Exam Uncertainty Reasoning in Knowledge

systems
Name: ......................................
January 20, 2012

Instructions:

ˆ Only use the white space below the questions for your answer. This space
should suffice. For questions 6,7,8 and 9, you may use the back of the
page.

ˆ Fill out your name on EVERY page.

ˆ Hand in ALL the pages of the exam, even if you did not fill on the questions
on some pages.

ˆ The fractions after the questions indicate their relative importance for the
exam.

ˆ This is an open book. You may also use a calculator though that it is not
really necessary.

ˆ The marks are indicated after the question. This is for a total of 60 points.

1
 Name: .............................................

Questions:

1. Consider the following Bayesian network.

(a) List all the variables that are d-separated from F given E. (/2)

(b) List all the variables that are d-separated from F given E and K. (/2)

(c) Specify the Markov blanket of H. (/2)

2
 Name: .............................................

2. In your nuclear power station, there is an alarm that senses when a tem-
perature gauge (sensor) exceeds a given threshold. The gauge measures
the temperature of the core. Consider the Boolean variables A (alarm
sounds), FA (alarm is faulty), and FG (gauge is faulty), and the multival-
ued nodes G (gauge reading) and T (actual core temperature).

(a) Draw a Bayesian network for this domain, given that the gauge is
more likely to fail when the core temperature gets too high. (/5)

(b) Is your network singly connected ? Why or why not ?(/2)

3
 Name: .............................................

(c) Suppose that there are just two possible actual and measured tem-
peratures, normal and high; the probability that the gauge gives the
correct temperature is x when it is working, but y when it is faulty.
Give the conditional probability table associated with G.(/2)

(d) Suppose the alarm works correctly unless it is faulty, in which case it
never sounds. Give the conditional probability table associated with
A. (/2)

4
 Name: .............................................

3. In a Hidden Markov Model, one can say that the past is independent of
the future given the present.

(a) Draw a Hidden Markov Model to define the variables of interest and
state formally what is meant by the statement.(/3)

(b) Explain why the formula (the statement) is correct. (/2)

5
 Name: .............................................

4. Consider the Bayesian network p(A, B, C, D) = p(C).p(D).P (A|C, D).p(B|A, D).

Is the set of dependencies {C ⊥ D, A ⊥ D, C ⊥ B|A} an independence
map, a dependence map or a perfect map. Justify your answer. (/3)

5. Consider the Markov network p(A, B, C, D) = Z1 f (C, A).f (A, B, D). Is

the set of dependencies {C ⊥ D|A, C ⊥ B|A} an independence map, a
dependence map or a perfect map. Justify your answer. (/3)

6
 Name: .............................................

1
7. Condider the factor graph p(a, b, c, d) = Z f (a, b)f (a, c)f (b, c)f (c, d).

(a) Draw the factor graph. (/2)

(b) This graph is multiply connected, which precludes sum-product in-

ference to compute p(d). However, if we know the value of b, we
can compute p(d|b) using the sum-product algorithm. Draw the fac-
tor graph in which to perform sum-product inference and explain in
three lines why the sum-product does apply to this factor graph. (/2)

(c) List all the messages (in the notation of the book) sent by the sum-
product algorithm. (/4)

(d) Explain how to compute p(d|b) based on these messages. (/2)

8
 Name: .............................................

8. (EM on belief networks)

Two doctors have been independently using two different tests to diagnose
whether a patient has the flu or not. At some point, the two doctors meet
to figure out which of the tests has the higher reliability.
Some of the patients did not agree to be tested twice, and some patients
did not follow-up to confirm whether they actually had the flu or not.
Nevertheless, the doctors do not want to waste any medical record, even if
some of the information is missing. The data they collected is as follows:

patient index test 1 outcome test 2 outcome Really has the flu?
1 positive missing yes
2 positive yes yes
3 missing negative no
4 positive negative yes
5 missing positive missing

(a) Draw the belief network that you would use to model this problem.
(/1)
(b) Assuming that the prior probabilities p(outcome test 1=neg|patient
has the flu=no) = 0.8, p(outcome test 1=neg|patient has the flu=yes)
= 0.15,p(outcome test 2=neg|patient has the flu=no) = 0.9, p(outcome
test 2=neg|patient has the flu=yes) = 0.25. and of having the flu is
50%. Describe in detail the updates for E and M steps for patient 1
and for at least one CPT. (/2)
(c) What is the value of the parameters p(outcome test 1=neg|patient
has the flu=no) and p(outcome test 1=neg|patient has the flu=yes)
after one full iteration of the EM algorithm using the whole dataset
(no detailed description required) ? (/3)