Artificial Intelligence, Computational Logic
DECOMPOSING ABSTRACT DIALECTICAL FRAMEWORKS
Sarah Gaggl and Hannes Strass
Pitlochry, 12th September 2014
Motivation
•
Computational complexityof semantics for ADFs is in general higher than for AFs [Strass and Wallner, 2014].•
Algorithms based onSCC-recursive schemafor AF semantics show significant performance gain [Cerutti et.al. KR 2014].•
We propose a similar approach based on a recursive decomposition along SCCs.•
Allows to definecf2andstage2semantics for ADFs.Motivation
•
Computational complexityof semantics for ADFs is in general higher than for AFs [Strass and Wallner, 2014].•
Algorithms based onSCC-recursive schemafor AF semantics show significant performance gain [Cerutti et.al. KR 2014].•
We propose a similar approach based on a recursive decomposition along SCCs.•
Allows to definecf2andstage2semantics for ADFs.Main Difference to AFs
1 Acceptance conditions of statements insub-frameworksmay still depend on statements not contained in sub-framework.
2 Elimination ofredundanciesfrom links and acceptance formulas.
3 Propagationof truth values to subsequent SCCs.
Agenda
1 Introduction and Background
– Abstract Dialectical Framework (ADFs)
2 Decomposing ADFs – Sub-Frameworks – Redundancies – Reduced Frameworks
– Decomposition-based Semantics
3 Conclusion and Future Work
ADFs - The Formal Framework
•
Like AFs, use graph to describe dependencies among nodes.•
Unlike AFs, allow individual acceptance condition for each node.•
Assignst(rue) orf(alse) depending on status of parents.Definition
Anabstract dialectical framework(ADF) is a tupleD= (S,L,C)where
•
Sis a set ofstatements(positions, nodes),•
L⊆S×Sis a set oflinks,•
C={Cs}s∈Sis a set of total functionsCs:2par(s)→ {t,f}, one for each statements.Csis calledacceptance conditionofs.Semantics
Definition
Letϕbe a propositional formula over vocabularySand for anM⊆Slet v:M→ {t,f,u}be athree-valued interpretation.
Thepartial valuationofϕbyvisϕv=ϕ[p/t:v(p) =t][p/f:v(p) =f].
Definition
LetD= (S,L,C)be an ADF. A three-valued interpretationvis
•
conflict-freeiff for alls∈Swe have:– v(s) =timplies thatϕvsis satisfiable, – v(s) =fimplies thatϕvsis unsatisfiable;
•
naiveiff it is≤i-maximal with respect to being conflict-free;Where≤iis apartial orderover the truth values (resp. interpretations), i.e.
u<itandu<if.
Semantics ctd.
Definition
LetD= (S,L,C)be an ADF. Thepartial valuationofϕbyvis ϕv=ϕ[p/t:v(p) =t][p/f:v(p) =f].
A three-valued interpretationvis
•
conflict-freeiff for alls∈Swe have:– v(s) =timplies thatϕvsis satisfiable, – v(s) =fimplies thatϕvsis unsatisfiable;
•
naiveiff it is≤i-maximal with respect to being conflict-free;Example
a
¬c
b
¬a
c
¬b
v={a7→f,b7→u,c7→t}is conflict-free, asϕva=¬tis unsatisfiable and ϕvc=¬bis satisfiable.
Sub-Frameworks and Redundancies
a
¬c
b
¬a
c
¬b
d
c∨f
e
d∧f
f
e
•
independent setindD(∅) ={a,b,c}=M0•
independent moduloM0:indD(M0) ={a,b,c,d,e,f}=S•
Mindependent set:sub-frameworkD|M = (M,L∩(M×M),{ϕs}s∈M)Sub-Frameworks and Redundancies
a
¬c
b
¬a
c
¬b
d
c∨f
e
d∧f
f
e
•
independent setindD(∅) ={a,b,c}=M0•
independent moduloM0:indD(M0) ={a,b,c,d,e,f}=S•
Mindependent set:sub-frameworkD|M = (M,L∩(M×M),{ϕs}s∈M)•
Redundanciescan change dependencies between statements.•
If(r,s)is redundant thenrhas no influence on the truth value ofϕswhatsoever.
Example
Considerϕs=a∨(b∧c)and the interpretationv={a7→u,b7→f,c7→u}.
ϕvs=a∨(f∧c)
Sub-Frameworks and Redundancies
a
¬c
b
¬a
c
¬b
d
c∨f
e
d∧f
f
e
•
independent setindD(∅) ={a,b,c}=M0•
independent moduloM0:indD(M0) ={a,b,c,d,e,f}=S•
Mindependent set:sub-frameworkD|M = (M,L∩(M×M),{ϕs}s∈M)•
Redundanciescan change dependencies between statements.•
If(r,s)is redundant thenrhas no influence on the truth value ofϕswhatsoever.
Example
Considerϕs=a∨(b∧c)and the interpretationv={a7→u,b7→f,c7→u}.
ϕvs=a∨(f∧c)≡a chas no influence
Reduced ADF
Given an ADFD= (S,L,C), an independent setM⊆Sand an interpretation v:M→ {t,f,u}. The ADFDreduced withvonMis obtained by:
•
adapt the acceptance condition of statementsto – t(resp.f) ifv(s) =t(resp.v(s) =f) – ¬sifv(s) =u– partial valuationϕvsfor remaining statements and ifris redundant in ϕvs, replacerwitht
•
remove redundant links•
add links{(s,s)|v(s) =u}Reduced ADF
Given an ADFD= (S,L,C), an independent setM⊆Sand an interpretation v:M→ {t,f,u}. The ADFDreduced withvonMis obtained by:
•
adapt the acceptance condition of statementsto – t(resp.f) ifv(s) =t(resp.v(s) =f) – ¬sifv(s) =u– partial valuationϕvsfor remaining statements and ifris redundant in ϕvs, replacerwitht
•
remove redundant links•
add links{(s,s)|v(s) =u}Procedure
For a semanticsσand an ADFD, we obtain theσ2interpretations recursively by applyingσ2(D) =σ2(indD(∅),D)by:
1 Start with all statementsindependent modulo∅, i.e.M0=indD(∅)
2 Compute allσ-interpretations of sub-frameworkD|M
0 3 For eachσ-interpretationwofD|M
0compute thereduced ADF
4 Call Step 1 with reduced ADF andM1=indD(M0).
Example
a
¬c
b
¬a
c
¬b
d
c∨f
e
d∧f
f
e nai2(D) =nai2(indD(∅),D)
1 indD(∅) ={a,b,c}=M0
Example
a
¬c
b
¬a
c
¬b
d
c∨f
e
d∧f
f
e nai2(D) =nai2(indD(∅),D)
1 indD(∅) ={a,b,c}=M0
2 Then we obtainnai(D|M
0) ={v0,v1,v2}:
v0={a7→u,b7→t,c7→f}, v1={a7→f,b7→u,c7→t}, v2={a7→t,b7→f,c7→u}.
Example
a
¬c
b
¬a
c
¬b
d
c∨f
e
d∧f
f
e nai2(D) =nai2(indD(∅),D)
1 indD(∅) ={a,b,c}=M0
2 v1={a7→f,b7→u,c7→t}
Example
a
f
b
¬b
c
t
d
t∨f
e
d∧f
f
e nai2(D) =nai2(indD(∅),D)
1 indD(∅) ={a,b,c}=M0
2 v1={a7→f,b7→u,c7→t}
3 Reduced ADF witht∨f ≡t, thus link(f,d)is redundant
Example
a
f
b
¬b
c
t
d
t∨t
e
d∧f
f
e nai2(D) =nai2(indD(∅),D)
1 indD(∅) ={a,b,c}=M0
2 v1={a7→f,b7→u,c7→t}
3 Reduced ADF
Example
a
f
b
¬b
c
t
d
t∨t
e
d∧f
f
e nai2(D) =nai2(indD(∅),D)
1 indD(∅) ={a,b,c}=M0
2 v1={a7→f,b7→u,c7→t}
3 Reduced ADF
4 Call Step 1 with reduced ADFD1andM1=indD1(M0).
Example
a
f
b
¬b
c
t
d
t∨t
e
d∧f
f
e nai2(M1,D1)
1 M1=indD1(M0) ={a,b,c,d}
Example
a
f
b
¬b
c
t
d
t∨t
e
d∧f
f
e nai2(M1,D1)
1 M1=indD1(M0) ={a,b,c,d}
2 nai(D|M
1) =v3=v1∪ {d7→t}
Example
a
f
b
¬b
c
t
d
t
e
t∧f
f
e nai2(M1,D1)
1 M1=indD1(M0) ={a,b,c,d}
2 nai(D|M
1) =v3=v1∪ {d7→t}
3 Reduced ADF
Example
a
f
b
¬b
c
t
d
t
e
t∧f
f
e nai2(M1,D1)
1 M1=indD1(M0) ={a,b,c,d}
2 nai(D|M
1) =v3=v1∪ {d7→t}
3 Reduced ADF
4 Call Step 1 with reduced ADFD2andM2=indD2(M1)
Example
a
f
b
¬b
c
t
d
t
e
t∧f
f
e nai2(M2,D2)
1 M2=indD2(M1) ={a,b,c,d,e,f}=S
Example
a
f
b
¬b
c
t
d
t
e
t∧f
f
e nai2(M2,D2)
1 M2=indD2(M1) ={a,b,c,d,e,f}=S
2 nai(D2) ={v4,v6}:
v4=v3∪ {e7→t,f 7→t}, v6=v3∪ {e7→f,f7→f}.
Main Theorem
Theorem
1. Letσ∈ {cfi,adm,pre,com,mod}. Thenσ≤σ2. 2. Letσ∈ {nai,stg}. Thenσ6≤σ2. 3. Letσ∈ {cfi,nai,adm,pre,com,mod}. Thenσ2≤σ.
4. Letσ∈ {stg}. Thenσ26≤σ.
Conclusion
•
We proposed a decomposition schema for semantics for ADFs.•
We introducednai2andstg2for ADFs.•
Due to the relation of ADFs to logic programms we also getnai2andstg2 semantics for LPs.•
Also in the paper:composingADFs.Future Work
•
Analysis of the complexity of the approach.•
Implementation.•
Studysplittingsof ADFs andequivalences.References
Leila Amgoud and Claudette Cayrol and Marie-Christine Lagasquie and Pierre Livet, On Bipolarity in Argumentation Frameworks
International Journal of Intelligent Systems 23(10): 1062–1093 (2008)
P. Baroni, F. Cerutti, M. Giacomin and G. Guida.
AFRA: Argumentation Framework with Recursive Attacks.
Int. J. Approx. Reasoning 52(1): 19–37 (2011).
Baroni, P., Giacomin, M., and Guida, G. (2005).
Scc-recursiveness: A general schema for argumentation semantics.
Artif. Intell., 168(1-2):162–210.
G. Brewka and S. Woltran.
Abstract Dialectical Frameworks.
KR 2010, pp. 102–111, AAAI Press, 2010.
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.
S. Modgil.
Reasoning about Preferences in Argumentation Frameworks.
Artif. Intell. 173(9-10): 910–934 (2009).
H. Strass and J. Wallner.
Analyzing the Computational Complexity of Abstract Dialectical Frameworks via Approximation Fixpoint Theory.
KR 2014, AAAI Press, 2014.