Lecture 3

Question 1

What does a search algorithm do (in generic terms)?

Answer 1

A search algo takes a search problem as input and returns a solution or indication of failure.

Question 2

In what kind of environment is the solution to any problem a fixed solution of actions?

Answer 2

In an environment that is fully observable, deterministic, and known.

Question 3

What is an admissible heuristic?

Answer 3

One that never overestimates the cost to reach the goal. Thereofore, it is optimistic.

Question 4

How do we decide which node from the frontier to expand next?

Answer 4

Using a search function, such as best-first, breadth-first, or depth-first search.

Question 5

What are the 4 ways to measure the performance of a search algorithm?

Answer 5

1. Completeness
2. Cost Optimality
3. Time Complexity
4. Space Complexity

Question 6

What is the difference between an informed and uninformed search?

Answer 6

Informed knows how far the node is from the goal at any given step of the algorithm, whereas uninformed never has any idea of its distance from the goal.

Question 7

How can you implement a BFS using Best-First Search?

Answer 7

f(n) = n.depth

Using the node’s depth as the evaluation function.

Question 8

How can you implement a DFS using Best-First Search?

Answer 8

f(n) = -n.depth

Using the negative of the node’s depth as the evaluation function.

Question 9

How can you implement a Dijkstra’s using Best-First Search?

Answer 9

f(n) = n.pathCost

Using the node’s path cost as the evaluation function.

Question 10

What could go wrong if you use greedy best-first search for path-finding?

Answer 10

It is prone to getting stuck in local optima or dead ends. Because the algorithm always prioritizes nodes that appear to be closest to the goal, it may not explore other potentially useful paths or nodes that are further away from the goal but could ultimately lead to a better solution.

Question 11

Is A* cost-optimal?

Answer 11

It depends on the heuristic function. If the heuristic function is admissible or consistent, A* is cost-optimal. Otherwise it may not be cost-optimal.

Question 12

What is the performance difference between running hill-climbing search (or simulated-annealing) for k times and beam search of size k?

Answer 12

hill-climbing search with 5 threads could have essentially the same result. (the 4 that are bad will stay until the end because there is no way to prune them)

in local beam-search, bad options are pruned from the beginning, and you’re taking the k best

Question 13

Considering the vanilla implementation of MINIMAX, discuss the disadvantages and possible solutions to fix them.

Answer 13

Memory and speed problems due to the algorithm’s requirement to expand a node on the tree all the way to the bottom of the tree.
- Alpha-Beta pruning can help speed up the algorithm by preventing some nodes from even being calculated.
- Depth-limit will solve the problem. We use heuristics to estimate the possible score for non-goal states.

Question 14

What is a competitive Environment?

Answer 14

An environment in which two or more agents have conflicting goals.

Question 15

Adversarial agents can be considered one of which 3 environment stances?

Answer 15

• an economy
• part of the environment (making it non-deterministic)
• explicitly modelled

Question 16

(Game Theory) Define transition model.

Answer 16

Defines the state resulting from taking action ‘a’ in state ‘s’

RESULT(s, a);

Question 17

(Game Theory) Define terminal test.

Answer 17

A test which is true when the game is over and false otherwise.

IN_TERMINAL(s);

Question 18

(Game Theory) What is a terminal state?

Answer 18

States where the game has ended.

Question 19

(Game Theory) What does a utility function define?

Answer 19

The final numeric value (utility score) of player ‘p’ when the game ends in terminal state ‘s’

UTILITY(s, p);

Question 20

(Game Theory) Which components make up the state space graph?

Answer 20

The initial state, ACTIONS function, and RESULTS function.

Question 21

(Game Theory) What is a state space graph?

Answer 21

A graph where the vertices are states, the edges are moves, and a state might be reached by multiple paths.

Question 22

(Game Theory) How do you determine which move to make using a state space graph?

Answer 22

By superimosing a game tree over part of the state space graph.

Question 23

(Game Theory) What is a game tree?

Answer 23

A search tree that follows every sequence of moves all the way to a terminal state.

Question 24

What is pruning?

Answer 24

Removing/ignoring large parts of a game tree that make no difference to the outcome.

Question 25

Describe a Weighted Linear Evaluation Function.

Answer 25

Computes numerical contributions from each feature (f _i) and then combines them using weights w _i from experience to find the total value.

Question 26

What are the weaknesses of Alpha-Beta Search (that are addressed by Monte Carlo Search Tree)?

Answer 26

1. Alpha-Beta Search is limited in moves when the game tree has a large branching factor
2. The evaluation function is difficult to define for complex games like Go

(where material value is not a strong enough indicator and most positions are in flux until the endgame)

Question 27

How does a Monte Carlo Search Tree strategy assign value to states (instead of a heuristic evaluation function)?

Answer 27

The value of a state is estimated as the average utility over a number of simulations of complete games starting from the state.

Question 28

What does a simulation (aka playout or rollout) do in the Monte Carlo Search Tree?

Answer 28

A simulation repeatedly chooses moves for one player then the other until a terminal position is reached.

(The rules of the game, not heuristics, determine the score and the winner.)

Question 29

What does the Monte Carlo Search selection policy do?

Answer 29

Determines which states to utilize by selectively focusing computational resources on Exploration or Exploitation of the game tree.

Question 30

What are the four steps performed every iteration of a Monte Carlo Tree Search?

Answer 30

1. Selection
2. Expansion
3. Simulation
4. Back-propagation

Question 31

What are the disadvantages of the Monte Carlo tree search?

Answer 31

1. It’s likely that a single move can change the course of the game, but MC search might fail to consider it because of its stochastic nature
2. There are game states that are “obviously” a win to humans or an evaluation function, but MC search will still take many moves in a playout to verify the winner.

Question 32

Problem-solving Agent

Answer 32

Plans ahead to find a sequence of actions that form a path to a goal state when the correct action is not immediately obvious.

Question 33

4 Step Problem-solving Process

Answer 33

• Goal Formulation
• Problem Formulation (desc. of states and actions necessary to reach goal)
• Search
• Execution