00 positive Bewertungen00 negative Bewertungen

15 Ansichten12 SeitenApr 23, 2008

© Attribution Non-Commercial (BY-NC)

PS, PDF, TXT oder online auf Scribd lesen

Attribution Non-Commercial (BY-NC)

Als PS, PDF, TXT **herunterladen** oder online auf Scribd lesen

15 Ansichten

00 positive Bewertungen00 negative Bewertungen

Attribution Non-Commercial (BY-NC)

Als PS, PDF, TXT **herunterladen** oder online auf Scribd lesen

Sie sind auf Seite 1von 12

IRIT / Universite Paul Sabatier CRIL / Universite d'Artois

118 route de Narbonne rue de l'Universite - S.P. 16

31602 Toulouse - France 62307 Lens - France

lang@irit.fr marquis@cril.univ-artois.fr

and Donini, 2000]). Inconsistency can also be dealt

In this paper, a fairly general framework for with. In this case, trivialization is avoided by weaken-

reasoning from inconsistent propositional be- ing the set of consequences that can be derived from

lief bases is dened. Variable forgetting is the belief base; this can be achieved either by con-

used as a basic operation for weakening be- sidering an approximation by below of classical entail-

liefs so as to restore consistency. The key ment (like in many paraconsistent logics, see [Besnard

notion is that of recoveries, which are sets and Hunter, 1998] [Hunter, 1998] for a survey) or by

of variables whose forgetting enables restor- weakening the input belief base while keeping classi-

ing consistency. Several criteria for dening cal entailment. This last technique is at work in so-

preferred recoveries are proposed, depending called coherence-based approaches to paraconsistent

on whether the focus is laid on the relative inference (see e.g. [Rescher and Manor, 1970] [Fagin

relevance of the variables or the relative en- et al., 1983] [Ginsberg, 1986] [Brewka, 1989], [Baral

trenchment of the pieces of information (or et al., 1991] [Pinkas and Loui, 1992] [Benferhat et al.,

both). Our framework encompasses several 1993]), as well as in belief merging (see e.g. [Libera-

previous approaches as specic cases, includ- tore and Schaerf, 1998] [Revesz, 1997] [Konieczny and

ing reasoning from preferred consistent sub- Pino Perez, 1998] [Lin and Mendelzon, 1999]).

sets, and some forms of information merging While some of these approaches take account for the

and belief revision. Interestingly, the gain relative importance of pieces of information, they do

in exibility and generality oered by our not handle the relative importance of variables in the

framework does not imply a complexity shift problem at hand. This is problematic in many situa-

compared to these specic cases. tions where some variables are less central than others,

especially when some variables are meaningful only in

the presence of others. Thus, it makes no sense to rea-

1 INTRODUCTION son about whether John's car is grey if there is some

strong con ict about whether John actually has a car.

Nontrivial processing of inconsistent sets of formulas is In order to address this issue, we dene in the follow-

an important topic in articial intelligence. Both the ing a new framework for reasoning from inconsistent

complexity of the problem and its signicance are re- propositional belief bases, using variable forgetting as a

ected by the number of approaches to paraconsistent basic operation for weakening beliefs. Belief bases are

reasoning that can be found in the literature under var- viewed as (nite) vectors of formulas, conjunctively in-

ious names, like paraconsistent logics, belief revision, terpreted. W.l.o.g., each formula is assumed to be is-

argumentative inference, information merging, model sued from a specic source of information. The key

tting, arbitration, knowledge integration, knowledge notion of our approach is that of recoveries, which

purication, etc. are sets of variables whose forgetting enables restor-

Corresponding to these approaches are many dier- ing consistency. Several criteria for dening preferred

ent mechanisms to avoid trivialization. For instance, recoveries are proposed, depending on whether the fo-

inconsistency can be removed by identifying wrong cus is laid on the relative relevance of the variables or

pieces of information through knowledge-gathering ac- the relative entrenchment of the pieces of information

(or both). Our framework oers several advantages noted 9V:, is a quantied boolean formula over PS ,

compared with many existing approaches to paracon- equivalent to a formula from PROPPS that can be in-

sistent reasoning. First of all, it is quite general ductively dened as follows:

and exible. Especially, it enables to model situations

where some sources of information are considered more 9;: ;

reliable than others in an absolute way, but also rel- 9fxg: x 0 _ x 1 ;

atively to some topics of interest. Equity between 9(fxg [ V ): 9V:(9fxg:).

some sources of information can also be achieved by

imposing to forget the same variables in the sources. For example, 9fag::a ^ b b and 9fag:(a _ b) >.

Accordingly, our framework encompasses several pre-

vious approaches as specic cases, including reasoning 9V: represents the logically strongest consequence

from preferred consistent subsets, and forms of infor- of that is independent of V , which means that there

mation merging and belief revision. Interestingly, the exists a formula 0 from PROPPS s.t. 0 and

gain in generality and exibility oered by our frame- V ar( 0 ) \ V = ;. Accordingly, forgetting a set of vari-

work does not imply a complexity shift compared with ables within a formula leads to weaken it. To be more

these specic cases. precise, if V W holds, then 9V: j= 9W: holds.

Moreover, is consistent if and only if 9V ar(): is

The rest of this paper is organized as follows. Formal valid (see [Lang and Marquis, 1998]).

preliminaries including a presentation of forgetting are

given in Section 2. Our framework is presented in Sec- Many characterizations of variable forgetting, together

tion 3. Computational issues are discussed in Section with complexity results, are reported in [Lang et al.,

4. Section 5 concludes the paper. Proof sketches can 2001]. For instance, for every V PS and ev-

be found in an appendix. ery formula , we clearly have 9V: 9V :, where

V = V \ V ar().

2 FORMAL PRELIMINARIES 3 REASONING FROM

2.1 PROPOSITIONAL LOGIC PREFERRED RECOVERIES

PROPPS denotes the propositional language built up 3.1 BELIEF BASES AND RECOVERIES

from a nite set PS of symbols, the Boolean constants

> (true) and ? (false), and the usual connectives. Let us rst make precise the belief bases we are inter-

V ar() denote the set of propositional variables oc- ested in.

curring in the formula . x 0 (resp. x 1 ) denotes Denition 2 (belief base) A belief base B is a

the formula obtained by replacing in every occur- vector h1 ; : : : ; ni of n formulas from PROPPS ,

rence of variable x by ? (resp. >). where n is a positive integer.

An interpretation ! is an assignment of a truth value

to each variable of PS.

= 2PS is the set of all inter- B = h1; : : : ; ni is conjunctively interpreted, so that

pretations. Formulas are interpreted in the classical it is said to be inconsistent if and only if f1; : : : ; ng

way. Every nite set of formulas is interpreted con- is inconsistent. Each i (1 i n) identies a source of

junctively. Mod() denotes the set of models of . Fi- information and i denotes the piece of belief conveyed

nally, ! and !0 being two interpretations, Diff(!; !0 ) by source i. Note that it can be the case that a formula

is the set of propositional variables assigned dierent occurs more than once in B, which can be used to

truth values by ! and !0 . model the situation where two dierent sources (or

more) give the same information.

2.2 FORGETTING The key notions of our approach are forgetting vectors

and recoveries. A forgetting vector consists of a collec-

Our approach to restore consistency is based on vari- tion of sets of variables to be forgotten in each formula

able forgetting, also known as projection or marginal- from the belief base. These sets of variables need not

ization. Variable forgetting can be dened as follows be identical, but they should obey some constraints

(see [Lin and Reiter, 1994] [Lang and Marquis, 1998] bearing on the forgetting process:

for more details):

Denition 3 (forgetting context) A forgetting

Denition 1 (forgetting) Let be a formula from context C for a belief base B = h1 ; : : : ; ni is a triple

PROPPS and V PS . The forgetting of V in , hF; G; H i where:

F = hFi ii=1:::n where, for each i, Fi PS is the be blue without forgetting whether it should be red: ei-

set of variables that can be forgotten in i (i.e., ther we forget about the colour of the swimming pool

variables in PS n Fi must not be forgotten in i); or not. Similarly, either we forget about the expenses,

i.e., we forget fc0; c1; c2g or not (i.e., we forget none

G = hGi ii=1:::n where, for each i, Gi is a directed of them).

graph on Fi ; an arc (v; v0 ) in Gi means that v0 is

meaningful only in the presence of v, i.e., that if Clearly enough, the preferences of the group are

v is forgotten in i then v0 must be forgotten as jointly inconsistent. This scenario can be encoded in

well in i; our framework using the following belief base B =

h1 ; 2; 3; 4; 5i and forgetting context C :

H PS f1 : : :ng f1 : : :ng is a set of triples 1 = (s ) (sr sb )) ^ (:s ) :sr ^ :sb )

(v; i; j) where v 2 Fi \ Fj and i =

6 j , meaning that ^(c2 , (s ^ t)) ^ (c1 , (s t)) ;

if v is forgotten in i, then it must be forgotten as ^(c0 , (:s ^ :t))

well in j .

2 = (c0 _ c1) ^ (s ) sr );

F, G and H impose some constraints over the way the 3 = (s _ t) ^ (s ) sb );

pieces of belief i from B can be weakened. Thus, if 4 = s;

x 62 Fi , forgetting x in i is forbidden. This is helpful 5 = s ^ t;

to model the situation where source i is fully reliable F1 = ;, F2 = F3 = F4

to what concerns x, i.e., what i says about x must be = F5 = fs; t; sr ; sb ; c0; c1; c2g;

taken for sure. Gi enables to express that variables can G1 = ;, G2 = G3 = G4 = G5

be signicant only in presence of others so forgetting = f(s; sr ); (s; sb ); (sr ; sb); (sb ; sr );

the latter imposes to forget the former. Finally, H (c0; c1); (c0 ; c2); (c1; c0);

can be used to force some sources of information to (c1; c2); (c2 ; c0); (c2; c1)g;

be considered on equal terms w.r.t. weakening. For H = ;.

instance, if i and j are together inconsistent and

consistency can be recovered by forgetting x in i, then Given that F1 = ;, 1 expresses the integrity con-

it could be expected that x should also be forgotten in straint according to which the swimming pool must

j .

be either red or blue if it is constructed, as well

Example 1 (inspired from [Konieczny and Pino as the logical denition of the induced price from

Perez, 1998]). As a matter of illustration, let us con- the number of equipments built. 2 (resp. 3,

sider the following preference merging scenario. Sup- 4, 5) encodes the preferences of the co-owners.

pose that a group of four co-owners of a residence tries G2 (= G3 = G4 = G5) ensures that sr and sb are ir-

to agree about building a new tennis court t and/or a relevant if there is no agreement on s, that sr is irrele-

new swimming pool s. If it is constructed, the swim- vant if and only if sb is, and that for any i; j 2 f0; 1; 2g,

ming pool can be either red (sr ) or blue (sb ). If both ci is irrelevant if and only if cj is. In the situa-

a tennis court and a pool { respectively, one of them, tion where all co-owners must be considered on equal

none of them { are (is) constructed, the induced cost is terms w.r.t. the set of variables to be forgotten, H

2 money units (c2 ) { respectively, 1 unit (c1), nothing must be changed to (fsr ; sbg f2; 3; 4; 5g f2; 3; 4; 5g)

c0 ). The rst co-owner would not like to spend more [(fc0; c1; c2g f2; 3; 4; 5g f2; 3; 4; 5g).

than 1 unit, and prefers a red swimming pool, should it As expected, our denition is very general and many

be constructed. The second co-owner would like to im- dierent forgetting contexts can be considered. In

prove the quality of the residence by the construction of some situations, the variables forgotten in each of the

at least a tennis court or a swimming pool and would pieces of information must be identical so that all

prefer a blue swimming pool, should it be constructed. sources of information are considered on equal terms.

The third co-owner just prefers a swimming pool to be This can be captured by considering homogeneous con-

built, whatever its colour. The fourth co-owner would texts, where a forgetting context C = hF; G; H i for

like both a swimming pool (of whatever colour) and a B = h1 ; : : : ; ni is said to be homogeneous if and

tennis court to be built. only if for every i, j 2 1 : : :n, we have Fi = Fj and

Of course, if there is no agreement about whether the Gi = Gj , and H = PS f1 : : :ng f1 : : :ng. A sim-

swimming pool is to be constructed, any preference ple case when C should not be homogeneous is when

concerning its colour must be ignored. Furthermore, it some of the pieces of information must be left intact:

is meaningless forgetting about whether the pool should for such a i, we set Fi = ;. This is useful to encode

integrity constraints (or revision formulas), i.e., for- fact that the ith source of information is fully unre-

mulas that are required to be true (just like 1 in the liable, so it should not be taken into account at all

example above). Another non-homogeneous forgetting (hence the corresponding i should be removed from

context for B = h1 ; : : : ; ni is its standard forgetting B).

context, noted CS and given by Fi = PS, Gi = ; for

every i 2 1 : : :n, and H = ;. In this context, every 3.2 PREFERRED RECOVERIES

variable can be forgotten in any piece of belief, and

variables can always be forgotten in an independent Generally, many recoveries for a belief base given a

way. forgetting context are possible, but most of the time,

We can now dene our notion of recovery: some of the corresponding projections are more ex-

pected than others. Discriminating them calls for a

Denition 4 (recovery) Let B = h ; : : : ; ni be notion of preferred recovery. Many preference criteria

1

a belief base and C be a forgetting context for B . A can be considered so as to capture in formal terms sev-

recovery vector (or recovery for short) V~ for B given eral intuitions about the way the pieces of information

C is a vector V~ = hV1 ; : : : ; Vni of subsets of V ar(B)1 should be merged.

s.t.: For instance, we may prefer the recoveries V~ =

hV1 ; :::; Vni for B = h1 ; : : : ; ni in which the set Vi of

for every i 2 1 : : :n, Vi Fi , and variables forgotten in some of the i's are as close to

each other as possible2, and eventually coincide (which

for every i 2 1 : : :n, for any (v; v0 ) 2 Gi, v 2 Vi always is the case whenever homogeneous contexts are

implies v0 2 Vi , and considered). We may also prefer recoveries that lead

for every (v; i; j) 2 H , v 2 Vi implies v 2 Vj , and to forget minimal sets of variables, where minimality

nally may be dened using a preference criterion induced by

some priorities or penalties on the set of variables. Fi-

the set B j V~ = f9Vi :i j i 2 1 : : :ng is consistent. nally, we may prefer recoveries that preserve as much

as preferred beliefs as possible (like in the coherence-

B j V~ is called the projection of recovery V~ on B given based approaches to paraconsistent reasoning).

C . RC (B) denotes the set of all recoveries for B given In order to capture these various preferences in formal

C . B is said to be recoverable w.r.t. C if and only if terms, RC (B) can be equipped with a preordering (i.e.,

RC (B) = 6 ;. a re exive and transitive relation) v. For every vector

Example 1 (continued) Here is a list of recoveries V~ , V~ 0 of length n, V~ p V~ 0 means that Vi Vi0 for

for B given C . For each recovery V~ and every i 2 each i 2 1 : : :n.

1 : : :5 we compute 9Vi :i, and nally B jV~ . Denition 5 (preference relation) Given a be-

All the vectors V~ 1 to V~ 9 considered on the table 1 are lief base B = h1 ; : : : ; ni and a forgetting context C

recoveries for B given C , while the following ones are for it, a preference relation is a re exive and transi-

not recoveries for B given C : tive relation v on RC (B) satisfying the monotonicity

property:

h;; ftg; ftg; ftg; ftgi, because 9V1 :1 ^ 9V2 :2 ^

9V3 :3 ^ 9V4 :4 ^ 9V5 :5 is inconsistent; 8V~ , V~ 0 2 RC (B), if V~ p V~ 0 then V~ v V~ 0.

h;; fsg; fsg; fsg; fsgi, because it does not satisfy We note V~ ⊏ V~ 0 for V~ v V~ 0 and V~ 0 6v V~ , and V~ V~ 0

the constraints of C . for V~ v V~ 0 and V~ 0 v V~ .

By construction, replacing the beliefs of B by the pro- It is easy to prove that for any V~ , V~ 0 2 RC (B), if

jection on B of any of its recoveries is sucient to V~ p V~ 0 then B j V~ j= B j V~ 0 holds. Accordingly,

restore consistency, provided that B is recoverable. In among the recoveries from RC (B), the minimal ones

the general case, it may happen that RC (B) is empty, w.r.t. v = p lead to projections that preserve as

because some variables must not be forgotten. An- much information as possible given C . This is helpful

other important sucient condition for a belief base to represent some inertia principle in our framework.

not to be recoverable is when it incorporates some in-

consistent i. In some situations, this can re ect the

2

For instance, if both V~ = hfag; fb; cgi and V~ = 0

1 We note V ar (h ; : : : ; i) = Sn V ar ( ).

hfag; fa; cgi are recoveries for B , V~ should (in some con-

0

Table 1:

recovery 9V1 :1 9V2 :2 9V3 :3 9V4 :4 9V5:5 B j V~

V~ = h;; fs; sb; sr g; fs; sb; sr g;

1

1 c0 _ c1 > > t 1 ^ :s ^ t

fs; sb ; sr g; fs; sb; sr gi

V~ 2 = h;; ft; sb; sr g; ft; sb; sr g; 1 c0 _ c1 > s s 1 ^ s ^ :t

ft; sb; sr g; ft; sb; sr gi

V~ 3 = h;; fc0; c1; c2; sb ; sr g;

fc0 ; c1; c2; sb; sr g; 1 > s_t s s^t 1 ^ s ^ t

fc0 ; c1; c2; sb; sr g;

fc0; c1; c2; sb ; sr gi

V~ 4 = h;; fsb; sr g; fsb; sr g; ;; ftgi 1 c0 _ c1 s_t s s 1 ^ s ^ :t

~V 5 = h;; fc0; c1; c2g; fsb; sr g; ;; ;i 1 s ) sr s_t s s^t 1 ^ s ^ t ^ sr

V~ 6 = h;; fsb; sr g; ;; ;; ftgi 1 c0 _ c1 (s _ t) ^ (s ) sb ) s s 1 ^ s ^ :t ^ sb

V~ 7 = h;; ;; ;; fs; sb; sr g; fs; sb; sr gi 1 (c0 _ c1)^ (s _ t) ^ (s ) sb ) > t 1 ^ :s ^ t

(s ) sr )

V~ 8 = h;; fc0; c1; c2; sb; sr g; ;; ;; ;i 1 > (s _ t) ^ (s ) sb ) s s^t 1 ^ s ^ t ^ sb

V~ 9 = h;; ;; fsb; sr g; ;; ftgi 1 (c0 _ c1)^ s_t s s 1 ^ s ^ :t ^ sr

(s ) sr )

symmetric, total function H : INn ! IN s.t. 0(;) = 0

V~ . Accordingly, B j V~ is at least as logically strong

4 6 and 8X, Y PS, if X Y , then 0 (X) 0 (Y ), and

as B j V~ .

4 H(0; :::; 0) = 0, and (V~ ) = H(0 (V1 ); :::; 0(Vn )).

Depending on the problem at hand, many other prop- We are now in position to dene the family of inference

erties on v can be imposed so as to capture various relations jCv that can be dened in our framework by

intuitions about the result of the merging process. letting C and v to vary. We call them forget inference

relations.

Thus, given B = h1; : : : ; ni, a preference relation v

is said to satisfy Denition 6 (skeptical forget inference) Given

a belief base B and a forgetting context C for it, let v

the binaricity property if and only if 8V~ , V~ 0 2 be a preference relation on RC (B) (possibly induced by

RC (B), if (for every i 2 1 : : :n, (Vi 6= ; , Vi0 6= a ranking function ).

;)), then V~ V~ 0 .

A recovery V~ 2 RC (B) is said to be preferred

the gathering property if and only if 8V~ 2 RC (B), (w.r.t. v) if and only if it is minimal in its

~V hS V~ ; : : :; S V~ i, where S hV1 ; : :~: ; Vni = set w.r.t. v, i.e., there is no V~ 0 2 RC (B) s.t.

VS1 [ : : : [S Vn (obviously, it is required that V~ 0 ⊏ V~ .

h V~ ; : : :; V~ i 2 RC (B) for this notion to make Let be any formula from PROPPS . is said

sense). to be (skeptically) inferred from B w.r.t. v, de-

noted by B jCv , if and only if for any preferred

In many cases, it is desirable to assume that v is recovery V~ (w.r.t. v), we have B j V~ j= .

a complete preordering. In this situation, we can

equivalently represent v by a ranking function from Since skeptical inference is considered, among the pre-

RC (B) to IN such that (h;; ;; :::; ;i) = 0, and 8V~ , ferred recoveries for B, only the maximal ones w.r.t.

V~ 0 2 RC (B), if V~ p V~ 0 , then (V~ ) (V~ 0). The p are relevant for inference, in the sense that the

preference relation v induced by is the total pre- other ones can be ignored without modifying the in-

ordering dened by 8V~ 2 RC (B), V~ v V~ 0 if and only ference relation jCv . Indeed, such maximal elements

if (V~ ) (V~ 0). We say that a ranking function is correspond to the logically weakest projections.

anonymously decomposable if and only if there exists

a total function 0 : 2PS ! IN and a non-decreasing, Denition 7 (prudent forget inference) Given

a belief base B and a forgetting context C for it, let that the set of all maxcons inference relations is con-

v be a preference relation on RC (B) (possibly in- tained in the set of forget inference relations, and

duced by a ranking function ). Let be any for- therefore that the latter family is at least as general as

mula from PROPPS . is a prudent consequence the former. In particular, if preference over recoveries

of B w.r.t. v, denoted B jCvprudent , if and is dened by 8V~ , V~ 0 2 RC (B), V~ v V~ 0 if and only

if fi j Vi 6= ;g fi j Vi0 6= ;g then B jCv if and

S

only if B j V~prudent

v

j= , where V~prudent

v

is a re-

S

hSV~ 2min(R (B);v) V1 ; : : : ; SV~ 2min(R (B);v) Vni of all

covery for B , dened as the the pointwise union

clusion) consistent subset S of B. Other criteria such

preferred recoveries for B w.r.t. v (especially, the re-

C C

that maximum cardinality, \discrimin" or \leximin",

lation does not hold whenever V~prudent

v

does not satisfy

or minimum penalty, can be recovered as well. The

the constraints imposed by C ).

assumption that each i (1 i n) in B is consistent

is necessary (and sucient) to ensure that B is recov-

Example 1 (cont'd) If v = p is chosen as erable given its standard forgetting context. It can be

a preference relation, we have min(RC (B); v) =

made w.l.o.g., just because B and B n fi j 1 i n

and i is inconsistentg have the same consistent sub-

fV~ 5 ; V~ 6; V~ 7 ; V~ 8; V~ 9 g. Accordingly, V~prudent

v

= sets. As an immediate consequence of Proposition 1,

h;; fc0; c1; c2; sb; sr g; fsb; sr g; fs; sb; sr g; fs; sb; sr ; tgi. our framework encompasses other important frame-

We obtain B jCv 1 ^(s_t) and B jCvprudent 1 ^(s_t). works, like supernormal default theories with priori-

ties [Brewka, 1989] and syntax-based belief revision

3.3 ON THE GENERALITY OF OUR [Nebel, 1992], as specic cases.

FRAMEWORK

We now show how several well-known paraconsistent 3.3.2 Belief Merging

inference relations belong to the family of forget rela- A merging operator (see e.g. [Liberatore and Schaerf,

tions. 1998] [Revesz, 1997] [Konieczny and Pino Perez, 1998]

[Lin and Mendelzon, 1999]) maps any nite collec-

3.3.1 Reasoning from Preferred Consistent tion (multiset) of consistent propositional formulas

Subsets B = f1; : : : ; ng, with n being unxed, into a new

Let us rst recall the denition of inference drawn from consistent propositional formula Merge(1 ; : : : ; n).

preferred consistent subsets, also called maxcons in- When n is xed, we dene a n-merging operator

ference relation (see e.g. [Rescher and Manor, 1970] Mergen (1 ; : : : ; n) as the restriction of Merge to

[Fagin et al., 1983] [Ginsberg, 1986] [Brewka, 1989] n-uples of formulas. In order to simplify nota-

[Baral et al., 1991] [Pinkas and Loui, 1992] [Benferhat tions, we will write Merge(1 ; : : : ; n) instead of

et al., 1993]). Let B = f1; : : : ; ng be a nite set of Mergen (1 ; : : : ; n) in the following.

formulas s.t. each i (1 i n) is consistent, and A special class of n-merging operators is the class of

a preference pre-ordering on subsets of B s.t. 8X; decomposable merging operators induced by (pseudo-

Y B, if X Y , then X Y 3. Then B j8 if and )distances. Let d :

! IN be a total func-

only if for any preferred (w.r.t. ) consistent subset tion satisfying 8!, !0 2

, d(!; !0) = 0 if and

X of B, we have X j= . only if ! = !0 , and d(!; !0 ) = d(!0 ; !). Let

be a total function from INn to IN, symmetric and

Proposition 1 j8 is a maxcons inference relation monotonic in each of its arguments. For every B =

if and only if there exists a preference relation v f1; : : : ; ng s.t. each i (1 i n) is consis-

on RC (B) satisfying binaricity and s.t. for any , tent, the decomposable n-merging operator Merged

f ; : : : ; ng j8 if and only if h ; : : : ; ni jCv . induced by d and is dened by Merged (B) =

S

S

1 1

f! 2

j d(!; B) is minimalg, where d(!; B) =

The rst direction of Proposition 1 expresses that un- (d(!; 1); : : : ; d(!; n)) when B = f1 ; : : : ; ng and

der the assumption of binaricity and standard forget- d(!; i) = min! j=i d(!; !0). The n-merging inference

0

ting context, inference from preferred recoveries comes relation jM d; induced by d and is dened by B jM d;

down to inference from preferred maximal consistent if and only if Merged (f1 ; : : : ; ng) Mod( ).

subsets. The other direction of Proposition 1 states In order to characterize the decomposable n-merging

3

Among usual preference relations, we nd set inclusion: inference relations as forget relations, we need to fo-

X Y if and only if X Y , and cardinality: X Y if cus on the dierential ones, i.e., those based on a dif-

and only if jX j jY j. ferential distance, where a distance d is said to be

dierential if and only if there exists a total func- Proposition 3 Let C be the forgetting context for

tion f : 2PS ! IN satisfying (i) f(;) = 0 and (ii) h ; i s.t. F = G = G = H = ; and F = PS .

1 1 2 2

8A,B PS, if A B, then f(A) f(B), and s.t. Let vS (resp. vD ) be a preference relation on RC (B)

8!, !0 2

, d(!; !0) = f(Diff(!; !0 )). s.t. 8V~ , V~ 0 2 RC (B), V~ vS V~ 0 if and only if V V 0

2 2

(resp. V~ vD V~ 0 if and only if jV j jV 0 j). Provided

Proposition 2 f ; : : : ; ng jMd; is a decomposable

2 2

1 that is consistent, we have:

and dierential n-merging inference relation if and

only if there exists a complete preference relation v for any , S j= if and only if h ; i jCvS ;

on RC (B), induced by a ranking measure satisfy-

S

f ; : : : ; ng jM

1 d; if and only if h ; : : : ; ni jv

1

S C for any , D j= if and only if h ; i jCvD ;

The rst direction of Proposition 2 expresses that un- for any , W j= if and only if

der the assumptions of completeness of v , anonymous h ; i jCvprudent

S .

decomposability, and standard forgetting context, for-

get inference comes down to inference from a merging 3.4 And Many Others!

operator. The other direction expresses that a partic-

ularly interesting subclass of inferences from merging Up to now, we showed in this section that reason-

operators is contained in the set of forget inference ing from preferred consistent subsets and some forms

relations. In particular, usual arbitration and major- of belief merging and of belief revision are particu-

ity merging operators (see [Konieczny and Pino Perez, lar cases of forget inference relations. Now, there are

1998]) are recovered by letting (1) i (A) = jAj for each many other forget inference relations which do not be-

i 2 1 : : :n, which implies that the induced distance d long to any of the previous families (provided that no

is the Hamming distance between interpretations, and new propositional symbols are added).

(2) = max or = +, respectively.

A rst example is the family of homogeneous forget

3.3.3 Belief Revision inference relations. Such relations can be obtained

by requiring the forgetting context to be homogenous,

Our framework also includes some approaches to or requiring the preference relation v to satisfy the

model-based4 belief revision, including Satoh's S gathering property (which is less demanding). When

[Satoh, 1988] (or equivalently, the operator given in homogeneous inference relations are considered, it is

[Delgrande and Schaub, 2000]), Dalal's D [Dalal, enough to dene a preference relation on subsets of

1988] and Weber's W [Weber, 1986], provided that PS (instead of vectors of subsets of PS). Many such

the belief set under consideration is not PROPPS (i.e., relations can be obtained by letting the preference rela-

it can be encoded by a consistent formula ). tion on 2PS vary, in quite the same manner as maxcons

inference relations are obtained by letting the prefer-

Given two formulas and , we have: ence relation on subsets vary. We may for instance

Diff min ( ; ) = min(fDiff(!; !0 ) j ! 2 Mod( ); minimize theSset of Sforgotten variables (i.e., V~ v V~ 0 if

!0 2 Mod()g; ); and only if V V~ 0), or minimize the number of

~

jjDiff min ( ; )jj = min(fjDiff(!; !0 )j j ! 2 forgotten variables, or, more generally, make use of a

Mod( ); !0 2 Mod()g; ); predened penalty function or priority preordering on

Mod( S ) = f! 2 Mod( ) j 9!0 2 Mod() s.t. variables.

Diff(!; !0 ) 2 Diff min ( ; )g;

Mod( D ) = f! 2 Mod( ) j 9!0 2 Mod() s.t. Here is a variety of conclusions that can be achieved

jDiff(!; !0 )j = jjDiff min ( ; )jjg. through a judicious choice of v on Example 1.

Mod( W ) = f! 2 Mod( ) j 9!0 2 Mod() s.t. Example 1 (continued) Consider the following

S! Diff

and !0 coincides except possibly on variables from

min ( ; )g. preference relations and the induced inference relations

jCv . We dene the mapping k : 2V ar B ! IN by

( )

k = kc + ks + kt where kc (V ) = 3 if fc ; c ; c g V ,

0 1 2

The inference relations induced by S , D and W are kc (V ) = 0 otherwise ; ks (V ) = 3 if fs; sb; sr g V ,

forget ones since: ks(V ) = 1 if V \ fs; sb; sr g = fsb ; sr g, and ks(V ) = 0

otherwise ; kt(V ) = 2 if t 2 V , kt(V ) = 0 otherwise.

4

As evoked previously, it incorporates as well the usual

syntax-based ones. v1 =p ;

What should be concluded from such inconsistent

X

5 X

5

bases, especially when no information about the (abso-

v2 : V~ v2 V~ if and only if

0 k (V i ) k(Vi )

0

i=1 i=1

are available? There is no absolute answer to this

v3 : V~ v3 V~ 0 if and only if S V~ S V~ 0;

question! The result depends on various assumptions

on the merging process, and the variety of conclusions

v4 : V~ v4 V~ 0 if and only if k(S V~ ) k(S V~ 0 );

obtained through many paraconsistent inference rela-

tions, as reported in the table, exemplies it.

v5 : V~ v5 V~ 0 if and only if The very important point here is that our framework

fi j Vi =6 ;g fi j Vi0 6= ;g. enables to capture each of them:

The results are synthesized in table 2. v1 is the point- v1 corresponds to inference from maximal (w.r.t.

wise inclusion relation. v2 is a merging inference re- or cardinality) consistent subsets.

lation. v5 is a maxcons inference relation. v3 and v2 (resp. v3 ) corresponds to distance-based merg-

v4 correspond to homogeneous inference relations: v3 ing where d is the Hamming distance and = +

minimizes the set of variables forgotten everywhere (resp. = max).

while v4 minimizes their cumulated cost.

v4 = the closure of p under gathering.

A second family of forget inference relations which

does not degenerate into one of the previous types of v5 = p under the standard forgetting context.

inference is obtained in the situation where reliability

between sources is a matter of topic5, as explained in Compared with the weakening by inhibition mecha-

the following example: nism at work in the coherence-based approaches, weak-

ening by variable forgetting is typically less drastic; in-

Example 2 Let be B = h ; i, with = a^b, =

1 2 1 2 deed, instead of inhibiting a whole formula { or equiv-

:a^:b. Let us consider the standard forgetting context alently, replacing it by > { it is possible to keep all its

for B . Assume that source 1. is more reliable to what consequences that are not involved in any contradic-

concerns a than to what concerns b and conversely, tion (see e.g. the piece of information c in Example

source 2. is more reliable to what concerns b than to 3). Subsequently, more information can be preserved.

what concerns a. To capture it in our framework, the

following ranking function can be considered (among 4 COMPUTATIONAL ISSUES

P (hV1; V2i) = 0(V1) + 0(V2)

other possible choices):

where 0 (Vi ) = v2Vi ki (v) and k1(a) = 2, k1(b) = 1,

k2(a) = 1, k2 (b) = 2. Then the (unique) preferred We now investigate the computational complexity of

recovery for B is hfbg; fagi and therefore we have forget inference relations. We suppose the reader fa-

B jCv a ^ :b. miliar with computational complexity (see e.g. [Pa-

padimitriou, 1994]), and especially the complexity

Finally, in order to illustrate the
exibility oered by classes NP, p2 and p2 of the polynomial hierarchy.

our framework, let us consider the two following bases: We have obtained the following results:

Example 3 Let B = h ; i, and let B 0 = h0 ; 0 i,

1 2 1 2

Proposition 4

with: 1. Given a belief base B and a forgetting context C

1 = (a _ b) ^ c 1 = a ^ b

0

for it, determining whether B is recoverable is NP-

2 = :a ^ :b 2 = :a ^ :b

0

complete.

v1 1 _ 2 a,b 2. Provided that the preference relation v on RC (B)

v2 (:a _ :b) ^ c > can be tested in polynomial time, the inference

v3 (a _ b _ c) ^ (:a _ b) a 6, b

v4 c ^ (:a _ :b) > problem associated to jCv is in p2 .

v5 c ^ (:a _ :b) > 3. If moreover v is induced by a ranking function

computable in polynomial time, the inference

5 Here, a topic simply is a propositional variable x and

problem associated to jCv is in p2.

a source i conveys some information about it as soon as

i is not independent of x, i.e., i 6 9fxg:i . For a very

dierent approach where topics are taken into account, see These results show that the gain in generality and

[Cholvy, 1995].
exibility oered by our framework does not induce

Table 2:

preference relation preferred recoveries j

that are maximal w.r.t. p

C

v

v1 V~ 5 , V~ 6 , V~ 7 , V~ 8 , V~ 9 1 ^ (s _ t)

v2 V~ 6 , V~ 9 1 ^ (s ^ :t)

v3 V~ 1 , V~ 2 , V~ 3 s_t

v4 V~ 1 , V~ 2 , V~ 6 1 ^ (s t)

v5 h;; ;; ;; F4 ; F5 i, h;; ;; F3; ;; F5i, h;; F2 ; ;; ;; ;i 1 ^ (s _ t) ^ (s ^ t ) sb ) ^ (s ^ :t ) sr )

approaches (such as maxcons inferences and forms of

belief revision), that are already p2 -complete. This 2. If moreover the set fi j 1 i ng of formu-

holdsp as well for distance-based belief merging which las from B belongs to a class of formulas that are

is 2-complete in the general case (see [Liberatore and tractable for clausal entailment and stable for new

Schaerf, 2000] [Konieczny et al., 2002]), while the prob- variable renaming6 , the inference problem asso-

lem of forget inference with a complete preference re- ciated to jCv (or jCvprudent ) is in P when CNF

lation induced by a polytime ranking function is not queries are considered.

harder. Hardness results for our forget inference rela-

tions (in the general case) can be easily obtained from Interestingly, several well-known classes that are

known results for maxcons inference relations, belief tractable for the satisability problem are also stable

revision or merging (see e.g. [Eiter and Gottlob, 1992] for new variable renaming; For example, this is the

[Cayrol et al., 1998] [Nebel, 1998] [Konieczny et al., case for the class of Krom formulas (CNF formulas

2002]), thanks to the forget-based characterizations of consisting of binary clauses) and for the class of CNF

such approaches, as reported in Section 3.3. Horn renamable formulas.

In the general case (especially, when v is not induced

by a ranking function), our complexity result show

that two independent sources of complexity must be 5 CONCLUSION

dealt with. One of them lies in the possibly expo- We have proposed a new framework for reasoning from

nential number of preferred recoveries which must be inconsistent belief bases, which proceeds by selecting

taken into account. The other one originates from subsets of variables to be forgotten in the pieces of in-

the complexity of classical entailment in propositional formation, respecting some xed constraints. We have

logic. shown that this framework encompasses many existing

One way to circumvent this intractability consists in frameworks as specic cases, namely reasoning from

compiling the belief base [Cadoli and Donini, 1997] preferred consistent subsets, forms of belief merging

[Coste-Marquis and Marquis, 2000]. Such a com- and of belief revision.

pilation may consist in computing o-line the set As evoked in the introduction, our approach is based

min(RC (B); v) - or only the maximal elements of it on classical entailment so that the logical contents of

w.r.t. p . In this situation, the source of complexity the belief base is weakened instead of the inference

due to the number of preferred recoveries disappears relation itself. This gives to our family of forget re-

(in practice, this is signicant only when the number is lations some pros and some cons. Thus, if the belief

small enough, which is always the case when prudent base consists of a single inconsistent formula, our ap-

inference is considered); accordingly, the complexity proach is not adapted (as well as the coherence-based

of (skeptical or prudent) inference goes down to coNP. approaches and belief merging, which are also based

The other source of complexity can also be removed on classical entailment). Paraconsistent logics based

by imposing some restrictions on the belief base B. In on subclassical entailment relations can then be used

this situation, on-line forget inference is tractable. with prot. Conversely, since weakening bears on the

Proposition 5 6 This means that for every formula from P ROP

PS

1. Provided that min(RC (B); v) - or only the max- and for every substitution that renames some occurrences

of variables from into new variables (not occurring in ),

imal elements of it w.r.t. p - is part of the in- if belongs to the class under consideration, this is also

put, the inference problem associated to jCv (or the case for ().

inference relation in paraconsistent logics, it is uni- Transactions on Knowledge and Data Engineering,

form and the fact that information comes from several 3(2):208{220, 1991.

sources is not exploited. For instance, it is not possible [Benferhat et al., 1993] S. Benferhat, C. Cayrol,

to derive b from fa ^ (:a _ b); :ag using the paracon-

sistent logic LPm [Priest, 1991] (this is because :a _ b D. Dubois, J. Lang, and H. Prade. Inconsistency

is a consequence of :a). management and prioritized syntax- based entail-

ment. In Proceedings of the 13th International Joint

The idea of weakening beliefs instead of throwing them Conference on Articial Intelligence (IJCAI'93),

out is not new. For instance, in [Bessant et al., pages 640{645, 1993.

2001], each formula belonging to a con ict is weak-

ened through the introduction of a new propositional [Besnard and Hunter, 1998] Ph. Besnard and

symbol. This way, inconsistency is removed. Then, A. Hunter. Handbook of Defeasible Reasoning

non-monotonic inference is conducted over the result- and Uncertainty Management Systems, volume 2,

ing base so as to prefer the introduced symbols as false. chapter Introduction to actual and potential contra-

This enables a non-trivial treatment of iterated re- dictions, pages 1{11. Kluwer Academic Publishers,

vision. Unfortunately, this approach comes down to 1998.

the well-known WIDTIO approach (\When In Doubt [Bessant et al., 2001] B. Bessant, E. Gregoire, P. Mar-

Throw It Out") [Winslett, 1990] when only one revi- quis, and L. Sas. Frontiers of Belief Revision, vol-

sion is performed. Indeed, in this situation, weakening ume 22 of Applied Logic Series, chapter Iterated

by variable introduction does not enable more infor- syntax-based revision in a nonmonotonic setting,

mation to be preserved than if beliefs were removed. pages 369{391. Kluwer Academic, 2001.

Our approach does not suer from this drawback be-

cause weakening is achieved through forgetting instead [Brewka, 1989] G. Brewka. Preferred subtheories: an

of variable introduction. This also enables classical en- extended logical framework for default reasoning.

tailment to be performed from the resulting base (i.e., In Proceedings of the 11th International Joint Con-

after forgetting) instead of nonmonotonic inference. ference on Articial Intelligence (IJCAI'89), pages

Finally, weakening formulas by variable forgetting 1043{1048, Detroit (MI), 1989.

can be useful in many practical situations. For in- [Cadoli and Donini, 1997] M. Cadoli and F.M.

stance, if a series of tests has to be performed so Donini. A survey on knowledge compilation. AI

as to \solve" inconsistency [Lang and Marquis, 2000; Communications, 10:137{150, 1997. Printed in

Liberatore and Donini, 2000], preferred recoveries give 1998.

some hints about the variables that should be tested [Cayrol et al., 1998] C. Cayrol, M.-C. Lagasquie-

rst, namely, the most important ones (w.r.t. the Schiex, and Th. Schiex. Nonmonotonic reasoning:

graph G in C ) that have to be forgotten. Similarly, from complexity to algorithms. Annals of Mathe-

in a decision-theoretic context where the i 's are no matics and Articial Intelligence, 22(3{4):207{236,

longer beliefs but preferences of several agents, vari- 1998.

ables that do not need to be forgotten are these the

agents agree on, and preferred recoveries help nding [Cholvy, 1995] L. Cholvy. Automated reasoning with

the variables about which negociation should start. merged contradictory information whose reliabil-

ity depends on topics. In Proceedings of the Eu-

Acknowledgements ropean Conference on Symbolic and Quantitative

Approaches to Reasoning and Uncertainty (EC-

The authors wish to thanks Philippe Besnard and SQARU'95), pages 125{132, Fribourg, 1995.

Torsten Schaub for many helpful discussions. The sec- [Coste-Marquis and Marquis, 2000] S. Coste-Marquis

ond author has been partly supported by the IUT de and P. Marquis. Compiling stratied belief bases. In

Lens, the Universite d'Artois, the Region Nord/Pas- Proceedings of the 14th European Conference on Ar-

de-Calais under the TACT-TIC project, and by the ticial Intelligence (ECAI'00), pages 23{27, Berlin,

European Community FEDER Program. 2000.

[Dalal, 1988] M. Dalal. Investigations into a theory

References a knowledge base revision: preliminary report. In

Proceedings of the 7th National Conference on Ar-

[Baral et al., 1991] C. Baral, S. Kraus, and J. Minker. ticial Intelligence (AAAI'88), pages 475{479, St

Combining multiple knowledge bases. IEEE Paul (MN), 1988.

[Delgrande and Schaub, 2000] J. Delgrande and [Liberatore and Donini, 2000] P. Liberatore and F.M.

T. Schaub. A consistency-based model for belief Donini. Verication programs for abduction. In Pro-

change: Preliminary report. In Proceedings of the ceedings of the 14th European Conference on Arti-

17th National Conference on Articial Intelligence cial Intelligence (ECAI'00), pages 166{170, Berlin,

(AAAI'98), pages 392{398, 2000. 2000.

[Eiter and Gottlob, 1992] Th. Eiter and G. Gottlob. [Liberatore and Schaerf, 1998] P. Liberatore and

On the complexity of propositional knowledge base M. Schaerf. Arbitration (or how to merge knowl-

revision, updates and counterfactuals. Articial In- edge bases). IEEE Transactions on Knowledge and

telligence, 57:227{270, 1992. Data Engineering, 10(1):76{90, 1998.

[Liberatore and Schaerf, 2000] P. Liberatore and

[Fagin et al., 1983] R. Fagin, J.D. Ullman, and M.Y. M. Schaerf. Brels: A system for the integration

Vardi. On the semantics of updates in databases. of knowledge bases. In Proceedings of the 7th

In Proceedings of the 2nd ACM Syposium on Prin- International Conference on Knowledge Repre-

ciples of Database Systems (PODS'83), pages 352{ sentation and Reasoning (KR'00), pages 145{152,

355, 1983. Breckenridge (CO), 2000.

[Ginsberg, 1986] M.L. Ginsberg. Counterfactuals. Ar- [Lin and Mendelzon, 1999] J. Lin and A.O. Mendel-

ticial Intelligence, 30:35{79, 1986. zon. Knowledge base merging by majority. In Dy-

namic Worlds: From the Frame Problem to Knowl-

[Hunter, 1998] A. Hunter. Handbook of Defeasible edge Management. Kluwer, 1999.

Reasoning and Uncertainty Management Systems, [Lin and Reiter, 1994] F. Lin and R. Reiter. Forget

volume 2, chapter Paraconsistent logics, pages 11{ it! In Proceedings of the AAAI Fall Symposium on

36. Kluwer Academic Publishers, 1998. Relevance, pages 154{159, New Orleans (LA), 1994.

[Konieczny and Pino Perez, 1998] S. Konieczny and [Nebel, 1992] B. Nebel. Belief Revision, chapter

R. Pino Perez. On the logic of merging. In Proceed- Syntax-based approaches to belief revision, pages

ings of the 6th International Conference on Knowl- 52{88. Number 29 in Cambridge Tracts in Theoreti-

edge Representation and Reasoning (KR'98), pages cal Computer Science. Cambridge University Press,

488{498, Trento, 1998. 1992.

[Nebel, 1998] B. Nebel. Belief Revision, volume 3 of

[Konieczny et al., 2002] S. Konieczny, J. Lang, and Handbook of Defeasible Reasoning and Uncertainty

P. Marquis. Distance-based merging: a general Management Systems, chapter How hard is it to re-

framework and some complexity results. In Pro- vise a belief base?, pages 77{145. Kluwer Academic,

ceedings of the 8th International Conference on 1998.

Knowledge Representation and Reasoning (KR'02),

Toulouse, 2002. This issue. [Papadimitriou, 1994] Ch. H. Papadimitriou. Compu-

tational complexity. Addison{Wesley, 1994.

[Lang and Marquis, 1998] J. Lang and P. Marquis. [Pinkas and Loui, 1992] G. Pinkas and R.P. Loui.

Complexity results for independence and denabil- Reasoning from inconsistency: a taxonomy of prin-

ity. In Proceedings of the 6th International Confer- ciples for resolving con ict. In Proceedings of the

ence on Knowledge Representation and Reasoning 3rd International Conference on Knowledge Repre-

(KR'98), pages 356{367, Trento, 1998. sentation and Reasoning (KR'92), pages 709{719,

Cambridge (MA), 1992.

[Lang and Marquis, 2000] J. Lang and P. Marquis. In

search of the right extension. In Proceedings of the [Priest, 1991] G. Priest. Minimally inconsistent LP.

7th International Conference on Knowledge Repre- Studia Logica, 50:321{331, 1991.

sentation and Reasoning (KR'00), pages 625{636, [Rescher and Manor, 1970] N. Rescher and R. Manor.

Breckenridge (CO), 2000. On inference from inconsistent premises. Theory

[Lang et al., 2001] J. Lang, P. Liberatore, and P. Mar- and Decision, 1:179{219, 1970.

quis. The complexity of propositional independence. [Revesz, 1997] P. Z. Revesz. On the semantics of arbi-

Technical report, IRIT, 2001. (submitted for publi- tration. International Journal of Algebra and Com-

cation). putation, 7(2):133{160, 1997.

[Satoh, 1988] K. Satoh. Nonmonotonic reasoning by Sketch of proof of Proposition 3:

minimal belief revisions. In Proceedings of the Inter- We give the proof in the case of S . As a direct

national Conference on Fifth Generation Computer consequence of Corollary 3 from [Lang et al., 2001],

Systems (FGCS'88), pages 455{462, Tokyo, 1988. for every ! 2 Mod( ), there exists !0 2 Mod()

s.t. Diff(!; !0 ) = S if and only if ! j= ^ 9S:

[Weber, 1986] A. Weber. Updating propositional for- holds. Subsequently, we have f! 2 Mod( ) j 9!0 2

mulas. In Proceedings of the 1st Conference on Ex- Mod() s.t. Diff(!; !0 ) 2 Diff min ( ; )g =

pert Database Systems, pages 487{500, 1986. f! 2 Mod( ^ 9S:) j S 2 Diff min ( ; )g. Now,

S 2 Diff min ( ; ) if and only if S is a minimal (w.r.t.

[Winslett, 1990] M. Winslett. Updating logical ) set of variables s.t. ^ 9S: is consistent. This is

databases. Cambridge Tracts in Theoretical Com- equivalent to state that h;; S i is a preferred recovery

puter Science. Cambridge University Press, 1990. for h ; i w.r.t. vS given C , and the conclusion

follows. The proof for the case D is similar, with

Appendix: proof sketches cardinality playing the role that set inclusion plays

for S . We omit the case of W , the proof of which is

Sketch of proof of Proposition 1: not hard to establish.

((): Let v be a preference relation over RC (B)

satisfying binaricity. Let B = h ; : : : ; ni. For any Sketch of proof of Proposition 4: The membership

S

1

subset S of f ; :::; ng, we dene V~S = hV ; : : : ; Vni

1 1 proof for Point 1 goes as follows. Let B = h ; : : : ; ni

1

where for each i 2 1 : : :n, Vi = ; if i 2 S and Vi = Fi be a belief base and let C = hF; G; H i be a forget-

otherwise. Now, we dene a preference relation ting context for B. The following nondeterministic

on consistent subsets of f ; :::; ng by X Y if

1 algorithm can be used to determine whether B is

and only if V~X v V~Y . It can then be shown that recoverable: 1. Guess n subsets V1; : : : ; Vn of V ar(B)

f ; :::; ng j8 if and only if B jCv .

1

S

and n + 1 interpretations !, !1; : : : ; !n over V ar(B);

()): Let be a preference relation over con- 2. For each i 2 1 : : :n, check that: (a) Vi Fi ; (b) for

sistent subsets of f ; :::; ng. For any vector

1 every (v; v0 ) 2 Gi , if v 2 Vi , then v0 2 Vi ; (c) !i j= i;

V~ = hV ; : : : ; Vni of subsets of PS, let SV~ be the

1

(d) ! coincides with !i on V ar(i ) n Vi .

subset of f ; :::; ng dened by SV~ = fi j Vi = ;g.

1

Due to lack of space we omit hardness (easy to

Let us now dene the preference relation v over establish) and points 2 and 3.

RC (B) by V~ v V~ 0 if and only if SV~ SV~ 0 . Note

that v satises binaricity. Now, it can be shown that

S

f ; :::; ng j8 if and only if h ; :::; ni jCv .

1 1

S

V~ = hV ; : : : ; Vni be any preferred recov-

1

ery from min(RC (B); v). Forget inference

Sketch of proof of Proposition 2: W~

from B comes down to classical inference from

((): Assume that v is a complete preference V 2min R B ;v B j V~ . The key point is the equiv-

( C( ) )

relation on RC (B) induced by a ranking function

S

V that f9V : ; : : : ; 9Vn:ng j= if and

alence stating 1 1

only if ni i [new(Vi )] j= , where i [new(Vi )] is

H s.t. (V~ ) = H(0 (V1); : : : ; 0(Vn )). Let us dene

=1

the formula obtained by renaming in an uniform way

the merging operator jM d; induced from the pseudo- in i every variable v 2 Vi by the new variable vi ,

distance dened by d(!; !0) = 0(Diff(!; !0 )) and provided that does not contain any new variable

the aggregation function = H. It can then be vi , i 2 1 : : :n. While the formula corresponding to

proven that B jMd; if and only if B jv .

C

S

B j V~ after such a renaming process is not logically

()) Let B jM d; be a decomposable and dif- equivalent to it7, it is query-equivalent to it (i.e.,

ferential n-merging inference relation induced equivalence in preserved over the original language,

by pseudo-distance d and aggregation func- only), but this is sucient for our query answering

tion , i.e., d(!; B) = ni=1 (d(!; i)) and purpose.

d(!; !0 ) = f(Diff(!; !0 ). Let us consider the

forget inference induced from the standard for-

getting context and the function dened by

(V~ ) = H(0(V1 ); : : : ; 0(Vn )) where 0 = f and 7

In [Lang et al., 2001], we show that in the general case

H = . Then we prove that we get (!) = d(!; B), and unless NP \ coNP P/poly, there exists no proposi-

from which it follows that B jM d; = B jv .

C S

tional formula equivalent to 9V:, s.t. the size of is

polynomially bounded in the size of plus the size of V .

## Viel mehr als nur Dokumente.

Entdecken, was Scribd alles zu bieten hat, inklusive Bücher und Hörbücher von großen Verlagen.

Jederzeit kündbar.