Page 1
University of BresciaDipartimento di Elettronica per l'Automazione
Knowledge Engineering and Human-Computer Interaction Research Group
© 2009 Federico Cerutti <[email protected] >
An Argumentation-based An Argumentation-based Approach to Modelling Decision Approach to Modelling Decision Support Contexts with What-If Support Contexts with What-If
CapabilitiesCapabilities
Pietro Baroni, Federico Cerutti, Massimiliano Giacomin and Giovanni GuidaPietro Baroni, Federico Cerutti, Massimiliano Giacomin and Giovanni Guida
AAAI 2009 Fall Symposium SeriesThe Uses of Computational Argumentation
Arlington, November, 5, 2009
Page 2
Slide 2 © 2009 Federico Cerutti <[email protected] >
“Good advice”
The advice should be presented in a form which can be readily understood by decision makers
There should be ready access to both information and reasoning underpinning the advice
If decision support involves details which are unusual to the decision maker, it is of primary importance that s/he can discuss these details with his advisor
Girle et al., 2003
Page 3
Slide 3 © 2009 Federico Cerutti <[email protected] >
Transparency about the advice
Practical reasoningabout “what to do”
Knowledgerepresentation
Computation ofoutcomes
Page 4
© 2009 Federico Cerutti <[email protected] >
Knowledge representationKnowledge representation➢ Computation of outcomes
➢ Conclusions and future works
Page 5
© 2009 Federico Cerutti <[email protected] >
Argument (and attack) schemes
Use of argument scheme to represent the knowledge Structure which contains the information in favour of
a given conclusion
Introduction of a possible modelling of conflicts by “attack scheme”
Structure which contains the information in favour of a given conflict
Page 6
© 2009 Federico Cerutti <[email protected] >
The main concepts
Circumstance: a state of the world Fact: a particular circumstance assumed to be true Goal: a state of the world we want to achieve Action: support for the achievement of a goal Preference: “[…] a greater liking for one alternative
over another or others […]” Value: “Worth or worthiness […] in respect of rank
or personal qualities” Must Value: a value that we commit to promote
Page 7
Slide 7 © 2009 Federico Cerutti <[email protected] >
Example (1)
Treatment for a patient threatened by blood clotting Goal: obtaining a low platelet adhesion The available knowledge base concerning
treatments: Administer Aspirin (value of Safety) Administer Chlopidogrel (value of Safety) Do nothing (value of Cost)
Page 8
Slide 8 © 2009 Federico Cerutti <[email protected] >
PAS: A2Circumstances: given patient's situationAction: we should administer aspirinGoal: reducing blood clottingValue: SafetySign: +
Formal counterpart (1)Practical Args (from Atkinson et al.)
Page 9
Slide 9 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (1)The Attacks among PAS
PAtS1: αSource: A1Target: A2Conditions:A1.action and A2.action are incompatible
Page 10
Slide 10 © 2009 Federico Cerutti <[email protected] >
VAS: V1Value: Cost
Formal counterpart (1)The Values
Page 11
Slide 11 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (1)The Defences from the Values
VDeS1: βSource: V2Target: αConditions:α.target.value = V2.value, α.source.value ≠ V2.value
PAtS1: α…
Page 12
Slide 12 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (1)The Defences from the Values
VDeS1: βSource: V2Target: αConditions:α.target.value = V2.value, α.source.value ≠ V2.value
VDeS2: γSource: V1Target: βConditions:β.source ≠ V1, β.target.source.value ≠ V1.value
PAtS1: α…
Page 13
Slide 13 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (1)The Defences from the Values
VDeS1: βSource: V2Target: αConditions:α.target.value = V2.value, α.source.value ≠ V2.value
VDeS2: γSource: V1Target: βConditions:β.source ≠ V1, β.target.source.value ≠ V1.valueVDefence: β
Defending: A2Defended: V2
VDefence: γDefending: A1Defended: V1
PAtS1: α…
Page 14
Slide 14 © 2009 Federico Cerutti <[email protected] >
Example (2)
From patient's file we learn that he has a history of gastritis
Then we should not administer Aspirin without a proton pump inhibitor
In fact, it gives rise to risk of ulceration And it will demote the value of Safety As far as we know, no proton pump inhibitor is
available
Page 15
Slide 15 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (2)A PAS with negative sign...
PAS: A4Circumstances: proton pump unavailableAction: we should not administer aspirinGoal: risk of ulcerationValue: SafetySign: -
Page 16
Slide 16 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (2)...and the relative attacks
PAtS2: δSource: A4Target: A2Conditions:A4.action = ¬ A2.action
Page 17
Slide 17 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (2)...and the relative attacks
VAAtS: εSource: A4Target: βConditions: A4.circumstance = β.defended.circumstance, A4.action = ¬ β.defended.action, A4.goal = β.defended.goal, A4.value = β.defended.value, A4.sign = -, β.defended.sign = +
PAtS2: δSource: A4Target: A2Conditions:A4.action = ¬ A2.action
Page 18
Slide 18 © 2009 Federico Cerutti <[email protected] >
Example (3)
Suppose now that the assumption that no proton pump inhibitor is available reveals to be false
Suppose also that between aspirin and chlopidogrel a doctor prefers to administer aspirin because it is in stock and immediately available
Page 19
Slide 19 © 2009 Federico Cerutti <[email protected] >
FAS: A5Circumstances: a proton pump is available
Formal counterpart (3)A Fact
Page 20
Slide 20 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (3)An “undercut”
FAtS: ζSource: A5Target: A4Conditions: A5.circumstances= ¬ A4.circumstances
Page 21
Slide 21 © 2009 Federico Cerutti <[email protected] >
PRAS: P1Preferred: A2Notpreferred: A3
Formal counterpart (3)A preference
Page 22
Slide 22 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (3)The greater liking
PAtS: η...
FAtS: θSource: P1Target: ηConditions: P1.preferred = η.target, P1.notpreferred = η.source
Page 23
Slide 23 © 2009 Federico Cerutti <[email protected] >
Example (4)
Determine the ultimate decision outcome Achieve the goal of reducing blood clotting It promotes the value of Safety We must promote the value of Safety
Page 24
Slide 24 © 2009 Federico Cerutti <[email protected] >
MAS: MV2Value: Safety
Formal counterpart (4)The Must Value
Page 25
Slide 25 © 2009 Federico Cerutti <[email protected] >
Formal counterpart (4)What-If Scenario
MAtS2: κSource: MV2Target: γConditions: MV2.value = γ.target.source.value,MV2.value ≠ γ.source.value
VDeS2: γ...
Page 26
Slide 26 © 2009 Federico Cerutti <[email protected] >
Small SummaryArgument Scheme Taxonomy
Practical Argument Scheme Factual Argument Scheme Value Argument Scheme Preference Argument Scheme Must Argument Scheme
Page 27
Slide 27 © 2009 Federico Cerutti <[email protected] >
In the example
Page 28
Slide 28 © 2009 Federico Cerutti <[email protected] >
Small SummaryAttack Scheme Taxonomy (1)
PAtS1: incompatible actions PAtS2: “rebuttal” VAtS : incompatible values VDefence (VDeS[1-2]): a value protects both the
arguments which promote it and the attacks sourced from that arguments
VAAtS: if a practical argument P suggests not to perform an action A since it demotes a value V, if P will be considered acceptable, then V cannot defend the argument whose action is A
Page 29
Slide 29 © 2009 Federico Cerutti <[email protected] >
In the example (1)
Page 30
Slide 30 © 2009 Federico Cerutti <[email protected] >
Small SummaryAttack Scheme Taxonomy (2)
FAtS: “undercut” PRAtS: someone told us that an attack cannot be
considered since an external preference MAtS1: an instance of Must Argument Scheme has
to protect the related Value argument against the incompatible values
MAtS2: an instance of Must Argument Scheme has to protect the instances of VDefence which start from the related Value argument
Page 31
Slide 31 © 2009 Federico Cerutti <[email protected] >
In the example (2)
Page 32
© 2009 Federico Cerutti <[email protected] >
Computation of outcomesComputation of outcomes➢ Knowledge representation
➢ Conclusions and future works
Page 33
Slide 33 © 2009 Federico Cerutti <[email protected] >
Argumentation Framework for Decision Support Problem
Page 34
Slide 34 © 2009 Federico Cerutti <[email protected] >
Argumentation Framework withRecoursive Attacks (AFRA)
Page 35
Slide 35 © 2009 Federico Cerutti <[email protected] >
From AFDSP to AFRA
Page 36
Slide 36 © 2009 Federico Cerutti <[email protected] >
AFRA: Defeat relation
Page 37
Slide 37 © 2009 Federico Cerutti <[email protected] >
AFRA: Admissibility
Page 38
Slide 38 © 2009 Federico Cerutti <[email protected] >
AFRA: Preferred Extension
Page 39
Slide 39 © 2009 Federico Cerutti <[email protected] >
Recalling the example...
Page 40
Slide 40 © 2009 Federico Cerutti <[email protected] >
...and the preferred extension
Page 41
© 2009 Federico Cerutti <[email protected] >
Conclusions and Conclusions and future worksfuture works
➢ Knowledge representation➢ Computation of outcomes
Page 42
Slide 42 © 2009 Federico Cerutti <[email protected] >
Conclusions
Preliminary investigation about formalisation of decision support problems
Three main contributions: The role of attack schemes Attacks to attacks in practice Support to “What-if” reasoning
Page 43
Slide 43 © 2009 Federico Cerutti <[email protected] >
Future works
Knowledge representation: Enhancing attacks schemes Ontological status of attacks Multiple What-if Situations
Computation of outcomes Further investigation on the theoretical bases of AFRA Argumentation semantics in this context
Page 44
Slide 44 © 2009 Federico Cerutti <[email protected] >
Open questions
The notion of attack scheme: soundness and usefulness
Attack schemes and critical questions What-if only w.r.t. Values