Multi-Agent SystemsChapter 2

Adapted (with permission) from Adina Magda Florea

adina@wpi.edu

2

Candy Quiz(not for points, just for candy)

• What do each of the following mean?• KQML

• ACL

• Ontology

• Illocution

• KSE

• FIPA

• KIF

3

Candy Quiz(not for points, just for candy)

• What do each of the following mean?• KQML Knowledge Query Manipulation Language

• ACL Agent Communication Language

• Ontology specification of objects, concepts and relationship

• Illocution intended meaning - performative

• KSE Knowledge Sharing Efort

• FIPA Foundation for Intelligent Physical Agents

• KIF Knowledge Interchange Format

1. The nature of communication

Human communication Communication is the intentional exchange of information brought about

by the production and perception of signs drawn from a shared system of conventional signs (AIMA, Russell&Norvig) language

Communication seen as an action (communicative act) and as an intentional stance (want something to happen)

Component steps of communicationSpeaker Hearer Intention Perception Generation- create Analysis Synthesis- organize Disambiguation

Incorporation

4

Syntax-how structuredSemantics-symbol meaningPragmatics-interpretation

Artificial Communication low-level language vs high-level languages direct communication vs. indirect communication

Computer communication shared memory message passing

Agent communication/ MAS communicationCommunication in MAS = more than simple communication,

implies interactionThe environment provides a computational infrastructure

where interactions among agents take place. The infrastructure includes protocols for agents to communicate and protocols for agents to interact

5

Communication protocols = enables agents to exchange and understand messages

Interaction protocols = enable agents to have conversations, i.e., structured exchanges of messages. Can trace series of responses.

Aim Communication enables agents to:coordinate their actions and behaviorattempt to change state of the other agentsattempt to get other agents perform actions

6

Communication infrastructure blackboard (shared memory) or message-based connected or connection-less (email) point-to-point (1), multicast (some), broadcast (all)

push or pull (information given to you or requested)

synchronous or asynchronous

8

Dimensions of Meaning(think of examples of each)

• Descriptive/Prescriptive: describe phenomenon vs proscribe behavior (I am cold)

• Personal/conventional: different meaning to individual

• Semantics/Pragmatics: interpreted differently than intended: mental states

• Contextuality: cannot be understood in isolation• Coverage: language express necessary concepts• Identity: meaning is dependent on individual• Cardinality: private message interpreted

differently than public message

2. Indirect communication2.1 Signal propagation - Manta, A. Drogoul 1993

An agent sends a signal, which is broadcast into the environment, and whose intensity decreases as the distance increases

At a point x, the signal may have one of the following intensities (why?)

V(x)=V(x0)/dist(x,x0) V(x)=V(x0)/dist(x,x0)2

9

x0

Agent A

(stimulus triggersbehavior P)

Agent B(stimulus triggersbehavior P)

Topological differenceslead to social differences

2.2 Trails - L. Steels, 1995agents drop "radioactive crumbs" making trailsan agent following a trail makes the trail faint until it disappears

Reactive agentsReactive agents

Stimulus

2.3 Blackboard systems, Barbara Hayes-Roth, 1985

Blackboard = a common area (shared memory) in which agents can exchange information, data, knowledge

Agents initiates communication by writing info on the blackboard Agents are looking for new info, they may filter it Agents must register with a central site for access authorization Blackboard = a powerful distributed knowledge paradigm Agents = Knowledge sources (KS)

10

Cognitive agentsCognitive agents

KS

KS

KS

KS

Control

Blackboard

events

Pending KS Activations

3. Direct communication

Sending messages = method invocationThe Actor language

– an Actor executes a sequence of actions in reply to the received message

Exchange of partial plans– distributed planning

ACL = Agent Communication Languages

communication as action - communicative acts11

3.1 Agent Communication Languages Concepts (distinguish ACLs from RPC, RMI or CORBA, ORB):

ACLs handle propositions, rules, and actions instead of objects with no associated semantics

An ACL message describes a desired state in a declarative language, rather than a procedure or method invocation

ACLs are mainly based on BDI theories: BDI agents attempt to communicate their BDI states or attempt to alter others BDI state

ACLs are based on Speech Act Theory Agent behavior and strategy drive

communication and lead to conversations

12

ACL

Content language

Ontology

Ontology =formal definition of a body of knowledge. The most typical ontology used in building agents involves a structural component. Essentially a taxonomy of class and subclass relations coupled with definitions and relationships between things. (Jim Hendler)

An ontology is analogous to a data base organization, not the contents of the database.

Example: To express the idea that a block is a physical object we might use the FOPL expression:

x (Block x) (PhysicalObject x)To state relationships between classes:x,y,z (instanceOf x y ) (subclassOf y z) (instanceOf x z)To define a relationship On(X,Y) (domain On PhysicalObject) (range On PhysicalObject)

14

Origins of ACLsOrigins of ACLs

Knowledge Sharing Effort - DARPA, 1990 External Interface Group - interaction between KBS - KQML Interlingua – language for communicating between independent

agents. KIF (not designed as a language to be used by humans) Shared, Reusable Knowledge Bases – Ontolingua (set of tools written in Lisp for analyzing and

translating ontologies. Uses KIF parser and syntax checker. Web-based)

Theory of Speech ActsJ. Austin - How to do things with words, 1962,

J. Searle - Speech acts, 1969Treats communication as an action – made to further agent’s intentionsChanges world in a way analogous to physical actionsNeed formal representation so planning systems can reason about

them.

A speech act has 3 aspects: locution = physical utterance by the speaker illocution = the intended meaning of the utterance by the speaker

(performative) prelocution = the action that results from the locution

Alice told Tom: "Would you please close the door“

15

16

Human Communication: Intent may be ambiguous

• Colleague states “I am cold”

• What is the meaning?

• For computers, state performative (KQML) or Communicative Act (FIPA) so no ambiguity (except in content itself).

17

Illocutionary aspect - several categories(According to speech act theory)

Assertives, which inform: the door is shut Directives, which request: shut the door, can pelicans fly? Commissives, which promise something: I will shut the door Permissive, which gives permission for an act: you may shut the

door Prohibitives, which ban some act: do not shut the door Declaratives, which causes events: I name you king of Cache

Valley Expressives, which express emotions and evaluations: I am happy

18

KQML – Knowledge Query and Manipulation LanguageA high-level, message-oriented communication language and protocol for information exchange, independent of content syntax (KIF, SQL, Prolog,…) and application ontology

KQML separates: semantics of the communication protocol (domain independent) semantics of the message (domain dependent)

3 (conceptual) layers

18

Content

Communication

Message

Describes low levelcommunication parameters:- identity of sender and receiver- a unique id associated with the communication

Core of KQML- identity of the network protocol with which to deliver the message- speech act or performative Optional- content language- ontology

19

Syntax LISP-like list of attribute/value pairs

(ask-one :sender joe :receiver ibm-stock :reply-with ibm-stock :language PROLOG // of content :ontology NYSE-TICKS // define terminology :content (price ibm ?price) )

19

(tell :sender willie:receiver joe:reply-with block1:language KIF:ontology BlockWorld:content (AND (Block A)(Block B)

(On A B)) )1. Query performatives: ask-one, ask-all (want all answers to question), ask-if, stream-all (multiple response version of ask-all)

(stream-all :sender willie:receiver ibm-stock:content (price ?VL ?price ) )

(standby:content (stream-all:content (price ?VL ?price) )

ask-one(P)

tell(P)

tell(P1)

stream-all(P)

tell(P2)eos

tell(P1,P2,...)

ask-all(P)

B

B

B

A

A

A

20

4. Generic informational performatives:

tell (I know it), untell, insert (add to your VKB), delete, ...

tell(P)

untell(P)

delete(P)

insert(P)

2. Generator performatives:standby (be ready to respond), ready, next, rest, discard, generate,...

6. Network performatives:register, unregister, forward, route, ...

3. Response performatives:reply, sorry ...

5. Capability performatives:advertise, subscribe, recommend...

Facilitator

A

A

B

B

21

Knowledge Sharing Effort KSE

• Darpa funded in 1990

• Resulted in– KQML– KIF (knowledge interchange format) looks like

first order logic (and, or, forall, exists) cast into Lisp-like notation

• Used for content part of KQML message

22

Example of KQML Dialog(evaluate :sender A :receiver B :language

KIF :ontology motors

:reply-with q1

:content (val (torque m1))

(reply

:sender B :receiver A :language KIF

:ontology motors :in-reply-to q1

:content (= (torque m1) (scalar 12 kgf))

23

Another example of KQML

(stream-about // multiple responses wanted

:sender A :receiver B :language KIF :ontology motors

:reply-with q1 :content m1)

(tell :sender B :receiver A :in-reply-to q1

:content (=(torque m1) (scalar 12 kgf))

(tell :sender B :receiver A :in=reply-to q1

:content (= (status m1) normal))

(eos :sender B :receiver B :in-reply-to q1) //end of stream

24

In the next KQML example

• A advertises that it is willing to accept subscriptions related to m1.

• B responds by subscribing

• A sends message sequence about m1

• A untells the previous fact and gives new information

• eos ends the stream of messages

25

(advertise :sender A :language KQML :ontology K10:content (stream-about :language KIF :ontology motors :content m1)))(subscribe :sender B :receiver A :reply-with s1:content (stream-about :language KIF :ontology motors : content m1))(tell :sender A :receiver B :in-reply-to s1 :content (=(torque m1)(scalar 12 kgf)))(tell :sender A :receiver B :in-reply-to s1 :content (=(status m1) normal))(untell :sender A :receiver B :in-reply-to s1 :content (=(torque m1)(scalar 12 kgf)))(tell :sender A :receiver B :in-reply-to s1 :content (=(torque m1)(scalar 15 kgf)))(eos :sender A :receiver B :in-reply-to s1)

26

Facilitator agent= an agent that performs various useful communication services:maintaining a registry of service names (Agent Name Server)forwarding messages to named servicesrouting messages based on contentmatchmaking between information providers and clientsproviding mediation and translation services

subscribe(ask(P)) tell(P)

tell(P)

advertise(ask(P))

tell(P)

recruit(ask(P))

ask(P)

recommend(ask(P)) advertise(ask(P))

tell(P)

ask(P)

reply(B)

A A

A

B

B

B

A wants to get all suitableagents to respond to ask(P).B has told F it can do ask(P). F tells B to tell A

FF

F

F

27

Semantics of KQML (Labrou & Finin)Use preconditions and post conditions that govern the use of a performative + the final state for the successful performance of the performativeUses propositional attitudes: belief Bel(A,P), knowledge Know(A,S), desire Want(A,S), intentions Int(A,S)Preconditions: the necessary states for an agent to send a performative and for the receiver to accept it and successfully process it; if the preconditions do not hold, the most likely response is error or sorryPostconditions - describe the state of the sender after successful utterance of a performative and of the receiver after the receipt and processing of a messageCompletion condition - the final state after a conversation has taken place and that the intention (associated with the performative that started the conversation) has been fulfilled

28

tell(A,B,X)A states to B that A believes the content X to be true, Bel(A,X)

Pre(A): Bel(A,X) Know(A, Want(B, Know(B,S))) no unsolicited infoPre(B): Int(B, Know(B,S))

where S may be any of Bel(B,X) or Bel(B,X)Post(A): Know(A, Know(B, Bel(A,X))) Post(B): Know(B, Bel(A,X))Completion: Know(B, Bel(A,X))

advertise(A,B,M)A states to B that A can and will process the message M from B, if it receives one

commisive actPre(A): Int(Proc(A,M))Pre(B): NONEPost(A): Know(A, Know(B, Int(A, Proc(A,M)))Post(B): Know(B, Int(A, Proc(A,M)))Completion: Know(B, Int(A, Proc(A,M)

29

Criticisms of KQML

• not fixed performatives – different implementations don’t interoperate

• transport mechanisms (ways of getting a message from A to B) not precisely defined

• semantics not rigorous – not adhered to• missing commissives (make commitment)• performative set too large and ad hocResulted in FIPA

30

3.3 FIPA - ACLFoundation for Intelligent Physical Agents, 1996

Goal of FIPA = make available specifications that maximize interoperability across agent-based systems. SL formal languageSyntax similar to KQML(inform :sender Agent1

:receiver Agent2 :content (price good2150)

:in-reply-to round-1: reply-with bid03: language S1:ontology hp-auction:reply-by 10:protocol offer:conversation-id conv-1 )

FIPA communicative actsInformatives - query_if, subscribe, inform, inform_ifconfirm, disconfirm, not_understoodTask distribution - request, request_whenever, cancel, agree,refuse, failureNegotiation - cfp (call for proposal - initiate), propose, accept_proposal, reject_proposal

31

FIPA - SemanticsSL (Semantic Language) - a quantified, multi-modal logic, with modal operators:B - beliefD - desireU - uncertain belief – neither believes or its negatation, but believes more likely than its negation.PG - persistent goal, but will not necessarily plan to bring it aboutBif - agent believes or disbelieves Uif - agent has an uncertain belief of or an uncertain belief of ¬

Given that one of the most frequent and serious criticisms of KQML is the lack of adequate semantics, it is not surprising that the developers of FIPA felt it important to give comprehensive formal semantics to their language.

32

The semantics of a communicative act is specified as a set of SL's formulae that describe:

Feasibility preconditions - the necessary conditions for the sender - the sender is not obliged to perform the CA

Rational effect - the effect that an agent can expect as a result of performing the action; it also typically specifies conditions that should hold true of the recipient

The receiving agent is not required to ensure that the expected effect comes about

The sender can not assume that the rational effect will necessary follow

Semantics of inform:

<i, inform(j, )>

Pre: Bi Bi (Bifj Uifj )

Post: Bj

33

Syntax of request:

<k, request(j, )>

Pre: Bk Agent( ,j) BkIj(Done )

Post: Done()

Agent( ,j) means that the agent of action is j (j is the one who can perform the action)

Done( ) means that the action has been done

Effect: agent k is requesting agent j to do action and agent k believes that the agent to do the action is j and it believes that j does not currently intend to do the action. The effect is that j will do it.

34

Using ACLs in MASAny MAS that is to use an ACL must provide:a finite set of APIs for composition, sending, and receiving ACL messagesan infrastructure of services that assist agents in naming, registration, and basic facilitation services (finding other agents that can do things for your agent)code for every reserved message type that takes the action prescribed by the semantics for the particular application;the code depends on the application language, the domain, and the details of the agent system using the ACL

35

History of DAML (one possible content language)

HTML defines grammar for interspersing documents with markup commands.

HTML is fixed – can’t introduce tags

Suppose you want to sell CDs and put price information on the web.

HTML can markup with layout information, but browsers won’t know how to interpret content. What is price? What is title?

XML was developed to provide semantic markup – helps computers to process information in document by allowing user defined tags.

DAML (Darpa markup language) is based on XML

36

4. Communication content

Content languages– KIF– Prolog– Clips– SQL– FIPA-SL, FIPA-CCL, FIPA-KIF– DAML– XML Xtensible Markup Language

Ontologies

ACL

Content language

Ontology

DARPA Agent Markup LanguageThe DAML Program officially began in August 2000.The goal of the DAML effort is to develop a languageand tools to facilitate the concept of the Semantic Web.

Interaction protocols

enable agents to have conversations, i.e., structured exchanges of messages

Finite automataConversations in KQMLPetri nets - more in distributed planning lecture

FIPA IP standards:– FIPA-query, FIPA-request, FIPA-contract-net,

Finite state automataNodes represent states

Arrows represent actions

38

A:B<<ask(do P)

B:A<<accept(do P)

B:A<<refuse(do P)

B:A<<result(do P)B:A<<fail(do P)

proposeS(P)

acceptR(P) rejectR(P)

counterR(P)counterS(P)

acceptS(P) rejectS(P)Winograd, Flores, 1986

COOL, Barbuceanu,95

39

Conversations in KQMLUse Definite Clause Grammars (DCG) formalism for the specification of conversation policies for KQML performativesDCGs extend Context Free Grammars in the following way:non-terminals may be compound termsthe body of the rule may contain procedural attachments, written as "{" and "}" that express extra conditions that must be satisfied for the rule to be valid

Ex: noun(N) [W], {RootForm(W,N), is_noun(N)}

S s(Conv, P, S, R, inR, Rw, IO, Content), {member(P, [advertise, ask-if]}

s(Conv, ask-if, S, R, inR, Rw, IO, Content) [ask-if, S, R, inR, Rw, IO, Content] |[ask-if, S, R, inR, Rw, IO, Content], {OI is inv(IO)},

r(Conv, ask-if, S, R, _, Rw, OI, Content)

r(Conv, ask-if, R, S, _, inR, IO, Content) [tell, S, R, inR, Rw, IO, Content] |problem(Conv, R, S, inR, _, IO)

Labrou, Finin, 1998

40

DAA wants to do P,A cannot do P

Request do(P)

Refuse do(P)

Accept/request do(P)

Fail to do(P)

Notification of end(P)

DB

AR1

AR2

FA2FA1

BR

FBFailure Satisfaction

Impossibleto do(P)

B does not wantto do(P)

B is willingto do(P)

Completed(P)

Petri netsPetri net = oriented graph with 2 type of nodes:places and transitions;there are moving tokens through the net - representation of dynamic aspect of processes. Stores state.Tokens are moved from place to place, following firing rules.A transition T is enabled if all the input places P of T contains a token A marking (state) is a distribution of tokens over places.

Ferber, 1997

40

41

ReferencesM. Huhns, L. Stephens. Multiagent systems and societies of agents. In Multiagent Systems - A Modern Approach to Distributed Artificial Intelligence, G. Weiss (Ed.), The MIT Press, 2001, p.79-120.M. Wooldrige. Reasoning about Rational Agents. The MIT Press, 2000, Chapter 7Y. Labrou, T. Finin. Semantics and conversations for an agent communication language. In Readings in Agents, M. Huhns & M. Singh (Eds.), Morgan Kaufmann, 1998, p.235-242.J. Ferber - Multi-Agent Systems. Addison-Wesley, 1999, Chapter 6T. Finnin, R. Fritzson - KQML as an agent communication language. In Proc. of the Third International Conference on Information and Knowledge Management (CIKM'94), ACM Press, 1994.M. Singh. Agent communication languages: Rethinking the principles. IEEE Computer, Dec. 1998, p.40-47.Y. Labrou, T. Finnin, Y. Peng. Agent communication languages: The current Landscape. IEEE Computer, March/April 1999, p. 45-52.FIPA97. "Agent Communication Language" Specification FIPA, 11/28/97

DARPA KSE http://www-ksl.stanford.edu/knowledge-sharing/KQML http://www.cs.umbc.edu/kqml/KIF http://logic.stanford.edu/kif/Ontolingua http://www-ksl-svc.stanford.edu:5915/&service=frame-editorFIPA http://www.fipa.org/DAML http://www.daml.org/