BrahimChaib-draa
LavalUniversity,ComputerScienceDepartment
PavillonPouliot,Ste-Foy,PQ,Canada,G1K7P4
chaib@ift.ulaval.ca
Abstract. This paperproposes to seeagent communicationlanguage
(ACL)asajointactivityandnotasthesumofthespeaker'sandhearer's
(speech) acts. In this paper, a conversation in the context of ACL is
viewed asajointactivitywhichcanberealizedas sequencesofsmaller
actions,manyofwhicharethemselvesjointactions.Socialagentswhich
participate tothis joint activity haveto coordinatetheir joint actions.
Ineach joint act, the participants face a coordination problem: which
actionsareexpected?Theanswertothisquestionproposedhere,isbased
oncomplexnotionsascollectiveintention,jointplan,jointcommitments
andthenotionofcommonground.
1 Introduction
Multi-agent systems (MAS) are the subject of an area of research studying
systems made up of multiple heteregenous intelligent software entities (called
agents)wherecompetitioncoexistenceorcooperationispossiblebetweenthem.
MASdiersfrom distributedproblem solving inthesense that thereisnocom-
mon global goalto be solved which is known at design time; onthe contrary,
a multi-agent system is generally peopled by dierent agents having dierent
purposes.
Inrecentyearstheinterestinmulti-agentsystems(MAS)hasgrowntremen-
dously, andtodaymulti-agenttechnology isbeingused in alargerange ofim-
portantindustrialapplicationareas.Theseapplicationsrangesfrominformation
managementthroughindustrialprocesscontroltoelectroniccommerce.Allthese
applicationshaveonethingincommon.Agentsmustbeabletotalktoeachother
todecidewhatactiontotakeandhowthisactioncanbecoordinatedwithoth-
ers' actions. The language used for this exchange is the agent communication
language(ACL).
Traditional approachesof ACL viewconversationsasthe sum of speaker's
and hearer'sspeech acts.This approachis counterintuitivein thesense itdoes
notreect ourintuitionaboutconversationswhich aresocial activities. Inthis
paper we consider a conversation as a joint activity which can be realized as
sequencesofsmalleractions,manyofwhicharethemselvesjointactions.Social
agents which participate to this joint activity have to coordinate their joint
actions. Ineach joint act, theparticipantsface acoordination problem: which
commonground.ThisistheapproachthatwehaveadoptedforACL, andthat
wedetailhere.
2 Agent Communication Language (ACL): An Overview
Arstattempttocometoastandardizedagentcommunicationlanguage(ACL)
cameforthfromtheARPAknowledgesharingprojectandproducedKQML.In
the context of this project, researchersdeveloped twomain components: (1)a
representation languagefor the contentsof messages(called KnowledgeInter-
change FormatKIF),whichis anextensionof rst-orderlogic; and(2)acom-
munication language KQML (Knowledge Query and Manipulation Language)
whichconsists ofaset of communicationprimitivesaimingto supportinterac-
tionamongagentsinMAS.KQMLincludesmanyperformativesofspeechacts,
allassertives(i.e.whenit statesafact) ordirectives(i.e.when itreectscom-
mandorrequest),whichagentsusetoassertfacts,requestqueriesorsubscribe
to services.A sampleKQMLmessagehas thefollowingsyntax (tell:senderA :
receiverB:contentsnowing),thatistheagentAtellstoagentBthatthepropo-
sitionitisrainingistrue.ThesemanticsofKQMLpresupposesthat eachhas
itsownvirtualKB(knowledgebase).Intheseconditions,tellingPcorresponds
toreportingthatPisinitsKB;askingforPisattemptingtoextractPfromthe
addressee's KB, etc. Up till now KQML is the onlyACL that is implemented
and (widely) used (at least in the academic world). Recently, the authors of
KQMLgaveitasemantics issuedfromthetheoreticalfoundationofSearleand
Vanderveken[11].
MorerecentlyanothereorttocometoastandardACLhasstartedthrough
the FIPA initiative. This eort brings Arcol to bear on ACL, a language de-
veloped by France Télécom [2]. In Arcol the set of primitives is smaller than
in KQML (primitivesin Arcol canbecomposed) andthis set also includes as-
sertivesordirectivesasinKQML.ArcolhasaformalsemanticsbasedonCohen
and Levesque's approach on speech acts [4]. Conversely to KQML, in Arcol,
agent Acan tell agent B that Ponly if A believes that P and believes that B
does not believe P. Thus, Arcol gives preconditions on communicative acts as
speciedbyitssemantics.
Although some work has been done on the semantics of individual speech
acts in KQMLand Arcol,littleis known aboutthe semanticsof conversations
andtherelationsbetweenspeechactsandtheconversationsofwhichtheyform
apart.Infact,ACLmustbeviewedasasort ofconversationbetweensoftware
agentsand not asa set of speech acts. In this sense,the semantics of ACL is
the semanticsof a conversationthat cannot be reduced to the conjunction or
composition ofsemanticsof itsspeech acts.
Inthecontextofmulti-agentsystems,conversationsbetweenagentsareused
tosharetasksandresultsasintheWinogradandFlores'conversationforaction
(WFcfa)(showninFig.1). Althoughwehaveconcernsabouttheadequacyofthis
Initialnode
Finalnode 3
7
8
6
9
4 5
2 1
S:Reject
S:Withdraw H:Renege
S:Decl are
S:Request
H:Promise
S:Withdraw H:Assert
H:Reject
S:Withdraw
H:Withdraw S:Counter
H:Counter
S:Accept
Fig.1.WinogradandFlores'Conversationforaction[18].
communicationprotocolamongsoftwareagents.Indeed,thismodelisadequate
forrequeststhat softwareagentsmakeofeachother andthat humansmakeof
agentsystems[1].
Inthe WFcfa model (see Fig.1), the circles (nodes) representthe states of
thecomplexspeechact,andthearcsrepresentspeechactsthatcausetransitions
fromstatetostateinthecomplexrequestact.Theauthorsofthismodelassert
that (i)states 5,7,8and9represent nalstatesof thecomplexact, (ii)state
1 initiates the conversation and, (iii) other states (2, 3, 4and 6) representan
intermediatestateof thecomplexrequest.
ConversationsbetweenagentsarealsousedinthecontractNetwhereagents
coordinate their activities through contracts to accomplish specic tasks. An
agentactingasamanager,decomposesitscontract(thetaskorproblemitwas
assignedwith)intosub-contractstobeaccomplishedbyotherpotentialcontrac-
toragents.Foreachsubcontractthe managerannouncesataskto thegroupof
agents.Theseagentsreceiveandevaluatetheannouncementandthosewiththe
appropriateresources,expertise,andknowledgereplyto themanagerwith bids
thatindicatetheirabilitytoachievetheannouncedtask.Themanagerevaluates
thebidsithasreceivedandawardsthetasktothemostsuitableagent,calledthe
contractor. Finally, managerand contractorexchangeinformation together dur-
ingtheaccomplishmentofthetask. Inthese conditions,wesee inthisprotocol
thefollowingkeysteps:1) themanagerannouncesatask; 2)agentsreplywith
bids;3)themanagerawardsthetasktothecontractor;5)themanagerandthe
contractorexchangeinformationabouttheresultsofthetask.
Marketmechanismsconstituteanotheraspectofcoordinationbetweenagents
z
9
z
9
R esult Bid
Taskassignment Callforbids
(x) (y)
Fig.2.MessagesExchangedintheContractNet(after[13])
mutualselectionorbyusingsomeauctionmechanismsasEnglish,Dutch,double-
auction,Vickery,etc.
Anotherimportantaspectofconversationsbetweenagentsturnsaroundthe
matchmaking problem. Matchmaking is based on a cooperative partnership
between information providers and consumers, assisted by an intelligent facil-
itator utilizing a knowledge sharing infrastructure [10]. Information providers
take an activerole in nding specic consumers by advertising their informa-
tioncapabilities to amatchmaker.Conversely,consumerssend requests forde-
sired informationto thematchmaker,whichin turns, attempts to identify any
advertisementsthat are relevant to the requests and notify the providers and
consumersasappropriate.
Fromalargerperspective,interactionsbetweensocialagentsmaybeconsid-
eredas takingplacewithin conversations,andconversationmaybeviewedasa
joint activity[3]which canberealizedassequencesofsmalleractions,manyof
whicharethemselvesjointactions.Socialagentswhich participateto thisjoint
activityhavetocoordinatetheirjointactions.Ineachjointact,theparticipants
faceacoordinationproblem:whichactionsareexpected?
Tosumup,ACLsareconversationsandtherefore
1. theiranalyzemuststartfromconversationstospeechacts;
2. theirformalsemanticsshouldemphasizeonsocialconcepts(asjointactivity,
jointactions,collectiveintention)andthecommunground reectingtheir
commonbackground;
ThiscanhelptoelaborateasemanticsofACL,sincethereisalackofconsen-
sus onthesemanticsof thecommunicationbetweenagents.Thisalso canhelp
theconversationsofwhich theyform apart.
Inourapproach,communicationsbetweenagentsareconsidered asconver-
sation types with goal-dened with constraintson the allowable contributions
andfairlyxedturnorder.Theseconversationtypesaremorestructureoriented
and in this sense they contrast with conversationsin general,which are more
processoriented.Theinteractionsinthislatertypeofconversationsarenotpre-
dictable because,ontheonehand,theydepend ontheinteractions' individual
historiesand experiences,andontheotherhand,theyarecontinuallyenlarged
and modied during the process of conversation. In the next section, we will
detail these aspects of structure orientedvs process orientedand explain why
ACLisastructure oriented.
3 Agent Communication Language is Structure Oriented
Discourses,andinparticularconversations,onlycomeintoexistencethroughthe
interactionoftwoormorepeoplewhomayhavecompletelydiverginggoalsand
intentions.Atthe beginning of aconversationitis very oftennotclear to any
ofitsparparticipantshowlongitisgoingandwhatitsnaloutcomewillbe.A
lotofworkthathasappearedduringthelasttwentyyearsorsoinfactassumes
that both text and discourses are structural entities requiring analytical tools
borrowed from sentence grammar. Proponents of approaches that fall within
this categoryclaimthat discoursesconsist ofwell-denedunits of alowerlevel
inverymuchthesamewayassentencesconsistofclausesofphrases,phrasesof
wordsandsoon.Inadditionthereareruleswhichgovernpossibleandimpossible
sequences ofthese units,just assentence grammardistinguishes between well-
formed and ill-formed sequences of words or phrases.In this case, there must
be amapping procedure which relates actual utterance,as they are produced
in discourses, to their underlying actions. In fact, such a mapping procedure
cannotbefoundbecauseutterancesmayhavemorethanonefunction(iftheyare
ambiguousor moreorlessintentionallyindeterminateas totheir illocutionary
force).
It should follow, then, that the structural tools which are used in sentence
grammars cannot be adopted for the analysis of conversations. In general, a
conversationisacomplexprocessbasedontheinteractionbetweenparticipants.
Wenowdetailthis view.
Participantsinaconversationhaveintheirbrainsalargesetofassumptions,
knowledgeandbeliefs.Assumptions,knowledgeandbeliefsabouttheworldthey
livein,abouttheir particularsociety,aboutthemselves,andabouttheirexpe-
riences. They also have assumptions and beliefs about each other. They use
these assumptions and beliefsaccording to therelevancetheory asestablished
bySperberandWilson(e.g.[14,15]).Inthistheory,everyutterancecomeswith
atacitguaranteeofitsownoptimalrelevance,thatistosay,thespeakeralways
assumesthatwhatheorshehastosaywillbeworththeaddressee'swhiletopro-
assumptions,knowledgeandbeliefscorrespondtosomeextenttothenotionsof
background+networkaspostulatedbySearle[12]ortothenotionofcommun
ground asintroducedbyClark[3].AccordingtoSearle,thenetworkcomprises
thepropositionalbeliefsthataredirectlynecessarytounderstandanutterance,
whereas the backgroundcomprisesthose assumptionsand beliefs that are too
basictobeneededdirectlyfortheinterpretationofutterancesbutwhicharenec-
essaryifthenetworkpropositionsarespeltoutindetails.Infact,thedistinction
betweenthesetwonotionsisnotimportantheresincebothreectassumptions
and beliefs and they only dier by their degrees of manifestness. Similar tho
these twonotions,isthecommunground notionintroducedbyClark[3].Ac-
cording to Clark,common groundisasine qua non foreverythingwe dowith
othersastheactivityofconversationsthatheconsidersasajointactivity.
Inlightoftheseconsiderations,relevancetheorycanaccountforconversation
if they are taken to be processes rather than structures, processes in which
the aimsof theparticipantscanchange orshift. Thus, what is relevantto the
participants changes continuallyalong withthe changingset ofbackground or
commungroundassumptions.
Levinson [9] introduced the useful notionof activity type, which is more
general than discourse type because it also includes social events to which
verbalexchangesareonlyincidental,asforinstanceagameofsoccerorthetask
ofrepairingacomputerin alab.
Itakethenotionofanactivitytypetorefertoafuzzycategorywhosefocal
members aregoal-dened,socially constituted,bounded,events with con-
straintsonparticipants,setting,andsoon,butaboveallonthekindsofal-
lowablecontributions.Paradigmexampleswouldbeteaching,doctor-patient
consultation,a round-tabledebate,ajob interview,ajuralinterrogation,a
footballgame, ataskin aworkshop, adinnerparty, andso on. ([9]p.368,
hisemphasis).
Thenotionsgoal-denedandconstraintsdonotapplytoallactivitytypes
ordiscoursetypeinthesamemanner.Forinstance,adoctor-patientconsultation
certainlyhasawell-deneddominantgoalandthereareagreatnumberofcon-
straintsonallowablecontributions.Ontheotherhand,conversationsingeneral
havefewconstraints,andtheirdominantgoalareill-dened.Thereare further
variables,concomitantwiththese assummarizedin Tab.1,takenfrom [7].
Aswesee,conversationsnaturallytendverymuchtowardsthelefthandside
of the scales in Tab. 1, whereas exchanges foran auction, interviews forjobs,
orcourtroomexaminationsaresituated towardstheoppositeend.Oneobvious
dierencebetweenthetwoistherolesoftheparticipants,andconcomitantwith
this,theturnorder andthetypesofturnthatmayoccur.
Inthecaseofagentcommunicationlanguageandaswecanseethroughthe
examplesofsection2:
1. therearemanyconstraintsonallowableagents'contributions;
ProcessOriented Structure Oriented
Fewconstraintson Manyconstraintson
allowablecontribution allowablecontributions
Multipleandmainly?goals Fewandmainlydominantgoals
Relativelyfreeturnorder Relativelyxedturnorder
Rolesofparticipantsnot Rolesofparticipants
clearlydened clearlydened
Contributionlargelydetermined Contributionoftennotdetermined
bypreviouscontributions bypreviouscontributions
Localorganizationprinciples Globalorganizationprinciples
3. allagentspresentin interactionareassignedclearlydenedroleswithinthe
proceedingsand the rightto communicate and theallowablecontributions
areinseparablylinkedwiththeindividualroles;
4. relevancecan'taccountsincetheaimsofparticipantsdonotchangeorshift.
4 Agent Comunication Language is a joint activity
Searle [12]suggeststhat foraminimalunderstanding ofconversationasathe-
oretical entity, it must be seen asan expression of shared intentionality. This
kindofintentionalitytranscends theconjunctionofthespeaker'sindividualin-
tentional states,becauseinfact itistheintentionalityofaninclusiveWe(such
aswhenpushingatabletogether,anactinwhichIpushthetableasanintrinsic
contributionto thefact ofpushingthetabletogether).Searle putsforwardthe
ideathatsharedintentionality(i.e.,jointorcollectiveintention,orwe-intention)
isaanimportantfactorintheexplanationofanyformofsocialbehavior.
More precisely, Searle claims that all conversations are forms of collective
intentionality. Heconsiders this collective intentionality asaprimitive concept
whichdoesnotreduce toindividual intentionalityplusmutualknowledge.The
high level of the collective intentionality we are doing such and such allows
within it for individual intentionality of the form I am doing so and so as
partof ourdoingsuch and such. Thus, theindividual intentionalityis partof
the collective intentionality. In this case, conversationsare forms of collective
intentionalityand the We-intention of We aretalking aboutthe price of X'
allowsfordieringIintentions,e.g..formyIintention:Ioeryou$5forX
andforyourI-intentionIrefuseyouroer.
TheClark's notionofjoint project[3]canservetomakeclearthisnotionof
jointintention.ClarkstartedfromthenotionofanactivitytypeofLevinsonas
previouslyintroducedandconsiderslanguageasonetypeofjointactivity,onein
which languageplaysanespeciallyprominentrole. Inconversations,according
to Clark,peopleaccomplishonepieceat time.Theydothat viajointprojects:
projected by one of its participants and taken up by the others. Joint projects
requireparticipantsinconversationstocommittodoingthingswitheachother.
Thus,converselytoautonomousactionswhereindividualshavetobewillingand
abletodo,jointprojectneedthecommitmentofalltheparticipants.
Inthese conditions,wecanstate:
Thesis1:Jointintentionsarejointprojects+jointcommitments
Thesis2:Jointprojectsareplanstoperformsomeactivityjointly
Jointintentions,Jointplans (or jointprojects) and jointcommitmentscan
belinkedthroughtheplanterminologybythefollowingTuomela'sthesis[17]:
Someagents(sayA
i
;:::;A
j
;:::;A
m
)haveformedthe(agreement-based)
jointintention to perform X i eachof them(a) hasa plan toperformX
jointly;(b)hascommunicatedthisacceptancetotheothers,and(c)because
of(a)and(b)itisatruemutualbeliefamongA
i
;:::;A
j
;:::;A
m
thatthey
are jointly committed toperforming X and thatthere is or will be a part
or share (requiring at leastpotential contribution)of X for each agent to
performthatheaccordinglyisorwillbecommittedtoperforming
Infact,Tuomeladistinguishedbetweenthejointintentionandthenotionof
acceptingandendorsingaplan toactjointly.Forhim, havingajointintention
amounts to acceptingand endorsing a plan, provided the agents' have in the
simple core cases at least, communicated their acceptance to each other and
havebecomejointlycommittedtocarryingitout[16].
Noticethat in thecaseofconversationsprocessorientedthejointintention
cantaketime toemerge sincethis jointintentionisnotgiveninadvancebut
builtupinthecourseofandbecauseoftheconversation.Itisinfacttheobject
of continuous negotiation and evaluation by theinterlocutors. In conversation
structure oriented in contrast, there is apurpose and topic from which agents
can enter under theumbrellaof ajoint intentionveryeasily. Inthecasewhere
protocols are used in ACL, we canassume that each protocol is some sort of
jointproject onwhich agentshavetobecommittediftheywantto useit.In
thiscaseitiseasyforagentstoknowwhichcommunicationprotocol touse.
Whenagentstakepartin jointplan (oractivity),theyperformavarietyof
jointactions.Manyofthesejointactions,ortheirpart,arecommunicativeacts
throughwhichtheygetotherstounderstandwhattheymean.Asthesecommu-
nicativeactsarelinkedtospeechacts,weshouldconsiderhereanewperspective
ofspeechacts,aperspectivewhich isnotconcernedwithonlythesender'sper-
spectiveasisthecaseodtraditionalspeechacts.Asnewperspectivesofspeech
acts,weshouldconsider(1)theroleofanyspeechactinaconversationbyknow-
inghowitisrelatedtootheractsintheconversation(particularlyrelationsacts
betweenactsinwhatspeaker(orhearer)say,andrelationsbetweenactsinwhat
speakerand hearer say); (2) the generation of all potential illocutionary acts
speaker'meaning.
As ourapproach views ACL asajoint activity, wesuggest to consider the
followingthesis:
Thesis3:speaker'smeaningisatypeofintentionthat canbedischargedonly
throughjointactions.
Thesis4: Illocutionaryactsand perlocutionary actsareaccomplishedonly as
jointactions.
Searle's second constructiveproposalconcerns the importance for atheory
ofconversationofthenotionofbackground.Thisnotionisnecessaryforthemu-
tualunderstandingbytheinterlocutorsofconversationalsequences.Inorderto
understand asequence, onehasto embed theintrinsic semanticsof utterances
in anetworkoffundamentalbeliefsandknowledgewhichservesasthebedrock
onwhichunderstandingisbased.Thisbackgroundfunctionsasanecessarycon-
textinanycommunicativeinteraction,andconversationalrelevanceisnecessary
measuredwithreferenceto it.
Here also Clark oers similar notion that he called commun ground. This
notionhasaclearerprole,and specially makesthebackgroundnotionempir-
icallyoperational.ForClark,Most joint activitiesget realized assequences of
smalleractions,manyofwhicharethemselvesjointactions.Thisisalsothecase
of conversations where participants have to coordinate their joint actions. In
each jointact, theparticipantsface a coordination problem: whichactions are
expected?Tosolvethisproblem,Clarkproposesacoordinationdevicebasedon
thesolutionthat ismostsalient,prominent,orconspicuouswithrespecttothe
commungroundoftheparticipants.
Accordingto Clark,commongroundisasinequa nonforeverythingwedo
withotherfromthebroadestjointactivitytothesmallestactionsthatcomprise
them.FortwoagentsAandBtoactjointly,theyhavetocoordinatewhatthey
doandwhentheydoit.Andtocoordinate,theyhaveto appeal,ultimately, to
theircurrentcommonground.Atthesametime,witheveryjointactionAandB
perform,theyaddtotheircommonground.Thisishowaexiblecommunication
betweenhumanprogress.
Wecando thesamefor communicationbetweenagents,in thesense where
agentsenter aconversation,bypresupposingcertain commongroundandwith
eachjointactioneachspeechacts,forexampletheytrytoaddtoit.Todothat
they need to keep track of their common ground as it accumulates increment
byincrement.It isclearthat in thecaseofACL, acommon groundshouldin-
clude(1)rulesforthecoordinationofjointactions;(2)conventions,obligations
[5], norms and social rules; (3) shared expertise; etc. In addition to the com-
mon ground,eachagenthasherown orpersonalgroundwhichreect herown
Wehavepresentedhereanewperspectiveof AgentCommunicationLanguage,
aprespectivewhich viewsitasaconversationandthereforeasasocialactivity.
Evidently,otherresearchershavefocusedonconversationconsideredasasocial
activity: (1) Cohen and Levesque have proposed persistent goal and mutual
belief[4];(2)Groszhasproposedasharedplan[6];(3)SinghhasproposedJoint
commitments[13].Wehavetakenthesameroadandintroducedheresomenew
ideastakenmainlyfrom SearleandClarkwork:collectiveintention, jointplan,
jointcommitmentsandcommonground.
Acknowledgments:
ThisresearchwassupportedbytheNaturalSciencesandEngineeringResearch
Council ofCanada(NSERC), bytheSocialSciencesand Humanities Research
Council(SSHRC)ofCanada,bytheFondspourlaFormationdesChercheurset
l'aideàlaRecherche(FCAR)duQuébecandinpart,bytheGermanResearch
CentreforAI(DFKI).
References
1. Bradshaw, J. M., S. Duteld, B.Carpenter, and Robinson, T. KAoS: A Generic
ArchitectureforAerospaceApplications.Proc.oftheCIKM'95-WorkshoponIntel-
ligentInformationAgents,MD,USA,1995.
2. Breiter,P.andM.D.Sadek.Arationalagentasakernelofacooperativedialogue
system: implementinga logicaltheoryofinteraction.inProc.ECAI96Workshop
Agent Theories, Architectures, and languages, SpringerVerlag, Berlin, 1996, pp.
261-276.
3. Clark,H.H.UsingLanguage,CambridgeUniversityPress,1996.
4. Cohen, P. and H. Levesque. Performatives in a rationally based speech act the-
ory,Proc.28th Annual Meeting of the Association for Computational Linguistics,
Pittsburg,PA.1990.pp.79-88.
5. Dignum, F.and Linder, B.Modelling social agents: communicationas Action.in
J. P.Müller, M. J.Wooldridge andN. R.Jennings (Eds). Intelligents Agents III,
SpringerVerlag,1996,pp.205-218.
6. Grosz, B. and S. Kraus. Collaborative plans for complex group action.Articial
Intelligence,vol.86, 1996,pp.269-357.
7. Jucker,A.H.Conversation:structureorprocess?inH.ParretandJ.Verschueren,
(eds).(On)SearleonConversation,BenjaminsPub.,Philadelphia,1992,pp.729.
8. Labrou,Y.andT.Finin.Semanticsandconversationsforanagentcommunication
language.inM.Huhnsand M.Singh(eds).Readingin Agents, MorganKaufman,
SanMateo,Calif.,1988.pp.235-242.
9. Levinson,S.C.Activitytypesandlanguage.Linguistics,17,1979,pp.365-399.
10. Patil, R. S., R. Fikes, P. F. Patel-Schneider, D. McKay, T. Finin, G. Thomas
andR.Neches.TheDARPAknowledgesharingeort:progressreport.Proc.ofthe
Third Int.Conf on Principles of KnowledgeRepresentation and Reasoning, 1992,
Univ.Press.1985.
12. Searle, J.:Conversation. inH. Parret and J.Verschueren, (eds). (On)Searle on
Conversation,BenjaminsPub.,Philadelphia,1992,pp.729.
13. Singh,M.MultiagentSystems:aTheoreticalFrameworkforIntentions,Know-How,
andCommunications.SpringerVerlag.Heidelberg,1994.
14. Sperber,D.andD.Wilson.Ironyandtheuse-mentiondistinction.inP.Cole(ed),
RadicalPragmatics,AcademicPress,NY,1981,pp.295-318.
15. Sperber,D.andD.Wilson.Relevance.CambridgeMA:HarvardUniversityPress,
1986.
16. Tuomela, Raimo. The Importance of Us: A Philosophical Study of Basic Social
Notions.StanfordUniv.Press,Stanford,CA.
17. Tuomela,Raimo.Philosophyanddistributedarticialintelligence:thecaseofjoint
intention.InG.M.P.O'HareandN.R.Jennings,(eds).Foundations ofDistributed
ArticialIntelligence,Wiley&Sons,Inc.1996,pp.487-503.
18. Winograd, T. and F. Flores. Understanding Computers and Cognition: A New
FoundationforDesign.Ablex,Norwood,NJ,1986.