Transcript Chapter 9 Inference in first

Chapter 8
Inference in first-order logic
Inference in FOL,
removing the quantifiers
 i.e., converting KB to PL
 then use Propositional inference


which is easy to do
Inference rules for quantifiers

In KB, we have
A rule called Universal Instantiation (UI)
substituting a ground term for the variable
 SUBST( , ) denotes


the result of applying the substitution  to the
sentence 
v 
SUBST({v/g}, )

Here the examples are

{x / John}, {x / Richard}, {x / Father(John)}
Existential Instantiation:

For any sentence , variable v, and
constant k that does not appear elsewhere
in the KB:
v 
SUBST({v/k}, )
x Crown(x)  OnHead(x, John)
If C1 does not appear elsewhere in the KB
 then we can infer:
Crown(C1)  OnHead(C1, John)

Reduction to Propositional
Inference
Main idea

for existential quantifiers
find a ground term to replace the variable
 remove the quantifier
 add this new sentence to the KB


for universal quantifiers
find all possible ground terms to replace the variable
 add the set of new sentences to the KB

apply UI to the first sentence
from the vocabulary of the KB
 {x / John}, {x / Richard} – two objects
 we then obtain

view other facts as propositional variables

use inference to induce Evil(John)
Propositionalization
The previous techniques
applying this technique to
every quantified sentence in the KB
 we can obtain a KB consisting of
propositional sentences only
 however, this technique is very inefficient

in inference
 it generates other useless sentences

Unification and Lifting
Unification
a substitution θ such that applying on two
sentences will make them look the same
 e.g., θ = {x / John} is a unification
 applying on x King ( x)  Greedy( x)  Evil( x)
 it becomes

x King ( John)  Greedy( John)  Evil( John)

and we can conclude the implication

using King(John) and Greedy(John)
Generalized Modus Ponens (GMP)
The process capturing the previous steps
A generalization of the Modus Ponens

also called the lifted version of M.P.
For atomic sentences pi , pi', and q,
there is a substitution  such that
 SUBST(, pi)= SUBST(, pi'), for all i:

Generalized Modus Ponens
Unification
UNIFY, a routine which
takes two atomic sentences p and q
 return a substitution 

that would make p and q look the same
 it returns fail if no such substitution  exists


Formally, UNIFY(p, q)=
where SUBST(, p) = SUBST(, q)
  is called the unifier of the two sentences

Standardizing apart
UNIFY failed on the last sentence
in finding a unifier
 reason?
 two sentences use the same variable name



even they are having different meanings
so, assign them with different names
internally in the procedure of UNIFY
 = standardizing apart

MGU
Most Generalized Unifier
there may be many unifiers for two sentences
 which one is the best?

the one with less constraints
 e.g., UNIFY(Knows(John, x), Knows(y, z))
 one unifier = { y/John, x/John, z/John }
 another = {y/John, z/x} – the best


even if z and x are not yet found
Forward and backward chaining
Forward chaining
start with the sentences in KB
 generate new conclusions that



in turn allow more inferences to be made
usually used
when a new fact is added to the KB
 and we want to generate its consequences

First-order Definite clauses

the normal form of sentences for FC
can contain variables
 either atomic or
 an implication whose




antecedent is a conjunction of positive literals
consequent a single positive literal
Not every KB can be converted into a set of
definite clauses, but many can

why? the restriction on single-positive-literal
Example
We will prove that West is a Criminal
First step
translate these facts as first-order definite
clauses
 next figure shows the details

For forward chaining, we will have two
iterations
The above is the proof tree generated
No new inferences can be made at this
point using the current KB

such a KB is called a fixed point of the
inference process
Backward chaining
start with something we want to prove

(goal/query)
find implication sentences

that would allow to conclude
attempt to establish their premises in turn
normally used

when there is a goal to prove (query)
Backward-chaining algorithm
This is better to illustrate with a proof tree
One remark

backward chaining algorithm uses
composition of substitutions
 SUBST(COMPOSE(θ1, θ2), p)



= SUBST(θ2, SUBST(θ1, p))
it’s used because
different unification are found for different goals
 we have to combine them.

Resolution
Modus Ponens rule
can only allow us to derive atomic conclusions
 {A, A=>B}├ B

However, it is more natural
to allow us derive new implication
 {A => B, B => C} ├ A=>C, the transitivity
 a more powerful tool: resolution rule

CNF for FOL
Conjunctive Normal Form

a conjunction (AND) of clauses
each of them is a disjunction (OR) of literals
 the literals can contain variables


e.g.,
Conversion to CNF

6 steps
Skolemize
process of removing 
 i.e., translate x P(x) into P(A), A is a new
constant
 If we apply this rule in our sample, we have

which is completely wrong
 since A is a certain animal
 To overcome it, we use a function to represent any
animal, these functions = Skolem functions

Drop universal quantifiers

all variables now are assumed to be
universally quantified
Fortunately, all the above steps can be
automated
Resolution inference rule
Example proof
Resolution proves that KB |= by
proving KB  α unsatisfiable, i.e., empty clause
 First convert the sentences into CNF

next figure, empty clause
 so, we include the negated goal  Criminal(West)

Example proof
This example involves
skolemization
 non-definite clauses
 hence making inference more complex

Informal description
The following answers: Did Curiosity kill the cat?
First assume Curiosity didn’t kill the cat.
Add it into KB
Empty clause, so the assumption is false
Resolution strategies
Resolution is effective
but inefficient
 because it is like forward chaining
 the reasoning is randomly tried

There are four general guidelines in
applying resolution
Unit preference
When using resolution on two sentences
 one of the sentences must be a unit clause
 (P, Q, R, etc.)
The idea is to produce a shorter sentence:
 e.g., P  Q => R and P
 will produce Q => R
 hence reduce the complexity of the clauses

Set of support
Identifying a subset of sentences from the KB
 Every resolution combines a sentence

from the subset
 and another sentence from the KB


The resolvent (conclusion) of the resolution
is added to the subset
 and continue the resolution process


How to choose this set?

a common approach: the negated query



to prove the query, assume negative
and prove the contradiction
advantage: goal-directed
Input resolution

Every resolution combines

one of the input sentences (facts)




from the query
or the KB
with some other sentence
Next fig
For
each resolution,
at
least one of the sentences from the query or KB
Subsumption (inclusion, 包含)
 eliminates
all sentences
 that
are subsumed by (i.e., more specific
than) an existing sentence in the KB
 If
P(x) is in KB, x means all arguments
 then
we don’t need to store the specific
instances of P(x): P(A), P(B), P(C) …,
 Subsumption
helps keep the KB small