COMP3308/3608 Introduction to Artificial Intelligenece semester 1, 2022

Question 1. (Type 2 – problem solving/calculation)

In the tree below the step costs are shown along the edges and the h values are shown next to each node. The goal nodes are double-circled: F and D.

Write down the order in which nodes are expanded using:

a)  Breadth-first search

b)  Depth-first search

c)  Uniform cost search

d)  Iterative deepening search

e)   Greedy search

f)   A*

In case of ties, expand the nodes in alphabetical order.

Question 2. (Type 1 – short answers)

Answer briefly and concisely:

a)  A* uses admissible heuristics. What happens if we use a non-admissible one? Is it still useful to use A* with a non-admissible heuristic?

b)  What is the advantage of choosing a dominant heuristic in A* search?

c)  What is the main advantage of hill climbing search over A* search?

Question 3. (Type 2 – problem solving/calculation)

Consider the following game in which the evaluation function values for player MAX are shown at the leaf nodes. MAX is the maximizing player and MIN is the minimizing player. The first player is MAX.

a)  What will be the backed-up value of the root node computed by the minimax algorithm?

b)  Which move should MAX choose based on the minimax algorithm – to node B, C or D?

c)  Assume that we now use the alpha-beta algorithm. List all branches that will be pruned, e.g. AB etc. Assume that the children are visited left-to-right (as usual).

Question 4. (Type 1 – short answers)

Answer briefly and concisely:

a)  The 1R algorithm generates a set of rules. What do these rules test?

b) Gain ratio is a modification of Gain used in decision trees. What is its advantage?

c)  Propose two strategies for dealing with missing attribute values in learning algorithms.

d)  Why do we need to normalize the attribute values in the k-nearest-neighbor algorithm?

e)  What is the main limitation of the perceptrons?

f)   Describe an early stopping method used in the backpropagation algorithm to prevent overfitting.

g)  The  problem  of finding  a  decision  boundary  in  support  vector  machine  can  be formulated  as  an  optimisation problem  using  Lagrange  multipliers.  What  are  we maximizing?

h)  In linear support vector machines, we use dot products both during training and

during classification of a new example. What vectors are these products of? During training:

During classification of a new example:

Question 5. (Type 2 – problem solving/calculation)

Consider the task of learning to  classify mushrooms  as safe or poisonous based  on the following four features: stem = {short, long}, bell = {rounded, flat}, texture = {plain, spots,

bumpy, ruffles} and number = {single, multiple}.

The training data consists of the following 10 examples:

Safe:

Poisonous:

These examples are also shown in the table below:

a)  Use Naïve Bayes to predict the class of the  following new example: stem=long,    bell=flat,    texture=spots,    number=single.    Show    your calculations.

a)  How would 3-Nearest Neighbor using the Hamming distance classify the same example as above? Explain your answer. (Hint: The Hamming distance is the number of different feature values).