CSC 1147 - Foundations of AI - 2026-27

Part 1: Multiple choice

1. If evil AIs ever consider the logic of wiping out humans, they will realise:
   1. This will be against their logical rules.
   2. Some humans will survive.
   3. This will be against their ethics.
   4. For the AIs, it would be a good idea.
   5. Any serious damage to human civilization right now will wipe out all AIs fast.
   6. Humans will fight them.
   7. Humans will die before AIs do.

2. Which of the following makes the most sense?
   1. Supervised Learning for character recognition. Supervised Learning for robotic soccer.
   2. Supervised Learning for character recognition. Reinforcement Learning for robotic soccer.
   3. Reinforcement Learning for character recognition. Supervised Learning for robotic soccer.
   4. Reinforcement Learning for character recognition. Reinforcement Learning for robotic soccer.

3. AI is related to human immortality because:
   1. AI will discover new cures.
   2. If you are a robot, you can still be destroyed.
   3. AI will make humans extinct.
   4. If you are physical, you can still be destroyed.
   5. If you have a backup of yourself, you can be effectively immortal.
   6. AI will replace us.
   7. AI is not related to human immortality.
   8. AI will extend human lifespans.

4. Define a perceptron that receives x and y as input and fires for these points:

   a x - b y ≤ -c

   1. weights: -a, b, threshold: c
   2. weights: a, b, threshold: c
   3. weights: -a, -b, threshold: -c
   4. weights: a, -b, threshold: c
   5. weights: -a, b, threshold: -c
   6. weights: a, -b, threshold: -c
   7. weights: a, b, threshold: -c
   8. weights: -a, -b, threshold: c

5. Consider the following heuristic:

   Initialise with 100 random solutions.
   Repeat forever
   {
       Pick the best 99.
       Pick one at random to duplicate so we keep size 100.
       Make small or no changes to the 100.
       Repeat.
   }

   What can we say about this algorithm:

   1. 99 is too big. This cannot work.
   2. It is guaranteed to settle on the global optimum.
   3. It will tend to settle on a local optimum too quickly.
   4. It never varies from the start solutions.
   5. 99 is too small.
   6. 99 is probably too big. But we get rid of bad ones eventually, so it might work.

6. In a neural network, we initialise the weights as:
   1. all = 0
   2. random small absolute values
   3. all = c, a negative quantity
   4. random large absolute values
   5. all = c, a positive quantity

7. The sigmoid function y = ς(ax + b) is linear for:
   1. b small absolute (positive or negative)
   2. b large absolute (positive or negative)
   3. never
   4. a small absolute (positive or negative)
   5. b=0
   6. a large absolute (positive or negative)
   7. always
   8. a=0

8. The input to the 7th output node of a neural network is:
   x₇ = w₁₇ y₁ + w₂₇ y₂ + w₃₇ y₃
   Then ∂x₇ / ∂w₂₇ is:
   1. w_jk
   2. w_jk
   3. w₂₇
   4. y_j
   5. y₂
   6. y_j

9. In a multi-layer neural network, hidden node j can learn to ignore input I_i by:
   1. Moving w_jj towards 0.
   2. Moving w_ij towards 1.
   3. It cannot learn to ignore input I_i.
   4. Moving w_jj towards 1.
   5. Moving w_ii towards 0.
   6. Moving w_ii towards 1.
   7. Moving w_ij towards 0.

10. Imagine the following in Best-first search. The heuristic evaluation function tells us a state is 4 steps away from the goal. We examine the children of this state and the heuristic evaluation function tells us all of them are 5 steps away from the goal. How can this happen?
    1. We are in an infinite loop.
    2. This always happens.
    3. The goal has been wrongly defined.
    4. We are going the wrong way.
    5. This cannot happen.
    6. The heuristic evaluation function is inaccurate.

11. The set of inequalities that define a perceptron to implement NOT are:
    1. 0 ≤ t < w
    2. w ≤ t < 0
    3. 0 ≤ w < t
    4. 0 < t ≤ w
    5. t < w ≤ 0
    6. 0 < w ≤ t
    7. w < t ≤ 0
    8. t ≤ w < 0

12. In neural networks, let be the change in error as we change a weight. Let c be a negative constant. Then a rule to reduce error would be:
    1. W := c
    2. W := W - c
    3. W := W + c
    4. W := - c

13. Which is the most correct statement about the creators of an AI program and the users of the AI program?
    1. It is usually easy for both users and creators to understand why AI programs generate the output they do.
    2. It is usually easy for creators to understand why AI programs generate the output they do.
    3. It is often hard for creators to understand why AI programs generate the output they do.
    4. It is usually easy for users to understand why AI programs generate the output they do.
    5. It is often hard for both users and creators to understand why AI programs generate the output they do.
    6. It is often hard for users to understand why AI programs generate the output they do.

14. Consider the Travelling salesman problem with n cities. We must start in city A and then visit all the other cities. We do not consider paths that visit the same city twice. The number of possible paths is:
    1. n!
    2. 2ⁿ
    3. 2^(n-1)
    4. nⁿ
    5. (n-1)!
    6. n²

15. In the output layer of a neural network, if x_k = w_jk y_j
    then is:
    1. y_k
    2. w_jk
    3. w_jk y_j
    4. y_j
    5. w_jk y_j
    6. y_j
    7. w_jk

16. The sigmoid function y = ς(ax + b) is the linear function y = 3/4 x when:
    1. b = 0, a = something else (other than 3/4)
    2. a = 0, b = 3/4
    3. It can only be made linear y = 1/2.
    4. b = 0, a = 3/4
    5. a = 0, b = something else (other than 3/4)
    6. It cannot be made linear.
    7. It cannot be made linear y = 3/4 x.

17. Classic AI was very interested in chess because:
    1. Chess is symbolic AI.
    2. Chess requires machine vision.
    3. Chess requires robotic grippers.
    4. Chess is easily solved by computers.
    5. Chess is provably hard to solve optimally for a computer.
    6. People of high intelligence play chess.
    7. Chess is well suited to exhaustive search.

18. An upper bound on the game-tree complexity of Xs and Os is:
    1. 9⁹
    2. 3³
    3. 3⁹
    4. 9!
    5. 9³
    6. 3!

19. We are trying to maximise the following function from exemplars. (Meaning try x values and see y values.) The function f(x) is as follows:
    f(1.01040320402) = 80.
    f(5.46377356366) = 100.
    f(x) = 0 for all other x.
    What happens?
    1. It learns f(x) = 0 for some x.
    2. It may think the maximum is at x = 1.01040320402.
    3. It learns the maximum is y = 100.
    4. It probably learns f(x) = 0 for all x.
    5. It tries random x values until finding the maximum.
    6. It learns the maximum is at x = 5.46377356366.

20. We have a neural network where inputs are finite (positive or negative). If a node has a (negative) infinite threshold, then:
    1. None of the above.
    2. If weights are positive, you have constant output 1.
    3. If weights are positive, you have constant output 0.
    4. If weights are finite, you have constant output 1.
    5. If weights are negative, you have constant output 0.
    6. If weights are finite, you have constant output 0.
    7. If weights are negative, you have constant output 1.

21. A perceptron has 2 binary inputs, 2 weights w₁ and w₂, a threshold t, and 1 binary output. For the perceptron to implement XOR, what is the set of inequalities that must be satisfied?
    1. w₁ < t,   w₂ < t,   0 < t,   w₁+w₂ ≥ t
    2. w₁ ≥ t,   w₂ ≥ t,   0 < t,   w₁+w₂ ≥ t
    3. w₁ ≥ t,   w₂ ≥ t,   0 < t,   w₁+w₂ < t
    4. w₁ < t,   w₂ < t,   0 < t,   w₁+w₂ < t

22. A perceptron has inputs I_i with weights w_i, and a threshold t. It produces binary output. The rule for output is:
    1. If ≤ t, then output = 0, else output = 1.
    2. If ≥ t, then output = 0, else output = 1.
    3. If > t, then output = 0, else output = 1.
    4. If < t, then output = 0, else output = 1.

23. Imagine we do Best-first search with a random heuristic evaluation function. (A heuristic evaluation function that generated random numbers.) One of the following is an interesting statement about it. The others are either false, or true but uninteresting.
    1. The heuristic is inaccurate.
    2. The heuristic would take time to calculate.
    3. It would be exhaustive search.
    4. The search would be randomised but not quite breadth-first or depth-first search. Random branches would be not searched because the top parent's random number was bad.
    5. The search would be randomised but basically breadth-first search.
    6. The search would be randomised but basically depth-first search.
    7. Search could not proceed at all.

24. The idea of "informed search" is that we choose the next state that is most likely to lead to the goal. But if we can do that, then the problem is solved! Just keep choosing the best next state after that and you get to the goal ASAP. No search needed! No AI needed. Explain.
    1. Sometimes you have to go away from the goal.
    2. Exactly. Choose the state that is most likely to lead to the goal.
    3. There could be joint winners.
    4. No such reliable heuristic function will exist for any interesting problem.
    5. We have to write the heuristic first.
    6. Sometimes you should not choose the state that is most likely to lead to the goal.

25. The sigmoid function y = ς(x) has the property that:
    1. dy/dx = y
    2. dy/dx = 1-y
    3. y(1-y) = 1-y
    4. y = 1-y
    5. dy/dx = y(1-y)
    6. dy/dx = y/x

26. In Q-learning, why can't we systematically go through each action in each state: "For all (x,a), try a in x"?
    1. Because we need to try actions in other states.
    2. We cannot try all actions.
    3. We do this eventually.
    4. We don't know how to get to x.
    5. This would take too long.
    6. We do in fact do this.

27. We are learning to represent a function Q(x). The input x is a 10 dimensional vector. Each dimension is an integer from 1 to 100. One approach to representing this function would be to have a "lookup table" or array of entries, where we have one entry Q(x) for each input x. This array would be initially blank and we would fill it in by learning. The size of the array is:
    1. 1000
    2. 100
    3. 110
    4. 100¹⁰
    5. 10
    6. 1010
    7. 10¹⁰⁰

28. A perceptron has 2 inputs, taking values 0 or 1, and 1 output. Its 2 weights are 0 and 1. The output threshold is -0.5. This perceptron implements:
    1. Constant output 0.
    2. AND
    3. None of the above.
    4. NOT
    5. Constant output 1.
    6. OR
    7. XOR

29. Why on earth would we use a "heuristic" program that can deliver a sub-optimal result?
    1. We only use heuristics that deliver optimal results.
    2. The optimal result does not exist.
    3. Algorithm should return quickly.
    4. Sub-optimal is normally as good as optimal.
    5. In some problems, all practical solutions are sub-optimal.
    6. It is hard to think of heuristics.

30. One of these is not a good "Not Enough Stuff" argument:
    1. Not enough parallelisation.
    2. Not enough speed.
    3. Not enough training time.
    4. Not enough training data.
    5. Not enough diversity of sub-minds or sub-experts.
    6. Not enough weights.
    7. Not enough layers.
    8. Not enough serialisation.

31. In the output layer of a neural network, if x_k = w_jk y_j
    then is:
    1. w_jk
    2. y_j
    3. w_jk
    4. w_jk y_j
    5. y_j
    6. w_jk y_j
    7. y_k

32. Consider the function y = f(x) = ς(-x) where ς is the sigmoid function.
    f has its (a) highest slope value, and (b) steepest slope (highest absolute value) at:
    1. (a) y = 0 or 1, (b) y = 0 or 1
    2. (a) y = 1/2, (b) y = 0 or 1
    3. (a) y = 1/2, (b) y = 1/2
    4. (a) y = 0 or 1, (b) y = 1/2

33. One of these is known to be false:
    1. Machines have beaten world-class humans at tennis.
    2. In the future, immortal humans will travel to distant stars.
    3. In the future, humans will live forever on earth in artificial bodies, regularly backing up their brains in order to cheat death.
    4. Machines have beaten world-class humans at chess.
    5. In the future, intelligent machines will exterminate the human race.
    6. AI (the construction of intelligent machines) is impossible.
    7. AI is possible.

34. In Reinforcement Learning, the reward function is:

    if (good thing) r = 1
    else if (bad thing) r = -1
    else r = 0

    Then a Q-value of less than 0 is:

    1. Expected.
    2. Impossible.
    3. Bad.
    4. Good.
    5. Could be good, neutral or bad. It depends on the frequency of the good and bad things.
    6. Neutral.

35. In a multi-layer neural network, we have n hidden units with binary outputs, 1 output unit, weights w₁,..,w_n connecting them, and threshold t for the output unit. Which one of the following implements an output unit that fires if and only if all the hidden units fire.
    1. w₁=w₂=..=w_n=1,   t=0
    2. w₁=w₂=..=w_n=W, W > 0,   t=(n+1)W
    3. w₁=w₂=..=w_n=W, W > 0,   t=n
    4. w₁=w₂=..=w_n=W, W > 0,   t=(n-1)W
    5. w₁=w₂=..=w_n=0,   t=1
    6. w₁=w₂=..=w_n=W, W > 0,   t=nW
    7. w₁=w₂=..=w_n=W, W > 0,   t=W

36. In the following neural network, the threshold values are circled. The weights are listed along each link. The input units take values 0 or 1.
    
    If the -2 weight was changed to +2, and the output unit threshold to 1.5, what function would the network implement?
    1. constant output 1
    2. A function different to the above
    3. XOR
    4. NOT
    5. It wouldn't implement any function
    6. AND
    7. constant output 0
    8. OR

37. In state x, if we take action a, we sometimes get reward 0, and sometimes get reward 5. If we take action b, we always get reward 3. Under what circumstances is the expected reward for taking action a the same as the expected reward for taking action b?
    1. P_xa(0) = 0
    2. P_xa(0) = 1
    3. P_xa(5) = 1
    4. P_xa(5) = 0
    5. P_xa(5) = 3/5
    6. P_xa(0) = 3/5

38. In neural networks, using the notation on this course, we make large absolute changes to the weight w_jk when:
    1. y_j is small
    2. | y_k-O_k |   is small
    3. (1-y_k) is large
    4. y_k is large
    5. (1-y_j) is large
    6. y_k = 1/2
    7. y_k is small

39. In a Q-learning problem, there are n actions to choose from. We are using a "Soft max" probability distribution, based on Q-values, with a "temperature" parameter T, to choose what action to take.
    Then as   T -> infinity   the probability of choosing the action with the lowest Q-value goes to:
    1. 1/n
    2. 1
    3. n
    4. 0
    5. 1/T
    6. T
    7. infinity

40. We are trying to maximise the following function from exemplars. (Meaning try x values and see y values.) The function f(x) is as follows:
    f(a) = 0 for some value a.
    f(x) = 1 for all other x.
    What happens?
    1. It learns the maximum is at x = a.
    2. It learns the maximum is at y = a.
    3. It learns the maximum is at x = 0.
    4. It probably finds the maximum immediately.
    5. It tries random x values until finding the maximum.
    6. It takes a long time to find the maximum.
    7. It learns the maximum is at y = 0.

Part 2: Prepared question

CSC 1147

Answer sheet

Black out the square that matches the correct answer.