Skip to content

Artificial Intelligence — MCQ (Multiple Choice Questions)

Questions 1–30 include a few broad AI-application questions. Questions 31–70 concentrate on the five modules represented by the 87 source PDFs: agents, search, knowledge representation, uncertainty, and neural networks.

Topic 1: Intelligent Agents

1. The Turing test is best described as: a) A benchmark measuring the IQ of a computer program b) A test of whether a machine can exhibit intelligent behavior indistinguishable from a human, judged via text dialogue c) A benchmark for chess-playing programs d) A measure of a robot's reaction time

2. The PEAS descriptor for an agent stands for: a) Performance, Environment, Actuators, Sensors b) Performance measure, Environment, Actuators, Sensors c) Problem, Environment, Actuators, Sensors d) Performance, Evaluation, Actuators, Sensors

3. An environment is called "fully observable" when: a) The agent has actuators for every action b) The agent's sensors give it complete information about the state of the environment c) There is no uncertainty in outcomes d) The environment is deterministic

4. Breadth-first search is guaranteed to find an optimal solution when all step costs are: a) Equal b) Zero c) Positive d) Negative

5. Iterative deepening search (IDS) combines: a) BFS's speed with DFS's low memory b) DFS's speed with BFS's optimality c) DFS's low memory with BFS's completeness and optimality d) A higher branching factor with a lower depth limit

6. Which search strategy is complete on a finite tree but NOT guaranteed complete on an infinite tree? a) Breadth-first search b) Uniform-cost search c) Depth-first search d) Iterative deepening search

7. A search is guaranteed to be optimal and complete if its heuristic h(n) is:* a) Consistent b) Admissible c) Greedy d) Always zero

8. Greedy best-first search selects the next node based on: a) f(n) = g(n) + h(n) b) g(n), the cost so far c) h(n), the heuristic estimate to the goal d) h*(n), the true cost to the goal

9. Hill-climbing search can fail to reach the goal because of: a) Plateaus and ridges only b) Local maxima, plateaus, and ridges c) Reaching the global maximum too quickly d) A branching factor that is too low

Topic 4: Constraint Satisfaction

10. The constraint "A ≠ B" on two variables is: a) A unary constraint b) A binary constraint c) A ternary constraint d) An n-ary constraint

11. Backtracking search for CSPs commonly combines backtracking with: a) Forward checking only b) Arc consistency (such as AC-3) combined with backtracking search c) Min-conflicts search only d) Hill climbing

12. Arc consistency (AC-3) works by: a) Removing variables that are not needed b) Removing values from a variable's domain that cannot be extended to a consistent solution c) Adding new constraints between variables d) Merging two domains into one

Topic 5: Knowledge Representation and Logic

13. Propositional logic can represent facts about: a) Specific objects and their relations, including quantifiers b) Statements that are simply true or false, with no internal structure c) Temporal ordering of events d) Uncertainty and probabilities

14. First-order (predicate) logic adds to propositional logic: a) Truth tables b) Predicates and quantifiers (∀ and ∃) c) Logical connectives d) Probabilities and certainty factors

15. Semantic networks and frames are examples of: a) Logical inference algorithms b) Knowledge representation schemes c) Heuristic search methods d) Machine learning algorithms

Topic 6: Logical Inference

16. Forward chaining proceeds: a) Backward from a goal, using rules b) Forward from known facts, using rules c) Forward from goals, using facts d) Backward from facts, using rules

17. Backward chaining is primarily used for: a) Computing all logical consequences of a set of facts b) Proving a specific goal by working backward from it c) Optimizing a plan d) Clustering similar data

18. Resolution is a refutation-complete inference rule for: a) Propositional and first-order logic b) Probabilistic models c) Decision trees d) Neural networks

Topic 7: Uncertainty and Probability

19. Bayes' theorem lets us compute: a) P(hypothesis | evidence) from P(evidence | hypothesis) and the prior P(hypothesis) b) P(evidence | hypothesis) from P(hypothesis | evidence) c) The variance of the evidence d) Whether two variables are independent

20. A rational agent operating under uncertainty should select the action maximizing: a) Expected utility b) The probability of the single best outcome c) Immediate reward d) The number of possible outcomes

21. A Bayesian network represents a joint probability distribution using: a) A set of conditional probability tables and a directed acyclic graph b) A single flat probability table c) A correlation matrix d) A simple rule list

Topic 8: Machine Learning

22. Supervised learning differs from unsupervised learning in that it learns from: a) Unlabeled data b) Labeled examples (input-output pairs) c) Rewards and punishments d) Discovered clusters

23. A decision tree chooses the best attribute to split on by maximizing: a) Entropy b) Information gain (the reduction in entropy) c) Gini impurity d) Variance

24. In a multi-layer perceptron (MLP) with hidden layers, the learning algorithm is: a) Gradient descent b) Backpropagation (chain rule + gradient descent) c) K-means clustering d) Support vector optimization

Topic 9: Natural Language Processing

25. The task of labeling each word with its part of speech (noun, verb, etc.) is: a) Parsing b) Lexical segmentation c) Part-of-speech tagging d) Named-entity recognition

26. The task of finding names of people, organizations, and locations is: a) Part-of-speech tagging b) Parsing c) Named-entity recognition d) Sentiment analysis

27. Machine translation between languages most directly relies on models that learn: a) Clusters of similar words b) Sequence-to-sequence mappings (alignments) between source and target language c) Decision trees over word counts d) A* search paths

Topic 10: Robotics and Expert Systems

28. The DENDRAL expert system is a classic example in which domain? a) Medical diagnosis b) Chemical structure elucidation c) Chess playing d) Flight route scheduling

29. In an intelligent agent, the "performance measure" is: a) The sensor inputs the agent receives b) A reward/criterion defining good behavior c) The actuator outputs the agent produces d) The full world state

30. A neural network with a layer of hidden units between the input and output layers is a: a) Single-layer perceptron b) A linear classifier c) A multi-layer perceptron (MLP) d) A decision stump

Answers (key only)

Reveal answers after attempting the questions

1-b, 2-b, 3-b, 4-a, 5-c, 6-c, 7-b, 8-c, 9-b, 10-b, 11-b, 12-b, 13-b, 14-b, 15-b, 16-b, 17-b, 18-a, 19-a, 20-a, 21-a, 22-b, 23-b, 24-b, 25-c, 26-c, 27-b, 28-b, 29-b, 30-c.

PDF-aligned extended practice: Questions 31–70

No.QuestionABCD
31A rational agent chooses an action expected to maximize its:Sensor countPerformance measureProgram sizeKnowledge-base rows
32In PEAS, actuators are used to:Perceive the environmentAct on the environmentDefine utility onlyStore beliefs
33A stochastic environment is one where:Outcomes may be uncertainEvery state is visibleNothing changesThere is one agent only
34A model-based reflex agent maintains:Internal stateOnly condition-action rules with no memoryA complete plan alwaysNeural weights only
35A search problem includes initial state, actions, goal test, and:Path costCitation styleActivation functionDatabase schema
36BFS expands the shallowest node and uses a:LIFO stackFIFO queuePriority by h onlyRandom list
37UCS expands the node with smallest:DepthPath cost g(n)Heuristic h(n)Branching factor
38Iterative deepening repeatedly runs:Depth-limited DFSA*Hill climbingBeam search only
39A consistent heuristic satisfies h(n) ≤:g(n) alwaysc(n,n') + h(n')0 onlyGoal depth
40An admissible heuristic never:UnderestimatesOverestimates the true remaining costEquals zeroUses domain knowledge
41Greedy best-first search ignores:h(n)g(n)The goalFrontier nodes
42A* evaluates nodes with:g(n)+h(n)g(n)−h(n)h(n) onlyDepth only
43Simulated annealing can escape a local optimum by:Sometimes accepting worse movesRemoving successorsSetting h=0Using only BFS
44In minimax, MAX assumes MIN will choose the:Best outcome for MAXWorst outcome for MAXRandom outcomeShallowest node
45Alpha–beta pruning changes minimax's:Final optimal valueNumber of evaluated nodes, not the resultUtility functionPlayer order
46A CSP consists of variables, domains, and:ConstraintsHeuristics onlyTruth tables onlyRewards
47MRV chooses the variable with:Fewest legal remaining valuesMost legal valuesLowest path costHighest utility
48Forward checking removes values made inconsistent by:A new assignmentThe final goal onlyBackpropagationA random restart
49Propositional resolution operates on:ClausesFrames onlyProbabilitiesWeights
50Modus ponens derives Q from P and:Q→PP→Q¬PP∨Q
51A semantic network represents knowledge mainly as:Nodes and labeled linksA numeric matrix onlySource codeSearch queue
52In a frame, a slot stores a:Property/valueSearch frontierProbability normalizationGradient only
53A script represents:A stereotyped event sequenceA single propositionA neuronA utility table
54An ontology formally defines:Concepts and relationships in a domainOnly individual factsA search treeTraining epochs
55Inheritance in a semantic net lets a subclass receive:Properties from a superclassSearch costsPriors onlyActivation values
56`P(AB)` means probability of:B given AA given BA and B independent
57Two events are independent if:`P(AB)=P(A)` when defined`P(AB)=0`
58Bayes' denominator P(E) acts as a:Normalizing constantLoss functionSearch costLearning rate
59Naive Bayes assumes features are conditionally independent given the:ClassSearch nodeTime step onlyUtility
60A Bayesian network must be a:Directed acyclic graphComplete undirected graphSearch tree onlyCyclic list
61An HMM includes hidden states and:Observable emissionsNo transitionsOnly deterministic rulesSearch heuristics
62Expected utility combines outcome utilities with their:ProbabilitiesSyntaxDepthsFrame slots
63A perceptron computes a weighted sum plus bias, then applies an:Activation functionOntologyGoal testQuantifier
64A single-layer perceptron cannot solve:ANDORXORLinearly separable data
65The sigmoid output lies between:−1 and 10 and 10 and infinity−infinity and infinity
66ReLU is commonly max(0, z) and helps reduce:Saturation for positive inputsAll overfittingDataset sizeSearch depth
67Gradient descent updates a weight generally in the direction:Of the gradientOpposite the gradientOf maximum lossOf random labels only
68Backpropagation efficiently computes gradients using the:Chain ruleBayes ruleResolution ruleMinimax rule
69One epoch is:One pass through the training dataOne neuron update onlyThe final testOne hidden unit
70Overfitting means strong training performance but weak:Generalization to unseen dataActivationSyntaxSearch completeness

Extended answer key

31-B, 32-B, 33-A, 34-A, 35-A, 36-B, 37-B, 38-A, 39-B, 40-B, 41-B, 42-A, 43-A, 44-B, 45-B, 46-A, 47-A, 48-A, 49-A, 50-B, 51-A, 52-A, 53-A, 54-A, 55-A, 56-B, 57-A, 58-A, 59-A, 60-A, 61-A, 62-A, 63-A, 64-C, 65-B, 66-A, 67-B, 68-A, 69-A, 70-A.

Built from Markdown with VitePress.