Homework 4 Health Behaviours

COMP-424: Artificial intelligence
Homework 4
General instructions.
• Unless otherwise mentioned, the only sources you should need to answer these questions are your
course notes, the textbook, and the links provided. Any other source used should be acknowledged with
proper referencing style in your submitted solution.
• Submit a single pdf document containing all your pages of your written solution on your McGill’s
myCourses account. You can scan-in hand-written pages. If necessary, learn how to combine many pdf
files into one.

.PLEASE CHECKOUT this website FOR MORE HOMEWORK SOLUTIONS
Question 1: Health Behaviours
Consider the following causal graphical model involving three Bernoulli random variables, which is a
simple model of health status and behaviours: H (health status), C (cautious behaviour), D (disease).
People’s health status influences whether they adopt cautious behaviour, and
their health status together with their behaviour influence their probability of
disease.
We collect an observational dataset of a population of 1000 people:
H C D #instances
0 0 0 10
0 0 1 13
0 1 0 44
0 1 1 28
1 0 0 689
1 0 1 124
1 1 0 84
1 1 1 8
a) Using maximum likelihood estimation, estimate the observed conditional probabilities of disease
given cautiousness: P(D=1|C=1) and P(D=1|C=0)
b) Suppose that we can intervene and perfectly persuade people to be cautious or not. Estimate
P(D=1|do(C=1)) and P(D=1|do(C=0)) using maximum likelihood estimation. What is the relative
risk reduction of adopting cautious behaviour? (RRR = 1 – (P(D=1|do(C=1)) / P(D=1|do(C=0)))
c) A talk show host on TV points to your results in part a) to say that there is no point in being
cautious. How do you rebut this argument?
H
C
D
Question 2: Fire Hazard
Consider the Bayes Net shown here, with all Bernoulli variables,
which involve building types and their risk of fire. Having a fire has
a utility of -1000 if the building was insured, but has a utility of -
50000 if the building was not insured (or the insurance claim is
denied). Having insurance when there is no fire has a utility of -50,
and not having insurance and no fire has a utility of 0.
Use the principle of Maximum Expected Utility and Value of
Information to answer the following questions. For parts a)-c) assume the insurance
company pays for 100% of the cases.
a) Given no information about whether a building is residential or whether it is
occupied, how much should the insurance company charge to insure the
building to break even, in utility points?
b) How much should the company charge for insurance if you know for certain that the building
is commercial (R=F) and occupied (O=T) to break even?
c) How much should they charge if the building is residential (R=T)?
d) A company is offering cheaper insurance but has a reputation of rejecting 25% of insurance
claims. How much should they charge for this insurance, to make it competitive with the
insurance offered by the more reliable company? (Hint: Set the cost of the new insurance to
have the same MEU as the other insurance.)
R (F) R (T)
0.25 0.75
O (F) O (T)
0.15 0.85
R O F (T)
F F 0.005
F T 0.02
T F 0.005
T T 0.01
Residential
(R)
Occupied
(O)
Fire
(F)

.PLEASE CHECKOUT  this websiteFOR MORE HOMEWORK SOLUTIONS
Question 3: Bandits
Consider the following 6-armed bandit problem. The initial value estimates of the arms are given by Q =
{1, 2, 2, 1, 0, 3}, and the actions are represented by A = {1, 2, 3, 4, 5, 6}. Suppose we observe that each
lever is played in turn: (from lever 1 to lever 6, and then start from lever 1 again):
𝐴𝑡 = ((𝑡 − 1)𝑚𝑜𝑑 6) + 1 (1)
We also observe that the rewards 𝑅𝑡
seem to fit the following function:
𝑅𝑡 = 2 cos [
𝜋
6
(𝑡 − 1)] (2)
So, the first two action-reward pairs are 𝐴1 = 1, 𝑅1 = 2, and 𝐴2 = 2, 𝑅2 = √3.
a) Show the estimated Q values from 𝑡=1 to 𝑡=12 of the trajectory using the average of the observed
rewards, where available. Do not consider the initial estimates as samples.
b) It turns out the player was following an 𝜖-greedy strategy, which just happened to coincide with
the scheme described above in (1) for the first 12 time steps. For each time step t from 1 to 12,
report whether it can be concluded with certainty that a random action was selected.
c) Suppose now we continue to visit the levers iteratively as in (1), and that the observed rewards
continue to fit the pattern established by (2). Is there a limiting expected reward 𝑄
∗
(𝑎) for each
action 𝑎 ∈ 𝐴 as 𝑡 approaches infinity? Justify your answer.