the sg family

24
The SG family Different kinds of knowledge facts, rules and constraints • Generic deduction problem: given a KB K =(…) and a SG Q, is Q deducible from K? E.g. SG formalism: K={facts} deduction = projection Different formalisms obtained depending on the composition of K and the definition of deduction

Upload: breanna-leach

Post on 01-Jan-2016

17 views

Category:

Documents


0 download

DESCRIPTION

The SG family. Different kinds of knowledge facts , rules and constraints Generic deduction problem: given a KB K =(…) and a SG Q , is Q deducible from K ?.  Different formalisms obtained depending on the composition of K and the definition of deduction. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: The  SG  family

The SG family

• Different kinds of knowledge

facts, rules and constraints

• Generic deduction problem: given a KB K =(…) and a SG Q, is Q deducible from K?

E.g. SG formalism: K={facts} deduction = projection

Different formalisms obtained depending on the composition of K and the definition of deduction

Page 2: The  SG  family

SG

SR+rules

SGC

+constraints

SRC SEC

SREC

facts

inference rules evolution rules

Page 3: The  SG  family

Rules

« Every researcher is member of a project »

« The relation near is symetrical (on locations) »

Researcher member ProjectR1

Location Locationnear

near

R2

• A rule expresses knowledge of form «if Hypothesis then Conclusion»

• NB: a fact can be seen as a rule with empty hypothesis

Page 4: The  SG  family

SG bicoloré

• Un SG bicoloré est un SG muni d'une coloration de ses sommets avec deux couleurs {0,1}

• On impose que le sous-graphe induit par les sommets de couleur 0 soit un SG syntaxiquement correct (i.e. si un sommet relation est de couleur 0, tous ses voisins aussi)

sinon l'application de la règle peut produire un SG "mal formé"

Page 5: The  SG  family

Person

Employee

role emp

Compagny

Manager

emp

Person role

agent

Decision

object

Salary poss

Sommet frontière : sommet concept de couleur 0 ayant au moins un voisin (relation) de couleur 1

Hypothèse de la règle : sous-graphe de couleur 0

Conclusion de la règle :sous-graphe de couleur 1+ sommets frontières

Page 6: The  SG  family

Rule application

Office:#124

• A rule is applicable to a fact G if there is a projection from its hypothesis to G

• The result of the rule application is obtained by adding its conclusion to G according to (each frontier node c of the conclusion is merged with (c))

(+normalization if necessary)

…near nearOffice:#125

Location Locationnear

near

R2

G

nearG’

Page 7: The  SG  family

Person

Employee

role

R

emp

Compagny

Manager

emp

Person role

agent

Decision

object

Salary poss

GMan:J role emp

CarBuilder: P

Manager

Employee

Manager

Tc

Page 8: The  SG  family

Logical semantics

X (Y H[XY] Z C[XZ])

Researcher member Project

x (Res(x) y Project(y) member(x,y))

x y

Person worksIn Compagny

Salary

poss x ( y Person (x) (Comp(y) worksIn(x,y))

z (Salary(z) poss(x,z)))

X : frontier node variablesY : other variables of color 0 nodesZ : variables of color 1 nodes

x

XY(… Z…)

Page 9: The  SG  family

Définition usuelle des règles

• Un SG = un lambda-SG sans sommet distingué (n = 0)

• Règles et contraintes : couples de lambda-SGs avec même nombre de sommets distingués

( c11 ... c1n G1, c21 ... c2n G2)

• Un lambda SG est obtenu à partir d'un SG G en distinguant certains sommets concepts génériques c1 ... cn, n 0. On note c1 ... cn G

• c1 ... cn G) : comme G) mais en laissant libres les variables associées à c1 ... cn

Page 10: The  SG  family

R = ( c11 ... c1n G1, c21 ... c2n G2)

• Application de R sur un SG G selon une projection : G1 -> G

• on ajoute G2 à G

• puis on fusionne chaque c2i avec (c1i)

• Interprétation logique :

• on associe la même variable à c1i et c2i notons x1 ... xn ces variables

• on construit la formule :x1...xn(( c11 ... c1n G1) ( c11 ... c1n G1))

Voir cas "hypothèse vide" et "conclusion vide"

Page 11: The  SG  family

SG bicoloré versus couple de lambda-SG

Personne : *x aPourEnfant Personne

Parent : *x

SI

ALORS

(Plaçons-nous dans le cadre des types conjonctifs)

Personne aPourEnfant Personne

ParentIl faut que la conclusion puissecomporter des liens de coréférence

Page 12: The  SG  family

SR : facts + rules

Deduction problem: given a KB K ={facts, rules} and a SG Q, is Q deducible from K,i.e. is there a sequence of rule applications leading to a SG answering Q (ie Q projects to it)?

K

ruleapplications

Q

Forward and backward chaining mechanisms

facts

Page 13: The  SG  family

The forward and backward chaining mechanisms are sound and complete :

Q deducible from K iff (Q) logically deducible from (K)

Soundness and completenessof graph operations

[completeness up to normality conditions

for forward chaining]

(S), (facts), (rules)

Page 14: The  SG  family

Decidability ?

• Deduction in SR is only semi-decidable

• Decidable specific cases?

• SR is a computation model , ie one can simulate a Turing machine (représentation du problème de l'arrêt d'une MdT – pour une entrée particulière – dans SR)

ex: range-restricted rule : no generic concept in conclusion (frontier nodes excepted)

(1) build the full SG F from the KB K F exists!

(2) check whether Q projects to F

Page 15: The  SG  family

Plus généralement :

• Observation : une application de règle est inutile si elle produit un graphe équivalent au graphe d'origine

• Def : un SG G est dit plein (full) par rapport à un ensemble de règles R si toute application d'une règle de R sur G produit un graphe équivalent à G.

• Pté : étant donnés G et R , s'il existe une dérivation menant à un graphe plein, alors la forme irredondante de ce graphe est unique (modulo isomorphisme)

• Def : ensemble de règles à expansion finie : tel que pour tout G, il existe une séquence (finie) d'applications de règles menant à un graphe plein.

En ce cas, déduction décidable

Page 16: The  SG  family

SG

SR+rules

SGC

+constraints

SRC SEC

SREC

facts

inference rules evolution rules

Page 17: The  SG  family

Office

SGC: facts and constraints

Positive constraint C+ Negative constraint C-

in in

worksWith

Person Person

Office

in

HeadOfGroup

Office

in

Secretary

near

« The boss office must be nearall secretary offices »

« Persons working togethershould not share an office »

• A constraint expresses knowledge of form « if A is found so must B » (C+) « if A is found B must not » (C-)

Page 18: The  SG  family

"Idea" : G satisfies a positive constraint C+ if every projection from Condition(C+) to G can be extended to a projection from C+ to G

Office

in

HeadOfGroup

Office

in

Secretary

near

C+

Office:#3

in

HeadOfGroup

Office:#2

in

Secretary:K.

near

GSecretary:L.

in Office:#10

near

nearnear

on verra par la suite que cette définition doit être précisée

Page 19: The  SG  family

Office

in in

worksWith

Person Person

C-

Def: G satisfies a negative constraint C- if no projection from Condition(C-) to G can be extended to a projection from C- to G

worksWith

Researcher:K. Researcher

G

Office:#3

in

Office

in

There is no projection from C- to G

Page 20: The  SG  family

But previous definition is not good enough

t

C+

r t

G1

r t

G2

r

t

G1 and G2 are equivalent. Thus they should be both consistent or both inconsistent w.r.t. C+

Page 21: The  SG  family

Definition :

G satisfies a positive constraint C+ if

every projection from Condition(C+) to irr(G) (the irredundant form of (G)) can be extended to a projection from C+ to irr(G).

Ce problème de graphes équivalents qui ne se comportent pas de la même façon face à une contrainte ne se pose pas avec les contraintes négatives. Pourquoi ?

Page 22: The  SG  family

• Def: two constraints C1 and C2 are equivalent if any SG that satisfies C1 also satisfies C2, and reciprocally.

• Any negative constraint can be colored in 1 (interdiction part) yielding an equivalent constraint.

• Any negative constraint can be transformed into an equivalent positive constraint

" if C- must [NotThere]"

• Let C- be a negative constraint. Let G1 and G2 be equivalent SGs. G1 satisfies C- iff G2 does.

Les contraintes négatives peuvent donc être vues comme un cas particulier de contraintes positives

Page 23: The  SG  family

worksWith

Office

in in

worksWith

Person Person

C- C'-

Office

in in

Person Person

Office

in in

worksWith

Person Person

C+NotThere

Page 24: The  SG  family

SGC-Consistency: Given a KB K, is K consistent, ie do the facts satisfy the constraints?

If K is not consistent, nothing can be deduced from it

KB K = {facts, constraints}

• Complexity

- Consistency is 2P-complete

- If negative constraints only, co-NP-Complete

SGC-Deduction: Given a consistent KB K and a SG Q, is Q deducible from K (is there a projection from Q to facts)?

[ Variante : Etant donnés K et Q, a-t-on :(1) K consistante et (2) Q se déduit de K ? ]

SGC : facts and constraints