2005/07/27Commitment-based Conversations 1 Multi-Agent System Communication Paradigms Rob Kremer University of Calgary Department of Computer Science Calgary, CANADA Abstract: Multi-Agent Systems seem like a good idea -- solve problems by using several different specialized agents who cooperate to reach a solution. Aside from obvious problems (such as disagreements, inconsistencies, and composition), communication is a significant impediment. Agents may be written by different people in different styles, with different objectives. Simple communication protocols tend to be brittle and fail because not all agents my behave identically, and my respond in unanticipated ways. Attempts to formalize and standardize communication, such as the one the Foundation for Intelligent Physical Agents (FIPA) developed using BDI (Belief, Desire, Intension) semantics, encountered difficulties that made implementation and conformance difficult. Some of these problems stem from the FIPA standard's requirement that all agents model the beliefs, desires, and intentions of all other agents, which makes for a rather unnaturally omniscient society of agents. One way to address this issue is to neglect the modeling of the "minds" of other agents, and model only the social commitments agents undertake (or fulfill) in the course of their conversations. To achieve this, conversations are modeled as sequences of speech acts, where rules or policies (which may be considered social norms) translate from the speech acts to the instantiation or deletion of social commitments among the participant agents.

03/06/2016CPSC /599.68: Agent Communications 2 Simple Conversation: naive What time is it? It’s 2:00

03/06/2016CPSC /599.68: Agent Communications 3 Simple Conversation: more complex What time is it? It’s 2:00 Just a sec, I’ll check Thanks

03/06/2016CPSC /599.68: Agent Communications 4 Go ask someone else. Simple Conversation: failure What time is it? I don’t know, my watch stopped (didn’t hear)

03/06/2016CPSC /599.68: Agent Communications 5 State Machine Eg: “call for proposals” A MIP-net (Multi-agent Interaction Protocol net) – combining two A-nets (Agent nets) and one IP- net (Interaction Protocol net). Reference: Sea Ling & Seng Wai Loke. A Formal Compositional Model of Multiagent Interaction. AAMAS’03 – International Conference on Autonomous Agents and Multi- Agent System, July 14-18, 2003, Melbourne, Australia. ACM, Also available: 000/860791/p1052- ling.pdf?key1=860791&key2= &coll=&dl=ACM&CFID= &CFTOKEN= /860791/p1052- ling.pdf?key1=860791&key2= &coll=&dl=ACM&CFID= &CFTOKEN= Ling & Loke 2003, p.1053

03/06/2016CPSC /599.68: Agent Communications 6 BDI Belief –Environment Desire –Goals Intention –The current desire(s) chosen by the selection function with the beliefs and desires as input

03/06/2016CPSC /599.68: Agent Communications 7 Motivation Agents aren’t objects – therefore they must communicate in order to influence other agents (as opposed to invoking methods on other agents) We treat such communications just like other actions, and call it “speech act theory” 1,2 1 John Austin. How to Do Things With Words. Oxford University Press: Oxford, England, John Searle, Speech acts: An Essay in the Philosophy of Language. Cambridge University Press: Cambridge, England, 1969.

03/06/2016CPSC /599.68: Agent Communications 8 Example The state I might want to bring about is that “you believe it’s raining in London”: B you Weather(london, raining) but just because I send you “Weather(london, raining)” in a message doesn’t necessarily mean that you will come to believe it (I may be a compulsive liar, or very bad at weather forecasts)

03/06/2016CPSC /599.68: Agent Communications 9 Speech Acts Performative verbs –Eg: request, inform, promise Aspects: –Locutionary act: making an utterance –Illocutionary act: action performed in saying something –Perlocution: the effect of the act Felicit conditions: –There exist an accepted procedure, and circumstances and persons must be specified –Must be executed correctly and completely –Must be sincere, and must have uptake

03/06/2016CPSC /599.68: Agent Communications 10 Example Request Normal I/O conditions: –Hearer not deaf –Not uttered in a play or film Preparatory conditions: –Speaker must correctly choose the act –Hearer must be able –Speaker must believe the hearer is able –Hearer wouldn’t have done the act anyway Sincerity conditions: –Not sarcasm, etc.

03/06/2016CPSC /599.68: Agent Communications 11 Types of Speech Acts Representatives: commits the speaker to the truth of a proposition –Eg: Informing Directives: attempt to get the hearer to do something –Eg: Requesting Commissives: commit to a course of action –Eg: Promising Expressives: express some psychological state –Eg: Thanking Declarations: effects some change in an institutional state of affairs –Eg: Declaring war

03/06/2016CPSC /599.68: Agent Communications 12 Rational Action Speech act theory can be considered as a specialization of a more general theory of rational action 1 Eg: “A request is an attempt on the part of spkr, by doing , to bring about a state where, ideally, (i) addr intends  (relative to the spkr still having that goal, and addr still being helpfully inclined to spkr), and (ii) addr actually eventually does , or at least brings about a state where addr believes it is mutually believed that it wants the ideal situation” 1,p P. R. Cohen and H. J. Levesque. Rational interaction as the basis for communication. In P. R. Cohen, J. Morgan, and M. E. Polleck (eds), Intentions in Communications, pp The MIT Press: Cambridge, MA, 1990.

03/06/2016CPSC /599.68: Agent Communications 13 FIPA Messages performativethe type of the communicative act of the message senderthe sender of the message receiverthe intended recipients of the message reply-tosubsequent messages in this conversation thread are to be directed to the agent named contentthe content of the message; equivalently denotes the object of the action languagethe language in which the content parameter is expressed encodingthe specific encoding of the content field ontologythe ontology(s) used to give a meaning to the symbols in the content expression protocolthe interaction protocol that the sending agent is employing with this ACL message conversation-idan expression used to identify the sequence of communicative acts that form a conversation reply-withan expression that will be used by the responding agent to identify this message in-reply-toan expression that references an earlier action to which this message is a reply reply-bya time and/or date expression indicating the latest time the sending agent wants a reply See FIPA standard SC00061GFIPA standard SC00061G

03/06/2016CPSC /599.68: Agent Communications 14 FIPA Performatives (informally) accept-proposalaccepting a previous proposal agreeagreeing to perform some action cancelinform another agent that the agent no longer need perform some action call-for-proposalcalling for proposals to perform an action confirminforms a given proposition is true disconfirminforms a given proposition is false failurean action was attempted but failed Informa given proposition is true inform-ifinform whether or not a proposition is true inform-refinform the object which corresponds to a descriptor not-understooddid not understand what the receiver just did propagatepass a message on proposesubmitting a proposal to perform an action proxypass on an embedded message query-ifasking whether or not a proposition is true query-refasking for the object referred to refuserefusing to perform an action reject-proposalrejecting a proposal during negotiation requestrequest to perform some action request-whenrequest to perform some action when some proposition becomes true request-wheneverrequest to perform some action each time the proposition becomes true subscriberequesting to notify of the value of a reference whenever the object changes See FIPA standard SC00037JFIPA standard SC00037J

03/06/2016CPSC /599.68: Agent Communications 15 FIPA Semantics FP: B i φ  ￢ B i (Bif j φ  Uif j φ) RE: B j φ Rational Effect Semantic content Feasibility Precondition Agent Proposition U j φ  U j ￢ φ (where U j means “j is uncertain about φ”) Bjφ  Bj￢φBjφ  Bj￢φ “j believes φ”

03/06/2016CPSC /599.68: Agent Communications 16 FIPA inform semantics Formally: FP: B i φ  ￢ B i (Bif j φ  Uif j φ) RE: B j φ Example: (inform :sender (agent-identifier :name i) :receiver (set (agent-identifier :name j)) :content "weather (today, raining)“ :language Prolog)

03/06/2016CPSC /599.68: Agent Communications 17 FIPA request semantics Formally: FP: FP(a) [i\j]  B i Agent (j, a)  ￢ B i I j Done (a) RE: Done (a) FP(a) [i\j] denotes the part of the FPs of a which are mental attitudes of i. Example: (request :sender (agent-identifier :name i) :receiver (set (agent-identifier :name j)) :content "open \"db.txt\" for input“ :language vb)

03/06/2016CPSC /599.68: Agent Communications 18 Plan Preconditions: P Predicate Postconditions: P Predicate Actions: Seq Action hd:Plan  Action // returns head(Plan.Action) tail:Plan  Plan // returns {Pre,Post,tail(Actions)}

03/06/2016CPSC /599.68: Agent Communications 19 BDI Agent BDIagent B: P Bel D: P Des I: P Int getNextPercept:   Percept brf: P Bel  Percept  P Bel options: P Bel  P Int  P Des filter: P Bel  P Des  P Int  P Int plan: P Bel  P Int  Plan execute:Plan

03/06/2016CPSC /599.68: Agent Communications 20 Basics First order model logic with identity Three primitive attitudes: B i pi (implicityly) believes p U i pi is uncertain about p but thinks that p is more likely than ¬p C i p(choice) i desires that p currently holds Other attitudes: PG i pi has p as a persistent goal (desire?) I i pi has the intention to bring about p

03/06/2016CPSC /599.68: Agent Communications 21 Actions a 1 ; a 2 sequence in which a 2 follows a 1. a 1 | a 2 nondeterministic choice in which either a 1 happens or a 2 happens, but not both. Feasible(a,p) a can take place and, if it does, p will be true just after that. (Feasible(a)  Feasible(a, True) Possible(p)   a.Feasible(a,p). Done(a,p)a has just taken place and p was true just before that. (Done(a)  Done(a,True) Agent(i,a)i is the only agent that ever performs (in the past, present or future) the actions a. Single(a)a is not a sequence. ¬Single(a 1 ;a 2 ), but Single(a 1 |a 2 ) iff Single(a 1 )/\Single(a 2 )

03/06/2016CPSC /599.68: Agent Communications 22 Abbreviations Bif i p  B i p \/ B i ¬p Bref i l x  (x)   y.B i ( l x  (x) = y). Agent i believes that it knows the (x which is) . Uif i p  U i p \/ U i ¬p Uref i l x  (x)   y.U i ( l x  (x) = y). AB n,i,j p  B i B j B i …p. n is the number of B operators alternating between i and j.

03/06/2016CPSC /599.68: Agent Communications 23 Property: intending to achieve a RE (  x. B i a k =x)// there exists an action /\ RE(a k )=p // who’s RE is p /\ ¬C i ¬Possible(Done(a k ))// that i thinks should be done  (I i p  I i Done(a 1 | … |a n )// then if i intends p, then i intents one of the act that can achieve it Where a 1,..., a n are all the acts of type a k “If I intent to break the vase, then I intent to either drop it, or smash it with a hammer”

03/06/2016CPSC /599.68: Agent Communications 24 Property: satisfiability of intent  = I i Done(a)  B i Feasible(a) \/ I i B i Feasible(a) If agent i intends a, then it needs to believe a is feasible or at least have the intent to discover if a is feasible “If I intend to build a perpetual motion machine, then I have to believe it’s possible or to at least discover if it’s possible.”

03/06/2016CPSC /599.68: Agent Communications 25 Property: intent of act implies intent of RE  = I i Done(a)  I i RE(a) If agent i intends a, then it also intents the RE of a “If I intend to drop the vase, then I also intend to break the vase”

03/06/2016CPSC /599.68: Agent Communications 26 Property: observing an act  = B i (Done(a) /\ Agent(j,a)  I j RE(a)) If agent i observes j doing a, then i will come to believe the j intends the RE of a. “If I see Jane hammering the vase, then I believe that Jane intends to break the vase”

03/06/2016CPSC /599.68: Agent Communications 27 BDI Issues Sometimes end up with expressions like “I believe that you believe that I believe that you believe that…” Calls on agents to have a omniscient view of all the other agents FIPA is based on BDI semantics

03/06/2016CPSC /599.68: Agent Communications 28 Social Commitments Many utterances imply some sort of “conversational” social commitment Eg: a request commits the receiver to reply Other social commitments are negotiated Eg: “wash my car” Basic Agent body: When an agent observes (or sends or receives) a message, it uses policies (rules) as social norms that generate (or delete) social commitments Agents spend their free time trying to fulfill the social commitments for which it is a debtor

2005/07/27Commitment-based Conversations29 FIPA Performatives inform cancelrequest query-refcall-for-participation inform-refquery-ifrequest-when request-whenever propose agree subscribe confirm failurenot-understoodreject-proposalrefuse disconfirm accept-proposal proxypropagate

2005/07/27Commitment-based Conversations30 FIPA Performatives inform cancelrequest query-refcall-for-participation inform-refquery-ifrequest-when request-whenever propose agree subscribe confirm failurenot-understoodreject-proposalrefuse disconfirm accept-proposal proxypropagate performative ack affirmative-reply reply negative-reply reply-propose-discharge Added catagories

2005/07/27Commitment-based Conversations31 FIPA Performatives inform cancelrequest query-refcall-for-participation inform-refquery-ifrequest-when request-whenever propose agree subscribe confirm failurenot-understoodreject-proposalrefuse disconfirm accept-proposal proxypropagate Arranged in a lattice performative ack affirmative-reply reply negative-reply reply-propose-discharge Added catagories

2005/07/27Commitment-based Conversations32 Extended FIPA Performatives inform cancelrequest query-refcall-for-participation inform-refquery-ifrequest-when request-whenever propose notifyagree subscribe nack confirm timeoutfailurenot-understoodreject-proposalrefuse disconfirm accept-proposal proxypropagate propose-discharge done Arranged in a lattice performative ack affirmative-reply reply negative-reply reply-propose-discharge Added catagories

2005/07/27Commitment-based Conversations33 Policies P-informcommits the addressee to acknowledge P-ackreleases informed agents of the commitment to acknowledge P-requestcommits the proposed agents to reply P-counteroffercommits addressees to reply P-replyreleases proposed agents of the commitment to reply and releases counteroffered agents of the commitment to reply P-agreean acceptance realizes the shared uptake of proposed/counteroffered commitments P-donereleases accepted agents of the commitment earlier agreed to

2005/07/27Commitment-based Conversations34 Policies  Commitment Operators Performative Ack Reply Inform Request Agree Confirm Performatives Commitment-operator Operator: add Commitment: (receiver,sender,ack) Commitment-operator Operator: delete Commitment: (receiver,sender,reply) Commitment-operator Operator: delete Commitment: (sender,receiver,ack) Commitment-operator Operator: add Commitment: (reciever,sender,reply) Commitment-operator Operator: add Commitment: (receiver,sender,content) P-inform P-ack P-request P-reply P-agree Commitment-operator Operator: delete Commitment: (receiver,sender,content) P-confirm PoliciesCommitment Operators action reply ack communication-act Unspecified action Social Commitments

2005/07/27Commitment-based Conversations35 Example: Informally Alice Bob (performative: request, content: attend(Bob,x)) Can you attend this meeting? (performative: agree, content: request|attend(Bob,x)) Sure... (performative: propose, content: discharge|attend(Bob,x)) I’m here (performative: accept-proposal, content: discharge|attend(Bob,x)) Thanks for coming. (performative: ack, content: agree|request|attend(Bob,x)) (nod) (performative: ack, content: accept-proposal|discharge|attend(Bob,x)) (nod) (performative: ack, content: discharge|attend(Bob,x)) (nod)

2005/07/27Commitment-based Conversations36 Example: Perf. Lattice and Commitments reply-propose-discharge(Alice,Bob,x) act(Bob,Alice,x) propose-discharge(Bob,Alice,x) Alice Bob request inform reply agree informack propose/discharge reply informack agree accept-proposal reply informack reply(Bob,Alice,x) ack(Bob,Alice,x) ack ack(Bob,Alice,x) ack ack(Alice,Bob,x) ack ack(Alice,Bob,x) ack

2005/07/27Commitment-based Conversations37 Example: Implementation Details reply(Bob,Alice,x) act(Bob,Alice,x) reply-propose-discharge(Alice,Bob,x) propose-discharge(Bob,Alice,x) Alice Bob reply request/* agree/request|* propose/discharge|* agree accept-proposal/propose/discharge|* reply inform evaluate(Alice,Bob,x) decide(Bob,Alice,x)` consider(Alice,Bob,x) accept(Bob,Alice,x) inform ack ack(Bob,Alice,x) ack ack(Bob,Alice,x) ack ack(Alice,Bob,x) ack ack(Alice,Bob,x) ack

2005/07/27Commitment-based Conversations38 Conclusions Arranging performatives in a lattice simplifies interpretation Messages (performatives) → policies → commitment operators → shared social commitments Easily observable by 3 rd parties Agents do not have the be implemented in the SC style (eg. could be BDI internally) Turn taking arrises naturally