LACADFandroid

JGridComputing(2017)11f20;BACKGROUND-COLOR:#4ae2f7">5:343–31f20;BACKGROUND-COLOR:#4ae2f7">56DOI10.
1007/s10723-017-9406-2AutonomousContext-BasedServiceOptimizationinMobileCloudComputingPiotrNawrocki·BartlomiejSniezynskiReceived:9February2016/Accepted:4July2017/Publishedonline:20July2017TheAuthor(s)2017.
ThisarticleisanopenaccesspublicationAbstractAstheconceptofmergingthecapabilitiesofmobiledevicesandcloudcomputingisbecom-ingincreasinglypopular,animportantquestionarises:howtooptimallyscheduleservices/tasksbetweenthedeviceandthecloud.
Themainobjectiveofthispaperistoinvestigatethepossibilitiesforusingadecisionmoduleonmobiledevicesinordertoautonomouslyoptimizetheexecutionofserviceswithintheframe-workofMobileCloudComputingwhiletakingcon-textintoaccount.
Anovelmodelofthedecisionmodulewithlearningcapabilities,service-orientedarchitecture,andserviceselectionoptimizationalgo-rithmareproposedtosolvethisproblem.
Toachieveautonomous,onlinelearningonmobiledevices,weapplysupervisedlearning.
Informationaboutthecon-text,taskdescription,thedecisionmadeanditsresultssuchascalculationtimeorpowerconsumptionarestoredandformtrainingdataforasupervisedlearningalgorithm,whichupdatestheknowledgeusedbythedecisionmoduletodeterminetheoptimalplacefortheexecutionofagiventypeoftask.
Toverifythesolutionproposed,service-orientedmobileprocessingsystemsP.
Nawrocki()·B.
SniezynskiFacultyofComputerScience,ElectronicsandTelecommunications,DepartmentofComputerScience,AGHUniversityofScienceandTechnology,al.
A.
Mickiewicza30,30-01f20;BACKGROUND-COLOR:#4ae2f7">59Krakow,Polande-mail:piotr.
nawrocki@agh.
edu.
plB.
Sniezynskie-mail:bartlomiej.
sniezynski@agh.
edu.
plformultimediafileconversionhavebeendevelopedandseriesofexperimentshavebeenexecuted.
Resultsshowthatthedecisionmodulehasbecomemoreeffi-cientinassigningthetasktoeitherthemobiledeviceorcloudresources.
KeywordsServiceoptimization·Mobiledevice·Cloud·Context-aware·Supervisedlearning1IntroductionandMotivationInrecentyears,theimportanceofmobiledevicesincomputersystemshasincreased[1].
Thedevelopmentofmobilephones,whichatfirstwereonlymeantforvoicecommunication,hastakenaturninthedirec-tionofmulti-functiondevicesknownassmartphones,whichcombinethefunctionalitiesofbothphonesandcomputersandadditionallyincorporatevarioussensors.
Alongsidehardwaredevelopment,softwarebegantodevelopaswell;moreandmoreprogramsstartedtoleveragetheresourcesofmobiledevices,andespeciallytheCPU,memoryandbattery.
WhenitcomestoCPUandmemorycapabilities,mobiledeviceshavebecomemorelikecomputers;however,therestillaresomecharacteristicsthatdifferentiatethem,e.
g.
relativelysmallscreens(evenintablets),batterypowersupplyandcommunicationmodulesthatrelysolelyonthewirelesstechnology.
Currently,oneofthemainareasindistributedsys-temsiscloudcomputing[2].
However,otherdirections344P.
Nawrocki,B.
Sniezynskiarealsobeingexplored,likevisualtoolsformanag-ingcomputations[3]orusingnovelmiddlewarefordistributedcomputing[4].
Alongwiththedevelopmentofcloudcomputingandmobilesystems,whichareveryspecificmiddle-ware,researchintotheimplementationofthisconceptinthemobiledeviceenvironmentbegan.
OneofthemostimportantpublicationsinthisareaistheworkofKumarandLu[1f20;BACKGROUND-COLOR:#4ae2f7">5]whoresearchthepotentialenergysavingsonmobiledevicesthroughoffloadingcomputationtothecloud.
Inanotherpaper[6],theyextendtheirresearchonthepossibilityofimprovingperformancethroughoffloadingcomputa-tionfromthemobiledevicetothecloud.
However,inoffloadingframeworksthereisanadditionalover-headrelatedtocomponentmigrationatruntime.
Inpaper[7],anoveldistributedEnergyEfficientCompu-tationalOffloadingFramework(EECOF)isproposedfortheprocessingofintensivemobileapplicationsinMCC.
Thesolutionfocusesonleveragingapplicationprocessingservicesincloudcomputingwithmini-malinstancesofcomputationallyintensivecomponentmigrationatruntime.
ThroughtheuseofEECOF,theamountsofdatatransmittedandenergyconsumedarereducedincomputationaloffloadingforMCC.
Onthebasis,interalia,oftheresearchdescribedabove,theMobileCloudComputing(MCC)paradigm[8],whichenablesthemigrationofindividualservices(tasks[9],data)frommobiledevicestothecloud[10],emerged.
OnthebasisofthegeneralMCCconcept,open(e.
g.
OpenMobster,Clonecloud[11])andcom-mercial(e.
g.
PerfectoMobile)implementationsofthesolutionhavebeendeveloped.
Manysamplescenariosandapplicationsusingmobiledevicesandcloudcom-putinghavebeendescribedinliterature[12].
AuthorsDinhetal.
,in[13],haveconductedasurveyoftypi-calMCCapplications,suchasmobilecommerce[14],mobilelearning[11f20;BACKGROUND-COLOR:#4ae2f7">5],mobilehealthcare,allowingeasyandeffectiveaccesstothepatients'medicalrecords[16]orenablingthemonitoringofthepatients'con-ditionwhileathome[17],andalsomobilegaming,enablingtheusertoplaywhilesimultaneouslysend-ingcertaingametasksthatrequirethemostcomputingresources(suchasthegraphicrenderingprocess)tothecloud[18].
In[13],theauthorsalsodescribemanyexamplesofotherpracticalapplicationsusingMCC,includingamobilelocalizationserviceandavoice-basedortag-basedsearchingservice,amongothers.
However,inthatpapermultimediaaspectsaremissing(e.
g.
image,faceandtextrecognitionusingamobiledeviceandMCC).
Theseapplicationscon-taincalculationsthatrequireafairlylargeamountofcomputationalpowerandthereforecanbeadvan-tageouslyoffloadedtothecloud.
Apartfromtheexamplesmentionedabove,anincreasingnumberofapplicationsrelatedtosocialnetworks[19]andusingmobiledevicesensors(suchasGPS)[20]usetheMCCparadigm.
ThemainobjectiveofthestudiesdescribedinthispaperwastooptimizetheexpensesofexecutionofservicesintheMCCenvironment[21].
Theopti-mizationprocessisperformedautonomouslyonthemobiledevicetotakeintoaccounttheheterogeneityofmobilesystems,environmentalcontextandfluctu-atingconditions.
OptimizationcriteriarepresentedbytheexpensescorrespondtotheQualityofExperience(QoE)parameter[22]andmayinvolve,interalia,exe-cutiontime,energyconsumptionbythemobiledeviceandusersatisfaction.
Wedemonstratethatthisopti-mizationcanbeensuredbyusinganoveldecisionmodulewithlearningcapabilitiesintheMobileCloudComputingenvironment.
Importantly,weshowthattoprovideautonomousonlinelearningonamobiledevice,supervisedlearn-ingmaybeapplied.
Itisnotatraditionalapproachbecauseautonomousonlinelearning,usuallyconsid-eredinthecontextofagent-basedsystems,istypicallyrealizedusingreinforcementlearning(seesurveys[23,24]).
However,ourresearchdemonstratesthatwheretheparameterspaceislarger,supervisedlearn-ingisfasterthanreinforcementlearning[21f20;BACKGROUND-COLOR:#4ae2f7">5].
Thesolutionproposedwasappliedinthedevel-opmentofthecasestudyenvironmentrelatedtotheservice-orientedmobileprocessingsystemformulti-mediafileconversion.
Experimentalresultsshowthatitispossibletooptimizetheutilizationofresources(executiontime)bymovingselectedtasksperformedwithintheframeworkofindividualservicestothecloud.
Ourapproachappliesnewtechnologiesandisageneralsolutionatthesametime.
Itispossibletouseittooptimizeanyservice-orientedsysteminwhichthereisachoicewithrespecttotheplaceofexecu-tionofagivenservice,andexecutioncharacteristics(suchascalculationtimeorenergyconsumption)arepredictable.
Thispaperisstructuredasfollows:Section2con-tainsadescriptionofrelatedwork,Section3isAutonomousContext-BasedServiceOptimizationinMobileCloudComputing341f20;BACKGROUND-COLOR:#4ae2f7">5concernedwithemployingthelearningdecisionmod-uleintheprocessofoptimizingtheserviceselectionstrategyonmobiledevices,Section4describesindetailthedecisionmodule,Section1f20;BACKGROUND-COLOR:#4ae2f7">5presentsthecasestudy,Section6describesperformanceevaluationandSection7containstheconclusion.
2RelatedWorkTheMCCparadigmyieldsmanybenefits[13],espe-ciallyformobiledevices.
Oneofthemostvitaladvan-tagesisenhancingbatterylifewithoutthenecessitytoreplacemobiledevicehardwareandsoftware.
Thepossibilityofmigratingcomplexservices/taskstothecloudresultsindecreasedenergyconsumptionbytheCPUand,asaresult,increasedbatteryuseefficiency.
Anotherimportantadvantageisimprovedreliability.
Savingdatainthecloudreducestheriskofdatalossandenablestheintroductionofadditionalfunction-alities,suchascopyrighteddigitalcontentandvirusscanning.
Nonetheless,allMCCscenariosandapplicationsmustaccommodatethelimitationsresultingfromthenatureofmobilesystems,whichnecessarilyrelyonthewirelesstransmissiontechnology[26]andtakeintoaccounttheoptimaldeploymentofsoftwareintheMCCenvironment[27].
Apartfrompossibleprob-lemsrelatedtowirelessnetworkcommunications(lowbandwidthordataloss),anotherimportantchallengeforMCCissecurity.
Providingtheproperlevelofdatasecurity[28]andprivacy-awarecommunicationmechanism[29]arekeyaspectsofMCC,especiallyinthecontextofe-health[30]orfinanceandbankingapplications.
In[31],theauthorsdealwiththeissueofensuringconfidentialityofdatainMCCthroughtheuseofencryptionandappropriatecryptographicmethods.
Thechoiceofencryptionmethodsshouldtakeintoaccountthelimitedresourcesofmobiledevicesandthepeculiarcharacteristicsofthetech-nologyusedforwirelesscommunication.
Thisarti-cledescribesaCloud-Manager-basedRe-encryptionScheme(CMReS),whichcombinesthecharacteris-ticsofmanager-basedre-encryptionandcloud-basedre-encryptionforprovidingbettersecurityserviceswithaminimumofprocessingburdenonmobiledevices.
Experimentresultsdemonstratethatthesolu-tionproposedresultsinasignificantimprovementinturnaroundtime,energyconsumptionandresourceutilizationonthemobiledeviceascomparedtoexist-ingre-encryptionschemes.
ThepossibilitiesofemployingtheMCCparadigminordertooptimizeservicesonmobiledevicesfornumerousapplicationsandscenarios[8]havebeendiscussedinliterature.
In[32,33]theauthorspresentthepossibilityofusingthisparadigminimagepro-cessingforanapplicationthatreadstext(e.
g.
descrip-tionsofexhibitsatSouthKoreanmuseums[33])andtranslatesitintothelanguageofchoiceusingopti-calcharacterrecognition(OCR).
Ifitisnotpossibletotranslatethetextonthemobiledevice,thetaskisshiftedtothemobilecloudwherethetextistranslatedandtheresultissentbacktotheuser.
Anotherexampledescribedin[12]isaserviceformanagingmultimediafiles,whichcanbecollectedfromnumerousmobiledevicesandcombinedintoasinglefilepresentingtheimagefromdifferentanglesandperspectives.
Inthe"lostchild"scenario,theauthordescribesasit-uationwhereachildismissinganditispossibletocollectandsendrecords(orphotos)fromdiffer-entusers'mobiledevicestothemobilecloudandtogathercomprehensiveinformationonthemissingper-son.
AnexampleofanothersystemusingMCC,whichprovidesmultimediaservices,is[34].
Thedevelopedmobility-awareframeworkformobilecloudstreamingservicesmakesdynamic,optimizedserverselectionfunctionsavailableinordertosupportusermobility.
InthecontextoftheuseofMobileCloudCom-puting,themanagementofcloud-basedresourcesisanimportantissue.
Oneaspectofresourcemanage-mentisoptimizationthatshouldconcernnotonlythemobiledevicebutthecloudaswell.
In[31f20;BACKGROUND-COLOR:#4ae2f7">5],theauthorsdemonstratethatitispossibletooptimizeresourceusageinMCCbyapplyingcommonpatternsusedintraditionalcloudcomputing.
Animportantissueinservice-orientedsystemsisaccountingforthecontextinwhichtheserviceinquestionisexecuted.
Anexampleofsuchasolutionisacontext-awareservicediscoveryframeworkbasedonvirtualpersonalspace(VPS),whichisdescribedin[36].
Inthisframework,themiddlewareembeddedinthedeviceprovidespersonalizedresponses.
AnothersolutionistheMobiByte[37]context-awareappli-cationmodel,whichusesmultipledataoffloadingtechniquesinordertosupportawiderangeofappli-cations.
Inthismodel,manyimportantaspectssuchasenergyefficiency,performance,generality,contextawarenessandprivacyaretakenintoaccount.
346P.
Nawrocki,B.
SniezynskiTheincreaseinuserrequirementsrelatedtomobileaccesstoavarietyofresourceshasresultedinthegrowingpopularityoftheVehicularMobileCloud.
In[38],adaptiveLearningAutomatabasedContentionAwareDataForwarding(LACADF)isproposed.
Thissolutionisevaluatedindifferentnetworkscenarioswithrespecttomultipleparameters(suchasthrough-putanddelay)byvaryingthedensityandmobilityofvehicles.
Asofyet,therehavenotbeenmanystudiesofMCCandtheuseofautonomousmachinelearningalgorithmsexecutedonlineintheprocessofoptimiz-ingtheserviceselectionstrategyonmobiledevices.
AsimilartopicconcerningtheoptimizationofthemobileenvironmentusingMCCisinvestigatedin[39].
Inthepaper,theauthorsemploygeneticalgo-rithmsintheoptimizationprocess;however,theydonote.
g.
considerenergyaspects(mobiledevicebatterylife),focusingsolelyoncomputationalcomplexityandrequirementsconcerningthememoryallocatedtopar-ticularservices.
OptimizationofSOAsystemswithmachinelearningisdiscussedin[40].
However,thispaperdoesnotconsiderthemobileenvironmentandanotherlearningstrategyisused(reinforcementlearn-ing).
Agent-basedapproachtoSOAoptimizationisalsodiscussedin[41].
Supervisedlearningisappliedtogenerateastrategyforchoosingaservicetoexe-cutetasks.
Inthisresearch,mobiledeviceswerenottakenintoaccount.
ThereareseveralothersolutionsthatuseMCCandmachinelearningalgorithmsbutinthosesolutions,someelementsrelatedtolearningarealwaysanalyzedofflineincontrastofoursolu-tionwherethelearningprocessonlytakesplaceonlineonamobiledevice.
In[42],mobiledevicelocationispredictedusingsupervisedlearningwithaslid-ingwindow.
Threemachinelearningalgorithmsareapplied:1NN,C4.
1f20;BACKGROUND-COLOR:#4ae2f7">5,andvoting(basedon1NNandC4.
1f20;BACKGROUND-COLOR:#4ae2f7">5).
Inthatpaper,theproblemoftaskallocation(whichweareworkingon)isnotstudiedexten-sivelyandallexperimentsareperformedofflineonthedatageneratedpreviously.
Inanotherpaper[43],machinelearningalgorithmsareusedtopredicttaskexecutiontimeinMCC.
Thatsolutionconsistsoftwocomponents:anofflineone(developingataskper-formancemodelforeachdevice)andanonlineone(applyingthemodellearnedtothecurrentcase).
Inoursolution,notaskperformancemodelneedstobedevelopedofflineforeachmobiledevice.
OneofthemostimportantpapercoveringMCCsolutionsinthecontextofmachinelearningalgorithmsis[44].
Inthatpaper,theauthorspresentaconceptwherepowersav-ingispossiblethankstothedynamicadaptationofdatatransferandnetworkinterfaceparameters,whichisconductedautomaticallywithoutuserintervention,usingmachinelearningalgorithms.
Inexperiments,theauthorsused1f20;BACKGROUND-COLOR:#4ae2f7">5mobiledevices(withtheAndroidsystem)todefineusermodels.
Theanalysisofdatafromthosedeviceswasperformedoffline.
Owingtothelackofonlineanalysisoftheindividualservicesandthecontextoftheirexecution,thatsolution,whichwasdevelopedusingafewpredefinedmodels,doesnotenablepowerconsumptionoptimizationinthecaseofreal-worldmobileapplicationsandservices.
Anotherconceptthatallowspowerconsumptiontobereducedispresentedin[41f20;BACKGROUND-COLOR:#4ae2f7">5].
However,thatsolutiondoesnotuseMCCbutrathercloudletsthatallowtheoffloadingofservicestonearbymobileresource-richdevices.
Oneofthefewsolutionsthatenablesonlineanal-ysisistheMALMOSframework[46],whichmakesitpossibletodecide,withtheapplicationofmachinelearningtechnologies,whentooffloadapplicationsfromthemobiledevicetothecloud.
Fortheoffload-ingitself,thesolutionusestheDPartnerenviron-ment(Java-basedon-demandoffloadingframework).
Inordertoproperlyteachthesystemwhereitshouldrunapplications,thesolutionusesanonlinetrainingmechanism.
Thesystemworksintwomodes:onlinetrainingphaseandruntimeschedulingphase.
Intheformer,traininginformationiscollectedbyexecut-ingthesametasklocallyandremotely.
Asaresult,aclassifierislearnedthatdetermineswherethetaskshouldbeexecuted.
Inthelatterphase,thetaskisscheduledonthebasisofthelearnedmodel.
Inoursolution,machinelearningisappliedtolearnmod-elsfortasksthatareusedtopredictcalculationtimeandenergyconsumptioninagivencontext.
Thepre-dictionsarethenusedtoestimatecostsoflocalorcloudexecutionandthelessexpensivedecisionischo-sen.
Asaresult,thereisnospeciallearningmode,becausetrainingdatacanbecollectedduringnormalruntime.
Toestimatetrainingphaseduration,anadap-tiveonlinetrainingmechanismisproposedin[46].
Inoursolution,suchastrategyisnotrequiredandthesystemissimpler.
Intests,threemachinelearningalgorithmswereused:instance-basedlearning,per-ceptronandna¨veBayes.
TheresultsdemonstratedanimprovementintheprocessofselectingthelocationAutonomousContext-BasedServiceOptimizationinMobileCloudComputing347forexecutingapplicationsascomparedtosolutionsthatdidnotutilizemachinelearningmechanisms.
However,thosetestswereperformedusinglaptopanddesktopcomputerswithsimulatedworkloadsandnetworkbandwidths,whileinourexperimentsweusedpubliccloudsandreal-lifemobile/WiFinetworkconnections.
Furthermore,thesolutionmentionedcompletelyfailstoaddresstheimportantaspectofenergyconsumptionduringthelaunchingoftheappli-cationandtransferringthedatatothecloudandalsotheamountofenergyusedbytheonlinelearningmechanism.
Anothersolutionthatenablesonlineanalysishasbeendevelopedbytheauthorsofaservice-orientedcontext-awarerecommendersystem[47];itutilizeslearningagentsinanMCCenvironment.
However,thesystemhasbeendevelopedforservicerecom-mendationsonlyandisnotauniversalsolutionthatwouldallowforautonomouscontext-basedserviceoptimizationinMCC.
Alltheexamplesdescribedassumethattheuserhasdecidedtosendthetaskordatafromhisorhermobiledevicetothemobilecloudinordertoperformapartic-ularoperation.
However,weassumethat,incontrasttotheaforementionedexamples,themobiledevicealsohastheoptiontoperformtheserviceinquestionbutitmightprovemorecost-effectivetosendthetask/datatothemobilecloud,performtheoperationsrequiredandreturntheresultstothedevice.
Thisassumptionmakesitpossibletooptimizetheoperationofmobiledevicesonthebasisofvariouscriteriasuchasenergyconsumption,theamountofdatatransferred(andtherelatedcharges)andexecutiontime.
Atthesametime,wehaveconductedresearchintotheuseofthedeci-sionmoduleandthesupervisedlearningprocessinconnectionwithmakingdecisionsontherelevantcir-cumstancesandtheservices(tasks,data)thatshouldbesentto,andexecutedin,themobilecloud.
3OptimizationofServiceSelection:Context,ExecutionResultsandModelsAnimportantquestionrelatedtoMCCistheopti-mizationoftheserviceselectionstrategyonmobiledevices.
Increasinglyoften,itispossibletoperformcomplexservices(tasks)onmobiledevices(thanksto,interalia,theirgreaterprocessingpowerandmem-orysize),butthisincreasesenergyconsumptionatthesametime(shorteningbatterylife).
Thisiswhy,apartfromcaseswheremobiledeviceshavetodelegateser-vices(tasks)tothecloud,theMCCparadigmshouldtakeintoconsiderationsituationsinwhichservices(tasks)mayeitherbemigratedtothecloudtopreservetheresourcesofthemobiledeviceormaybeexecutedlocally.
Thedecisiononthismigrationmaybemadebytheuser,e.
g.
whereheorsheisnotsatisfiedwiththeperformanceofaserviceortheoutcome.
How-ever,theprocessmayalsobeautomatedandonlineadaptationbyapplyingmachinelearningispossible.
Thereforeasolutionthatusesadecisionmodulethatisabletomonitortheenvironmentand,onthatbasis,tomakedecisionsabouttheplacewhereaservice(task)istobeexecuted,hasbeenproposed.
ThedecisionmoduleoperatingonamobiledeviceintheMCCparadigmmonitorstheenvironmentandcollectsinformationon:–thetasktobeexecutedbytheserviceinques-tion,includingitstype,keyarguments,estimateddatainput/outputsize,estimatedexecutiontime,thecostofperformingthecomputationandthetimewhentheresultisneeded;–thecostofperformingtheserviceinthecloud,affectingtheassessmentofthecost-effectivenessofservice;–thelocationofthedevice(domestic/roaming),indicatingwhetherthemobiledeviceusesthedatatransmissionserviceofferedbylocalmobilecar-riers(lowercosts),ormustuseroamingservicesabroad(highertransmissioncosts);–possibledeviceconnectionmodes(Wi-Fi,2G/3G/4G)andconnectionquality,affectingnet-workthroughputbetweenthemobiledeviceandthecloud.
Thetypeofconnectioncanalsoaffectthepowerconsumptionofthemobiledevice;–batterystatus,determininghowlongthemobiledevicecanoperateandhowlongtheservicecanbeperformedonthisdevice;–thecurrenttimeanddate(includingdayofweek,holidays),affectingtheabilitytotakeadvantageofbetterdatatransmissionratesortotrans-mit/receivedataduringperiodswhenthetelecom-municationoperator'sinfrastructureislessbusy;–readingsofsensorssuchastheaccelerometer,lightsensor,etc.
fordeterminingthestatusofthemobiledevice(devicemovement,ambientlighting).
348P.
Nawrocki,B.
SniezynskiThisinformationformsacontext,onthebasisofwhichthemodulemakesdecisionswhenandwheretoperformtheservice(locallyorinthemobilecloud).
Aftercompletingthetask,themoduleassessesitsdecision,consideringoneormoreexecutionresultssuchas:–mobiledevicepowerconsumption;–thetimespentwaitingfortheresult;–theuser'ssatisfaction(theusercouldoverridethemodule'sdecision,whichmeansthatheorshedoesnotagreewithit);–costs(e.
g.
chargesrelatedtodatatransferorusingcloudresources).
Theseresultsareusedtoformulateoptimizationcrite-ria(see(4)).
Themodulegathersexperienceandupdatesitsstrategy,applyingasupervisedlearningalgorithm.
Thegeneratedknowledgestoredinthemodulemayinclude:–modelsoftheuser'sbehaviorthatenableittoassesstheimpactoffactorssuchashisorherloca-tion/connectionaccessibility/abilitytochargethedevice;–modelsofestimatedoutcomesofperformingaservicelocally/inthecloud(energyconsumption,time).
Thesemodelsmaybeusedtoimprovetheestimatesofdecisionconsequences.
4DecisionModule:Model,ArchitectureandAlgorithmBasedontheassumptionsmentionedintheprevi-oussection,wewouldliketoproposeasolutionforserviceselectionadaptationwhichappliessupervisedlearning,executedon-line,onamobiledevice,incon-trastto[36,38]citedabove,inwhichotheroptimiza-tiontechniquesareusedor[41]inwhichMCCisnotconsidered.
Thedecisionmodule,locatedonamobiledevice,worksasfollows:itgetstaskandcontextasaninputandappliesknowledgetomakeadecisionwherethetaskshouldbeexecutedinthegivencon-text.
Thesedatatogetherwithexecutionresultsarestoredastrainingdata.
Thesedataare,inturn,aninputforasupervisedlearningalgorithm,whichreturnstheknowledgeusedtomakethedecisionabouttasks.
LetusdefinetheDecisionModuleasatuple:DM=(A,GK,TD,Dec),(1)whereAisasetofattributes,whichareusedtodescribetasksandthecontext,GKisgeneratedknowledge,andTDistrainingdata,whichisasetofexamples.
Theaimofthemoduleistoreturnadeci-siond∈Dec={d1,d2,.
.
.
dns},whichcorrespondstoengagingoneofnsservices.
Inputdataforthemoduleisapairx=(t,c)rep-resentingthetasktwhichshouldbeexecutedinthecontextc.
TheprocessingmoduledescribesitwithobservationattributesO={o1,o2,.
.
.
on}A,whichyieldsxO=(o1(x),o2(x)on(x)),i.
e.
adescriptionoftheProblem.
Next,usingknowledgestoredinGKitsolvestheProblembyselectingd∈Dec,whichhastheminimumpredictedcost.
IfGKisempty,disrandomized.
Thedecisiondisthenappliedandthetaskisrunusingthecorrespondingservice(e.
g.
locallyorinthemobilecloud).
Aftertheexecution,themoduleobtainsexecutionresultsr,whicharedescribedbyRes={r1,r2,.
.
.
rm}Aattributes(e.
g.
whetherexecutionwassuccessfules(x,d),batteryconsumptionb(x,d),calculationtimect(x,d)anduser'sdissatisfactiondis(x,d),whichmaybemeasuredbyobservingiftheuseroverrodethemodule'sdecision).
Therefore,thesetofallattributesusedtodescribetheinputdataisasumofOandResA=O∪Res.
(2)ThemodulestorestheseresultstogetherwithxOanddecisiondinTD.
Thereforethecompleteexam-plestoredinTDhastheformxA∪Dec=(o1(x),o2(x)on(x),r1(x,d),r2(x,d),.
.
.
rm(x,d),d).
(3)ThemodelstopredictRvaluesareconstructedusingsupervisedlearningalgorithmsandstoredinGK.
Thesemodelsinfluencethechoiceofthedeci-sion.
TheremaybeseveralmodelsstoredinGK.
Usingvaluepredictionsri∈R,themodulemayrateitsdecisionsd∈Decbycalculatingpredictedexpensese(x,d):e(x,d)=mi=1wiri(x,d),(4)wherewiareweightsoftheresultri.
Bychoosingtheweightsonesetstheprioritiesofthecriteria.
AsAutonomousContext-BasedServiceOptimizationinMobileCloudComputing349Fig.
1Thedecisionmodulearchitecturereflectingitslearningcapabilitiesaresult,thesystemisflexibleanduniversal,sinceitmaybeadjustedtotheuser'srequirements.
Themoduleshouldselectthedecisionforwhichexecutionispredictedtobesuccessfulandtheexpenseispredictedtobethelowest.
Theinternalstructureofthemoduleshouldreflectitslearningcapabilities.
ThearchitectureispresentedinFig.
1;threemaincomponentsmaybedistinguished:–ProcessingComponent,whichisresponsibleforbasicactivitiessuchastheprocessingofinputdata,storingtrainingdata,executingthelearningprocessandleveragingtheknowledgelearned;–LearningComponent,whichisresponsibleforexecutingthelearningalgorithmsandprovidinganswerstoproblemsusingtheknowledgelearned;–TrainingData(TD),whichprovidesstoragefortheexamples(experience)usedinlearning;–GeneratedKnowledge(GK),whichprovidesstoragefortheknowledgelearned(models).
Thesecomponentsinteractinthefollowingway:theProcessingComponentreceivesContextdataanddescriptionoftheTask(parameterslistedintheprevi-oussubsection),formulatesaProblembydescribinginputdatawithavailableattributesOandappliesGeneratedKnowledgetogetthePredictionfortheProblem,whichisusedtochoosetheDecision.
Afterexecutingthetask,Resultsareobserved.
AdescriptionoftheProblem,DecisionandResultsisanexam-ple,see(3).
TheProcessingModuledecideswhatexamplesshouldbestoredinTDstorage.
Currentlyallexamplesarestored,butitisalsopossibletostoreoutliers(e.
g.
withextremelyhighcomputationtimes)tosavespace.
Whenrequired(e.
g.
periodi-callyorwhenTDcontainsmanynewexamples),itcallstheLearningComponenttoexecutethelearn-ingalgorithmthatgeneratesnewknowledgefromTD.
TheknowledgelearnedisstoredintheGKbase.
Thealgorithm,whichisexecutedintheProcessingComponent,ispresentedinFig.
2.
Atthebeginning,GKandTDaresettoempty(lines2–3).
Next,ifthereFig.
2AlgorithmoftheProcessingComponentallowingtheoptimizationofthedecisionstrategyusingonlinesupervisedlearning31f20;BACKGROUND-COLOR:#4ae2f7">50P.
Nawrocki,B.
Sniezynskiisnolearnedknowledge,thedecisionisrandomized(lines1f20;BACKGROUND-COLOR:#4ae2f7">5–6).
Elsethereissomeknowledge,thereforeProblemissettoadescriptionoftheobservedcon-textandtask,xO(line9).
Next,GKisusedtoselectthebestdecisionfortheProblem(line10).
Thetaskisexecutedintheselectedservice(line12).
Resultsoftheexecutionareobserved(line13)andexample(xA∪Dec)isstoredintheTD(line14).
Afterpro-cessingagivennumberoftasks(line11f20;BACKGROUND-COLOR:#4ae2f7">5),LearningComponentiscalledtogeneratenewknowledgefromTDandthelearnedknowledgeisstoredinGK(lines16–17).
TheformofknowledgestoredinGKdependsonthelearningalgorithmutilized.
Itmayhaveanexplicitform,e.
g.
rules,adecisiontreeoraBayesianmodelinthecaseofsupervisedlearning.
Itmayalsobestoredinlower-levelformsuchasparametersrepresentingalinearregressionmodel,adecisionvaluefunctionoraneuralnetworkapproximatorofsuchafunctionifreinforcementlearningisapplied.
1f20;BACKGROUND-COLOR:#4ae2f7">5Service-OrientedMobileProcessingSystemInordertoverifythepossibilityofusingadecisionmodulewithlearningcapabilitiestooptimizetheprac-ticalexecutionofservicesintheMCCenvironment,theauthorsdevelopedaService-orientedMobilePro-cessingSystem(SMPS)thatenablestheprocessingofmultimediafiles.
Varyingdemandsforcomputationalpowerfordifferenttasksandthenecessitytotransfervaryingamountsofdatainthecaseofcloudprocess-ingmakethiscasearepresentativedomainfortestingthesolutionproposed.
IntheSMPS,multimediafilesmaybeconvertedusingamultimediadataconversionserviceinthecloudenvironmentordirectlyonthemobiledevice.
Atapplicationruntime,dataconcerningconversionefficiencywerecollectedwithregardtodifferentcon-ditions:thesizeofthefile,thecodecused,tasktypeandwheretheconversiontookplace.
Basedonthosecontextdata,theapplication(DecisionModule)selectedtheconversiontypeandplacethatwouldofferbetterexpectedefficiency.
Executionresultsaresub-mittedtotheDecisionModuletoprovidedataforopti-mization.
ThearchitectureofthesolutiondevelopedisshowninFig.
3.
ThemobileapplicationwasdevelopedfortheAndroidoperatingsystem.
InordertoimplementthefunctionalityusedinthetestsconductedtheauthorspickedtheGoogleCloudEndpointssolution.
ThisisatechnologyenablingeasycommunicationbetweenmobileandwebappswithabackendoperatingintheGoogleAppEnginecloud[48].
Thesolutioninvolvesbothclientlibrariesandserverones,whichareavail-ablefortheJava,Python,PHPandGolanguages.
SincethetestmobileapplicationwasdevelopedfortheAndroidoperatingsystem,theauthorsdecidedtouseJavabothfortheclientonthemobiledeviceandforthebackendoperatinginthecloud.
Inordertoconvertmultimediafiles,thejcodeclibrarywasused,whichcanoperatebothonamobiledevicewiththeAndroidoperatingsystemandintheGoogleAppEnginecloud.
Thismadeitpossibletocompareconversiontimesonthemobiledeviceandinthecloudwherethesamemechanismwasusedforencodingmultimediafiles.
Thejcodeclibrarywasusedtoimplementthefollowingconversiontypes:–fromtheH.
264AVCformat(MP4container)totheAppleProRes422format(Proxy)(MP4container);–fromtheH.
264AVCformat(MP4container)toaseriesofPNGimages(whereevery1f20;BACKGROUND-COLOR:#4ae2f7">5th,11f20;BACKGROUND-COLOR:#4ae2f7">5thor21f20;BACKGROUND-COLOR:#4ae2f7">5thframewasencoded).
Fig.
3ArchitectureoftheService-orientedMobileProcessingSystemAutonomousContext-BasedServiceOptimizationinMobileCloudComputing31f20;BACKGROUND-COLOR:#4ae2f7">51Table1Experimentresultsforana¨veBayesclassifierRound12341f20;BACKGROUND-COLOR:#4ae2f7">5678Minimumexecutiontime,s31f20;BACKGROUND-COLOR:#4ae2f7">5893402282730133411f20;BACKGROUND-COLOR:#4ae2f7">5194026832861Maximumexecutiontime,s4701f20;BACKGROUND-COLOR:#4ae2f7">51f20;BACKGROUND-COLOR:#4ae2f7">5227407444774760296746903743Averageexecutiontime,s3676429731f20;BACKGROUND-COLOR:#4ae2f7">5033378422121f20;BACKGROUND-COLOR:#4ae2f7">53836843181f20;BACKGROUND-COLOR:#4ae2f7">5Standarddeviation9967801f20;BACKGROUND-COLOR:#4ae2f7">5917631f20;BACKGROUND-COLOR:#4ae2f7">571432841387TheAndroidoperatingsystemversionoftheWekalibrary[49],WekaforAndroid[1f20;BACKGROUND-COLOR:#4ae2f7">50]wasusedtoimplementlearningalgorithmsandtoapplytheknowledgelearned.
Inthedomainconsidered,thesolutionproposedmaybespecifiedindetailasfollows.
Thetasktisdescribedbythreeattributes:typerepresentingtasktype(withtwovalues:codecconversionorframeselection),framegap∈{1f20;BACKGROUND-COLOR:#4ae2f7">5,11f20;BACKGROUND-COLOR:#4ae2f7">5,21f20;BACKGROUND-COLOR:#4ae2f7">5}representingthenumberofframestobeskippedduringframeselec-tion,andlength∈{1,2,9}representingthedurationofthemovieprocessed(measuredinsec-onds).
Thecontextcisdescribedbythetimenumericattributerepresentingtheminuteofthedayandthenominalconnectionattributeshowingatypeofcon-nection(LTE,HSPA+,HSPA,EDGE),henceO={type,framegap,length,time,connection}.
ResultsaredescribedbythesuccessfulBooleanattributeshowingwhethertheexecutionwassuccessfulornot,batteryrepresentingbatteryusageandthenumericcalctimeattributerepresentingcalculationtimeinmillisec-onds.
ThereforeRes={successful,battery,calctime}.
Dec={d1=l,d2=cl},whichcorrespondstoengaginglocalorcloudresources.
Duringlearning,aclassifierGKctislearned.
ItisusedtopredictcalctimefromO∪Decattributes.
Thecalctimeisdiscretizedinto7equalranges.
Asaresult,GK=(GKct).
6PerformanceEvaluationInthissectionwedescribeexperimentswhosegoalwastocheckhowwelloursolutionperformedinareal-lifeenvironment.
Everyexperimentconsistsofaseriesoftaskpackagesbeingexecuted.
Everytaskpackageisacombinationofselectedtaskparametersandsystemstatevalues.
Thismeansthateachtestpackageinvolvesmeasuringtheconversiontimeandbatteryusageforthemultimediafileinquestiononthemobiledeviceandusingcloudforvariousfilesizes(withthemaximumfilesizebeing61f20;BACKGROUND-COLOR:#4ae2f7">50kB),variousconversiontypesandnetworkconnectionstatuses.
Duringtheseriesofexperiments,examples(xA∪Dec)arestoredinTD.
Theycontaininforma-tionaboutthecontext,task,results(computationtimeandbattery)anddecision.
Aftereverytaskpackage,thedecisionmoduleinitiatesthelearningprocessandbuildstheGKctclassifier(usingsupervisedlearning),whichisusedinthenextroundtoprocessthetaskpackage.
Duringtheroundn,thelearningmoduleusestheexamplescollectedinrounds1.
.
.
n1astrain-ingdata.
Whentheseriesiscompleted,TDandGKareclearedandthenextseriesisexecutedtocol-lectstatisticaldata.
Iftaskexecutionresultsinfailure,taskexecutiontimeandbatteryusagearesettothemaximumsuccessfulvalueobserved.
Theexperimentconsistedofexecutingtestsinseriesofeightrounds,whichwererepeatedtentimes.
Beforetheexperimentspresentedhere,preliminarytestswithhighernumbersofroundswereperformed.
Theresultsdemonstratedthatincreasingthenumberofroundsdoesnotyieldasignificantimprovementinlearningresults,whileitmakesexperimentsmuchmoredifficult(becauseofthecloudtransferquota)andmoretimeconsumingaswell.
Wehaveexecutedourexperimentswithdefaultset-tingsforalllearningalgorithms.
Therefore,parametersTable2ExperimentresultsforarandomforestclassifierRound12341f20;BACKGROUND-COLOR:#4ae2f7">5678Minimumexecutiontime,s2790367111f20;BACKGROUND-COLOR:#4ae2f7">51f20;BACKGROUND-COLOR:#4ae2f7">56176316172911f20;BACKGROUND-COLOR:#4ae2f7">520641733Maximumexecutiontime,s3882489728232071440431f20;BACKGROUND-COLOR:#4ae2f7">58833162246Averageexecutiontime,s34964218234219333030326828221933Standarddeviation61262368611f20;BACKGROUND-COLOR:#4ae2f7">56139433766727431f20;BACKGROUND-COLOR:#4ae2f7">52P.
Nawrocki,B.
SniezynskiTable3ExperimentresultsforaneuralnetworkRound12341f20;BACKGROUND-COLOR:#4ae2f7">5678Minimumexecutiontime,s3141434924972110181f20;BACKGROUND-COLOR:#4ae2f7">50163918121718Maximumexecutiontime,s3701f20;BACKGROUND-COLOR:#4ae2f7">51f20;BACKGROUND-COLOR:#4ae2f7">571f20;BACKGROUND-COLOR:#4ae2f7">5331f20;BACKGROUND-COLOR:#4ae2f7">571f20;BACKGROUND-COLOR:#4ae2f7">5261724002621f20;BACKGROUND-COLOR:#4ae2f7">533242117Averageexecutiontime,s34341f20;BACKGROUND-COLOR:#4ae2f7">5024290223072138201f20;BACKGROUND-COLOR:#4ae2f7">582462191f20;BACKGROUND-COLOR:#4ae2f7">51Standarddeviation2837031f20;BACKGROUND-COLOR:#4ae2f7">5872712761f20;BACKGROUND-COLOR:#4ae2f7">510778208havethefollowingvalues.
Forna¨veBayes,theuseKernelEstimatoranduseSupervisedDiscretizationparametersaresettofalse.
Fortherandomforest,bagSizePercentis100,maximumtreedepthisunlim-ited,breakTiesRandomlyandcalcOutOfBagaresettofalse,thenumberofiterations(trees)is100,andnumFeaturesis0(whichmeansthatthenumberofrandomlyselectedattributesislog2(#predictors)+1).
Fortheneuralnetwork,thehiddenLayersparam-eterissettothe"a"value,meaningthatthereisasinglehiddenlayerwithfourneurons,trainingtimeisequalto1f20;BACKGROUND-COLOR:#4ae2f7">500epochs,learningrateis0.
3andmomen-tumis0.
2.
Thefollowingthreeparametersaresettotrue:nominalToBinaryFilter,normalizeAttributesandnormalizeNumericClass.
Ineachtest,EDGE/WiFinetworkcommunication,theGoogleAppEnginecloudandthesupervisedlearningmethodwereused.
Theexperimentwascar-riedoutusingtheSamsungGT-i91f20;BACKGROUND-COLOR:#4ae2f7">501f20;BACKGROUND-COLOR:#4ae2f7">5mobiledevice.
Thismobiledevicehasthefollowingspecifications:SoC–QualcommAPQ8064TSnapdragon600,CPU–quad-core1.
9GHzKrait300,GPU–Adreno320,RAM–2GBandstorage–16GB.
Forexperimentsrelatedtotheexecutiontimeforthemultimediafile,thedecisionmoduletakesintoaccountcalculationtimectincostcalculations(weightswb=wd=0,wct=1).
Inthefirstexperiment,ana¨veBayeslearn-ingalgorithmisusedtolearntheGKctclassifier.
TheresultsofthosetestsareshowninTable1.
Applicationofmachinelearningusingthena¨veBayesclassifiersignificantlyincreasesexecutionspeedforthetasksrequested.
Themeantaskexecutiontimeinsubse-quentroundsoftestsusingtheknowledgegainedasaresultofmachinelearningshowsadownwardtrend.
Thedifferencebetweenmeanvaluesfromthefirstandthelastroundisabout1f20;BACKGROUND-COLOR:#4ae2f7">500seconds(about13%).
However,accordingtothet-Studenttest,thisdiffer-enceisnotstatisticallysignificant(p-valueisequalto0.
0831).
Inthesecondexperiment,arandomforestclassi-fierisappliedtogenerateGKct.
TheresultsofthosetestsareshowninTable2.
Inthiscasethedifferencebetweenmeanvaluesfromthefirstandthelastroundsislarger–over1,1f20;BACKGROUND-COLOR:#4ae2f7">500seconds(41f20;BACKGROUND-COLOR:#4ae2f7">5%).
Thet-Studenttestdemonstratesthatthedifferenceisstatisticallysignificant(thep-valueislessthan0.
0001).
Inthethirdexperiment,aback-propagationneuralnetworkwasusedtomodelGKct.
TheresultsofthosetestsareshowninTable3.
Thedifferencebetweenmeanvaluesfromthefirstandthelastroundsisslightlysmallerthanfortherandomforest.
Itisalmost1,1f20;BACKGROUND-COLOR:#4ae2f7">500seconds(43%).
Thet-StudenttestdemonstratesFig.
4ComparisonoftheaverageexecutiontimesinsubsequentroundsforthreemachinelearningalgorithmsappliedtoselecttheappropriatelocationforrunningamultimediaserviceAutonomousContext-BasedServiceOptimizationinMobileCloudComputing31f20;BACKGROUND-COLOR:#4ae2f7">53Fig.
1f20;BACKGROUND-COLOR:#4ae2f7">5Comparisonoftheaveragenumberofserviceexecutionsinthecloudinsubsequentroundsforthreemachinelearningalgorithms(aspercentages)thatthedifferenceisstatisticallysignificant(thep-valueislessthan0.
0001).
AcomparisonoftheresultsofallthreeexperimentsispresentedinFig.
4.
Aswecansee,therandomfor-estandneuralnetworkhavesimilarfinalperformance.
However,theneuralnetworkismorestable.
Na¨veBayesistheworst.
Thereasonisthatitisunabletorepresentthedependenciesbetweeninputattributes.
Toexplainthevaryingexecutiontimes,wecheckedhowthelocationforexecutingtheservicechanges.
Figure1f20;BACKGROUND-COLOR:#4ae2f7">5presentstheaveragenumbersofmultime-diadataconversionservicestart-upswithinthecloud(asapercentage)insubsequentroundsforallthreelearningalgorithms(randomforest,na¨veBayesandneuralnetwork).
Thenumberoftotalservicecallsintheroundwas40(100%).
Itcanbenoticedthat,basedonFigs.
4and1f20;BACKGROUND-COLOR:#4ae2f7">5,theshortestserviceexecutiontimesoccurwhenabouthalfofcallsareexecutedonthemobiledevice(round8).
However,itshouldberememberedthatexecutingtheservicelocally,onamobiledevice,cancausehigherloadonthedevice,withtheresultingincreaseinbatteryusageandreduc-tionintheoperationtimeofthemobiledevice.
Whatismore,executingalltasksfromthepackageonthemobiledeviceresultsinamuchlongeraverageexecu-tiontime(1f20;BACKGROUND-COLOR:#4ae2f7">5,1f20;BACKGROUND-COLOR:#4ae2f7">563seconds).
Thismeansthatdecisionsabouttaskallocationarenottrivial.
Wehavealsoperformedexperimentsmeasuringmobiledeviceenergyconsumption(usingthePow-erTutorsoftware[1f20;BACKGROUND-COLOR:#4ae2f7">51])duringtheexecutionoftasks.
Twoclassifiersarelearned:GK={GKct,GKb},withtheformerdiscussedaboveandthelatterbeingusedtopredictbatteryusagefromO∪Decattributes.
Thebatteryattributeisalsodiscretizedinto7equalranges.
Duringthesetests,thedecisionmoduletakesintoaccountcalculationtimectandbatteryusagebincostcalculations(weightswb=0.
1f20;BACKGROUND-COLOR:#4ae2f7">5,wct=0.
1f20;BACKGROUND-COLOR:#4ae2f7">5,wd=0).
TheresultsofthosetestsareshowninFig.
6.
AsweFig.
6Comparisonofaveragebatteryusageinsubsequentroundsforthethreemachinelearningalgorithmsusedtoselecttheappropriatelocationforrunningamultimediaservice31f20;BACKGROUND-COLOR:#4ae2f7">54P.
Nawrocki,B.
Sniezynskicansee,theapplicationofthemachinelearningalgo-rithmreducesbatteryusageforthetasksrequested.
Theaveragereductioninbatteryusagebetweenthefirstandlastroundsisequalto41%forna¨veBayes,33%forrandomforest,and16%forneuralnetwork.
Accordingtothet-Studenttest,thesedifferencesarestatisticallysignificantforna¨veBayes(p-valueisequalto0.
0001)andrandomforest(p-valueisequalto0.
0002).
Forneuralnetwork,thedifferenceisnotstatisticallysignificant(p-valueisequalto0.
01f20;BACKGROUND-COLOR:#4ae2f7">594),whichmayresultfromthedefaultsettingsforthisclassifierintheWekalibrary.
Experimentalresultsshowthattheexecutiontimeofmultimediadataconversionservicesandbatteryusageweresignificantlydecreased.
Itsuggeststhattheautonomouscontext-basedserviceoptimisationmethodwiththeuseoflearningalgorithms,whichwehaveproposed,maybeappliedtoimproveQualityofExperienceinreal-lifescenarios.
7ConclusionsInthispaper,wehaveproposedanovelsolutionforautonomouscontext-basedserviceoptimizationintheMCCenvironmentwhichincludesaformalmodelofthelearningdecisionmodule,service-orientedarchi-tecture,andserviceselectionoptimizationalgorithm.
Thesolutionisbasedonananalysisofpossibleapproaches.
AfteraninvestigationwehaveselectedappropriateoptimizationmethodsintheMCCenvi-ronment.
Theexperimentsrelatedtotheoptimiza-tionofvideofileprocessingserviceshavedemon-stratedthatthemainobjectiveofourstudieshasbeenachievedandtheexpensesoftheexecutionofservicesintheMCCenvironmenthavebeenopti-mized.
ThelearningdecisionmoduleproposedforselectingservicesintheMCCenvironmenthasbeenabletooptimizethelocationwheretasksaretobeexecuted.
Thesupervisedlearningapproachworksverywellandallowsmobiledevicestolearnthemodelsonlineandinanautonomousmanner.
Theresultsdemon-strateastatisticallysignificantdecreaseinthetimerequiredfortheexecutionofconversionservicesowingtotheautomaticselection(usinglearningmeth-ods)ofthelocation(mobiledevice/cloud)wheretheconversionofmultimediadataistobeperformed.
Thedecisionmodulewasabletoautonomouslycollecttrainingdata,whichwastheinputforthelearningalgorithmexecutedonlineonthemobiledevice.
Summingup,byapplyingnewtechnologies,suchassupervised,on-linelearningperformedintheMCCenvironmentwehaveintroducedanovelapproachtooptimizingtheexecutionofservicesandtheoperat-ingtimeofmobiledevicesinthisenvironment.
Thesolutionisageneraloneandallowsustooptimizevar-iousservice-orientedsystemswherethelocationforexecutingagiventaskisflexible.
Furtherresearchinthisareawillincludesimilarstudiesfordifferenttypesofserviceswithlearningtechniques,takingintoaccountcontextdatafromthewiderangeofsensorsinstalledinmobiledevicessuchasGPSmodules,accelerometersandlightsensors.
Researchondiscoveringchangesinconditionswillbecarriedout.
Wealsoplantoapplyotherlearn-ingalgorithmsandinvestigatewhethertheexchangeoflearnedknowledgebetweendecisionmoduleswillprovebeneficial.
AcknowledgementsTheresearchpresentedinthispaperwassupportedbythePolishMinistryofScienceandHigherEdu-cationunderAGHUniversityofScienceandTechnologyGrant11.
11.
230.
124.
WethankMagorzataPa˙zek,JakubCzy˙zewskiandDanielOlszowskiforassistancewithimplementationandtesting.
OpenAccessThisarticleisdistributedunderthetermsoftheCreativeCommonsAttribution4.
0InternationalLicense(http://creativecommons.
org/licenses/by/4.
0/),whichpermitsunre-stricteduse,distribution,andreproductioninanymedium,providedyougiveappropriatecredittotheoriginalauthor(s)andthesource,providealinktotheCreativeCommonslicense,andindicateifchangesweremade.
References1.
Roberts,J.
R.
,Incorporated,M.
:MobileTechReport2014:Technologynewsfrom2013andpredictionsandinsightsabout2014.
MindwarmIncorporated(2014)2.
Buyya,R.
,Yeo,C.
S.
,Venugopal,S.
,Broberg,J.
,Brandic,I.
:CloudcomputingandemergingITplatforms:vision,hype,andrealityfordeliveringcomputingasthe1f20;BACKGROUND-COLOR:#4ae2f7">5thutility.
Futur.
Gener.
Comput.
Syst.
21f20;BACKGROUND-COLOR:#4ae2f7">5(6),1f20;BACKGROUND-COLOR:#4ae2f7">599–616(2009)3.
Cushing,R.
,Koulouzis,S.
,Belloum,A.
,Bubak,M.
:Applyingworkflowasaserviceparadigmtoapplica-tionfarming.
ConcurrencyandComputation:PracticeandExperience26(6),1297–1312(2014)4.
Cushing,R.
,Putra,G.
H.
H.
,Koulouzis,S.
,Belloum,A.
,Bubak,M.
,DeLaat,C.
:Distributedcomputingonanensembleofbrowsers.
IEEEInternetComput.
17(1f20;BACKGROUND-COLOR:#4ae2f7">5),1f20;BACKGROUND-COLOR:#4ae2f7">54–61(2013)AutonomousContext-BasedServiceOptimizationinMobileCloudComputing31f20;BACKGROUND-COLOR:#4ae2f7">51f20;BACKGROUND-COLOR:#4ae2f7">51f20;BACKGROUND-COLOR:#4ae2f7">5.
Kumar,K.
,Lu,Y.
-H.
:Cloudcomputingformobileusers:canoffloadingcomputationsaveenergyComputer43(4),1f20;BACKGROUND-COLOR:#4ae2f7">51–1f20;BACKGROUND-COLOR:#4ae2f7">56(2010)6.
Kumar,K.
,Liu,J.
,Lu,Y.
-H.
,Bhargava,B.
:Asurveyofcomputationoffloadingformobilesystems.
Mob.
Netw.
Appl.
18(1),129–140(2013)7.
Shiraz,M.
,Gani,A.
,Shamim,A.
,Khan,S.
,Ahmad,R.
W.
:Energyefficientcomputationaloffloadingframeworkformobilecloudcomputing.
JournalofGridComputing13(1),1–18(2011f20;BACKGROUND-COLOR:#4ae2f7">5)8.
Fernando,N.
,Loke,S.
W.
,Rahayu,W.
:Mobilecloudcomputing:asurvey.
FutureGener.
Comput.
Syst.
29(1),84–106(2013)9.
Ma,R.
K.
K.
,Wang,C.
-L.
:Lightweightapplication-leveltaskmigrationformobilecloudcomputing.
In:IEEE26thInternationalConferenceonAdvancedInformationNet-workingandApplications(AINA),2012,pp.
1f20;BACKGROUND-COLOR:#4ae2f7">51f20;BACKGROUND-COLOR:#4ae2f7">50–1f20;BACKGROUND-COLOR:#4ae2f7">51f20;BACKGROUND-COLOR:#4ae2f7">57(2012)10.
Nawrocki,P.
,Sobon,M.
:Publiccloudcomputingforsoft-wareasaserviceplatforms.
Comput.
Sci.
11f20;BACKGROUND-COLOR:#4ae2f7">5(1),89–103(2014)11.
Chun,B.
-G.
,Ihm,S.
,Maniatis,P.
,Naik,M.
,Patti,A.
:Clonecloud:elasticexecutionbetweenmobiledeviceandcloud.
In:ProceedingsoftheSixthConferenceonCom-puterSystems,EuroSys'11,pp.
301–314.
ACM,NewYork,NY,USA(2011)12.
Satyanarayanan,M.
:Mobilecomputing:thenextdecade,pp.
1f20;BACKGROUND-COLOR:#4ae2f7">5:1–1f20;BACKGROUND-COLOR:#4ae2f7">5:6.
ACM,NewYork,NY,USA(2010)13.
Dinh,H.
T.
,Lee,C.
,Niyato,D.
,Wang,P.
:Asurveyofmobilecloudcomputing:architecture,applications,andapproaches.
Wirel.
Commun.
Mob.
Comput.
13(18),11f20;BACKGROUND-COLOR:#4ae2f7">587–1611(2013)14.
Yang,X.
,Pan,T.
,Shen,J.
:On3gmobilee-commerceplat-formbasedoncloudcomputing.
In:3rdIEEEInternationalConferenceonUbi-MediaComputing(U-Media),2010,pp.
198–201(2010)11f20;BACKGROUND-COLOR:#4ae2f7">5.
Li,J.
:Studyonthedevelopmentofmobilelearningpro-motedbycloudcomputing.
In:2ndInternationalConfer-enceonInformationEngineeringandComputerScience(ICIECS),2010,pp.
1–4(2010)16.
Doukas,C.
,Pliakas,T.
,Maglogiannis,I.
:Mobilehealth-careinformationmanagementutilizingcloudcomputingandandroidos.
In:EngineeringinMedicineandBiologySociety(EMBC),2010AnnualInternationalConferenceoftheIEEE,pp.
1037–1040(2010)17.
Tang,W.
-T.
,Hu,C.
-M.
,Hsu,C.
-Y.
:Amobilephonebasedhomecaremanagementsystemonthecloud.
In:3rdInternationalConferenceonBiomedicalEngineeringandInformatics(BMEI),2010,vol.
6,pp.
2442–2441f20;BACKGROUND-COLOR:#4ae2f7">5(2010)18.
Wang,S.
,Dey,S.
:Renderingadaptationtoaddresscommu-nicationandcomputationconstraintsincloudmobilegam-ing.
In:GlobalTelecommunicationsConference(GLOBE-COM2010),2010IEEE,pp.
1–6(2010)19.
Karadimce,A.
,Davcev,D.
:Buildingcontext-richmobilecloudservicesformobilecloudapplications.
In:Mesquita,A.
,Peres,P.
(eds.
)Proceedingsofthe2ndEuropeanCon-ferenceonSocialMedia2011f20;BACKGROUND-COLOR:#4ae2f7">5:ECSM2011f20;BACKGROUND-COLOR:#4ae2f7">5,pp.
1f20;BACKGROUND-COLOR:#4ae2f7">501f20;BACKGROUND-COLOR:#4ae2f7">5–1f20;BACKGROUND-COLOR:#4ae2f7">513(2011f20;BACKGROUND-COLOR:#4ae2f7">5)20.
Ye,Z.
,Chen,X.
,Li,Z.
:Videobasedmobilelocationsearchwithlargesetofsiftpointsincloud.
In:Proceedingsofthe2010ACMMultimediaWorkshoponMobileCloudMediaComputing,MCMC'10,pp.
21f20;BACKGROUND-COLOR:#4ae2f7">5–30.
ACM,NewYork,NY,USA(2010)21.
Ahmed,E.
,Gani,A.
,Sookhak,M.
,Hamid,S.
H.
A.
,Xia,F.
:Applicationoptimizationinmobilecloudcomputing.
J.
Netw.
Comput.
Appl.
1f20;BACKGROUND-COLOR:#4ae2f7">52(C),1f20;BACKGROUND-COLOR:#4ae2f7">52–68(2011f20;BACKGROUND-COLOR:#4ae2f7">5)22.
Ivesic,K.
,Skorin-Kapov,L.
,Matijasevic,M.
:Cross-layerQoE-drivenadmissioncontrolandresourceallocationforadaptivemultimediaservicesinLTE.
J.
Netw.
Comput.
Appl.
46(0),336–31f20;BACKGROUND-COLOR:#4ae2f7">51(2014)23.
Panait,L.
,Luke,S.
:Cooperativemulti-agentlearning:thestateoftheart.
Auton.
Agent.
Multi-AgentSyst.
11,2001f20;BACKGROUND-COLOR:#4ae2f7">5(2001f20;BACKGROUND-COLOR:#4ae2f7">5)24.
Stone,P.
,Veloso,M.
:Multiagentsystems:asurveyfromthemachinelearningperspective.
Auton.
Robot.
8,341f20;BACKGROUND-COLOR:#4ae2f7">5–383(2000)21f20;BACKGROUND-COLOR:#4ae2f7">5.
Sniezynski,B.
:Astrategylearningmodelforautonomousagentsbasedonclassification.
Int.
J.
Appl.
Math.
Comput.
Sci.
21f20;BACKGROUND-COLOR:#4ae2f7">5(3),471–482(2011f20;BACKGROUND-COLOR:#4ae2f7">5)26.
Bagchi,S.
:Thesoftwarearchitectureforefficientdis-tributedinterprocesscommunicationinmobiledistributedsystems.
JournalofGridComputing12(4),611f20;BACKGROUND-COLOR:#4ae2f7">5–631f20;BACKGROUND-COLOR:#4ae2f7">5(2014)27.
Verbelen,T.
,Stevens,T.
,DeTurck,F.
,Dhoedt,B.
:Graphpartitioningalgorithmsforoptimizingsoftwaredeploymentinmobilecloudcomputing.
Futur.
Gener.
Comput.
Syst.
29(2),41f20;BACKGROUND-COLOR:#4ae2f7">51–41f20;BACKGROUND-COLOR:#4ae2f7">59(2013).
Specialsection:Recentadvancesine-Science28.
Khan,A.
N.
,MatKiah,M.
L.
,Khan,S.
U.
,Madani,S.
A.
:Towardssecuremobilecloudcomputing:asurvey.
FutureGener.
Comput.
Syst.
29(1f20;BACKGROUND-COLOR:#4ae2f7">5),1278–1299(2013)29.
Lin,H.
,Xu,L.
,Mu,Y.
,Wu,W.
:Areliablerecommendationandprivacy-preservingbasedcross-layerreputationmech-anismformobilecloudcomputing.
FutureGener.
Comput.
Syst.
(2014)30.
Huang,D.
,Zhou,Z.
,Xu,L.
,Xing,T.
,Zhong,Y.
:Securedataprocessingframeworkformobilecloudcom-puting.
In:IEEEConferenceonComputerCommunica-tionsWorkshops(INFOCOMWKSHPS),2011,pp.
614–618(2011)31.
Khan,A.
N.
,MatKiah,M.
L.
,Ali,M.
,Shamshirband,S.
,Khan,A.
U.
R.
:Acloud-manager-basedre-encryptionschemeformobileusersincloudenvironment:ahybridapproach.
J.
GridComput.
13(4),61f20;BACKGROUND-COLOR:#4ae2f7">51–671f20;BACKGROUND-COLOR:#4ae2f7">5(2011f20;BACKGROUND-COLOR:#4ae2f7">5)32.
Huerta-Canepa,G.
,Lee,D.
:Avirtualcloudcomputingproviderformobiledevices.
In:Proceedingsofthe1stACMWorkshoponMobileCloudComputing&Services:SocialNetworksandBeyond,MCS'10,pp.
6:1–6:1f20;BACKGROUND-COLOR:#4ae2f7">5.
ACM,NewYork,NY,USA(2010)33.
Cheng,R.
K.
B.
alan.
J.
,Satyanarayanan,M.
:ExploitingRichMobileEnvironment.
TechnicalReportTechnicalReportCarnegieMellonuniversity-CS-01f20;BACKGROUND-COLOR:#4ae2f7">5-199.
CarnegieMellonUniversity(2001f20;BACKGROUND-COLOR:#4ae2f7">5)34.
Liu,R.
,Yuan,X.
,Xu,J.
,Chen,J.
,Zeng,Y.
,Cao,M.
,Liu,J.
,Xu,L.
,Fang,Q.
:Anovelserverselectionapproachformobilecloudstreamingservice.
Simul.
Model.
Pract.
Theory1f20;BACKGROUND-COLOR:#4ae2f7">50,72–82(2011f20;BACKGROUND-COLOR:#4ae2f7">5).
SpecialIssueonResourceMan-agementinMobileClouds31f20;BACKGROUND-COLOR:#4ae2f7">5.
Nawrocki,P.
,Reszelewski,W.
:Resourceusageoptimiza-tioninmobilecloudcomputing.
Comput.
Commun.
99(C),1–12(2017)31f20;BACKGROUND-COLOR:#4ae2f7">56P.
Nawrocki,B.
Sniezynski36.
Park,K.
-L.
,Yoon,U.
H.
,Kim,S.
-D.
:Personalizedservicediscoveryinubiquitouscomputingenvironments.
IEEEPervasiveComput.
8(1),1f20;BACKGROUND-COLOR:#4ae2f7">58–61f20;BACKGROUND-COLOR:#4ae2f7">5(2009)37.
Khan,A.
U.
R.
,Othman,M.
,Khan,A.
N.
,Abid,S.
A.
,Madani,S.
A.
:Mobibyte:anapplicationdevelopmentmodelformobilecloudcomputing.
J.
GridComput.
13(4),601f20;BACKGROUND-COLOR:#4ae2f7">5–628(2011f20;BACKGROUND-COLOR:#4ae2f7">5)38.
Kumar,N.
,Iqbal,R.
,Misra,S.
,Rodrigues,J.
J.
P.
C.
:Bayesiancoalitiongameforcontention-awarereliabledataforwardinginvehicularmobilecloud.
Futur.
Gener.
Com-put.
Syst.
48,60–72(2011f20;BACKGROUND-COLOR:#4ae2f7">5).
SpecialSection:BusinessandIndustrySpecificCloud39.
Liu,Q.
,Jian,X.
,Hu,J.
,Zhao,H.
,Zhang,S.
:Anoptimizedsolutionformobileenvironmentusingmobilecloudcom-puting.
In:1f20;BACKGROUND-COLOR:#4ae2f7">5thInternationalConferenceonWirelessCom-munications,NetworkingandMobileComputing,2009.
Wicom'09,pp.
1–1f20;BACKGROUND-COLOR:#4ae2f7">5(2009)40.
Skalkowski,K.
,Zielinski,K.
:Automaticadaptationofsoasystemssupportedbymachinelearning.
In:Camarinha-Matos,L.
M.
,Tomic,S.
,Graca,P.
(eds.
)TechnologicalInnovationfortheInternetofThings,vol.
394ofIFIPAdvancesinInformationandCommunicationTechnology,pp.
61–68.
Springer,Berlin(2013)41.
Sniezynski,B.
:Agent-basedadaptationsystemforservice-orientedarchitecturesusingsupervisedlearning.
ProcediaComputerScience29,101f20;BACKGROUND-COLOR:#4ae2f7">57–1067(2014)42.
Anagnostopoulos,T.
,Anagnostopoulos,C.
,Hadjiefthymi-ades,S.
:Mobilitypredictionbasedonmachinelearning.
In:2011IEEE12thInternationalConferenceonMobileDataManagement,vol.
2,pp.
27–30(2011)43.
Shi,C.
,Pandurangan,P.
,Ni,K.
,Yang,J.
,Ammar,M.
,Naik,M.
,Zegura,E.
:Ic-cloud:ComputationOffloadingtoanIntermittently-ConnectedCloud.
Technicalreport(2013)44.
Donohoo,B.
K.
,Ohlsen,C.
,Pasricha,S.
,Xiang,Y.
,Ander-son,C.
:Context-awareenergyenhancementsforsmartmobiledevices.
IEEETrans.
Mob.
Comput.
13(8),1720–1732(2014)41f20;BACKGROUND-COLOR:#4ae2f7">5.
Zhang,Y.
,Niyato,D.
,Wang,P.
:Offloadinginmobilecloudletsystemswithintermittentconnectivity.
IEEETrans.
Mob.
Comput.
14(12),21f20;BACKGROUND-COLOR:#4ae2f7">516–21f20;BACKGROUND-COLOR:#4ae2f7">529(2011f20;BACKGROUND-COLOR:#4ae2f7">5)46.
Eom,H.
,Figueiredo,R.
,Cai,H.
,Zhang,Y.
,Huang,G.
:Malmos:machinelearning-basedmobileoffloadingsched-ulerwithonlinetraining.
In:2011f20;BACKGROUND-COLOR:#4ae2f7">53rdIEEEInternationalConferenceonMobileCloudComputing,Services,andEngineering,pp.
1f20;BACKGROUND-COLOR:#4ae2f7">51–60(2011f20;BACKGROUND-COLOR:#4ae2f7">5)47.
Nawrocki,P.
,Sniezynski,B.
,Czyzewski,J.
:Learningagentforaservice-orientedcontext-awarerecommendersysteminheterogeneousenvironment.
ComputingandInformatics31f20;BACKGROUND-COLOR:#4ae2f7">5(1f20;BACKGROUND-COLOR:#4ae2f7">5),1001f20;BACKGROUND-COLOR:#4ae2f7">5–1026(2016)48.
Malawski,M.
,Kuzniar,M.
,Wojcik,P.
,Bubak,M.
:Howtousegoogleappengineforfreecomputing.
IEEEInternetComput.
17(1),1f20;BACKGROUND-COLOR:#4ae2f7">50–1f20;BACKGROUND-COLOR:#4ae2f7">59(2013)49.
Witten,I.
H.
,Frank,E.
:Datamining:Practicalmachinelearningtoolsandtechniqueswithjavaimplementations.
MorganKaufmann(1999)1f20;BACKGROUND-COLOR:#4ae2f7">50.
Marsan,R.
J.
:Wekaforandroid.
https://github.
com/rjmarsan/weka-for-android1f20;BACKGROUND-COLOR:#4ae2f7">51.
Zhang,L.
,Tiwana,B.
,Qian,Z.
,Wang,Z.
,Dick,R.
P.
,Mao,Z.
M.
,Yang,L.
:Accurateonlinepoweresti-mationandautomaticbatterybehaviorbasedpowermodelgenerationforsmartphones.
In:ProceedingsoftheEighthIEEE/ACM/IFIPInternationalConferenceonHardware/SoftwareCodesignandSystemSynthesis,CODES/ISSS'10,pp.
101f20;BACKGROUND-COLOR:#4ae2f7">5–114.
ACM,NewYork,NY,USA(2010)