Scribd
Upload
23 views

Exam2021-2022 (Jan C)

This document contains a practice exam for a machine learning principles course. It consists of 9 multiple choice and short answer questions covering topics like decision trees, Bayesian networks, reinforcement learning, and more. The questions test understanding of concepts like model selection, conditional independence, off-policy learning, and estimating probabilities from data.

Uploaded by

Danilo Toapanta
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
23 views

Exam2021-2022 (Jan C)

This document contains a practice exam for a machine learning principles course. It consists of 9 multiple choice and short answer questions covering topics like decision trees, Bayesian networks, reinforcement learning, and more. The questions test understanding of concepts like model selection, conditional independence, off-policy learning, and estimating probabilities from data.

Uploaded by

Danilo Toapanta
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

H0E96a Principles of Machine Learning – Exam

2022-01-13, 17:00–20:00∗

Question 1 Given a hypothesis space L of binary classifiers. Given that the space contains
a finite amount n of classifiers, prove that V C(L)  log2 (n)

Question 2 The first image is a binary dataset. The other images1 show decision surfaces
(coloured by certainty) for different classifiers trained on these data. Connect the classifiers:
Linear SVM Naive Bayes k-NN Random forest Decision tree

Question 3 Given the data in the table2 below, perform one step of decision tree learning,
and answer the following questions. Treat missing values as seen in the lectures.
1. What is the dataset’s class entropy?
2. What is the information gain of the chosen test?
3. What class would be predicted for an input (pp,yes,?)?

idx. age astigmatic dioptre lens


1 young yes normal soft
2 young no reduced none
3 young no reduced none
4 ? yes normal none
5 pre-presbyopic yes ? soft
6 pre-presbyopic no reduced none
7 ? yes ? none
8 presbyopic yes normal soft
9 presbyopic ? normal soft
10 presbyopic yes reduced none
∗
Recalled and transcribed to LATEX by Thomas Bauwens.
1
Transcription note: Beware that on the actual exam, these were in grayscale and downsampled.
2
Transcription note: This table was reconstructed based on total recall of the attribute proportions. The
rows and labels could differ from those on the exam, but calculating the first two answers should be identical.

1
Question 4 Given the following pair of clauses.

p(X) q(X, Y ), q(Y, X)


p(X) q(X, Y ), q(Y, Z), q(Z, X)

1. How many literals does their `gg have?

2. What clause do we get after reducing the result under theta-subsumption?

Question 5 We have a binary dataset of 1600 instances with a class skew of 0.0625. On
this dataset, classifier has a positive precision of 0.8 and a TPR of 0.4.

1. Calculate the balanced accuracy of the classifier.3

2. Calculate the FPR of the classifier.

Question 6 We trained two classifiers on the same binary dataset. Their decision bound-
aries are shown.4

1. Which model will have better accuracy on the training data? (2-3 sentences)

2. Which model will have better accuracy if we use a bigger set of testing data? (2-3
sentences)

3
Transcription note: Indeed, “balanced accuracy” was never seen in class, so try to think about what it
could be and why this dataset lends itself to balancing. We don’t make the rules :/.
4
Transcription note: The actual figure doesn’t seem to be on Google. It was basically two images of the
same twin point cloud (of about 1000 points altogether), where one image had a decision boundary with
long vertical and horizontal stretches, and the other had a decision boundary that also ran horizontally or
vertically, but more fine-grained, sometimes surrounding a single point.

2
Question 7 Execute the PC algorithm for a Bayesian network of variables A, B, C, D using
the below graphical model as an oracle for (un)conditional independences.
1. Give all the sets Sxy produced by the algorithm.
2. Give the final model.

C F D

Question 8 hAngelika’s question of death based on the 231-slide lecture.i You are given
a table of 20 artworks, along with their artist, and whether the artworks are expensive or
cheap. You are also given which of those artists is famous.5 You are also given some prior
percentages: of 320 artworks and 100 artists, we know P (fame) = 2%, P (exp | fame) = 40%,
P (exp | not fame) = 10%. Given that expensiveness follows from fame,
1. Draw a par-factor graph over all artworks and all artists.
2. Estimate the probabilities for famous, expensive given famous and expensive given
non-famous, using data from both the new table as well as the prior data.
3. Draw a par-factor graph for the situation where expensive artworks imply fame.
4. Can you estimate the probabilities for expensive, famous given expensive and famous
given non-expensive, or are we missing prior percentages to get there?
Question 9 Answer the following questions with 1
5
A4 text at most per answer.
1. What is learnt in model-based reinforcement learning?
2. What is learnt in model-free reinforcement learning?
3. What is off-policy learning?
4. What is the difference between passive and active model-free reinforcement learning?
5
Transcription note: The exact data don’t matter, so the reader is asked to come up with their own.

You might also like