MP-Closure in Nonmonotonic Reasoning

• MP-Closure is a nonmonotonic reasoning formalism that refines defeasible inference by incorporating aspect-indexed preferences to handle exceptions.
• It combines rational closure with a lexicographic approach to enable selective, aspect-by-aspect inheritance of defaults in both Description Logics and propositional settings.
• MP-Closure provides a modular and skeptical reasoning framework that balances computational complexity with precise default inheritance in knowledge-based systems.

Multipreference Closure (MP-Closure) is a formal nonmonotonic reasoning mechanism designed for knowledge bases with defeasible conditionals, especially within Description Logics (DLs) and KLM-style propositional logics. It addresses the limitations of rational closure by incorporating aspect-indexed preferences and lexicographically refined default prioritization, enabling the inheritance of properties on an aspect-by-aspect basis even in the presence of exceptions. MP-Closure is formally situated between rational closure and lexicographic closure with respect to logical strength, and provides a sound, skeptical basis for reasoning under exceptions in knowledge representation systems (Giordano et al., 2018, Giordano et al., 2019).

1. Foundations and Formal Semantics

MP-Closure operates on knowledge bases containing two types of information: strict axioms and defeasible conditionals. Within DLs enriched with a typicality operator $T(\cdot)$, the knowledge base comprises concept inclusions (for strict assertions) and typicality inclusions of the form $T(C) \sqsubseteq D$, indicating that "typical $C$s satisfy $D$" (Giordano et al., 2018).

Enriched interpretations for MP-Closure model "aspects" (concepts on the right of some $T(\cdot) \sqsubseteq \cdot$) via a family of preference relations $\langle <_{A_1}, ..., <_{A_n}\rangle$ on the domain $\Delta$, one per aspect, in addition to a global preference $<$. These are irreflexive, transitive, modular, and well-founded. The interpretation evaluates $T(C)$ as the set of $<$-minimal elements of $C^I$.

A model satisfies the knowledge base if:

1. Strict inclusions hold extensionally ($C^I \subseteq D^I$).
2. For each $T(C)\sqsubseteq A$, both $<$-minimal and $<_{A}$-minimal $C$-elements are in $A$.
3. For strongly enriched (S-enriched) models, aspectwise preferences propagate to global for specificity: if $x<_{A_i}y$ for $A_i$ and no $j$ with $y<_{A_j}x$, then $x<y$; further, global preference orders individuals by the rank of the defaults they violate, with more specific violated defaults taking precedence.

Minimal canonical S-enriched models are selected by lex minimality: first, all $k_{A_i}(x)$ (aspect ranks) must be minimized componentwise, then global $k(x)$. Canonicality requires that every consistent concept conjunction is realized in the domain (Giordano et al., 2018).

2. MP-Closure Construction and Algorithmic Characterization

The multipreference closure relies on an overview of classical rational closure ranking and a lexicographically refined mechanism for identifying maximally compatible default sets (Giordano et al., 2018, Giordano et al., 2019).

Rational Closure Preliminaries

Defeasible inclusions are partitioned into levels $D_i$ based on their rational closure rank, determined by the iterated exceptionality of antecedents via a sequence $E_0, E_1, ...$ as in standard rational closure.

Maximal Compatible Sets and Entailment

Given a query $T(B)\sqsubseteq D$, compute $k = \mathrm{rank}(B)$. Consider all subsets $S$ of defeasible axioms satisfying:

• No axiom of rank $<k$ in $S$ is inconsistent with $B$ at level $E_k$.
• $S$ is maximal under the lex-inclusion order; for $S', S$ that differ at the greatest $i$, $S'\cap D_i \subset S\cap D_i$.

For each such maximal $S$, $T(B)\sqsubseteq D$ is entailed by the MP-Closure iff $E_k \cup$ materialized $S$ entails $T(B)\sqsubseteq D$ under ALC+TR semantics.

Pseudocode outline (Giordano et al., 2018):

procedure MP_Closure_Entails(KB, query T(B)⊑D): compute rational-closure levels E_0,…,E_n k ← rank(B) S_sets ← all maximal compatible S wrt < at level k for each S in S_sets: if not (E_k ∪ Materialize(S) ⊢_{ALC+TR} T(B)⊑D): return false return true

In the propositional setting, the analogous approach considers maximal MP-bases $D'\subseteq D$ for an antecedent $A$, where maximality is with respect to the MP-seriousness lex-ordering on sets of violated defaults (Giordano et al., 2019).

3. Comparative Analysis with Related Closures

The table below situates MP-Closure with respect to standard alternatives.

Closure Type	Strength Relative to MP	Distinguishing Feature
Rational Closure	Weaker	All-or-nothing inheritance; subset of MP consequences
Lexicographic Closure	Stronger	Compares by cardinality of violated defaults per rank
Relevant Closure	Weaker	MP strictly stronger

Rational closure enforces that exceptionality w.r.t. any default precludes inheritance of all defaults for that aspect, resulting in all-or-nothing outcomes. MP-Closure, in contrast, allows independent inheritance along aspects unless directly blocked. Lexicographic closure further strengthens this by counting the number of satisfied defaults at highest ranks, whereas MP-Closure discriminates only by set-inclusion at each rank, not cardinality (Giordano et al., 2018, Giordano et al., 2019).

Overriding-based approaches such as DEN rely on explicit normality and allow for KB-repair or inconsistency in case of unresolved conflicts; MP-Closure maintains skeptical reasoning and never collapses typicality to contradiction (Giordano et al., 2018).

4. Properties, Postulates, and Complexity

MP-Closure is preferential in the KLM sense (satisfying LLE, RW, REF, AND, OR, CM), but not rational—it fails Rational Monotonicity. Its semantic characterization is via the unique minimal MP-model ($N_{MP}$) obtained from minimal canonical rational closure models by refining $<$ to $<_{MP}$ according to violation tuples by rank (Giordano et al., 2019).

Entailment checking in the DL setting is EXPTIME-complete on ALC+TR due to the need to enumerate all maximal $S$, with possible exponential explosion in subset enumeration. In the propositional case, reasoning requires O($|D|$) classical entailment checks for ranking, but potentially explores exponentially many MP-bases. Deciding $K \models_{MP} A \rightsquigarrow B$ lies in the second level of the Boolean hierarchy ($\text{co-NP}^{\text{NP}}$) (Giordano et al., 2018, Giordano et al., 2019).

MP-Closure is not complete for minimal S-enriched semantics; some S-enriched entailments fall outside its scope due to further discrimination among defaults of equal rank via the global rank $k(\cdot)$.

5. Illustrative Example

Consider the propositional knowledge base:

• $d_1$: $Student \rightsquigarrow \neg PayTaxes$
• $d_2$: $Student \rightsquigarrow Young$
• $d_3$: $Employee \rightsquigarrow (Young \wedge PayTaxes)$
• $d_4$: $Employee \wedge Student \rightsquigarrow Busy$

with ranks: $d_1,d_2,d_3$ at 0; $d_4$ at 1.

For $A = Employee \wedge Student$ and $B = Young$, MP-bases are:

• $D'_1=\{d_1,d_2,d_4\}$ (consistent with $A$, maximally preferred)
• $D'_2=\{d_3,d_4\}$ (also consistent and incomparable with $D'_1$)

Both $D'_1 \cup \{A\}$ and $D'_2 \cup \{A\}$ entail $Young$, so $$K\models_{MP}(Employee\wedge Student)\rightsquigarrow Young$$.

For $B=\neg PayTaxes$:

• $D'_1 \cup \{A\}$ entails $\neg PayTaxes$
• $D'_2 \cup \{A\}$ does not (due to $d_3$ imposing $PayTaxes$)

Thus, $$K \not\models_{MP}(Employee\wedge Student)\rightsquigarrow \neg PayTaxes$$. This demonstrates the non-all-or-nothing aspect inheritance provided by MP-Closure (Giordano et al., 2019).

6. Practical Considerations and Extensions

Implementation challenges for MP-Closure revolve around the enumeration and pruning of maximal sets $S$ or MP-bases, each of which requires consistency and entailment tests. Approximation strategies, as in skeptical closure variants, may reduce complexity by working with prioritized single bases when the independence assumption is too strict. The framework can be extended to more expressive DLs (e.g., SHIQ, SROIQ) if rational closure ranking is obtainable (Giordano et al., 2018).

A plausible implication is that MP-Closure, by offering a fine-grained, aspect-sensitive approach, addresses practical requirements for modularity and cautious inheritance in ontological reasoning under exceptions, while maintaining soundness with respect to its intended minimal model semantics.

7. Summary and Significance

Multipreference Closure is a nonmonotonic reasoning paradigm that refines the inheritance of defeasible properties in knowledge bases by leveraging aspect-indexed preference relations and lexicographic comparison on the sets of violated defaults. It is technically situated strictly between rational closure and lexicographic closure, providing a skeptical, modular, and sound procedure for capturing nuanced patterns of exception and inheritance that classical all-or-nothing approaches cannot represent. MP-Closure's formal properties and complexity position it as a principled and robust option for ontology management and KLM-style default reasoning in both DLs and propositional settings (Giordano et al., 2018, Giordano et al., 2019).