A Geometric Semantics for Agent Interaction Protocols

1
A Geometric Semantics forAgent Interaction
Protocols

• Peter McBurney
• Department of Computer Science
• University of Liverpool
• Liverpool L69 7ZF
• p.j.mcburney_at_csc.liv.ac.uk
• (Joint work with Simon Parsons, Brooklyn College,
  CUNY, New York.)
• Presentation to
• Condensed Matter Physics Group
• Imperial College, London
• 29 October 2003

2
We are on the verge of a revolution . . .

• Computational devices and systems will soon be
• Everywhere
• Interconnected
• Always active
• Intelligent and autonomous.
• Software systems will thus be
• Situated
• Responsive to and influential upon their
  environment
• Open
• Computational entities will enter and leave these
  environments continually
• Autonomous
• Entities and systems will be goal-directed and
  exhibit autonomous behaviour
• Systems and sub-systems will have multiple
  threads of control, not one.

3
Autonomous intelligent software agents

• It helps to conceive of computer systems as
  consisting of interacting autonomous entities.
• A software agent is a computational entity with
  (some degree of)
• Social awareness
• Proactive behaviour towards defined goals
• Reactive behaviour in response to its environment
  
• Decision-making autonomy.
• (Wooldridge Jennings 1995)
• Some applications
• Air Traffic Control systems (agents representing
  aircraft and controllers)
• Electronic commerce (agents representing buyers,
  sellers, others)
• Management of utility networks (telecoms,
  electricity, etc)
• Provisioning of complex products and services
  (e.g. telecoms services)
• Management of fleets (vehicles, satellites, SCADA
  devices, etc).

4
Two key research problems

• How to design agents
• The most common approach is based on the
  Philosophy of Intention and Rational Agency
  (Bratman, Pollock)
• e.g In the BDI model, agents are assumed have
  three types of mental states Beliefs, Desires,
  and Intentions.
• Considerable work has focused on formalizing
  these models using dialects of modal logic
  (epistemic, temporal, deontic, etc) or formalisms
  adopted from argumentation theory.
• How to design Multi-Agent Systems (MAS)
• How may agents interact with one another?
• How may they make joint decisions?
• I will consider agent interaction languages in
  this talk.

5
How to humans interact?

• By means of language
• So, an obvious first step to designing agent
  interaction mechanisms is to consider the design
  of artificial languages for agent interaction.
• Types of Agent Communications Languages
• Generic ACLs
• Dialogue Game Protocols
• Auction Mechanisms.
• Following the philosophy of language, agent
  languages designers usually distinguish between
  two layers of communicated messages
• The topics of conversation (which may be
  represented in a suitable logical language)
• eg It is raining
• The illocutions which communicate something about
  these topics, eg
• QUESTION(raining)
• INFORM(raining)
• DEMAND(raining).

6
Generic ACLs

• Two major proposals
• USA DARPAs Knowledge Query and Manipulation
  Language (KQML)
• Arose from attempts to merge multiple knowledge
  bases
• Focus was information-sharing between
  knowledgeable agents.
• www.cs.umbc.edu/kqml/
• Foundation for Intelligent Physical Agents ACL
  (FIPA ACL)
• Arose from an automated purchase transaction
  system at France Telecom
• Focus was negotiation of tasks between expert
  agents
• FIPA is a computer industry standards body for
  agent technologies.
• www.fipa.org

7
FIPA Agent Communications Language (FIPA ACL)

• FIPA ACL has 22 illocutions
• e.g. inform, query-if, request, agree, refuse.
• Each has a defined syntax
• (inform
• sender (agent-identifiername j)
• receiver (agent-identifiername i)
• content
• weather (today, raining)
• language Prolog)
• The origins of FIPA ACL in knowledge-sharing and
  contract negotiations are apparent
• 11 of the 22 illocutions concern requests for or
  transmissions of information
• 4 involve negotiation (e.g. cfp, propose,
  reject-proposal)
• 6 involve performance of action (e.g. refuse,
  request)
• 2 involve error-handling of messages (e.g.
  failure).

8
Problems with FIPA ACL

• The language implicitly assumes eternal
  connections between the agents
• Where are the illocutions for entering and
  leaving dialogues?
• Where are the illocutions for permitting or
  contesting participation?
• As befits a language for knowledge-sharing, the
  semantics impose sincerity
• Agents cannot utter beliefs they do not hold.
• As befits a language for contract negotiations,
  the underlying (implicit) argumentation theory is
  simplistic.
• There are no illocutions for contesting
  statements, or for requesting or giving reasons
  for claims, or for structuring dialogue.
• The participants incur no dialectical
  obligations.
• The language does not readily support
  self-transformation
• How may an agent express a change of its beliefs?
  
• The absence of an explicit argumentation theory
  causes a state-space explosion
• Any illocution may follow any other Disruptive
  behavior is not precluded.
• Dialogue Game Protocols have been proposed as a
  solution to this problem.

9
Dialogue Game Protocols

• Games between two or more participants where
  each moves by making utterances, subject to
  some rules.
• Origins in Philosophy
• Aristotle and medieval philosophers
• Revived for the study of supposedly fallacious
  reasoning (Hamblin 1970, MacKenzie 1979)
• Proof theory for intuitionistic classical logic
  (Lorenzen 1959)
• Applied to quantum physics (Mittelstaedt 1979).
• Within computer science, applied to
• Modeling human dialogues in computational
  linguistics
• Software development processes
• Modeling legal reasoning
• Man-machine dialogues (e.g. for automated
  tutoring systems)
• Protocols for agent dialogues.

10
A DG Protocol is defined in terms of

• A language of statements (the topics of the
  dialogue)
• Usually expressed in some logical language (e.g.
  propositional logic, FOL, etc).
• A set of illocutions instantiated with the
  statements
• eg assert(p), accept(p), contest(p).
• Combination rules, defining the circumstances in
  which each instantiated illocution may be uttered
• eg It may not be possible to assert a statement
  and then its negation.
• Termination Rules, defining the circumstances in
  which dialogues terminate.
• Rules for creating and combining commitments
• Commitment Stores publicly-accessible sets of
  statements, holding the commitments incurred by
  participants.
• Dialogic and external (semantic) commitments, and
  rules for their combination.

11
An influential typology of dialogues

• Doug Walton and Erik Krabbe (1995) have proposed
  a typology of human dialogues, based on the
  information known to participants at
  commencement their respective objectives and
  the purpose(s) of the dialogue.
• Information-seeking dialogues
• One participant seeks the answer to a question
  which it believes another knows.
• Inquiry dialogues
• All participants collaborate to find the answer
  to a question which no one knows.
• Persuasion dialogues
• One participant seeks to persuade other(s) to
  endorse a statement.
• Negotiation dialogues
• Participants seek to divide a scarce resource.
• Deliberation dialogues
• Participants collaborate to decide a course of
  action in some situation.
• Eristic dialogues
• Participants quarrel to vent perceived
  grievances, as a substitute for physical fighting.

12
Formal Dialogue-Game Protocols

• Agent interaction protocols have been designed
  for
• Inquiry dialogues (McBurney Parsons 2001)
• Persuasion dialogues (Dignum, Dunin-Keplicz
  Verbrugge 2000)
• Negotiation dialogues (Amgoud, Parsons Maudet
  2000 Sadri, Toni Torroni 2001 McBurney, van
  Eijk, Parsons Amgoud 2003)
• Deliberation dialogues (Hitchcock, McBurney
  Parsons 2001).
• These protocols are more constrained than are
  generic Agent Communications Languages
• Rules govern combinations of locutions agents
  usually cannot say just anything at anytime.
• Usually, the protocol is designed with a specific
  purpose in mind, and informed by an explicit
  theory of argument.

13
Example locutions in a Dialogue Game Protocol

• Locutions for a deliberation dialogue (to jointly
  decide a course of action)
• open_dialogue(Pi, q?)
• enter_dialogue(Pj, q?)
• propose(Pi, type, t)
• assert(Pi, type, t)
• prefer(Pi, a, b)
• ask_justify(Pj, Pi, type, t)
• move(Pi, action, a)
• retract(Pi, locution)
• withdraw_dialogue(Pi,q?)
• where
• Pi, Pj are participating agents
• type ? question, goal, constraint, perspective,
  fact, action, evaluation
• and there are various constraints on, and impacts
  of, utterance of these locutions.
• (Hitchcock, McBurney Parsons 2001)

14
Example (continued)

• For this protocol, the purpose is joint practical
  reasoning
• For a group of participants to jointly decide on
  an action, or course of action, in some situation
• Or, at least, to decide if they have a joint
  responsibility for such a decision.
• The theory of argument made explicit was Harald
  Wohlrapps retroflexive argumentation model
  (1998)
• Here, proposed actions and suggested
  justifications are both modified iteratively, in
  the light of reflections on each.
• For example
• The law should allow euthanasia, since this would
  permit people in terminal pain to die.
• But such a law could be abused by (say) evil
  doctors or relatives.
• Thus the law should allow euthanasia only under
  some conditions, for example, that two
  independent doctors agree.
• Etc.

15
Auction mechanisms

• The simplest communications protocols are the
  mechanisms of commerce
• Auction mechanisms
• Mechanisms for negotiations
• Cake-cutting algorithms, etc.
• Called Game-Theoretic Mechanisms in AI.
• At the simplest, these involve illocutions for
• Proposing a deal (a division of some scarce
  resource)
• Accepting or rejecting a proposed deal
• (And possibly also) Entering and leaving the
  interaction.
• Because of the rise of e-commerce, these
  mechanisms have been much studied within Computer
  Science/AI of late.
• See Agent-Mediated e-Commerce Workshop series
  (Springer).

16
Examples of GT protocols

• Auction Mechanisms
• English (ascending) auctions
• Dutch (descending) auctions
• Vickrey (second-price) auctions.
• Combinatorial auctions
• Bidders may bid on any combination of a set of
  items.
• Continuous Double Auctions (k-CDA)
• Multiple buyers and sellers make bids and asks
  (respectively)
• Transaction price is a function (with parameter
  k) of bid and ask prices
• Used in most organized stock and commodity
  exchanges.
• Monotonic Concession Protocol
• 2 participants
• Participants may propose (make an offer),
  counter-propose, accept a proposal, or withdraw.
• Proposals must always concede, relative to
  previous proposals.

17
Relationship between types of interaction
protocols
Generic ACLs Dialogue Game Theoretic
Game (Auction) Protocols
Mechanisms
Increasing constraints on utterances
Increasing expressiveness
18
Key Research Challenges

• Defining the philosophies underlying agent
  societies
• e.g. Argumentation theories philosophies of
  democracy etc.
• Automation of Inquiry, Deliberation and Command
  dialogues
• We have defined protocols for the conduct of
  these dialogues.
• Key challenge How are possible
  hypotheses/action-options generated?
• Developing a formal, mathematical theory of
  interaction protocols
• To understand the space of protocols in its
  entirety, and to understand the relationship
  between two or more protocols.
• Currently under development
• Johnson, McBurney Parsons
• Drawing on Category Theory and Algebraic
  Topology.
• Understanding the relationships between local and
  global properties
• How to achieve dialogue-level properties (e.g.
  fast termination) using only local levers (e.g.
  locution-combination rules)?

19
Semantics for ACLs

• Linguistic theory distinguishes between
• Syntax of a language its words, phrases,
  sentences and grammar
• Semantics of a language what meanings are
  assigned to the words, phrases sentences
• Pragmatics of a language how the words, phrases
  and sentences and are used in conversation.
• Within mathematical logic, the Wittgenstein-Tarski
  an view of semantics is as a mapping from the
  legal formulae or sentences of a logical language
  to truth-values.
• Truth Values may be viewed as mathematical
  objects, eg 0,1.
• Model Theory studies the objects which are
  semantics for logical languages and their
  relationships to one another, as abstract
  mathematical objects.
• In Theoretical Computer Science, there are
  several types of semantics
• Axiomatic
• Operational
• Denotational
• Game-Theoretic.

20
Semantics of ACLs

• Considerable work on defining semantics of
  individual utterances
• Less work on semantics of dialogues under a given
  protocol
• No work yet on semantics of protocols
• My work is intended to develop a formal semantics
  of protocols
• To be able to determine if two protocols are the
  same or not
• To understand the relationship between syntactic
  form of a protocol and the properties of the
  dialogues conducted under it.
• This relationship is not continuous.

21
Axiomatic Semantics

• An axiomatic semantics articulates the
  pre-conditions and post-conditions of an
  utterance
• The semantics define the pre-conditions required
  for an utterance to be validly made, and the
  post-conditions which occur upon its utterance.
• This is usually done in a formal logical
  language, such as First-Order Logic.
• FIPA ACL has been given a formal, axiomatic
  semantics using speech act theory from the
  philosophy of language.
• Speech acts are utterances which are intended to
  change the world in some way.
• I name this ship, The Queen Elizabeth.
• I declare you man and wife.
• Austin 1955, Searle 1969.
• The speech act semantics for FIPA ACL links
  utterances to the private mental states of the
  participants.
• Their Beliefs, Uncertain Beliefs, and Intentions.
  
• This semantics has been formalized using modal
  epistemic logic.
• Bretier, Cohen, Levesque, Perrault, Sadek (1979,
  1990, 1997).

22
For example inform

• Suppose agent A informs agent B that It is
  raining.
• Required Pre-conditions Before a valid
  utterance by A
• A must believe It is raining,
• A must not already believe that B has any belief
  regarding whether or not it is raining (i.e. A
  must believe that B has an uncertain belief about
  this matter)
• and
• A must desire that B also comes to believe It is
  raining.
• Post-conditions Upon receipt by B of such an
  utterance by A
• B must believe that A believes It is raining
• and
• B must believe that A desires that B believes It
  is raining.
• Note that following the utterance by A, B may or
  may not adopt the belief It is raining.

23
Operational Semantics

• An operational semantics treats the utterances in
  an agent interaction as programming commands on
  some large, virtual machine
• The commands acts to change the state of this
  virtual machine.
• We can therefore view the utterances as functions
  which cause state transitions.
• From a formal axiomatic semantics we can define
  an operational semantics, which indicates the
  state transitions for every possible utterance.
• Does the virtual machine include the mental
  states of the interacting agents?
• An operational semantics has been defined for a
  dialogue game protocol for consumer purchase
  negotiations.
• McBurney, van Eijk, Parsons Amgoud 2003.

Utterance
Prior state of machine
Subsequent state of machine
24
Game Semantics

• To each formulae in a language is associated a
  game
• Usually between 2 imaginary players Proponent
  Opponent
• A formula is considered to be true iff a
  designated player (usually Proponent) has a
  winning strategy in the associated game.
• Example Ehrenfreucht-Fraisse games
• To assess whether two collections of objects are
  isomorphic, allow each player to select objects
  in turn.
• One player seeks to show the objects selected are
  in 11 relationship, the other player that this
  is not so.
• Used in model theory, and also recently in
  theoretical computer science to give a semantics
  for some programming languages.

25
Denotational Semantics

• Each formulae is mapped to some object in a
  mathematical space
• E.g. Mapping logical formulae to the set True,
  False or 0,1.
• The standard semantics for modal logic languages
  is the Possible Worlds semantics
• Due to Leibniz, Kanger (1957), Kripke
  (1959/1962), Hintikka (1962) (and Everett 1957)
• This is a collection of states of the world, at
  each of which some propositions are true and some
  not.
• Some worlds are connected by accessibility
  relationships, indicating (for example) that it
  is possible to move from one world-state to
  another.

26
Negotiation and Deliberation

• Deliberation Dialogues are dialogues over
  possible actions (or courses of action)
• Negotiation dialogues are a special case of
  Deliberations, where the actions are intended to
  divide some scarce resource.
• Deliberations typically involve one or more
  participants making proposals for action, which
  all parties then consider.
• We assume that the interaction protocol enables
  participants to
• Suggest proposals for action
• Accept or reject proposals which have been
  suggested
• Express a preference between two suggested
  proposals
• Commit to execute a specific proposal.
• We also assume that time is represented by a set
  common to all participants which is countable,
  and that exactly one utterance occurs at each
  time-point.

27
A category-theoretic semantics

• At each time point t
• We specify a proto-category representing the
  public utterances in the dialogue up to that time
• Called the Dialogue (or Public) Store
• Objects Proposed actions
• Arrows Expressed preferences between actions.
• We specify a proto-category for each participant
• Called the Private Store of the Participant
• Objects Possible actions under consideration by
  the Participant
• Arrows Determined preferences between actions
• One distinguished object ND (No Deal),
  representing termination of the deliberation
  without an agreement on an action being reached.
• For these entities to be categories, the
  participants preferences must be transitive.

28

Private Store Participant 1
Private Store Participant 2
Dialogue (Public) Store
Time t gt 8
29
Current Work

• Formalize this semantics, and study the
  mathematical properties of these structures.
• Not much work in CT on linked sequences of
  categories.
• Represent common deliberation and negotiation
  protocols in this way.
• Identify categorical constructs analogous to
  decision-mechanisms in deliberations and
  negotiations
• Decisions internal to the participants
• Judgment aggregation decisions in the dialogue
  (eg voting).

30
Further reading

• Agent-Enabled Computing
• M. Luck, P. McBurney and C. Preist (2003) Agent
  Technology Enabling Next Generation Computing.
  AgentLink II Network of Excellence.
• Available from www.agentlink.org
• M. J. Wooldridge (2002) Introduction to
  Multi-Agent Systems (Wiley)
• M. J. Wooldridge (2000) Reasoning About
  Rational Agents (MIT Press).
• Game-theoretic Interaction Mechanisms
• J. S. Rosenschein G. Zlotkin (1994) Rules of
  Encounter (MIT Press)
• S. Kraus (2001) Strategic Negotiation in
  Multiagent Environments (MIT Press).
• Agent Communications Languages and Dialogue Game
  Protocols
• www.fipa.org
• www.cs.umbc.edu/kqml/
• M-P. Huget (Editor) (2003) Communication in
  Multi-Agent Systems Agent Communication
  Languages and Conversation Policies. (Springer,
  LNAI 2650).
• F. Dignum (Editor) (2003) Advances in Agent
  Communication. (Springer LNAI 2922) (forthcoming).

31
Finally . . .

• Thank you for inviting me and for listening!

32
Combining dialogues of different types

• Most real human dialogues are complex
  combinations of primary types
• e.g. Analysis of environmental risk of new
  technologies involves combinations of
  Information-seeking, Information-Provision,
  Inquiry, Persuasion, Negotiation, Deliberation,
  Command, and even Eristic dialogues.
• There are two proposals for formalisms to
  represent combinations of agent dialogues
• Reeds Dialogue Frames (1998) can represent
  iterated, sequential embedded dialogues.
• This formalism is neutral regarding the syntax
  used in each dialogue.
• McBurney Parsons ADF (2002) can represent
  iterated, sequential, parallel embedded
  dialogues.
• This formalism is a dialect of Dynamic Modal
  Logic, and is potentially generative, i.e. it can
  be used generate many types of dialogues
  automatically.
• Both formalisms permit the incorporation of new
  primary types of dialogues.

33
Semantic Verification

• Problem How to verify that an agent using an
  ACL conforms to the (private) semantics of that
  ACL?
• i.e. How to verify that an agent really believes
  (or prefers or intends) what it says it does?
• Proposed Partial Solutions
• Social Semantics (Singh)
• Have agents profess their beliefs and intentions
  publicly
• Then check their subsequent utterances for
  consistency against these professions.
• Semantic Contestability (McBurney Parsons)
• Allow participants to question and contest each
  others statements
• Require agents to provide justifications for
  assertions (of beliefs, preferences, intentions)
  and allow argument over these justifications
• There is a connection here with the
  verificationist theory of truth of Michael
  Dummett and Crispin Wright.

34
Automation of ACL dialogues

• Agent interactions to jointly decide use of
  shared resources have used
• Theories of Persuasion
• Adopted from psychology (Abelson 1960, 1970)
• Example Sierra, Jennings, Noriega Parsons
  1998.
• Agents offer threats/rewards to persuade others
  to adopt proposals
• Acceptance/rejection based on relative positions
  in a social hierarchy.
• Argumentation Theory
• Parsons, Sierra Jennings 1998.
• An agent generates a proposal by constructing an
  argument (a tentative proof) for an intention it
  has, and communicating this to the other
  participants.
• The other agents attempt to counter this
  argument, and only accept it if they fail to
  counter it.
• Uses the Logic of Argumentation of Cancer
  Research UK.

35
Automation of DG dialogues (1)

• Negotiation dialogue protocol of Amgoud, Parsons
  Maudet 2000
• 7 Locutions assert, accept, question, challenge,
  request, promise, refuse.
• Locutions may be instantiated with propositions
  and arguments for propositions.
• Agents vested with an argumentation mechanism, to
  generate arguments for propositions and to accept
  or reject arguments received from other agents.
• Not quite automatic.
• Negotiation dialogue protocol of Sadri, Toni
  Torroni 2001.
• Based on Amgoud, Parsons and Maudet 2000.
• 6 Locutions accept, challenge, request, promise,
  refuse, justify.
• Agents co-operate to agree the use of
  possibly-scarce resources.
• Agents vested with abductive logic mechanisms
  (if-then rules).
• These determine which locution should next be
  uttered, based on the most recent locution
  uttered and the current status of the agents
  resources knowledge base.
• No theoretical grounding for these if-then rules.

36
Automation of DG dialogues (2)

• A Consumer Purchase Transaction Protocol
• A protocol for purchase negotiations for consumer
  durables, based on a standard decision model from
  marketing theory (Roberts Lilien 1993).
• 11 illocutions
• open_dialogue
• enter_dialogue
• seek_info
• willing_to_sell
• desire_to_buy
• prefer
• refuse_to_buy
• refuse_to_sell
• agree_to_buy
• agree_to_sell
• withdraw_dialogue.

37
Automation of DG dialogues (2) (continued)

• Based on the marketing theory decision model, we
  have defined semantic decision mechanisms for the
  participating agents, e.g.
• Seek_Information
• Provide_Information
• Assess_Options, etc.
• Agents vested with these decision mechanisms and
  using the protocol may engage in automated
  dialogues.
• This is proven by defining an Operational
  Semantics for the protocol, a formal definition
  of the locutions in terms of their effects on the
  interaction state-space.
• Protocol due to
• McBurney, van Eijk, Parsons Amgoud 2003.