Académique Documents
Professionnel Documents
Culture Documents
c G. Lakemeyer
Gerhard Lakemeyer
xIn(x,box) xIn(x,box) Light(x) heaviest-block = blockA heaviest-block = favourite(john) Express knowledge that is incomplete Expressive power = what can be left unsaid
LKB/SS-2014
2 / 18
Knowing Who
Consider Teaches(ai-course,father(bill)) Would like to know Who is the father of Bill?
c G. Lakemeyer
Might have father(bill) = best-friend(george) But then Who is the best friend of George? Might have father(bill) = mister-smith But Who is Mister Smith?
LKB/SS-2014
3 / 18
Standard Names
Two options:
1
say you never really know who teachers are. Just a collection of Teaches(ai-course,t),
c G. Lakemeyer
ti = tj ,
2
tj = tk
partition domain into equivalence classes named #1,#2,#3, . . . know who = know which equivalence class father(bill) = #27
LKB/SS-2014
The Logic L
L is is a rst-order logic with function and predicate symbols of any arity equality
c G. Lakemeyer
standard names as the domain of discourse Note: What makes L special is that that the domain of discourse is xed and also part of the language. Caveat: Basic knowledge of classical rst-order logic is assumed. We will go over the denitions rather quickly and focus on using logic for KR.
LKB/SS-2014
5 / 18
Terms in L
write: f( ) as f
#
Standard names: 1,#2, . . . (Standard names are syntactically like constants. We often call standard names simply names.) Ground terms: no variables Primitive terms: f (n1 , ...., nk ) with ni standard names
LKB/SS-2014
6 / 18
Well-Formed Formulas in L
Atoms: P(t1 , ..., tk ) ti terms, P predicate symbol ground atom: no variables primitive atom: P(n1 , ...nk )
c G. Lakemeyer
LKB/SS-2014
7 / 18
Domain of quantication
Assumption: domain isomorphic to the set of names names refer uniquely names cover domain
c G. Lakemeyer
Not true for constants So can assume a single countably innite domain Two consequences:
1
in order to determine the denotation of a term it sufces to nd its co-referring standard name. can understand quantication substitutionally (e.g. to satisfy xP(x) it sufces to nd a name n such that P(n) is satised.)
LKB/SS-2014
8 / 18
Will determine co-referring name for every ground term truth value for every sentence
c G. Lakemeyer
Need only assign to primitives w[f (n1 , ...nk )] Names w[P(n1 , ...nk )] {1,0} Example: to determine f (g(n), m) for standard names n and m get w[g(n)] = n (the co-referring name n of g(n)) then w[f (n , m)].
LKB/SS-2014
9 / 18
|t|w
c G. Lakemeyer
|n|w = n |f (t1 , ...tk )|w = w[f (n1 , ...nk )], where ni = |ti |w
LKB/SS-2014
10 / 18
Sentence Satisfaction:
w |=
w |= P(t1 ...tk ) iff w[P(n1 ...nk )] = 1 w |= (t1 = t2 ) iff |t1 |w = |t2 |w w |= iff w |= w |= ( ) iff w |= and w |=
x w |= x iff w |= n for every name n.
where ni = |ti |w
c G. Lakemeyer
LKB/SS-2014
11 / 18
A sentence is valid (write |= ) iff w |= , for every w (iff {} is not satisable) implies (write |= ) iff {} is unsatisable These determine the logic L. The main focus is on valid sentences and nite .
LKB/SS-2014
12 / 18
Properties of Logic L
Except for =, names: |= iff is a theorem of ordinary rst-order logic Names: |= #1 = #3 Equality |= x1 y(y = x1 ) |= x1 x2 y(y = x1 ) (y = x2 ) |= x1 x2 x3 y(y = x1 ) (y = x2 ) (y = x3 ) etc.
c G. Lakemeyer
LKB/SS-2014
13 / 18
Therefore: innite domain. Use typed quantication for (nite) subsets of the domain. instead of
c G. Lakemeyer
x(P(x) ) Non-compactness
xP(x) (x = #1 x = #2 . . . x = #27)
4
There is a unsat such that every nite subset is sat: = {xP(x), P(#1), P(#2), P(#3), . . .} Because can name every domain element.
LKB/SS-2014
14 / 18
An Axiom System
Why do we care? another view of valid sentences What is not a reason (for us) to look at axioms?
c G. Lakemeyer
procedural interpretation (proof theory for generating the valid sentences) Analogy: care about language use grammar as another description separate from building recognizers
LKB/SS-2014
15 / 18
Universal generalization
Innite version:
x x From # 1 , #2 , . . . infer x x But: if I can prove n and there is nothing special about n,
then conclude x
c G. Lakemeyer
Not quite right: #1 = #2 is valid, but dont want x(x = #2) as theorem
LKB/SS-2014
16 / 18
Universal generalization
Innite version:
x x From # 1 , #2 , . . . infer x x But: if I can prove n and there is nothing special about n,
then conclude x
c G. Lakemeyer
Not quite right: #1 = #2 is valid, but dont want x(x = #2) as theorem Correct version:
x x From n , . . . , n , infer x 1 k
LKB/SS-2014
16 / 18
Proof Theory
Axioms A1. ( ) A2. ( ( )) (( ) ( ))
c G. Lakemeyer
LKB/SS-2014
17 / 18
Proof Theory
Axioms A1. ( ) A2. ( ( )) (( ) ( ))
c G. Lakemeyer
A3. ( ) (( ) ) A4. x( ) (x x ) A5. x tx A6. (n = n) (n = m), for distinct n and m Rules of Inference MP. From , ( ), infer .
x x UG. From n , . . . , n , infer x if ni range over names in and one not in . 1 k
LKB/SS-2014
17 / 18
Proof Theory
Axioms A1. ( ) A2. ( ( )) (( ) ( ))
c G. Lakemeyer
A3. ( ) (( ) ) A4. x( ) (x x ) A5. x tx A6. (n = n) (n = m), for distinct n and m Rules of Inference MP. From , ( ), infer .
x x UG. From n , . . . , n , infer x if ni range over names in and one not in . 1 k
Theorem: |= iff
LKB/SS-2014
17 / 18
Example Proof
8. yx(x = y f (x) = f (y)) 7,UG (Applications of Axioms A1A3 and Rule MP are not shown for simpicity.)
LKB/SS-2014
18 / 18