Between Belief Bases and Belief Sets: Partial Meet Contraction Yuri D. Santos and Marcio M. Ribeiro and Renata Wassermann University of Sao˜ Paulo - Brazil [email protected] [email protected] [email protected] Abstract satisfied). Hence, in belief base contraction there is a dis- tinction between the sentences in the base and the sentences In belief revision, the result of a contraction should inferred from the base. Especially in computer science, be- be a logically closed set (closure) such that the in- lief base revision is very useful since computing the logical put is not inferred (success). Traditionally, these closure of a set may be hard if possible at all. two postulates depend on the same consequence Operations in belief bases, however, have some undesirable operator. properties due to the fact that equivalent sentences may be The present work investigates the consequences of treated differently. In the present work we explore the conse- using two different consequence operators for each quences of keeping both closure and success where the conse- of these desiderata: the classic Cn to compute suc- quence operations in these two postulates do not necessarily cess and a second weaker operator to compute the coincide. More precisely, we assume that closure is guaran- closure. We prove a representation theorem and teed only for a weaker consequence operation. This weaker some other properties for the new contraction. notion of consequence may be easy to compute and at the same time useful to avoid some undesirable consequences of 1 Introduction belief base operations. Throughout this paper, we consider the language of classi- Belief revision is the branch of knowledge representation that cal propositional logic, closed under the usual boolean con- deals with the dynamics of epistemic states. In their semi- nectives. nal paper [Alchourron´ et al., 1985], Alchourron,´ Gardenfors¨ We call consequence operator a function that takes sets of and Makinson proposed to represent the epistemic state of an formulas into sets of formulas. A consequence operator C is agent as a logically closed set of propositions called belief Tarskian if and only if it satisfies: sets. The authors focus on three main epistemic operations over belief sets: expansion, contraction and revision. Expan- (inclusion) A ⊆ C(A) sion is the simple addition of a proposition to the epistemic state, contraction is the removal of a proposition and revision (idempotence) C(A) = C(C(A)) is the consistent addition of a proposition. Each of this oper- ations is characterized by a set of rationality postulates. (monotonicity) If A ⊆ B then C(A) ⊆ C(B) Consider, for example, the following postulates for con- Cn denotes the classical consequence operator and ` de- traction: notes the associated relation: A ` α stands for α 2 Cn(A). (closure) K − α = Cn(K − α) Lowercase Latin letters (p, q, r) stand for atoms, lower- case Greek letters (α; β) stand for formulas, uppercase Latin (success) If α2 = Cn(;), then α2 = Cn(K − α) letters (A, B, K) stand for sets of formulas. The rest of the paper is structured as follows. A brief Closure guarantees the resulting epistemic state to be rep- overview of classical belief revision is given in Section 2. resented as a logically closed set of propositions, i.e., the re- Section 3 presents pseudo-contractions, the kind of operation sult of contracting a belief set is also a belief set. Success we are exploring in this paper. In Section 4, we describe our guarantees that the removed proposition α is no longer im- proposal, depicting it with examples in Section 5. Section 6 plied by the new belief set, unless it is a tautology. highlights some previous works that are related to ours. We Hansson in [Hansson, 1992a] suggests a generalization of delineate our conclusions in Section 7. Proofs for theorems the AGM framework where the epistemic state is not nec- and propositions are found in the Appendix. essarily closed under logical consequences. Although in his work epistemic states are represented as arbitrary sets of sen- tences called belief bases (i.e. closure is not necessarily sat- 2 AGM Paradigm isfied), the removal of a sentence in contraction is still eval- In the AGM paradigm [Alchourron´ et al., 1985], belief states uated against the closure of the belief base (i.e. success is are typically represented by sets of sentences closed under logical consequence, the so-called belief sets. Three change (uniformity) If for all B0 ⊆ B, α 2 Cn(B0) if and only if operations were initially defined: expansion, contraction and β 2 Cn(B0), then B − α = B − β revision. Expansion is the simple addition of a new sentence, Furthermore, we have the following representation theo- followed by logically closing the resulting set, i.e, K + α = rem: Cn(K [ fαg). In the case of contraction, we have a class of operations, all delimited by rationality postulates that they Theorem 4 [Hansson, 1992b] An operation is a partial meet should satisfy. Let K be a belief set and α and β be formulas. contraction over belief bases if and only if it satisfies the pos- The original AGM basic postulates for contraction are: tulates of success, inclusion, relevance and uniformity. (closure) K − α = Cn(K − α) For logically closed sets, both characterizations are equiv- alent (this is true for classical logics, for other cases, cf. (success) If α 62 Cn(;), then α 62 K − α [Ribeiro et al., 2013]) in the sense that all initial AGM pos- tulates are consequence of the postulates for bases [Hansson, (inclusion) K − α ⊆ K 1999]. (vacuity) If α 62 K, then K − α = K 3 Pseudo-Contractions Contraction operators on belief sets can be generated from (recovery) K ⊆ (K − α) + α operations on belief bases. As an example, one can define a contraction operator for belief sets as K ÷ α = Cn(B − α), (extensionality) Cn(α) = Cn(β) K −α = K −β If , then where K = Cn(B) and − is a base contraction operator. If − is a partial meet base contraction, ÷ satisfies five of the six Note that in the presence of closure, we can use α 62 K −α AGM postulates, but not recovery. or α 62 Cn(K − α) interchangeably in the success postulate. In [Hansson, 1989] a weakening of the inclusion postulate AGM revision is usually defined from contraction and ex- was proposed, called logical inclusion. pansion by means of the Levi identity: (logical inclusion) Cn(B − α) ⊆ Cn(B) K ∗ α = (K − :α) + α Hansson has suggested to call operations that satisfy suc- Therefore, in this paper we will focus on contraction. cess and logical inclusion pseudo-contractions. Besides defining rationality postulates for contraction, Al- Nebel has proposed a pseudo-contraction for bases that chourron´ et al.[1985] have also proposed a construction for generates a contraction that satisfies all the six AGM postu- a contraction operation. Partial meet contraction is based on lates [Nebel, 1989]. the notion of a remainder set, the set of all maximal subsets Definition 5 Let V B be the conjunction of all elements of that do not imply the element that is to be contracted. For- B. Nebel’s pseudo-contraction for the set B is the operator mally: − such that for all sentences α: Definition 1 Let B be a set and α a formula. The remainder set B?α is such that X 2 B?α if and only if: B if α 2 Cn(;) B − α = • X ⊆ B T γ(B?α) [ fα ! V Bg otherwise • X 0 α Although the belief set operation generated from Nebel’s • For all sets Y , if X ⊂ Y ⊆ B, then Y ` α pseudo-contraction satisfies all the AGM postulates, it adds unnecessary information to the base. As shown in [Ribeiro Definition 2 A function γ is a selection function for the set and Wassermann, 2008], it suffices to add fα ! V B0g, B if and only if: where B0 = BnT γ(B?α). As already noted in [Ribeiro and • If B?α 6= ; then ; 6= γ(B?α) ⊆ B?α Wassermann, 2008], there is no other intuition behind Nebel’s • Otherwise, γ(B?α) = fBg operation than maintaining recovery, a postulate which has been deemed as polemic already since the 80’s [Makinson, Definition 3 Let γ be a selection function for a set of sen- 1987]. tences B. The partial meet contraction of B by a sentence α T In this work we want to further explore the possibility of is given by B − α = γ(B?α). working with belief bases with logical inclusion, allowing for The partial meet operation previously defined can be ap- some syntax independence without having to resort to belief plied directly over belief bases, in such a way that it satisfies sets. the following postulates: (success) If α2 = Cn(;), then α2 = Cn(B − α) 4 Between Belief Sets and Belief Bases As stated earlier, the direct application of partial meet con- (inclusion) B − α ⊆ B traction over closed belief sets and over belief bases creates problems of practical (computational infeasibility) and the- (relevance) If β 2 B n (B − α), then there is a B0 such oretical (syntax dependence) nature, respectively. One of that B − α ⊆ B0 ⊆ B, α2 = Cn(B0), but α 2 Cn(B0 [ fβg) the aims of this study is to assess the effects of doing the traditional partial meet contraction on belief bases closed With a proof that is very similar to that of the representa- by a consequence operation that is between the classical tion theorem for partial meet contraction on bases (which can consequence operator and the identity (i.e., the base itself).