Professional Documents
Culture Documents
Distance Based Merging: A General Framework and Some Complexity Results
Distance Based Merging: A General Framework and Some Complexity Results
net/publication/221393284
CITATIONS READS
72 47
3 authors, including:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Pierre Marquis on 22 July 2015.
arguments; more formally, a \fun
tion" f is an family Example 1 Assume for example that we want to
f = ffn j n 2 INg of n-ary fun
tions from IN to IN.
Slightly abusing notations, we write f (x ; : : : ; xn ) instead merge E = fK1; K2 ; K3; K4 g under the integrity
on-
of fn (x ; : : : ; xn ) sin
e this
an never be ambiguous. straints IC = >, where
1
1
{ K1 = fa; b;
; a ! :bg, 3.2 INSTANCIATING OUR
FRAMEWORK
{ K2 = fa; bg, Let us now instantiate our framework and fo
us on
some simple families of distan
es and aggregation fun
-
{ K3 = f:a; :bg, tions.
Denition 4 (some distan
es) Let !1 , !2 2W be
{ K4 = fa; a ! bg. two interpretations.
In this example, K1 knows that
holds; sin
e this { The drasti
distan
e dD is dened by
dD (!1 ; !2 ) = 0 if !1 = !2 ,
pie
e of information is not involved in any
ontra- 1 otherwise
di
tion, it
an prove sensible to be
ondent in K1
about the truth of
. Model-based merging opera- { The Hamming distan
e dH is dened by
tors
an not handle this situation: in
onsistent knowl- dH (!1 ; !2) = jfx 2 P S j !1 (x) 6= !2(x)gj
edge bases
an not be taken into a
ount. Thus, pro- { Let q be a total fun
tion from P S to IN . The
vided that the Hamming distan
e between interpre- weighted Hamming distan
e dHq indu
ed by q is
tations is
onsidered, the operator 4 [Revesz, 1997;
Lin and Mendelzon, 1999; Konie
zny and Pino Perez, dened by P
dHq (!1 ; !2 ) = x2P S j !1 (x)=6 !2(x) q(x)
1999℄ gives a merged base whose models are: fa; b; :
g
and fa; b;
g; the operator 4Gmax [Konie
zny and Pino These distan
es satisfy the requirements imposed in
Perez, 1999℄ gives a merged base whose models are: Denition 3.
f:a; b; :
g, f:a; b;
g, fa; :b; :
g, and fa; :b;
g. In
any of these two
ases, nothing
an be said about The Hamming distan
e is the most usual distan
e
on-
the truth of
in the merged base, whi
h is
ounter- sidered in model-based merging6. It is very simple to
intuitive sin
e no argument against it
an be found in express, but one has to keep in mind that it is very
the input. sensitive to the representation language of the problem
Syntax-based operators render possible the exploita- (i.e., the
hoi
e of propositional symbols) and that nu-
tion of in
onsistent knowledge bases, but they do not merous others distan
es
an be used. Weighted Ham-
are about the distribution of information. Consider ming distan
es are relevant when some propositional
the two standard syntax-based operatorsS[Baral et al.,
symbols are known as more important than others.
1992℄, sele
ting the maximal subsets of mi=1 Ki (one Denition 5 (some aggregation fun
tions)
w.r.t. set in
lusion and the other one w.r.t.
ardinal-
ity). On the previous example, the rst one returns { Let q be a total fun
tion from f1; : : : ; ng to IN
a merged base equivalent to
and the se
ond one to s.t. q(1) = 1 whenever n = 1. The weighted sum
^:a. So, a is in the result for none of these two oper- WSq indu
ed by q is dened by WSq (e1 ; : : : ; en ) =
ators, whereas a holds in three over four input bases. P
i=1 q (i)ei .
n
Our distan
e-based operators a
hieve a
ompromise { Let q be a total fun
tion from f1; : : : ; ng to IN s.t.
between model-based operators and syntax-based op- q(1) = 1 whenever n = 1, and q(1) 6= 0 in any
erators, by taking into a
ount the way information is
distributed and by taking advantage of the informa- P
ase. The ordered weighted sum OWSq indu
ed by
q is dened by OWSq (e1 ; : : : ; en ) = ni=1 q(i)e(i)
tion stemming from in
onsistent knowledge bases. For where is a permutation of f1 : : : ng s.t. e(1)
instan
e, our operator 4dD;sum;sum (
f. Se
tion 3.2) e(2) : : : e(n) .
gives a merged base whose single model is fa; b;
g, and
4dD ;sum;lex returns a merged base whose models are q is a weight fun
tion, that gives to ea
h formula (resp.
f:a; b;
g and fa; :b;
g. So, with any of these two op-
erators, we
an dedu
e that
holds after the merging. knowledge base) 'i (resp. Ki) of index i its weight q(i)
Moreover, these operators exhibit typi
al merging be- denoting the formula (resp. knowledge base) reliabil-
haviours. The rst one is a majority operator: sin
e ity. With the slight dieren
e that q is normalized (but
three of four bases agree on a, a holds in the result. without requiring that q(1) = 1 whenever n = 1), the
The se
ond one is an arbitration operator; being more latter family is well-known in multi-
riteria de
ision
onsensual, it gives that only one of a or b holds, to be 6
In this
ontext, it is also
alled Dalal distan
e [Dalal,
as
lose as possible to ea
h of the knowledge bases. 1988℄.
4dD ;max;max = >
4 dD ;max;sum ,4 , 4dH ;max;sum
dD ;max;lex = a^b
4dD ;sum;max = :b
4dD ;sum;sum = (:a ^ :b) _ (a ^ b ^
)
4dD ;sum;lex = :a ^ :b
4Hd ;sum;max , 4dH ;sum;lex = a ^ :b ^
4Hd ;max;max ; 4dH ;max;lex = (:a ^ b ^
) _ (a ^ :b ^
)
4dH ;sum;sum = a^
Figure 1: Example 2
making under the terminology \Ordered Weighted Av- theory). Noti
eably, the usual aggregation fun
tions
erages" (OWAs) [Yager, 1998℄. When q(i) = 1 for ev- used in these elds are all polynomially
omputable,
ery i 2 1 : : : n, WSq is the usual sum (and OWSq as whi
h makes the following
omplexity results appli
a-
well). When q(1) = 1 and q(2) = : : : = q(n) = 0 then ble when instantiating f and g with these fun
tions.
OWSq (e1 ; : : : ; en) = max(e1 ; : : : ; en ). Note that fun
tions su
h as the purely utilitarian
For the se
ond aggregation step g, it is relevant to
on- sum or weighted sum allow for
ompensations between
sider the well-known leximax ordering whi
h
ompares s
ores (and lead to majority-like operators), while the
two ve
tors of s
ores by fo
using on the largest s
ores egalitarian fun
tions max and lex do not.
of ea
h ve
tor, and in
ase of equality, on the se
ond By letting the parameters d, f and g vary in these re-
largest s
ores, and so on. For the sake of homogeneity, spe
tive sets, several merging operators are obtained;
we reformulate the leximax ordering so as to
ompare some of them were already known and are thus re
ov-
aggregated s
ores rather than ve
tors of s
ores. This ered as spe
i
ases in our framework, and others are
an be done thanks to a spe
i
aggregation fun
tion new operators. Thus,V4dD;max;max is the basi
merg-
OW Sq : ing operator, giving E ^ IC if
onsistent and IC
Denition 5.1 (leximax) otherwise. 4dD;max;sum is the drasti
merging oper-
Let M be an upper bound of the s
ores d(!; Ki )7 , i.e., ator whi
h amounts to sele
t the models of IC sat-
for any ! we have d(!; Ki ) < M . Now, let q(i) = isfying the greatest number of knowledge bases from
M n i for all i. The rank order on ve
tors of s
ores E . It is equivalent to the drasti
majority operator
indu
ed by OWSq is the leximax ordering, abbreviated as dened in [Konie
zny, 2000℄ when working with de-
by lex8 . du
tively
losed knowledge bases. 4dD;sum;sum
or-
responds to the interse
tion operator of [Konie
zny,
Using the leximax aggregation for the rst aggregation 2000℄. 4dD;WSq ;max
orresponds to an operator used
step (f ) would also be possible, but leads to rather in [Lafage and Lang, 2000℄ in a group de
ision
ontext.
lengthy te
hni
al tri
ks to be dened properly in
ase When singleton knowledge bases are
onsidered 9 { re-
where the se
ond aggregation fun
tion g is not purely
all that in this
ase f is irrelevant { every 4dH ;f;max
ordinal (i.e., g dierent from max and leximax) and we operator is a 4Max operator [Revesz, 1997℄, every
ignore this possibility here (see the long version of the 4dH ;f;sum operator is a 4 operator [Revesz, 1997;
paper [Konie
zny et al., 2001℄). Lin and Mendelzon, 1999; Konie
zny and Pino Perez,
1999℄, and every 4dH ;f;lex operator is a 4GMax oper-
All these fun
tions satisfy the requirements imposed in ator [Konie
zny and Pino Perez, 1999℄. Still with sin-
Denition 3; all of them are symmetri
but weighted gleton knowledge bases, taking d = dD and f = WSq ,
sum (ex
ept when q is uniform). 4dH ;f;WSq is a penalty-based merging (where one
Many other possible
hoi
es for f and g
an be found minimizes the sum of the penalties q(i) atta
hed to
in the literature of multi-
riteria de
ision making (and the Ki's) [Pinkas, 1995℄, and taking d = dD and
to a smaller extent in the literature of group de
ision f = W MAXq (dened by W MAXq (x1 ; : : : ; xn ) =
maxi=1:::n min(q(i); xi )) we get10 a possibilisti
merg-
For instan
e, when d = dH and f = max
7
P we
an
hoose ing operator [Benferhat et al., 2000℄.
M = jP S j + 1; when d = dH and f = we
an
hoose
M = jP S j + 1.
2 9
Or, equivalently, when ea
h Ki is repla
ed by fV Ki g
Namely,
8
0 0 we have> e0 OWS q (e ; : : : ; en ) before merging.
OWSq (e ; : : : ; en ) i (e ) or (e = e0 and
1
10
The s
ales used for s
ores are dierent but it is obvious
e > e0 ) or et
.
1 (1) 0 (1) (1) 0 (1)
We will now illustrate the behaviour of these dierent point in this framework is that, unlike
lassi
al model-
operators on an example. based merging operators, the
onne
tor \;" is not the
same that the
onne
tor \^".
Example 2 Consider the following knowledge set
E = fK1; K2 ; K3 ; K4g that we want to merge under
the integrity
onstraints IC = >. 4 COMPUTATIONAL
COMPLEXITY
{ K1 = fa ^ b ^
; a ! :bg,
Let us now turn to the
omplexity issue. We obtained
{ K2 = fa ^ bg, the following result:
{ K3 = f:a ^ :b; :bg, Proposition 1 Let 4d;f;g be a distan
e-based merg-
{ K4 = fa; a ! bg. ing operator. Given a knowledge set E and two for-
mulas IC and :
The result of the merging of E a
ording to the dier-
ent operators with d 2 fdD ; dH g, f 2 fmax; sumg and { If d, f and g are
omputable in polynomial time,
g 2 fmax; sum; lexg under no
onstraints (i.e. IC= IC (E ) j= holds is
then determining whether 4d;f;g
>) is indi
ated gure 1. See table 1 for an example in p2 .
of
al
ulation with the 4dH ;sum;lex operator. In this
table the interpretation (1; 0; 0) for example is the one { If d, f and g are
omputable in polynomial time
mapping a to true and b and
to false. The result of and are polynomially bounded, then determining
the merging 4d>H ;sum;lex (E ) is the interpretation that IC (E ) j= holds is in 2 .
whether 4d;f;g p
Table 3: Complexity results (d = dH ) stead of
ardinality and \lo
ated" at the 2nd level of
f=g max sum lex WSq OWSq PH)
annot be en
oded in polynomial time as distan
e-
based operators (unless PH
ollapses).
max p2p-
p2p-
p2p -
p2p-
p2p -
sum 2 -
2 -
2p -
2p-
2p -
WSq p2p -
p2p -
2p -
2p-
2p -
5 LOGICAL PROPERTIES
OWSq 2 -
2 -
2 -
2 -
2 -
Sin
e we aim at investigating the logi
al properties of
our family of merging operators, a set of properties
must rst be
onsidered as a base line. In [Konie
zny
Sket
hes of the proofs are given in the Appendix and Pino Perez, 1999℄, a study of logi
al properties
(again, see [Konie
zny et al., 2001℄ for fully detailed that \good" merging operators should satisfy (in the
proof). It is worth adding that in the
ase d = dHq ,
ase where all the knowledge bases are equally reli-
p2 -hardness still holds whenever E is a singleton fK g, able) is
arried on. The following set of postulates was
K is a singleton f'g and ' is a
onjun
tion of variables proposed:
(in this
ase, neither f nor g plays a signi
ant role in
the elaboration of the distan
e to E ). As to the
ase Denition 6 (IC merging operators) Let E , E1 ,
d = dH , p2 -hardness still holds when ea
h expli
it E2 be knowledge sets, K1 , K2 be
onsistent knowledge
belief is a
onjun
tion of variables, and p2 -hardness bases, and IC , IC1 , IC2 be formulas from P ROPP S .
results hold whenever E is a singleton fK g, K is a 4 is an IC merging operator i it satises the follow-
singleton f'g and ' is a
onjun
tion of variables. ing postulates:
Looking at the tables above, we
an observe that the (IC0) 4IC (E ) j= IC
hoi
e of the distan
e d has a great in
uen
e on the (IC1) If IC is
onsistent, then 4IC (E ) is
onsistent
omplexity results. Thus, whenever d = dH or d = V E is
onsistent with IC , then
dHq , the
omplexity results for inferen
e from a merged
4IC (E ) V E ^ IC
(IC2) If
base
oin
ide whenever f (or g) is a WSq fun
tion or
a OWSq fun
tion. This is no longer the
ase when
d = dD is
onsidered. (IC3) If E1 E2 and IC1 IC2 , then
Together with Proposition 1, the
omplexity of many 4IC1 (E1 ) 4IC2 (E2 )
model-based merging operators already pointed out in (IC4) If K1 j= IC and K2 j= IC , then 4IC (K1 t
the literature are derived as a by-produ
t of the pre- K2 ) ^ K1 is
onsistent i 4IC (K1 t K2 ) ^ K2
vious
omplexity results. To the best of our knowl- is
onsistent
edge, the
omplexity of su
h operators has not been (IC5) 4IC (E1 ) ^ 4IC (E2 ) j= 4IC (E1 t E2 )
identied up to now11, hen
e this is an additional
on- (IC6) If 4IC (E1 ) ^ 4IC (E2 ) is
onsistent , then
tribution of this work. We
an also note that, while 4IC (E1 t E2 ) j= 4IC (E1 ) ^ 4IC (E2 )
the
omplexity of our distan
e-based operators is not
very high (rst level of PH, at most), nding out sig- (IC7) 4IC1 (E ) ^ IC2 j= 4IC1^IC2 (E )
ni
ant tra
table restri
tions seems a hard task sin
e
intra
tability is still the
ase in many restri
ted sit- (IC8) If 4IC1 (E ) ^ IC2 is
onsistent, then
uations. Finally, our results show that some syntax- 4IC1^IC2 (E ) j= 4IC1 (E )
based merging operators (based on set in
lusion in- Two sub-
lasses of IC merging operators have also
11
However, (4 (E ) j= ) 2
an be re-
dH ;sum;sum ? p been dened. Majority operators that aim at resolving
IC
overed from a
omplexity result given in [Liberatore and
2
on
i
ts by listening the majority wishes, and arbitra-
S
haerf, 2000℄, page 151. tion operators that have a more
onsensual behaviour:
Denition 7 (majority and arbitration) A ma- Con
erning the operators examined in the previous
jority operator is an IC merging operator that satises se
tion, we have identied the following properties:
the following majority postulate:
Proposition 5 4d;f;g satises the logi
al properties
(Maj) 9n 4IC (E1 t E2 n ) j= 4IC (E2 ) stated in Tables 5 and 6. Sin
e all these operators are
An arbitration operator is an IC merging operator that already known to satisfy (IC0), (IC1), (IC2), (IC7)
satises the following postulate: and (IC8) (
f. Proposition 3), we avoid repeating su
h
4IC1 (K1 ) 4IC2 (K2 ) 9
> postulates here. For more readability, postulate (ICi)
(Arb) IC 4IC1 ,:IC2 (K1 t K2 ) (IC1 , :IC2 ) = ) is noted i and M (resp. A) stands for (Maj) (resp.
1 6j= IC2 > (Arb)).
IC2 6j= IC1
;
4IC1 _IC2 (K1 t K2 ) 4IC1 (K1 )
Table 5: Logi
al properties (d = dD )
See [Konie
zny and Pino Perez, 2002; Konie
zny and f=g max sum lex WSq
Pino Perez, 1999℄ for more explanations about those max 3,4,5,A 3,4,5,6,M,A 5,6,M
two postulates and the behaviour of the two sub- sum 5,A 5,6,M 5,6,A 5,6,M
lasses.
WSq OWSq 5,A 5,6,M 5,6,A 5,6,M
We have the following result:
Proposition 3 4d;f;g satises (IC0), (IC1), (IC2),
(IC7), (IC8). The other postulates are not satised in Table 6: Logi
al properties (d = dH or d = dHq )
the general
ase.
f=g max sum lex WSq
Clearly enough, it is not the
ase that every distan
e- max 5,A 5,6,M 5,6,A 5,6,M
based merging operator is an IC merging operator (not sum 5,A 5,6,M 5,6,A 5,6,M
satisfying some postulates is deliberate sin
e we want WSq OWSq 5,A 5,6,M 5,6,A 5,6,M
to give some importan
e to the syntax in order to take
into a
ount in
onsistent knowledge bases). Let us in-
trodu
e some properties to be satised by aggregation The tables above show our operators to exhibit dier-
fun
tions f : ent properties. We remark that among our operators,
only 4dD;max;sum satises all listed properties. Fail-
1) f (x1 ; : : : ; xn ) = 0 i x1 = : : : = xn = 0 ing to satisfy (IC3) (irrelevan
e to the syntax) in many
(minimality)
ases is not surprising, sin
e we want to allow our oper-
ators to take syntax into a
ount. (IC4) imposes that,
2) If '1 ^ : : : ^ 'n is
onsistent, then when merging two knowledge bases, if the result is
on-
f (d(w; '1 ); : : : ; d(w; 'n )) = f (d(w; '1 ^ : : : ^ 'n )) sistent with one knowledge base, it has to be
onsistent
(and) with the other one { this fairness postulate is irrele-
vant when working with non-symmetri
operators (so,
3) For any permutation , f (x1 ; : : : ; xn ) = unsurprisingly, it is not satised for g = WSq ). This
f ((x1 ; : : : ; xn )) (symmetry) postulate is not satised by any operator for whi
h d
4) If f (x1 ; : : : ; xn ) f (y1 ; : : : ; yn ), then is Hamming distan
e sin
e
ardinalities of the knowl-
f (x1 ; : : : ; xn ; z ) f (y1 ; : : : ; yn; z ) (
omposition)
edge bases have an in
uen
e on f , and more gener-
ally, it is hardly satisable when working with syntax-
5) If f (x1 ; : : : ; xn ; z ) f (y1 ; : : : ; yn ; z ), then dependent operators. (IC5) and (IC6) are related to
f (x1 ; : : : ; xn ) f (y1 ; : : : ; yn ) (de
omposition) Pareto dominan
e in so
ial
hoi
e theory and are really
important for multi-sour
e aggregation; so it is worth
Now, if one wants to re
over the full set of postulates noting that almost all operators satisfy them (only op-
(IC0)-(IC8): erators for whi
h g = max do not satisfy (IC6)).
We do not put the operators with g = OWSq in the ta-
Proposition 4 A distan
e-based merging operator bles be
ause they gather many aggregation fun
tions
4d;f;g satises (IC0)-(IC8) if and only if the fun
tion and so they do not satisfy a lot of logi
al properties.
f satises (minimality) and (and), and the fun
tion g Moreover, some properties (as (IC5) and (IC6)) re-
satises (minimality), (symmetry), (
omposition) and quire to be able to
ope with knowledge sets of dier-
(de
omposition). ent sizes, whereas g = OWSq operators have to spe
ify
exa
tly the size of the knowledge sets. It is possible Proof : It is suÆ
ient to
onsider the following
to generalize the denition of those operators to
ope non-deterministi
algorithm:
with these
ases but it is out of the s
ope of this paper.
1. guess an interpretation ! and N interpretations
!i;j (i = 1::m, j = 1::ni ) over V ar(E [ fIC g),
6 CONCLUSION
where N = i=1::nni is the total number of for-
mulas 'i;j in E ;
The major
ontribution of this paper is a new frame-
work for propositional merging. It is general enough to 2.
he
k that ! j= IC and that !i;j j= 'i;j for all
en
ompass many existing operators (both model-based i = 1::m and all j = 1::ni ;
ones and syntax-based ones) and to enable the deni- 3.
ompute d(!; !i;j ) for all i and all j ;
tion of many new operators (symmetri
or not). Both
the logi
al properties and the
omputational proper- 4.
ompute d(!; Ki) for all i;
ties of the merging operators pertaining to our frame-
work have been investigated. Some of our results are 5.
ompute d(!; E ) and
he
k that d(!; E ) k.
large-s
ope ones in the sense that they make sense un-
der very weak
onditions on the three parameters that This algorithm runs in polynomial time in the size of
must be set to dene an operator in our framework. the input (E , IC and k represented in binary) sin
e
By instantiating our framework and
onsidering sev- d, f , g are
omputable in polynomial time.
eral distan
es and aggregation fun
tions, more rened
results have also been obtained.
This work
alls for several perspe
tives. One of them Lemma 2 If for all ! 2 W the value of d(!; E ) is
onsists in analyzing the properties of the distan
e- bounded by the value h(jE j + jIC j) (where h is a fun
-
based operators that are a
hieved when some other tion with values in IN) then the value min!j=IC d(!; E )
aggregation fun
tions or some other distan
es are
on-
an be
omputed using dlog h(jE j + jIC j)e
alls to an
sidered. For instan
e, suppose that a
olle
tion of for- NP ora
le.
mulas of interest (topi
s) is available. In this situation,
the distan
e between !1 and !2
an be dened as the Proof : min = min!j=IC d(!; E )
an be
omputed
number of relevant formulas on whi
h !1 and !2 diers using binary sear
h on f0; : : : ; h(jE j + jIC j)g with
(i.e., su
h that one of them satises the formula and at ea
h step a
all to an NP ora
le to
he
k whether
the other one violates it). Several additional distan
es min!j=IC d(!; E ) k (that is in NP from lemma 1).
ould also be dened and investigated (see e.g. [Lafage Sin
e a binary sear
h on f0; : : : ; h(jE j + jIC j)g needs
and Lang, 2001℄ for distan
es based on Choquet inte- at most dlog h(jE j + jIC j)e steps, the result follows.
gral).
A
knowledgements Point 1. of Proposition 1
If d, f and g are
omputable in polynomial time,
The third author has been partly supported by the then for every knowledge set E and every ! 2 W ,
IUT de Lens, the Universite d'Artois, the Region the binary representation of d(!; E ) is bounded by
Nord/Pas-de-Calais under the TACT-TIC proje
t, p(jE j + jIC j), where p is a polynomial. Hen
e, the
and by the European Community FEDER Program. value of d(!; E ) is bounded by 2p(jEj+jICj). From
lemma 2, we
an
on
lude that min
an be
omputed
using a polynomial number of
alls to an NP ora
le.
Appendix: Proof sket
hes of the Now, let E be a knowledge set, IC be a formula, k
omplexity results be an integer and be a formula, it
an be shown
that the problem of determining whether there
Sket
h of Proof of Proposition 1 : These results exists a model ! of IC su
h that d(!; E ) = k and
are
onsequen
es of the two following lemmata: su
h that ! 6j= is in NP. So we
an show that
IC (E ) 6j= using rst a polynomial number of
4d;f;g
Lemma 1 Let k be an integer; if d, f and g are
om-
alls to an NP ora
le in order to
ompute min, and
putable in polynomial time, then the problem of de- then using an additional
all to an NP ora
le in order
termining whether min!j=IC d(!; E ) k given IC , E to determine whether there exists a model ! of IC s.t.
and k is in NP. d(!; E ) = min and ! 6j= . Hen
e the membership to
p2 for this problem, and hen
e for its
omplement.
ardinality-maximizing base revision ÆC (The-
orem 5.14 from [Nebel, 1998℄) sin
e we have
Point 2. of Proposition 1 dICD ;f;g (ff'1g; : : : ; f'ngg) f'1; : : : ; 'ng ÆC IC
When d, f and g are polynomially bounded, the proof for any hf; gi 2 fhmax; maxi; hmax; lexi;
is similar to the one of point 1., but the
omputation hsum; maxi; hsum; sumig. Sin
e sum is a spe-
of min!j=IC d(!; E ) needs only a logarithmi
number
i
OW Sq fun
tion, the
orresponding results
of steps sin
e h isp polynomially bounded, hen
e the still hold in the
ases (f = OWSq ; g = max) and
membership to 2 . (f = max; g = OWSq ).
{ table 2,
ase d = dD , f = OWSq , g = sum: p2 -
hardness is established by
onsidering pthe follow-
Sket
h of Proof of Proposition 2 : ing polynomial redu
tion from the 2 -
omplete
problem max-sat-asgodd [Wagner, 1987℄. max-
sat-asgodd is the following de
ision problem:
1. Membership given a propositional formula s.t. V ar() =
Membership-to-p2 results are dire
t
onsequen
es of fx1 ; : : : ; xn g and a stri
t ordering x1 < x2 < : : : <
Proposition 1 sin
e both distan
es and aggregation xn on V ar() indu
ing the lexi
ographi
ordering
fun
tions
an be
omputed in polynomial time. 4 on
, is the greatest model ! of w.r.t. 4 su
h
Membership-to-p2 results are also
onsequen
es of that !(xn) = 1? We just give here the redu
tion:
Proposition 1, ex
ept those for whi
h f or g is an to s.t. V ar() = fx1 ; : : : ; xn g, we asso
iate the
OW Sq (in
luding lex) when the drasti
distan
e dD tuple M () = hE; IC; i, where E = fKi j i 2
is
onsidered; these
ases are brie
y dis
ussed now: 1 : : : ng,VIC = , = xn and for ea
h i 2 1 : : : n,
Ki = f kn=1 +2 j x j j 2 1 : : : n + 2 ig (ea
h K
i i
{
ase d = dD , f = max and g = OWSq . We rst
ontains n + 2 i formulas that are synta
ti
ally
establish that d(!; E )
an only take only a poly- distin
t but all equivalent to xi ), and we
onsider
nomial number of dierent values, and that this the OWSq fun
tion f indu
ed by q s.t. q(1) = 1
set of possible values
an be
omputed in poly- and for every j > 1, q(j ) = 2i 2.
nomial time. Indeed, if kE (!) is the number of
belief basesPK i from E s.t. ! j= Ki , we have { table 2, p2 -hardness results in the
ase f = sum:
d(!; E ) = m i=k(!)+1 qi ; whi
h makes jE j + 1 dif-
hardness in the
ase (d = dD ; f = sum; g = lexp)
ferent values,
omputable in polynimial time. The is easily derived by taking advantage of the 2 -
rest of the proof is similar to the proof of mem- hardness result in the
ase where ea
h Ki is a sin-
bership to p2 in the
ases where g and g are poly- gleton redu
ed to a
onjun
tion of atoms (hen
e f
nomially bounded, the dieren
e being here that is irrelevant), g = lex and d = dH . Sin
e sum is a
the minimal value min = min!2
d(!; E ) is
om- spe
i
WSq fun
tion and lex is a spe
i
OWSq
puted through binary sear
h using the pre
om- fun
tion, this hardness result
an be extended to
puted jE j +1 dierent possible values for d(!; E ). the rest of the table, ex
ept for the
ases where
f is a WSq fun
tion and g 2 fmax; sumg) and
{ the
ase d = dD , f = OWSq and g = max is sim- where g is a WSq fun
tion. In the latter
ase, the
ilar, the main dieren
e is that d(!; E )
an only p2 -hardness of linear base revision ÆL (Theorem
take at most maxi21:::n
ard(Ki ) dierent values. 5.9 from [Nebel, 1998℄)
an be used to obtain the
desired result: indeed, it is suÆ
ient to
onsider
Finally, as to the basi
merging operator belief bases Ki redu
ed to singletons; we have
(dD ; max; max), determining whether a formula 4dICD ;f;g (fK1 ; : : : ; Kng) fK1; : : : ; Kng ÆL IC ,
is a logi
al
onsequen
e of the merged base E given where g is the weighted sum indu
ed by q s.t.
IC
an be a
hieved using the following algorithm: q(i) = 2n i , and ea
h Ki is viewed as the unique
if sat(E [ fIC g) then return(unsat(E [ fIC; :g)) formula it
ontains. Here, the preferen
e ordering
else return(unsat(fIC; :g)), whi
h shows member- over fK1; : : : ; Kng is s.t. K1 < K2 < : : : < Kn.
ship of the de
ision problem to BH2. { table 2,
ase (d = dD ; f = g = max): it is suf-
ient to
onsider the following polynomial re-
2. Hardness: du
tion M from sat-unsat: to a pair of for-
mulas h'; i whi
h do not share variables (this
{ table 2, p2 -hardness results: they are di-
an be assumed without loss of generality), we let
re
t
onsequen
es of hardness results for M (h'; i) = hE = '; IC = new; = ' ^ new ^
: i where new is a new variable and we
he
k [Baral et al., 1992℄ C. Baral, S. Kraus, J. Minker, and
that h'; i 2 sat-unsat i is a logi
al
onse- V. S. Subrahmanian. Combining knowledge bases
quen
e of the merged base E given IC .
onsisting of rst-order theories. Computational In-
telligen
e, 8(1):45{71, 1992.
{ table 3, p2 -hardness results. They still hold in the
situation where E
ontains only one belief base [Benferhat et al., 2000℄ S. Benferhat, D. Dubois,
K and K itself
ontains only one formula that is S. Ka
i, and H. Prade. En
oding information fu-
a
onjun
tion of atoms. This merely shows that sion in possibilisti
logi
: a general framework for
our hardness result is independent from f and g rational synta
ti
merging. In Pro
. of ECAI'00,
(sin
e they are irrelevant whenever E and K are pages 3{7, 2000.
singletons) but is a
onsequen
e of the distan
e [Blo
h and Hunter, 2001℄ I. Blo
h and A. Hunter, ed-
that is used (Hamming). Indeed, in this restri
ted
ase, 4dICH ;f;g (fK g) is equivalent to K ÆD IC where itors. Fusion: General Con
epts and Chara
teris-
ÆD is Dalal's revision operator. The fa
t that the ti
s, volume 16 of International Journal of Intelli-
inferen
e problem from K ÆD IC is p2 -hard (even gent Systems. Wiley, 2001. Spe
ial Issue on Data
in the restri
ted
ase where K is a
onjun
tion of and Knowledge Fusion.
atoms)
on
ludes the proof (see Theorem 6.9 from [Dalal, 1988℄ M. Dalal. Investigations into a theory
[Eiter and Gottlob, 1992℄). of knowledge base revision: preliminary report. In
{ table 3, p2 -hardness results. We show that these Pro
. of AAAI'88, pages 475{479, 1988.
2 -hardness results hold in the restri
ted
ase
p
[Eiter and Gottlob, 1992℄ T. Eiter and G. Gottlob.
where ea
h Ki is a singleton, redu
ed to a
on- On the
omplexity of propositional knowledge base
jun
tion of literals (hen
e f is irrelevant) when revision , updates, and
ounterfa
tuals. Arti
ial
g = lex by the following polynomial redu
tion M Intelligen
e, 57(2-3):227{270, 1992.
from max-sat-asgodd: to any propositional for-
mula s.t. V ar() = fx1;V: : : ; xi+1n g we asso
iate [Hansson, 1998℄ S. O. Hansson. Revision of belief sets
M () = hE = fKi = fxi ^ j2=n i+1
V newj g j i 2 1 and belief bases. Handbook of Defeasible Reason-
.. ng; IC = ^ j=2 :newj ; = xn i where ea
h
2n ing and Un
ertainty Management Systems, Vol. 3,
newj (j 2 2 .. 2n) is a new variable. pages 17{75, 1998.
{ table 4. We show that p2 -hardness holds in the [Konie
zny and Pino Perez, 1999℄ S. Konie
zny and
very restri
ted
ase where E
ontains only one R. Pino Perez. Merging with integrity
onstraints.
belief base K and K itself
ontains only one for- In Pro
. of ECSQARU'99, LNAI 1638, pages 233{
mula that is a
onjun
tion of atoms. This merely 244, 1999.
shows that our hardness result is independent [Konie
zny and Pino Perez, 2002℄ S. Konie
zny and
from f and g (sin
e they are irrelevant when- R. Pino Perez. On the frontier between arbitration
ever E and K are singletons) but is a
onsequen
e and majority. In Pro
. of KR'02, 2002.
of the family of distan
es that is used (weighted
Hamming). This is done by the following poly- [Konie
zny et al., 2001℄ S. Konie
zny, J. Lang,
nomial redu
tion M from max-sat-asgodd: to and P. Marquis. Distan
e-based merg-
any s.t. V ar() V = fx1; : : : ; xng we asso
iate ing: a general framework and some
om-
M () = hE = ff ni=1 xi gg; IC = ; = xn i and plexity results. Te
hni
al report, IRIT,
the weighted Hamming distan
e dHq indu
ed by ftp://ftp.irit.fr/pub/IRIT/RPDMP/DBMC.ps.gz,
q s.t. 8i 2 1 .. n, q(xi ) = 2n i . 2001.
[Konie
zny, 2000℄ S. Konie
zny. On the dieren
e
between merging knowledge bases and
ombining
them. In Pro
. of KR'00, pages 135{144, 2000.
[Lafage and Lang, 2000℄ C. Lafage and J. Lang. Log-
Referen
es i
al representation of preferen
es for group de
ision
[Baral et al., 1991℄ C. Baral, S. Kraus, and J. Minker. theory. In Pro
. of KR'00, pages 457{468, 2000.
Combining multiple knowledge bases. IEEE [Lafage and Lang, 2001℄ C. Lafage and J. Lang.
Transa
tions on Knowledge and Data Engineering, Propositional distan
es and preferen
e representa-
3(2):208{220, 1991. tion. In Pro
. of ECSQARU'01, pages 48{59, 2001.
[Liberatore and S
haerf, 1998℄ P. Liberatore and
M. S
haerf. Arbitration (or how to merge knowl-
edge bases). IEEE Transa
tions on Knowledge and
Data Engineering, 10(1):76{90, 1998.
[Liberatore and S
haerf, 2000℄ P. Liberatore and
M. S
haerf. Brels: a system for the integration of
knowledge bases. In Pro
. of KR'00, pages 145{152,
2000.
[Lin and Mendelzon, 1999℄ J. Lin and A. O. Mendel-
zon. Knowledge base merging by majority. In Dy-
nami
Worlds: From the Frame Problem to Knowl-
edge Management. Kluwer, 1999.
[Nebel, 1989℄ B. Nebel. A knowledge level analysis of
belief revision. In Pro
. of KR'89, pages 301{311,
1989.
[Nebel, 1998℄ B. Nebel. How hard is it to revise a belief
base? Handbook of Defeasible Reasoning and Un
er-
tainty Management Systems, Vol. 3: Belief Change,
pages 77{145, 1998.
[Papadimitriou, 1994℄ C. H. Papadimitriou. Compu-
tational Complexity. Addison-Wesley, 1994.
[Pinkas, 1995℄ G. Pinkas. Reasoning, nonmonotoni
-
ity and learning in
onne
tionist networks that
ap-
ture propositional knowledge. Arti
ial Intelligen
e,
77:203{247, 1995.
[Revesz, 1997℄ P. Z. Revesz. On the semanti
s of arbi-
tration. International Journal of Algebra and Com-
putation, 7(2):133{160, 1997.
[Wagner, 1987℄ K. W. Wagner. More
ompli
ated
questions about maxima and minima, and some
lo-
sures of NP. Theoreti
al Computer S
ien
e, 51:53{
80, 1987.
[Yager, 1998℄ R. R. Yager. On ordered weighted av-
eraging aggregation operators in multi-
riteria de
i-
sion making. IEEE Transa
tions on Systems, Man
and Cyberneti
s, 18:183{190, 1998.