Revising Beliefs Through Arguments: Bridging the Gap Between Argumentation and Belief Revision in MA

1
Revising Beliefs Through ArgumentsBridging the
Gap BetweenArgumentation and Belief Revision in
MAS

• Fabio Paglieri
• University of Siena
• Cristiano Castelfranchi
• ISTC-CNR, Roma

ArgMAS 2004 New York July 19, 2004
2
Introduction Belief Revision and Argumentation

• 1st underlying claim Belief revision (BR)
  and argumentation strategies (Arg) are best
  understood / should be studied within the same
  conceptual framework.
• Belief revision the way in which an agent
  changes its own mind, i.e. its own beliefs.
• Argumentation the way in which an agent changes
  other agents mind, by influencing their beliefs
  through communication.

• Two sides (cognitive and social) of the same
  epistemic coin.
• Related works belief change and communication
  (Galliers 1992), belief revision and defeasible
  reasoning (Pollock, Gillies 2000 Falappa,
  Kern-Isberner, Simari 2002).
• In ArgMAS 2004 argumentation-based model of
  beliefs and belief change (Capobianco, Chesñevar,
  Simari), argumentation as belief-monitoring
  (Adler).
• 2nd underlying claim In order to capture
  and model Arg, we need BR formalisms with a
  proper degree of structural complexity.

3
Introducing Data-oriented Belief Revision (DBR)

• Our model is based on a fundamental distinction
  between data and beliefs.
• Data are information available to the agent (i.e.
  gathered and stored in his mind), without and
  before any commitment to their reliability.
• Beliefs are those data that the agent accepts as
  reliable bases for action, decision and specific
  reasoning tasks, e.g. inference, prediction and
  explanation.

• In this framework, data are selected (i.e. either
  accepted or rejected as beliefs) on the ground of
  their informational properties. Thus, changes
  over time in the outcomes of the selection
  process determine belief revision in other
  words, BR is an emerging effect of data
  manipulation.
• We call this process Data-oriented Belief
  Revision (DBR).

4
Introducing Data-oriented Belief Revision (DBR)

• DBR is based on a conceptual model of epistemic
  processing far more complex than the original AGM
  scheme (Alchourrón, Gärdenfors, Makinson 1985)

5
Introducing Data-oriented Belief Revision (DBR)

• Informational properties of data (Castelfranchi
  1996 Paglieri 2004) are
• Relevance a measure of the pragmatic utility of
  the datum, i.e. number and values of the
  (pursued) goals for which the datum is
  needed/useful
• Credibility a measure of the number and values
  of all supporting data, contrasted with all
  conflicting data, down to external and internal
  sources
• Importance a measure of the epistemic
  connectivity of the datum, i.e. number and values
  of the data that the agent will be forced to
  reconsider, should he reconsider that single one
• Likeability a measure of the motivational appeal
  of the datum, i.e. number and values of the
  (pursued) goals that are directly fulfilled by
  that datum.
• All of these are relational properties thus, in
  DBR data are organized in networks, called Data
  Structures.

6
Introducing Data-oriented Belief Revision (DBR)

• Belief selection active data (i.e. data
  candidate as beliefs) are either accepted or
  rejected as beliefs on the basis of their
  credibility and/or importance and/or likeability,
  depending by the selection parameters of that
  particular agent (i.e. the informational
  properties that he considers most crucial in
  assessing the reliability of a given datum).

• Belief selection in DBR is performed by a
  mathematical system formed by a condition C, a
  threshold k and a function F. C and k together
  determine whether the datum is accepted or not as
  belief, while F assigns a value of strength to
  the corresponding belief (if any).
• Given a datum f with credibility cf, importance
  if and likeability lf, let B represent the
  agents belief set and Bsf the belief f with
  strength s. Then the general form of belief
  selection is
• If C(cf, if, lf) k then Bsf ? B
• If C(cf, if, lf) gt k then Bsf ? B with sf
  F(cf, if, lf)

7
Introducing Data-oriented Belief Revision (DBR)

• Data structures consist in data (nodes) linked
  together by characteristic relations (link). We
  define three basic relations

8
Introducing Data-oriented Belief Revision (DBR)

• Example Rhett was aware that his beloved
  Scarlett was supposed to take the US637 flight
  from Atlanta to New Orleans today, to pay a visit
  to her elderly wet nurse Mamy. Watching the news,
  Rhett is informed that there has been a terrible
  crash during the landing of that airplane and all
  passengers died. Frantic, Rhett calls Mamy, who
  tells him to get a grip of himself and stop
  blabbering, since Scarlett arrived safe and sound
  at her home two hours ago.

DATA a Scarlett was on the flight US637
today b all passengers of today flight US637
died in a crash g Scarlett is right now at
Mamys home, safe and sound d Scarlett is
dead RELATIONS (a b) ? d, g ? d
9
Modeling Argumentation in DBR

• Representation results in Data Structures
• argumentation through plausibility
• Toulmins model of argument
• defeasible reasoning.
• Expressivity results in DBR jump
• contradiction management
• local vs. global argumentation strategies.

10
Modeling Argumentation in DBR

• A crucial feature in argumentation is
  plausibility, i.e. how much the claim of the
  arguer fit in with the pre-existing beliefs of
  the audience.

• In DBR, plausibility-based arguments work on the
  importance of the datum that the arguer wants to
  defend. Importance of the datum can be
  manipulated employing two different cognitive
  strategies
• Self-evident datum the new datum is presented as
  following from what the audience already knew
  the datum has not yet been inferred, but it might
  have been, and the audience is likely to remark
  Sure! Of course! Obviously! etc.
• Explanatory datum the new datum is presented as
  supporting and explaining data already available
  to the audience since such explanation was
  missing so far, it produces reactions like Now
  I see! Thats why! I knew it! etc.

11
Modeling Argumentation in DBR

• Toulmins model can be easily represented in DBR
  as a peculiar Data Structure, as follows

12
Modeling Argumentation in DBR

• In DBR is it possible to represent three
  different types of defeaters
• direct (rebutting) defeaters data contrasting
  the C node
• premise defeaters data contrasting the D node
• undercutting defeaters data contrasting the W
  node (i.e. rebuttals).

• Example John is innocent of the murder of his
  wife (claim) because he loved her much (data) and
  usually (qualifier) people do not murder the ones
  they love (warrant), since murder implies hate
  towards the victim (backing).
• Direct defeater John had been seen shooting his
  wife.
• Premise defeater John had a secret affaire with
  another woman.
• Undercutting defeater Jealousy can make you
  kill the ones you love most.

13
Modeling Argumentation in DBR

• While AGM approaches to BR exclude contradictions
  in principle from the agents belief set,
  argumentation proved to be a successful tool for
  handling contradictions (e.g. Amgoud, Cayrol
  2002 De Rosis et al. 2000 Pollock, Gillies
  2000).

• In DBR we distinguish three types of mutually
  conflicting information
• data contrast this is not at all problematic,
  but rather a beneficial relation, since it allows
  the agent to gather negative evidence on the
  contrasting claims
• implicit contradiction the belief set of a
  resource-bounded agent is not closed under
  deduction, hence he might harbor (and ignore)
  some implicit contradictions and making
  explicit such implicit contradictions is a
  well-known argumentative strategy
• explicit contradiction co-occurring beliefs that
  mutually contradict each other.
• The bottom-line is contradictions need to be
  solved only if they arise at the level of
  beliefs. This is rare in DBR, but not impossible.
  However, the agent is not safe from contradiction
  by some benevolent law of nature, but he is
  rather equipped to handle contradictions
  efficiently.

14
Modeling Argumentation in DBR

• There is a relevant distinction between
• local argumentation the agent aims to persuade
  his audience of a claim or set of claims, hence
  addressing specific beliefs (as in the examples
  discussed so far)
• global argumentation the agent aims to make the
  audience accept a different way of thinking,
  hence modifying their belief revision procedures
  (e.g. political campaign, commercial advertising,
  religious proselytism).

• In DBR, local argumentation targets and
  manipulates specific nodes, relations or
  structures in the audiences data network and/or
  belief set, while global argumentation targets
  the epistemic parameters of the audience, trying
  to change their setting according to the arguers
  goal.

15
Modeling Argumentation in DBR

• Unless I see in his hands the print of the
  nails, and place my finger in the mark of the
  nails, and place my hand in his side, I will not
  believe. (St John, 20 25)

16
Modeling Argumentation in DBR
Have you believed because you have seen me?
Blessed are those who have not seen and yet
believe.
Meaning you should be readier to believe without
much evidence i.e. changing your whole attitude
in belief revision. GLOBAL ARGUMENTATION
17
Modeling Argumentation in DBR
18
Current and Future Works

• To refine the DBR model, e.g. by exploring in
  detail the assessment of data properties
  (Paglieri 2004), information update
  (Castelfranchi 1997 Fullam 2003), data mapping
  and inferential processing of beliefs.
• To move towards implementation in agent-based
  cognitive and social simulation, e.g. within the
  AKIRA framework (Pezzulo, Calvi 2004),
  characterizing DBR itself as a distributed system
  (Dragoni, Giorgini 2003).
• To compare our theoretical predictions over
  belief change and argumentation with empirical
  findings in experimental psychology.
• To explore the possible relationship between DBR
  and Truth-Maintenance Systems (TMS see Doyle
  1979 Huns, Bridgeland 1991).
• To provide more systematic connections with other
  argumentation models, especially within the MAS
  community.
• For instance, wrt Capobianco, Chesñevar, Simari
  2004dialectical databases vs. data
  structuresas precompiled knowledge as
  precompiled structured informationperception v
  s. information updateas flawless and
  prioritized as source-depending (reliability)

19
Current and Future Works

• 2nd underlying claim In order to capture
  and model Arg, we need BR formalisms with a
  proper degree of structural complexity.