Font size
WorksheetsWorksheet Questions Extraction
Total questions: 84
Worksheet time: 42mins
A learning task is defined by ⟨T, P, E⟩. If a system improves its performance at playing chess with experience, then P refers to:
Winning strategy
Percentage of games won
Number of board states
Training dataset size
Which of the following is NOT a valid perspective of machine learning?
Statistical
Algorithmic
Biological
Deterministic logic
A concept is represented as a Boolean function over attribute space. This representation is most commonly used in:
Reinforcement learning
Concept learning
Clustering
Dimensionality reduction
In concept learning, noise in training data primarily affects:
Hypothesis representation
Learning rate
Consistency of hypothesis
Target function
If the hypothesis space contains the target concept, the learner is said to be:
Complete
Sound
Realizable
Optimal
The version space represents:
All hypotheses consistent with training data
All possible hypotheses
Only the most general hypothesis
Only the most specific hypothesis
The S-boundary in Candidate Elimination represents:
Maximally general hypotheses
Minimally general hypotheses
Maximally specific hypotheses
Minimally specific hypotheses
When a positive example is misclassified by S, the algorithm will:
Remove S
Generalize S minimally
Specialize G
Discard the example
When a negative example is covered by G, the algorithm will:
Generalize G
Remove S
Specialize G minimally
Expand version space
Candidate Elimination fails when:
Hypothesis space is finite
Training data contains noise
Data is linearly separable
Version space is non-empty
Inductive bias refers to:
Error in data
Prior assumptions made by the learner
Sampling bias
Training bias
Without inductive bias, a learner can:
Generalize perfectly
Learn any function
Only memorize training data
Always overfit
Occam’s Razor is an example of:
Statistical bias
Computational bias
Inductive bias
Representation bias
A learner preferring linear hypotheses over nonlinear ones is an example of:
Search bias
Representation bias
Sample bias
Noise bias
Which learning algorithm has no explicit inductive bias?
Decision Trees
KNN
Naïve Bayes
Neural Networks
Decision tree learning performs a:
Depth-first search
Greedy search
Exhaustive search
Random search
Information Gain is based on:
Variance
Entropy
Mean squared error
Probability density
A highly unbalanced dataset may cause decision trees to:
Underfit
Ignore minority class
Increase depth unnecessarily
Fail to converge
Which measure reduces bias toward multi-valued attributes?
Information Gain
Gini Index
Gain Ratio
Chi-square
Pruning in decision trees helps to:
Increase training accuracy
Reduce variance
Which representation allows disjunctions but not conjunctions?
Decision trees
Propositional logic
Linear classifiers
Conjunctive hypotheses
The expressive power of a hypothesis space determines:
Learning speed
Generalization ability
Whether the target can be represented
Training time
A more expressive hypothesis space increases the risk of:
Underfitting
Overfitting
Bias
Data leakage
Which hypothesis representation is most interpretable?
Neural networks
Decision trees
SVM
KNN
Conjunctive hypothesis spaces are limited because they cannot represent:
AND relations
OR relations
Linear boundaries
Boolean logic
A heuristic is used to:
Guarantee optimality
Reduce search space
Increase hypothesis space
Eliminate noise
Heuristic search in learning primarily trades off:
Bias and variance
Optimality and efficiency
Accuracy and recall
Precision and speed
Which search strategy is used in decision tree learning?
Backtracking search
Greedy heuristic search
Uniform-cost search
Hill climbing
A heuristic that always chooses the locally optimal split may lead to:
Global optimum
Local optimum
Overfitting only
Infinite loop
Search in hypothesis space is exponential mainly due to:
Noise
Dimensionality
Sample size
Overfitting
If both S and G converge to a single hypothesis, the learner has:
Failed
Generalized
Learned the target concept
Overfitted
A large version space indicates:
High confidence
Insufficient data
Perfect learning
Low hypothesis complexity
Adding irrelevant attributes to a decision tree dataset usually:
Improves accuracy
Reduces tree depth
Increases overfitting risk
Has no effect
Which learning setting is most affected by inductive bias?
Supervised
Unsupervised
Reinforcement
All learning settings
A learner that memorizes all examples suffers from:
High bias
High variance
Underfitting
Low complexity
Candidate Elimination assumes the training data is:
Probabilistic
Noisy
Noise-free
Continuous
Which of the following is a search bias?
Limiting hypothesis space
Preferring shorter trees
Using entropy
Removing noise
A decision tree with depth equal to number of attributes implies:
Underfitting
Balanced learning
Possible overfitting
Optimal learning
If entropy before and after split remains same, information gain is:
1
−1
0
Maximum
Learning problems with continuous attributes often require:
Discretization
Elimination
Encoding
Normalization only
Which factor primarily controls decision tree complexity?
Learning rate
Depth
Feature scaling
Dataset size
A hypothesis consistent with all examples may still fail due to:
Noise
Overgeneralization
Poor generalization
Low bias
Greedy learning algorithms are preferred because they:
Guarantee optimality
Reduce computational cost
Avoid bias
Remove noise
Which component defines what is learnable?
Training set
Hypothesis space
Learning rate
Loss function
Increasing training examples generally causes version space to:
Expand
Remain same
Shrink
Randomize
Decision trees naturally handle:
Missing values
Only numeric data
Only Boolean data
Only balanced datasets
A heuristic that is admissible ensures:
Faster learning
Optimal solution
No overfitting
Noise tolerance
Which is a limitation of decision tree learning?
Interpretability
Handling non-linear boundaries
High variance
Greedy search
Bias-variance tradeoff mainly affects:
Training accuracy
Model selection
Feature scaling
Data cleaning
A learner that always predicts the majority class demonstrates:
High variance
High bias
Optimal learning
Noise fitting
Learning Problems & Concept Learning: A learning problem where the output is a continuous real value is best categorized as:
Classification
Clustering
Regression
Reinforcement
Learning Problems & Concept Learning: If the training accuracy is high but test accuracy is low, the model is likely:
Underfitting
Overfitting
Well-generalized
Noise-free
Learning Problems & Concept Learning: In concept learning, the target function represents:
Learner hypothesis
Optimal hypothesis
True mapping from instances to labels
Training data distribution
Learning Problems & Concept Learning: Which factor determines whether a hypothesis is consistent?
Hypothesis size
Dataset size
Agreement with all training examples
Prediction confidence
Learning becomes impossible if:
Hypothesis space is infinite
No inductive bias is present
Training data is large
Learning rate is small
Version Space Theory: The version space shrinks when:
More hypotheses are added
More training examples are observed
Hypothesis space expands
Noise is added
Version Space Theory: If the S-boundary becomes empty, it indicates:
Successful learning
No hypothesis fits positive examples
Overgeneralization
Complete version space
Version Space Theory: If G-boundary becomes empty during learning, it implies:
Overfitting
No hypothesis can explain negative examples
Learning failure
Noise-free data
Candidate Elimination maintains:
Only S
Only G
Both S and G
Only most recent hypothesis
Which operation is applied to S when encountering a negative example?
Generalization
Specialization
No change
Deletion
Inductive Bias & Generalization: Bias introduced by restricting hypothesis space is known as:
Search bias
Statistical bias
Representation bias
Sampling bias
Inductive Bias & Generalization: A learner that prefers smaller trees has bias toward:
Accuracy
Simplicity
Completeness
Variance
Inductive Bias & Generalization: Which of the following best explains generalization?
The ability of a model to perform well on unseen data by applying learned patterns
Memorizing the training set to achieve perfect fit on observed examples
Fitting highly complex models to capture every fluctuation in the training data
Relying solely on large datasets without considering inductive bias
Bias-variance tradeoff implies:
Lower bias always improves accuracy
Lower variance always improves accuracy
Optimal balance is required
Bias and variance are independent
Increasing model complexity generally:
Increases bias
Decreases variance
Decreases bias and increases variance
Improves generalization always
Which criterion selects the attribute that best separates classes?
Entropy
Information Gain
Probability
Variance
Decision tree learning is an example of:
Lazy learning
Eager learning
Reinforcement learning
Unsupervised learning
A tree with zero depth corresponds to:
Perfect classifier
Random classifier
Single node tree
Overfitted model
Which stopping condition helps avoid overfitting?
All attributes used
Node purity threshold
Maximum depth restriction
All of the above
Which scenario leads to the deepest decision tree?
Few attributes, balanced classes
Many irrelevant attributes
Linearly separable data
Small dataset
A hypothesis space that cannot represent the target function leads to:
Overfitting
Underfitting
High variance
Noise sensitivity
Which representation can model XOR easily?
Linear classifier
Decision tree
Single-layer perceptron
Naïve Bayes
Increasing hypothesis expressiveness affects learning by:
Reducing training error
Increasing search complexity
Increasing overfitting risk
All of the above
Which representation supports recursive partitioning?
Neural networks
Decision trees
Rule-based systems
Linear regression
A hypothesis that predicts the same output for all inputs has:
Zero bias
Zero variance
High bias
High complexity
Heuristic search is preferred because exhaustive search is:
Impossible
Inaccurate
Computationally expensive
Unreliable
Greedy algorithms may fail due to:
High bias
Getting trapped in local optima
Overfitting
Noise
Search space in learning refers to:
Input space
Feature space
Hypothesis space
Output space
A heuristic that underestimates cost is called:
Consistent
Inadmissible
Admissible
Complete
Which factor increases heuristic effectiveness?
Randomness
Domain knowledge
Noise
High dimensionality
If adding training data does not reduce error, the likely cause is:
Overfitting
High bias
Data leakage
Noise
A learning algorithm that updates hypothesis after every example is:
Batch learning
Online learning
Unsupervised learning
Lazy learning
Decision tree pruning mainly addresses:
Bias
Variance
Noise
Data imbalance
If hypothesis space is too small, the learner suffers from:
Overfitting
High variance
Underfitting
Noise sensitivity
