Epistemic Description Logic

Origin. Donini et al. (1992). Combine DL with epistemic operators. Knowledge about ABox: K. Introspection: what is known. Foundation for ontology-based epistemic reasoning.

Models. DL knowledge bases with epistemic operator. K φ: φ is known (true in all models). Nonmonotonic via epistemic closed world. Multi-agent variants.

Formalism.

Epistemic operator: K C: concept C is known K-concepts: add K to DL concept language

Extended concepts: C ::= A | ¬C | C ⊓ D | ∃R.C | K C Epistemic on top of ALC (typically).

Semantics: Epistemic interpretation: set of DL models. E ⊨ K C(a) iff for all I ∈ E: I ⊨ C(a) "Known in all compatible models."

Epistemic closed world: K-minimal model: assume K φ for minimal φ. Nonmonotonic: closed-world on K.

Example: TBox: Employee ⊑ Person ABox: Employee(john) K Person(john): known (derivable) ¬K HasBonus(john): not known (open world)

Query: K ∃HasProject.Important(john)? "Is it known John has an important project?"

Complexity: Epistemic ALC: EXPTIME (same as ALC). With closed-world: higher complexity. Undecidable for some combinations.

Symbols.

Symbol	Unicode	Name	Meaning
K	—	Knows	Epistemic operator on concepts and roles
⊓	U+2293	Intersection	Concept and
⊑	U+2291	Subsumption	Subclass
⊨	U+22A8	Entails	Epistemic satisfaction in (I, E)
≡	U+2261	Equivalence	Collapse of iterated K
¬	U+00AC	Negation	Concept complement; ¬K is failure to know
∃	U+2203	Existential restriction	Some R-successor in C
∈	U+2208	Membership	I ∈ E: an interpretation in the epistemic state
::=	—	BNF definition	Grammar of epistemic concepts
\|	—	Alternation	Separates alternatives in the grammar
E	—	Epistemic state	Set of interpretations the agent cannot rule out
I	—	Interpretation	A single member of E
ALCK	—	System name	ALC extended with K

Metatheory. ALCK extends ALC with an operator K read over epistemic interpretations (I, E), where E is a set of interpretations sharing one fixed countably infinite domain of standard names, so that terms are rigid designators and distinct names denote distinct things. (K C) is interpreted as the intersection of the extensions of C across every member of E, which makes the accessibility relation universal and gives K the properties of S5: iterating collapses, K K C ≡ K C and ¬K C ≡ K ¬K C. Donini, Lenzerini, Nardi, Nutt and Schaerf (1992, 1998) showed the resulting logic decidable, and the argument is a reduction rather than a new model construction: since E is taken to be the models of the knowledge base, K C(a) holds exactly when the knowledge base entails C(a), so an epistemic concept is evaluated by a finite sequence of ordinary ALC entailment tests. The epistemic layer therefore adds nothing above ALC with a general TBox, which is ExpTime-complete. Requiring E to be maximal yields a nonmonotonic entailment relation over a classical ontology.

Applies to. Ontology queries. Incomplete knowledge. Data integration. Closed-world reasoning. Multi-agent ontologies.

Limitations. K is a single agentless operator with a universal accessibility relation, so it models exactly one reasoner: there is no agent index, no nesting of one agent's knowledge inside another's, and the S5 collapse means iteration adds nothing. Knowledge is identified with entailment from the knowledge base, so logical omniscience is not an idealization but the definition — everything the ontology entails is known, and an agent that fails to draw a consequence has no representation. The semantics needs a fixed domain of standard names and rigid designation, which imposes a unique name assumption that the underlying description logic does not make, so ALCK is not a conservative extension of a DL in which two names may denote the same individual. K permits closed-world queries without closing the world in the ontology: constraints written with K restrict answers, not models, so an ABox violating them stays consistent.