EduConceptHub

Start MCQs

Artificial Intelligence

Artificial Intelligence

Introduction

Explore how computer systems represent knowledge, search for solutions and learn from data.

200 AI MCQs available: Seven practice sets with four options, correct answers and explanations. Questions 31–60 · 61–90 · 91–120 · 121–150 · 151–175 · 176–200. The first 30 questions appear below the study overview.

Topic-Wise Overview

AI Foundations

AI applications, intelligent systems and knowledge representation

Search and Planning

Uninformed and informed search, heuristics and local search

Games and Decisions

Minimax and alpha-beta pruning

Learning and Language

Learning from examples, neural networks and natural language processing

Practical Exploration

Illustrative AI methods using Python and applied case studies

Study Guidance

Start with the core definitions, follow practical examples and test your understanding with the 200 topic-wise MCQs and explained answers provided below. Additional short-question collections and step-by-step worked solutions will be developed progressively.

Explore Other Subjects

Browse Computer Science subjects

Artificial Intelligence MCQs: Set 1 (Questions 1–30)

Choose an answer, then reveal the correct option and explanation. This is the first set toward a 150–200 question bank.

Question 1. What is the main aim of artificial intelligence?

  1. To make all computers physically larger
  2. To eliminate every algorithm
  3. To store only numerical data
  4. To build systems that perform tasks requiring aspects of human intelligence
Show correct answer and explanation

Correct answer: D. To build systems that perform tasks requiring aspects of human intelligence

AI develops computational systems that can perceive, reason, learn or act in ways useful for intelligent tasks.

Question 2. Which is an example of narrow AI?

  1. A system that solves every intellectual task
  2. A machine with unrestricted human-level capabilities
  3. A chess-playing program
  4. A device without software
Show correct answer and explanation

Correct answer: C. A chess-playing program

A chess program is designed for a specific task rather than general intelligence.

Question 3. In an intelligent agent, sensors are primarily used to:

  1. Perceive the environment
  2. Execute actions
  3. Store the entire internet
  4. Replace the performance measure
Show correct answer and explanation

Correct answer: A. Perceive the environment

Sensors provide observations or percepts about the environment.

Question 4. An actuator allows an agent to:

  1. Receive percepts only
  2. Define a database schema
  3. Compute probabilities without data
  4. Act on its environment
Show correct answer and explanation

Correct answer: D. Act on its environment

Actuators implement actions chosen by an agent.

Question 5. A rational agent selects actions to:

  1. Maximize expected performance based on available information
  2. Always choose the shortest-looking action
  3. Ignore its percept history
  4. Avoid using any model
Show correct answer and explanation

Correct answer: A. Maximize expected performance based on available information

Rationality concerns expected performance given the percepts, knowledge and available actions.

Question 6. Which task environment is fully observable?

  1. The agent receives no percepts
  2. An agent can observe all relevant aspects of the current state
  3. Every action has a random outcome
  4. The agent cannot distinguish states
Show correct answer and explanation

Correct answer: B. An agent can observe all relevant aspects of the current state

Full observability means sensors provide all information relevant to choosing an action at a given time.

Question 7. A deterministic environment is one in which:

  1. The next state is fixed by the current state and action
  2. Actions have no possible outcomes
  3. Every action is equally likely
  4. The agent has no goal
Show correct answer and explanation

Correct answer: A. The next state is fixed by the current state and action

Deterministic transition rules uniquely determine the next state from the current state and action.

Question 8. Which search method uses a FIFO frontier?

  1. Depth-first search
  2. Breadth-first search
  3. Greedy best-first search
  4. Hill climbing
Show correct answer and explanation

Correct answer: B. Breadth-first search

Breadth-first search expands nodes in first-in, first-out order.

Question 9. Which search method typically uses a LIFO stack?

  1. Depth-first search
  2. Uniform-cost search
  3. Breadth-first search
  4. A* with a priority queue
Show correct answer and explanation

Correct answer: A. Depth-first search

Depth-first search follows the most recently discovered frontier node first.

Question 10. Uniform-cost search selects the frontier node with the:

  1. Largest depth
  2. Smallest path cost so far
  3. Highest heuristic value
  4. Most children
Show correct answer and explanation

Correct answer: B. Smallest path cost so far

Uniform-cost search expands the node with minimum accumulated cost g(n).

Question 11. A heuristic function h(n) commonly estimates:

  1. The number of stored variables
  2. The remaining cost from node n to a goal
  3. The number of computers
  4. The exact search time in seconds
Show correct answer and explanation

Correct answer: B. The remaining cost from node n to a goal

A heuristic guides search by estimating cost from the current node to a goal.

Question 12. The evaluation function for standard A* search is:

  1. f(n) = g(n) − h(n)
  2. f(n) = g(n) × h(n)
  3. f(n) = h(n) / g(n)
  4. f(n) = g(n) + h(n)
Show correct answer and explanation

Correct answer: D. f(n) = g(n) + h(n)

A* combines known path cost g(n) with estimated remaining cost h(n).

Question 13. An admissible heuristic is one that:

  1. Always overestimates goal cost
  2. Has the same value at every node
  3. Never overestimates true remaining optimal cost
  4. Always equals zero at every node
Show correct answer and explanation

Correct answer: C. Never overestimates true remaining optimal cost

For a minimization problem, admissibility means h(n) does not exceed the true optimal remaining cost.

Question 14. Which algorithm is designed for adversarial two-player games?

  1. K-means
  2. Apriori
  3. Dijkstra’s algorithm
  4. Minimax
Show correct answer and explanation

Correct answer: D. Minimax

Minimax evaluates choices assuming that an opponent chooses moves to minimize the first player’s utility.

Question 15. Alpha-beta pruning is used to:

  1. Change the rules of a game
  2. Train a classifier
  3. Convert text to speech
  4. Avoid evaluating branches that cannot affect the minimax decision
Show correct answer and explanation

Correct answer: D. Avoid evaluating branches that cannot affect the minimax decision

Alpha-beta bounds can eliminate game-tree branches without changing the minimax result.

Question 16. In propositional logic, P ∧ Q is true when:

  1. At least one proposition is true
  2. Exactly one proposition is false
  3. Both propositions are false
  4. Both P and Q are true
Show correct answer and explanation

Correct answer: D. Both P and Q are true

Conjunction requires both constituent propositions to be true.

Question 17. Which connective represents logical implication?

  1. P → Q
  2. P ∧ Q
  3. P ∨ Q
  4. ¬P
Show correct answer and explanation

Correct answer: A. P → Q

The symbol → represents ‘if P, then Q’.

Question 18. A knowledge base in an AI system stores:

  1. Represented facts and rules used for reasoning
  2. Only executable machine instructions
  3. Only image pixels
  4. Physical sensor hardware
Show correct answer and explanation

Correct answer: A. Represented facts and rules used for reasoning

Knowledge bases contain formal representations that support inference.

Question 19. Which inference rule derives Q from P and P → Q?

  1. Modus ponens
  2. De Morgan’s law
  3. Commutativity
  4. Resolution by itself
Show correct answer and explanation

Correct answer: A. Modus ponens

Modus ponens permits inferring Q when P and the implication P → Q are known.

Question 20. In supervised learning, training examples usually include:

  1. Only unlabeled inputs
  2. Inputs paired with target labels or values
  3. Only random guesses
  4. Only model weights
Show correct answer and explanation

Correct answer: B. Inputs paired with target labels or values

Supervised learning uses labeled input–target examples to learn a predictive mapping.

Question 21. Which task is a classification problem?

  1. Predicting tomorrow’s exact temperature
  2. Estimating a house price
  3. Assigning an email to spam or not spam
  4. Calculating a numerical average
Show correct answer and explanation

Correct answer: C. Assigning an email to spam or not spam

Spam detection predicts a category, making it a classification task.

Question 22. Which task is a regression problem?

  1. Assigning a news article to a category
  2. Predicting a continuous house price
  3. Grouping unlabeled customers
  4. Checking a Boolean formula
Show correct answer and explanation

Correct answer: B. Predicting a continuous house price

Regression predicts a numerical quantity, such as a house price.

Question 23. Unsupervised clustering is primarily used to:

  1. Group similar unlabeled observations
  2. Assign known target labels during training
  3. Guarantee causal explanations
  4. Solve every search problem
Show correct answer and explanation

Correct answer: A. Group similar unlabeled observations

Clustering seeks structure in data without requiring predefined class labels.

Question 24. Overfitting occurs when a model:

  1. Cannot fit any training data
  2. Contains no parameters
  3. Performs well on training data but generalizes poorly
  4. Uses a test set correctly
Show correct answer and explanation

Correct answer: C. Performs well on training data but generalizes poorly

An overfit model learns patterns or noise specific to training data that do not transfer well.

Question 25. Why is a test set kept separate from model training?

  1. To increase the number of labels
  2. To guarantee perfect predictions
  3. To remove all bias automatically
  4. To estimate performance on unseen data
Show correct answer and explanation

Correct answer: D. To estimate performance on unseen data

Held-out test data provide a more credible estimate of generalization when not used to tune the model.

Question 26. In binary classification, a true positive means:

  1. Positive was predicted but actual was negative
  2. Positive was predicted and the actual class was positive
  3. Negative was predicted but actual was positive
  4. Negative was predicted and actual was negative
Show correct answer and explanation

Correct answer: B. Positive was predicted and the actual class was positive

True positives are correctly predicted members of the positive class.

Question 27. Precision is defined as:

  1. TP / (TP + FN)
  2. TN / (TN + FP)
  3. TP / (TP + FP)
  4. (TP + TN) / TP
Show correct answer and explanation

Correct answer: C. TP / (TP + FP)

Precision is the fraction of predicted positives that are actually positive, provided the denominator is nonzero.

Question 28. Recall is defined as:

  1. TP / (TP + FP)
  2. FP / (FP + TN)
  3. TN / (TN + FN)
  4. TP / (TP + FN)
Show correct answer and explanation

Correct answer: D. TP / (TP + FN)

Recall measures the proportion of actual positives successfully identified.

Question 29. The main purpose of an activation function in a neural network is to:

  1. Ensure all weights are zero
  2. Replace the training dataset
  3. Provide nonlinearity in the network’s computation
  4. Remove every prediction error
Show correct answer and explanation

Correct answer: C. Provide nonlinearity in the network’s computation

Nonlinear activation functions allow multilayer networks to represent nonlinear relationships.

Question 30. Natural language processing primarily concerns:

  1. Only computer hardware design
  2. Mechanical motion planning
  3. Database storage capacity
  4. Human language text and speech
Show correct answer and explanation

Correct answer: D. Human language text and speech

NLP develops methods for understanding, analyzing and generating human language.

Artificial Intelligence MCQs: Set 2 (Questions 31–60)

Choose your answer and reveal the explanation. Study each question before moving to the next.

Question 31. In reinforcement learning, an agent primarily learns through:

  1. Labeled examples for every action
  2. Only prewritten logical rules
  3. Rewards received from interactions
  4. Deleting training data
Show correct answer and explanation

Correct answer: C. Rewards received from interactions

An agent uses reward feedback from interactions to learn a policy for future actions.

Question 32. What is a policy in reinforcement learning?

  1. A list of target labels
  2. A physical sensor
  3. A mapping from states to actions
  4. The number of network layers
Show correct answer and explanation

Correct answer: C. A mapping from states to actions

A policy specifies the action to select in each state, possibly probabilistically.

Question 33. The exploration–exploitation tradeoff involves choosing between:

  1. Increasing memory and reducing power
  2. Trying less-known actions and using currently promising actions
  3. Classification and compression
  4. Training and deleting data
Show correct answer and explanation

Correct answer: B. Trying less-known actions and using currently promising actions

Exploration gathers information while exploitation uses current knowledge to gain reward.

Question 34. A decision tree classifier makes predictions using:

  1. Only unlabeled samples
  2. An array without branches
  3. A necessarily diagonal matrix
  4. A hierarchy of feature-based tests
Show correct answer and explanation

Correct answer: D. A hierarchy of feature-based tests

The input follows branches determined by feature tests until reaching a prediction.

Question 35. The k-nearest neighbors classifier commonly predicts using:

  1. Labels of nearby training examples
  2. An adversarial game tree
  3. Only the average of all labels
  4. A hand-written grammar
Show correct answer and explanation

Correct answer: A. Labels of nearby training examples

KNN uses nearby labeled examples according to a chosen distance metric.

Question 36. What is the Euclidean distance from (0, 0) to (3, 4)?

  1. 5
  2. 4
  3. 7
  4. 25
Show correct answer and explanation

Correct answer: A. 5

The distance is the square root of 3 squared plus 4 squared, or 5.

Question 37. A confusion matrix summarizes:

  1. Randomly generated model weights
  2. Website passwords
  3. Predicted and actual classification outcomes
  4. Shortest paths in a network
Show correct answer and explanation

Correct answer: C. Predicted and actual classification outcomes

Each cell counts predictions for a particular actual class and predicted class.

Question 38. In binary classification, a false negative occurs when:

  1. The model predicts positive and the actual class is positive
  2. The model predicts positive but the actual class is negative
  3. The model predicts negative but the actual class is positive
  4. The model predicts negative and the actual class is negative
Show correct answer and explanation

Correct answer: C. The model predicts negative but the actual class is positive

A false negative is a positive case missed by the classifier.

Question 39. Which is a typical use of a validation set?

  1. Selecting hyperparameters
  2. Reporting an untouched final test score
  3. Replacing all training observations
  4. Guaranteeing zero model error
Show correct answer and explanation

Correct answer: A. Selecting hyperparameters

Validation data help compare model settings without using the final held-out test set.

Question 40. Gradient descent adjusts parameters in the direction of:

  1. The positive gradient of the loss
  2. The negative gradient of a loss function
  3. A random label for each example
  4. The largest feature value
Show correct answer and explanation

Correct answer: B. The negative gradient of a loss function

For a sufficiently small learning rate, stepping along the negative gradient tends to reduce a differentiable loss locally.

Question 41. What does the learning rate control in gradient descent?

  1. The number of class labels
  2. The number of input records
  3. The size of parameter updates
  4. The meaning of precision
Show correct answer and explanation

Correct answer: C. The size of parameter updates

The learning rate scales the update step based on the gradient.

Question 42. A loss function is used to:

  1. Measure prediction error or model mismatch
  2. Convert sensors into actuators
  3. List all search nodes
  4. Guarantee an unbiased dataset
Show correct answer and explanation

Correct answer: A. Measure prediction error or model mismatch

Training minimizes an objective measuring how far model predictions are from desired outcomes.

Question 43. In a neural network, weights primarily:

  1. Specify the number of dataset rows
  2. Store the testing date
  3. Prevent all overfitting
  4. Scale contributions from inputs to units
Show correct answer and explanation

Correct answer: D. Scale contributions from inputs to units

Weights determine the strength of connections used to compute activations.

Question 44. Backpropagation is used to:

  1. Sort examples alphabetically
  2. Find the shortest path in any graph
  3. Create labels without human input
  4. Compute gradients of neural-network loss with respect to parameters
Show correct answer and explanation

Correct answer: D. Compute gradients of neural-network loss with respect to parameters

Backpropagation applies the chain rule to calculate gradients efficiently.

Question 45. Which function maps a real number to a value strictly between 0 and 1?

  1. Identity function
  2. Square function
  3. Sigmoid function
  4. Absolute-value function
Show correct answer and explanation

Correct answer: C. Sigmoid function

The logistic sigmoid is 1/(1+e^(-x)), which lies strictly between zero and one for any finite real x.

Question 46. A convolutional neural network is especially suited to:

  1. Only solving symbolic equations
  2. Replacing all kinds of search
  3. Processing image-like spatial data
  4. Guaranteeing perfect forecasts
Show correct answer and explanation

Correct answer: C. Processing image-like spatial data

Convolutions exploit local spatial patterns and shared filters, which are useful for image tasks.

Question 47. What is the role of tokenization in natural language processing?

  1. Encrypting network traffic
  2. Training without data
  3. Breaking text into units such as words or subwords
  4. Finding a shortest road route
Show correct answer and explanation

Correct answer: C. Breaking text into units such as words or subwords

Tokenization segments text into units that a language-processing system can handle.

Question 48. An embedding represents a word or token as:

  1. A physical circuit board
  2. A binary class label only
  3. A numerical vector
  4. A fixed-length handwritten sentence
Show correct answer and explanation

Correct answer: C. A numerical vector

Embeddings map tokens into numeric vector spaces used by models.

Question 49. In machine translation, the input and output are generally:

  1. Texts in different languages
  2. An image and a motor signal
  3. Two physical temperatures
  4. Only file names
Show correct answer and explanation

Correct answer: A. Texts in different languages

Machine translation maps text in a source language to text in a target language.

Question 50. Which mechanism lets a transformer model weigh the relevance of different input tokens?

  1. Attention
  2. Bubble sort
  3. Binary search
  4. Alpha-beta pruning
Show correct answer and explanation

Correct answer: A. Attention

Attention computes context-dependent weights over token representations.

Question 51. A generative AI model is designed to:

  1. Store only exact copies of every training record
  2. Remove the need to evaluate outputs
  3. Guarantee factual correctness
  4. Produce new content based on learned patterns
Show correct answer and explanation

Correct answer: D. Produce new content based on learned patterns

Generative models create outputs such as text, images or audio conditioned on learned distributions and prompts.

Question 52. An AI hallucination refers to:

  1. A computer overheating
  2. A camera losing power
  3. A successful proof verified independently
  4. An output that presents unsupported or false information as though it were valid
Show correct answer and explanation

Correct answer: D. An output that presents unsupported or false information as though it were valid

A generated statement can sound confident yet be unsupported or incorrect.

Question 53. Which practice helps evaluate fairness in a classification system?

  1. Compare relevant error rates across affected groups
  2. Use only overall accuracy
  3. Ignore missing data
  4. Remove all performance tests
Show correct answer and explanation

Correct answer: A. Compare relevant error rates across affected groups

Examining subgroup performance can reveal disparities hidden by aggregate metrics.

Question 54. Data leakage occurs when:

  1. Data are stored in a database
  2. A model receives legitimate training labels
  3. An experiment uses a separate test set correctly
  4. Information unavailable at prediction time improperly influences training or evaluation
Show correct answer and explanation

Correct answer: D. Information unavailable at prediction time improperly influences training or evaluation

Leakage can make estimated performance unrealistically high by revealing information that should not be available.

Question 55. A common privacy risk in AI development is:

  1. Exposing personal details from training or user data
  2. Using numerical matrices
  3. Splitting a dataset for testing
  4. Computing a mean
Show correct answer and explanation

Correct answer: A. Exposing personal details from training or user data

Sensitive data can be exposed through collection, storage, outputs or model behavior.

Question 56. In graph search, a cycle can cause difficulty because:

  1. All graphs are trees
  2. Every edge has negative cost
  3. Repeated states may be revisited indefinitely without safeguards
  4. No state has neighbors
Show correct answer and explanation

Correct answer: C. Repeated states may be revisited indefinitely without safeguards

Cycle detection or a visited-state mechanism can prevent unproductive repeated expansions.

Question 57. Which search strategy uses h(n) alone to prioritize frontier nodes?

  1. Uniform-cost search
  2. Breadth-first search
  3. Depth-first search
  4. Greedy best-first search
Show correct answer and explanation

Correct answer: D. Greedy best-first search

Greedy best-first search prioritizes the estimated remaining cost h(n), without directly accounting for accumulated cost.

Question 58. A consistent heuristic h satisfies, for every edge from n to n’ with cost c:

  1. h(n) > c + h(n’)
  2. h(n) ≤ c + h(n’)
  3. h(n) = 2c for every edge
  4. h(n) is always negative
Show correct answer and explanation

Correct answer: B. h(n) ≤ c + h(n’)

Consistency is the triangle inequality for heuristic estimates along an edge; goal heuristics are usually set to zero.

Question 59. If events A and B are independent, then P(A ∩ B) equals:

  1. P(A)P(B)
  2. P(A)+P(B)
  3. P(A)/P(B)
  4. 1−P(A)
Show correct answer and explanation

Correct answer: A. P(A)P(B)

For independent events the probability that both occur is the product of their probabilities.

Question 60. Which statement correctly describes Bayes’ theorem when P(B)>0?

  1. P(A|B)=P(B|A)P(A)/P(B)
  2. P(A|B)=P(A)+P(B)
  3. P(A|B)=P(B)/P(A)
  4. P(A|B)=1−P(B|A)
Show correct answer and explanation

Correct answer: A. P(A|B)=P(B|A)P(A)/P(B)

Bayes’ theorem relates a posterior conditional probability to likelihood, prior and evidence.

Artificial Intelligence MCQs: Set 3 (Questions 61–90)

Study each question, select your answer, and reveal the correct option with a concise explanation.

Question 61. Which distribution models successes in a fixed number of independent Bernoulli trials?

  1. Exponential
  2. Binomial
  3. Continuous uniform
  4. Cauchy
Show correct answer and explanation

Correct answer: B. Binomial

A binomial variable counts successes across a fixed number of independent trials with a common success probability.

Question 62. What is the expected value of a Bernoulli variable with success probability p?

  1. 1−p
  2. p²
  3. 1/p
  4. p
Show correct answer and explanation

Correct answer: D. p

The variable is one with probability p and zero otherwise, so its expectation is p.

Question 63. What is the probability of rolling a number greater than 4 on a fair six-sided die?

  1. 1/3
  2. 1/6
  3. 1/2
  4. 2/3
Show correct answer and explanation

Correct answer: A. 1/3

Outcomes 5 and 6 qualify, so the probability is 2/6 = 1/3.

Question 64. What does P(A|B) denote?

  1. Probability of A given B
  2. Probability of B given A
  3. Probability of A plus B
  4. Probability that A is impossible
Show correct answer and explanation

Correct answer: A. Probability of A given B

Conditional probability is the probability of A when B is known to occur, provided P(B) > 0.

Question 65. In a Bayesian network, nodes represent:

  1. Only computers
  2. Program keywords
  3. File folders
  4. Random variables
Show correct answer and explanation

Correct answer: D. Random variables

Nodes denote random variables, and directed edges encode probabilistic dependencies.

Question 66. The graph underlying a Bayesian network must be:

  1. A complete undirected graph
  2. A graph with directed cycles
  3. A directed acyclic graph
  4. A single path
Show correct answer and explanation

Correct answer: C. A directed acyclic graph

Bayesian networks use directed acyclic graphs (DAGs).

Question 67. Monte Carlo methods commonly estimate quantities by:

  1. Using random samples
  2. Sorting only
  3. Deleting observations
  4. Applying minimax
Show correct answer and explanation

Correct answer: A. Using random samples

Monte Carlo estimation uses random sampling to approximate numerical quantities.

Question 68. Shannon entropy measures:

  1. Processor frequency
  2. Number of labels only
  3. Uncertainty in a probability distribution
  4. Physical memory size
Show correct answer and explanation

Correct answer: C. Uncertainty in a probability distribution

Entropy quantifies uncertainty in a probability distribution.

Question 69. What is the entropy of a fair binary outcome, in bits?

  1. 0
  2. 0.5
  3. 2
  4. 1
Show correct answer and explanation

Correct answer: D. 1

Using −Σp log₂p with two probabilities of 1/2 gives 1 bit.

Question 70. Information gain in decision trees usually measures:

  1. Training duration
  2. Entropy reduction from a split
  3. Maximum tree depth only
  4. Number of features deleted
Show correct answer and explanation

Correct answer: B. Entropy reduction from a split

Information gain is parent entropy minus the weighted average entropy after splitting.

Question 71. Which technique is used to identify named people, locations, or organizations in text?

  1. Pathfinding
  2. Image sharpening
  3. Named entity recognition
  4. Gradient clipping
Show correct answer and explanation

Correct answer: C. Named entity recognition

Named entity recognition identifies and classifies mentions of entities in language.

Question 72. What is sentiment analysis most directly concerned with?

  1. Determining expressed opinion or emotional polarity
  2. Counting graph edges
  3. Finding a minimum spanning tree
  4. Choosing a CPU
Show correct answer and explanation

Correct answer: A. Determining expressed opinion or emotional polarity

Sentiment analysis estimates attitudes or polarity conveyed by text.

Question 73. A language model commonly estimates:

  1. Exact future economic events
  2. Probabilities of token sequences
  3. Circuit resistance only
  4. The number of sensors
Show correct answer and explanation

Correct answer: B. Probabilities of token sequences

Language models assign probabilities to tokens or sequences in a linguistic context.

Question 74. What is a prompt in a generative AI application?

  1. The model’s physical processor
  2. A guarantee of accuracy
  3. A password encryption algorithm
  4. Input instructions or context supplied to a model
Show correct answer and explanation

Correct answer: D. Input instructions or context supplied to a model

A prompt provides context or instructions that condition generated output.

Question 75. Retrieval-augmented generation typically combines:

  1. External document retrieval with generation
  2. Only sorting and filtering
  3. Only compression and encryption
  4. Breadth-first and depth-first search
Show correct answer and explanation

Correct answer: A. External document retrieval with generation

RAG retrieves relevant source material and passes it as context to a generative model.

Question 76. A common reason to cite retrieved sources in an AI answer is to:

  1. Guarantee zero errors
  2. Avoid all training
  3. Allow readers to inspect supporting evidence
  4. Increase parameter count
Show correct answer and explanation

Correct answer: C. Allow readers to inspect supporting evidence

Citations let readers check source support; they do not by themselves guarantee truth.

Question 77. Which of these is an example of adversarial input for an image classifier?

  1. A carefully altered image intended to cause misclassification
  2. A properly labeled training image
  3. An ordinary validation split
  4. A larger monitor
Show correct answer and explanation

Correct answer: A. A carefully altered image intended to cause misclassification

Adversarial examples deliberately perturb inputs to induce incorrect predictions.

Question 78. The main role of regularization in machine learning is to:

  1. Discourage overly complex solutions and improve generalization
  2. Ensure training accuracy is exactly 100%
  3. Remove the need for test data
  4. Make features identical
Show correct answer and explanation

Correct answer: A. Discourage overly complex solutions and improve generalization

Regularization penalizes or constrains model complexity to reduce overfitting risk.

Question 79. In L2 regularization, the penalty is based on:

  1. The sum of class labels
  2. Only prediction count
  3. The sum of squared parameter values
  4. A graph’s vertex degrees
Show correct answer and explanation

Correct answer: C. The sum of squared parameter values

L2 regularization typically adds a coefficient times the sum of squared weights to the loss.

Question 80. Dropout in neural-network training randomly:

  1. Deletes all test examples
  2. Deactivates selected units during training
  3. Changes the target labels permanently
  4. Replaces optimization with search
Show correct answer and explanation

Correct answer: B. Deactivates selected units during training

Dropout masks selected activations during training and can reduce reliance on particular units.

Question 81. Which metric uses the harmonic mean of precision and recall?

  1. Accuracy
  2. F1 score
  3. Mean squared error
  4. Entropy
Show correct answer and explanation

Correct answer: B. F1 score

F1 equals 2PR/(P+R) when the denominator is nonzero.

Question 82. A receiver operating characteristic (ROC) curve plots:

  1. Training loss against epoch only
  2. Precision against model size
  3. Entropy against memory
  4. True-positive rate against false-positive rate
Show correct answer and explanation

Correct answer: D. True-positive rate against false-positive rate

The ROC curve varies the decision threshold and plots sensitivity versus false-positive rate.

Question 83. What is class imbalance in a classification dataset?

  1. Some classes have many more examples than others
  2. Each class has exactly equal examples
  3. The inputs have different units
  4. All predictions are correct
Show correct answer and explanation

Correct answer: A. Some classes have many more examples than others

Class imbalance occurs when class frequencies differ substantially.

Question 84. Which method is commonly used to partition data into multiple evaluation folds?

  1. Alpha-beta pruning
  2. Tokenization
  3. K-fold cross-validation
  4. A* search
Show correct answer and explanation

Correct answer: C. K-fold cross-validation

K-fold cross-validation trains and evaluates repeatedly using different held-out partitions.

Question 85. A feature in machine learning is:

  1. An input attribute used to make predictions
  2. The final accuracy score
  3. The only possible output label
  4. The duration of training
Show correct answer and explanation

Correct answer: A. An input attribute used to make predictions

Features are variables or representations that provide information to a predictive model.

Question 86. Feature scaling can be useful because:

  1. It always makes data unbiased
  2. Features with very different numeric ranges can affect some algorithms disproportionately
  3. It guarantees perfect classification
  4. It eliminates the test set
Show correct answer and explanation

Correct answer: B. Features with very different numeric ranges can affect some algorithms disproportionately

Distance-based and gradient-based methods can be sensitive to differing input scales.

Question 87. In principal component analysis (PCA), the first principal component maximizes:

  1. Classification accuracy for every dataset
  2. Number of labels
  3. Nonlinearity of a neural network
  4. Variance of projected centered data among unit directions
Show correct answer and explanation

Correct answer: D. Variance of projected centered data among unit directions

The first principal component is the direction of greatest sample variance in centered data.

Question 88. Which algorithm divides observations into k clusters by iteratively updating centroids?

  1. Minimax
  2. A*
  3. Naive Bayes
  4. K-means
Show correct answer and explanation

Correct answer: D. K-means

K-means alternates assigning observations to nearest centroids and recomputing cluster means.

Question 89. In an AI system, explainability is concerned with:

  1. Adding unnecessary parameters
  2. Hiding training limitations
  3. Eliminating privacy safeguards
  4. Making model decisions or behavior understandable
Show correct answer and explanation

Correct answer: D. Making model decisions or behavior understandable

Explainability aims to communicate reasons, influential factors or mechanisms behind outputs.

Question 90. Human oversight is especially important for high-stakes AI because:

  1. All AI outputs are guaranteed correct
  2. Model errors can have consequential real-world effects
  3. Human review always removes every risk
  4. Testing is unnecessary
Show correct answer and explanation

Correct answer: B. Model errors can have consequential real-world effects

High-stakes uses need appropriate review, evaluation and avenues to challenge or correct consequential decisions.

Artificial Intelligence MCQs: Set 4 (Questions 91–120)

Study each question, then reveal the correct answer and a short explanation.

Question 91. In classical planning, a goal test determines whether:

  1. All actions have identical costs
  2. A heuristic is always zero
  3. The environment contains no uncertainty
  4. The current state satisfies the desired conditions
Show correct answer and explanation

Correct answer: D. The current state satisfies the desired conditions

A goal test checks whether the current state meets the specified goal conditions.

Question 92. A planning action’s preconditions describe:

  1. Only the cost of searching
  2. The final test accuracy
  3. The number of neurons
  4. Conditions that must hold for the action to be applicable
Show correct answer and explanation

Correct answer: D. Conditions that must hold for the action to be applicable

Preconditions specify what must be true before an action can be executed.

Question 93. In a state-space search problem, a successor function:

  1. Measures only network bandwidth
  2. Always returns the goal state
  3. Generates states reachable through applicable actions
  4. Labels all training data
Show correct answer and explanation

Correct answer: C. Generates states reachable through applicable actions

A successor function specifies possible next states from the current state.

Question 94. A goal-based agent differs from a simple reflex agent mainly because it:

  1. Never perceives its surroundings
  2. Considers desired future states when selecting actions
  3. Has no action choices
  4. Must use supervised learning
Show correct answer and explanation

Correct answer: B. Considers desired future states when selecting actions

Goal-based agents evaluate actions in terms of achieving explicitly represented goals.

Question 95. What is an important limitation of hill climbing?

  1. It always explores every state
  2. It requires a labeled dataset
  3. It guarantees a globally optimal result
  4. It can get trapped at local optima or plateaus
Show correct answer and explanation

Correct answer: D. It can get trapped at local optima or plateaus

Hill climbing uses local improvements and may stop before finding the global optimum.

Question 96. Simulated annealing sometimes accepts a worse move mainly to:

  1. Ensure every move is optimal
  2. Remove randomization entirely
  3. Help escape local optima during search
  4. Convert a graph into a tree
Show correct answer and explanation

Correct answer: C. Help escape local optima during search

Accepting occasional worse moves can help search move beyond local optima.

Question 97. Which property means a search algorithm finds a solution whenever one exists under stated assumptions?

  1. Consistency of a dataset
  2. Precision
  3. Completeness
  4. Classification
Show correct answer and explanation

Correct answer: C. Completeness

Completeness refers to eventually finding a solution when one exists, within the algorithm’s assumptions.

Question 98. What does optimality mean for a search algorithm?

  1. It always uses the least memory
  2. It returns a least-cost solution under the stated cost model
  3. It never explores unnecessary states
  4. It guarantees zero computation time
Show correct answer and explanation

Correct answer: B. It returns a least-cost solution under the stated cost model

An optimal search algorithm finds a solution with minimum total path cost.

Question 99. In constraint satisfaction problems (CSPs), a variable domain is:

  1. Only the solution cost
  2. A list of output images
  3. The set of values the variable may take
  4. The number of search algorithms
Show correct answer and explanation

Correct answer: C. The set of values the variable may take

A CSP has variables, their possible values, and constraints restricting allowable combinations.

Question 100. A solution to a constraint satisfaction problem must:

  1. Assign every variable a value that satisfies all constraints
  2. Assign the same value to all variables
  3. Minimize the number of variables regardless of constraints
  4. Ignore domain restrictions
Show correct answer and explanation

Correct answer: A. Assign every variable a value that satisfies all constraints

A complete valid assignment must satisfy each constraint.

Question 101. The minimum remaining values heuristic selects the variable with:

  1. The fewest currently legal values
  2. The most expensive domain values
  3. The largest numerical identifier
  4. The least important constraint
Show correct answer and explanation

Correct answer: A. The fewest currently legal values

MRV prioritizes the most constrained unassigned variable.

Question 102. Arc consistency in a binary CSP requires that:

  1. All variables have equal domains
  2. Every constraint is an equality
  3. Every value in one variable’s domain has supporting values in each neighboring variable’s domain
  4. No variable can be assigned
Show correct answer and explanation

Correct answer: C. Every value in one variable’s domain has supporting values in each neighboring variable’s domain

For each directed constraint arc, every remaining value must have some compatible value in the neighboring domain.

Question 103. Forward checking after a CSP assignment typically:

  1. Deletes all constraints
  2. Removes inconsistent values from neighboring unassigned variables’ domains
  3. Assigns every remaining variable randomly
  4. Calculates classification recall
Show correct answer and explanation

Correct answer: B. Removes inconsistent values from neighboring unassigned variables’ domains

Forward checking prunes values that conflict with the latest assignment.

Question 104. A Markov decision process (MDP) is commonly specified using:

  1. Only features and class labels
  2. States, actions, transition probabilities, rewards and a decision horizon or discount
  3. Only a sorted list of integers
  4. Only an image and a caption
Show correct answer and explanation

Correct answer: B. States, actions, transition probabilities, rewards and a decision horizon or discount

An MDP models sequential decision making with stochastic transitions and reward feedback.

Question 105. The Markov property means that:

  1. Every action is deterministic
  2. The agent remembers no information
  3. The future state distribution depends on the current state and action rather than the complete past, when the state is sufficient
  4. Rewards must be positive
Show correct answer and explanation

Correct answer: C. The future state distribution depends on the current state and action rather than the complete past, when the state is sufficient

A suitably defined state captures information needed to determine next-state probabilities.

Question 106. In an MDP, a value function commonly represents:

  1. The number of inputs in a dataset
  2. A word’s grammatical category
  3. Expected cumulative reward under a policy
  4. A network’s clock speed
Show correct answer and explanation

Correct answer: C. Expected cumulative reward under a policy

A value function evaluates expected accumulated rewards from a state under a specified policy.

Question 107. Q-learning estimates:

  1. Only the marginal probability of a word
  2. Action values for state–action pairs
  3. The Euclidean norm of every vector
  4. The number of game players
Show correct answer and explanation

Correct answer: B. Action values for state–action pairs

Q-learning learns estimates of expected discounted return for choosing actions in states.

Question 108. In a discounted reward model, a discount factor γ between 0 and 1:

  1. Makes all rewards negative
  2. Changes state labels into classes
  3. Eliminates exploration
  4. Reduces the relative weight of rewards further in the future
Show correct answer and explanation

Correct answer: D. Reduces the relative weight of rewards further in the future

The return weights rewards received t steps later by powers of γ.

Question 109. A partially observable Markov decision process (POMDP) requires an agent to reason about:

  1. Only fully known states
  2. A deterministic sequence with no sensors
  3. Only independent classification labels
  4. Uncertain hidden states using observations or beliefs
Show correct answer and explanation

Correct answer: D. Uncertain hidden states using observations or beliefs

In a POMDP, the true state is not directly observed, so a belief state summarizes uncertainty.

Question 110. In a POMDP, a belief state is:

  1. A guaranteed true proposition
  2. A probability distribution over possible environment states
  3. A list of all actions already executed
  4. The neural-network loss alone
Show correct answer and explanation

Correct answer: B. A probability distribution over possible environment states

Belief states represent uncertainty about the unobserved state.

Question 111. A hidden Markov model is widely used for:

  1. Exactly solving every NP-hard problem
  2. Constructing website navigation menus
  3. Modeling sequences with hidden states and observed emissions
  4. Representing only deterministic functions
Show correct answer and explanation

Correct answer: C. Modeling sequences with hidden states and observed emissions

An HMM models hidden state transitions and probabilistic observations over time.

Question 112. The Viterbi algorithm is commonly used to:

  1. Cluster unlabeled numerical data
  2. Find a most probable hidden-state sequence in a hidden Markov model
  3. Compute a binary classification F1 score
  4. Search a game tree with minimax
Show correct answer and explanation

Correct answer: B. Find a most probable hidden-state sequence in a hidden Markov model

Viterbi uses dynamic programming to find a highest-probability hidden-state path.

Question 113. Which property of naive Bayes simplifies calculating class-conditional likelihoods?

  1. All features have equal values
  2. Conditional independence of features given the class
  3. Every class has probability one
  4. No training data are needed
Show correct answer and explanation

Correct answer: B. Conditional independence of features given the class

Naive Bayes assumes features are conditionally independent given the target class.

Question 114. In a medical test example, sensitivity is another name for:

  1. Precision
  2. Recall or true-positive rate
  3. Specificity
  4. False-positive rate
Show correct answer and explanation

Correct answer: B. Recall or true-positive rate

Sensitivity is TP divided by TP plus FN, provided the denominator is nonzero.

Question 115. Specificity in binary classification is:

  1. TN / (TN + FP)
  2. TP / (TP + FP)
  3. TP / (TP + FN)
  4. FP / (FP + TN)
Show correct answer and explanation

Correct answer: A. TN / (TN + FP)

Specificity measures the proportion of actual negative cases correctly identified.

Question 116. For a numerical regression model, mean squared error averages:

  1. Absolute model weights only
  2. Correct classification labels
  3. Probabilities of all tokens
  4. Squared differences between predictions and actual values
Show correct answer and explanation

Correct answer: D. Squared differences between predictions and actual values

MSE is the average of (prediction minus target) squared over evaluated examples.

Question 117. Which statement about correlation and causation is correct?

  1. Correlation always proves causation
  2. Negative correlation means no relationship
  3. Correlation alone does not establish that one variable causes another
  4. A high correlation guarantees an unbiased study
Show correct answer and explanation

Correct answer: C. Correlation alone does not establish that one variable causes another

Observed association can arise from confounding, reverse causation or other mechanisms.

Question 118. A counterfactual question in causal reasoning asks:

  1. Which item appears first alphabetically
  2. How many records are in a file
  3. What would have happened under a different action or intervention
  4. What is the largest eigenvalue only
Show correct answer and explanation

Correct answer: C. What would have happened under a different action or intervention

Counterfactual reasoning compares an actual outcome with a hypothetical outcome under another intervention.

Question 119. An interpretable decision rule such as ‘if income exceeds a threshold’ is an example of:

  1. A convolutional filter
  2. An unobservable sensor
  3. A human-readable rule-based explanation
  4. A probability distribution by itself
Show correct answer and explanation

Correct answer: C. A human-readable rule-based explanation

Rule-based systems express decisions through explicit conditions and conclusions.

Question 120. Which is a responsible way to deploy an AI system in education?

  1. Assume its answers are always correct
  2. Evaluate accuracy, privacy, bias and provide avenues for human review
  3. Publish student records without consent
  4. Remove all teacher oversight
Show correct answer and explanation

Correct answer: B. Evaluate accuracy, privacy, bias and provide avenues for human review

Educational deployments should test quality and fairness, protect personal information and retain suitable human oversight.

Artificial Intelligence MCQs: Set 5 (Questions 121–150)

Choose an option and reveal the correct answer and its explanation.

Question 121. What is a perceptron?

  1. A method for storing images only
  2. A graph-search stopping condition
  3. A form of database normalization
  4. A simple artificial neuron that computes a weighted input and applies an activation rule
Show correct answer and explanation

Correct answer: D. A simple artificial neuron that computes a weighted input and applies an activation rule

A perceptron combines weighted features and a bias before producing an output using an activation rule.

Question 122. For a linear model y = wx + b, what does b represent?

  1. The learning rate
  2. The input variable
  3. The bias or intercept
  4. The number of training samples
Show correct answer and explanation

Correct answer: C. The bias or intercept

The intercept b shifts the linear function vertically independently of x.

Question 123. Which activation function outputs max(0, x)?

  1. Softmax
  2. Sigmoid
  3. ReLU
  4. Tanh
Show correct answer and explanation

Correct answer: C. ReLU

The rectified linear unit (ReLU) returns zero for negative inputs and x for nonnegative inputs.

Question 124. What does the softmax function produce from a vector of finite real-valued scores?

  1. A single Boolean value
  2. Nonnegative values that sum to one
  3. A sorted list of features
  4. Only negative values
Show correct answer and explanation

Correct answer: B. Nonnegative values that sum to one

Softmax exponentiates scores and normalizes them, producing strictly positive values that sum to one.

Question 125. Which loss function is widely used for multiclass classification?

  1. Mean absolute percentage error only
  2. Cross-entropy loss
  3. Graph path cost
  4. Shannon coding length only
Show correct answer and explanation

Correct answer: B. Cross-entropy loss

Cross-entropy compares predicted class probabilities with target class distributions.

Question 126. An epoch in neural-network training means:

  1. One new class label
  2. One complete pass through the training dataset
  3. A fixed number of network layers
  4. The time required to make one prediction
Show correct answer and explanation

Correct answer: B. One complete pass through the training dataset

An epoch normally refers to processing the entire training set once during training.

Question 127. A mini-batch in stochastic optimization is:

  1. A subset of training observations used for one parameter update
  2. The full test set
  3. A model’s final weights
  4. A set of only misclassified examples
Show correct answer and explanation

Correct answer: A. A subset of training observations used for one parameter update

Mini-batch training estimates gradients from a subset of examples rather than the whole dataset.

Question 128. Which optimizer maintains adaptive per-parameter learning-rate estimates using moving averages?

  1. Breadth-first search
  2. K-means
  3. Adam
  4. Minimax
Show correct answer and explanation

Correct answer: C. Adam

Adam combines estimates of first and second moments of gradients to adapt updates for individual parameters.

Question 129. Vanishing gradients can make training deep networks difficult because:

  1. Model outputs become integers
  2. All training examples disappear
  3. Early layers receive very small gradient signals
  4. Every gradient becomes infinite
Show correct answer and explanation

Correct answer: C. Early layers receive very small gradient signals

Repeated small derivatives can shrink gradients propagated toward earlier layers.

Question 130. What is transfer learning?

  1. Reusing knowledge from a model trained on one task for a related task
  2. Copying test labels into training data
  3. Replacing learned weights with random numbers after deployment
  4. Using only unsupervised training
Show correct answer and explanation

Correct answer: A. Reusing knowledge from a model trained on one task for a related task

Transfer learning adapts pretrained representations or parameters to a new, often related task.

Question 131. Fine-tuning a pretrained model generally involves:

  1. Removing its entire learned representation
  2. Guaranteeing identical predictions on every input
  3. Converting all data into images
  4. Updating some or all pretrained parameters on new task data
Show correct answer and explanation

Correct answer: D. Updating some or all pretrained parameters on new task data

Fine-tuning continues optimization of pretrained parameters using task-specific examples.

Question 132. What is the purpose of an attention mask in a transformer?

  1. To measure fairness
  2. To increase input image brightness
  3. To replace all tokens with zeros
  4. To restrict which token positions can attend to others
Show correct answer and explanation

Correct answer: D. To restrict which token positions can attend to others

Attention masks control permitted token-to-token interactions, such as blocking future tokens in causal generation.

Question 133. In autoregressive text generation, the next token is predicted based on:

  1. Only future unseen tokens
  2. Random hardware temperatures
  3. A fixed response independent of the prompt
  4. The previously available token context
Show correct answer and explanation

Correct answer: D. The previously available token context

Autoregressive models condition each next-token prediction on the prefix that is already available.

Question 134. What does temperature usually control when sampling tokens from a language model?

  1. The processor’s physical temperature
  2. The number of vocabulary words
  3. The text’s reading grade
  4. The sharpness or randomness of the next-token probability distribution
Show correct answer and explanation

Correct answer: D. The sharpness or randomness of the next-token probability distribution

Sampling temperature rescales logits; lower positive values concentrate probability on high-scoring tokens.

Question 135. In top-k sampling, the model chooses the next token from:

  1. All possible character combinations
  2. The k highest-scoring candidate tokens
  3. Exactly k previous user messages
  4. The k longest training documents
Show correct answer and explanation

Correct answer: B. The k highest-scoring candidate tokens

Top-k sampling restricts selection to the k candidates with greatest model scores before renormalizing.

Question 136. What is a typical limitation of a fixed context window?

  1. It prevents any numerical computation
  2. It automatically verifies factual claims
  3. It stores all prior conversations indefinitely
  4. Only a bounded amount of input and generated context can be processed at once
Show correct answer and explanation

Correct answer: D. Only a bounded amount of input and generated context can be processed at once

A finite context window limits the total number of tokens considered by the model in a single pass.

Question 137. Which statement about model training and inference is correct?

  1. Both always require labeled data
  2. Inference always changes all model weights
  3. Training is identical to displaying a webpage
  4. Training updates model parameters; inference uses parameters to produce outputs
Show correct answer and explanation

Correct answer: D. Training updates model parameters; inference uses parameters to produce outputs

Model training adjusts parameters, while ordinary inference generates predictions without modifying them.

Question 138. What does an out-of-distribution input mean?

  1. A row with no identifier
  2. An input different in important ways from the data distribution used to train or assess the model
  3. An example with a correctly predicted label
  4. A file stored outside a folder
Show correct answer and explanation

Correct answer: B. An input different in important ways from the data distribution used to train or assess the model

Distribution shift can make prior performance estimates unreliable on substantially different inputs.

Question 139. Which approach can help detect distribution shift?

  1. Comparing deployed input statistics with reference data
  2. Never monitoring production data
  3. Assuming test data represent all future cases
  4. Automatically trusting every new prediction
Show correct answer and explanation

Correct answer: A. Comparing deployed input statistics with reference data

Monitoring changes in feature or output distributions can signal that deployment data differ from reference data.

Question 140. In model calibration, a well-calibrated classifier should:

  1. Be correct about 80% of the time on predictions assigned roughly 80% confidence
  2. Have exactly equal numbers of classes
  3. Always be correct when confident
  4. Output only probabilities of zero or one
Show correct answer and explanation

Correct answer: A. Be correct about 80% of the time on predictions assigned roughly 80% confidence

Calibration means predicted probabilities correspond approximately to observed frequencies over comparable predictions.

Question 141. What is a false-positive rate?

  1. TP/(TP+FN)
  2. FP/(FP+TN)
  3. TN/(TN+FP)
  4. FP/(FP+TP)
Show correct answer and explanation

Correct answer: B. FP/(FP+TN)

The false-positive rate is the fraction of actual negatives incorrectly predicted as positive.

Question 142. Which evaluation method is most appropriate for time-series forecasting?

  1. Use chronological training and evaluation to avoid looking into the future
  2. Randomly include future labels as training features
  3. Train on all future observations and test on the past only
  4. Skip evaluation after training
Show correct answer and explanation

Correct answer: A. Use chronological training and evaluation to avoid looking into the future

Time-aware splits preserve the ordering of observations and help prevent information from the future leaking into training.

Question 143. What is early stopping in machine learning?

  1. Halting before processing any data
  2. Ending training when a monitored validation criterion stops improving sufficiently
  3. Always stopping after exactly one batch
  4. Removing the validation set
Show correct answer and explanation

Correct answer: B. Ending training when a monitored validation criterion stops improving sufficiently

Early stopping uses a held-out validation metric to decide when continued training may no longer help generalization.

Question 144. An ensemble model combines:

  1. Predictions from multiple models or learners
  2. Only metadata fields
  3. One training example and one label
  4. An input and an output without learning
Show correct answer and explanation

Correct answer: A. Predictions from multiple models or learners

Ensembles aggregate model outputs, often improving predictive robustness or accuracy.

Question 145. In random forests, diversity between trees is commonly encouraged by:

  1. Using the identical split sequence for every tree
  2. Fixing all feature values to zero
  3. Bootstrapped samples and random feature selection
  4. Disabling all randomness
Show correct answer and explanation

Correct answer: C. Bootstrapped samples and random feature selection

Random forests commonly use bootstrap data samples and random subsets of candidate features when building trees.

Question 146. What does bagging stand for?

  1. Bayesian graph generation
  2. Binary gradient grouping
  3. Balanced action guidance
  4. Bootstrap aggregating
Show correct answer and explanation

Correct answer: D. Bootstrap aggregating

Bagging fits learners on bootstrap samples and aggregates their outputs.

Question 147. How does boosting generally build an ensemble?

  1. By discarding all previous models
  2. By training each learner on identical random outputs
  3. By adding learners sequentially to address remaining prediction errors
  4. By removing training targets
Show correct answer and explanation

Correct answer: C. By adding learners sequentially to address remaining prediction errors

Boosting adds weak learners sequentially, often emphasizing residual errors or difficult training examples.

Question 148. What is the objective of anomaly detection?

  1. Classify each word by part of speech
  2. Identify unusual observations that deviate from expected patterns
  3. Calculate all graph shortest paths
  4. Choose a game move using minimax
Show correct answer and explanation

Correct answer: B. Identify unusual observations that deviate from expected patterns

Anomaly detection looks for rare or atypical data points relative to learned normal behavior.

Question 149. Why should AI-generated citations be checked?

  1. Models can produce inaccurate or nonexistent bibliographic details
  2. Every generated reference is always authentic
  3. Citation checking reduces the model’s training set
  4. All journal titles are interchangeable
Show correct answer and explanation

Correct answer: A. Models can produce inaccurate or nonexistent bibliographic details

Generated titles, authors, page numbers or links may be fabricated or mismatched, so they should be verified against reliable records.

Question 150. Which practice best supports reproducible machine-learning research?

  1. Publishing only a claimed accuracy score
  2. Using a different hidden test set for every comparison without tracking it
  3. Avoiding descriptions of methods
  4. Documenting datasets, preprocessing, splits, model settings and evaluation procedures
Show correct answer and explanation

Correct answer: D. Documenting datasets, preprocessing, splits, model settings and evaluation procedures

Clear experimental details allow others to understand and attempt to reproduce the reported results.

Artificial Intelligence MCQs: Set 6 (Questions 151–175)

Try each question before revealing its correct answer and explanation.

Question 151. Which algorithm finds shortest paths from a source in a graph with nonnegative edge weights?

  1. K-means
  2. Minimax
  3. Dijkstra’s algorithm
  4. Backpropagation
Show correct answer and explanation

Correct answer: C. Dijkstra’s algorithm

Dijkstra’s algorithm finds shortest paths for nonnegative edge weights.

Question 152. For an unweighted graph, which search finds a path with the fewest edges?

  1. Depth-first search
  2. Breadth-first search
  3. Hill climbing
  4. Greedy search
Show correct answer and explanation

Correct answer: B. Breadth-first search

Breadth-first search explores nodes by increasing edge distance.

Question 153. What is a heuristic in informed search?

  1. The exact memory used
  2. A random action
  3. A training label
  4. An estimate that guides search
Show correct answer and explanation

Correct answer: D. An estimate that guides search

A heuristic estimates relevant quantities such as remaining path cost.

Question 154. If A* uses h(n)=0 for all nodes, it behaves like:

  1. Uniform-cost search
  2. Depth-first search
  3. Greedy best-first search
  4. Beam search
Show correct answer and explanation

Correct answer: A. Uniform-cost search

With h(n)=0, A* prioritizes path cost g(n) alone.

Question 155. Which statement describes a monotone (consistent) heuristic?

  1. It always increases after every search step
  2. It cannot equal zero
  3. It always overestimates true cost
  4. Its estimated cost satisfies a triangle inequality over transitions
Show correct answer and explanation

Correct answer: D. Its estimated cost satisfies a triangle inequality over transitions

Consistency requires h(n) ≤ c(n,n’) + h(n’) and h(goal)=0.

Question 156. What is the branching factor of a search tree?

  1. The tree’s deepest level
  2. The number of goals
  3. The cost of the root
  4. The number of successors per node, often expressed as a maximum or typical value
Show correct answer and explanation

Correct answer: D. The number of successors per node, often expressed as a maximum or typical value

Branching factor describes how many children nodes typically have.

Question 157. Which formalism is useful for representing ‘all students are learners’?

  1. Only propositional logic
  2. An image matrix
  3. First-order predicate logic
  4. A sorting algorithm
Show correct answer and explanation

Correct answer: C. First-order predicate logic

First-order logic can quantify over objects and use predicates such as Student(x) → Learner(x).

Question 158. The symbol ∀ in predicate logic means:

  1. There exists
  2. For all
  3. Not
  4. Implies
Show correct answer and explanation

Correct answer: B. For all

Universal quantification asserts a formula for every individual in the domain.

Question 159. The symbol ∃ in first-order logic means:

  1. For all
  2. And
  3. There exists
  4. If and only if
Show correct answer and explanation

Correct answer: C. There exists

Existential quantification asserts that at least one domain element satisfies a formula.

Question 160. Which logical equivalence follows from De Morgan’s law?

  1. ¬(P∧Q) ≡ ¬P∨¬Q
  2. ¬(P∧Q) ≡ ¬P∧¬Q
  3. P∨Q ≡ P∧Q
  4. ¬P ≡ P
Show correct answer and explanation

Correct answer: A. ¬(P∧Q) ≡ ¬P∨¬Q

Negating a conjunction produces the disjunction of the negations.

Question 161. A tautology is a proposition that is:

  1. Always false
  2. True only if two variables agree
  3. True under every truth assignment
  4. Neither true nor false
Show correct answer and explanation

Correct answer: C. True under every truth assignment

A tautology evaluates to true for every assignment of its variables.

Question 162. Which technique proves a conclusion by showing the premises and its negation are inconsistent?

  1. Proof by contradiction
  2. Regression
  3. K-means
  4. Feature scaling
Show correct answer and explanation

Correct answer: A. Proof by contradiction

Deriving a contradiction from premises plus the negated conclusion establishes the conclusion.

Question 163. Unification in logic-based AI seeks:

  1. Substitutions making logical expressions match
  2. The largest image
  3. Minimum training epochs
  4. A neural activation
Show correct answer and explanation

Correct answer: A. Substitutions making logical expressions match

Unification finds compatible substitutions for terms.

Question 164. Which is an example of a production rule?

  1. A column of numbers
  2. A vector norm
  3. IF fever AND cough THEN consider respiratory illness
  4. A file checksum
Show correct answer and explanation

Correct answer: C. IF fever AND cough THEN consider respiratory illness

Production rules express condition–conclusion patterns used in rule-based systems.

Question 165. An expert system’s inference engine primarily:

  1. Applies rules to known facts to derive conclusions
  2. Stores physical sensors
  3. Compiles operating systems
  4. Generates random labels
Show correct answer and explanation

Correct answer: A. Applies rules to known facts to derive conclusions

The inference engine uses the knowledge base to perform reasoning.

Question 166. Forward chaining commonly starts from:

  1. A final hypothesis only
  2. An untrained model
  3. Known facts and applies applicable rules
  4. A game tree’s leaves
Show correct answer and explanation

Correct answer: C. Known facts and applies applicable rules

Forward chaining is data-driven reasoning from facts toward inferred conclusions.

Question 167. Backward chaining is generally:

  1. A type of image compression
  2. A gradient optimizer
  3. Goal-driven reasoning that seeks support for a query
  4. A method of clustering
Show correct answer and explanation

Correct answer: C. Goal-driven reasoning that seeks support for a query

Backward chaining works backward from the proposed goal to subgoals or premises.

Question 168. In fuzzy logic, membership grades typically lie in:

  1. All negative integers
  2. Only {0,2}
  3. All complex numbers
  4. The interval from 0 to 1
Show correct answer and explanation

Correct answer: D. The interval from 0 to 1

Fuzzy membership functions assign degrees between zero and one inclusive.

Question 169. What does a membership grade of 0.7 indicate in a fuzzy set?

  1. A 70% probability in every interpretation
  2. A guaranteed truth
  3. A sample size of seven
  4. A membership degree of 0.7 under the chosen fuzzy-set definition
Show correct answer and explanation

Correct answer: D. A membership degree of 0.7 under the chosen fuzzy-set definition

Fuzzy membership describes degree of belonging, which is not automatically a probability.

Question 170. A knowledge graph usually represents information as:

  1. Entities and relations among them
  2. Unordered pixel arrays
  3. Only scalar labels
  4. A single probability
Show correct answer and explanation

Correct answer: A. Entities and relations among them

Knowledge graphs encode connected entities and their relationships.

Question 171. In recommender systems, collaborative filtering commonly uses:

  1. Only a user’s password
  2. No historical interactions
  3. Only graph shortest paths
  4. Patterns of user–item interactions
Show correct answer and explanation

Correct answer: D. Patterns of user–item interactions

Collaborative filtering learns preferences from patterns across users and items.

Question 172. Content-based recommendation primarily relies on:

  1. The number of servers
  2. Attributes of items and user preferences
  3. Only random recommendations
  4. A minimax utility tree
Show correct answer and explanation

Correct answer: B. Attributes of items and user preferences

Content-based recommenders match item features with user preference profiles.

Question 173. A recommender system’s cold-start problem arises when:

  1. The graph has a cycle
  2. New users or items have too little interaction history
  3. The loss reaches zero
  4. All labels are balanced
Show correct answer and explanation

Correct answer: B. New users or items have too little interaction history

Insufficient history makes it difficult to estimate preferences for newcomers.

Question 174. Which term describes a robot estimating its position and building a map simultaneously?

  1. SLAM
  2. Tokenization
  3. Cross-validation
  4. Question answering
Show correct answer and explanation

Correct answer: A. SLAM

Simultaneous localization and mapping estimates pose while constructing a map.

Question 175. What is the main purpose of path planning in robotics?

  1. Translate text
  2. Estimate a classifier’s precision
  3. Find a feasible route from start to goal while respecting constraints
  4. Store training labels
Show correct answer and explanation

Correct answer: C. Find a feasible route from start to goal while respecting constraints

Path planning produces a route through the robot’s operating space.

Artificial Intelligence MCQs: Set 7 (Questions 176–200)

Try each question before revealing its correct answer and explanation.

Question 176. What do computer vision object detectors generally predict?

  1. Only sentence sentiment
  2. Only a single network weight
  3. The probability of a graph cycle
  4. Object classes and locations in an image
Show correct answer and explanation

Correct answer: D. Object classes and locations in an image

Object detectors identify object categories and their spatial locations, often with bounding boxes.

Question 177. Semantic segmentation assigns:

  1. One numerical score to the entire dataset
  2. A category label to individual image pixels
  3. Only filenames to pictures
  4. A single path cost to every object
Show correct answer and explanation

Correct answer: B. A category label to individual image pixels

Semantic segmentation labels image regions at pixel level.

Question 178. What is the difference between image classification and object detection?

  1. They always produce identical outputs
  2. Detection cannot use neural networks
  3. Classification requires GPS
  4. Detection also localizes objects
Show correct answer and explanation

Correct answer: D. Detection also localizes objects

Classification typically predicts image-level class labels, whereas detection predicts object locations.

Question 179. Optical character recognition (OCR) is used to:

  1. Convert text in images into machine-readable text
  2. Estimate discounted rewards
  3. Solve graph coloring directly
  4. Detect model calibration
Show correct answer and explanation

Correct answer: A. Convert text in images into machine-readable text

OCR extracts characters and words from image-based documents.

Question 180. Speech recognition maps:

  1. Spoken audio to a text transcription
  2. Text to a rendered picture
  3. Probabilities to graph edges
  4. A route to coordinates only
Show correct answer and explanation

Correct answer: A. Spoken audio to a text transcription

Automatic speech recognition converts spoken language in audio into text.

Question 181. Text-to-speech synthesis maps:

  1. An image to a bounding box
  2. A reward into a policy
  3. Text into generated speech audio
  4. A map into a graph
Show correct answer and explanation

Correct answer: C. Text into generated speech audio

Text-to-speech systems synthesize spoken audio from written text.

Question 182. A multimodal AI system can process:

  1. Multiple data modalities such as text and images
  2. Only numerical tables
  3. Only text documents
  4. Only graph edges
Show correct answer and explanation

Correct answer: A. Multiple data modalities such as text and images

Multimodal models handle and connect different forms of input or output.

Question 183. In a transformer, positional information is important because:

  1. All tokens are identical
  2. Self-attention alone does not inherently represent token order
  3. Positions replace model weights
  4. Sequences never have order
Show correct answer and explanation

Correct answer: B. Self-attention alone does not inherently represent token order

Position information helps models distinguish token order.

Question 184. What does a token vocabulary specify?

  1. The number of GPUs
  2. All future outputs
  3. Only punctuation marks
  4. The set of token IDs and corresponding units recognized by the tokenizer
Show correct answer and explanation

Correct answer: D. The set of token IDs and corresponding units recognized by the tokenizer

A tokenizer defines mappings between units of text and token identifiers.

Question 185. In retrieval systems, cosine similarity compares:

  1. Two dates
  2. Directions of two vectors via their normalized dot product
  3. Two tree depths
  4. The difference between model sizes
Show correct answer and explanation

Correct answer: B. Directions of two vectors via their normalized dot product

Cosine similarity is the dot product divided by the product of vector norms when vectors are nonzero.

Question 186. Which vector pair is orthogonal in the usual Euclidean inner product?

  1. (1,1) and (1,1)
  2. (1,0) and (0,1)
  3. (1,0) and (2,0)
  4. (1,2) and (2,4)
Show correct answer and explanation

Correct answer: B. (1,0) and (0,1)

The dot product of (1,0) and (0,1) is zero.

Question 187. What is the cosine similarity of identical nonzero vectors?

  1. 0
  2. 1
  3. −1
  4. Undefined for all vectors
Show correct answer and explanation

Correct answer: B. 1

Identical nonzero vectors have zero angle, so cosine similarity is one.

Question 188. A nearest-neighbor vector search is commonly used in:

  1. Hardware voltage control only
  2. Changing a domain name
  3. Semantic document retrieval
  4. Boolean simplification only
Show correct answer and explanation

Correct answer: C. Semantic document retrieval

Vector search can find semantically similar embeddings in document retrieval.

Question 189. A key reason to separate retrieval and generation stages in RAG is to:

  1. Guarantee no hallucinations
  2. Provide relevant external context to support responses
  3. Avoid storing any document
  4. Force every answer to be a single word
Show correct answer and explanation

Correct answer: B. Provide relevant external context to support responses

Retrieval supplies task-relevant evidence that a generator can use, though verification remains necessary.

Question 190. What is prompt injection?

  1. Malicious instructions embedded in untrusted content to redirect an AI system
  2. A statistical test for variance
  3. A way to raise a learning rate
  4. A hardware failure
Show correct answer and explanation

Correct answer: A. Malicious instructions embedded in untrusted content to redirect an AI system

Prompt injection attempts to make a model follow attacker-controlled instructions found in lower-trust material.

Question 191. Which practice helps mitigate prompt injection?

  1. Treat retrieved pages as untrusted data and enforce instruction boundaries
  2. Trust all external text as system instructions
  3. Disable user authorization checks
  4. Allow tool actions without validation
Show correct answer and explanation

Correct answer: A. Treat retrieved pages as untrusted data and enforce instruction boundaries

Separating untrusted content from trusted instructions and checking permissions reduces injection risks.

Question 192. What is model drift?

  1. A mathematical proof that never changes
  2. Degradation or change in model behavior over time due to shifting conditions
  3. A guaranteed improvement with age
  4. Identical predictions across all inputs
Show correct answer and explanation

Correct answer: B. Degradation or change in model behavior over time due to shifting conditions

Changes in populations, features or target relationships can cause model performance to drift.

Question 193. Why is maintaining a versioned dataset valuable?

  1. It ensures a perfect model
  2. It eliminates all legal requirements
  3. It prevents every duplicate record
  4. It improves traceability and reproducibility of experiments
Show correct answer and explanation

Correct answer: D. It improves traceability and reproducibility of experiments

Dataset versioning records which inputs were used for training and evaluation.

Question 194. What is a model card?

  1. An optimizer hyperparameter
  2. A hardware expansion card
  3. A substitute for independent testing
  4. Documentation describing a model’s intended uses, evaluation and limitations
Show correct answer and explanation

Correct answer: D. Documentation describing a model’s intended uses, evaluation and limitations

Model cards communicate model details, performance characteristics and limitations.

Question 195. What is the primary purpose of a privacy-preserving data minimization policy?

  1. Keep all records forever
  2. Collect only data needed for a defined purpose
  3. Publish raw identifiers
  4. Ignore consent
Show correct answer and explanation

Correct answer: B. Collect only data needed for a defined purpose

Data minimization limits unnecessary collection and exposure of personal information.

Question 196. Which statement about fairness metrics is accurate?

  1. One metric guarantees fairness in every context
  2. Different fairness criteria can conflict depending on data and decision setting
  3. Fairness never needs domain expertise
  4. Accuracy automatically ensures fairness
Show correct answer and explanation

Correct answer: B. Different fairness criteria can conflict depending on data and decision setting

Fairness criteria measure different aspects, and some cannot be satisfied simultaneously under certain conditions.

Question 197. What is the purpose of red-teaming an AI application?

  1. To choose a brand color
  2. To probe weaknesses and failure modes using adversarial tests
  3. To sort documents
  4. To remove all evaluation
Show correct answer and explanation

Correct answer: B. To probe weaknesses and failure modes using adversarial tests

Red-teaming systematically challenges systems to uncover vulnerabilities and unsafe behavior.

Question 198. What should researchers report when comparing AI models?

  1. Only the most favorable test run
  2. Evaluation protocol, datasets, metrics and meaningful limitations
  3. Only a final headline metric
  4. Only the largest parameter count
Show correct answer and explanation

Correct answer: B. Evaluation protocol, datasets, metrics and meaningful limitations

Transparent comparisons depend on clear experimental setups and acknowledged limitations.

Question 199. Why can a mathematically correct AI explanation still be inadequate?

  1. It may omit assumptions or fail to justify key logical steps
  2. Correct mathematics never needs assumptions
  3. All valid results require code
  4. Every concise proof is wrong
Show correct answer and explanation

Correct answer: A. It may omit assumptions or fail to justify key logical steps

A reliable mathematical argument must specify applicable assumptions and justify transitions.

Question 200. Which action best supports trustworthy use of AI-generated educational MCQs?

  1. Publish all automatically without review
  2. Use the same answer option every time
  3. Check the question, options, answer and explanation for correctness and ambiguity
  4. Avoid correcting errors
Show correct answer and explanation

Correct answer: C. Check the question, options, answer and explanation for correctness and ambiguity

Reviewing all components helps identify incorrect keys, ambiguous distractors and misleading explanations.

Privacy Policy
Powered by CookieAdmin