A Comprehensive Analysis of the cf2 Argumentation Semantics:
From Characterization to Implementation
Sarah Alice Gaggl
Institute of Informationsystems, Vienna University of Technology
Vienna — March 4, 2013
Motivation
Argumentation is one of the major fields inArtificial Intelligence (AI).
Applicationsin diverse domains (legal reasoning, multi-agent systems, social networks, e-government, decision support).
Concept ofabstract Argumentation Frameworks (AFs)[Dung, 1995]
is one of the most popular approaches.
Argumentsand a binaryattackrelation between them, denoting conflicts, are the only components.
Numeroussemanticsto solve the inherent conflicts by selecting acceptable sets of argument.
Admissible-basedversusnaive-basedsemantics.
Motivation ctd.
cf2
Semantics
is based ondecompositionof the framework along itsstrongly connected components (SCCs)[Baroni et al., 2005];
does not require to defend arguments against attacks;
allows totreat cyclesin a moresensitive waythan other semantics;
isnot well studied, due to quite complicated definition.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 2
Motivation ctd.
cf2
Semantics
is based ondecompositionof the framework along itsstrongly connected components (SCCs)[Baroni et al., 2005];
does not require to defend arguments against attacks;
allows totreat cyclesin a moresensitive waythan other semantics;
isnot well studied, due to quite complicated definition.
Goals of the Thesis
Answer-set programming encodingsforcf2. Alternative characterization.
Outline
1 Background on abstract argumentation frameworks and semantics
2 Alternative characterization ofcf2
3 Combiningcf2and stage semantics
4 Redundancies and strong equivalence
5 Computational complexity
6 Implementations
7 Conclusion
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 3
Argumentation Framework
Abstract Argumentation Framework [Dung, 1995]
Anabstract argumentation framework (AF)is a pairF= (A,R), whereA is a finite set of arguments andR⊆A×A. Then(a,b)∈Rifaattacksb. Argumenta∈AisdefendedbyS⊆A(inF) iff, for eachb∈Awith (b,a)∈R,Sattacksb.
Example
Semantics
Semantics for AFs
LetF= (A,R)andS⊆A, we say
Sisconflict-freeinF, i.e.S∈cf(F), if∀a,b∈S: (a,b)6∈R;
S∈cf(F)is maximal conflict-free ornaive(inF), i.e. S∈naive(F), if
∀T ∈cf(F),S6⊂T.
Example
naive(F) ={{a,d,g},{a,c,e},{a,c,g}}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 5
Semantics ctd.
Naive-based Semantics
LetF= (A,R)andS⊆A. LetS+R =S∪ {b| ∃a∈S, s. t.(a,b)∈R}be therangeofS. Then, a setS∈cf(F)is
astableextension (ofF), i.e. S∈stable(F), ifS+R =A; stageinF, i.e. S∈stage(F), if for eachT ∈cf(F),S+R 6⊂TR+.
Example
Semantics ctd.
Admissible-based Semantics
Then,S∈cf(F)isadmissibleinF, i.e.S∈adm(F), if eacha∈Sis defended byS; apreferredextension (ofF), i.e.S∈pref(F), ifS∈adm(F)and for eachT ∈adm(F),S6⊂T.
Example
adm(F) ={∅,{a},{c},{d},{a,c},{a,d}},pref(F) ={{a,c},{a,d}}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 7
cf2 Semantics
One of the SCC-recursive semantics introduced in [Baroni et al., 2005].
Naive-basedsemantics.
Handlesodd- andeven-lengthcyclesin auniformway.
Can accept arguments out of odd-length cycles.
Can accept arguments attacked byself-attackingarguments.
Satisfies most of theevaluation criteriaproposed in [Baroni and Giacomin, 2007].
cf2 Semantics
One of the SCC-recursive semantics introduced in [Baroni et al., 2005].
Naive-basedsemantics.
Handlesodd- andeven-lengthcyclesin auniformway.
Can accept arguments out of odd-length cycles.
Can accept arguments attacked byself-attackingarguments.
Satisfies most of theevaluation criteriaproposed in [Baroni and Giacomin, 2007].
Further Notations, letF= (A,R)
SCCs(F): set ofstrongly connected componentsofF, CF(a): the unique setC∈SCCs(F), s.t.a∈C,
F|S= ((A∩S),R∩(S×S)): sub-frameworkofFw.r.t.S, F|S−S0 =F|S\S0,F−S=F|A\S.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 8
cf2 Semantics ctd.
Definition (D
F(S))LetF= (A,R)be an AF andS⊆A. An argumentb∈Ais
component-defeatedbyS(inF), if there exists ana∈S, such that (a,b)∈Randa∈/CF(b). The set of arguments component-defeated by SinFis denoted byDF(S).
cf2 Semantics ctd.
Definition (D
F(S))LetF= (A,R)be an AF andS⊆A. An argumentb∈Ais
component-defeatedbyS(inF), if there exists ana∈S, such that (a,b)∈Randa∈/CF(b). The set of arguments component-defeated by SinFis denoted byDF(S).
cf2
Extensions [Baroni et al., 2005]
LetF= (A,R)be an argumentation framework andSa set of arguments.
Then,Sis acf2extensionofF, i.e.S∈cf2(F), iff S∈naive(F), in case|SCCs(F)|=1;
otherwise,∀C∈SCCs(F),(S∩C)∈cf2(F|C−DF(S)).
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 9
cf2 Semantics ctd.
S∈cf2(F)
iff,
S∈naive(F), in case|SCCs(F)|=1;
otherwise,∀C∈SCCs(F),(S∩C)∈cf2(F|C−DF(S)).
Example
S={a,d,e,g,i},S∈cf2(F)?
cf2 Semantics ctd.
S∈cf2(F)
iff,
S∈naive(F), in case|SCCs(F)|=1;
otherwise,∀C∈SCCs(F),(S∩C)∈cf2(F|C−DF(S)).
Example
S={a,d,e,g,i},S∈cf2(F)? C1={a,b,c},C2={d}, C3={e,f,g,h,i}andDF(S) ={f}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 10
cf2 Semantics ctd.
S∈cf2(F)
iff,
S∈naive(F), in case|SCCs(F)|=1;
otherwise,∀C∈SCCs(F),(S∩C)∈cf2(F|C−DF(S)).
Example
S={a,d,e,g,i},S∈cf2(F)? C1 ={a,b,c},C2={d}, C3={e,f,g,h,i}andDF(S) ={f}.
cf2 Semantics ctd.
S∈cf2(F)
iff,
S∈naive(F), in case|SCCs(F)|=1;
otherwise,∀C∈SCCs(F),(S∩C)∈cf2(F|C−DF(S)).
Example
S={a,d,e,g,i},S∈cf2(F)? C4={e},C5={g},C6={h},C7={i}
andDF|{e,g,h,i}({e,g,i}) ={h}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 10
Alt. Characterization of cf2
Original definition ofcf2is rathercumbersometo be directly encoded inASPdue to therecursive computationof different sub-frameworks.
Inalternative characterizationwe shift therecursionto a certainset of arguments.
This enables to directly guessa setS;
checkwhetherSis anaive extensionof a certaininstanceofF.
Alt. Characterization of cf2 ctd.
Separation
An AFF= (A,R)is calledseparatedif for each(a,b)∈R, there exists a path frombtoa. We define[[F]] =S
C∈SCCs(F)F|C and call[[F]]the separationofF.
Example
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 12
Alt. Characterization of cf2 ctd.
Separation
An AFF= (A,R)is calledseparatedif for each(a,b)∈R, there exists a path frombtoa. We define[[F]] =S
C∈SCCs(F)F|C and call[[F]]the separationofF.
Example
Alt. Characterization of cf2 ctd.
Reachability
LetF= (A,R)be an AF,Ba set of arguments, anda,b∈A. We say that bisreachableinFfromamoduloB, in symbolsa⇒BF b, if there exists a path fromatobinF|B.
Definition (∆
F,S)
For an AFF= (A,R),D⊆A, and a setSof arguments,
∆F,S(D) ={a∈A| ∃b∈S:b6=a,(b,a)∈R,a6⇒A\DF b},
and∆F,S be the least fixed-point of∆F,S(∅).
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 13
Alt. Characterization of cf2 ctd.
Reachability
LetF= (A,R)be an AF,Ba set of arguments, anda,b∈A. We say that bisreachableinFfromamoduloB, in symbolsa⇒BF b, if there exists a path fromatobinF|B.
Definition (∆
F,S)
For an AFF= (A,R),D⊆A, and a setSof arguments,
∆F,S(D) ={a∈A| ∃b∈S:b6=a,(b,a)∈R,a6⇒A\DF b},
Alt. Characterization of cf2 ctd.
cf2
Extensions [Gaggl and Woltran, 2012]
Given an AFF= (A,R).
cf2(F) ={S|S∈naive(F)∩naive([[F−∆F,S]])}.
Example
S={a,d,e,g,i},S∈naive(F).
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 14
Alt. Characterization of cf2 ctd.
cf2
Extensions [Gaggl and Woltran, 2012]
Given an AFF= (A,R).
cf2(F) ={S|S∈naive(F)∩naive([[F−∆F,S]])}.
Example
S={a,d,e,g,i},S∈naive(F).
Alt. Characterization of cf2 ctd.
cf2
Extensions [Gaggl and Woltran, 2012]
Given an AFF= (A,R).
cf2(F) ={S|S∈naive(F)∩naive([[F−∆F,S]])}.
Example
S={a,d,e,g,i}, S∈naive(F),∆F,S(∅) ={f}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 14
Alt. Characterization of cf2 ctd.
cf2
Extensions [Gaggl and Woltran, 2012]
Given an AFF= (A,R).
cf2(F) ={S|S∈naive(F)∩naive([[F−∆F,S]])}.
Example
S={a,d,e,g,i}, S∈naive(F), ∆F,S({f}) ={f,h}.
Alt. Characterization of cf2 ctd.
cf2
Extensions [Gaggl and Woltran, 2012]
Given an AFF= (A,R).
cf2(F) ={S|S∈naive(F)∩naive([[F−∆F,S]])}.
Example
S={a,d,e,g,i}, S∈naive(F), ∆F,S({f,h}) ={f,h}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 14
Alt. Characterization of cf2 ctd.
cf2
Extensions [Gaggl and Woltran, 2012]
Given an AFF= (A,R).
cf2(F) ={S|S∈naive(F)∩naive([[F−∆F,S]])}.
Example
S={a,d,e,g,i}, S∈naive(F), ∆F,S={f,h},S∈naive([[F−∆F,S]]), thusS∈cf2(F).
Alt. Characterization of cf2 ctd.
cf2
Extensions [Gaggl and Woltran, 2012]
Given an AFF= (A,R).
cf2(F) ={S|S∈naive(F)∩naive([[F−∆F,S]])}.
Example
cf2(F) ={{a,d,e,g,i},{c,d,e,g,i},{b,f,h},{b,g,i}}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 14
Shortcomings of cf2
cf2produces questionable results on AFs with cycles of length≥6.
Example
cf2(F) =naive(F) ={{a,d},{b,e},{c,f},{a,c,e},{b,d,f}};
Shortcomings of cf2 and Stage
cf2produces questionable results on AFs with cycles of length≥6. The grounded extension is not necessarily contained in every stage extension.
Stage semantics does not satisfy directionality.
Example
stage(F) ={{a},{b}}butcf2(F) =ground(F) ={{a}}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 16
Combining cf2 and stage
Wecombinecf2and stage semantics [Dvorák and Gaggl, 2012a], by using theSCC-recursive schema of thecf2semantics and instantiate thebase casewithstagesemantics.
Combining cf2 and stage
Wecombinecf2and stage semantics [Dvorák and Gaggl, 2012a], by using theSCC-recursive schema of thecf2semantics and instantiate thebase casewithstagesemantics.
stage2
Extensions
For any AFF,stage2(F) ={S|S∈naive(F)∩stage([[F−∆F,S]])}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 17
stage2 Examples
For any AFF,stage2(F) ={S|S∈naive(F)∩stage([[F−∆F,S]])}.
Example
stage2(F) =cf2(F) ={{a}}, wherestage(F) ={{a},{b}}.
stage2 Examples
For any AFF,stage2(F) ={S|S∈naive(F)∩stage([[F−∆F,S]])}.
Example
stage2(F) =cf2(F) ={{a}}, wherestage(F) ={{a},{b}}.
stage2(G) =stage(G) ={{a,c,e},{b,d,f}}, but cf2(G) =naive(F) ={{a,d},{b,e},{c,f},{a,c,e},{b,d,f}}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 18
Relations between Semantics
Redundancies
Argumentation is adynamic reasoning process.
Whicheffecthasadditional informationw.r.t. a semantics?
Which informationdoes not contribute to resultsno matter which changes are performed?
Identification ofkernelsto removeredundant attacks [Oikarinen and Woltran, 2011].
Definition
Two AFsFandGarestrongly equivalentto each other w.r.t. a semantics σ, in symbolsF≡σs G, iff for each AFH,σ(F∪H) =σ(G∪H).
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 20
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 21
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
H = (A∪ {d,x,y,z},
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
LetE={d,x,z},E∈cf2(F∪H)butE6∈cf2(G∪H).
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 21
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
LetE={d,x,z},E∈cf2(F∪H)butE6∈cf2(G∪H).
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 21
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
LetE={d,x,z},E∈cf2(F∪H)butE6∈cf2(G∪H).
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 21
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
Strong Equivalence w.r.t. cf2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
LetE={d,x,z},E∈cf2(F∪H)butE6∈cf2(G∪H).
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 21
SE w.r.t. cf2 and stage2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
No matter which AFsF6=G, one can always construct anHs.t.
cf2(F∪H)6=cf2(G∪H);
SE w.r.t. cf2 and stage2
Theorem
For any AFsFandG,F≡cf2s GiffF =G.
No matter which AFsF6=G, one can always construct anHs.t.
cf2(F∪H)6=cf2(G∪H);
Forstage2semantics alsostrong equivalence coincides with syntactic equivalence.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 22
SE w.r.t. cf2 and stage2
No matter which AFsF6=G, one can always construct anHs.t.
cf2(F∪H)6=cf2(G∪H);
Forstage2semantics alsostrong equivalence coincides with syntactic equivalence.
Succinctness Property
An argumentation semanticsσsatisfies thesuccinctness propertyor is maximal succinctiff no AF contains a redundant attack w.r.t.σ.
Comparing Semantics w.r.t. SE
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 23
Complexity Analysis
Ver Cred Skept Exists¬∅
naive inP inP inP inP stable inP NP-c coNP-c NP-c cf2 inP NP-c coNP-c inP stage coNP-c ΣP2-c ΠP2-c inP stage2 coNP-c ΣP2-c ΠP2-c inP Table:Computational complexity of naive-based semantics.
Implementation
Reduction-based Approach
Answer-set Programming (ASP) encodings forcf2andstage2. Saturation vs.metaspencodings forstage2.
All encodings incorporated in the system ASPARTIX[Egly et al., 2010].
Direct Approach
Labeling-basedalgorithms forcf2andstage2.
Web-Application
http://rull.dbai.tuwien.ac.at:8080/ASPARTIX
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 25
Conclusion
Alternative characterizationforcf2to avoid the recursive computation of sub-frameworks.
stage2semantics overcomes problems ofcf2.
Strong equivalencew.r.t.cf2(resp.stage2) coincides with syntactic equivalence.
Provided the missingcomplexity resultsforcf2(resp.stage2).
Implementationin terms ofASPandlabeling-based algorithms.
Future Work
Further relations to other semantics likeintertranslatability.
Optimizationsof ASP encodings.
Development ofappropriate instantiation methodsfor naive-based semantics.
Other combinationsof semantics in the alternative characterization, likesem(F) ={S|σ(F)∩τ([[F−∆F,S]])}.
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 27
Baroni, P. and Giacomin, M. (2007).
On principle-based evaluation of extension-based argumentation semantics.
Artif. Intell., 171(10-15):675–700, 2007.
Baroni, P., Giacomin, M., and Guida, G. (2005).
Scc-recursiveness: A general schema for argumentation semantics.
Artif. Intell., 168(1-2):162–210.
Dung, P. M. (1995).
On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games.
Artif. Intell., 77(2):321–358.
Dvorák, W. and Gaggl, S. A. (2012)
Incorporating stage semantics in the scc-recursive schema for argumentation semantics.
In Proceedings of the 14th International Workshop on Non-Monotonic Reasoning(NMR 2012), 2012.
Dvorák, W. and Gaggl, S. A. (2012)
Computational aspects of cf2 and stage2 argumentation semantics.
In Bart Verheij, Stefan Szeider, and Stefan Woltran, editors, Proceedings of the 4th International Conference on
Computational Models of Argument(COMMA 2012), volume 245 of Frontiers in Artificial Intelligence and Applications, pages 273–284. IOS Press, 2012.
Egly, U., Gaggl, S. A. and Woltran, S. (2010)
Answer-set programming encodings for argumentation frameworks.
Argument and Computation, 1(2):144–177, 2010.
Gaggl, S. A. and Woltran, S. (2012).
Tractable Fragments
cf2 stage2 stable stage
Credσacycl inP inP P-c P-c
Skeptσacycl inP inP P-c P-c Credσeven−free NP-c coNP-h P-c ΣP2-c Skeptσeven−free coNP-c coNP-h P-c ΠP2-c
Credσbipart inP inP P-c P-c
Skeptσbipart inP inP P-c P-c Credσsym inP inP/ΣP2∗ inP inP/ΣP2∗ Skeptσsym inP inP/ΠP2∗ inP inP/ΠP2∗ Table:Complexity results for special AFs (∗with self-attacking arguments).
Sarah A. Gaggl, TU Vienna Comprehensive Analysis of cf2 Semantics 29