Difficulty: Medium
How can an admissible heuristic affect A* tree search?
A Always find every path
B Guarantee an optimal solution
C Guarantee constant time
D Eliminate all memory use
Answer: B. Guarantee an optimal solution. For A* tree search, an admissible heuristic supports optimal solution cost.
Greedy best-first search ranks nodes using what?
A Estimated remaining cost
B Training loss
C Accumulated path cost
D Branching factor only
Answer: A. Estimated remaining cost. Greedy best-first search selects by the heuristic estimate alone.
Uniform-cost search prioritizes which quantity?
A Heuristic error
B Estimated remaining cost
C Accumulated path cost
D Number of successors
Answer: C. Accumulated path cost. It expands a path with the lowest cost accumulated so far.
Alpha-beta pruning removes branches that cannot change what?
A Search goal test
B Model parameters
C Minimax decision
D Training labels
Answer: C. Minimax decision. Alpha-beta pruning avoids evaluating branches irrelevant to the minimax choice.
Minimax assumes the opponent chooses which move?
A The lowest-cost move
B Its best move
C A random move
D The shallowest move
Answer: B. Its best move. Minimax models an opponent that also chooses optimally for its side.
Bayes’ rule revises a probability using what?
A Model size
B Training duration
C Tree depth
D Observed evidence
Answer: D. Observed evidence. Bayes’ rule updates a prior belief using evidence and likelihood.
Naive Bayes assumes features are conditionally independent given what?
A Loss value
B Class label
C Training epoch
D Search state
Answer: B. Class label. The naive assumption is conditional independence of features given the class.
What operation does an internal decision-tree node perform?
A Feature test
B Final prediction
C Cluster assignment
D Reward update
Answer: A. Feature test. An internal node tests a feature and selects a branch.
How does a random forest combine its predictions?
A Only nearest neighbors
B Many decision trees
C One linear model
D A single rule base
Answer: B. Many decision trees. A random forest aggregates predictions from multiple decision trees.
Support vector machines seek which type of separator?
A Nearest centroid
B Minimum-depth tree
C Highest-reward path
D Maximum-margin separator
Answer: D. Maximum-margin separator. A support vector machine seeks a separating boundary with a large margin.
K-means assigns each point to its closest what?
A Decision boundary
B Class label
C Search frontier
D Cluster centroid
Answer: D. Cluster centroid. Each k-means point is assigned to the nearest current cluster centroid.
Principal component analysis creates directions emphasizing what?
A Shortest path cost
B Highest reward
C Greatest data variance
D Smallest label count
Answer: C. Greatest data variance. Principal components order directions by the variance they explain.
Regularization mainly discourages what?
A Additional training data
B Correct predictions
C Excessive model complexity
D Validation splits
Answer: C. Excessive model complexity. Regularization penalizes complexity to help control overfitting.
During dropout training, selected activations are what?
A Always increased
B Sorted by value
C Copied to labels
D Randomly set to zero
Answer: D. Randomly set to zero. Dropout randomly disables units during training to reduce reliance on them.
Evaluation leakage occurs when what enters training?
A More training examples
B Additional input features
C Test-set information
D A larger batch
Answer: C. Test-set information. Using test information during training contaminates evaluation.
K-fold cross-validation commonly summarizes scores using what?
A The first fold
B Their maximum
C The training loss
D Their average
Answer: D. Their average. The usual k-fold score is the average across held-out folds.
Why can severe class imbalance mislead accuracy?
A Majority guesses may score well
B Accuracy ignores all labels
C It requires continuous outputs
D It counts only false positives
Answer: A. Majority guesses may score well. A model can appear accurate by predicting the majority class repeatedly.
On a ROC curve, which rates are plotted?
A True-positive and false-positive rates
B Precision and accuracy
C Bias and variance
D Loss and recall
Answer: A. True-positive and false-positive rates. ROC curves show true-positive rate versus false-positive rate across thresholds.
For a calibrated model, predicted probabilities should match what?
A Observed outcome frequencies
B Feature magnitudes
C Search costs
D Training-set size
Answer: A. Observed outcome frequencies. Calibration means predicted probabilities align with observed event frequencies.
Why scale features before distance-based learning?
A Larger reward values
B Fewer class labels
C Comparable feature magnitudes
D Deeper decision trees
Answer: C. Comparable feature magnitudes. Scaling prevents large-unit features from dominating distance calculations.
Raising a binary classification threshold usually yields what?
A More training labels
B Fewer positive predictions
C A larger feature set
D More positive predictions
Answer: B. Fewer positive predictions. A higher threshold generally requires stronger scores to predict the positive class.
For binary logistic regression, sigmoid maps a score to what?
A A cluster identifier
B A tree depth
C Probability from 0 to 1
D A reward sequence
Answer: C. Probability from 0 to 1. The sigmoid function maps a real-valued score into the interval from zero to one.
Data augmentation expands training variation through what?
A Test-label copying
B Label-preserving transformations
C Random class removal
D Loss rescaling
Answer: B. Label-preserving transformations. Augmentation creates varied training examples while preserving their intended labels.
Concept drift describes which change?
A Changing input-target relationships
B A deeper network
C A smaller training batch
D A fixed class vocabulary
Answer: A. Changing input-target relationships. Concept drift occurs when the relationship underlying data changes over time.
Reinforcement learning balances which two aims?
A Training and testing
B Precision and recall
C Bias and variance
D Exploration and exploitation
Answer: D. Exploration and exploitation. An agent must trade off trying actions and using known rewarding actions.