeswhicharepublishedbytheserviceprovider,maybepositiveornegative.WeusethevectorQs={q1（s）,q2（s）,…,qr（s）}torepresenttheQoSvaluesofservices,andthefunctionqi（s）determinesthepublishedvalueofthei-thattributeoftheservices.Then,theQoSvectorforacompositeserviceconsistingofn,n[1,||

||]servicecomponentsCS={s1,s2,…,sn}isdefinedasQCS={q1（CS）,q2（CS）,…,qr（CS）},wheretheqi（CS）istheestimatedend-to-endvalueofthei-thQoSattributeandcanbecomputedbyaggregatingthecorrespondingvaluesofcomponentservices.

Definition1.（FeasibleSelection）:

Foragivenabstractprocess

andavectorofglobalQoSconstraintsC={c1,c2,…,cm},1mr,whichrefertotheuser’srequirementsandexpressedintermsofavectorofupper（orlower）boundsfordifferentQoScriteria,weconsideraselectionofconcreteservicesCStobeafeasibleselection,iffitcontainsexactlyoneserviceforeachserviceclassSiof

anditsaggregatedQoSvaluessatisfytheglobalQoSconstraints,i.e.q1（CS）ck,k[1,m].

InordertoevaluatetheoverallqualityofagivenfeasibleselectionCS,autilityfunctionUisusedwhichmapsthequalityvectorQCSintoasinglerealvalueanddefinesasfollows:

（1）

with

beingtheweightofqktorepresentuser’spriorities,

and

（2）

beingtheminimumandmaximumaggregatedvaluesofthek-thQoSattributeforcompositeserviceCS,andFdenotinganaggregationfunctionthatdependsonQoScriteriashownasintable1.

Definition2.（ServiceSelection）:

andavectorofglobalQoSconstraintsC={c1,c2,…,cm},1mr,theserviceselectionistofindthefeasibleselectionthatmaximizestheoverallutilityfunctionUvalue.

2.2Theantcolonyoptimizationalgorithm

Innature,foragerantscommunicateindirectlybydepositingandsensingpheromonetrails.Thissetsupapositivefeedbackloopthatreinforcespromisingpaths.TheACOalgorithmisinspiredbythisbehaviorofrealants,inwhichtheartificialantscompleteaseriesofwalksofadatastructure,knownasaconstructiongraph.Theylaypheromonetrailsonthisgraphedgesandchoosetheirpathwithrespecttoprobabilitiesthatdependonpheromonetrailsandthesepheromonetrailsprogressivelydecreasebyevaporation.Inmostcases,pheromonetrailsareupdatedonlyafterhavingconstructedacompletepathandnotduringthewalk,andtheamountofpheromonedepositedisusuallyafunctionofthequalityofthepath.Furthermore,theprobabilityforanartificialanttochooseanedgeoftendependsnotonlyonpheromones,butalsoonsomeproblem-specificlocalheuristics.TheclassicalACOalgorithmisshowasAlgorithm1.

Ateachcycle,eachantconstructsacompleteassignmentandthenpheromonetrailsareupdatedincludingthepheromonedepositingandevaporating.Thefitisafitnessfunctionusedtoevaluateanassignment.WecanseethatwhenusingtheACOmeta-heuristictosolveanewcombinatorialoptimizationproblem,oneofthemaintasksistomodeltheproblemasthesearchofafeasibleminimumcostpathoveraweightedgraph,wherethefeasibilityisdefinedwithrespecttoasetofconstraints.

3.Theclusteringbasedantcolonyalgorithmforserviceselection

Obviously,foranapplicationrequestwithnserviceclassesandlcandidateservicesperclass,therearelnpossiblecombinationstobeexamined.So,whenthenumberoffunctionally-equivalentservicesofferedbecomeslarge,howtoeffectivelyshrinkingthesolutionspaceandmakethesearchquicklygorightdirectionisveryimportant.IntheCASSalgorithm,askylinequeryprocessisusedtofilteringthecandidatesrelatedeachserviceclass,andanunsupervisedclusteringprocessisintroducedtopartitiontheskylineservicesinperserviceclass.Thenadirectedclusteringgraphisconstructedbasedonclusteringresulttoabstractthesearchspace,andusedtoguidetheantsglobalsearching.

Definition3.（SkylineServices）:

TheskylineofaserviceclassS,denotedbySLS,comprisesthesetofthoseservicesinSthatarenotdominatedbyanyotherservice,i.e.,SLS={i∈S|¬

∃j∈S;

j≺i}.WeregardtheseservicesastheskylineservicesofS.

Definition4.（Dominance）:

ConsideraserviceclassS,andtwoservicesi,jS,characterizedbyasetofQofQoSattributes.idominatesj,denotedasi≺j,ifiisasgoodorbetterthanjinallparametersinQandbetterinatleastoneparameterinQ,i.e.∀k∈[1,|Q|]:

qk（x）≤qk（y）and∃k∈[1,|Q|]:

qk（x）<

qk（y）.

Sincenota