©CopyrightJASSS
AlexisKirkeandEduardoMiranda(2015)
AMulti-AgentEmotionalSocietyWhoseMelodiesRepresentitsEmergentSocialHierarchyandAreGeneratedbyAgentCommunications
JournalofArtificialSocietiesandSocialSimulation 18(2)16<http://jasss.soc.surrey.ac.uk/18/2/16.html>
Received:29-Nov-2013Accepted:26-Oct-2014Published:31-Mar-2015
Abstract
Inthisarticleamulti-agentsystemispresentedwhichgeneratesmelodypitchsequenceswithahierarchicalstructure.Theagentshavenoexplicitmelodicintelligenceandgeneratethepitchesasaresultofartificialemotionalinfluenceandcommunicationbetweenagents,andthemelody'shierarchicalstructureisaresultoftheemergingagentsocialstructure.Thesystemisnotamappingfrommulti-agentinteractionontomusicalfeatures,butactuallyutilizesmusicfortheagentstocommunicateartificialemotions.Eachagentinthesocietylearnsitsowngrowingtuneduringtheinteractionprocess.Experimentsarepresenteddemonstratingthatdiverseandnon-trivialmelodiescanbegenerated,aswellasahierarchicalmusicalstructure.
Keywords:SocialNetworks,Music,Emotion
Introduction
1.1 Thegenerationofnovelmusicisattheheartofmanycomputer-aidedcomposition(CAC)systems.Withoutsomewayofgeneratingnewmaterial,aCACwillchurnoutthesamematerialtimeaftertime.Toavoidthis,manysystemsutilizerandomnumbers.Amorerecentalternativeisthegenerationofcomplexstructureswhichareorderedbutunpredictable.PopulartypesofsystemsthatgeneratestructureswithsuchcomplexityarefoundinthefieldofartificiallifeorA-Life(Brown2002).A-Lifeinvestigatessystemsrelatedtolife,theirprocesses,andevolution;itdoesthismostoftenthroughcomputersimulationsandmodels–forexamplecellularautomata.ManyA-lifesystemshavetwoelementsincommonwithhavemadethemattractivetocomposersforuseinCAC:theygeneratecomplexdatawithorderandstructure,andtheyinspirecomposersbythevarietyofpatternsinthedata(Panzarasa&Jennings2006).SoalthoughA-Lifesystemscangenerateunexpectedbehaviour,thereisaninherentorder–theyarenotsolelyrandom.Thisisoftencalledemergentbehaviour.
1.2 Onefieldwhichhasalargeintersectionwithartificiallifeismulti-agentsystems(MAS),whichisoneofthe2keyareasutilizedinthisarticle.EachagentinanMASisadigitalentitywhichcaninteractwithotheragentstosolveproblemsasagroup,thoughnotnecessarilyinanexplicitlyco-ordinatedway.Whatoftenseparatesagent-basedapproachesfromnormalobject-orientedormodularsystemsistheiremergentbehaviour(Dahlstedt&McBurney2006).Thesolutionoftheproblemtackledbytheagentsisoftengeneratedinanunexpectedwayduetotheircomplexinteractionaldynamics,thoughindividualagentsmaynotbethatcomplex.AswiththeapplicationofotherA-LifesystemsinCAC,thesesocialdynamicscanbebothartisticallyfunctional–forexampleeachagentinanensemblecancontributeamotiforplayanartificialinstrumentinapieceofmusic;orartisticallymotivational,inspiringanalgorithmiccomposertoproducethemusicofartificialsocieties.
1.3 Inthisarticleamulti-agentsystemispresentedwhichgeneratesmelodypitchsequenceswithahierarchicalstructure.Theagentshavenoexplicitmelodicintelligenceandgeneratethepitchesasaresultofartificialemotionalinfluenceandcommunicationbetweenagents,andthemusic'shierarchicalstructureisaresultoftheemergingagentsocialstructure.Anotherkeyelementisthatthesystemisnotamappingfrommulti-agentinteractionontomusicalfeatures,butactuallyutilizesmusicfortheagentstocommunicateartificialemotions.Eachagentinthesocietylearnsitsowngrowingtuneduringtheinteractionprocess.
RelatedWork
2.1 Anumberofsystemswithsimilaritiestotheoneinthispaperarenowexaminedindetail.Beforedoingthat,abriefoverviewofmoregeneralmulti-agentmusicsystemsisgivenusingTable1.Thesearenotexaminedindetailbutthetableisdesignedtogivequickfamiliaritywithanumberofkeyissuesfoundinmusicalmulti-agentsystems.Thefieldswillnowbeexplained.Complexitydescribesthelevelofprocessinginindividualagents,howcomplexarethey?Homog/HetindicateswhethertheagentsintheMASarehomogeneousorheterogeneous–i.e.doagentsallstartoutthesame,oraresomedifferent?Commindicateswhethertheagentscommunicate,andifsodotheydoitsynchronouslyorasynchronously;i.e.dotheytakeitinturnstocommunicateandprocess,ordotheydoitconcurrently?InitialHierarchydescribeswhetherthereisahierarchyofplanning/controlfortheagents;aresomeagentsdependentonothers?Cansomeagentscontrolothers?Tuneindicateswhetherthesystemgeneratesmultiplecompositionalternativeswhenitcompletesprocessing,orasinglecomposition.Real-timedescribeswhetherwhentheagentsareactivated,themusicgeneratedinreal-time.Sizegives–whereavailableandrelevant–thenumber,oraveragenumber,ofagentsinthesystem.FinallyModel/Funcindicateswhetherthesystemisdesignedsolelytomodelsomeelementofmusic,orasacomputer-aidedcompositionsystem.Manyoftheabovepropertiesarealsokeydefiningfeaturesofnon-musicalMAS.
http://jasss.soc.surrey.ac.uk/18/2/16.html 1 20/10/2015
2.2 Thesysteminthisarticleisanon-realtimesystemwhichworkswithasmalltomediumnumberofagents–i.e.nothundredsofagents,itgeneratesmultipletunesinparallel,anditisfocusedoncomputer-aidedcompositionnotonmodellingthecompositionprocessormusicalculture.
Table1:MusicalMulti-AgentSystems
System Complexity Homog/Het
Comm Tune InitialHierarchy
Realtime
Size Model/Func
SwarmMusic(Blackwell&Bentley2002)
Low Het No 1 Flat Y 21 F
AntColonyMusic((Clairetal.2008) Low Homog No 1 Flat Y FSwarmOrchestra(Bisig&Neukom2008)
Low Homog No 1 Flat Y F
SocietyofMusicAgents(Beyls2007 Low Homog Sync 1 Flat N FMMAS(Wulfhostetal.2003) Higher Het ASync 1 Flat Y 8 FMusicalAgents(Fonseka2000) Higher Het Async 1 Flat Y FAndante(Ueda&Kon2003) Higher Het Async 1 Flat Y FVirtuaLatin(Murray-Rustetal.2005) Higher Het Sync 1 Hierarchy N 1 FMAMA(Murray-RustandSmaill2005) Higher Het ASync 1 Hierarchy Y FKineticEngine(Eigenfeldt2009) Higher Het ASync 1 Hierarchy Y FCinBalada(Sampaioetal.2008) Higher Het ASync 1 Flat N FAALIVE(Spiceretal.2003) Higher Het ASync 1 Hierarchy Y F
2.3 Thesystemswhichareclosesttotheoneinthisarticle(andnotlistedinTable1)arenowexaminedinmoredetail.TheDahlstedtandMcBurney(2006)systemusesagentswhichhavedifferentexplicitgoalsthatrepresentdifferentpartsoftheprocessofmusiccomposition.Anexampleisgivenofanagentwhosegoalistoreducesoundobjectdensityifthepopulationofthesystem'ssoundlandscapebecomestoocluttered;anotherisgivenofanagentwhodoestheopposite.Bothagentswouldtakeintoaccountthemusicalcontextwhiledoingthis.Theresearchersexplicitlyintendtoutiliseemergencetogenerateinterestingmusic.Thisisasimilaritywiththesysteminthisarticle,thoughkeydifferencesare:theDahlstedtandMcBurneyagentsactonasinglemusiccompositiontogether,whereasagentsinthisarticleeachhavetheirownrepertoireswhichcandevelopinparallel,anddonothaveexplicitanddistinctgoals.
2.4 Miranda's(2002)systemgeneratesmusicalmotifsinawaydesignedtostudytheevolutionofculture.Inthiscasetheagentsuseatwo-wayimitationproceduretobondsocially.Agentscanstorearepertoireoftunesandhaveabasicbiologicalmodelofanadaptivevoiceboxandauditorysystem.Agentspickotheragentstointeractwithrandomly.
2.5 WhentwoagentsAandBinteractthefollowingprocessoccurs:ifagentAhastunesinitsrepertoireitpicksonerandomlyandsingsit,ifnotthenitsingsarandomtune.Thesetunesarethreenoteslonganddonotgrowinlength.AgentBcomparesthetunefromAtoitsownrepertoireandifitfindsonesimilarenough,playsitbacktoagentBasanattemptedimitation.ThenagentBmakesajudgementabouthowgoodtheimitationis.Ifitissatisfiedwiththeimitationitmakesa"re-assuring"noisebacktoagentA,otherwiseitdoesnot.BasedonthesuccessoftheimitationAgentsAandBupdatetheirrepertoiresandtheirvoiceboxsettingstotrytoimprovetheirchancesofsociallybondinginlaterinteractions–e.g.bydeletingorre-enforcingtunes,ormakingrandomdeviationstotheirvoiceboxparameters.Theaimofthesystemistoseehowtherepertoireisgeneratedandaffectedundersuchsocialpressures.Asaresultofthesocialbondinginteractionsacommunityrepertoirewasfoundtoemerge.
2.6 Gongetal.(2005)producedasimplemusicgeneratingsystemwithasimilarpurposetoMiranda( 2002)–investigatingtheemergenceofmusicalculture.Theagentsstartwithasetofrandommotifs,togetherwithdifferentagentsbeingequippedwithdistinctbutverysimpleaestheticevaluationfunctions(forrhythm,pitch,etc.).Anagentplaysitstunetoanotheragentandifthesecondagentfindsthetuneunpleasant,itmodifiesit(basedonitsmusicalevaluation),andplaysitbacktothefirstagent.Ifthefirstagentthinksthemodifiedtuneisbetterthanitsoriginal,itdeletesitsoriginalandstoresthemodifiedversion.Asagentsinteractthisleadsto"morepleasant"motifsemerging.Also,usinganinteraction-historymeasure,thesociallinkbetweenfirstandsecondagentisstrengthenedsothattheyaremorelikelytointeractinthefuture.Howeverifthefirstagentdoesnotpreferthemodifiedtunetoitsownversion,itdiscardsitandthelinkbetweenthetwoagentsisnotstrengthened.Itwasfoundthatintheemergentsocialnetworktheagentstendedtoclusteraccordingtotheiraestheticpreferencefunction.Thissystemhasacoupleofsimilaritiestotheoneinthisarticle:itutilizesMASsocialnetwork/trusttechniques(Ramchurnetal.2004)todecidewhointeractswithwhom,andineachinteractionagentsvarytheirrepertoirebasedontheiropinionoftheotheragent'srepertoire.Thekeydifferencesbetweenthissystemandtheoneinthisarticleisthatagentsinthisarticlehavenoexplicitevaluativemelodicintelligence,andtheycanextendthenumberofnotesintheirrepertoire;andfinallythesocialnetworkinthisarticleisusedtogeneratehierarchicalmusicstructurewithinanagent'srepertoirenottoexperimentwiththeclusteringofagentsaccordingtotheirrepertoires.
2.7 TheA-Rhythm(Martins&Miranda2007)systemsetsouttoexaminetheapplicationofmulti-agentsystemstoalgorithmiccomposition.Currentreportsfocuson,likeMirandaandGongetal.,investigatingtheemergenceofsocialclusters,andaresolelybasedonrhythmicrepertoire.A-Rhythmhassomesimilaritiestothesysteminthisarticle:theagentscommunicateandprocessoneatatimeserially,ratherthaninparallel,andtheirmusicalcontentgrowslonger.HoweverA-Rhythmfocusesonrhythm,i.e.isnon-pitched.Alsothesimilaritymeasuresaremoredirectlybasedonthemusic,ratherthanaffectivecontentofthemusic.FinallyA-Rhythmusesmeasuresforthepopularityofrhythmsinanagent'srepertoire,butnotforthepopularity/trustofagents.Agentsinthissystemcantransformtheirrepertoiresbasedoninteraction–usingcertainrhythmictransformationrules,ratherthantheaffective-basedtransformationsusedinthisarticle.Anumberofexperimentsaredonebasedondifferentinteractionapproaches,andtheresultingpopulationandrepertoiredynamicsareexamined,showingthepotentialfortheemergenceofstructuredrhythmicrepertoires.
Multi-agentAffectiveSocialCompositionSystem:MASC
Overview
http://jasss.soc.surrey.ac.uk/18/2/16.html 2 20/10/2015
3.1 Thesysteminthisarticle–themulti-agentaffectivesocialcompositionsystem(MASC)–isnowpresentedinoverview.AgentsinMASCareinitializedwithatunecontainingasinglenote,andovertheinteractionperiodeachagentbuildslongertunesthroughinteraction.Figure1showsanoverviewrepresentationofacollectionofagents.
Figure1.SixMASCagentsinavarietyofaffectivestateswithoneagentperforming.
3.2 Thefollowingaresomeofthekeyfeaturesofthesystem.MASCusuallyconsistsofasmall-mediumsize–2to16–collectionofagents,butcanbemore.EachagentcanperformmonophonicMIDItunesandlearnmonophonictunesfromotheragents.Anagenthasanaffectivestate,anartificialemotionalstatewhichaffectshowitperformsthemusictootheragents;e.g.a"happy"agentwillperformtheirmusicmore"happily".Anagent'saffectivestateisinturnaffectedbytheaffectivecontentofthemusicperformedtoit;e.g.if"sad"musicisperformedtoahappyagent,theagentwillbecomealittle"moresad".Agentscanbemadetoonlylearntunesperformedtothemiftheaffectivecontentofthetuneissimilarenoughtotheircurrentaffectivestate;learnedtunesareaddedtotheendoftheircurrenttune.Agentsdevelopopinions/trustofotheragentsthatperformtothem,dependingonhowmuchtheotheragentscanhelptheirtunesgrow.Theseopinionsaffectwhotheyinteractwithinthefuture.
AffectiveModels
3.3 Beforegoingintothedatastructureswithineachagentindetail,theissueofaffectivemodelswillbecovered.ThereisavarietyofapproachesforaffectiverepresenationwhichcanbebroadlydividedintheDimensionaltypeandtheCategorytype(Zentneretal.2008).Categoryapproachesrangefrombasicemotiondefinitions–whichassumesthatsomeemotionsaremorefundamentalthanothersandattemptstolistthese;tothemoreeverydayemotionlabelsystems–whichdonotattempttocategorizebasedonanemotionhierarchy.Arecentcategory-basedapproachforemotionistheGenevaEmotionMusicScales(GEMS)approach(Zentneretal.2008)whichattemptstoprovidecategoriesoptimalformusicalemotion.This
http://jasss.soc.surrey.ac.uk/18/2/16.html 3 20/10/2015
isdonebyfirstinvestigatingthroughpsychologicaltestswhichsortsofemotionaremostcommonlyexpressedtopeoplebymusic.AlthoughGEMSdoesgetuserstoscorethecategorywithanintegerfrom1to5,thefactithasupto45categoriesputsitmoreintherealmofcategoricalthanthedimensionalsystemsnowdiscussed.
3.4 TheDimensionalapproachtospecifyingemotionutilizesann-dimensionalspacemadeupofemotion"factors".Anyemotioncanbeplottedassomecombinationofthesefactors.Forexample,inthispaper,twodimensionsareused:ValenceandArousal(Lang1995).Inthismodel,emotionsareplottedonagraphwiththefirstdimensionbeinghowpositiveornegativetheemotionis(Valence),andtheseconddimensionbeinghowintensethephysicalarousaloftheemotionis(Arousal).ThisisshowninFigure2.Justascategoryapproacheswouldnotclaimtolistallpossibleemotions,sodimensionalapproachesdonotclaimtobecomplete.Itisnotknownifemotionscanbepinpointedbasedonuniqueindependentdimensions.Otherdimensionalapproachesincludethethreedimensionalvalence/arousal/dominancesystem(Oehmeatal.2007).InthiscaseValenceandArousalhavethesamemeaningasinthe2Dversion.Howeverinthe2DapproachFearandAngerarebothlowvalence,higharousal.Inthe3Dversion,Dominancedifferentiatesemotionssuchasanger(highdominance)andfear(lowdominance);angercanbeseenasmoreofanactiveemotion,fearasmoreofare-activeone.TherearealsoexamplessuchasCanazzaetal.(2004)whereatask-specificmood-spaceisconstructedforexpressiveperformanceusingexperimentsandprinciplecomponentanalysis.Inthatparticularcasethedimensionsarenotexplicit.
Figure2.TheValence/ArousalModelofEmotion
AgentDataStructures
3.5 Eachagentcontainsthreedatastructures.Thefirstisanagenttune,amonophonictuneinMIDIformat.Thesecondisanagentaffectivestate–anumberpair[valence,arousal]representingtheartificialaffectivestateoftheagentbasedonthevalence/arousalmodelofaffectivity.Thisisthemostcommondimensionalaffectiverepresentationincomputermusic.AshasbeenmentionedValencereferstothepositivityornegativityofanemotion–e.g.ahighvalenceemotionisjoyorcontentment,alowvalenceoneissadnessoranger.Arousalreferstothearousalleveloftheemotion–forexamplejoyhasahigherarousalthanhappiness,thoughbothhavehighvalence,andangerahigherarousalthansadness,thoughbothhavelowvalence.Alinguisticelementneedstobeclarified.TheuseofaffectivelabelssuchashappyandsadareusedtoassistclarityinintroducingthereadertotheconceptsofMASC;theyarenotmeanttobetakenliterally.Forexamplehappyreferstoaregionofhighvalenceandarousalvalues,andsadreferstoaregionoflowvalenceandarousalvalues.Thesamegoesforanywordswhichmayseemtoimplythatagentshaveanykindofpersonification,ordeeperintentionalorbiologicalmodel.Suchlanguageismerelyashorthandforclarifyingfunctionality.
3.6 Thethirdandfinalagentdatastructureisaninteractioncoefficientlist,whichisalistofinteractioncoefficientsofalltheotheragentsinthecollection.Thesearenon-negativefloatingpointnumberswhichmeasurehowpopulartheagentfindseachoftheotheragents.Theconceptofinteractioncoefficientisusedheretoattempttocreateemergentcompositionalhierarchies,aswillbedemonstrated.Anotherwayofthinkingofinteractioncoefficientatthispointistoconsideranimaginedmotivationforanagent.TheaimofMASCis–startingwitheachagenthavingasinglenote–tobuildactualmelodies.Soanagentshouldwantnotes.AnagentA'sinteractioncoefficientmeasureofanother,sayAgentB,isbasedonthenotecountandnumberofperformancesithasaddedfromBtoitsowntune.
3.7 Anagentalsohasanumberofinternalprocessingfunctions.Theperformanceoutputchoicefunctioninvolvesanagentchoosingwhotoperformto,basedontheagent'sinteractioncoefficientlistofotheragents.Itwillonlyperformtoagentsitfindsusefulenough.Theperformanceoutputtransformfunctioninvolvestheagentplayingitssinglestoredtuneasaperformancetoanotheragent,withmusicalfeaturesbasedonitsowncurrentaffectivestate.Theperformanceinputestimatefunctionallowstheagenttoestimatetheaffectivecontentofatuneperformedtoitbyanotheragent,andadjustitsowninternalaffectivestatebasedontheaffectivecontent.Anagent'sperformanceinputchoicefunctioninvolvesitdecidingwhethertostore
http://jasss.soc.surrey.ac.uk/18/2/16.html 4 20/10/2015
aperformancefromanotheragent,andisbasedon:(a)theaffectivecontentofthatperformance,and(b)howmanynotesareinthelisteningagent'scurrenttune–anagenthasafinitetunelengthmemorywhichcanfillup.Theperformanceinputinteractioncoefficientfunctionletstheagentupdateitsinteractioncoefficientmeasureofanotheragentbasedonthatagent'sperformance.Finallytheperformanceinputaddfunctionletstheagentstoreaperformancebyconcatenatingittotheendofitscurrenttune.AnexampleinteractioncycleisshowninFigure3.Thiscycleisrepeateduntilthedesiredcompositionalresultisreached.
Figure3.ExampleInteractionCycle
PerformanceOutputTransformFunction
4.1 Thefunctionforperformanceoutputtransformwillnowbeexaminedinmoredetail.Beforeperformingitstunetoanotheragent,anagentwilltransformitstuneinacompositionalway.
CompositionalTransforms
4.2 Twotypesofcompositionaltransformationsareapplied–linearfeaturetransformsandakeymodetransformintoCmajororCminor.Theaimofthisworkwasittoinvestigatetheeffectsofmulti-agentemergenteffects,ratherthanthetransformationsthemselves–hencethetransformationswerekeptassimpleaspossible,forgoingnon-linearityandpsychophysicalaccuracy.Theycanbecomparedtothesimplisticlinearrulesusedinswarmorflockingsystems(Reynolds1987).Clearlysuchlinearrulesareanover-simplificationofbird/insectbiologyandpsychology.However,whatisofinterestisthatsuchsimplerulescancreatesuchcomplexdynamics.Theunderlyingelementsbeingsimplifiedherearetherelationshipsbetweenmusicandemotion.Thisareahasbeenmeta-surveyedinLivingstoneetal.(2010),whichthenestablishedaseriesofrulesfortransformingmusictoexpressemotions.Theserulesareshownintable2.
Table2:TransformationRulesproposedin(Livingstoneetal.2010)
EmotionLabel Valence Arousal FeaturesTempo Loudness Pitch Keymode
Happy Higher Higher Increase10BPM Increase5db +4 MajorAngry Lower Higher Increase10BPM Increase7db 0 MinorSad Lower Lower Decrease15BPM Decrease5db -4 Minor
http://jasss.soc.surrey.ac.uk/18/2/16.html 5 20/10/2015
Tender Higher Lower Decrease20BPM Decrease7db +4 Major
4.3 Equations(1)to(4)showtheresultoftransformingtheseintosimplifiedlinearrulesformulti-agentsysteminteraction.Theprocessofthissimplificationisexplainedbelow,aftertheirpresentation.WhenanagentAisabouttoperformandhasaparticularlevelofvalence,writtenvalenceA,andarousal,writtenarousalA,itwillfirstcompositionallytransformitsstoredtunebasedontheeffectsofequations(1)to(4).Theprimedvaluesonthelefthandsideoftheequationsarethedefiningfeaturesofthecompositionallytransformedmusic,andareusedtounambiguouslygenerateatransformedMIDIfile.Thepre-transformationvaluesIOIi(A),duri(A),loudi(A),andpitchi(A)are:theinter-onsetintervalbetweennoteiandthenextnotei+1,thenotedurationinseconds,theMIDIloudness,andMIDIpitchofthei-thmusicalnoteofAgentA'sstoredtune.Thethetavalues–θonset,θloud,andθpitch–definetheaffectivesensitivityofthetransformation–i.e.howmucheffectachangeinAgentA'svalenceorarousalwillhaveonthetransformation.Theyarethemaximumvariationpercentagebarsaroundthecurrentfeaturevalue.
IOIi(A)'=IOIi(A)(1−θIOIarousalA (1)
duri(A)'=duri(A)(1−θIOIarousalA) (2)
(3)
(4)
4.4 ForexampleifθIOIis0.25,thenbyEquation(1)theonsetwillvaryfrom25%belowitscurrentvalueto25%aboveitscurrentvaluewhenarousalvariesfrom-1to1.IfatransformationgoesabovethemaximumMIDIvalue(127)thenitissetto127.Similarlyifitgoesbelow1itissetto1.NotethatθIOIisusedbothforonsetsanddurationsothatasgapsbetweennotesareincreasedordecreased,thedurationofthesamenotesisincreasedanddecreasedbythesameamount.
4.5 ThemappingbetweenTable2andEquations(1)to(4)isexplainedasfollows.Asmalleraverageinter-onsetintervalinapieceofmusicleadstoahighertempo(Dixon2010),andequation(1)meansthatahigherarousalwillcreateasmallerinter-onsetinterval,andthusahighertempo.ThisapproximatelycapturesthelineardynamicsofthetempocolumninTable2wheretempochangesinthesamedirectionasarousal,butisnotaffectedbyvalence.Durationin(2)ischangedproportionaltointer-onsetinterval–sothatwhennotesareclosertogether(duetoahighertempo)theywillbeproportionallyshorter,aswouldbeexpected.
4.6 Equation(3)isalinearsimplificationofthefactthatTable2showshowchangingvalencetendstocauseloudnesstoincreaseinthesamedirection,andchangingarousalalsotendstocauseloudnesstoincreaseinthesamedirection.Equation(4)isbasedonthefactthatTable2showspitchtobechangedbychangesinvalenceandarousal,butslightlymoresobychangesinvalence.
4.7 Table2wasnotoriginallyaprescriptionforalinearmodel,inthesenseitsaysnothingaboutwhathappensinbetweenthefourstates.Evenifthevalence/arousalmodelofemotionwascomplete(whichclearlyitisnot)therewouldcertainlybenon-linearbehaviorbetweentheoriginandthefourpointsreferencedinthistable.Thisisonereasonwhynoattemptwasmadeintheequationstocalculatepreciselinearcorrelationcoefficients.
4.8 SotheequationsarenotmeanttobeanaccuratelinearmodelofthebehaviorinTable2,butthesimplestpossiblelinearmodelofmusicfeaturesandemotioninformedbyTable2.Themodeldoesnotclaimthatallhighpitchedmusicishighervalence,orthatalllowtempomusicislowarousal.Contra-examplescanbefoundforbothoftheseinmusic:forexamplehighpitchedmelancholyviolins,orslowbutgrandandinspiringorchestralpieces.Therearehowevernocompletemodelsformusicandemotion,justastherearenocompletemodelsforbirdandinsectflocking.Forexample,youcankeepaddingorremovingbirdsfromaflockingmodel,andtheflockwillincreaseordecreaseinsizeinanunlimitedway–obviouslyacontra-exampleoftheflockingmodel.Henceitisarguedthattheincompletenessoftheabovelinearmodelisacceptableforthepurposesoftheworkherepresented.
4.9 Onemusicfeaturewhichcannotbechangedlinearlyiskeymode,soadifferent–butstillsimple–approachisused.ItislargelybasedonTable2butwithoneadjustment.Forpositiveemotionamajorkeyisutilizedandfornegativevalencewithnegativearousal(e.g.sadness),aminorkeyisutilized.Fornegativevalenceandpositivearousal–e.g.angerorfear–eachnoteinthetuneistransformedtoCminorthenmovedalternatelyupordownasemitone;thisisdesignedtoinjectanatonalelementtothemusic.Forexamplethesequence"CEbDFEbGC"wouldbecome"DbDEbEEGbDb".Thisisbasedontheideathatfearandangercanberepresentedbyatonality(Chongetal2013).ThiswillimpacttheeffectofEquation(4)whichalsoraisesandlowerspitchduetovalence.Howeverthechangesinpitchduetovalencein(4)–intheexperimentsdetailedlater–areofasignificantlygreaterorderthanonesemi-tone.Thustheimpactoftheatonaltransformationisminimalon(4).Alsoequation(4)isalinearsimplification,sothereisnoclaimtoitbeingaccuratetowithinasemi-toneintermsofitsvalencerepresentation.
4.10 Thetransformisalgorithmicanddeterministic–itsearcheseithersideofthecurrentnotesforanoteinthenewmodewhichdoesnotviolateaMIDIboundary–i.e.notoutoftheMIDI128parameterrange.SosupposeanagentAhasstoredatunefromahappyagentwhichisamajorkey.IfagentAthenperformsitstunewhilesaditwillconvertallofitstune,includingthemajorpartitreceivedfromanotheragent,intotheminormode.Thecurrentversioninthisarticlehasnoabilityforactualkeycompositionfunctionality,hencethereasonforusingonlyCmajorandCminor.
TuneAffectiveEstimationFunction
5.1 Alinearequationisusedtomodelthelisteningagent's,sayagentB,affectiveestimateofaperformancebyagentA–thisisshowninequations(5)and(6).
valenceEstB=xpmean(pitchA)+xlmean(loudA)+xkkeyModeA+xIOImean(IOIA)+x0
(5)
arousalEstB=ypmean(pitchA)+ylmean(loudA)+yIOImean(IOIA)+y0
(6)
http://jasss.soc.surrey.ac.uk/18/2/16.html 6 20/10/2015
IntheseequationspitchAandloudArefertotheaverageMIDIpitchandMIDIloudnessofanagentA'sperformanceheardbyB.keyModeArepresentsthekeymodeofA'stuneestimatedusingakeyprofile-basedalgorithm(Krumhansl&Kessler1982)definedashavingvalue2foraminorkey,and1foramajorkey.Akeyprofileisavectorof12elements,oneforeachnoteinthescale.Eachkeyhasitsownprofile.Pitcheswhichfitwellintoascalehaveahigherweightinginitskeyprofilevector,andthosewhichfitlesswellhavealowerweighting.TheseweightingswerecalculatedforeachkeyfromperceptualexperimentsinKrumhanslandKessler(1982).Thustheycanbeusedtofindwhetheraparticularsetofnotesfitsbestintoamajororaminorkey.
5.2 ThexandycoefficientsintheEquationsareconstantsestimatedbylinearregression.Theseareestimatedinaone-offprocessasfollows.Asetof1920randomMIDIfileswasgenerated,ofrandomlengthsbetween1and128notes.EachMIDIfilewastransformedfor10givenandequallyspacedvalenceandarousalvaluesbetween−1and1usingtransformationequations(1)to(4),andkeymodetransformations.
5.3 ThenalinearregressionwasrunontheresultingtransformedMIDIfilesagainsttheknownarousalandvalencevalues–basedonequations(5)and(6).Theresultingcoefficientsweretestedonaseparatesetof1920transformedrandomfiles,andtheaveragepercentageerrorswere10%forvalenceand9%forarousal.Theseareconsideredtobesufficientlyaccurategiventhatactualhumanmusicalemotionrecognitionerrorratescanbeashighas23%andotherfarmorecomplexartificialmusicalemotiondetectionsystemshaveratessuchas81%(Legaspietal.2007).Theactualcoefficientsforpitch,loudness,keymodeandIOIwererespectivelyinequations5and6forx=[−0.00214,0.012954,1.1874,−0.6201]andy=[0.003025,0.052129,−1.4301,0.59736];withtheadditiveconstantsforxandyrespectivelyof0.61425and−4.5185.
5.4 Thelinearestimatorisusedintwoaspectsoftheagents–firstlyforanagenttodecidewhetherornottoaddaperformancetoitsowntune,andsecondlyforanagenttobeinfluencedbytheapproximatedemotionalcontentofaperformanceithasheard.Equations(7)and(8)belowareusedtoupdatethevalenceandarousalofagentBafteraperformancefromagentA.Theγ(gamma)constant–between0and1–defineshowsensitiveanagentistoaffectivestatechange–i.e.theamountofchangetovalenceandarousal.Ifitissetto1thenthenewvalenceandarousalvalueswillbetotallychangedtotheestimatedvaluesoftheperformancetheagenthasjustheard.Avalueof0willleadtonochange.Valuesbetween0and1willcausetheestimatetohaveaproportionallygreatereffect.
valence'B=(1−γv)valenceB+γvvalenceEstA (7)
arousal'B=(1−γa)arousalB+γaarousalEstA (8)
5.5 OncetheagentBhasdecidedwhetherornottoappendtheperformancefromA–andifso,hasdoneso–itwillupdateitsvalenceandarousalbasedonEquations(7)and(8).Infuture,whenitnextperformsatune,itwilltransformitbasedonitsnewvalenceandarousalstate.Itisdesignedsothatthroughthisseriesofupdatingaffectivestatesandtheagenttunecommunicationandsystem,newmusicalstructureswillemerge.
5.6 Itisworthtakingamomenttodiscusstheaboveprocess,inparticulartheaffectivecommunicationandestimationelements.AnalternativewouldhavebeenforthesystemtodirectlytransmitthevalenceandarousalofagentAtoagentB,ratherthancomputingcoefficientsforEquations(5)and(6)andgoingthroughtheprocessofestimatingvalenceandarousalfromthemusicfeatures.Theestimatingthecoefficientsfor(5)and(6)couldbeseenasquitearecursiveprocess:firstanapproximationwasmadeofhowvalenceandarousalcanbecommunicatedthroughmusicfeatures,andthenanapproximationwasmadeofhowmusicfeaturescommunicatevalenceandarousal.Wouldn'tithavebeensimplertojustcommunicatethevalenceandarousaldirectly;or–ifwewantedtoobservethemusicaleffects–stillperformthecompositionaltransforms,whichcommunicatingthevalenceandarousaloftheperformingagentdirectly?
5.7 Thesesimplificationswouldhaveremovedtheartificiallifefoundationfromtheformulation.Themusicwouldhavenotbeenpartoftheprocess,butmerelyinfluencedbytheprocess.Whatmakestheformalismofinterestisthatthemusicispartoftheprocess,andthusanycreativeresultsareemergentfromtheprocess.Furthermore,hadtheagentscommunicatedtheirvalenceandarousaldirectly,withsayarandomcommunicationerror,thenitwouldbecomealessnovelmulti-agentsystem.ItwouldhavebeenanMASinwhicheachagenthadtwonumericparametersandagentswhosenumericparameterswereclosewouldtheninfluenceeachother'sparametersmorestrongly.Thisisatypeofartificiallifesystemthathasbeenstudiedmanytimesbefore,andleadstoagentstendingtoclusterintogroupsofsimilarparametervalues–i.e.similarvalenceandarousalvalues.Byputtingthemusicatthecentreoftheparametercommunicationnotonlydoesitcreateanovelartificiallifeprocessbutalsomakesthemusicemergenttotheprocess.ThussomethingislearnedaboutthedynamicsofanewALsystem,andalsoabouthowapplicablethatsystemmightbetoalgorithmiccomposition.
5.8 Attheheartoftheexperimentsarethequestions:canmusicalcommunicationinmulti-agentsystemsbeusedasaformofalgorithmiccomposition?Music,incommunicationterms,ismostcommonlyconsideredaformofemotionalcommunication(Juslin&Laukka2004).Pastresearchhasmostcommonlylinkedmusicalfeaturestocommunicatedemotion.Howeverthiscommunicationisimperfect:peopledetectemotioninmusicambiguously,andcommunicateinemotionambiguously.Sothesystemasdesignedcapturesmanyelementsofmusicalcommunicationinamorerealisticwaythatsimplycommunicatingvalenceandarousalwithanerror.Itgeneratesmusicalfeaturesbasedontheemotionbeingcommunicated.Thelisteningagentisthenaffectedbythemusicalfeatures.Thiseffectisthesimplestnon-directmethodpossiblebasedonmusicalfeatures:alinearmodelbasedonthenon-invertibleequations(1)to(4).This–althoughsignificantlylesssimple–itistheequivalentintheflockingexampleofagent'smovementandtheirperceptionofeachother'smovement.However,unlikemovement,theprocesshereisnotanimmediatelyvisibleone;itconcernstheagents'emotionswhicharenotsosimplyobservableorcommunicable.Infactevenwhenhumanbeingsattempttocommunicateemotionsdirectly,therearelimitations;whichisoneofthereasonsthattheartsareoftenconsideredtobeformsofemotionalexpression,abletocommunicateinwaysthatlanguageisnot(Juslin&Laukka2004).
PerformanceInputInteractionCoefficientFunction
6.1 BeforeanAgentAperformstoanAgentBitcomparesitsinteractioncoefficientmeasureofAgentBtotheaverageofitsinteractioncoefficient(IC)forotheragents:
IC(A,B)>mean[IC(A,allagents)] (9)
whereIC(A,B)isA'sinteractioncoefficientmeasureofB.Ifitisnot,thenitdoesnotperformtoAgentBandmovesontothenextagent.TheincreaseinInteractionCoefficientisproportionaltothelengthoftuneithasadded.SothemorenotesinAgentB'spastperformances,thegreateritsinteractioncoefficientwillbeviewedbyAgentA.IfagentAaddsatunefromagentBoflengthNnotesthen:
http://jasss.soc.surrey.ac.uk/18/2/16.html 7 20/10/2015
IC(A,B)=IC(A,B)+d.N (10)
6.2 Theparameterdisaconstantcalledtheinteractioncoefficientupdaterate.Thiscanbevisualisedasanagent'sbasicresourcesbeingtunes–sothemorenotesinanagent'stune,thegreateritspotentialinteractioncoefficienttootheragents.Howevertheactualreasonforincludinginteractioncoefficientfunctionality,andmakinginteractioncoefficientproportionaltothenumberofnotesinaperformingagent'stuneisprimarilytogenerateaninteraction/socialhierarchyamongsttheagentswhichinfluencesthehierarchicalstructureofthecomposedmusic.Bearinginmindthatanagentwillonlyperformtootheragentswithahighenoughinteractioncoefficient,itcanbeseenthatagentswhichperformmorethanlistenwilltendtohavelowerinteractioncoefficients.Furthermoreagentswhichmostlylistenandstorewillhavelongertunesandhigherinteractioncoefficients;andagentswithhigherinteractioncoefficientswilltendtobeselectedaslistenersmoreoften
6.3 Sothesystemisdesignedtoturntheagentpopulationintoasetofagentswhotendtoperformandhaveshortertunes,andasetofagentswhotendtolistenandstore.Theaimisforlowerinteractioncoefficientagentstobefocusedonprovidinglowerelements–i.e.shorterelements–ofthemusicalhierarchy.
AnExampleCycle
6.4 Anexamplecyclewillnowbeshown.Inthisexampleathreeagentsystemisexamined.Agent1istheperformerandstartsbyconsideringperformingtoAgent2;Agent1'smeasureofAgent2'sinteractioncoefficientisverylowinthisexample;Agent1'smeasureofAgent3'sinteractioncoefficientisveryhigh;Agent1'saffectivestateishighvalenceandhigharousal–i.e.happy;andAgent3'saffectivestateislowvalenceandlowarousal–i.e.sad.
6.5 Asthecyclingstarts,becauseAgent1'sinteractioncoefficientofAgent2isverylow,Agent1doesnotevenperformtoAgent2.ItselectsthenextAgentiteratively.Agent3isselectedbecauseagentsareorderedbynumericallabel.Agent1'sviewofAgent3'sinteractioncoefficientisveryhigh–soAgent1performsitstuneT1adjustingittomakeithappierbecauseofitshighvalenceandarousalstate,givingaperformanceP1.
6.6 Agent3estimatestheaffectivecontentofAgent1'sperformanceP1andgetsaresultofhighvalenceandarousal–i.e.itestimatesitisahappyperformance.BecauseAgent3'saffectiveestimateofAgent1'stuneishighvalenceandarousalbutAgent3'sstateislowvalenceandarousal–i.e.verydifferenttohappy–Agent3discardsAgent1'stune.HoweverAgent3stilladjustsitownsaffectivestatetowardsitsestimateoftheaffectivecontentofperformanceP1i.e.itbecomesalittlemorehappy.NeitherAgentmakesanyadjustmenttotheirinteractioncoefficientmeasuressincenoperformanceswerestored.NextAgent2becomestheperformer,andthefirstagentisiterativelychosentolisten–i.e.Agent1.
Experiments
7.1 Theissueofhowtoevaluateanalgorithmiccompositionisbynomeansagreedintheresearchcommunity.Parametric/Example-basedinvestigationsarecommonininvestigatingalgorithmiccompositionsystems(e.g.,Beyls2007;Fonseka2000;Anders2007).SuchexperimentsweredoneanalysinghowMASCoutputrespondedtovariousparameterchanges,givingobjectiveinformationonMASC'sbehaviourforapotentialuser.Suchexperimentsareimportantbecausetheyprovideinsightintothedynamicsofthesystem.Itshouldbenotedthatinthissystemthereareanumberofparameterswhichneedtobeset;itisbeyondthescopeofthisarticletodescribeallofthem.Thosenotmentionedexplicitlyhereweresettodefaultvalues.
MelodyGeneration
7.2 Ahelpfulwaytoindicatethatthissystemcanproducenon-trivialmelodies,inspiteofitslackofmelodicintelligence,istoexploretheoutputspaceforsomedifferentinitialaffectivestates.UsuallyinaMASonewouldwanttoperformastatisticalanalysisofbehaviour,butitisanunsolvedprobleminalgorithmiccompositionastowhatstatisticsaremusicallyrelevant(Freitasetal.2012).Soinsteadspecificmusicalresultsarepresented.Figures4to8showagent6'sfinaltune,froman8agentsystemrunfor10cycles.Thesimilaritythresholdwas1,theinteractioncoefficientsystemwasswitchedoff,andvalenceandarousalupdaterates-gammainequations(7)and(8)–were0.001.AgentswereinitialisedeachwiththesinglenoteofmiddleC,ofduration0.5seconds.Fourinitializingstateswereusedwithdifferentlevelsfor[Valence,Arousal]pairs.Thesewere:Happy=[0.5,0.5];Sad=[−0.5,−0.5];Angry=[−0.5,0.5];Tender=[0.5;−0.5].Thefirstlettersofeachstatewereusedtoindicatetheagentinitialstates.ForexampleTTTTTTHHis6Tenderand2Happy.
Figure4.8AgentsAAAAAAAA,10Cycles,and8AgentsAAAAAASS,10Cycles
http://jasss.soc.surrey.ac.uk/18/2/16.html 8 20/10/2015
Figure5.8AgentsSSSSSSSS,10Cycles,and8AgentsSSSSSSAA,10Cycles
Figure6.8AgentsSSSSSSHH,10Cycles,and8AgentsSSSSSSTT,10Cycles
Figure7.8AgentsHHHHHHSS,10Cycles,and8AgentsHHHHHHTT,10Cycles
http://jasss.soc.surrey.ac.uk/18/2/16.html 9 20/10/2015
Figure8.8AgentsTTTTTTSS,10Cycles,and8AgentsTTTTTTHH,10Cycles
7.3 Eachverticaltickisasemi-tone.Precisepitchvalueshavebeenremovedfromthegraphstoaidreadability,andbecauseabsolutepitchvaluesarenotthekeyelementhere,buttherelativeinterplayofstructureswhichitwillnowbearguedindicatenon-triviality.Toclarifywhythisbroadrangeoftunesisconsiderednon-trivialasmelodies,anumberofelementsarenowhighlighted.Musichasbeengenerateduptoandover50secondslong.Ifthetuneswereonly3to6notesorafewsecondslong,thesewouldbetrivial.Themelodiesarenotjustsimpledirectionalpitchpatterns,likesinglerepeatednotes,uniformlyrisingorfallingpatterns,orrepeated"zig-zags".Themelodiesarenotjustgroupsofrepeatednotes,e.g.5notesatonepitch,then4notesatanother,etc.Thepitchesvarymuchmorethanthat.Howevertheydonotvaryallthetime–therearetimeswhennotesarerepeated2or3times,asonewouldexpectinmelodies.Timingrepetitionisnottoohighortoolow.Melodieshavebeengeneratedwherethetempodoesnotstayconstant,butalsothetempodoesnotsimplyseemtovaryrandomly.Itwouldseemoddifnoteonlyraisedorloweredbyonepitchinalltunes.Inmusictherearesometimeslargerjumps.Howeveritwouldalsoseemoddifallthepitchchangeswerelarge.Ithasbeenshownthattunescanbegeneratedwhichavoidthesetwoextremes.Finally,themelodiescontainrecognizablenotegroupingswhicharerepeatedandtransformedtodifferentpitchesandtimings.Thisisexpectedbywesternlistenerswhousuallylooktoidentifyamotifstructure.
7.4 Ideallyitwouldbedesiredtohavesomemorescientificmeasureofnon-trivialityforthemelodies,howeverthereisnosuchagreedmeasure.Therehavebeensomeattemptstousemeasureslikeentropybutnoconclusiveresultshavebeenobtained(Kirke&Miranda2007).Anotherapproachmightbetousemeasuresofcomplexity,asitwasstatedatthestartofthispaperthatsuchcomplexitywasamajormotivationinusingartificiallifesystemsforcreativity.Howeverthereisnoconclusiveworkonartisticallyinformedmeasuresofcomplexity.Soalthoughtheaboveapproachisnon-scientific,itisfairlydetailedanddoescapturemanyelementswhichacomposerwouldconsidertobekeytonon-trivialityofmusic.
7.5 Insummarythefollowingexampleofafullcompositioncanbeheardhttp://cmr.soc.plymouth.ac.uk/alexiskirke/mapc.html
7.6 Thiscompositionhasbeenputthroughacomputersystemforexpressiveperformancetomakeitmorelistenablethroughasequencer(Kirke&Miranda2009).
7.7 Itwasstatedatthebeginningofthispaperthatanaimwasthatthenon-trivialmelodypitchstructureswouldbedevelopedbyagentswithoutexplicitmelodicknowledge.Givenwehavenowclaimedtheproductionofnon-trivialmelodystructures,wewillexaminetheclaimofnoexplicitmelodicknowledge.Thisdoesnotmeanheagentshavenomusicalknowledge.Infacttheagentshaveasignificantamountofhand-craftedmusicalknowledgeinthemusicaltransformationandaffectiveestimationformulae.However,thisknowledgedoesnotincludehowtosequentiallyconstructmelodies.Itincludeshowtochangekeymodes,timingsandpitches,butnomusicalrulesaboutwhichmusicalfeaturesshouldfollowwhich.Thebasisofmelodicstructureiswhichpitchesfollowandwhichnotetimingsfollowwhich.Thisisasignificantgapintheagents'explicitknowledgeaboutmusicwhichcanbereducedbycarefulhand-craftingoftherules,butnotremovedwithoutincludingorderingconstraintsinthemusicalequations.Thustheorderingofnotesinthissystememergesasaresultofthesocialinteractionsbetweentheagents.
7.8 Itwouldbeinformativetoseehowsimplethemusicalrulescouldhavebeenmadebeforethetunesfailedtobenon-trivial(withnon-trivialityformulatedbasedontheprocessdescribedearlierinthissection).Thismightbringthewholesystemcloserinphilosophytotheflockingsimulationsdiscussed.Howeverthisisbeyondthescopeofthiscurrentwork.
AgentAffectiveTuneandStateAdaptation
7.9 IthelpstounderstandtheMASCdynamicsmoreclearlybylookinginamoregeneralwayathowmusicfeaturesareeffectedbyagentinitialaffectivestates,aswellashowagentaffectivestatesarechangedasisdoneinFigures9and10.Notethattheseexperimentsinvolvedswitchingoffthesimilaritythresholdaswell,soastofocuspurelyonaffectiveinteractiondynamics.Thesefiguresexaminetheresultmusicfeaturesafter10cyclesofinterationofMASC,forthesystemof8agentsabove,aswellasforasmallersystemof2agents.ThepitchspreadisthedistancebetweenthehighestandlowestMIDIpitchinthefinaltuneaveragedacrossallagents,similarlywiththemedianpitch.Theaveragekeyisfoundusingakeyprofilingalgorithm(Krumhansl&Kessler1982).Figures9and10alsohavearrowstohighlighttheprogressionoffeaturesasinitialaffectivestatesarechanged.Theyfurthermorehavedashedellipsestohighlighthowclosetogethertheresultingfeaturesofthe2agentsystemaretotheir"equivalent"8agentsystem.
7.10 AkeyelementofMASCwhichwouldbenewformanycomposersutilizingit,wouldbetheassigningofinitialaffectivestatestotheindiviualagents.ThusthemoreintutitvetheresultsofsuchassignmentsthemoreusefulMASCcouldbe.Thismayatfirstseemcontradictorytowhatwassaidinitiallycautioningthereaderearlierontheirinterpretationoftheourusesofthewords"Happy"and"Sad"inrelationtoagents.Howeverthiscautionwastopreventintepretationthattheagentsthemselvesweresomehow"Happy"or"Sad".Theuseoflabelssuchastheseasashorthandforarousalandvalenceconfigurationsisstillvalid,andhelpfultothecomposerusingthissystem.Figures9and10provideinsightintotheeffectsofthisprocess.Itcanbeseenthatbroadlythereareunderstandablepatterns/trajectories,intermsofmusicfeatureresults,whentheinitialaffectivestatesarechanged.Forexampleincreasingthenumberofhappyagentsincreasesmedianpitchandreducespitchspread.
http://jasss.soc.surrey.ac.uk/18/2/16.html 10 20/10/2015
Figure9.EffectsofInitialAffectiveStatesin8AgentSystemonaverageIOIandKey,8CyclesRun
7.11 ThetunesinFigures8and9generatedbythe8agentsystem,togetherwithsixmoreinvolvingfurthercombinationsofH,S,AandT,wereplayedto10listeners.Itwasfoundthatwhenatleast6ofthe8agentshadthesameinitialvalenceandarousal,thenthelistenershada71%chanceofdetectingthatsamevalenceinthefinaltune,andan82%chanceofdetectingthatsamearousalinthetune.Althoughthesmallnumberofparticipantsmeansthattheresultsarenotverysignificant,theyarecertainlyindicativethat–giventhesupportoftheremainderoftheparametricevaluation–itisworthdoingasubstantialsetoflisteningteststoevaluatethispotentialaffectivepattern.ThiswouldfurtherhighlighttheabilityofMASCtobeusedinarelativelyintutivewayasacomputer-aidedcompositiontool.
Figure10.EffectsofInitialAffectiveStatesin8AgentSystemonPitchMedian/Spread,8CyclesRun
InteractionCoefficientandMusicalHierarchy
7.12 Ashasbeenmentioned,theinteractioncoefficientprocessisdesignedasanattempttoregulatetunegrowthinsuchawaythatcertainagentswillbecometuneprovideragentsandsomewillbecometunereceiveragents,thuscreatingahierarchyintheagent'sinteractionstructurewhichis
http://jasss.soc.surrey.ac.uk/18/2/16.html 11 20/10/2015
hopefullyreflectedinthehierarchalstructureofthemelodies.Inthisexperiment,an8agentsystemwasusedwithequallyspreadagentinitialaffectivestates,300notememorysize,32cycles,affectivesimilaritythresholdof1,pitchupdaterateof0.1,andIOIandloudnessupdateratesof0.5;valenceandarousalupdateratesweresetto0.1.Theinteractioncoefficientupdateratewassetto0.2;andtheinteractioncoefficientthresholdwassetto0.9.Figure11showstheevolutionofanagent'sinteractioncoefficientaveragedacrossallotheragents.Sothetopgraphshowstheaverageview/trustthatagents2to8haveofagent1byaveragingtheircoefficientvaluesforagent1.
7.13 After32cyclesthenumberofnotesthatagents1to8haveisrespectively:291,102,102,102,102,102,18,and5.ThesetunescanbeseeninFigure12.LookingatFigure11,itcanbeseenthatthisrelatestotheinteractioncoefficient–thehigheranagent'sfinalinteractioncoefficientthehigheritsnotecount.Table3showsineachcyclewhichagentsanagentreceivestunesfrom.Soforexampleincycle1,agents2to6receivetunesfromagent1.Incycle2agent1receivesatunefromagent2,butnootheragentsreceivetunes.
7.14 InTable3itcanbeseenthatthelowerinteractioncoefficientagentstendtogiveouttunes,whilethehigherinteractioncoefficientagentstendtoreceivetunes.Thelowernumberedagentshavehigherinteractioncoefficientbecauseoftheorderingofagentinteractionineachcycle.Thelowernumberedagentswillbereceiversfirst,andhaveachancetobuildupthesizeoftheirtunes.Thenwhentheybecomeperformers–givers–theloweragentswillreceivelargetunesfromthemandtheirinteractioncoefficientwillincreaseasaresult.
7.15 Toseehowthiscreatesthehierarchicalstructure,considerthatbyTable3Agent1'sfinaltunecouldbewrittenasaconcatenationofsub-tunes1021324354657629310411512613217318whereeachnumberindicatestheagentwhoperformed,andthesubscriptsarethecyclenumbers;anagent'stunevariesoverdifferentcycles–e.g.318isnotthesame32.BecausetheMASisaclosedsystem,alltunesinthisstructurearetheresultofatransformationonanotheragent'stune.
7.16 Soforexample21=2010'and32=3010''and76=7010'''.HeretheprimesrepresenttransformationsonAgent1'stuneduetoAgent1'saffectivestateatthetime.InthenextroundoftunesbeinggiventoAgent1thisgives29=2177'188'=(2010')77'188'
7.17 ThisexpansioncanbecontinueduntilthereisafulldescriptionofAgent1'stunesbasedonthewayinwhichthetunegrows.Thisdescriptionwillshowthebuildingstructureofthetune.Itwillnotnecessarilyshowtheperceptualstructureofthetune–thisisnotclaimed,butitwillshowhowthetunewasbuiltfromthemotifsandphrasesandsoforthofotheragents.Thisstructureisclearlyafunctionoftheagentinteractionhierarchy,andashasbeenseenthishierarchyisstronglyinfluencedbytheInteractionCoefficientfunctionality.
7.18 Thesediagramshighlightacharacteristicofthesystem–itssequentialupdaterule.Becauseagentsarealwaysupdatedinthesameorder,theagentswithalowerindexnumbertendtodevelopmuchlargertunesandhaveahighinteractioncoefficient.Anasynchronoussystemsimulationcouldhavebeenutilizedtoavoidthis–wherethenextagenttointeractisrandomlyselected.Howeverthiswouldhavemovedawayfromtheprocessofalgorithmic–i.e.non-random–compositionprocesseswhichthisworkwasdesignedtobuildupon.
http://jasss.soc.surrey.ac.uk/18/2/16.html 12 20/10/2015
Figure11.ChangeinmeanInteractionCoefficient(x-axis=numberofinteractions)
http://jasss.soc.surrey.ac.uk/18/2/16.html 13 20/10/2015
Figure12.Tunesafter32cycles
Table3:PatternofInteraction
ConclusionsandFutureWork
8.1 Amulti-agentsystemMASChasbeenpresentedwhichisaproofofconceptthatamulti-agentsystemcandevelopnon-trivialmelodypitchstructuresthroughaffectiveinteractionofagentswithoutexplicitmelodicknowledge.Agentshavenoexplicitknowledgeofwhichnotesshouldfollowwhich,howtheyshouldrepeat,andsoforth.Anagent'sonlycompositionalknowledgeofmusicisitsabilitytoextractaffectivedatafromthewholeandimposeaffectivefeaturesonthewhole.MASCalsodemonstratesthatmulti-agentsocialstructurescangeneratemusicalstructureonthematicandsectionallevelsaswellasonanoteorphraselevel.Thereweretwodemonstrations.Itwasdemonstrateddiagrammaticallyhowtheinteractionstructurewouldrelatetothemusicalstructure.Thenanexamplewasgivenshowingthemusicalstructurebuildingupandhowitrelatedtotheagents'socialstructure.AfinalcontributionofMASCisthelinearmusic-emotionanalyzingmodelwhichtakesasinputamonophonicMIDIfileandestimatesitsaffectivecontent.
8.2 Intermsoffuturework,thelisteningtestsperformedwerefairlybasic,usingasmallnumberofsubjectsandthusonlyindicative.MoreextensivetestsareneededtosupporttheabilityofMASCtobeusedinarelativelyintutivewayasanaffectivecomputer-aidedcompositiontool.SuchtestscouldalsobeusedtoexaminethedifferencebetweenthegenerativestructureinMASCtunes,andtheperceivedstructureforhumanlisteners.Thiswouldhelptoclarifytheeffectivenessoftheinteractioncoefficientapproachtohierarchicalstructuregeneration.
8.3 OnthesubjectofInteractionCoefficient,thereareothercontextsthatcouldbeinvestigatedtoinfluencefutureagentinteractions,besidestheirpastinteractionslists.Forexampleanagentcouldhavetimevaryingtrajectoriessetbythecomposer,whichcouldbiaselementsliketheirarousalandvalence,ortheextenttowhichtheiraffectivestatetransformsthemusictheyperform.ThiswouldprovideadditionalcontextsforthecomposertouseincontrollingtheMASoutput.
8.4 Anotherelementoffutureworkistheadditionofindeterminacy.Itwasdesiredtoexaminethemulti-agentsystemasaformofalgorithmiccomposition–hencekeepingthewholesystemdeterministic.Forexampleinpreviousworktheauthorshaveutilizedsemi-randomcommunicationerrorsbetweenagents,contributingtochangesintheirtunesandtransformations(Kirke&Miranda2011).Ashasalreadybeenmentioned,theuseofindeterminacywouldalsoallowforthesimulationofasynchronouscommunication–sothatitisnotalwaysthesameagentswhobeginthesingingcycle.
Acknowledgements
http://jasss.soc.surrey.ac.uk/18/2/16.html 14 20/10/2015
ThisworkwaspartiallyfinanciallysupportedbytheEPSRCfundedproject"LearningtheStructureofMusic,"grantEPD063612-1.
References
ANDERS,T.(2007).ComposingMusicbyComposingRules:DesignandUsageofaGenericMusicConstraintSystem.Thesis/Dissertation,QueensUniversity,Belfast.
BEYLS,P.(2007).InteractionandSelf-organisationinaSocietyofMusicalAgents.ProceedingsofECAL2007WorkshoponMusicandArtificialLife(MusicAL2007),Lisbon,Portugal.
BLACKWELL,T.&Bentley,P.(2002).ImprovisedMusicwithSwarms.Proceedingsofthe2002WorldCongressonEvolutionaryComputation,Honolulu,Hawaii.[doi:10.1109/CEC.2002.1004458]
BISIG,D.&Neukom,N.(2008).InteractiveSwarmOrchestra,anArtificialLifeApproachtoComputerMusic.Proceedingsofthe11thGenerativeArtConference,Milan,Italy.
BROWN,A.(2002).OpportunitiesforEvolutionaryMusicComposition.Proceedingsof2002ACMC,Melbourne,Australia.
CANAZZA,S.,DePoli,G.,Drioli,C.,Rodà,A.,&Vidolin,A.(2004).Modelingandcontrolofexpressivenessinmusicperformance.TheProceedingsoftheIEEE92,686–701.[doi:10.1109/JPROC.2004.825889]
CHONG,H.,Jeong,E.&Kim,S.(2013).Listeners'PerceptionofIntendedEmotionsinMusic.InternationalJournalofContents9(4),78–85.[doi:10.5392/IJoC.2013.9.4.078]
CLAIR,R.,Moncharce,N.&Slimane,M.(2008).Interactionsbetweenanartificialcolonyofmusicalantsandanimpairedhumancomposer:towardsaccessiblegenerativearts.Proceedingsofthe11thGenerativeArtConferenceMilan,Italy.
DAHLSTEDT,P.&McBurney,P.(2006).Musicalagents:TowardComputer-AidedMusicCompositionUsingAutonomousSoftwareAgents.Leonardo39,469–470[doi:10.1162/leon.2006.39.5.469]
DIXON,S.(2010).AutomaticExtractionofTempoandBeatFromExpressivePerformances.JournalofNewMusicResearch30(1),39–58.[doi:10.1076/jnmr.30.1.39.7119]
EIGENFELDT,A.(2009).TheEvolutionofEvolutionarySoftware:IntelligentRhythmGenerationinKineticEngine.ApplicationsofEvolutionaryComputing.Springer,Heidelberg,498-507.[doi:10.1007/978-3-642-01129-0_56]
FONSEKA,J.(2000).MusicalAgents.Thesis/Dissertation,MonashUniversity
FREITAS,A.,Guimarães,F.,andBarbosa,R.(2012).IdeasinAutomaticEvaluationMethodsforMelodiesinAlgorithmicComposition.Proceedingsofthe2012SoundandMusicComputingConference,Copenhagen,Denmark.
GONG,T.,Zhang,Q.,&Wu,H.(2005).Musicevolutioninacomplexsystemofinteractingagents.Proceedingsofthe2005IEEECongressonEvolutionaryComputation,Edinburgh,UK.[doi:10.1109/CEC.2005.1554815]
JUSLIN,P.&Laukka,P.(2004).Expression,perception,andinductionofmusicalemotion:areviewandaquestionnairestudyofeverydaylistening.JournalofNewMusicResearch33,216–237.[doi:10.1080/0929821042000317813]
KIRKEA.&Miranda,E.R.(2007).EvaluatingMappingsforCellularAutomataMusic,InProceedingsofECAL2007WorkshoponMusicandArtificialLife,Lisbon,Portugal.
KIRKE,A.&Miranda,E.R.(2009).ASurveyofComputerSystemsforExpressiveMusicPerformance.ACMComputingSurveys.42(1).[doi:10.1145/1592451.1592454]
KIRKE,A.&Miranda,E.(2011).ABiophysicallyConstrainedMulti-agentSystemsApproachtoAlgorithmicCompositionwithExpressivePerformance.InCA-lifeforMusic,Wisconsin:A-REditions,Inc.,165–195.
KRUMHANSL,C.&Kessler,E.(1982).Tracingthedynamicchangesinperceivedtonalorganizationinaspatialrepresentationofmusicalkeys.PsychologicalReview,89,334–368.[doi:10.1037/0033-295X.89.4.334]
LANG,P.(1995).Theemotionprobe:Studiesofmotivationandattention.TheAmericanPsychologist,50,372–385.[doi:10.1037/0003-066X.50.5.372]
LEGASPI,R.,Hashimoto,Y.,Moriyama,K.,Kurihara,S.&Numao,M.(2007).MusicCompositionalIntelligencewithanAffectiveFlavor.ProceedingsofICIUII,USA.[doi:10.1145/1216295.1216335]
LIVINGSTONE,S.R.,Muhlberger,R.,Brown,A.R.&Thompson,W.(2010)."ChangingMusicalEmotion:AComputationalRuleSystemforModifyingScoreandPerformance."ComputerMusicJournal,13(41).[doi:10.1162/comj.2010.34.1.41]
MARTINS,J.M.&Miranda,E.R.(2007).EmergentRhythmicPhrasesinanA-LifeEnvironment.ProceedingsofECAL2007WorkshoponMusicandArtificialLife(MusicAL2007),Lisbon,Portugal.
MIRANDA,E.R.(2002)EmergentSoundRepertoiresinVirtualSocieties.ComputerMusicJournal,26(2),77–90.[doi:10.1162/014892602760137194]
MURRAY-RUST,A.&Smaill,A.(2005).MusicalActsandMusicalAgents.Proceedingsofthe5thOpenWorkshopofMUSICNETWORK.Vienna,Austria.
MURRAY-RUST,A.,Smaill,A.&Maya,M.C.(2005).VirtuaLatin-TowardsaMusicalMulti-AgentSystem.ProceedingsofSixthInternationalConferenceonComputationalIntelligenceandMultimediaApplications.LasVegas,NV,USA.
http://jasss.soc.surrey.ac.uk/18/2/16.html 15 20/10/2015
OEHME,A.,Herbon,A.,Kupschick,S.&Zentsch,E.(2007).Physiologicalcorrelatesofemotions,inProceedingsofthe2007ConferenceonArtificialIntelligenceandSimulationofBehaviour.Newcastle,UK.
PANZARASA,P.&Jennings,N.(2006).Collectivecognitionandemergenceinmulti-agentsystems.InCognitionandMulti-AgentInteraction,Cambridge:CambridgeUniversityPress,401–408.
RAMCHURN,S.,Huyunh,D.,&Jennings,N.(2004).Trustinmulti-agentsystems.TheKnowledgeEngineeringReview,19,1–25.[doi:10.1017/S0269888904000116]
REYNOLDS,C.(1987).Flocks,herds,andschools:Adistributedbehavioralmodel.ComputerGraphics21(4),25–34.[doi:10.1145/37402.37406]
SAMPAIO,P.A.,Ramalho,G.,&Tedesco,P.(2008).CinBalada:amultiagentrhythmfactory.JournaloftheBrazilianComputerSociety,14,31–49.[doi:10.1007/BF03192563]
SPICER,M.,Tan,B.,&Tan,C.(2003).ALearningAgentBasedInteractivePerformanceSystem.Proceedingsofthe2003InternationalComputerMusicConference.SanFrancisco,USA.
UEDA,L.K,&Kon,E.(2003).Andante:AMobileMusicalAgentsInfrastructure.Proceedingsofthe9thBrazilianSymposiumonComputerMusic.Campinas,Brazil.
WULFHOST,D.,Flores,L.V.,Nakayama,L.,Flores,C.D.,Alvares,O.C.,&Vicari,R.M.(2003).AnOpenArchitectureforaMusicalMulti-AgentSystem.ProceedingsofthesecondinternationaljointconferenceonAutonomousagentsandmultiagentsystems.Melbourne,Australia.[doi:10.1145/860575.860669]
ZENTNER,M.,Grandjean,D.&Scherer,K.(2008).EmotionsEvokedbytheSoundofMusic:Characterization,Classification,andMeasurement.Emotion,8(4),494–521.[doi:10.1037/1528-3542.8.4.494]
http://jasss.soc.surrey.ac.uk/18/2/16.html 16 20/10/2015
