Probabilistic Argumentation

Origin. Dung's abstract argumentation frameworks (1995) fix which arguments defeat which and ask which sets survive; the probabilistic extensions add uncertainty to that picture in two incompatible ways. Li, Oren and Norman (TAFA 2011) put a distribution over which arguments and attacks are actually present — the constellations approach. Hunter (2012, 2013) and Thimm (ECAI 2012) instead keep the graph fixed and read a probability on each argument as a degree of belief, constrained by postulates tied to the attack relation — the epistemic approach. Earlier, Haenni, Kohlas and Lehmann's probabilistic argumentation systems (2000) derived degrees of support from Dempster–Shafer theory rather than from a Dung graph.

Models. Two readings of the same numbers, answering different questions. On the constellations reading a probabilistic framework induces a distribution over ordinary sub-frameworks; classical semantics is applied to each, and the probability of an argument is the mass of sub-frameworks in whose extension it lies — so the number is the chance that an argument is accepted. On the epistemic reading the graph is certain and the number is belief in the argument's claim; the attack relation constrains the distribution rather than generating it, and Dung's semantics is recovered at the extremes when the degrees are pushed to 0 and 1.

Formalism.

Probabilistic AF: ⟨A, R, P⟩ where: A: arguments. R ⊆ A × A: attacks. P: probability distribution.

Argument probability: P(a): probability argument a holds. Uncertain premises. Graded.

Attack probability: P(a → b): probability attack succeeds. Uncertain defeat. Conditional.

Constellation approach: Distribution over subframeworks. Each subframework: subset of A, with the induced attacks. Semantics applied to each. Expected acceptance. For multiple-extension semantics, credulous and sceptical readings give different numbers.

Grounded probability: P_gr(a) = Σ_F P(F) · 1_{a ∈ gr(F)}. Sum over subframeworks. Weighted by probability. a in grounded extension.

Epistemic approach: Probability = degree of belief. Bayesian updates. Evidence changes P.

Epistemic postulates: Coherent: (a,b) ∈ R ⇒ P(a) ≤ 1 − P(b). Founded: P(a) = 1 for every unattacked a. Optimistic: P(a) ≥ 1 − Σ_{b : (b,a) ∈ R} P(b). Rational (Thimm): (a,b) ∈ R and P(a) > 0.5 ⇒ P(b) ≤ 0.5. Under coherence, {a : P(a) > 0.5} is conflict-free.

Likelihood-based: L(a | E) ∝ P(E | a) · P(a). Evidence E. Posterior from likelihood.

Social choice: Multiple agents. Aggregate probabilistic views. Judgment aggregation meets argumentation.

Li et al. semantics: P(a is IN) based on attackers. P(a is IN) = Π_{b attacks a} P(b is OUT). Product formula. Independence assumption.

Symbols.

Symbol	Unicode	Name	Meaning
P(a)	—	Argument probability	Chance of acceptance, or degree of belief
P_gr(a)	—	Grounded acceptance probability	Mass of subframeworks whose grounded extension contains a
A	—	Argument set	The arguments of the framework
R	—	Attack relation	R ⊆ A × A
F	—	Subframework	One realization of the uncertain graph
gr	—	Grounded extension	The least complete extension of a subframework
PAF	—	Probabilistic argumentation framework	The triple ⟨A, R, P⟩
E	—	Evidence	The observation conditioned on
L	—	Likelihood	Posterior support for an argument
⟨ ⟩	—	Angle brackets	Tuple notation for the framework
→	U+2192	Rightwards arrow	An attack, in P(a → b)
⊆	U+2286	Subset	R inside A × A
×	U+00D7	Multiplication sign	Cartesian product of A with itself
∈	U+2208	Element of	Membership of an argument in an extension
Σ	U+03A3	Greek capital sigma / Summation	Sum over subframeworks
Π	U+03A0	Greek capital pi / Product	Product over attackers in the Li et al. formula
1	—	Indicator	1 if the condition holds, 0 otherwise
∝	U+221D	Proportional to	Posterior up to normalization
·	U+00B7	Middle dot	Multiplication
⇒	U+21D2	Rightwards double arrow	Implication in the epistemic postulates
≤	U+2264	Less-than or equal to	Upper bound in the coherence postulate
≥	U+2265	Greater-than or equal to	Lower bound in the optimism postulate
\|	—	Vertical line / Given	Conditioning bar

Metatheory. The constellations construction is well-defined for grounded semantics, since every subframework has exactly one grounded extension and P_gr is therefore a genuine probability; for preferred and stable it is not, because a subframework may have several extensions or none, and credulous and sceptical readings yield different numbers from the same distribution. The cost is counting rather than reasoning: Fazzinga, Flesca and Parisi (2013) proved a dichotomy for computing the probability that a set is an extension: polynomial time for the admissible and stable semantics, FP^#P-complete for the complete, grounded, preferred and ideal semantics. Li, Oren and Norman's product formula P(a IN) = Π P(b OUT) is exact only on attack graphs that are trees, since it presupposes the attackers' statuses independent, which any shared ancestor or cycle destroys. The epistemic postulates connect back to Dung exactly: coherence alone forces {a : P(a) > 0.5} to be conflict-free, and a 0/1 valuation satisfies coherence and optimism iff the set it marks out is a stable extension, foundedness then being automatic.

Applies to. Uncertain reasoning. Risk assessment. AI. Decision making.

Limitations. The two readings measure different things and there is no translation between them: a constellations number is the chance an argument survives an uncertain graph, an epistemic number is belief in its claim. Neither is determined by the framework. The constellations approach needs a joint distribution over subsets of A ∪ R but is normally given only marginals, so independence is assumed exactly where it fails — arguments sharing a premise stand or fall together — and the subframework distribution is then wrong rather than approximate. The epistemic postulates are satisfiable but far from categorical: coherence, foundedness and optimism leave a convex set of distributions and no principle inside the theory picks one out. Multiple-extension semantics leave the probability undefined where no stable extension exists and ambiguous where several do. Because attacks are primitive rather than derived, the probability of an argument bears no computed relation to the probabilities of its premises; that connection must be supplied by a further independence stipulation the framework cannot check.