TheDynamicsofCollaborativeDesign:InsightsfromComplex
SystemsandNegotiationResearch
MarkKlein,1,*HirokiSayama,2PeymanFaratin3andYaneerBar-Yam41SloanSchoolofManagement,MassachusettsInstituteofTechnology,CambridgeMA02139,USA
2DepartmentofHumanCommunication,UniversityofElectro-Communications,1-5-1Chofugaoka,Chofu,
Tokyo182-8585,Japan
3LaboratoryforComputerScience,MassachusettsInstituteofTechnology,CambridgeMA02139,USA
4NewEnglandComplexSystemsInstitute,24Mt.AuburnStreet,CambridgeMA02138,USA
Abstract:Almostallcomplexartifactsnowadays,includingphysicalartifactssuchasairplanes,aswellasinformationalartifactssuchassoftware,organizations,businessprocesses,plans,andschedules,aredefinedviatheinteractionofmany,sometimesthousandsofparticipants,workingondifferentelementsofthedesign.Thiscollaborativedesignprocessistypicallyexpensiveandtime-consumingbecausestronginterdependenciesbetweendesigndecisionsmakeitdifficulttoconvergeonasingledesignthatsatisfiesthesedependenciesandisacceptabletoallparticipants.Recentresearchfromthecomplexsystemsandnegotiationliteratureshasmuchtooffertotheunderstandingofthedynamicsofthisprocess.Thispaperreviewssomeoftheseinsightsandofferssuggestionsforimprovingcollaborativedesign.KeyWords:collaborativedesign,dynamics,complexsystems,non-linear,negotiation
1.TheChallenge:CollaborativeDesignDynamics
Almostallcomplexartifactsnowadays,includingphysicalartifactssuchasairplanes,aswellasinforma-tionalartifactssuchassoftware,organizations,businessprocesses,plans,andschedules,aredefinedviatheinteractionofmany,sometimesthousandsofpartici-pants,workingondifferentelementsofthedesign.Thiscollaborativedesignprocessischallengingbecausestronginterdependenciesbetweendesigndecisionsmakeitdifficulttoconvergeonasingledesignthatsatisfiesthesedependenciesandisacceptabletoallparticipants.Currentcollaborativedesignapproachesareasaresulttypicallycharacterizedbyheavyrelianceonexpensiveandtime-consumingprocesses,poorincorporationofsomeimportantdesignconcerns(typicallylaterlife-cycleissuessuchasenvironmentalimpact),aswellasreducedcreativityduetothetendencytoincrementallymodifyknownsuccessful
designsratherthanexploreradicallydifferentandpotentiallysuperiorones.
Researchonnegotiationfocusesonunderstandingwhatlocalbehaviorsaretobeexpectedfrom(relativelysmallnumbersof)self-interestedagentsattemptingtocometoagreementsinthefaceofinterdependencies.Complexsystemsresearchcomplimentsthisperspectivebyattemptingtounderstandtheglobaldynamicsthatemergeasthecollectiveeffectofmanysuchlocaldecisions.Thesetwoperspectives,whenbroughttogether,havewebelievemuchtooffertoanunder-standingofthedynamicsofcollaborativedesign.Theremainderofthispaperisdedicatedtoexploringsomeoftheseinsights.
2.
AModelofCollaborativeDesign
*Authortowhomcorrespondenceshouldbeaddressed.E-mail:m_klein@mit.edu
Letusfirstestablishaworkingdefinitionofcollaborativedesign.Adesign(ofphysicalartifactssuchascarsandplanesaswellasbehavioralonessuchasplans,schedules,productionprocesses,orsoftware)canberepresentedasasetofissues(sometimesalsoknownasparameters)eachwithauniquevalue.Acompletedesignforanartifactincludesissuesthatcapturetherequirementsfortheartifact,thespecification
201
Volume11Number3September2003
1063-293X/03/030201–9$10.00/0DOI:10.1177/106329303038029
ß2003SagePublications
202M.KLEINETAL.
oftheartifactitself(e.g.thegeometryandmaterials),theprocessforcreatingtheartifact(e.g.themanufactur-ingprocess)andsoonthroughtheartifacts’entirelifecycle.Ifweimaginethatthepossiblevaluesforeveryissueareeachlaidalongtheirownorthogonalaxis,thentheresultingmultidimensionalspacecanbecalledthedesignspace,whereineverypointrepresentsadistinct(thoughnotnecessarilygoodorevenphysicallypossible)design.Thechoicesforeachdesignissuearetypicallyhighlyinterdependent.Typicalsourcesofinter-depen-dencyincludesharedresource(e.g.weight,cost)limits,geometricfit,spatialseparationrequirements,I/Ointerfaceconventions,timingconstraintsetc.Inalargeartifactlikeacommercialjettheremaybemillionsofcomponentsanddesignissues,hundredstothousandsofparticipants,workingonhundredsofdistinctdesignsub-spaces,allcollaboratingtoproduceacompletedesign.Somedesignsarebetterthanothers.Wecaninprincipleassignautilityvaluetoeachdesignandtherebydefineautilityfunctionthatrepresentstheutilityforeverypointinthedesignspace(thoughinpracticewemayonlybeabletoassesscomparativeasopposedtoabsoluteutilityvalues.Thegoalofthedesignprocesscanthusbeviewedastryingtofindthedesignwiththeoptimal(maximal)utilityvalue,thoughoftenoptimalityisabandonedinfavorof‘goodenough’.Thekeychallengeraisedbythecollaborativedesignofcomplexartifactsisthatthedesignspacesaretypicallyhuge,andconcurrentsearchbythemanyparticipantsthroughthedifferentdesignsubspacescanbeexpensiveandtime-consumingbecausedesignissueinterdependenciesleadtoconflicts(whenthedesignsolutionsfordifferentsubspacesarenotconsistentwitheachother).Suchconflictsseverelyimpactdesignutilityandleadtotheneedforexpensiveandtime-consumingdesignrework.Improvingtheefficiency,quality,andcreativityofthecollaborativeinnovativedesignprocessrequires,webelieve,amuchbetterunderstandingofthedynamicsofsuchprocessesandhowtheycanbemanaged.Inthenextsectionwewillreviewsomekeyinsightsthatnegotiationandcomplexsystemsresearchoffersforthispurpose.
3.
InsightsfromComplexSystemsand
NegotiationResearch
Acentralfocusofcomplexsystemsresearchisthedynamicsofdistributednetworks,i.e.networksinwhichthereisnocentralizedcontroller,soglobalbehavioremergessolelyasaresultofconcurrentlocalactions.Suchnetworksaretypicallymodeledasmultiplenodes,eachnoderepresentingastatevariablewithagivenvalue.Eachnodeinanetworktriestoselectthevaluethatmaximizesitsconsistencywiththeinfluencesfrom
theothernodes.Thedynamicsofsuchnetworksemergeasfollows:sinceallnodesupdatetheirlocalstatebasedontheircurrentcontext(attimeT),thechoicestheymakemaynolongerbethebestonesinthenewcontextofnodestates(attimeTþ1),leadingtotheneedforfurtherchanges.
Thenegotiationliteratureaddsthefollowingrefine-menttothismodel.Eachoneofthenodesisself-interested,i.e.attemptstomaximizeitsownlocalutility,atthesametimeitisseekingasatisfactorylevelofconsistencywiththenodesitisinterdependentwith.Acentralconcernofnegotiationresearchisdesigningtherulesofencounterbetweeninterdependentnodessuchthateachnodeisindividuallyincentedtomakedecisionsthatmaximizesocialwelfare,i.e.theglobalutilityofthecollectedsetoflocaldecisions.Inthiscase,wecandefineglobalutilitysimplyasthesumofnodeutilitiesplusthedegreetowhichtheinternodeinfluencesaresatisfied.
Isthisausefulmodelforunderstandingthedynamicsofcollaborativedesign?Webelievethatitis.Itisstraightforwardtomapthemodelofcollaborativedesignpresentedaboveontoanetwork.Wecanmapdesignparticipantsontonodes,whereeachparticipanttriestomaximizetheutilityofthesubsystemitisresponsiblefor,whileensuringitsdecisionssatisfyitsdependencies(representedasthelinksbetweennodes)withothersubsystems.Asafirstapproximation,itisreasonabletomodeltheutilityofadesignasthelocalutilityachievedbyeachparticipantplusameasureofhowwellallthedecisionsfittogether.Eventhoughreal-worldcollaborativedesignclearlyhastop-downele-mentsearlyintheprocess,thesheercomplexityofmanydesignartifactsmeansthateventuallynoonepersoniscapableofkeepingthewholedesigninhis/herheadandassessing/refiningitsglobalutility.Centralizedcontrolofthedesigndecisionsbecomesimpractical,sothedesignprocessisdominatedperforcebyconcurrentsubsystemdesignactivities(performedwithinthenodes)doneinparallelwithsubsystemdesignconsistencychecks(assessedbyseeingtowhatextentinternodeinfluencesaresatisfied).Wewillassume,forthepurposesofthispaper,thatindividualdesignersarereasonablyeffectiveatoptimizingtheirindividualsubsystems.
Thekeyfactordeterminingnetworkdynamicsisthenatureoftheinfluencesbetweennodes.Therearetwoimportantdistinctions:whethertheinfluencesarelinearornot,andwhethertheyaresymmetricornot.Wewillconsidereachoneofthesedistinctionsinturn,withanimportantsidetripintothenegotiationliteraturetounderstandthedilemmasraisedbythepresenceofself-interestedagents.Thiswillbefollowedbyadiscussionofsubdividednetworktopologies,andtheroleoflearning.Unlessindicatedotherwise,thematerialoncomplexsystemspresentedbelowisdrawnfrom[1].
TheDynamicsofCollaborativeDesign203
A.LinearversusNonlinearNetworks
NonlinearityProducesMultioptimumUtilityFunctions:Ifthevalueofnodesisalinearfunctionoftheinfluencesfromthenodeslinkedtoit,thenthesystemislinear,otherwiseitisnonlinear.Linearnetworkshaveasingleattractor,i.e.asingleconfigura-tionofnodestatesthatthenetworkconvergestowardsnomatterwhatthestartingpoint,correspondingtotheglobaloptimum.Theirutilityfunctionthushasasinglepeak.Thismeanswecanusea‘hill-climbing’approach(whereeachnodealwaysmovesdirectlytowardsincreasedlocalutility)becauselocalutilityincreasesalwaysmovethenetworktowardstheglobaloptimum.Nonlinearnetworks,bycontrast,arecharacterizedbyhaving‘bumpy’utilityfunctionswithmultiplepeaks(i.e.localoptima)andthusmultipleattractors.Akeypropertyofnonlinearnetworksisthatsearchfortheglobaloptimacannotbeperformedsuccessfullybypurehill-climbingalgorithms,becausetheycangetstuckinlocaloptimathataregloballysuboptimal.
Arangeoftechniqueshaveemergedthatareappropriateforfindingglobaloptimainmulti-optimautilityfunctions,allrelyingontheabilitytosearchpastvalleysintheutilityfunction.Stochasticapproachessuchassimulatedannealinghaveprovenquiteeffective[2].Simulatedannealingendowsthesearchprocedurewithatoleranceformovinginthedirectionoflowerutilitythatvariesasafunctionofavirtual‘tempera-ture’.Atfirstthetemperatureishigh,sothesystemisasapttomovetowardslowerutilitiesashigherones.Thisallowsittorangewidelyovertheutilityfunctionandpossiblyfindnewhigherpeaks.Sincehigherpeaksgenerallytendalsotobewiderones,thesystemwillspendmostofitstimeintheregionofhighpeaks.Overtimethetemperaturedecreases,sothealgorithmincreasinglytendstowardspurehill-climbing.Whilethistechniqueisnotprovablyoptimal,ithasbeenshowntogetclosetooptimalresultsinmostcases.ASocialDilemmawithSelf-interestedAgents:Annealingrunsintoadilemma,however,whenappliedtosystemswithself-interestedagents.Letusassumethatatleastsomeactorsare‘hill-climbers’,concernedonlywithmaximizingtheirlocalutilities,whileothersare‘annealers’,willingtoaccept,atleasttemporarily,lowerlocalutilitiesaspartoftheexploratoryprocess.Wecanuseasimulationapproachtoexplorewhathappens.Table1summarizestheresultsforsuchexperiments,
Table1.Annealingversushill-climbingagents.
Agent2Hill-climbs
Agent2Anneals
Agent1hill-climbs[0.86][0.86]0.73/0.740.99/0.51Agent1anneals
[0.86][0.98]0.51/0.99
0.84/.84
givingthelocalandglobalutilitiesachievedfordifferentpairingsofagentstrategiesinsimulatednonlinearnegotiations:
Inthetable,thecellvaluesarelaidoutasfollows:[ Theseresultsshowthat,whileannealersincreaseglobalutility,andarethereforehighlydesirable,annealersalwaysfareindividuallyworsethanhill-climberswhenbotharepresent.Hill-climbingisthusa‘dominant’strategy:nomatterwhatstrategytheotheragentuses,itisbettertobeahill-climber.Ifallagentsdothis,however,thentheyforegothehigherindividualutilitiestheywouldgetiftheybothareannealed.Individualstrategicconsiderationsthusdrivethesystemtowardsthestrategypairingwiththelowestutilityvalues. WhatcanbedoneaboutthisThispatternofutilityvaluesisaninstanceofawell-knownphenomenoningametheoryknownasthe‘‘prisoner’sdilemma’’[5].Ithasbeenshownthatthisdilemmacanbeavoidediftherearerepeatedinteractionsbetweenagents[6].Theideaissimple.Eachagentusesanannealingstrategyatfirst,butifitdeterminesthattheagentitisnegotiatingwithisusinghill-climbing,ititselfthenswitchestohill-climbingforitsfuturenegotiationswiththatagent,therebyforcingthembothintothe‘lose–lose’quadrantofTable1.Itturnsoutthatthis‘titfortat’approachincentsannealingbehaviorinallagents,assumingthattheynegotiatewitheachothermultipletimes.Thisideacanberefinedwiththeadditionofa‘reputationmechanism’,whereinagentsconsultadatabaseofpreviousnegotiations(inadditiontotheirindividualexperience)inordertodeterminewhethertheagenttheycurrentlyfacetendstobeanannealerorhill-climber.Ideally,however,wewouldprefertofindawaytoincentannealingbehaviorwithinthecontextofasinglenegotiation,withouttherequirementofmultipleinter-actions.Canthisbedone? Someapparentlyreasonableapproachesare,itturnsout,quiteineffective.Oneapproach,forexample,iswhatwecancall‘adaptive’annealing.Anegotiationtypicallyconsistsofarelativelylargenumberofoffersandcounter-offers,resultinginincreasinglybetterinterimagreementsthateventuallyareacceptedasfinalbybothparties.Anagentcouldthereforeinprincipleswitchinmidstreamfrombeinganannealertobeingahill-climberifitdeterminesthattheotheragentisbeingahill-climber.Determiningthestrategytypeoftheagentyouarenegotiatingwithisinfactrelativelyeasy:anannealertendstoacceptamuchhigherpercentageofinterimproposalsthanahill-climber.Theproblemwiththisapproachisthatdeterminingthetypeofanagentinthiswaytakestime.Oursimulationshaveshownthatthe 204M.KLEINETAL. divergenceinacceptanceratesbetweenannealersandhill-climbersonlybecomesclearaftermostoftheutilityhasbeencommitted,soitistoolatetofullyrecoverfromtheconsequencesofhavingstartedasanannealerifyounegotiatedwithahill-climber.Hill-climbingthereforeremainsthedominantstrategy.Anotherpossibilityisforannealerstosimplybelessconcessionary,i.e.lesswillingtoacceptutility-decreasinginterimagreements.Thisinfactallowsustoeliminatethepoorannealerpayoffsthatunderlietheprisoner’sdilemma,butonlyatthecostofradicallyreducedglobalutility.Inbothcases,weareunabletoincentagentstrategiesthatoptimizetheglobalutilityoftheoutcome. Resolvingtheprisoners’dilemmawithinthescopeofasinglenegotiationcanbeachieved,however,throughtheuseofwhatwecalla‘parity-enforcingannealingmediator’.Ratherthanrequiringthattheagentsanneal,wemovetheannealingintoathirdpartywecallamediator.Inthisapproach,possibleagree-mentsaregenerated(inourexperimentstheyweregeneratedbythemediator,butthisisanotacriticalpartofthescheme)andthenvotedonbythenegotiatingagents.Themediatorisakindofannealer:itisendowedwithatime-decreasingwillingnesstoatleasttemporarilyfollowupondesignproposalsthatoneorbothagentsvotedagainst.Agentsarefreetoremainhill-climbersintheirvotingbehavior,andthusavoidmakingharmfulconcessions.Themediator,byvirtueofbeingwillingtoprovisionallypursueutility-decreasingagreements,cantraversevalleysintheagents’utilityfunctionsandtherebyleadtheagentstowin–winsolutions.Paradoxically,usingamediatorthatoccasionallyignoresagentpreferencesleadstooutcomesthatarebetterforbothagents. Achievingmaximalglobalutilitiesinthisschemerequiresthatagentsbeabletoannotatetheirvoteswithstrengthinformation.Abinaryschemeissufficient,whereinagentsannotatetheiracceptvotesasbeingeitherstrongorweak.Thisallowsthepossibilityof‘over-rides’,whereinthemediatorpursuesaninterimagreementthatwasstronglypreferredbyoneagentandweaklyrejectedbyanother.Over-ridesareimportantbecausesuchagreementsarelikelytoincreaseglobalutility.Agentsmightofcoursebetemptedtoexaggerateinsuchcontexts,markingeveryvoteasbeingastrongone.Butthispossibilitycanbefoiledbyenforcingrunningparityonthenumberoftimeseachagentover-ridestheother.Thisworksforthefollowingreason.Onecanthinkofthisprocedureasgivingagents‘tokens’thattheycanusetogainover-rides.Atruthfulagentspendsitstokensexclusivelyonover-ridesthattrulyofferitastronglocalutilityincrease.Anexaggerator,ontheotherhand,willspendtokensevenwhentheutilityincrementitderivesisrelativelysmall.Attheendoftheday,thetruthfulagenthasspenditstokensmorewiselyandtobettereffect.Lessons:Howdotheseinsightsapplytocollaborativedesign?Generallyspeaking,linearnetworksrepresentaspecialcase(onlyatinyfractionofallpossibleinfluencerelationshipsarelinear),buttheyhaveprovenadequateformodelingwhathasbeencalledroutinedesign.Routinedesigninvolveshighlyfamiliarrequirementsanddesignoptions,asforexampleinautomobilebrakeortransmissiondesign[7].Inthesecontexts,designerscanusuallystartthedesignprocessnearenoughtothefinaloptimumthattheprocessactsasifithasasingleattractor.Previousresearchondesigndynamicshasfocusedonthisclassofdesignmodel,generatingsuchusefulresultsasapproachesforidentifyingdesignprocessbottlenecks[8]andforfine-tuningtheleadtimesfordesignsubtasks[9]. Rapidtechnologicalandotherchangeshavemadeitincreasinglyclear,however,thatmanyofthemostimportantcollaborativedesignproblems(e.g.concerningsoftware,biotechnology,orelectroniccommerce)involveinnovativedesign,radicallynewrequirements,andunfamiliardesignspaces.Itisoftenunclearhowtoachieveagivensetofrequirements.Theremaybemultipleverydifferentgoodsolutions,andthebestsolutionmayberadicallydifferentthananythathavebeentriedbefore.Forsuchcasesnonlinearnetworksseemtorepresentamoreaccuratemodelofthecollaborativedesignprocess. Thishasimportantconsequences.Oneisatendencytostaywithwell-knowndesigns.Whenautilityfunctionhaswidelyseparatedoptima,onceasatisfactoryoptimumisfoundthetemptationistosticktoit.Thisdesignconservatismisexacerbatedbythefactthatitisoftendifficulttocomparetheutilitiesforradicallydifferentdesigns.Wecanexpectthiseffecttobeespeciallyprevalentinindustries,suchascommercialairlinesandpowerplants,whicharecapital-intensiveandrisk-averse,sinceinsuchcontextsthecostofexploringnewdesigns,andtheimpactofgettingitwrong,canbeprohibitive.Anotherconsequenceisthatcollaborativedesignascurrentlypracticedisprobablyquitepronetogettingstuckinlocaloptimathatmaybesignificantlyworsethanradicallydifferentalternatives.Annealing-likeprocessespotentiallyapplicabletoaddressingthisproblemarewidelyusedinhumancollaborativedesignsettings.‘Brainstorming’,forexample,withitsemphasisonnotpruningcandidatesolutionstooquickly,canbeviewedasakindofannealing.Designersare,however,generallymuchmorestronglyencouragedtocreateagooddesignfortheirownsubsystems,thantoconcedetomakesomeoneelse’sjobeasier.Thisincentivestructureleadstothe‘‘prisoner’sdilemma’’describedabove. Theprisoner’sdilemmacan,aswehaveseen,beavoidedifweassumethatagentshavemultiplenegotiationencountersandusea‘titfortat’schemefordecidingwhentobeconcessionaryornot.Suchschemesareprobablyused,infact,bymanydesignersin TheDynamicsofCollaborativeDesign205 collaborativesettings.Therelativeinfrequencyofmajornegotiations,theabsenceofreputationdatabases,andhighturnoverinpersonnelmay,however,sabotagetheefficacyofsuchstrategies.Itseemslikely,inaddition,thatmanyengineersmakesomeuseoftheotherapproacheswedescribedabove,beingadaptiveorsimplyhighlysparinginhowmuchtheyconcede.Theseare,afterall,apparentlyreasonablestrategies.Theydonot,however,havethedesiredresultoffosteringthediscoveryofmoreoptimaloveralldesigns.Mediation,aswehaveseen,hasthepotentialofresolvingtheprisoner’sdilemma,anditinfacthasanimportantplaceincurrentcollaborativedesignpractice.Seniorengineers,andinsomecasesteamsofsuchengineers(sometimescalled‘‘changeboards’’)areoftencalledupontomediatesituationswheretheachievementofsatisfactoryglobalutilityappearstobethreatened.Engineerswiththatlevelofexperienceare,however,ascarceresource,sothistacticistypicallyreservedforonlythemostseriousproblems. Inbrief,itappearslikelythatcurrentcollaborativedesignpractice,particularlyforhighlyinnovativedesign,ispronetogettingstuckinunnecessarilysuboptimalsolutions.Wewilldiscusspossiblesolutionstotheseproblemsinthesection‘‘HowWeCanHelp’’below. B.SymmetricversusAsymmetricNetworks AsymmetryAllowsNonconvergence:Symmetricnet-worksareonesinwhichinfluencesbetweennodesaremutual(i.e.ifnodeAinfluencesnodeBbyamountXthenthereverseisalsotrue),whileasymmetricnetworksdonothavethisproperty.Asymmetricnetworks(iftheyhavecyclesinthem;seebelow)addthecomplicationofhavingdynamicattractors,whichmeansthatthenetworkdoesnotconvergeonasingleconfigurationofnodestatesbutrathercyclesindefinitelyaroundarelativelysmallsetofconfigurations.Letusconsiderthesimplestpossiblecyclicasymmetricnetwork:the‘oddloop’(Figure1): Thisnetworkhastwolinks:onewherenodeBinfluencesnodeAtohavethesamevalue,andanotherwherenodeAinfluencesnodeBtohavetheoppositevalue.Imaginebothnodeshavetheinitialvalue1,and -1AB+1Figure1.Thesimplestpossiblecyclicasymmetricnetwork–an‘oddloop’. updateeachotherinparallel.Thestatesofthetwonodeswillproceedasfollows: StateValueofNodeA ValueofNodeB Initialstate11State11À1State2À1À1State3À11State 4 1 1 Afteronetimestep(State1)nodeAwillcausenodeBto‘flip’toÀ1,andnodeBwillleavenodeAunchanged.Afteraseconditeration(State2)nodeAleavesnodeBunchanged,butnodeBcausesthevalueofnodeAtoflip.Ifwetracethisfarenoughwefindthatthesystemreturnstoitsinitialstate(State4)andthuswillrepeatadinfinitum.Morecomplicatedasymmetricnetworkswillproducedynamicattractorswithmorecomplicatedshapes,includingoneswherestatesareneverexactlyrepeated,buttheupshotisthesame:thesystemwillnotconverge.Onecanalwaysofcoursestopthesystematsomearbitrarypointalongitstrajectory,butthereisnoguaranteethatthedesignutilityatthatpointwillbebetterthanthatatanyotherpointbecausethesystem,unlikethesymmetriccase,doesnotnecessarilyprogressmonotonicallytowardshigherutilityvalues.Thiscanbeunderstoodinthefollowingway.Everyutilityfunctioncan,inprinciple,be‘compiled’intoa(symmetric)networkthatwillprogressmonotonicallytowardshigherutilityvaluesaslongastheindividualnodesperformlocaloptimization.Theopposite,however,isnottrue.Therearemanynetworks(includingmostasymmetricones)thatdonotcorrespondtoanywell-formedutilityfunction,sotheirsequencesofstatesclearlycannotbeviewedasprogressingtowardsautilityoptimum[1]. Ifanetworkisacyclichowever(alsoknownasafeed-forwardnetwork,whereinanodeisneverabletodirectlyorindirectlyinfluenceitsownvalue),ithasawell-definedutilityfunctionandthuswillnothaveadynamicattractor. Lessons:Howdoesthisapplyincollaborativedesignsettings?Traditionalserializedcollaborativedesignisanexampleofanasymmetricfeed-forwardnetwork,sincetheinfluencesallflowuni-directionallyfromtheearlierproductlifecyclestages(e.g.design)tolaterones(e.g.manufacturing)withonlyweakfeedbackloopsifany.Insuchcontextstheattractorsshouldbestaticandconvergenceshouldalwaysoccur,givensufficienttime.Insuchsettingswemaynot,however,expectparticularlyoptimaldesigns.Itistypicallyverydifficult,giventheboundedrationalityofhumanbeings,fordesignersearlierinthedesignlifecycletoensurethatthedesignerslateroninthelifecyclewillbeabletoproduce 206M.KLEINETAL. near-optimalsolutionsfortheirverydifferentbuthighlydependentproblems.Thisisinfacttherationalunder-lyingtheadoptionofconcurrentengineeringapproaches.‘Pure’concurrentengineering,wherealldesigndisciplinesarerepresentedonmultifunctionaldesignteams,encourageroughlysymmetricinfluencesbetweentheparticipantsandthuscanalsobeexpectedtohaveconvergentdynamicswithstaticattractors.Currentcollaborativedesignpractice,however,isahybridofthesetwoapproaches,andthusislikelytohavethecombinationofasymmetricinfluencesandinfluenceloopsthatproducesdynamicattractorsandthereforenonconvergentdynamics. This,moreover,isafundamentalproblem.Asnotedabove,itisinprinciplestraightforwardtocomputetheproperinternodeinfluencesgivenaglobalutilityfunction.Indesignpractice,however,wedonotknowtheglobalutilityfunction,especiallyoncewehavereachedtherealmofdetaileddesign.Thespaceofpossibledesigns,andthecostofcalculatingtheirindividualutilityvalues,issimplytoolarge.Atbesttheglobalutilityfunctionisrevealedtousincrementallyaswegenerateandcomparedifferentcandidatedesigns.Theinflu-encerelationshipsbetweendesignersare,asaresult,invariablydefineddirectlybasedonexperienceandourknowledgeofdesigndecisiondependencies.Butsuchaheuristicapproachcaneasilyleadtothecreationofinfluencenetworksthatdonotinstantiateawell-formedutilityfunction,andthusdisplaydynamicattractors. Dynamicattractorswerefoundnottohaveasignificanteffectonthedynamicsofatleastsomeroutine(linear)collaborativedesigncontexts[9],butmayprovemoresignificantininnovative(nonlinear)collaborativedesign.Itmayhelpexplain,forexample,whyitsometimestakessomanyiterationstofullypropagatechangesincomplexdesigns[10].C. SubdividedNetworks SubdivisionCanSpeedConvergence:Anotherimpor-tantpropertyofnetworksiswhetherornottheyaresubdivided,i.e.whethertheyconsistofsparselyinter-connected‘clumps’ofhighlyinterconnectednodes.Whenanetworkissubdivided,nodestatechangescanoccurwithinagivenclumpwithonlyminoreffectsontheotherclumps.Thishastheeffectofallowingthenetworktoexploremorestatesmorerapidly.Ratherthanhavingtowaitforanentirelargenetworktoconverge,wecanrelyinsteadonthemuchquickerconvergenceofanumberofsmallernetworks,eachoneexploringpossibilitiesthatcanbeplacedindifferingcombinationswiththepossibilitiesexploredbytheothersubnetworks[11]. Lessons:Thiseffectisinfactwidelyexploitedindesigncommunities,whereitisoftenknownas modularization.Thisinvolvesintentionallycreatingsubdividednetworksbydividingthedesignintosubsystemswithpredefinedstandardizedinterfaces,sosubsystemchangescanbemadewithfeworanyconsequencesforthedesignoftheothersubsystems.Thekeytousingthisapproachsuccessfullyisdefiningthedesigndecompositionsuchthattheutilityimpactofthesubsysteminterdependenciesontheglobalutilityisrelativelylow,becausestandardizedinterfacesrarelyrepresentanoptimalwayofsatisfyingthesedependen-cies.Inmostcommercialairplanes,forexample,theengineandwingsubsystemsaredesignedseparately,takingadvantageofstandardizedenginemountstoallowtheairplanestousearangeofdifferentengines.Thisisalmostcertainlynottheoptimalwayofrelatingenginesandwings,butitisgoodenoughandsimplifiesthedesignprocessconsiderably.Iftheengine-winginterdependencieswerecrucial,forexampleifstandardenginemountshadadrasticallynegativeeffectontheairplane’saerodynamics,thenthedesignofthesetwosubsystemswouldhavetobecoupledmuchmorecloselyinordertoproduceasatisfactorydesign. D.Imprinting ImprintingCapturesSuccessfulInfluencePatterns:Onecommontechniqueusedtospeednetworkconver-genceisimprinting,whereinthenetworkinfluencesaremodifiedwhenasuccessfulsolutionisfoundinordertofacilitatequicklyfinding(similar)goodsolutionsnexttime.Acommonimprintingtechniqueisreinforcementlearning,whereinthelinksrepresentinginfluencesthataresatisfiedinasuccessfulfinalconfigurationofthenetworkarestrengthened,andthoserepresentingviolatedinfluencesweakened.Theeffectofthisistocreatefewerbuthigheroptimaintheutilityfunction,therebyincreasingthelikelihoodofhittingsuchoptimanexttime. Lessons:Imprintingisacrucialpartofcollaborativedesign.Theconfigurationofinfluencesbetweendesignparticipantsrepresentsakindof‘social’knowledgethatisgenerallymaintainedinanimplicitanddistributedwaywithindesignorganizations,intheformofindividualdesigner’sheuristicsaboutwhoshouldtalktowhomwhenaboutwhat.Whenthisknowledgeislost,forexampleduetohighpersonnelturnoverinanengineeringorganization,theabilityofthatorganiza-tiontodocomplexdesignprojectsiscompromised.Itshouldbenoted,however,thatimprintingreinforcesthetendencywehavealreadynotedfororganizationsinnonlineardesignregimestosticktotried-and-truedesigns,byvirtueofmakingthepreviously-foundoptimamoreprominentinthedesignutilityfunction,andthusmaybecounter-indicatedforchallengesrequiringhighlyinnovativedesigns. TheDynamicsofCollaborativeDesign207 4.HowCanWeHelp? Whatcanwedotoimproveourabilitytodoinnovativecollaborativedesign?Wewillbrieflyconsiderseveralpossibilitiessuggestedbythediscussionabove.Informationsystemsareincreasinglybecomingthemediumbywhichdesignparticipantsinteract,andthisfactcanbeexploitedtohelpmonitortheinfluencerelationshipsbetweenthem.Onecouldtrackthevolumeofdesign-relatedexchangesor(amoredirectmeasureofactualinfluence)thefrequencywithwhichdesignchangesproposedbyoneparticipantareacceptedasisbyotherparticipants.Thiscanbehelpfulinmanyways.Highlyasymmetricinfluencescouldrepresentanearlywarningsignofnonconvergentdynamics.Detectingalowdegreeofinfluencebyanimportantdesignconcern,especiallyonesuchasenvironmentalimpactthathastraditionallybeenlessvalued,canhelpavoidutilityproblemsdowntheroad.Arecordoftheinfluencerelationshipsinpreviousfailedandsuccessfuldesignprojectscanbeusedtohelpbettermanagefutureprojects.Thiswillrequirebeingabletodeterminewhichinfluenceswerecriticalinthesepreviousefforts.Ifalatehigh-impactproblemoccurredinasubsystemthathadalowinfluenceinthedesignprocess,forexample,thiswouldsuggestthattherelevantinfluencerelationshipsshouldbemodifiedinthefuture.Incentivemechanismscanbeputinplacethatrewardengineersnotjustforproducinggoodsubsystemdesigns,butalsoforparticipatinginwhatarebelievedtobeproductivepatternsofmutualinfluencewithotherdesigners.Notethatthishastheeffectofmakingacriticalclassofnormallyimplicitanddistributedknowledgemoreexplicit,andthereforemoreamenabletobeingpreservedovertime,aswellastransferredbetweenprojectsandevenorganizations. Informationsystemscanalsopotentiallybeusedtohelpassessthedegreetowhichthedesignparticipantsareengagedinroutine(i.e.optimization-driven)versusinnovative(i.e.highlyexploratory)designstrategies.Wecouldusesuchsystemstoestimateforexamplethenumberandvarianceofdesignalternativesbeingconsideredbyagivendesignparticipant.Thisisimportantbecause,aswehaveseen,aprematurecommitmenttoaroutinedesignstrategythatoptimizesagivendesignalternativecancausethedesignprocesstomissotheralternativeswithhigherglobaloptima.Trackingthedegreeofinnovativeexplorationcanbeusedtofine-tunetheuseofinnovation-enhancinginterventionssuchasincentives,competingdesignteams,introducingnewdesignparticipants,andsoon.Aswithsimulatedannealing,itwillprobablymakesensetoencouragemoreconcedingandexplorationearlyoninthedesignprocess,andgraduallytransitiontohill-climbingastimegoeson. Theprisoner’sdilemmaincentivestructurethatleadstosuboptimaldesignscanbeaddressedinatleasttwowaysthatareprobablyunder-utilizedincurrentpractice.Oneisbytheintroductionofreputationmechanisms.Ifwesimplymakeinformationavailableonwhichdesignershaveahistoryofconcedingsparingly,wearelikelytofindanincreaseinconces-sionarybehavior,andthereforeimproveddesignout-comes,evenintheabsenceofexplicit(e.g.salary)incentives.Anotherpossibilityisthewideruseofmediators.Mediatorsincollaborativedesigncontextshavetraditionallybeenseniorengineerscapableofdictatingthecontentofadesignoutcome.Recentworkonnegotiationalgorithmssuggests,however,thatmediatorscanbeeffectivebyguidingthedesignprocess,forexampleaswesuggestedabovebyoccasionallyhavingtheagentsfollowupondesignoptionsthatoneorbothrejected,andbyenforcingroughparityinthenumberofmixedwins.Process-orientedmediationdoesnotrequirethesamedepthofdomainexpertiseascontent-orientedmediation,anditisthereforelikelythatdesignerscanbetrainedtoprovidethatforeachother,andthatsuchmediationcanbecomemuchmorewidelyavailableasaresult. Finally,informationsystemscanbeusedtotrackthehistoryofdesignalternativesexploredandtherebydetectthedesignloopsthatindicateanonconvergentdesignprocess. 5.Conclusions Existingcollaborativedesignapproacheshaveyieldedsolidbutincrementaldesignimprovements,whichhasbeenacceptablebecauseoftherelativelyslowpaceofchangeinrequirementsandtechnologies.Considerforexamplethelast30yearsofdevelopmentinBoeing’scommercialaircraft.Whilemanyimportantadvanceshavecertainlybeenmadeinsuchareasasengines,materials,andavionics,thebasicdesignconcepthaschangedrelativelylittle.Consider,forexample,thefactthattheBoeing737(inaugurated1965)andtheBoeing777(1995)followessentiallythesamedesignconcept(seehttp://www.boeing.com).Futureradicallyinnova-tivedesignchallenges,suchashigh-performancecom-mercialtransport,willprobablyrequire,however,substantialchangesindesignprocesses. Thispaperhasbeguntoidentifywhatrecentresearchonnegotiationandcomplexsystemscanofferinthisregard.Thekeyinsightsarethatimportantpropertiesofcollaborativedesigndynamicscanbeunderstoodasreflectingtwobasicfacts:(1)collaborativedesignisakindofdistributednetwork,and(2)theagentsinthisnetworkareself-interestedandrespondtolocalincentives.Thisispowerfulbecausethismeansthatourgrowinggeneralunderstandingofnetworksand 208M.KLEINETAL. negotiationcanbeappliedtohelpusbetterunderstandandeventuallybettermanagecollaborativedesignregardlessofthedomain(e.g.physicalvs.informationalartifacts)andtypeofparticipants(e.g.humanvs.software-based). Thisinsightleadstoseveralothers.Mostprominentisthesuggestionthatweneedtofullyembraceaninfluences-andincentives-centricperspectiveonhowtomanagecomplexcollaborativedesignprocesses.Itiscertainlypossiblefordesignmanagerstohaveaverydirecteffectonthecontentofdesigndecisionsduringpreliminarydesign,whenarelativelysmallnumberofhigh-levelglobalutilitydrivendecisionsaremadetop-downbyasmallnumberofplayers.Butoncethedetaileddesignofacomplexartifacthasbeendistrib-utedtomanyplayers,theglobalutilityimpactoflocaldesignchangesistoodifficulttoassess,anddesigndecisionsaretoovoluminousandcomplextobemadetop-down,sothedominantdriversbecomelocalutilitymaximizationplusfitbetweentheselocaldesigndecisions.Inthisregimeencouragingtheproperinfluencerelationshipsandconcessionstrategiesbecomestheprimarytoolavailabletodesignmanagers.Ifthesearedefinedinappropriately,wecanendupwithdesignsthattaketoolongtocreate,donotmeetimportantrequirements,and/ormissopportunitiesforsignificantutilitygainsthroughmorecreative(far-ranging)explorationofthedesignspace. Acknowledgments ThisworkwassupportedbytheNationalScienceFoundationandtheDefenseAdvancedResearchProjectsAgency. References 1.Bar-Yam,Y.(1997).DynamicsofComplexSystems.,Vol.xvi,p.848,Reading,Mass.:Addison-Wesley. 2.Kirkpatrick,S.,Gelatt,C.D.andVecchi,M.P.(1983).OptimizationbySimulatedAnnealing,Science,220:671–680. 3.Klein,M.,Faratin,P.,Sayama,H.andBar-Yam,Y.(2002).NegotiatingComplexContracts,In:AutonomousAgentsandMulti-AgentSystems,BolognaItaly:AAAIPress. 4.Klein,M.,Faratin,P.andBar-Yam,Y.(2002).UsinganAnnealingMediatortoSolvethePrisoners’DilemmaintheNegotiationofComplexContracts,In:Agent-MediatedElectronicCommerce(AMEC-IV)Workshop,BolognaItaly:Springer. 5.Osborne,M.J.andRubinstein,A.(1994).ACourseinGameTheory,Vol.xv,p.352,Cambridge,Mass.:MITPress.6.Axelrod,R.(1984).TheEvolutionofCooperation,BasicBooks,NewYork,NY. 7.Brown,D.C.(1989).MakingDesignRoutine.In:ProceedingsofIFIPTC/WGonIntelligentCAD. 8.Smith,R.P.andEppinger,S.D.(1997).IdentifyingControllingFeaturesofEngineeringDesignIteration,ManagementScience,43(3):276–293. 9.Eppinger,S.D.,Nukala,M.V.andWhitney,D.E.(1997).GeneralizedModelsofDesignIterationUsingSignalFlowGraphs,ResearchinEngineeringDesign,9(2):112–123.10.Klein,M.(1994).Computer-SupportedConflictManage-mentinConcurrentEngineering:IntroductiontoSpecialIssue,ConcurrentEngineeringResearchandApplications,2(3). 11.Simon,H.A.(1996).TheSciencesoftheArtificial, CambridgeMAUSA:MITPress. Biographies MarkKlein MarkKleinisaPrincipalResearchScientistattheMITCenterforCoordinationScience,aswellasanAffiliateattheMITAILabandtheNewEnglandComplexSystemsInstitute.Hedoesresearchondevelopinginsightsandtoolstoenablemoreeffec-tivecoordinationingroupswithhumansand/orcompu-ter-basedagents. HirokiSayama Prof.SayamaobtainedhisD.Sc.in1999attheUniversityofTokyo,Japan.From1999to2002hewasapostdoctoralfellowattheNewEnglandComplexSystemsInstitute,Cambridge,Massachusetts,andworkedonvariousissuesintheoreticaltreatmentsofcomplexsystems,includingevolutionarybiology,multia-gentsystems,artificiallife,etc. TheDynamicsofCollaborativeDesign209 PeymanFaratin Dr.PeymanFaratiniscur-rentlyaPostdoctoralAssociateattheLaboratoryforComputerScienceatMITworkingonnegotiationproto-colsandalgorithmsforMulti-AgentSystems.HeobtainedhisPhDfromUniversityofLondononcomputationalmodelsofbargainingforagentsystems.HehasbeenavisitingscientistatMIT’s SloanSchoolofManagement,SpanishArtificialIntelligenceInstitute,NortelNetworks,andBritishTelecoms. YaneerBar-Yam ProfessorBar-YamisPresi-dentoftheNewEnglandComplexSystemsInstitute,anAssociateattheDepartmentofMolecularandCellularBiologyatHarvardUniversity,ChairmanoftheInternationalConferenceonComplexSys-tems,ManagingEditorofInterJournal,andauthorofthetext‘‘DynamicsofCom-plexSystems’’.Hisresearch isfocusedonformalizingcomplexsystemsconceptsandrelatingthemtoeverydayproblems. 因篇幅问题不能全部显示，请点此查看更多更全内容