EEL 5937 Agent communication EEL 5937 Multi Agent Systems Lotzi Bölöni

EEL 5937 Speech acts Most treatments of communication in multi-agent systems borrow their inspiration from speech act theory Speech act theories are pragmatic theories of language; they attempt to account for how language is used by people every day to achieve their goals and intentions. The origins of speech act theories are usually traced to J.L. Austin’s book “How to do things with words”. –This is a philosophy book with a linguistic approach –Not a computer science book!

EEL 5937 Speech acts (cont’d) Austin noticed that some utterances are rather like “physical actions”, that appear to change the state of the world Paradigm examples would be: –Declaring war –“I now pronounce you man and wife” But more generally, everything we utter is uttered with the intention of satisfying some goal or intention A theory of how utterances are used to achieve intentions is a speech act theory.

EEL 5937 Speech acts (cont’d) Searle (1969) identified various types of speech acts: Representatives –Such as informing: “It is raining” Directives –Attempts to get the hearer to do something, e.g. “please make the tea” Commisives –Which commit the speaker to doing something, e.g. “I promise to” Expressives: –Whereby a speaker expresses a mental state, e.g. “Thank you!” Declarations: –Such as declaring a war

EEL 5937 Speech acts (cont’d) There is some debate about whether this (or any!) typology of speech acts is appropriate In general, a speech act can be seen to have two components: –A performative verb (e.g. request, inform) –A propositional content (e.g. “the door is closed”)

EEL 5937 Relationship between the performative and content Performative = Request –Content = “The door is closed” –Speech act = “please close the door” Performative = Inform –Content = “The door is closed” –Speech act = “The door is closed!” Performative = Inquire –Content = “The door is closed” –Speech act = “Is the door closed?”

EEL 5937 Semantics of speech acts How can one define the semantics of a speech act? What is going to be the affect of the speech act to the world? –More exactly to the receiver? The sender agent can not (generally) force a receiver agent to accept some desired mental state. Different formalisms were proposed, depending on the representation of the world.

EEL 5937 Plan based semantics Cohen & Perrault (1979) defined the semantics of speech acts using the precondition-delete-add list formalism of planning research. Semantics for a request: request(s, r, a) Preconditions: –S believes r can do h »You don’t ask someone to do something unless you think they can do it –S believes h believes h can do a »You don’t ask someone unless they believe they can do it –S believes S wants a »You don’t ask someone unless you want it Postconditions: –H believes s believes s wants a »The effect is to make them aware of your desire.

EEL 5937 Agent communication languages We now consider the agent communication languages (ACL’s), standard formats for the exchange of messages. KQML FIPA-ACL

EEL 5937 KQML Developed by the ARPA knowledge sharing initiative Composed of two parts: Knowledge Query and Manipulation Language (KQML) –KQML is an “outer language”, that defines various acceptable “communicative acts” or performatives –Many critics say there were too many performatives (>40) Knowledge Interchange Format (KIF) –The content language usually used by KQML

EEL 5937 FIPA ACL, introduction FIPA (Foundation for Intelligent Physical Agents) non-profit organization –Companies like IBM, Hitachi, HP, British Telecom, Siemens –Universities and research institutes Started work on a program of agent standards, the most important being the ACL (1997)

EEL 5937 Basic structure of a FIPA message Performative –There are 20 different performatives in FIPA-ACL Housekeeping –Sender –Receiver –Reply-with –In-reply-to Content –The actual content of the message Language –The language in which the content is written Ontology –The ontology in which the message needs to be interpreted.

EEL 5937 Example (inform :sender agent1 :receiver agent2 :content (price milk 100) :language sl :ontology hlp-auction )

EEL 5937 Performatives in FIPA

EEL 5937 Inform and Request Inform and Request are the two basic performatives in FIPA ACL. All others are macro definitions, defined in terms of these. The meanings of inform and request are defined in two parts: –Pre-condition: what must be true in order for the speech act to succeed –Rational effect: what the sender hopes that the message will bring about.

EEL 5937 Inform The content is a statement Pre-conditions: –The sender holds that the content is true –Intends that the recipient believes the content –Does not already believe that the recipient is aware of whether the content is true or not.

EEL 5937 Request The content is an action. Pre conditions: –The sender intends the action content to be performed –Believes the recipient is capable of performing the action –Does not believe that the sender already intends to perform the action.

EEL 5937 FIPA ACL, Semantic Language, SL Used to define the semantics of FIPA ACL Quantified, multi-modal logic with operators for beliefs, desires, uncertain beliefs and intentions Can represent propositions, objects and actions

EEL 5937 Interaction protocols A standard set of messages needed to achieve a certain goal They are usually describing a time limited “conversation”

EEL 5937 Protocols standardized by FIPA FIPA Request FIPA Query FIPA Request When FIPA Contract Net FIPA Iterated Contract Net FIPA English Auction FIPA Dutch Auction FIPA Brokering FIPA Recruiting FIPA Subscribe FIPA Propose Interaction

EEL 5937

Missing from the specification Comprehensive treatment of the error conditions. What is happening in case of a timeout? In general, the time dimension of the messaging –How much do you wait before declaring an auction over? –How do you answer for offers coming in two late?

EEL 5937 Internal state of agents during a conversation Interaction protocols specify only the messages, not the state of the agents. The state of the agents should contain the agents perspective about a conversation (not necessarily correct one). The agent should know whether an incoming message is legal, and what does it mean. What information do you keep about a conversation? What data structure you use to keep the conversation state? –Finite state machine –Petri Net –Rules

EEL 5937 Managing multiple conversations When do you create a conversation? –Legal entry points How do you match messages to a conversation? –Hash table, or similar data structure When do you discard a conversation? –Legal termination points –Timeout (but you might keep the data around after the timeout as well). What about embedded or interdependent conversations? –Timeout calculation for embedded conversations