Download as pdf or txt
Download as pdf or txt
You are on page 1of 58

Ch.

15 Including Uncertainty in Deductive Databases {74{


VS
Including Uncertainty in Deductive
Databases
 Including Lattice-based (and hence also fuzzy) uncertainty in
Deductive Databases.
 Including Probabilistic modes of uncertainty in Deductive Databases.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {75{
VS
Lattice-based Extensions of Deductive
Databases: The GAP Paradigm
 GAP = \Generalized Annotated Program."
 Given any complete lattice, (L; v), we can de ne a logical
language within which GAPs may be written.
 This logical language has two parts:
{ Ordinary logic language: this is de ned in the usual way
from a set of constant symbols, function symbols, vari-
able symbols, and predicate symbols. Concepts of terms,
atoms, literals etc. are as usual.
{ Annotation language: this is a special lattice based lan-
guage, de ned in terms of the lattice (L; v) selected.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {76{
VS
Annotation language
1. Lattice constants: Every member of L is a lattice constant.
2. Lattice variables: We assume the existence of a set (in nite)
of lattice variables.
3. Lattice functions: We assume the existence of some set of
prede ned, implemented functions f1; : : :; fk on the lattice L.
4. Lattice terms: These are de ned inductively as follows:
(a) Every lattice variable is a lattice term.
(b) Every lattice constant is a lattice term.
(c) If t1 ; : : :; tn are lattice terms, and fi is an n-ary lattice
term, then fi(t1 ; : : :; tn) is a lattice term.

EX: Lattice = all real numbers with +1; 1 added, then ((V +
5)  (V 0 + 2)) div 4 is a lattice term.
So is (3 + 4).

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {77{
VS
GAP Language
 Mix of logical language and annotation language.
 If A is an atom in the logical language, and V is a lattice term,
then A : V is an annotated atom.
 EX: at(john; office) : 0:8 says (informally) that there is at
least an 80% certainty (not probability !) that John is in oce.
 If A0 : 0; : : :; An : n are annotated atoms, where 1; : : :;
n contain no lattice functions, then
A0 : 0 A1 : 1 ^    ^ An : n
is called an annotated clause.
 EX: Return to our face relation, and suppose our lattice is the
set of all closed intervals of [0; 1]. Then the following annotated
clause nds all people with whom Person1 was seen.
V1 + V2
seen(Person1; Person2) :
2

face(File; X1; Y1; X2; Y2; Person1) : V1 ^


face(File; X3; Y3; X4; Y4; Person2) : V2 ^
diff(X1; Y1; X2; Y2; X3; Y3; X4; Y4)
 Answer to above:
Person1 Person2 Unc
John Smith Mark Bloom V1+2 V2
John Smith Ted Lewis V1+2 V3
 GAP: nite set of annotated clauses.
Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved
Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {78{
VS
GAP Model Theory
 Interpretation: map from the Herbrand base of a GAP to
the set L
 EX: Take the GAP
p(a) : 1 :
p(b) : 0:8 :
q(X ) : V2 p(X ) : V:
Two possible interpretations are:
I (p(a)) = 1; I (p(b)) = 0:8; I (q(a)) = 0:7; I (q(b)) = 0:4
1 1 1 1

I (p(a)) = 1; I (p(b)) = 0:4; I (q(a)) = 0:2; I (q(b)) = 0:4


2 2 2 2

 Satisfaction:
1. I satis es a ground atom A :  i   I (A).
2. I satis es (F ^ G) i I satis es F and I satis es G.
3. I satis es (F _ G) i I satis es F or I satis es G.
4. I satis es F G i I satis es F or I does not satisfy G.
5. I satis es (8x:)F [x] i for all ground terms t from the
domain over which x ranges, I satis es F [x=t]. Here,
F [x=t] refers to the simultaneous replacement of all free
occurrences of x in F by t.
6. I satis es (9x:)F [x] i for some ground term t from the
domain over which x ranges, I satis es F [x=t].

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {79{
VS
 I satis es the above GAP, but I does not (in particular, it
1 2

does not satisfy the instance of the third rule whne X = a abd
V = 1.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {80{
VS
GAPs: Model Theory (Contd.)
 DEF: We write I vI 0 i for all ground atoms A, I (A)vI 0(A):
 EX:
I (p) = 0:3; I (q) = 0:4
I 0(p) = 0:5; I 0(q) = 0:45
I 00 (p) = 0:35; I 00 (q) = 0:3
Then: I vI 0, I 00 vI 0, I 6 vI 00 I 00 6 vI .
 Theorem: Set of all interpetations is a complete lattice under
v ordering.

t(I )(A) = tfJ (A) j J 2 Ig:


u(I )(A) = ufJ (A) j J 2 Ig:
 If I ; I satisfy a rule, then so must I u I .
1 2 1 2

 Theorem: Suppose I 0 is any interpretation. Then the set


fI ? j I ? satis es P and I 0vI ?g has a unique v-minimal ele-
ment. As a consequence, P has a unique v-minimal model.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {81{
VS
GAPs: Fixpoint Theory
 Given any GAP P , we associate with P , an operator TP that
maps Herbrand interpretations to Herbrand interpretations.
 TP (I )(A) = tf j A :  Body is a ground instance of an
annotated clause in P such that I satis es Body g.
 TP (I )(A) may be (ineciently) computed as follos:
1. Let S = ;.
2. First replace P by the (possibly in nite) GAP grd(P ) ob-
tained by instantiating each annotated clause in P in as
many ways as possible.
3. For each ground annotated clause in grd(P ) that has a
head of the form A : , do the following: if I satis es the
body of the clause, then put  in S .
4. Compute the least upper bound of all elements of S and
set TP (I )(A) to this value.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {82{
VS
An Example
 fEX: Suppose P is the GAP given by

p(a) : 1 :
p(b) : 0:8 :
q(X ) : V2 p(X ) : V:
Suppose I1; I2 are the interpretations given below:
I1(p(a)) = 1; I1(p(b)) = 0:8; I1(q(a)) = 0:7; I1(q(b)) = 0:4
I2(p(a)) = 1; I2(p(b)) = 0:4; I2(q(a)) = 0:2; I2(q(b)) = 0:4
Then TP (I1 ) is given by

TP (I )(p(a))
1 = 1
TP (I )(p(b))
1 = 0:8
TP (I )(q(a))
1 = 0:5
TP (I )(q(b))
1 = 0:4
However, TP (I2 ) is given by

TP (I )(p(a))
2 = 1
TP (I )(p(b))
2 = 0:8
TP (I )(q(a))
2 = 0:5
TP (I )(q(b))
2 = 0:2

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {83{
VS
Another Example
 P is given by:
p(a) : 0 :
p(X ) : V + 1
2
p(X ) : V:
q(X ) : 1 p(a) : 1:
 I is the interpretation such that I (p(a)) = I (q(a))
0 0 0

= 0.
 Ij = TPj (I0).
 Ijomega = tj<! Ij . For example, we can easily see that
I (p(a)) = 0:5; I (q(a)) = 0
1 1

I (p(a)) = 0:75; I (q(a)) = 0


2 2

I (p(a)) = 0:875; I (q(a)) = 0


3 3


0 1j
Ij (p(a)) = 1 @ 1 A ; Ij (q (a)) = 0
2

I! (p(a)) = 1; I! (q(a)) = 0
I! (p(a)) = 1; I! (q(a)) = 1
+1 +1

I! = I!
+2 +1

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {84{
VS
GAP Fixpoint Theory
 Theorem: TP is monotonic, i.e., if I vI , then TP (I )vTP (I ).
1 2 1 2

 Theorem: I satis es P i TP (I )vI .


 Theorem: fI j TP (I ) = I g has a unique, v-minimal element
called the least xpoint of TP , and denoted lfp(TP ).
 Theorem: For any ground atom A and any truth value , it
is the case that A :  is a logical consequence of the GAP P
i vlfp(TP )(A).

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {85{
VS
GAPs: Query Processing
 Atomic lattice constraint: Expression of the form tvt0
where t; t0 are lattice terms.
 Lattice constraint: Boolean combination of atomic lattice
constraints.
 Query: Expression of theform
(9X1 ; : : :; Xk )(A1 : 1 ^ : : : ^ An : n ) ^ C
where:
{ A :  ; : : :; An : n are annotated atoms
1 1

{ each i is a lattice element (not a lattice variable )


{ X ; : : :; Xk are all the variables occurring in A ; : : :; An
1 1

{ C is a lattice constraint (possibly empty).


 EX:
(9X; Y )p(X; Y ) : 1 ^ q (Y ) : 0:8:
\Find X; Y such that p(X; Y ) is true with certainty 1 or more
and q (Y ) is true with certainty 0:8 or more."
 Query Processing Problem: Find ground substitution  for the
variables in the query such that
(8)(A1  : 1 ^ : : : ^ An : n)
is a logical consequence of the GAP P .

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {86{
VS
GAPs: Examples of answers
 Take P to be:
p(X; Y ) : V q(X; Z ) : V ^ r(Z ) : V:
q(X; a) : 0:7 :
r(a) : 0:6 :
 Consider the query p(X; Y ) : 0:6.
  = fY = ag is an answe because (8X )p(X; a) : 0:6 is a
1

logical consequence of P .
 Consider the query p(X; Y ) : 0:7. 1 is not an answer to the
query because (8X )p(X; a) : 0:7 is not a logical consequence
of P .

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {87{
VS
GAPs:Annotated Resolvents
 Query Q:
(A1 : 1 ^ : : : ^ An : n ) ^ C
 Annotated clause C :
A: B :  ^    ^ Bm : m:
1 1

 Suppose Q; C are standardized apart.


 Suppose  is an mgu of Ai and A for some 1  i  n.
 Then:
(A :  ^ : : : ^ Ai : i ^ B :  ^    ^ Bm : m ^
1 1 1 1 1 1

Ai : i ^ : : : ^ An : n) ^ (C ^ iv)
+1 +1

is said to be an annotated resolvent of Q and C via mgu  .

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {88{
VS
GAPs:Annotated Resolvent Example
 P is:
p(X; Y ) : V q(X; Z ) : V ^ r(Z ) : V:
q(X; a) : 0:7 :
r(a) : 0:6 :
 Query: p(X 0; Y 0) : 0:6.
 Clause from P :
p(X; Y ) : V q(X; Z ) : V ^ r(Z ) : V:
 Uni er: X = X 0; Y = Y 0
 Annotated resolvent:
q(X 0; Z ) : V ^ r(Z ) : V ^ 0:6vV:

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {89{
VS
GAPs:Annotated Refutations
 Annotated Deduction: inite or in nite sequence
(Q0; C0 ; 0); : : :; (Qj ; Cj ; j ); : : :
where
1. Q0 = Q and
2. Cj is a renaming of an annotated clause in P for all j  0
and
3. Qj +1 is an annotated resolvent of Qj and Cj via mgu j .
 Annotated refutation of Q from P is a nite annotated
deduction
(Q0; C0; 0); : : :; (Qr 1; Cr 1 ; r 1)
(r  1) from P with respect to Q such that the annotated
resolvent of Qr 1 and Cr 1 via mgu r 1 is the empty query
(i.e., a query containing no annotated atoms, and a solvable
constraint part).

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {90{
VS
GAPs:Annotated Refutation Example
 P is:
p(X; Y ) : V q(X; Z ) : V ^ r(Z ) : V:
q(X; a) : 0:7 :
r(a) : 0:6 :
 Query: p(X 0; Y 0) : 0:6.
 Annotated refutation of this query is:
1. (Q ; Cl ; fX 0 = X; Y 0 = Y g).
0 1

2. (q (X; Z ) : 0:6 ^ r(Z ) : V ^ 0:6vV; Cl ; fZ = ag). 2

3. (r(a) : V ^ (0:6vV ^ V v0:7); Cl ; fg). 3

4. (0:6vV ^ V v0:7 ^ V v0:6; ; ).


 The last query above has no atoms (only a lattice constraint)
and this lattice constraint is solvable.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {91{
VS
>
f t

?
Figure 1: The lattice FOUR

Annotated Refutation is Not Complete


Two reasons for this:

 The rst reason is that the operator TP may lead to xpoints


at ordinals greater than ! , leading some atoms to have in nite
justi cations.
 The second reason is that the notion of deduction does not al-
low us to \merge" multiple derivations. For example, consider
a lattice such as the lattice FOUR and consider the simple GAP
P given by

p:t :
p:f :
 p : > is a logical consequence of this GAP.
 But there is no annotated refutation of this query from P .

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {92{
VS
GAPs: Fixpoint Reachability
 Given any function f on a complete lattice (L; v), we may
de ne the iterations of f as follows:

f " 0 = ?:
f " = f (f " ( 1)) for successor ordinals :
f "  = t <f " for limit ordinals :
 Fixpoint reachability: We say that a GAP P possesses
the xpoint reachability property if, whenever vlfp(TP )(A),
it is the case that vTP " k(A) for some integer k.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {93{
VS
GAPs: Reductants and Closure
 Suppose C ; C are, respectively:
1 2

A :
1 1 Body1:
A :
2 2 Body2:
 If A ; A are uni able via mgu , then the clause
1 2

(A :  t  1 Body1 ^ Body2)
1 2

is called a reductant of C1 ; C2.


 The closure of P is de ned to be <! (P ) where
1. <0 (P ) = P .
2. <j +1 (P ) = <j (P ) [ fC j there exist C1 ; C2 2 <j (P )
such that C is the reductant of C1; C2g.

newfoilGAPs: Reductants and Closure Example

 P:

p:t :
p:f :
 Reductant:
p:> :

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {94{
VS
 Closure:
p:t :
p:f :
p:> :

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.1 GAPs {95{
VS
Soundness and completeness of annotated
resolution
 Let P be any GAP and Q be a query.
 (Soundness) If there exists an annotated refutation of Q,
then Q is a logical consequence of P .
 ( Completeness) Suppose Q is a logical consequence of P
and P satis es the xpoint reachability condition. Then there
exists an annotated refutation of Q with respect to the closure
of P .
 More ecient procedures which avoid construction of closures
have been developed by:
{ Lu, Murray, and Rosenthal
{ Adali and Subrahmanian

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {96{
VS
Probabilistic Knowledge Bases
 Lattice based approach is not probabilistic.
 Like GAPs, Probabilistic KBs will have two languages { a log-
ical language identical to that for GAPs, and an annotation
language.
 The annotation language for Probabilistic KBs is di erent from
that for GAPs.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {97{
VS
Probabilistic KBs: Annotated Language
 Annotation variables: Range over the reals in the unit
interval [0; 1].
 Annotation functions: Maps (of di erent arities) from
[0; 1]n to [0; 1].
 Annotation items:  is called an annotation item if it is
one of the following:
1. a constant in [0,1], or
2. an annotation variable in the underlying language, or
3. of the form f (1; : : :; n), where f is an annotation func-
tion of arity n and 1 ; : : :; n are annotation items.
 Annotation Term: [ ;  ] is called an annotation (term)
1 2

if i(i = 1; 2) is an annotation item.


 EX: The following are annotation terms:
1. [0:1; 0:2]
2. [0:7; V2 ]
3. [ jV1 2 V2j ; jV1+2 V2j ]

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {98{
VS
Probabilistic KBs: Logical Language
 conj (BL) = fA ^ : : : ^ An j n  1 is an integer and
1

A ; : : :; An 2 BL and for all 1  i; j  n; i 6= j ) Ai 6= Aj g


1

 disj (BL ) = fA _: : :_Anjn  1 is an integer and A ; : : :; An 2


1 1

BL and for all 1  i; j  n; i =6 j ) Ai =6 Aj g,


where all the Ai's are ground atoms.
 Basic formula: Any member of conj (BL) [ disj (BL ).
 Thus, basic formulas are either conjunctions of atoms, or dis-
junctions of atoms, but not a mix of the two.
 bf (BL) denotes the set of all basic formula.
 Annotated Conjunctions/Disjunctions: If C is a con-
junction of atoms, not necessarily ground, and  is an annota-
tion, then C :  is called an annotated conjunction. Similarly,
if D is a disjunction, not necessarily ground, then D :  is
called an annotated disjunction.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {99{
VS
Examples
1. at(25; 26; robot) : [0:3; 0:4]: This annotated atom says that
the probability that the robot is at location (25; 26) lies in the
range 30% to 40%.
2. (at(25; 26; robot1) ^ at(25; 36; robot2)) : [0:2; 0:25]: This an-
notated conjunction says that the combined probability that
robot1 is at location (25; 26) and robot2 is at location (25; 36)
lies in the range 20% to 25%.
3. (at(25; 26; robot1) _ at(25; 36; robot2)) : [0:4; 0:45]: This an-
notated conjunction says that the combined probability that
either robot1 is at location (25; 26) or robot2 is at location
(25; 36) lies in the range 40% to 45%.
4. (at(25; 26; robot1) _ at(25; 36; robot2)) : [ jV1 2 V2j ; jV1+2 V2 j ]: This
says that the combined probability that either robot1 is at loca-
tion (25; 26) or robot2 is at location (25; 36) lies in the range
[ jV1 2 V2j ; jV1+2 V2 j ], where the values of V1 ; V2 are currently un-
known.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {100{
VS
Probabilistic KBs
 pf-clause:
F :
0 0 F :  ^ : : : ^ Fn : n
1 1

where:
 F ; : : :; Fn are basic formulas,
0

  ; : : :; n are annotations such that all the annotation vari-


0

ables that appear in 0 also appear in at least one of 1; : : :; n.


 p-clause: If F0 is an atom, then the clause is called a p-
clause.
 Pf-KB: A nite set of pf-clauses.
 p-KB: A nite set of p-clauses.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {101{
VS
Examples
 Consider the p-clause
object(a) : [max(0; V 2 0:1 ); V +2 0:1 ]
1 2
sensor1(a) : [V ; V ] ^ 1 1

sensor2(a) : [V ; V ]: 2 2

This rule says if sensor1 reports that object a is present with


probability V and sensor2 reports that object a is present with
1

probability V , then we may conclude that object a is present


2

with a probability in the range : [max(0; V1 : ); V2 : ].2


01 +0 1
2

 Consider the rule


object(a) : [max(0; V 2 0:1 ); V +2 0:1 ]
1 2
(sensor1(a) ^
sensor2(a)) : [V1; V2]:
This p-clause is slightly di erent from the previous one because
it says that we may conclude that object a is present with a
probability in the range : [max(0; V1 20:1 ); V2 +0
2
:1 ] if the joint
probability that sensor1 and sensor2 detect object a is in the
interval [V1; V2].
 The statement
(object(a) ^ object(b)) : [0:9; 1] :
is a simple pf-clause that says that the probability that objects
a and b will occur is over 90%.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {102{
VS
Probabilistic KBs: Fixpoint Theory
 World: Any Herbrand interpretation.
 World probability density function: Mapping, WP :
fW ; : : :;
1

Wk g ! [0; 1] such that for all j = 1; : : :; k, WP (Wj )  0


Xk
and WP (Wj ) = 1.
j =1
 Intuition: We hypothesize multiple possible worlds, but only
one of them is the \actual" world (at a given point in time).
Atoms are either true or false { uncertainty arises because we
don't know which atoms are true and which are false, and
hence, we don't know which of the hypothesized worlds re ects
the actual state of the world.
 Formula function: Mapping h : bf (BL) ! C [0; 1],
where C [0; 1] denotes the set of all closed subintervals of the
unit interval [0; 1].
 Ordering on Formula function: h1  h2 i 8F 2
bf (BL); h1(F )  h2(F ).
 Intuition: h1  h2 may be read as h2 is sharper than h1.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {103{
VS
Fixpoint Operator for pKBs
 Associate with P an operator TP that maps formula functions
to formula functions, de ned as follows:
1. For ground atoms A: TP (h)(A) = \f[`; u] j A : [`; u]
B1 : 1 ^ : : : Bk : k is a ground instance of some clause
in P such that for all 1  i  k, h(Bi )  ig.
2. For conjunctions A ^ B : TP (h)(A ^ B ) = TP (h)(A)
ig
TP (h)(B ), where
ig is the policy for computing conjunc-
tions under the policy of ignorance.
3. For disjunctions A _ B : TP (h)(A _ B ) = TP (h)(A) ig
TP (h)(B ), where ig is the policy for computing disjunc-
tions under the policy of ignorance.
 Procedural explanation of TP (h):
1. First, nd all ground instances of rules in P having a head
of the form A : [`; u] and whose bodies are satis ed by h,
i.e., if Bi : i occurs in the body of the ground rule, then
h(Bi) must be a subset of i.
2. Collect the intervals [`; u] for all rules satisfying the above
condition.
3. TP (h)(A) is the intersection of all the intervals collected
above.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {104{
VS
Example of TP for pKBs
P:

object(a) : [max(0; V 2 0:1 ); V +2 0:1 ]


1 2
sensor1(a) : [V ; V ] ^
1 1

sensor2(a) : [V ; V ]:2 2

sensor1(a) : [0:6; 0:6] :


sensor2(a) : [0:4; 0:5] :

h:

h(object(a)) = [0; 0]
h(sensor1(a)) = [0:9; 0:9]
h(sensor2(a)) = [0:8; 0:8]

TP (h):

TP (h)((object(a)) = [0; 0]
TP (h)(sensor1(a)) = [0:6; 0:6]
TP (h)(sensor2(a)) = [0:4; 0:5]

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {105{
VS
p-KBs: TP continued
 Iterations of TP :
TP " 0(A) = [0; 1]
TP " (A) = TP (TP " ) if is a successor ordinal with
as its predecessor
TP " (A) = tfTP " j < g if is a limit ordinal
 Theorem: If h ; h are formula functions such that h  h ,
1 2 1 2

then TP (h )  TP (h ).
1 2

 Theorem: If h ; h are formulaTP has a least xpoint, lfp(TP ),


1 2

and furthermore, there exists an ordinal such that lfp(TP ) =


TP " .

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {106{
VS
Fixpoint Operator for pf-KBs
3-stage de nition.

 Stage 1: In the rst stage, we will de ne an operator, called


SP , that will merely behave like TP , except that it will apply
to basic formulas in the heads of pf-clauses as well. However,
it will not attempt to derive probabilities associated with the
atomic parts of such basic formulas.
 Stage 2: Once SP (h) is computed, the next step is to treat
SP (h) as a set of events whose probabilities are known.
 Stage 3: We may then set up a linear program, denoted LP (SP (h))
for each basic formula, F 2 bf (BL ), in our language, we can
compute a lower bound cF and an upper bound dF on the
probability of F . VP (h)(F ) is then set to [cF ; dF ] and we are
done.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {107{
VS
Fixpoint Operator for pf-KBs: Stage 1
 Given h, let MF = f j F :  F :  V : : : V Fn : n is
1 1

in grd(P ),  is a ground substitution to the set of annotation


variables that occur in the body of the above clause such that
 6= ;, and 8i; 1  i  n; h(Fi)  ig.
 If MF =, then SP (h)(F ) = [0; 1].
 Otherwise, SP (h)(F ) = T MF .
 SP (h)(F ) does not take into account the fact that the proba-
bility of F may be a ected by our knowledge of the probabilities
of \atomic" parts of F . This is handled in stage 2.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {108{
VS
Fixpoint Operator for pf-KBs: Stage 2
 Let LP (SP (h)) be the linear program obtained, as described
when we studied Chapter 13.
 That is, for each formula F , create a constraint of the form
X
lb  pi  ub
wij=F
where SP (h)(F ) = [lb; ub] and pi is a variable denoting the
probability of world wi.
 In addition add the constraint
sumwi pi = 1
 And that pi  0 for all i.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {109{
VS
Fixpoint Operator for pf-KBs: Stage 3
 We now de ne an operator VP from formula functions to for-
mula functions as follows:
a. If LP (SP (h)) has solutions, then for all F 2 bf (BL),
VP (h)(F ) = [cF ; dF ],0where 1
BB X C
cF = minLP SP h ( ( )) B@ pj CCA and
B
0Wj j=F and Wj 22 L 1
BB X CC
dF = maxLP SP h ( ( )) B@
B
p jC A.
Wj j=F and Wj 22 L
Note that this step involves the straightforward use of any
linear programming package/algorithm, numerous imple-
mentations of which are available in the commercial mar-
ketplace.
b. Otherwise, if LP (SP (h)) is empty, then for all F 2 bf (BL ),
TP (h)(F ) = ;.
 BOTTOM LINE: Computation of VP (Pf-KBs) is much
harder than TP (fpr p-KBs). This is the price paid for intro-
ducing non-atomic basic formulas into the heads of rules.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {110{
VS
Example of VP
 P:
(a ^ b) : [1; 1] :
a : [0; 0] :
b : [1; 1] :
 Suppose h(F ) = [0; 1] for all basic formulas F .
 Stage 1: First, we compute SP (h) and notice that
SP (h)((a ^ b)) = [1; 1]
SP (h)(a) = [0; 0]
SP (h)(b) = [1; 1]
It is immediately apparent that there is an inconsistency here.
After all, the probability of the event (a ^ b) cannot be 1 if
the probability of a is 0, as the second equality above seems
to indicate.
 Stage 2: Set t up the linear program LP (SP (h)) as follows.
Suppose w1 is the world in which both a and b are false, w2 is
the world in which a is true, but b is false, w3 is the world in
which b is true, but not a, and w4 is the world in which both
a and b are true. Then the constraints generated are

p = 1
4

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {111{
VS
p +p 2 4 = 0
p 3 = 1
p +p +p +p
1 2 3 4 = 1
pi  1for i = 1; : : :; 4
pi  0for i = 1; : : :; 4
 Stage 3: As the set of constraints described above is unsolv-
able, VP (h)(F ) = ; for all basic formulas F .

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {112{
VS
Another example
 P:
(a _ b) : [1; 1] :
a : [0:2; 0:5] :
b : [0:4; 0:6] :
 Suppose h(F ) = [0; 1] for all basic formulas F .
 Stage 1: First, we compute SP (h) and notice that
SP (h)((a _ b)) = [1; 1]
SP (h)(a) = [0:2; 0:5]
SP (h)(b) = [0:4; 0:6]
 Stage 2: LP (SP (h)) is:
p +p +p =1
2 3 4 0:2  p + p  0:5 2 4

0:4  p + p  0:6
3 4 p +p +p +p = 1
1 2 3 4

0  pi  1 for i = 1; : : :; 4
 Stage 3: For each formula F , we may compute VP (h)(F ) by
nding cF and dF , where cF is the minimal (dF , the maximal)
value of the expression
X
pj
wj j=F
subject to the above constraints. The values of VP (h)(F ) for
some di erent F 's are given below:

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {113{
VS
F
VP (h)(F )
(a _ b) [1; 1]
a [0:4; 0:5]
b [0:5; 0:6]
(a ^ b) [1,1]

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {114{
VS
Pf-KB Fixpoint Theory
 Theorem: For all basic formulas F , and all formula functions
h, VP (h)(F )  SP (h)(F ).
 Theorem: The VP operator is monotonic.
 Theorem: Suppose P is any pf-KB. Then VP has a least
xpoint, denoted lfp(VP ). Furthermore, there exists an ordinal
such that VP " = lfp(VP ).

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {115{
VS
pf-KBs vs. pKBs
1. First, every iteration of the application of the VP operator re-
quires writing down a linear program that is exponential in size
(with respect to the number of ground atoms in the language).
This exacts a huge price. In contrast, no linear programs at all
need to be constructed when dealing with pKBs.
2. Second, for each basic formula F , when computing VP , we
need to solve two linear programming problems: one to obtain
a lower bound on F 's probability, and another to obtain an
upper bound. This needs to be done in every iteration toward
the xpoint. It is well known that linear programming packages
start falling apart when the number of variables (i.e., number of
worlds) and/or the number of constraints in the linear program
run into the tens of thousands. In our case, the number of
Herbrand interpretations could well be very large.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {116{
VS
Probabilistic KBs: Model Theory
 Prob-Interpretation: Let WP be a world probability den-
sity function. De ne probabilistic interpretation IWP is a map-
ping from all basic formulas to [0; 1] and
0 1
BB X CC
IWP (F ) = B@ B
p jCA:
Wj j=F and Wj 22 L

 EX: Suppose
WP (w ) 1 = 0:4
WP (w ) 2 = 0:1
WP (w ) 3 = 0:2
WP (w ) 4 = 0:3
Then
IWP (a) = 0:4
IWP (b) = 0:5
IWP (a _ b) = 0:6
IWP (a _ :b) = 0:8

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {117{
VS
Probabilistic KBs: Satisfaction
 I j=p F : [c ; d ] i I (F ) 2 [c ; d ]
1 1 1 1 1 1

 I j=p (F : [c ; d ] ^ : : : ^ Fn : [cn; dn]) i for all 1  j  n,


1 1 1

I j=p Fj : [cj ; dj ]

 I j=p F : [c ; d ] F : [c ; d ] VV: : : VVFn : [cn; dn] i


0 0 0 1 1 1

I j=p F : [c ; d ] or I 6j=p (F : [c ; d ] : : : Fn : [cn; dn])


0 0 0 1 1 1

 I j=p (9x)(C ) i I j=p (C (x=t)) for some ground term t,


where x represents an object variable, and C (x=t) denotes the
replacement of all free occurrences of x in C by t

 I j=p (8x)(C ) i I j=p (C (x=t)) for all ground terms t, where


x represents an object variable

 I j=p (9V )(C ) i I j=p (C (V=c)) for some c 2 [0; 1], where
V represents an annotation variable

 I j=p (8V )(C ) i I j=p (C (V=c)) for all c 2 [0; 1] such


6 ;, where V represents an annotation variable
that (V=c) =
occurring in annotation term . Here, if  is an annotation
term, then the notation (V=c) is the annotation term that

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {118{
VS
is obtained by replacing all occurrences of V in  by c and
evaluating the resulting term (if it is ground).

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {119{
VS
Probabilistic KBs: Relationship between
semantic structures
World Probability Probabilistic
Density Function Interpretation

Formula
Function

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {120{
VS
Fully de nedness
 Fully de ned FFs: Formula function h is fully de ned i
for all basic formulas F 2 bf (BL); ;  h(F )  [0; 1].

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {121{
VS
From Formula Functions to World Prob
Density Functions
 Suppose h is a formula function. Then the set of world proba-
bility density functions determined by h, denoted WPD(h), is
given by
WPD(h) = fWP j  is a solution of LP (h) and
WP (wi) = pi for all wig.
 Given h, we may compute the set, WPD(h), of world proba-
bility density functions associated with h using the following
steps:
1. Construct LP (h).
2. Find every solution of LP (h).
3. For each solution  of LP (h), the function WP de ned
by WP (wi) = pi is in WPD(h).

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {122{
VS
Model theoretic theoresm on Probabilistic
KBs
 Suppose h is a formula function. Then the set of probabilistic
interpretations, denoted PRI(h), is given by fIWP j WP 2
WPD(h)g.
 Theorem: Suppose P is a pf-KB and h is a formula func-
tion. Then LP (SP (h)) and LP (VP (h)) have the same set of
solutions.
 Theorem: Let h be a formula function. Then VP (h) is fully
de ned i WPD(h) 6= ;.
 Theorem: Suppose VP (h) is fully de ned. Then for all basic
formulas F , VP (h)(F ) is the smallest interval that contains
fI (F ) j I 2 PRI(h)g.
 Theorem: Suppose P is a pf-KB, and h is a fully de ned
formula function. Then each member of PRI(h) satis es P i
VP (h)  h.
 Theorem: Suppose P is a pf-KB. Then
1. VP has a unique least xpoint denoted by lfp(VP ).
2. For any annotated basic formula F : [cF ; dF ]: F : [cF ; dF ]
is a logical consequence of P i lfp(VP )(F )  [cF ; dF ].
 Theorem: Suppose P is a pf-KB. P is consistent (i.e., there
exists a probabilistic interpretation that satis es P ) i lfp(VP )
is fully de ned.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {123{
VS
Probabilistic KBs: Query Processing
 Concept of uni cation changes.
 (A ^ : : : ^ An) :  and C  (B ^ : : : ^ Bm) :  i
1 1 2 1 2

fAij1  i  ng = fBij1  i  mg.


 Similarly for disjunctions.
 EX:
Exp1 = (p(X; a) _ p(Y; b)):
Exp2 = (p(Z; Z ) _ p(c; W )):
 = fX = a; Z = a; Y = c; W = bg
1

 = fX = c; Z = b; Y = b; W = ag
2

 Bottom Line: MGU's are not necessarily unique !

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {124{
VS
Probabilistic KBs: Uni cation
 Let Uni(F ; F ) denote s the set of all uni ers of F ; F . Sup-
1 2 1 2

pose  ;  2 Uni(F ; F ). We say the following:


1 2 1 2

1.    i there exists a substitution such that  =  .


1 2 1 2

2.    i    and    . We use the notation []


1 2 1 2 2 1

to denote the equivalence class of , i.e., [] = f0 j   0,


where 0 2 Uni(F ; F )g. 1 2

3. []  [0] i   0.


4.  is a maximally general uni er (max-gu) of F1 ; F2 i there
does not exist 1 2 Uni(F1 ; F2 ) such that [] < [1 ].
 Theorem: If F ; F are basic formulas that are uni able,
1 2

then they have at least one max-gu.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {125{
VS
Resolution in pf-KBs
 Concept of closure and xpoint reacahbility condition used in
GAPs is also needed here.
 p-constraint: A conjunction of constraints having the form:
1  2 or 1 = 2, where i(i = 1; 2) is an annotation item.
 Constrained pf-clause: Of the form:
F :
0 0 F :  ^ : : : ^ Fn : n ^ ConC :
1 1

where
F : 0 F :  ^ : : : ^ Fn : n
0 1 1

is a pf-clause and ConC is a p-constraint.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {126{
VS
Compiled version of a pf-KB
 Similar to the closure of a GAP.
 LetP = f C ; : : :; Ck g, where for all 1  i  k, Ci  Fi :
1

[i; i] Bodyi and [i; i] is an annotation term.


1. We may associate a set, LP (P ), of linear constraints with
P as follows: for all0 1  i  k, the inequality
1
B X C
i  BB@ pj CCA  i
Wj j=Fi and Wj 22BL
is in LP (P ). In addition, LP (P ) contains the following
constraints: X
pj = 1; and
Wj 22BL
(8Wj 2 2BL ); pj  0:
2. The closure of P , denoted cl(P ), is now de ned as follows.
Let
cl(P ) = P S
f F : [F ; F ] Body1 V : : : V Bodyk V Consol j F 2
bf (BL); 0 1
BB X CC
F = minLP P
( ) @B p jCA;
B
0Wj j=F and Wj 22 L 1
BB X CC
F = maxLP (P ) @ B pj CA ;
Wj j=F and Wj 22BL
and Consol denotes the conjunction of constraints on the
annotation terms for LP (P ) to have solutions. g

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {127{
VS
Example of Closure
 P:
a : [0:3; 0:6] c : [0:3; 0:4]:
b : [0:4; 0:8] c : [0:1; 0:4]:
 cl(P ) is very large and contains a pf-clause having each of the
following basic formulas in the head: (a _ b), (a ^ b), (a _ c),
(a ^ c), (b _ c), (b ^ c), (a _ b _ c), and (a ^ b ^ c).
 cl(P ):
(a _ b) : [0:4; 1] c : [0:3; 0:4] ^ c : [0:1; 0:4]:
(a ^ b) : [0; 0:6] c : [0:3; 0:4] ^ c : [0:1; 0:4]:
(a _ c) : [0:3; 1] c : [0:3; 0:4] ^ c : [0:1; 0:4]:
(a ^ c) : [0; 0:6] c : [0:3; 0:4] ^ c : [0:1; 0:4]:
(b _ c) : [0:4; 1] c : [0:3; 0:4] ^ c : [0:1; 0:4]:
(b ^ c) : [0; 0:8] c : [0:3; 0:4] ^ c : [0:1; 0:4]:
(a _ b _ c) : [0:4; 1] c : [0:3; 0:4] ^ c : [0:1; 0:4]:
(a ^ b ^ c) : [0; 0:6] c : [0:3; 0:4] ^ c : [0:1; 0:4]:

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {128{
VS
Compiled Version of P
 redun(P ) = fF : [0; 1] j F 2 bf (BL)g.
 De ne the compiled version of P , denoted by com(P ), as
follows:
com(P ) = S cl(Q) for each subset Q of redun(P ) S grd(P ).
 Note: or each basic formula F , if F :  is a logical consequence
of P , then there exists a single clause C in com(P ) having a
ground instance, C , such that the head of C is of the form
F : 0 and   0.
 This makes it much easier to de ne a linear resolution-style
query processing procedure, similar to GAPs, that is sound
and complete.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {129{
VS
SLDp-Resolvents
 Query: If F :  ; : : :; Fn : n are annotated basic formulas,
1 1

and ConQ is a conjunction of constraints on the annotation


terms, then
9(F :  ^ : : : ^ Fn : n ^ ConQ)
1 1

is a constrained p-query.
 Clause C :
G : 0 0 G :  ^ : : : ^ Gm : m ^ ConC
1 1

 Suppose G and Fi are uni able with max-gu .


0

 SLDp-resolvent:
9((F :  V : : : V Fi : Vi VVG :  V : : : V Gm : m V Fi :
1 1 1 1 1 1 +1

i : : : Fn : n)
+1
V
ConC V ConQ V   i ) 0

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {130{
VS
SLDp-Refutations
 SLDp-Deduction: Sequence hQ ; C ;  i; : : :; hQr; Cr; ri; : : :,
1 1 1

where for all i  1, Ci is a renamed version of a clause in


com(P ); and Qi is an SLDp-resolvent of Qi and Ci through
+1

max-gu i.
 SLDp-Refutation: Finite SLDp-deduction
hQ ; C ;  i; : : :; hQn; Cn; ni
1 1 1

where the SLDp-resolvent of Qn and Cn via max-gu n, de-


noted by Qn+1, satis es the following two conditions:
1. The query part of Qn+1 is empty.
2. The constraint part of Qn+1 is satis able. 1 : : :n is called
the computed answer substitution.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997
Sec. 15.2 Probabilistic Knowledge Bases {131{
VS
Theorems
 DEF: A pf-KB P is compact if for all F 2 bf (BL), there
exists an integer F < ! such that lfp(TP )(F ) = TP "
F (F ).
 Theorem: SLDp-refutation is sound.
 Theorem: SLDp-refutation is complete for compact pf-KBs.

Zaniolo|Ceri|Faloutsos|Snodgrass|Subrahmanian| Zicari|All Rights Reserved


Advanced Database Systems Morgan Kaufmann Copyright
c 1997

You might also like