Automated Negotiationfor Provisioning Virtual Priv Networks usingFIPA-Compliant Agents P. Faratin 1N. ,R. Jennings
2
P. ,Buckle
3 and CSierra .
ate 4
1
Dept. of ElectronicEngineering, QueenMary andWes tfieldCollege, University oLondon, f LondonE14 NS, UK. Email:
[email protected] 2
Dept. of Electronics andComputer Science, Universi ty oSouthampton, f SouthamptonSO171BJ, UK. Email:
[email protected]
3
AdvancedIP Services andManagement, Nortel Network s, Harlow Labs, Harlow, UK. Email:
[email protected]
4
IIIAArtificial IntelligenceResearchInstitute, C 08193Bellaterra, Catalonia, Spain.
SIC-SpanishCouncil for ScientificResearch, Email:
[email protected]
ABSTRACT Thispaperdescribesthedesignandimplementation provisioningvirtualprivatenetworks.Theagentsa specificationandtheyareimplementedusingtheFI is focusedonthedesignandimplementationothe f
ofn egotiatingagentsforthetaskof ndtheirinteractionscomplywiththeFIPA PA-OSagentframework. Particular attention negotiationalgorithms.
1.INTRODUCTION Multi-agentsystem(MAS)researchfocuseso nhowa societyofautonomous,computational agentscaninteractino rdertosolvecomplex, real worldp roblemsthatareinherently distributed innature(Bond andGasser, 1998). Duetothed istributedandinter -relatednatureothe f problem, oneotfhekeyissuesinanyMASisthatofcoordin ation.Toachievecoordinationinagiven application, naumber of sub-problems need tboaed dressed (Weiss, 1999). Firstly, there inaseed todefinea languageanda protocol of interaction. The former specifies the syntax a nd semantics ofthecommunication.Thelatterspecifiesthenorm ative(orconstraining)rulesowho f cansay whatinthecourseothe f interaction.Secondly, th ereineed as tospecify h owagentswillreason aboutthecoordinationproblem.Thatis,thereias needtomodelhowagentswillrepresentand reasonabouttheactions,plansandknowledgeoth f emselvesandothers.Thirdly,thereineed as toidentify howcoherentsystem-widebehaviour can beattainedfrom agentsthat invariably h ave alocalandpartialviewoftheproblem.Finally,i nadditiontotheaforementionedmodelling problems,thereisalsoaneedtoactually engineerapracticalMASthroughthedesignof platforms andmethodologies (BondandGasser, 1998) . Theworkreportedhereaddressesalltheaforementi onedcoordinationsub-problems.The engineeringaspectsareaddressedbyadoptingthed omain-independentsoftwarestandard formulatedby theFoundationforIntelligentPhysicalAgents(FIP A).FIPAspecifiesthe normativestandardfor thesyntax andsemantics of thelanguageusedduring agentinteraction, as wellastheprotocolo finteraction.Agentsarethe nembeddedwithinthisenvironmentand interact,complyingwiththestandard,tosolvethe irdomainproblem.Theproblemdomainwe consideristhecompetitiveanddynamicprovisionin govirtual f privaten etworks(VPNs)by end
users. Agentsinthisscenarioreasonaboutthecoo modelthatwehaved evelopedandFIPA-compliant a n agentsinreachingagreementsandincoordinatingt computational andinformationb oundedness. Thecontributionothis f workitswofold.Firstly, OS)thatisbothpracticalandgeneric.Thisisthef FIPAstandard.Secondly,wepresentthedesignand suitablefor raangeoreal-world f applications. Th baseduponpracticalassumptionsandthatitincorp (includingconcessionmaking,makingtrade-offsbet thecourseoan fongoing negotiation).
rdinationprocess by using novel a negotiation egotiationprotocol.Thismodelaidsthe heirinteractions,whilerespectingtheir
wepresent na ewagentplatform (called irst,freelyavailableimplementationofthe utilisationofanegotiationmodelthatis ismodel advancesthestateothe f artin that it i orates wide a rangeonegotiation f techniques weenissuesandissuemodificationd uring
Theremainderotfhepaperisstructuredafsollows Section3introducestheVPNscenarioandsection4 OS.Thisscenarioitshenu sedtomotivateanddeta section6presents theconclusionsandfuturework
FIPA-
s
Section . 2presentstheFIPA-OSplatform. describesitsimplementationusingFIPAilthen egotiationmodelinsection5Finally, .
2.FIPA-OS ThepurposeofFIPAistopromotethedevelopmento technologies that maximiseinteroperability within functionitsoproduce specification a for anagent freetoproducetheir ownimplementationsothis f s andoperationcomplies with the published FIPA spec frameworks are interoperable. Within this context, themandatoryelementscontainedwithintheFIPAsp FIPA-OSdistributioncontainsclassfiles,Javasou simpletest agent toaccess theagentplatform serv
fspecificationsofgenericagent andacross agent based applications. Part of its enabling softwareframework. Contributorsare oftwareframeworkalsongaistsconstruction ification. In this way the individual software 1 FIPA-OSis an open source implementation of ecificationforagentinteroperability.The rcecodeanddocumentation. It alsoincludesa ices andsomevisualisationsoftware.
TheFIPA-OSarchitecturecanbenvisagedanasonmodel,entitiesinnon-adjacentlayerscanaccesse extendthearchitecture,noto nlybyappendingvalu agentsofacilitator r agentsontop)butinadditio ordeleted.Inadditiontothemandatorycomponents 2.1), theFIPA-OS d istributionincludes support for
strict layeredmodel. In non-strict a layered achotherdirectly.Thedeveloperisableto e-addedlayers(suchasspecialistservice n, lower or midlayerscanbreeplaced, modified oftheFIPAReferenceModel(seesection :
•
DifferenttypesoAgent f Shellsforproducingagent other using theFIPA-OS facilities;
•
Multi-layeredsupport for agent communication;
•
Messageandconversationmanagement;
•
Dynamicplatformconfigurationtosupportmultiple multipletypes of persistence(enabling integration
•
Abstractinterfaces andsoftwaredesignpatterns.
1
FIPA-OSisanagentframework,originatingfromre managedOpenSource[FIPA-OSURL]andicsurrently ACTS project FACTS[FACTS URL].
w s hichcanthencommunicatewitheach
communicationinfrastructuresand withlegacy p ersistencesoftware);
searchatNortelNetworks'HarlowLaboratoriesint usedwithinanumberofindustrialandacademicins
heUK.Iitsavailableas titutions,includingthe
2.1FIPA ReferenceModel TheFIPA referencemodelillustrates thecorecompo The agent reference model provides the normative fr operate. CombinedwiththeFIPA agentlifecycle, i for thecreation, operationandretirement of agent
nentsothe f FIPA-OS distribution (figure 1 ). amework within which FIPA agents exist and establishes t the logical and temporal contexts s.
Figure 1FIPA-OS : AgentFramework
TheDirectoryFacilitator(DF),AgentManagementSy stem(AMS)andAgentCommunication Channel(ACC)arespecifictypesoagents, f whichs upportagentmanagement. TheDFprovides "yellowpages" servicesto theragents.TheAMSan dACCsupportinter-agentcommunication usingFIPA’sagentcommunicationlanguage(ACL).Th eACCsupportsinteroperabilityb oth withinandacrossdifferentplatforms.TheInternal PlatformMessageTransport(IPMT)provides amessageroutingserviceforagentsonaparticula platform r thatmustbereliable,orderlyand adheretothedetailed requirements specifiedinse ction5 .2o(FIPA f 97 V2, Part 2). Together, the ACC,AMS,IPMTandDFformwhatwillbetermedthe AgentPlatform (AP).Theseare mandatory,normativecomponentsotfhemodel.Forf urtherinformationontheFIPAAgent Platform see(FIPA 97V2, Part 1, Agent Management) and(FIPA 98, Part 13). Inadditiontothe mandatory components of theFIPA includesanAgentShellthatcanbseeenaasnempt cancommunicatewitheachotherusingtheFIPA-OSf template.Detailsohf owthisAgentShellhasbeen paper aregiveninsection3 .1.
Reference Model, theFIPA-OS d istribution ytemplatefor anagent. Multipleagentsthat acilitiescanthenbep roducedfromthis usedtocreatetheagentsdescribedinthis
3.Designing VPN Provisioning Agents Thescenarioweconsideristhep rovisioningoapf Internet)as virtual a privatenetwork(VPN)foren enduserconsistsodifferent f combinationsorafa networkservicescanbebroadlyd ecomposedintothr agents. Thethreedomains andtheir representative 1. 2. 3.
ubliccommunicationnetwork(suchasthe du sers. Intherealworld, saingleservicetoan ngeonetwork f services(seefigure2).These eed omains,whichmayberepresentedb y agents are:
EndUser CommunicationDomain: PersonalCommunicationAgent (PCA). Network ServicesDomain: ServiceProvider DomainAgent (SPA). Network C onnectionDomain: NetworkProvider DomainAgent (NPA).
AServiceProviderhastheresponsibilitytotransl domainso f:i)theo verallnetworkservicedomain, enduser. Thegoalof allpartiesinvolveditsof service, cost andpenalty. NetMeeting Wrapper Agent (NMWA)
atethedifferingrequirementsothe f twoend andii)providingtherequiredservicetothe indtheb estdealsavailableintermsoquality f of
Outlook Wrapper Agent (OWA)
NetMeeting WrapperAgent (NMWA)
Outlook WrapperAgent (OWA)
ACL
ACL ACL
ACL Personal Communication Agent(PCA)
MeetingParameter Negotiation ACL
ACL ServiceNegotiation
ACL
Personal Communication Agent(PCA)
NPA1
SPA1 ACL ACL
NPA2 ACL
SPA2 ACL
NPA3 Connectivity Negotiation
Figure 2Inter : DomainAgentNegotiation
Therequirementsothe f serviceareestablishedb y considerthecaseinwhichtheenduserwishestos contactsotherusers’PCAsandschedulesvideocon a negotiateswiththeSPA toobtainthebestservice necessary parametersfor therequiredserviceandn solution. Theoverallproblemthenish owtoprovisionthese somecriteriasuchatsheprice, thetimeoqualit r theserviceproviders.Achievementofthisgoalmay amonganynumberosfervicep rovidersandcustomers servicep rovider whichboththecustomerando ther servicesfromrespectively.However,sinceeachgro mustberequested.Forexample,aservicep rovider service,orsaervicecustomermustbepersuadedto theo verallsystemcanbeviewedagroup as ointe f andconsumeservices tooneanother through serie a
thePCA withitsendu ser. Tobceoncrete, we etupavideoconferencemeeting.ThePCA ferencemeeting. TheinitiatingPCAthen dealavailable. The SPA then re-configures the egotiateswiththeNPAs toobtaintheoptimal
servicessothattheysatisfy,accordingto oythe f service, notonly thecustomersb utalso requireplanningandsolvingsubproblems Indeed, . thereareo ftenmorethanone serviceproviders canrequest services or sub upiasutonomous,servicesandinformation mustsomehowbepersuadedtoperforma accept laowerqualityoservice. f Therefore ractingandautonomousagentsthatprovide of smulti-lateral negotiations.
Additionally,theseagentsareoperatinginah ighl yd ynamicenvironment:servicesneedtobe updated,newonescomeonline,oldservicesarere movedandcurrentlyagreedservicesfail. Customer'srequirementsmayalsochange:n ewservic esmayberequired,servicesmaybe requiredsooner or later thaninitially anticipated higher , quality may b ecome more important, etc. Inall of thesecases, negotiationitshemeans of managing this complexity. New services become candidates of provisioning, thoseeffectedbtyhe f ailed services can breeprovisioned, and service conditionscanbedynamicallyconfiguredorreconfi gured.Inthecourseothese f negotiations, agentsarerequiredtomakeconcessionstooneanot her,tomaketrade-offsbetweenthe negotiationissuesandtosometimesbringnewissue isntothen egotiationino rdertofacilitate agreement making.
3.1ExtendingtheAgent Shell to construct Domain S
pecificAgents
AllagentsintheFIPA-OSdistributionextendthe AgentWorldAgentclass(partofthe aw package). Theinternalstructureoan f agentisno specified t andislefttotheindividualagent designer. Forexamplethefollowingobjectmodel( figure3illustrates ) howthep latformagents (AMS,DFandACC)havebeenconstructedbyinheriti ngandextendingtheagentshell, AgentWorldAgent. Allofthedomainspecificagents(PCA,SPA,NPA A , VWA,NMWA) describedinthis paper areconstructedisimila na fashion. r
Figure 3Object : Model for FIPA-OSPlatform Agents
Thekeyrolesinourapplicationarethoseothe fi 3.2),theservicep rovider(sub-section3 .3)andth theserolesaremappedtospecifickindsosfoftwar assumptions about thefuturenetwork environment: •
Typicallymultiplecompetingnetworkproviders(ISP auniquenetworkdomain,willallowsubscribingto during relatively a short periodotime. f
•
Serviceproviderswillh avethetaskofn egotiating selectingaspecificonetoactuallydeliverthese althoughtheISPsooperators r themselves canalso
nvolvedcustomers(orend-user)(sub-section en etworkp rovider(sub-section3 .4). Eachof eagents,takingintoaccountthefollowing ostelecom r operators), eachmanaging theirconnectivityservices,p ossibly withmultiplenetworkproviders rvice.Newplayerscouldtakethisrole, carry iout. t
3.2Personal Communication Agents ThePersonalCommunicationAgent(PCA)representst responsibleforenactingtheinterestsofitsendu scheduling meeting a and when selecting aanppropri thePCAsareresponsibleformanagingtheuser'sp r
heuserinthesystem.ThePCAis serwhennegotiatingwithotherPCAsfor ate SPA. To dtohis in the best way possible, ofileandpreferences.ThePCAisalso
and
responsibleforinteractingwiththirdpartysoftwa softwareinthiscase)athelocallevel(i.e.on following capabilities (moredetails of whicharep
re(CalendarManagerandVideoConference theuser'sterminal).ThePCAsupportsthe rovidedinsection4 ):
•
FIPAagentcommunicationprotocol(FIPAACLmessage andinteractionwithFIPAagentplatformcomponents Part 1).
•
FIPA agentnegotiationprotocols(fipa-iterated-con dialogueh andling (FIPA 97, V2, Part 2).
•
Userprofilemanagementfortheendusers’preferen videoconference(VC) configuration p arameters.
•
Interactionwiththirdpartysoftware.Thehandling platform specificandids escribedinmorein(FIPA
ofthisthirdpartysoftwareisagent 97, V2, Part3).
•
Serviceoffernegotiationcapabilitiesforissuesi description, andinformationabout thep articipants
ncluding:informationaboutevents,VC (including theorganiser).
handling)(FIPA97,V2,Part2) (DF,AMSandACC)(FIPA97,V2, tract-netandfipa-request) andsubsequent cesregardingschedulingmeetingsand
3.3ServiceProvider Agents TheServiceProvider Agent (SPA) canbceonsidered NPAs.Itexaminestheservicerequestfromthecoor (includingp riceandqualityoservice) f basedoni willbethesubjectofthenegotiationbetweenPCA successfully,theSPAwillsubsequentlystartn egot concludeintheselection othe f NPA that is most s totheservicerequestof thePCA. TheSPA supports
as V a PN service broker between PCAs and dinatingPCAandcomposesservice a offer tso wnservicecapabilities.Thisserviceo ffer andSPA.Ifthisn egotiationhascompleted iationswithseveralNPAs,whichwill uited tporovision therequired network, related thefollowing capabilities:
•
FIPAagentcommunicationprotocol(FIPAACLmessage andinteractionwithFIPAagentplatformcomponents Part 1).
•
FIPA agentnegotiationprotocols(fipa-iterated-con dialogueh andling (FIPA 97, V2, Part 2).
•
Serviceoffer composition, basedoset na of servic
ceapabilities.
•
Serviceandnetwork offer negotiationcapabilities description, price, penalty, andp articipants.
for issues including: event information, VC
handling),(FIPA97,V2,Part2) (DF,AMSandACC),(FIPA97,V2, tract-netandfipa-request) andsubsequent
3.4Network Provider Agents A NetworkProvider Agent(NPA)isresponsiblefor man offers arecomposed, basedo network capabilities whichcanbseupplied). Thesen etworkoffersareth andNPA. TheNPA supports thefollowing capabilitie
aging its ownnetwork d omain. Network (network related q uality oservice f parameters seubject of thenegotiationb etweentheSPA s:
ServiceDescription
rdf:Bag
serviceName:Literal serviceType:Literal
subClassOf subClassOf
ParticipantBag
VCDescription profile:VCProfileInfo security:SecurityInfo
EventBag
participant_li:ParticipantInfo
event_li:EventInfo
NodeBag node_li:NodeInfo
Service description:ServiceDescription events:EventBag price:PriceInfo penalty:PriceInfo
subClassOf
VPNService
Connection nodes:NodeBag qos:QosInfo
participants:ParticipantBag
EventInfo eventType:Literal startDateTime:Literal endDateTime:Literal duration:Literal
VCProfileInfo profileType:Literal frameRate:Literal viewFrameSize:Literal
QoSInfo availabilty:Float bandwith:Float NodeInfo ipAddress:Literal
PriceInfo amount:Float currency:Literal priceReference:Literal
ParticipantInfo userName:Literal email:Literal ipAddress:Literal
Figure 4:
SecurityInfo securityLevel:Literal
RDF Class Diagram for the ServiceOntology
•
FIPAagentcommunicationprotocol(FIPAACLmessage andinteractionwithFIPAagentplatformcomponents Part 1).
•
FIPA agentnegotiationprotocols(fipa-iterated-con dialogueh andling, (FIPA 97, V2, Part1).
•
Network o ffer composition, basedoset na of networ
ckapabilities.
•
Networkoffernegotiationcapabilitiesforissuesi description, price, andnodenames.
ncluding:Eventinformation,connection
handling)(FIPA97,V2,Part2) (DF,AMSandACC),(FIPA97,V2, tract-netandfipa-request) andsubsequent
3.5UsingtheFIPA ACL UnderstandinganACLmessagerequiresprocessingth positionwithinaparticularinteractionsequenceb understandingthetypeofcommunicationcalledaco messageandmayb earequestorstatementoffacto content andfinally understanding thesemantics of
emessagewithregardtoitstemporal etweentwoormoreagents.Thisinvolves mmunicationact,(itis pecifiedinthe query), r understandingthestructureothe f therequest.
AsACLcommunicationisorich,itiso ftenrepres (Finineal, t 1997).FIPA-OSsupportsACLcommunica conversation,ACLmessage,content(syntax)andont componentsn eedtobue sedb yeachagentandd iffer thesecomponentscanbesupportedateachlayerof flexibility ineeded s b ecauseiheterogeneous na wo thecontent differently.
entedamulti-tiered as layerinitso wnright tionusingfour mainsetsocomponents: f ology(contentsemantics).Notallof these entcombinationsodifferent f typesoeach f theFIPA-OSagentarchitecture.T his rld, different agents may encode and transport
FIPA-OSsupportsbothASCIIstringandXMLencoding oftheACLmessageusingthe appropriatedecoderandparser.Thereareseverals upportedencodingsfortheACLmessage contentincluding, FIPA SL0andFIPA SL1andthepr oposedFIPA-RDF specification(FIPA 99, Part18)forencodingthecontentinXML.Allagent isnthescenariodescribedinthispaper use theASCIIstringrepresentationforACLasdescribe din(FIPA97,V2,Part2)andFIPA-RDF encoding for thecontent expressions as describedi n(FIPA 99, Part18). TheVPNServiceProvisioningOntology,shownasan details of the issues over which the agents negotia in this document, arbitrary compositions of the iss beusedtoformthenegotiableServiceLevelAgreem interactionithe s FIPA-Iterated-Contract-Net proto
objectmodel(figure4),providesmore te. For the purposes of the negotiation d escribed ues illustrated itnhe object model presented can ents(SLAs). Moreover, thep rotocol for the col (figure5).
4.Implementing the Scenario EachofthePCA,SPAandNPAagenttypeshavebeen distribution.Thenegotiationabilitiesdescribedi additionalcomponent that extends theb asicfunctio eachothe f implementedagents is discussedbelow.
implementedusingtheFIPA-OS nsection5havebeenimplementedasan nality othe f Agent Shell. The functionality of
4.1PCA PCAsactonbehalfouf serstoarrangemutuallycon perhapslocation)for allparticipantswiththemin stressedthat althoughtheu ser interactionimini s belefttotheuser.WhenthePCA isstarted,itca interact witho ther agents. Specifically, it: •
Registers itself withtheAMS.
•
Registersitself withtheDF(Thisallowsotherage Identifier(GUID)(FIPA97,V2,Part1)sothatthe descriptionusedbythePCAcontainstheagenttype name" pca". Theownershipissettobetheemailaddressof represents.
venientmeetings(intermsotfimeand imum of interactionfrom theusers. It shouldb e mised, the final decision-making process should rriesoutna umberoftasksbeforeiistreadyto
ntstolookup thePCA’sGlobally Unique ycansendACLmessagestoiThe t). DF" facts-pca",servicetype" pca",service theusertowhichthePCA
Initiator,Participant
FIPA-Iterated-Contract-Net Initiator
constraint cfp,refuse*,notunderstood,propose, confirm,acceptproposal*,rejectproposal*,inform*, failure*
Participant cfp refuse
not-understood
propose reject-proposal
confirm reject-proposal
accept-proposal inform
failure
Figure 5The : FIPAStandardIteratedContractNetP •
SearchestheDFfortheexistenceofanOutlookCal belongs to theu ser.
•
SearchestheDFfortheexistenceoN af etMeetingW theu ser.
rotocol
endarWrapperAgent(OWA)that rapperAgent(NMWA)thatbelongsto
4.1.1CommunicationwithOWAandNMWA ThePCA is notitself designedtiontegrate with le agentsthatcanbeutilisedtocarryo utthenecess by thePCA aretheOWA andNMWA, which, respectivel andvideoconferencing tools. 4.1.2Negotiationwithother PCA’stoschedulemeet Uponreceiptofaninputfromtheenduserindicati PCA initiates negotiation a (initiating PCA is refe PCA’s(referredtoathe s RPCA’s). Thisnegotiation •
IPCA checks its calendar TheIPCAfirstensurestheuserisavailableatthe meeting. Thisias ccomplishedbyqueryingtheu ser negotiationphasehalts andanerror messageigen s
gacy systems, instead this task ileft s to wrapper aryinteractions. Thetwowrapperagentsused y,provideaccess totheuser’scalendar
ings ngthathewishestoarrangeameeting,the rredtoathe s IPCA) withtheinvitees’ “remote” canbseplitintofivedistinct stages: timestheywishtoattempttobooka ’sOWA.Iftheyarenotfree,thenthe erated.
•
IPCA sends call for proposals toall RPCAs TheIPCAdetermineswhichPCAsbelongtotheinvite dusersb yinterrogatingtheDFand sendsacallforproposals(CFP)toeach,including arangeopossible f timesatsowhento holdthemeeting. TheC FPalsocontainsinformatio nabouttheotherparticipants,although theirspecificdetailswillnotbeyetknown(e.g. whatIPaddresstheywouldbejoininga videoconferencefrom). WheninitiatingthenegotiationwiththeRPCAsthe ACLmessage encodingtheCFPindicatestherequiredInteraction Protocol(FIPA99,Part2 ),which constrainstheexpectedp erformativestobue sedin eachmessageexchange. Inthis case, it is theFIPA-Iterated-Contract-Net protocol.
•
RPCAs send proposals back totheIPCA TheRPCA’sresponseibs asedu pon: theCFP,thepre daytheywishtobeavailableformeetings)and,th rangessuggestedintheC FP. Thisprocessresultsi availableoa“failure”/”refuse” r beingsentbackt message, moredetailedinformationabout thep artic whichtheu serwilljointhemeetingandtheuser’s “refuse”/”failure”message(either duetotheOWA r or becausetheu ser isnot availableaall) t theOW
ferencesothe f user (e.g. whattimesof etimestheuserisavailableduringthe neither “apropose”withtimestheu ser is otheIPCA. Inthecaseothe f “propose” ipant is sentback (i.e. theIP address from videopreferences). Inthecaseotfhe efusing ofailing r tdoeal with the q ueries, A is closed.
•
IPCA evaluates theproposals OncetheIPCAhas receivedresponse a fromeachothe f RPCAs,providi ngthatatleastone RPCA has made proposal, a the actual meeting schedu ling process is executed. Otherwise the IPCA closes its OWA, andinforms theu ser of theo u tcome.
•
IPCA sends accept/reject messages totheRPCAs If m a eetingcannotbearrangedgiventheproposals made, “reject”messagesaresent toall of theRPCA’stoindicatethattheirproposalshavebe enrejected. Ifameetingcouldbe arranged,“reject”messagesarejustsenttotheRP CA’swhoseuserscannotattendand “accept”messagesaresenttotheRPCA’swhoseuser csan(alongwiththeprecisedetailsof themeeting, andtheother participants whowill be attending.
4.1.3NegotiationwithSPAs toarrangeVPN provisio Oncemeeting a timehasb eenarrangedandaleast t “accept”message, thePCA attempts tolocate numb a nonecanbelocated,thenthePCAwillsendarespo arrangedmeetingtimeandthep articipantswhocan indicating that V a PN has not beenp rovisionedfor Shouldaleast t oneSPA befound, theIPCA will beg ofservice(QoS)requiredforthemeeting(i.e.the definedbyeachu ser’sVCProfile,whichwereexchan negotiationfollowstheFIPA-Iterated-Contract-Net generatecounter-proposalsb asedupontheproposals tryingtoreach. Onceacompromisehasbeenreache agreementhasb eenmade,thePCAtheninformstheu arrangedmeeting.
ning oneRPCA hasrespondedfavourablytothe er of SPAs by sending query a totheDIfF. nsethroughthePCAGUI,indicatingthe attend,butwillincludewarning a message thetimeothe f meeting. in negotiating with the SPAs over the quality PCAwillaimtogettheclosestQoStothat gedduringthemeetingscheduling). This protocol,andthePCAhasamechanismto madebytheSPAsandtheQoStargetitis dbetweentheSPAandPCAandan ser,throughtheGUI,of thedetailsothe f
4.1.4StartingthevideoconferenceoncetheVPNhas
beencommissioned
Uponreceivingrequest a fromanSPAindicatingtha takes thefollowing steps:
the t VPNisbeingcommissioned,thePCA
•
IPCA configureslocal A/Vconferencinga pplication Theuser’sNMWAisinitialised,andthensentthe“ videoconferencingapplication(s)(i.e.NetMeeting/C anerror messageidisplayed s for theu ser.
•
IPCA informs RPCA’s that theconferenceiready s TheIPCAwillperformaDFlook-upofthenecessary messagewiththedetailsofthemeeting. Uponrece initialisetheparticipatinguser’sAVWA,andsend videoconferencingapplication(s)shouldcallintot AVWAhasinitiated. Oncethishasb eencompleteds inform/donemessagetoindicatesuccess,o therwise messagebaseduponthereasonwhy theuser cannot j
tohost themeeting start”commandtoconfigurethe amWiz)tohostthemeeting. If thisfails
RPCAs,thensendthema“request” iptofthismessage,anRPCAwill ita“call”commandinorderthatthe heconferencethattheIPCA’sassociated uccessfully,theRPCAreplieswithan itwillrespondwitharefuse/failure ointheconference.
4.2SPA SPAs negotiatewithanIPCA regarding the state and oftheservice,price,penalty)othe f servicethat collaboratingNPAstofindtheb estsolutionforth AnSPA hasaninterestinmaximising its profit.
conditions (e.g. typeoservice, f starting time theSPAistoprovide. Itthenn egotiateswith eprovisioningothe f servicetothecustomer.
4.2.1Register withAMSandDF TheSPAmustregisterwithanAMSand DF a beforei can t function. Thecurrentversionothe f SPA registers itself in the following manner: agent type “facts-spa”, ownership “www.agentworld.co.uk”, servicetype “spa”andservicen ame “spa”. 4.2.2Negotiates withPCA Whenameetingias greedamongthePCAs,theIPCAs endstheSPAsaservicerequest. The IPCAstartsthenegotiationoverservicecharacteri stics(includingprice,framerate,viewframe sizeandpenalty).The negotiationmayconsistoafnumberoifterationsf ollowingtheFIPAIterated-Negotiationprotocoluntilthenegotiation eithersucceedsofails. r Thenegotiationfails when the maximum time for negotiation othe f IPCA o SPA r is reached. The n egotiation succeeds whentheproposedservicecharacteristicsareagree dbetweentheIPCA andanSPA. 4.2.3Queries theDF for informationonavailableN Whentheservicecharacteristicsareagreed, theSP informationoavailable f NPAs.
PAs A send request a totheDFfor saearcho nthe
4.2.4Negotiates withNPA(s) WhentheDFreplieswithalistofNPAs,theSPAca agreedservice,andre-configurestheminimumandm negotiationtoensureaminimumprofit. TheSPAth proposals”(CFP)message,andstartsSPA-NPAnegoti subsetof theissues presentedinfigure4encoded
lculatesthenecessarybandwidthforthe aximumvaluesforpriceandpenaltyfor ensendseachoftheNPAsa“callfor ations. AnexampleCFPmessagewitha inXMLis givenbelow:
(cfp :sender spa3-spa@iiop://195.8.93.19:50/acc :receiver npa5-npa@iiop://195.8.93.19:50/acc :content ( SPA VideoConference Netmeeting ViedoConference Med Small 19990708T111100000Z +00000000T010000000 9.377499 85.0 7.4600005 ) :reply-with spa3-spa0 :in-reply-to spa3-spa0 :language FIPA-RDF :ontology vpn-service :protocol fipa-iterated-contract-net :conversation-id 931452074937)
If allof theSPA-NPA negotiationsfail, theSPA wi notbeprovided. However,if oneothe f SPA-NPA n withthePCA abouttheagreedserviceandiwill t t start time.
llinformtheIPCA thattheagreedservicecan egotiationssucceeds, theSPA willconfirm hensetitsinternaltimer toalarm attheagreed
4.2.5 InformsPCA ostarting f time
Whenthestarttimeothe f meetingias greed,theS agreedVPN andthat theIPCA starts theagreedAV c
PAwillrequestthattheNPAprovisionsthe onference.
4.3NPA NPAsareresponsiblefortheprovisioningothe f ne SPA. NPAsnegotiatewiththeSPAabouttermsandc Network Provider has aninterest inmaximising its
tworkconnectivity uponrequestsfromthe onditionsonf etworkconnectivity. A profit.
4.3.1NPARegisters withAMSandDF NPAsmustregisterwithanAMSandDF a b eforethey canfunction. Thecurrentversionothe f NPA registersitself with agent a type “facts-npa”, ownership “www.agentworld.co.uk”, service type “npa”andservicename “npa”. 4.3.2NPANegotiates withS PA WhenserviceprovisioningiasgreedbetweentheSPA NPAsaconnectionrequest(cfp). SPAstartsthen e (includeprice,bandwidth,availability andpenalty iterationsfollowing theFIPA-Iterated-Negotiation (whenthemaximumtimefor negotiationothe f SPA o succeeds whentheproposedconnectioncharacteristi
andIPCA,theSPAsendsselection a of gotiationoverconnectivitycharacteristics The ). negotiationmay consistof na umber of protocol until then egotiationsucceeds or fails NPA r has beenreached). Thenegotiation cs areagreedbtyheSPA.
5.AgentNegotiation Theprevioussectionshaveexplainedthesolutiont echnologythataddressestheACLandthe engineeringproblemfor M a AS. This sectionfocuses onthecoordinationnegotiationelement of theservicep rovisioningaspect of our system. Inp articular, wefocus onthereasoning model that the VPN agentsemployinordertoreachagreementsforprov idingandconsumingservicesfrom oneanother.Themodelisb asedonaformaldefinit ion,thedetailsandrationaleforthedesign choicesforwhicharecontainedin (Faratin etal 1998, , Faratin etal 1999, , Faratin etal 2000, , Jennings et al 2000, , Sierra et al 1999). , Thedecisionsfacedboyur agents are combination a
of:
•
offergenerationdecisions (whatinitialoffershouldbgeenerated?whatcoun begiveninsituationswheretheopponent's offer i unacceptable?), s and
•
evaluatory d ecisions (when should n egotiation baebandoned? and when sh bedeemedtohavebeenreached?).
Thesolutiontothesedecisionproblemsiscaptured evaluatoryd ecision(Evalcomponentsinfigure6i) offergenerationcomponents(orwhatisreferredto (Responsive, Tradeoff andIssueman. components in properties: •
thecomputational andinformational cost themechan
•
thesocial benefit of themechanism for theagents
teroffershould ould aangreement
intheagentarchitecture(figure6).The es xpandedonin(Faratin etal 1999).The asthemechanisms)o fthearchitecture figure6are ) d istinguishedbtyhefollowing ism incurs ontheagent
Thefirstproperty ifeature as thatdistinguishes our workfrom muchothe f worko ngametheory models(e.g. (RosencheinandZlotkin, 1994)). Thed esignothe f VPN negotiating agentis highly constrainedbythefactthatinrealworldapplicat ionsagentsareseldomcomputationallyand informationallyunbounded,anassumptionwhichunde rpinsmanyoftheclassicgametheory models.Theprovisioningoaf VPN serviceiasrealtimep rocess.Servicesarerequir edwithin tightschedulingwindowsand negotiation a mechanis m must respectthetimelimitsothe f agent. Furthermore,negotiationios nlysingle a elemento the f agent’sd eliberationsandsoimust t not consume disproportionate amounts ofcomputational r esource. In addition to being computationallyb ounded,agentsarealsoinformatio nb ound.Forexample, an IPCA agentwill
notknowthepreferences,resourcesoexecution r ti mechanism must findsolutions wheninformationips
melinesoaf rivate.
SPA agent.Thusnegotiation a
Thesecondfeaturerelatestotheconcernforthed esignofamechanismthatachievessome measureosocial f (orglobal)coherency fromlocal andd istributedprocessing.Since VPN agents areeconomical agents theno nesuchmeasureosoci f alcoherency ithe s sum of thevalue for each oftheagentsforano utcome(c.f.paretooptimald eals(Gorfman etal 1993)). , Nashiasnother, (Gorfman etal 1, 993).Usingthisfeature,wecandistinguishbet weenmechanismsthatare concerned with individual utility of the outcomes without concern for the social wel fare, and o nes that produce o utcomes that are both individually and jointly preferred btyhe agents. For example, ifad ealisrequiredverysoonthennegotiationbe tweenthe IPCA and SPA agentsids rivenb y concernfor daeal that is perhaps not socially opt imal but one that is agreeable bbyoth agents. On theo therhand,ifthereitsimetonegotiate,then thesamenegotiationmayinvolveb othagents searchingfordealsthataren otonlyindividually rationalbutmayalsobbe eneficialtotheother agent. Giventhesepropertieswehavedesignedandimpleme ntedthreenegotiationmechanisms (responsive,trade-offandissuemanipulationmecha nisms)whichd ifferentiallyimplementthese properties.Figure6describesthefunctional mode of ltheagent’sreasoningd uringn egotiation. Giventhenegotiationd eadline( Tmax),theopponent’slastoffer ( Y) andtheagent’slast offer ( X) (the three inputs into the negotiation architecture in figure 6all )three mechanisms simultaneously computeanewoffer( X’) (boxesResponsive,Trade-offandIssueman, infigure6).The mechanism’s evaluatory componentthenmakes thedec isiontoeither accept( A) or reject (R) the opponent’slastoffer Y,or counteroffer (X’),thenewcontract, totheopponent.Thefinal choice of whichmechanism’s suggestion to ffer is handled by the meta-strategy module. The processes involvedineachmechanism aredescribednext. 5.1ResponsiveMechanisms Responsivemechanismsmodelreactivenegotiationbe haviourtoanumberoef nvironmental factors. Theunderlying rationaleandmotivationfo the r designothese f mechanisms hasbeen the needtomodelconcessionarybehavioursthatareini tiatedbyprogressivelymoreimportant environmentalfactorsduringthecourseofnegotiat ionprocess.Forexample,if IPCAhas committedmanyresourcesduringitsn egotiationwit h SPAandthetimeothe f requiredvideo servicewithother RPCAsis oon,thenthe PCAmayprefersimpleandlesscostlydecision mechanisms that canresultinconcessions. Concessi onmay result inanagreement, and therefore not only free IPCA's resources, whichcanbuesedfor other activities, but also achieve the goal of establishing meeting a withthe RPCAs. Responsivemechanismsgenerateoffersbylinearlyc ombiningsimpledecayfunctions,called tactics (Faratin etal 1999). , Tacticsgeneratevaluesfor issuesu singo nly single a environmental criterion. Wehavedesignedthreefamilies of tacti cs: •
Time-dependenttactics m : odelincreasinglevelsofconcessionasthedead negotiationapproaches;
lineforthe
•
Resource-dependenttactics model : increasinglevelsoconcession f withdimini of resources;
shinglevels
•
Behaviour-dependenttactics :modelconcessionsbasedonthebehaviouroftheo negotiating party.
ther
Figure 6The : NegotiationArchitecture
Todeterminethebestcourseoaction, f anagentma nyeedtoconsider andassessmorethanjust oneenvironmentalcondition. Sinceeachtacticgen erates vaaluefor anissueusing only single a criterion,theconceptof strategy isintroducedtomodelthemodification,overtime, weights as thecriteria changetheir relativeimpor tanceinresponsetoenvironmentalchanges.
oftactic
5.2Trade-off Mechanisms Aconcessionmechanismgeneratescontractsthatare individuallyrational.Thatis,giventhe currentstateotfheenvironment, themechanismcomputeswhatisthebestofferanag entcan proposethatmaximises itsvalue. However, insomecasesthereiaslso need a for agents toact in amoresociallyresponsiblemanner.Thus,forexamp le,negotiationbetween NPAs tojointly provideanISservicetoa SPA agentisnaturally cooperative a activitywhenthe NPAs represent thesamen etworkoperator. Insuchcases,theagent asreconcernedbothwiththeoutcomeothe f negotiationforthemselvesandfortheirnegotiatio nopponent. Inshort,theycareaboutequity andsocialwelfare(Gorfman etal 1, 993),aswellastheirindividualutility.Acon cession mechanisminsuchcontextsiisnefficient, interms of possiblejointvaluegains (Rosencheinand Zlotkin,1994).Thisrequirementledutsodesignm odelsthatcanuncoverwin-winnegotiation solutions(Raiffa,1992),againinthepresenceof limitedknowledgeandcomputational boundedness.Win-Winnegotiationreferstobargai a ningsituationwherebothp artiessearchfor solutionsthat“squeezeout’’moregains(eithermu tuallyorindividually)thanthecurrently agreeddeal. Theparticularmechanismforwin-winnegotiationth atwed evelopedinthiscontextisthatof agentsmaking trade-offs(Faratin etal .,2000).Intuitively,tarade-offiswhereonepar tylowers its scores onsomen egotiationissues andsimultane ously d emands moreoonthers. Thus, an NPA mayacceptsaerviceolower f qualityiitis fchea per,or sahorterdeadlineiitreceives f higher a price.Suchmovementsareintendedtogenerateano fferthat,althoughothe f samevaluetothe proposer, may b enefitthen egotiationopponent and henceincreasetheoverallgainsbetweenthe twoagents.
Anagentwilldecidetomake trade-off a actionwhe nidoes t n otwishtodecreaseits aspirational level(denoted θ)for gaivenservice-orientedn egotiation.Thus, t heagentfirstneedstogenerate some/allofthepotentialcontractsforwhichire t ceivesthescoreof θ.Technically,itneedsto generatecontracts that lie otnhe iso-value (or in difference) curve for θ (Raiffa, 1992). Because all thesep otentialcontractshavethesamevaluefort heagent, itisindifferentamongstthem. Given thisfact,theaimofthetrade-offmechanismisto findthecontractthatismostpreferable(and henceacceptable) tothe negotiation opponent since this maximises the joint gain. More formally, aniso-curveids efinedafsollows: givenanaspira tional scoring value θ, theiso-curveset atlevel θfor agent aidefined s as: Iso(a θ)={x |Va(x)=θ} where Va(x) isthevalueocfontract x foragent a.Fromthisset,theagentneedstoselectthe contractthatmaximisesthejointgain.However,si nceanagentdoesn otknowitsopponent's utilityfunctionsomeformofapproximationins eed ed.Theh euristicwe mployitsoselectthe contract that is most “similar” ttohe opponent's l ast proposal (since this might be more acceptable totheopponent).Tocomputesimilarityweusethe conceptoffuzzysimilarity,(Zadeh,1971). Fuzzy similarity iasnapproximationh euristic, whi chsupportsreasoningothe f kind“ ifpitrue, s then is qclose tboeing true ”. This notion was chosen b ecause iallows t an agen to at pproximately modelthecloseness of twocontracts indecisionma king. Atrade-offisdefinedafsollows.Givenanoffer, x,fromagent ato b,andsubsequent a counter offer, y, from agent bto a, with θ =Va(x), taradeo ff for sent awithrespect to iyds efinedas: trade-off
(x,y) argmaxz a =
∈iso (a θS{)im(z,y)}
wherethesimilarity, Sim,betweentwocontractsisd efinedasaweightedco similarityothe f issues.Thesimilaritybetweentw ocontracts xand definedas: a Sim(x,y)= j ∈J wSim j j(x j,y j)
mbinationofthe yoverthesetofissues Jis
with {j ∈ J}w
j.
a = 1j
and
Simj being thesimilarity functionfor issue
5.3IssueManipulation Mechanisms Our other deliberationmechanism is theissueset m anipulation(Faratin et al 1999). ., Negotiation processesaredirectedandcentredaroundtheresol utionoconflicts f over saetofissues This J. setmay consistof oneomore r issues (distributed andintegrativebargaining respectively (Raiffa, 1992)). Forsimplificationwaessumetheo ntology ofthesetof possiblen egotiationissues(see figure4) J,is haredknowledgeamongsttheagents. Itisfu rtherassumedthatagentsbegin negotiationwithapre-specifiedsetof“core”issu es, J(core) ⊆J and , possiblyothermutually . to J(core) arenot permittedsincesome agreednon-coresetmembers, J(¬core} ⊆J Alterations featuressuchasthe priceofservicesaremandatory.However,elementsof J(¬core} canbe altereddynamically. Agentscanaddorremoveissu esinto J(¬core} astheysearchforn ew possibleandu pton ow unconsideredsolutions. IntheVPNscenario,agentsnegotiateoverbothcor eandnon-coreissues.Moreover,iits sometimesimportanttobeabletochangethesetof issuesino rdertomakeanagreementmore likelyorelevant. r Forexample,a SPA maybegin QoSnegotiationwitha NPAspecifyingo nly bandwidth.However,subsequently NPAmaydecidetoincludeintothe QoSnegotiationa packetloss issuewith high a value if SPA has demanded high a capacity bandwidth. Alternatively,
SPAmaydecidetoremovethe bandwidthissuefromthe QoSn egotiationwith NPAif IPCAhas changeditsdemandfromahighqualityvideoservic etoastandardaudioservice.Agents deliberate over how to combine these add and remove operators in manner a that maximises some measure ---suchasthecontractscore(Faratin etal 1999). , However,asearchofthetreeof possibleoperatorstofindtheoptimumsetoifssue smaybecomputationallyexpensive.To overcomethisp roblemweareinthep rocessoimpl f ementinganytimealgorithmsthatusethe negotiationtimelimits tocomputeapossibly , suboptimal, solution. 5.4MetaStrategies Theabove-mentionedmechanismsformallycapturethe decisionproblemso f VPNagentsinthe courseoservice f negotiation.However,sincethere arethreechoicesomechanism f anagentis nowfacedwiththed ecisionp roblemofwhichothe f threetochoosefrom.Thefunctionothe f meta-strategyreasoneristoresolvethisdecision problembyselectingoneothe f offersotfhe responsive,trade-off orissuemanipulationmechani sms.Themeta-strategy owhich f mechanism toselectcanbebasedonanyn umberofdecisionfa ctorssuchasthetimedeadlinetoreacha solution in egotiation, the concern for global opt imality othe f n egotiated solution, the b ehaviour of theopponentinn egotiation, negotiationiin ss talemate, etc. For example, if thetimetoreach a solutioniasdequate, VPN agentsmay adopt m a eta-strategy thatutilisesthe trade-off mechanism tosearchfor na egotiationdealthatincreasesthe globalbenefit. Ontheo ther hand, if thetimeto reach solution a ivery s short, then responsive a m echanism is more likely tfoind agreements that, althoughareolower f joint utility, are nonetheles better s than ndoeal. Alternatively, if negotiatio n isperceivedtobeindeadlockthenmodificationof thesetofissuesinnegotiationmaybreak deadlocks.Thechoiceruleofwhichmechanismtose lectiscurrentlybeingempirically investigatedinumber na of different negotiatione nvironments.
6.Conclusions Thispaper has describedthedesignandimplementat virtualprivaten etworks.Theimplementationibs as throughtheFIPA-OSplatform.Akeycomponentofth negotiationthattook placebetweenthevarioussta negotiationalgorithmswhe avedevelopedandimplem enableagents toengageinflexiblen egotiations an inthefaceovarying f environmentalfactors,issue mechanisms.
ionom af ulti-agentsystem for provisioning edu pontheFIPAstandardandwasrealised isapplicationwastheautomated keholder agents. Tothis end, we described the entedforthisapplication. Ouralgorithms idncorporate structures for concession making trade-offmechanismsandissuemanipulation
REFERENCES A.BondandL.Gasser(1988) ReadingsinDistributedArtificialIntelligence Publishers, Inc., SanMateo, California.
Morgan . Kaufmann
K.P.CorfmanandS.Gupta(1993) Mathematicalmodelsofg roupchoiceandnegotiation Handbook inOperations ResearchandManagement Scie nces, 5, 83-142. P.Faratin,C.Sierra,andN.R.Jennings(1998) Agents. Int. Journalof Robotics andAutonomous Systems2
NegotiationDecisionFunctionsforAutonomous 4(3-4) 159-182
P.Faratin,C.Sierra,N.R.JenningsandPBuckle . (1999) DesigningResponsivea ndDeliberative AutomatedNegotiators Proc. . AAAIWorkshoponNegotiation:SettlingConf lictsandIdentifying Opportunities, Orlando, FL, 12-18.
s.
P. Faratin, C. Sierra, and NR. .Jennings (2000) Using Similarity-Criteria tM o ake N egotiation Trade Offs. Proc. IntConf. OnMulti-Agent Systems(ICMAS-2000 ), Boston, USA.
-
FACTS URL: http://www.labs.bt.com/profsoc/facts/ T. Finin, Y. Labrou and JMayfield . (1997) Agents, Bradshaw JM (ed.), MITPress, 291-316.
KQML aan sagent communication language
In .Software
FIPA-OS URL: http://www.nortelnetworks.com/fipa-os FIPA97 V2, Part 2
Agent Communications October . 1998 URL:
FIPA97V2, Part 1
AgentManagement October . 1998URL:
FIPA97Part 3
Developer'sGuide October 1998URL:
FIPA99, Part 18
ContentLanguagelibrary
FIPA99, Part 2
http://www.fipa.org/spec/fipa97.html
Agent/Software Integration. October 1997 URL:
FIPA98, Part 13
http://www.fipa.org/spec/fipa97.html
http://www.fipa.org/spec/fipa97.html
http://www.fipa.org/spec/fipa98.html
January 2 000 URL:
AgentCommunication January 2000URL:
http://www.fipa.org/spec/fipa99.html http://www.fipa.org/spec/fipa99.html
N.R.Jennings,P.Faratin,T.J.Norman,P.O’Brie nandBOdgers . (2000) Business Process Management. Int. Journal of AppliedArtificial Intelligence H. Raiffa (1982)
TheArt andScienceoNegotiation f
J.S. RosencheinandGZlotkin . (1994)
AutonomousAgentsfor 14(2).
Harvard . University Press, Cambridge, USA.
Rules ofEncounter.
TheMITPress, Cambridge, USA.
C.Sierra,P.FaratinandN.R.Jennings(1999) DeliberativeAutomatedNegotiatorsUsingFuzzy Similarity.ProcEUSFLAT-ESTYLFJointConf.onFuzzyLogic,P almadeMallorca,Spain,155158. G. Weiss (1999) L.A. Zadeh(1971)
Multiagent Systems The , MITPress, Cambridge, Massachusetts. SimilarityRelationsandFuzzyOrderings.
InformationSciences, 3,177-200.