Intro AI - Exam 2

In the iterated prisoner's dilemma, in versions which allow many possible "partners" and some handling of reputation based on past behavior, - No one strategy dominates, as results depend mostly on what other agents are doing - 'mostly' cooperating but occasionally defecting can lead to the best results - always cooperating leads to the best results - always defecting leads to the best results

'mostly' cooperating but occasionally defecting can lead to the best results

Consider the following game tree. The current position is the root of the tree. The player about to move (squares) is trying to maximize the evaluation score, and the opponent (circles) is trying to minimize it. What is the backed-up value of the root? [] 0 0 0 [] [] [][] [] [][] [] [] 3,1 4,60,1 4,2 8,2 4,2 2,5 7,-2 3,4 - -2 - 4 - 1 - 12

4

What is the cause of the horizon effect in adversarial search? - Subtle threats can be completely undetectable until too late - It can be proven that no evaluation function can correctly assess every threat - Evaluation functions cannot assess certain types of positions reliably - A search cannot always distinguish between a bad outcome that's merely delayed, compared to one that's avoided completely

A search cannot always distinguish between a bad outcome that's merely delayed, compared to one that's avoided completely

Why are ensemble methods more accurate than single networks? - Even if each individual network is only about 50% correct, the consensus will often be correct - An ensemble can continue to improve even if most of its networks are already at 100% accuracy for training data - None of these are correct. - All of these are contributing factors - Different networks can train on different subsets of data

All of these are contributing factors

Poor or sub-optimal solutions in Genetic methods can be the result of - too small a population in genetic algorithms - cooling too rapidly in simulated annealing - More than one of these are possible causes, but not all of them - All of these are possible causes - culling the population too aggressively in genetic algorithms

All of these are possible causes

Which of the following can reduce overtraining effects? - Reducing the learning rate after each epoch - Dropping out a different subset of neurons for each batch - Applying regularization to reduce the magnitude of more aggressive weights - Adding noise to input data - All of these methods can reduce overtraining

All of these methods can reduce overtraining

A fundamental problem in game search is to reduce the size of the search space. This can be done by - take advantage of any left/right symmetry the game offers - Depending on the game, all of these could be possible options - tracking which positions we've already evaluated--we may have reached the same position by taking moves in a different order - Taking advantage of color symmetries; for example, if a position has a score of X, then the same position with the colors of the pieces reversed should be worth -X

Depending on the game, all of these could be possible options

In the game tree in the previous problem, which move should the player choose for best play by both sides? - The player to move (squares) should take the center move - The player to move (squares) should take the right-hand move - The player to move (squares) should take the left-hand move - Impossible to determine from the limited information provided.

The player to move (squares) should take the right-hand move

Suppose we have a function to evaluate a position that is rather optimistic; that is, whether the position is good or bad, the function's estimates are about 20% 'better' than the actual game state would indicate. - The errors in the evaluation function will probably lead the program to inconsistent play--sometimes over-aggressive, sometimes too cautious - The errors in the evaluation function will probably lead to more aggressive play - The errors in the function will probably lead to less aggressive play - This won't matter; moves will still sort into the correct order, so the best move can be selected - None of these are correct.

This won't matter; moves will still sort into the correct order, so the best move can be selected

Games of perfect information require only an initial belief to explore all possible moves using the ______________ algorithm. - minimax - alpha-beta - probcut - None of these are correct. - expectimax

minimax

In general, if the task demands a larger network, that task is likely ______________ complex and will take a ______________ time to train. - more / longer - less / shorter - less / longer - None of these are correct. - more / shorter

more / longer

Games in which everything about the game is equally accessible to both players are known as (from the text) - closed games - perfect-decision games - None of these are correct. - open games - perfect-information games

perfect-information games

The anticipated line of play representing best play by both sides is known as the - expectimax - minimax - principal continuation - optimum

principal continuation

The advantage of the alpha-beta algorithm compared to a full search is that alpha-beta - provides greater accuracy in ranking the exact merit of moves - reaches the same conclusions but searches fewer nodes - can use any of several evaluation functions - can find solutions a full search will overlook

reaches the same conclusions but searches fewer nodes

If a game is dice-based, we can estimate the value of a position by starting from the current position and 'playing out' to the end of the game several times, to see how frequent wins v. losses are. This is known as a - rollout - Venus simulation - outcome-based measurement - Vegas simulation

rollout

The purpose of the selection step in genetic algorithms is to - None of these are correct. - introduce new information into the system - ensure that the overall fitness will always improve - simulate the probability of survival due to selective pressure - give prioirty to items with the lowest scores

simulate the probability of survival due to selective pressure

In general, stochastic methods can provide - solutions that can be brought within as tight a bound of the true optimum as desired, if we are willing to expend the effort - solutions that are fairly good, but we have no way of knowing how close to the true optimum we are - a provably-optimum solution within tightly-bounded time constraints - an optimum solution in almost all cases, though proving that can be difficult - None of these are correct.

solutions that are fairly good, but we have no way of knowing how close to the true optimum we are

We are running a simulated annealing program and note that the number of modifications has been gradually decreasing and is currently much lower than when the program started. This is an indicator that - the algorithm is running normally - the temperature parameter is probably set too high - None of these are correct. - The temperature parameter should be reduced more quickly to help ensure a stable solution - The temperature parameter is probably set too low

the algorithm is running normally

The universal approximation theory of neural networks states that it is possible to construct an artificial neural network capable of estimating any function within a real-valued constraint. What is that constraint? - non-linear problems are poorly suited for a neural networks - gradient-descent methods must be used - the error of that approximation must be non-zero - to be useful, any approximation must have high flexibility - Monte Carlo methods are the most common approach

the error of that approximation must be non-zero

In carrying out search of a game tree, the backed-up score of a node is - an estimated answer produced under time pressure - the value of its parent node - the minimum or maximum of the values of its children - the average score of its children

the minimum or maximum of the values of its children

One way to improve the performance of genetic algorithms is to - set the mutation rate very high - provide hints about combinations likely to be beneficial - use a smaller population pool - use a larger population pool - set the mutation rate extremely low

use a larger population pool

The purpose of the crossover step in genetic algorithms is to - give priority to members with lower scores - None of these are correct. - introduce new information into the system - simulate different survival rates - find better combinations of existing chromosomes

find better combinations of existing chromosomes

Neural networks "learn" by a process of weight adjustment which balances the importance of selected features of the training set. Differences between the estimated output and the desired output are used to "propagate" error to those weights (connections) most responsible for that error.What is the name of the calculation that is used to determine the direction of increase or decrease in weight value relative to the magnitude of error? - None of these are correct. - Euclidian matrix - bias - derivative - gradient

gradient

A problem that only occurs outside the search radius isn't known to exist; thus if a problem can be pushed outside the search radius, it appears to just go away. This is the problem of the - hidden flaw effect - butterfly effect - oops effect - None of these are correct. - horizon effect

horizone effect

One way of improving learning is via momentum learning. This involves - alternating increasing and decreasing step size according to a fixed schedule - altering the step size or learning rate over several training runs - reducing the step size in correcting each node, as long as the steps are in the same direction - increasing the step size in correcting each node, as long as the steps are in the same direction - None of these are correct.

increasing the step size in correcting each node, as long as the steps are in the same direction

The purpose of the mutation step in genetic algorithms is to - None of these are correct. - introduce new information into the system - simulate different survival rates - find better combinations of existing chromosomes - ensure that the overall fitness will always improve

introduce new information into the system

The advantage of alpha-beta pruning is that - it evaluates each leaf node faster than a full search - None of these are correct. - it usually gives results similar to a full search with less evaluation per node - it reaches the same conclusions as a full search but examines fewer nodes - it allows immediately identifying the best possible move

it reaches the same conclusions as a full search but examines fewer nodes

One drawback of alpha-beta search is - None of these are correct. - it tends to be markedly slower than other methods - on the rare occasions its conclusions are incorrect, they tend to be extremely incorrect - its memory demands can grow without limit - its efficiency depends on the order moves are examined

its efficiency depends on the order moves are examined

One of the more common problems in backpropagation is the - None of these are correct. - multivariate gradient problem - non-analytic gradient problem - fixed gradient problem - vanishing gradient problem

vanishing gradient problem

Suppose our initial weights for a sigmoid network result in most of the neurons being saturated. We would expect this network to learn - with very poor precision - None of these are correct. - with very good precision - very slowly - very quickly

very slowly

We are running a genetic-algorithm program and note that while the average fitness is continuing to increase, the maximum fitness has held steady for several generations. This indicates - the well-known "phantom optimum" has trapped us in a loop; we should modify the fitness function and start again - we have fallen prey to the horizon effect; we should re-start the program - None of these are correct. - we might have found the true maximum, but we might also be at a plateau; we should let the program run longer. - we have stopped making progress and found the optimum; we should stop the program now

we might have found the true maximum, but we might also be at a plateau; we should let the program run longer.

One method of improving random search by changing small elements of the state is to - maintain a list of recently-changed elements and not allowing them to be changed again while on the list - None of these are correct. - change several elements at once - maintain strict randomness (the so-called 'drunkard's walk') to avoid introducing bias - change each element of the state in a fixed order

maintain a list of recently-changed elements and not allowing them to be changed again while on the list

One drawback of simulated annealing is that - very few problems can be cast correctly for it to be used - problems involvming maximazation can't be recast into minimization problems - the objective function cannot be complex or have multiple components - it still takes too long to be practical for most problems - None of these are correct.

None of these are correct.

In game theory, a Nash equilibrium is a strategy that - is dependent only on the values of the payoff matrix - is best for all parties - a strategy that neither party can change without harming their own chances - is sub-optimal if the game is zero-sum

a strategy that neither party can change without harming their own chances

In selecting a possible move, we can evaluate the positions that result from each move (not looking any farther ahead) and then taking the minimum or maximum, as appropriate. The result is that position's - None of these are correct. - backed-up score - expectimax score - heuristic score - principal continuation score

backed-up score

One common method of training a neural network involves randomly selecting a small set of samples, finding the average error for that sample, and back-propagating the the error generated by the forward pass, and adjusting the model weights to minimize the training-related loss for that group. This is known as? - cohort training - batch training - epoch selection - stochastic selection - None of these are correct.

batch training

One indicator of overtraining is - slowing or stopping of improvement with training data - declining performance with training data, continuing slow improvement with test data - None of these are correct. - abrupt sharp improvement in performance with training and test data - continued improvement with training data, declining performance with test data

continued improvement with training data, declining performance with test data

The most common type of network used for image classification is a(n) - feed-forward network - recurrent network - None of these are correct. - convolutional network - short long-term memory network

convolutional network