Uncertainty

Chapter 5

Artificial Intelligence
 This Lecture
• Probability
• Syntax
• Semantics
• Inference rules
• How to build a Naïve Bayes classifier
Fuzzy Logic in Real World
 Probability
Probabilistic assertions summarize effects of
Ignorance: lack of relevant facts, initial conditions, etc.
Laziness: failure to enumerate exceptions, qualifications, etc.

Subjective or Bayesian probability:

Probabilities relate propositions to one's own state of knowledge
e.g., P(A90 succeeds | no reported accidents) = 0.97

These are NOT assertions about the world, but represent belief about the
whether the assertion is true.

Probabilities of propositions change with new evidence: e.g.,

P(A90 | no reported accidents, 5 a.m.) = 0.99

(Analogous to logical entailment status; I.e., does KB |=  )

Making decisions under uncertainty
• Suppose I believe the following:
P(A 30 gets me there on time | ...) = 0.05
P(A 60 gets me there on time | ...) = 0.70
P(A 100 gets me there on time | ...) = 0.95
P(A 1440 gets me there on time | ...) = 0.9999

• Which action to choose?

• Depends on my preferences for missing flight vs.

airport cuisine, etc.

• Utility theory is used to represent and infer

preferences

• Decision theory = utility theory + probability theory

 Unconditional Probability
• Let A be a proposition, P(A) denotes the unconditional
probability that A is true.

• Example: if Male denotes the proposition that a

particular person is male, then P(Male)=0.5 mean that
without any other information, the probability of that
person being male is 0.5 (a 50% chance).

• Alternatively, if a population is sampled, then 50% of the

people will be male.

• Of course, with additional information (e.g. that the person

is a CS151 student), the “posterior probability” will likely
be different.
 Axioms of probability
For any propositions A, B
1. 0  P(A)  1
2. P(True) = 1 and
P(False) = 0
3. P(A  B) = P(A) +
P(B) - P(A  B)

de Finetti (1931): an agent who bets according to

probabilities that violate these axioms can be forced to
bet so as to lose money regardless of outcome.
 Syntax
Similar to PROPOSITIONAL LOGIC: possible worlds defined by
assignment of values to random variables.

Note: To make things confusing variables have first letter upper-case, and
symbol values are lower-case

Propositional or Boolean random variables

e.g., Cavity (do I have a cavity?)
Include propositional logic expressions
e.g., Burglary  Earthquake

Multivalued random variables

e.g., Weather is one of <sunny,rain,cloudy,snow>
Values must be exhaustive and mutually exclusive

A proposition is constructed by assignment of a value:

e.g., Weather = sunny; also Cavity = true for clarity
 Priors, Distribution, Joint
Prior or unconditional probabilities of propositions
e.g., P(Cavity) =P(Cavity=TRUE)= 0.1
P(Weather=sunny) = 0.72
correspond to belief prior to arrival of any (new) evidence

Probability distribution gives probabilities of all possible values

of the random variable.
Weather is one of <sunny,rain,cloudy,snow>
P(Weather) = <0.72, 0.1, 0.08, 0.1>
(normalized, i.e., sums to 1)
 Joint probability distribution
Joint probability distribution for a set of variables gives values for each
possible assignment to all the variables

P(Toothache, Cavity) is a 2 by 2 matrix.

Toothache=true Toothache = false
Cavity=true 0.04 0.06
Cavity=false 0.01 0.89

NOTE: Elements in table sum to 1 3 independent

numbers.
 The importance of
independence
• P(Weather,Cavity) is a 4 by 2 matrix of
values:
Weather = sunny rain cloudy snow

Cavity=True .072 .01 .008 .01

Cavity=False .648 .09 .072 .09

• But these are independent (usually):

– Recall that if X and Y are independent then:
– P(X=x,Y=y) = P(X=x)*P(Y=y)

Weather is one of <sunny,rain,cloudy,snow>

P(Weather) = <0.72, 0.1, 0.08, 0.1> and P(Cavity) = .1
 The importance of
independence
If we compute the marginals (sums of rows or columns), we see:

Weather = sunny rain cloudy snow marginals

Cavity=True .072 .01 .008 .01 .1

Cavity=False .648 .09 .072 .09 .9
marginals .72 .1 .08 .1

Each entry is just the product of the marginals (because they’re independent)!
So, instead of using 7 numbers, we can represent these by just recording 4
numbers, 3 from Weather and 1 from Cavity:
Weather is one of <sunny,rain,cloudy,snow>
P(Weather) = <0.72, 0.1, 0.08, (1- Sum of the others)>
P(Cavity) = <.1, (1-.1)>
This can be much more efficient (more later…)
 Conditional Probabilities
• Conditional or posterior probabilities
e.g., P(Cavity | Toothache) = 0.8
What is the probability of having a cavity given that the patient has a
toothache? (as another example: P(Male | CS151Student) = ??

• If we know more, e.g., dental probe catches, then we have

P(Cavity | Toothache, Catch) = .9 (say)

• If we know even more, e.g., Cavity is also given, then we have

P(Cavity | Toothache, Cavity) = 1
Note: the less specific belief remains valid after more evidence arrives,
but is not always useful.
• New evidence may be irrelevant, allowing simplification, e.g.,
P(Cavity | Toothache, PadresWin) = P(Cavity | Toothache) = 0.8
(again, this is because Cavities and Baseball are independent
 Conditional Probabilities
• Conditional or posterior probabilities
e.g., P(Cavity | Toothache) = 0.8
What is the probability of having a cavity given that the patient has a
toothache?
• Definition of conditional probability:
P(A|B) = P(A, B) / P(B) if P(B)  0

I.e., P(A|B) means,

our universe has
shrunk to the one
in which B is true.

P(A|B) is the ratio of

the red area to the
yellow area.
 Conditional Probabilities
• Conditional or posterior probabilities
e.g., P(Cavity | Toothache) = 0.8
What is the probability of having a cavity given that the
patient has a toothache?

• Definition of conditional probability:

P(A|B) = P(A, B) / P(B) if P(B)  0

Toothache=true Toothache = false

Cavity=true 0.04 0.06
Cavity=false 0.01 0.89

P(Cavity|Toothache) = .04/(.04+.01) = 0.8

 Conditional probability cont.
Definition of conditional probability:
P(A | B) = P(A, B) / P(B) if P(B) 
0

Product rule gives an alternative formulation:

P(A, B) = P(A | B)P(B) = P(B|A)P(A)

Example:
P(Male)= 0.5
P(CS151Student) = 0.0037
P(Male | CS151Student) = 0.9

What is the probability of being a male CS151 student?

P(Male, CS151Student)
= P(Male | CS151Student)P(CS151Student)
= 0.0037 * 0.9 = 0.0033
 Conditional probability cont.
A general version holds for whole distributions, e.g.,
P(Weather,Cavity) = P(Weather|Cavity) P(Cavity)
(View as a 4 X 2 set of equations, not matrix multiplication -
The book calls this a pointwise product)

The Chain rule is derived by successive application of

product rule:
P(X1,…Xn) = P(X1, …, Xn-1) P(Xn | X1, …, Xn-1)
= P(X1, …,Xn-2) P(Xn-1| X1 , …,Xn-2) P(Xn | X1,
…, Xn-1)
=…
n
=  P(Xi | X1, …, Xi-1)
i=1
 Bayes Rule
From product rule P(A B) = P(A|B)P(B) = P(B|A)P(A), we can
obtain Bayes' rule

P(B | A)P( A)
P( A | B) 
P(B)

Why is this useful???

For assessing diagnostic probability from causal probability:
e.g. Cavities as the cause of toothaches
Sleeping late as the cause for missing class
Meningitis as the cause of a stiff neck

P(Effect |
P(Cause | Effect) 
Cause)P(Cause)
P(Effect)
 Bayes Rule: Example
P(Effect| Cause)P(Cause)
P(Cause | Effect) 
P ( Effect )

Let M be meningitis, S be stiff neck

P(M)=0.0001
P(S)= 0.1
P(S | M) = 0.8

P(S |
P(M | S ) 
M )P(M ) 0.8

0.0001
 0.0008
 Bayes Rule: Another Example
Let A be AIDS, T be positive test result for AIDS
P(T|A)=0.99 “true positive” rate of test (aka a “hit”)
P(T|~A)= 0.005 “false positive” rate of test (aka a “false alarm”)
P(~T | A) = 1-0.99=.01 “false negative” or “miss”
P(~T|~A) = 0.995 “correct rejection”

Seems like a pretty good test, right?

 Bayes Rule: Another Example
Let A be AIDS, T be positive test result for AIDS
P(T|A)=0.99 “true positive” rate of test (aka a “hit”)
P(T|~A)= 0.005 “false positive” rate of test (aka a “false alarm”)
P(~T | A) = 1-0.99=.01 “false negative” or “miss”
P(~T|~A) = 0.995 “correct rejection”
Now, suppose you are from a low-risk group, so P(A) = .001, and you get a
positive test result. Should you worry?
Let’s compute it using Bayes’ rule: P(A|T) = P(T|A)*P(A)/P(T)
Hmmm. Wait a minute: How do we know P(T)?
Recall there are two possibilities: Either you have AIDS or you don’t.
Therefore: P(A|T) + P(~A|T) = 1.
P(~A|T) = P(T|~A)*P(~A)/P(T)
So, P(A|T) + P(~A|T) = P(T|A)*P(A)/P(T) + P(T|~A)*P(~A)/P(T) = 1
 Bayes Rule: Another Example
P(A|T) = P(T|A)*P(A)/P(T)
Hmmm. Wait a minute: How do we know P(T)?
Recall there are two possibilities: Either you have AIDS or you don’t.
Therefore: P(A|T) + P(~A|T) = 1.
Now, by Bayes’ rule:
P(~A|T) = P(T|~A)*P(~A)/P(T)
So,
P(A|T) + P(~A|T) = P(T|A)*P(A)/P(T) + P(T|~A)*P(~A)/P(T) = 1
 P(T|A)*P(A) + P(T|~A)*P(~A) = P(T)
 The denominator is always the sum of the numerators if they are exhaustive
of the possibilities - a normalizing constant to make the probabilities sum to 1!
 We can compute the numerators and then normalize afterwards!
 Bayes Rule: Back to the Example
Let A be AIDS, T be positive test result for AIDS
P(T|A)=0.99 “true positive” rate of test (aka a “hit”)
P(T|~A)= 0.005 “false positive” rate of test (aka a “false alarm”)
P(~T | A) = 1-0.99=.01 “false negative” or “miss”
P(~T|~A) = 0.995 “correct rejection”
Now, suppose you are from a low-risk group, so P(A) = .001, and you get a
positive test result. Should you worry?

P(A|T) P(T|A)*P(A) = 0.99*.001 = .00099

P(~A|T)  P(T|~A)*P(~A) = 0.005*.999 = 0.004995

P(A|T) = .00099/(.00099+.004995) = 0.165
Maybe you should worry, or maybe you should get
re-tested….
 Bayes Rule: Back to the Example II
Let A be AIDS, T be positive test result for AIDS
P(T|A)=0.99 “true positive” rate of test (aka a “hit”)
P(T|~A)= 0.005 “false positive” rate of test (aka a “false alarm”)
P(~T | A) = 1-0.99=.01 “false negative” or “miss”
P(~T|~A) = 0.995 “correct rejection”
Now, suppose you are from a high-risk group, so P(A) = .01, and you get
a positive test result. Should you worry?

P(A|T) P(T|A)*P(A) = 0.99*.01 = .0099

P(~A|T)  P(T|~A)*P(~A) = 0.005*.99 = 0.00495

P(A|T) = .0099/(.0099+.00495) = 0.667
=> You should worry!
 Bayes Rule: Another Example
Let A be AIDS, T be positive test result for AIDS
P(T|A)=0.99 “true positive” rate of test (aka a “hit”)
P(T|~A)= 0.005 “false positive” rate of test (aka a “false alarm”)
P(~T | A) = 1-0.99=.01 “false negative” or “miss”
P(~T|~A) = 0.995 “correct rejection”
Now, suppose you are from a low-risk group, so P(A) = .001, and you get a positive
test result. Should you worry?
Let’s compute it using Bayes’ rule: P(A|T) = P(T|A)*P(A)/P(T)
Hmmm. Wait a minute: How do we know P(T)?
Recall there are two possibilities: Either you have AIDS or
you don’t.
Therefore: P(A|T) + P(~A|T) = 1.
P(~A|T) = P(T|~A)*P(~A)/P(T)
So, P(A|T) + P(~A|T) = P(T|
A)*P(A)/P(T) +
P(T|~A)*P(~A)/P(T) = 1
 A generalized Bayes Rule

• More general version conditionalized on some

background evidence E

P(B | A, E)P( A | E)
P( A | B, E) 
P(B | E)
 Full joint distributions
A complete probability model specifies every entry in the
joint distribution for all the variablesX = X1, ..., Xn
i.e., a probability for each possible world X1= x1, ..., Xn=
xn

E.g., suppose Toothache and Cavity are the random

variables: Toothache=true Toothache = false
Cavity=true 0.04 0.06
Cavity=false 0.01 0.89

Possible worlds are mutually exclusive  P(w1  w2) = 0

Possible worlds are exhaustive  w1  …  wn is True
hence i P(wi) = 1
 Using the full joint distribution
Toothache=true Toothache = false
Cavity=true 0.04 0.06
Cavity=false 0.01 0.89

What is the unconditional probability of having a Cavity?

P(Cavity) = P(Cavity ^ Toothache) + P(Cavity^ ~Toothache)

= 0.04 + 0.06 = 0.1

What is the probability of having either a cavity or a Toothache?

P(Cavity  Toothache)
= P(Cavity,Toothache) + P(Cavity, ~Toothache) + P(~Cavity,Toothache)
= 0.04 + 0.06 + 0.01 = 0.11
 Using the full joint distribution
Toothache=true Toothache = false
Cavity=true 0.04 0.06
Cavity=false 0.01 0.89

What is the probability of having a cavity given that you

already have a toothache?
0.04
P(Cavity  Toothache) 
P(Cavity | Toothache)  P(Toothache)  0.04  0.8
0.01
 Normalization
Suppose we wish to compute a posterior distribution over random variable A
given B=b, and suppose A has possible values a1... am

We can apply Bayes' rule for each value of A:

P(A=a1|B=b) = P(B=b|
A=a1)P(A=a1)/P(B=b)
...
P(A=am | B=b) = P(B=b|
A=am)P(A=am)/P(B=b)

Adding these up, and noting that i P(A=ai | B=b) = 1:

P(B=b) = i P(B=b|
A=ai)P(A=ai)

This is the normalization factor denoted  =

1/P(B=b): P(A | B=b) =  P(B=b | A)P(A)

Typically compute an unnormalized distribution, normalize at end

e.g., suppose P(B=b | A)P(A) = <0.4,0.2,0.2>
then P(A|B=b) =  <0.4,0.2,0.2>
 Marginalization

Given a conditional distribution P(X | Z), we can create

the unconditional distribution P(X) by marginalization:

P(X) = z P(X | Z=z) P(Z=z) = z P(X, Z=z)

In general, given a joint distribution over a set of
variables, the distribution over any subset (called a
marginal distribution for historical reasons) can be
calculated by summing out the other variables.
 Conditioning
Introducing a variable as an extra condition:

P(X|Y) = z P(X | Y, Z=z) P(Z=z | Y)

Why is this useful??

Intuition: it is often easier to assess each specific circumstance, e.g.,
P(RunOver | Cross)
= P(RunOver | Cross, Light=green)P(Light=green |
Cross)
+ P(RunOver | Cross, Light=yellow)P(Light=yellow | Cross)
+ P(RunOver | Cross, Light=red)P(Light=red | Cross)
 Absolute Independence
• Two random variables A and B are (absolutely) independent iff
P(A, B) = P(A)P(B)

• Using product rule for A & B independent, we can show:

P(A, B) = P(A | B)P(B) = P(A)P(B)
Therefore P(A | B) = P(A)

• If n Boolean variables are independent, the full joint is:

P(X1, …, Xn) = i P(Xi)
Full joint is generally specified by 2n - 1 numbers, but when independent only n
numbers are needed.

• Absolute independence is a very strong requirement, seldom met!!

 Conditional Independence
• Some evidence may be irrelevant, allowing simplification, e.g.,

P(Cavity | Toothache, PadresWin) = P(Cavity | Toothache) = 0.8

• This property is known as Conditional Independence and can be expressed as:

P(X | Y, Z) = P(X | Z)
which says that X and Y independent given Z.

• If I have a cavity, the probability that thedentist’s probe catches in it doesn't

depend on whether I have a toothache:
1. P(Catch | Toothache, Cavity) = P(Catch | Cavity)
i.e., Catch is conditionally independent of Toothache given Cavity
This doesn’t say anything about P(Catch|Toothache)
The same independence holds if I haven't got a cavity:
2. P(Catch | Toothache, ~Cavity) = P(Catch | ~Cavity)
 Conditional independence contd.
Equivalent statements to
P(Catch | Toothache, Cavity) = P(Catch | Cavity) (*)

a. P(Toothache | Catch, Cavity) = P(Toothache | Cavity)

P(Toothache | Catch, Cavity)

= P(Catch | Toothache, Cavity) P(Toothache |
Cavity) / P(Catch | Cavity)
= P(Catch | Cavity) P(Toothache | Cavity) /
P(Catch | Cavity)

(from *)
= P(Toothache|Cavity)
P(Toothache,Catch | Cavity)
= P(Toothache | Catch,Cavity)
b. P(Toothache,Catch|Cavity) P(Catch | Cavity) (product rule)
= P(Toothache|
= P(Toothache | Cavity) P(Catch | Cavity)
Cavity)P(Catch|Cavity) (from 1a)
 Using Conditional Independence
Full joint distribution can now be written as
P(Toothache,Catch,Cavity)
= (Toothache,Catch | Cavity) P(Cavity)
= P(Toothache | Cavity)P(Catch | Cavity)P(Cavity) (from 1.b)

Specified by: 2 + 2 + 1 = 5 independent numbers

Compared to 7 for general joint
or 3 for unconditionally indpendent.