Skip to content
CS Exit Ethiopia
All courses

Introduction to Artificial Intelligence

Open the unified study guide and question reading lookup

Computer Science Exit Exam Study Guide · 235 questions

Read the question first. Select the eye beside the reading hint, study the cited module section, then reveal the answer. Each eye controls only its own block. All hints and answers start closed whenever this page opens.

Read the linked study-guide sections before revealing the answer. PDF page references count from the first page of the document.

235 of 235 questions

Question 1

Which one of the following is concerned with the meaning of a sentence in knowledge representation?

  • A. Structure
  • B. Computational aspect
  • C. Semantics
  • D. Syntax

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI24 · Represent knowledge for inference

Study-guide references

Correct answer: C — Semantics

Semantics assigns meaning to sentences; syntax restricts their legal construction (PDF p. 43). The stem asks about meaning rather than representation layout or computational efficiency.

Evidence in the complete study note

Knowledge representation encodes facts, objects, relations, events and rules so a machine can use them. Syntax defines legal expressions; semantics defines their meaning. A knowledge base stores assertions; an inference system derives consequences. Declarative knowledge says what is true; procedural knowledge specifies how to act.AI24

Choice by choice

A · incorrect — Structure concerns arrangement, not the stated meaning rule. See PDF p. 43.

B · incorrect — Computational aspects concern processing rather than the interpretation definition. See PDF p. 43.

C · correct — This is precisely the module definition of assigning meaning. See PDF p. 43.

D · incorrect — Syntax decides legal symbols and sentence construction. See PDF p. 43.

Question 2

Which one of the following agents works by searching and matching the rules defined by percent and stored internal state?

  • A. Model-based reflex agent
  • B. Goal-based agent
  • C. Simple reflex agent
  • D. Utility based agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI06 · Simple and model-based reflex agents

AI07 · Goals, utility and model updates

Study-guide references

Correct answer: A — Model-based reflex agent

Taking “percent” as a typographical error for percept, the stored internal state records percept history and the model specifies world evolution (PDF p. 14). That distinguishes model-based reflex selection from a simple current-percept rule.

Evidence in the complete study note

A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean. It has no persistent world-state memory. This is adequate when the current observation contains everything needed for its rule, but partial observability can make the same percept require different actions.AI06
A model-based reflex agent maintains internal state using the previous state, previous action, current percept, and models of world change and action effects. It then applies rules to that state. Both kinds can use condition–action rules; memory and a world model distinguish them.AI06
A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07

Choice by choice

A · correct — Combines a model and retained internal state. See PDF p. 14, 14.

B · incorrect — Goal-based selection needs a desired goal, absent here. See PDF p. 14, 14.

C · incorrect — Ignores percept history and therefore cannot account for the stated memory. See PDF p. 14, 14.

D · incorrect — Requires a utility measurement, absent here. See PDF p. 14, 14.

Question 3

Assume the entire state space and transition model are known for every task below, despite any hidden current state. Which problem/environment coupling is incorrect?

  • A. Contingency problem -> partially observable, non-deterministic
  • B. Multiple state problem deterministic, partially observable
  • C. Exploration problem -> partially observable, deterministic
  • D. Single state problem-> fully observable, deterministic

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI10 · Single-state, multiple-state and contingency problems

Study-guide references

Correct answer: C — Exploration problem -> partially observable, deterministic

A known deterministic model with incomplete observation can be handled by a belief state representing several possible current states. Unknown state spaces or transition models create the need for exploration; partial observability alone does not. Nondeterministic partial observation can require contingency branches, while full observation with known deterministic transitions supports a single-state formulation. The explicit knowledge assumption therefore makes the exploration pairing the incorrect one.

Evidence in the complete study note

With a known deterministic model and full observation, a single current state can be used. Without observation but with known deterministic transitions, a sensorless or multiple-state problem tracks a belief state: a set of possible states.AI10
Nondeterminism or partial observation may require a contingency plan with branches based on later percepts. Exploration or online search is required when the state space or transition model is unknown; observation alone does not identify this category. These classifications can overlap. A pairing based only on “partially observable, deterministic” cannot distinguish a known sensorless problem from an unknown exploration problem.AI10

Choice by choice

A · incorrect — This is a valid coupling: partial observation and nondeterminism can require contingency branches, so it is not the incorrect pairing.

B · incorrect — This is valid under the known-model assumption: incomplete observation can require a belief state, so it is not selected.

C · correct — This is the incorrect coupling: hidden state with a known deterministic model does not by itself require exploration.

D · incorrect — This is valid: full observation and known deterministic transitions support a single current state.

Question 4

Which one of the following is correct first order logic for the statement "There is someone standing at the door and is tall"?

  • A. ∀x (StandingAt (x; door )
  • B. ∀x ( StandingAt (x; door )^ tall(x))
  • C. ∃x ( StandingAt (x; door )^ tall(x))
  • D. ∃x ( StandingAt (x; door ) -> toll(x))

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI26 · First-order logic and quantifier scope

AI25 · Propositional logic and models

Study-guide references

Correct answer: C — ∃x ( StandingAt (x; door )^ tall(x))

“There is someone” requires ∃x, and “standing ... and ... tall” requires StandingAt(x,door) ∧ tall(x), so C asserts one witness with both properties (PDF pp. 46, 49). An implication does not assert its antecedent and therefore cannot assert that anyone is at the door.

Evidence in the complete study note

Use ∀x(Human(x)→Mortal(x)) for all humans, and ∃x(Student(x)∧Tall(x)) for at least one tall student. These are useful translation patterns, not rules forbidding other connectives. ∀x∃y Likes(x,y) allows a different y for each x; ∃y∀x Likes(x,y) requires one shared y. Negation swaps quantifiers: ¬∀x P(x) ≡ ∃x¬P(x). Only occurrences within the correct quantifier’s scope are bound.AI26
Propositional atoms represent whole true-or-false claims. ¬P negates; P∧Q requires both; P∨Q requires at least one; P→Q is false only when P is true and Q false; P↔Q requires matching truth values. Thus P→Q ≡ ¬P∨Q. De Morgan gives ¬(P∧Q) ≡ ¬P∨¬Q.AI25

Choice by choice

A · incorrect — Incomplete formula with a universal quantifier and no tall predicate. See PDF p. 49, 46.

B · incorrect — Claims everyone has both properties rather than at least one person. See PDF p. 49, 46.

C · correct — Existentially binds the same x in both conjuncts. See PDF p. 49, 46.

D · incorrect — Uses implication instead of conjunction and also spells the predicate toll. See PDF p. 49, 46.

Question 5

Which statements are incorrect? Interpret rational action as maximizing expected performance given available information; do not assume a successful action always exists. Select all that apply.

  • A. A rational agent know the result of its action, but unrealistic
  • B. Rational agent take action but not 100% sure of success
  • C. Rational agent act to achieve utmost success based on information
  • D. An omniscience agent takes actions for 100% sure of success

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI04 · Rationality and performance

Study-guide references

Correct answers (select all): A, D — A rational agent know the result of its action, but unrealistic; An omniscience agent takes actions for 100% sure of success

Rationality concerns choosing the action with the best expected performance given the agent’s information. It does not mean knowing actual outcomes in advance, so the first claim incorrectly attributes omniscience to a rational agent. Even omniscience would not ensure success if all available actions fail. The remaining statements correctly allow uncertainty while requiring the best expected choice under the available evidence.

Evidence in the complete study note

A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04
Omniscience would mean knowing actual action outcomes in advance. Rationality does not require omniscience or guaranteed success. Even complete outcome knowledge cannot create a successful action when every available action fails. Information gathering and exploration may therefore be rational actions.AI04

Choice by choice

A · correct — This is false: rationality does not imply knowing actual action outcomes in advance.

B · incorrect — This is true: a rational action need not guarantee success.

C · incorrect — This is true when utmost success means best expected performance given the available information.

D · correct — This is false: complete outcome knowledge cannot guarantee that a successful action exists.

Question 6

________ finds the shortest path to the goal based on cost.

  • A. Depth first search
  • B. Iterative deepening
  • C. Uniform cost search
  • D. Breath first search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI15 · Uniform-cost search and Dijkstra

AI13 · Breadth-first search

AI16 · Depth limits, iterative deepening and two directions

Study-guide references

Correct answer: C — Uniform cost search

Uniform-cost search explicitly seeks the goal path with lowest cumulative cost and uses a priority queue (PDF p. 21). “Shortest” here means minimum cost, under the usual nonnegative-cost setting; minimum depth alone is not sufficient when edge costs vary.

Evidence in the complete study note

UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15
With finite branching and each step at least ε>0, UCS is complete and optimal. Finite graphs permit nonnegative edges with proper duplicate handling. Dijkstra uses the same minimum-distance principle, often computing all reachable distances; UCS commonly stops at a goal and generates states on demand. Queue initialization is an implementation choice, not a fundamental distinction. Negative edges invalidate the usual finalized-distance argument.AI15
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
DLS runs DFS to depth ℓ and reports cutoff separately from failure. It misses goals deeper than ℓ. Iterative deepening runs DLS with limits 0,1,2,…; for finite b it finds the shallowest solution with O(b^d) time and O(bd) frontier space, optimal for equal step costs. Repeated shallow work is small compared with the deepest level.AI16

Choice by choice

A · incorrect — Follows depth and is explicitly nonoptimal on p. 19. See PDF p. 21, 21.

B · incorrect — Increases depth limits, not cumulative-cost priority. See PDF p. 21, 21.

C · correct — Directly minimizes cumulative path cost. See PDF p. 21, 21.

D · incorrect — Minimizes steps; equivalent to UCS only when all edge costs are equal. See PDF p. 21, 21.

Question 7

The data structure used in the standard implementation of Depth First Search is ?

  • A. Stack
  • B. Array
  • C. Linked List
  • D. Queue

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI14 · Depth-first search

AI13 · Breadth-first search

AI-S02 · Ordered lookup versus state-space traversal

Study-guide references

Correct answer: A — Stack

DFS uses a stack (PDF p. 18): the newest pending state is explored first. Arrays or linked lists could implement a stack, but they do not themselves specify the required access order.

Evidence in the complete study note

DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
Binary search exploits the order of a sorted sequence: a comparison narrows the remaining candidate interval instead of exploring arbitrary successor states. Array bisection therefore differs from breadth-first or depth-first traversal of a general state-space graph. A storage container such as an array does not by itself specify FIFO or LIFO frontier behavior.AI-S02

Choice by choice

A · correct — Explicitly named implementation structure. See PDF p. 18.

B · incorrect — May store a stack but does not specify LIFO behavior. See PDF p. 18.

C · incorrect — May implement a stack but is not the stated abstract structure. See PDF p. 18.

D · incorrect — FIFO queue is used for BFS on p. 20. See PDF p. 18.

Question 8

Among the following, which kind of algorithm is used in the Game tree to make decisions of Win/Lose?

  • A. Heuristic Search Algorithm
  • B. Min/Max Algorithm
  • C. Depth First Search Algorithm
  • D. Greedy Search Algorithm

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI22 · Games and minimax

AI23 · Alpha–beta and game evaluation

Study-guide references

Correct answer: B — Min/Max Algorithm

Minimax recursively backs up utilities, choosing maxima on MAX turns and minima on MIN turns, so it selects moves against an optimal opponent (PDF pp. 33–34). DFS is its traversal mechanism rather than the adversarial decision rule.

Evidence in the complete study note

Minimax assigns terminal utilities, takes minima at MIN nodes and maxima at MAX nodes, then chooses a maximum-valued root action. It gives the best guaranteed outcome against optimal opposition within the modeled tree. For stochastic outcomes add chance nodes weighted by probabilities.AI22
At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23

Choice by choice

A · incorrect — A broad category, not the specified alternating-player rule. See PDF p. 33, 34.

B · correct — Provides the game-tree decision procedure. See PDF p. 33, 34.

C · incorrect — Traverses the tree but alone does not compare opponent utilities. See PDF p. 33, 34.

D · incorrect — Chooses local heuristic promise rather than alternating MAX/MIN decisions. See PDF p. 33, 34.

Question 9

Agents' behavior can be best described by

  • A. Perception sequence
  • B. Agent function is performing
  • C. Sensors and Actuators
  • D. Environment in which agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

Study-guide references

Correct answer: B — Agent function is performing

The module evaluates behavior and performance through the agent function, a mapping from perception history to action (PDF p. 11). B is grammatically incomplete but identifies that function.

Evidence in the complete study note

A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03
An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03

Choice by choice

A · incorrect — Input history alone does not specify the selected actions. See PDF p. 11.

B · correct — Names the function that describes action behavior. See PDF p. 11.

C · incorrect — Describe input/output mechanisms, not the complete action mapping. See PDF p. 11.

D · incorrect — Provides context, not the mapping. See PDF p. 11.

Question 10

Which type of agent deals with happy and unhappy states?

  • A. Simple reflex agent
  • B. Model based agent
  • C. Utility based agent
  • D. Learning agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI07 · Goals, utility and model updates

AI06 · Simple and model-based reflex agents

AI08 · Learning-agent components

Study-guide references

Correct answer: C — Utility based agent

Happy/unhappy is an informal description of desirability. The utility function assigns each state a real number measuring how well actions achieve goals (PDF p. 15). The module does not literally use those emotional terms.

Evidence in the complete study note

A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07
A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean. It has no persistent world-state memory. This is adequate when the current observation contains everything needed for its rule, but partial observability can make the same percept require different actions.AI06
The performance element selects actions. The learning element changes the agent to improve future performance. The critic evaluates behavior against a performance standard and supplies feedback. The problem generator suggests exploratory actions likely to yield useful experience.AI08

Choice by choice

A · incorrect — Selects current-percept rules, not ranked desirability. See PDF p. 15.

B · incorrect — Tracks world state, without requiring desirability scores. See PDF p. 15.

C · correct — Uses the state preference measure described in the stem. See PDF p. 15.

D · incorrect — Improves through experience rather than defining state desirability. See PDF p. 15.

Question 11

What is the purpose of "Agent" in Artificial Intelligence?

  • A. Mapping of goal sequence to an action
  • B. Work without the direct interference of the people
  • C. Mapping of environment sequence to an action
  • D. Mapping of precept sequence to an action

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

AI04 · Rationality and performance

Study-guide references

Correct answer: D — Mapping of precept sequence to an action

The module defines Agent Function as a map from the percept sequence to an action (PDF p. 10). “Precept” is a spelling error for percept. D matches the definition in the current option order.

Evidence in the complete study note

A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03
A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04

Choice by choice

A · incorrect — Goals are not the input sequence in the definition. See PDF p. 10.

B · incorrect — Describes autonomy, not the function sought. See PDF p. 10.

C · incorrect — The environment itself is not a percept sequence. See PDF p. 10.

D · correct — Matches the stated percept-sequence-to-action mapping. See PDF p. 10.

Question 12

Here “exploratory learning” means discovering structure in data without target labels, oracle labeling or reward-driven interaction. Which learning regime is meant?

  • A. Supervised learning
  • B. Active learning
  • C. Unsupervised learning
  • D. Reinforcement learning
  • E. None of the Above

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI32 · Unsupervised, semi-supervised and active learning

AI33 · Reinforcement learning

Study-guide references

Correct answer: C — Unsupervised learning

Unsupervised learning seeks patterns such as clusters or associations without target labels. Active learning asks an oracle to label selected cases, supervised learning fits labeled input-target examples, and reinforcement learning improves behavior from reward-driven interaction. Specifying the feedback makes the intended use of exploratory precise. The term alone would have several usages, but the stated unlabeled structure-discovery task identifies a definite learning regime.

Evidence in the complete study note

Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
Semi-supervised learning combines some labels with additional unlabeled data. Active learning chooses informative cases for an oracle to label; it still uses labels. “Exploratory learning” is not an exclusive standard synonym: it can mean unsupervised exploration or active data acquisition, so define the feedback. A task named summarization or time-series analysis is not automatically supervised; its training targets determine that.AI32
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33

Choice by choice

A · incorrect — Supervised learning uses labeled targets, excluded by the stem.

B · incorrect — Active learning obtains oracle labels, excluded by the stem.

C · correct — Unsupervised learning discovers structure without target labels.

D · incorrect — Reinforcement learning uses reward-driven interaction, excluded here.

E · incorrect — The specified regime is present among the choices.

Question 13

Let F(x,y,t) mean that person x can fool person y at time t. Over nonempty person and time domains, which statement is equivalent to ∀x∃y∃t ¬F(x,y,t)?

  • A. Everyone can fool some person at some time
  • B. Every person has someone whom they cannot fool at any time.
  • C. No person can fool any person at any time.
  • D. No one can fool everyone all the time

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI26 · First-order logic and quantifier scope

AI44 · Interpret quantified statements and search assumptions

Study-guide references

Correct answer: D — No one can fool everyone all the time

For each person x, the formula requires a person y and a time t for which fooling fails. Moving negation across the quantifiers gives ¬∃x∀y∀t F(x,y,t): nobody can fool everyone all the time. The witness and time can depend on x. Requiring failure at every time, or forbidding all fooling events, would be stronger claims.

Evidence in the complete study note

Use ∀x(Human(x)→Mortal(x)) for all humans, and ∃x(Student(x)∧Tall(x)) for at least one tall student. These are useful translation patterns, not rules forbidding other connectives. ∀x∃y Likes(x,y) allows a different y for each x; ∃y∀x Likes(x,y) requires one shared y. Negation swaps quantifiers: ¬∀x P(x) ≡ ∃x¬P(x). Only occurrences within the correct quantifier’s scope are bound.AI26
The formula ∀x∃y∃t¬F(x,y,t) means that for each person there is a person and time at which fooling fails. By quantifier negation this is ¬∃x∀y∀t F(x,y,t), so nobody fools everybody all the time. Forbidding every successful instance is a much stronger claim than this formula.AI44

Choice by choice

A · incorrect — This asserts a successful fooling event for each person, rather than an unsuccessful event.

B · incorrect — This requires one witness person who cannot be fooled at every time, a stronger time scope than the formula requires.

C · incorrect — This forbids all fooling events, much stronger than requiring one unsuccessful event for each person.

D · correct — Negating someone fooling everyone at every time yields, for each person, someone and some time when fooling fails.

Question 14

A language system must decide which interpretation is intended when a sentence permits more than one meaning or grammatical attachment. Which NLP challenge does this describe?

  • A. Handling Tokenization
  • B. Handling Ambiguity of Sentences
  • C. Handling POS-Tagging
  • D. Handling Semantics of Sentences
  • E. All of the mentioned

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI37 · Language processing tasks

Study-guide references

Correct answer: B — Handling Ambiguity of Sentences

The difficulty described is ambiguity: a word or sentence can permit multiple meanings or structural attachments. Tokenization creates units, POS tagging assigns grammatical roles, and semantic analysis concerns meaning. These tasks can help resolve ambiguity, but they are not synonymous with the particular obstacle. The question therefore asks for the named challenge rather than ranking all NLP difficulties.

Evidence in the complete study note

Tokenization divides text into processing units; tokens may be words, punctuation or subwords. POS tagging assigns grammatical roles; named-entity recognition identifies people, organizations and locations. Parsing identifies grammatical structure; semantics concerns meaning; discourse links sentences; pragmatics uses context and speaker intent. Sentiment analysis predicts evaluative tone. Ambiguity can occur at every level: “bank” has multiple senses, and “I saw the man with a telescope” has multiple attachments.AI37
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37

Choice by choice

A · incorrect — Tokenization decides processing units; it does not name the existence of multiple interpretations.

B · correct — Ambiguity means that more than one interpretation is possible.

C · incorrect — POS tagging assigns grammatical categories rather than naming the competing interpretations.

D · incorrect — Semantics studies meaning generally; the specific obstacle described is multiplicity of interpretations.

E · incorrect — This includes tokenization, tagging and semantic analysis in general, which do not all name the specific multiple-interpretation obstacle.

Question 15

A goal-directed search strategy uses a problem-specific estimate of remaining cost to order current frontier candidates. Which category is it?

  • A. Uninformed search
  • B. Informed search
  • C. Circular traversal without heuristic estimates
  • D. Binary search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI-S02 · Ordered lookup versus state-space traversal

Study-guide references

Correct answer: B — Informed search

Informed search uses problem-specific heuristic information to estimate remaining cost or promise. Merely retaining the current state would not establish that category, because uninformed search also uses states, successors and known costs. The revised stem supplies the missing estimate. Binary search concerns an ordered interval rather than this state-space frontier, and circular traversal does not identify the heuristic-informed category being tested.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
Binary search exploits the order of a sorted sequence: a comparison narrows the remaining candidate interval instead of exploring arbitrary successor states. Array bisection therefore differs from breadth-first or depth-first traversal of a general state-space graph. A storage container such as an array does not by itself specify FIFO or LIFO frontier behavior.AI-S02

Choice by choice

A · incorrect — Uninformed search lacks the extra remaining-cost heuristic specified here.

B · correct — Informed search uses problem-specific heuristic estimates to guide the frontier.

C · incorrect — Traversal without heuristic estimates does not meet the specified remaining-cost-guided strategy.

D · incorrect — Binary search needs an ordered search interval, not merely goal-directed state exploration.

Question 16

An agent’s actions affect later decisions; identical state-action pairs can have different outcomes; and the world changes while the agent deliberates. Which list describes these three properties?

  • A. Deterministic, fully observable, static
  • B. Fully observable, sequential, deterministic
  • C. Partially observable, static, continuous
  • D. Sequential, non-deterministic, dynamic

Alternative version 1

  • A. Partially observable, non-deterministic, sequential, dynamic
  • B. Deterministic, sequential (non-episodic), dynamic, continuous
  • C. Static, episodic, continuous, partially observable
  • D. Fully observable, discrete, episodic, static

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: D — Sequential, non-deterministic, dynamic

The scenario establishes three independent dimensions. Earlier actions affect later decisions, so it is sequential. Outcomes are not fixed by the state-action pair, so it is non-deterministic. Change during deliberation makes it dynamic. It gives no complete observation or continuity specification. Classifying the stated properties avoids an unsupported universal ranking of how difficult different environments must be.

Evidence in the complete study note

Fully observable means sensors reveal relevant state; partially observable means relevant information is hidden or noisy. Deterministic means state and action determine the next state; stochastic means outcomes involve randomness. Strategic uncertainty about an opponent is distinct from dice-like randomness.AI09
Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · incorrect — Determinism and a static world contradict the uncertain outcomes and ongoing change.

B · incorrect — Sequential fits, but deterministic outcomes contradict the scenario.

C · incorrect — A static environment cannot change while the agent deliberates; this list also omits the stated action dependence.

D · correct — Sequential captures dependent decisions, non-deterministic captures uncertain outcomes, and dynamic captures change during deliberation.

Alternative version 1

Correct answer: A — Partially observable, non-deterministic, sequential, dynamic

Incomplete observations establish partial observability, uncertain transitions establish non-determinism, consequences for later decisions establish sequentiality, and change during deliberation establishes dynamics. The other lists contradict at least one of those conditions. No universal difficulty comparison is needed.

Fully observable means sensors reveal relevant state; partially observable means relevant information is hidden or noisy. Deterministic means state and action determine the next state; stochastic means outcomes involve randomness. Strategic uncertainty about an opponent is distinct from dice-like randomness.AI09

Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

A · correct — All four properties match the stated scenario.

B · incorrect — Determinism contradicts uncertain transition outcomes.

C · incorrect — Static and episodic contradict ongoing change and consequences for later decisions.

D · incorrect — Full observability, episodicity and static conditions each contradict the scenario.

Question 17

An agent which mainly operates based on a knowledge about "how things happen in the world" and representation of the current state based on percept history is

  • A. Simple reflex agent
  • B. Goal based agent
  • C. Utility based agent
  • D. Model based agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI06 · Simple and model-based reflex agents

AI07 · Goals, utility and model updates

Study-guide references

Correct answer: D — Model based agent

The wording reproduces the two model-based factors: knowledge of how the world changes and a current-state representation based on percept history (PDF p. 14). No explicit goal or utility comparison is requested.

Evidence in the complete study note

A model-based reflex agent maintains internal state using the previous state, previous action, current percept, and models of world change and action effects. It then applies rules to that state. Both kinds can use condition–action rules; memory and a world model distinguish them.AI06
A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07

Choice by choice

A · incorrect — Ignores history. See PDF p. 14, 14.

B · incorrect — Adds goals, which are not specified. See PDF p. 14, 14.

C · incorrect — Adds a utility measure, which is not specified. See PDF p. 14, 14.

D · correct — Both defining factors match directly. See PDF p. 14, 14.

Question 18

An evaluation technique of classifier accuracy where we randomly partition the data into k mutually exclusive subsets, each approximately equal size and, use Di as test set and others as training set at i-th iteration is ______.

  • A. Random sampling
  • B. Cross-validation
  • C. Stratified cross-validation
  • D. Holdout method

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI35 · Generalization and validation

Study-guide references

Correct answer: B — k-fold cross-validation.

AI35 describes exactly the rotation in the stem: build k models, each trained on all folds except one, evaluate on that held-out fold, then average the scores. The stem does not say class proportions are preserved, so stratification is not established. A single holdout split or generic random sampling does not specify this complete rotation.

Evidence in the complete study note

Train data fits parameters; validation data selects models or settings; a held-out test set estimates final performance. In k-fold validation, train k models, each on k−1 folds and evaluate on the remaining fold, then average scores. Stratification approximately preserves label proportions. Fit preprocessing within each training fold. Time-ordered data needs splits respecting time; repeated tuning on the final test set leaks information.AI35

Choice by choice

A · incorrect — Random sampling alone does not specify the disjoint k-fold rotation of training and test subsets.

B · correct — The described repeated use of each fold as the held-out subset is k-fold cross-validation.

C · incorrect — Stratification additionally preserves label proportions, a condition not supplied by equal fold sizes alone.

D · incorrect — A holdout method uses a fixed split rather than rotating through all k disjoint test folds.

Question 19

An intelligent agent is characterized by

  • A. Learning
  • B. Reasoning
  • C. All of them are answers
  • D. Perception

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI01 · AI capabilities and scope

AI03 · Agents, percepts and action

Study-guide references

Correct answer: C — All of them are answers

Perception, reasoning and learning all appear in the intelligent-behavior list (PDF p. 6). C aggregates those three capabilities; it does not imply every individual agent must implement every technique.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03

Choice by choice

A · incorrect — Listed capability, included in C. See PDF p. 6, 6.

B · incorrect — Listed capability, included in C. See PDF p. 6, 6.

C · correct — Collects the three listed capabilities. See PDF p. 6, 6.

D · incorrect — Listed capability, included in C. See PDF p. 6, 6.

Question 20

An agent type which has condition-action rule and memory to remind the current state of the world and its evolution is called

  • A. Simple reflex
  • B. Model based reflex
  • C. Goal based
  • D. Utility based

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI06 · Simple and model-based reflex agents

Study-guide references

Correct answer: B — Model based reflex

The internal state records percept history; updating it requires how the world evolves and how actions affect it (PDF p. 14). A model-based reflex agent therefore matches the rule-plus-memory description.

Evidence in the complete study note

A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean. It has no persistent world-state memory. This is adequate when the current observation contains everything needed for its rule, but partial observability can make the same percept require different actions.AI06
A model-based reflex agent maintains internal state using the previous state, previous action, current percept, and models of world change and action effects. It then applies rules to that state. Both kinds can use condition–action rules; memory and a world model distinguish them.AI06

Choice by choice

A · incorrect — Uses current percept and ignores the rest of the history. See PDF p. 14, 14.

B · correct — Has the stated model and internal memory. See PDF p. 14, 14.

C · incorrect — Would additionally require goal-directed lookahead. See PDF p. 14, 14.

D · incorrect — Would additionally require utility measurement. See PDF p. 14, 14.

Question 21

Assume that you are employed as a machine learning expert at the Ministry of Education (MoE). Which machine learning technique you would apply for the task of describing the existing higher education institutions as first generation, second generation and third generation.

  • A. Reinforcement Learning
  • B. Regression
  • C. Classification
  • D. Clustering

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

AI33 · Reinforcement learning

Study-guide references

Correct answer: C — Classification

First, second and third generation are supplied category names. Treating the task as assigning institutions to these predefined labels makes it classification, which the module lists under supervised learning (PDF p. 69). If categories were to be discovered without labels, clustering would instead be relevant.

Evidence in the complete study note

Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33

Choice by choice

A · incorrect — Requires action-dependent rewards, absent here. See PDF p. 69, 69.

B · incorrect — A regression output is not the category assignment requested. See PDF p. 69, 69.

C · correct — Matches assignment to the supplied categories. See PDF p. 69, 69.

D · incorrect — Discovers groupings from inputs rather than assigning known labels. See PDF p. 69, 69.

Question 22

Consider the definition of AI: "The effort to design and implement intelligent software that solves problems by identifying best action sequence to achieve the desired goal" can be views AI as

  • A. A system that acts humanly
  • B. A system that thinks humanly
  • C. A system that acts rationally
  • D. A system that thinks rationally

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI02 · Four approaches and the Turing test

Study-guide references

Correct answer: C — A system that acts rationally

The definition emphasizes the best action sequence for a goal. The acting-rationally view evaluates how a system acts and performs (PDF p. 6), rather than imitation of people or the internal laws of thought.

Evidence in the complete study note

Thinking humanly models human mental processes; thinking rationally derives conclusions by valid logic. Acting humanly seeks humanlike behavior; acting rationally selects actions expected to achieve objectives. Thought processes and observed behavior are different evaluation targets.AI02
The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02

Choice by choice

A · incorrect — Would emphasize humanlike external behavior. See PDF p. 6.

B · incorrect — Would model human thought processes. See PDF p. 6.

C · correct — Matches goal-directed best possible action. See PDF p. 6.

D · incorrect — Emphasizes inference and reasoning soundness rather than the action criterion. See PDF p. 6.

Question 23

What is completeness of search algorithms?

  • A. If there is a difference between the quality of goal states, the agent will find the best one
  • B. An estimate of how much information the algorithm needs to store for finding a solution
  • C. An estimate of how many search steps it takes to find a solution
  • D. Provided that a solution exists, the algorithm will find it

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI12 · Frontiers and evaluation criteria

AI13 · Breadth-first search

Study-guide references

Correct answer: D — Provided that a solution exists, the algorithm will find it

Completeness is the guarantee of finding a solution when an appropriately reachable solution exists; the BFS paragraph illustrates it using a finite-depth goal (PDF p. 20). It is distinct from finding the lowest-cost solution and from time or memory requirements.

Evidence in the complete study note

Completeness means finding a solution if one exists under stated assumptions; optimality means returning the best solution. Time counts work, often generated nodes; space counts simultaneous storage. Let b be branching factor, d shallowest goal depth, m maximum search depth, and ℓ a depth limit. Graph-search bounds in vertices V and edges E differ from implicit-tree bounds.AI12
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13

Choice by choice

A · incorrect — Defines optimality, not completeness. See PDF p. 20, 18.

B · incorrect — Describes space complexity. See PDF p. 20, 18.

C · incorrect — Describes time complexity. See PDF p. 20, 18.

D · correct — Matches solution-finding guarantee. See PDF p. 20, 18.

Question 24

Which one of the view of defining AI concerned with thought processing

  • A. Viewing AI as a system which has complete autonomous
  • B. Viewing AI as a system that think either humanly or think rationally
  • C. Viewing AI as a system that act either humanly or think rationally
  • D. Viewing AI as a system that act omnisciently to its environment

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI02 · Four approaches and the Turing test

AI04 · Rationality and performance

Study-guide references

Correct answer: B — Viewing AI as a system that think either humanly or think rationally

Thinking humanly models human cognition; thinking rationally examines ideal inference (PDF p. 5). Both concern thought processing, so B is the complete pairing.

Evidence in the complete study note

Thinking humanly models human mental processes; thinking rationally derives conclusions by valid logic. Acting humanly seeks humanlike behavior; acting rationally selects actions expected to achieve objectives. Thought processes and observed behavior are different evaluation targets.AI02
Omniscience would mean knowing actual action outcomes in advance. Rationality does not require omniscience or guaranteed success. Even complete outcome knowledge cannot create a successful action when every available action fails. Information gathering and exploration may therefore be rational actions.AI04

Choice by choice

A · incorrect — Autonomy alone does not specify thought processing. See PDF p. 5, 5.

B · correct — Includes both thought-oriented approaches. See PDF p. 5, 5.

C · incorrect — Mixes an action view with a thought view. See PDF p. 5, 5.

D · incorrect — Omniscience is not one of the four views. See PDF p. 5, 5.

Question 25

Which one of the following statement is false

  • A. Turing test is a mechanism to validate AI system that act humanly
  • B. Validating AI system that act rationally can be made by its performance
  • C. Validating AI system that thinks humanly is impossible
  • D. Validating AI system that think rationally can be made by using its reasoning logic

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI02 · Four approaches and the Turing test

Study-guide references

Correct answer: C — Validating AI system that thinks humanly is impossible

C makes an absolute impossibility claim. The module describes understanding and modeling human thought as the cognitive-science approach (PDF p. 5), and p. 7 explicitly says computers test theories of the human mind. That contradicts treating validation as inherently impossible.

Evidence in the complete study note

Thinking humanly models human mental processes; thinking rationally derives conclusions by valid logic. Acting humanly seeks humanlike behavior; acting rationally selects actions expected to achieve objectives. Thought processes and observed behavior are different evaluation targets.AI02
The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02

Choice by choice

A · incorrect — Turing’s imitation game tests humanlike behavior. See PDF p. 5, 5, 6.

B · incorrect — Rational-agent view focuses on actions and performance. See PDF p. 5, 5, 6.

C · correct — This is the false assertion requested: page7 explicitly says computers can test theories about how the human mind works, contradicting an absolute impossibility claim.

D · incorrect — This statement accords with the rational-thought view on page5: inference mechanisms, reasoning, soundness and completeness are assessment criteria. It is therefore not the false choice.

Question 26

Which one of the following is evaluating the degrees of success of an agent?

  • A. Performance measure
  • B. Action
  • C. Knowledge
  • D. Perception

Alternative version 1

  • A. Its Perception
  • B. Its Performance measure
  • C. Its action
  • D. Its prior kowled

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI04 · Rationality and performance

AI03 · Agents, percepts and action

Study-guide references

Correct answer: A — Performance measure

The performance measure determines how successful an agent is (PDF p. 10). Actions, knowledge and percepts influence behavior, but evaluating that behavior needs the criterion. In the alternative option order the answer is B.

Evidence in the complete study note

A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04
A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03

Choice by choice

A · correct — Explicit success criterion. See PDF p. 10.

B · incorrect — A selected operation, not its evaluation criterion. See PDF p. 10.

C · incorrect — Supports decisions but does not define success. See PDF p. 10.

D · incorrect — Supplies sensed inputs rather than the measure. See PDF p. 10.

Alternative version 1

Correct answer: B — Its Performance measure

Apply the same study-note rule to this alternative source set; the choice contrasts below show the derivation.

A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04

A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03

A · incorrect — Perception supplies input, not the success criterion (PDF p. 10).

B · correct — The performance measure determines success (PDF p. 10).

C · incorrect — An action is assessed by the criterion (PDF p. 10).

D · incorrect — Prior knowledge helps action selection, but is not the success criterion (PDF p. 12).

Question 27

The component of Expert system which provide interaction between user expert system and Expert system

  • A. Kowledge base
  • B. Inference engine
  • C. User interface
  • D. Human expert

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI40 · Expert systems and choosing methods

AI24 · Represent knowledge for inference

Study-guide references

Correct answer: C — User interface

The user interface explicitly provides interaction between the user and the expert system (PDF p. 64). Knowledge storage and rule application serve different roles.

Evidence in the complete study note

An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40
Knowledge representation encodes facts, objects, relations, events and rules so a machine can use them. Syntax defines legal expressions; semantics defines their meaning. A knowledge base stores assertions; an inference system derives consequences. Declarative knowledge says what is true; procedural knowledge specifies how to act.AI24

Choice by choice

A · incorrect — Stores domain knowledge. See PDF p. 64.

B · incorrect — Applies rules to derive conclusions. See PDF p. 64.

C · correct — Directly provides user-system interaction. See PDF p. 64.

D · incorrect — Provides expertise rather than the interaction component. See PDF p. 64.

Question 28

Which Terminology of Natural language processing study about meaning full components of word

  • A. Phonetics
  • B. Phonology
  • C. Morphology
  • D. Pragmatics

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI38 · Morphology, sounds and prosody

Study-guide references

Correct answer: C — Morphology.

AI38 defines morphology through prefixes, stems and suffixes, such as un-happi-ness and cat plus plural -s. These are the meaningful components of words requested by the stem. Phonetics concerns physical sounds, phonology sound patterns, and pragmatics intended meaning in context; none names the study of internal meaningful word parts.

Evidence in the complete study note

Morphology studies meaningful word parts: un-happi-ness contains a prefix, stem and suffix; cats contains cat and plural -s. It differs from tokenization, which identifies processing boundaries. Phonetics studies physical speech sounds; phonology studies a language’s sound patterns and contrasts.AI38
Prosody concerns stress, rhythm, timing, prominence and intonation across syllables and phrases. A pitch rise can help signal a question; shifting stress can change emphasis. These cues assist dialogue and speech processing. Pragmatics concerns intended meaning in context, not the internal components of words.AI38

Choice by choice

A · incorrect — Phonetics studies the physical properties of speech sounds rather than meaningful word components.

B · incorrect — Phonology studies a language’s sound patterns and contrasts rather than morpheme structure.

C · correct — Morphology analyzes meaningful internal parts such as prefixes, stems and suffixes.

D · incorrect — Pragmatics interprets intended meaning in context, not the internal components of a word.

Question 29

The application area of AI which focus on recognize objects based on patterns in the same way as the human visual system does.

  • A. Expert System
  • B. Speech recogition
  • C. Computer vison
  • D. Robotics

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI39 · Perception, vision and robotics

AI37 · Language processing tasks

Study-guide references

Correct answer: C — Computer vison

Computer vision extracts and interprets image information and explicitly includes object recognition (PDF p. 84). That is the application directly matching visual pattern recognition.

Evidence in the complete study note

Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37

Choice by choice

A · incorrect — Uses domain expertise, not specifically image interpretation. See PDF p. 84, 84.

B · incorrect — Converts spoken language rather than images. See PDF p. 84, 84.

C · correct — Directly addresses visual objects. See PDF p. 84, 84.

D · incorrect — Robotics may use vision but is the broader physical-agent domain. See PDF p. 84, 84.

Question 30

From steps of problem solving which one is concerned on looking various sequence of action which lead to the goal state, evaluate it and chose the sequence of optimal action?

  • A. Execute
  • B. Goal Formulation
  • C. problem formulation
  • D. Search

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI11 · Formulate before searching

Study-guide references

Correct answer: D — Search

The module defines a solution as an action sequence from start to goal and an optimal solution as the lowest-cost such solution (PDF p. 18). Searching is the process of finding/evaluating those sequences, so D matches the activity described.

Evidence in the complete study note

First formulate the goal; then formulate the problem by specifying initial state, legal actions, transition model, goal test, and action costs. Search selects a sequence; execution applies it. A state abstracts away irrelevant world details.AI11
In route finding, states are cities, actions are roads, successors are neighboring cities, and path cost is accumulated distance or time. A solution is a legal path ending at a goal; an optimal solution has minimum cost. Automatic programming or program synthesis searches for a program satisfying a specification; conditional plans and loops require richer representations than a fixed action list.AI11

Choice by choice

A · incorrect — Carries out a selected sequence rather than finding one. See PDF p. 18, 18.

B · incorrect — Specifies a desired outcome rather than searching paths. See PDF p. 18, 18.

C · incorrect — Sets up states/actions rather than choosing a path. See PDF p. 18, 18.

D · correct — Finds the goal-reaching sequence. See PDF p. 18, 18.

Question 31

Which types of rule used in phonology to interpretation of the fluctuation in stress and intonation across the sentences?

  • A. Phonetic rule
  • B. Phonemic rule
  • C. Prosodic rule
  • D. Morpheme rule

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI38 · Morphology, sounds and prosody

Study-guide references

Correct answer: C — Prosodic rules.

AI38 identifies stress and intonation as prosodic phenomena, alongside rhythm, timing and prominence. That directly matches the sentence-level variations in the question. Physical sound production and phonemic contrasts describe other aspects of speech; morphemes are meaningful word parts. The phrase phonology in the stem does not change the specific prosodic level requested.

Evidence in the complete study note

Morphology studies meaningful word parts: un-happi-ness contains a prefix, stem and suffix; cats contains cat and plural -s. It differs from tokenization, which identifies processing boundaries. Phonetics studies physical speech sounds; phonology studies a language’s sound patterns and contrasts.AI38
Prosody concerns stress, rhythm, timing, prominence and intonation across syllables and phrases. A pitch rise can help signal a question; shifting stress can change emphasis. These cues assist dialogue and speech processing. Pragmatics concerns intended meaning in context, not the internal components of words.AI38

Choice by choice

A · incorrect — Phonetic rules concern physical speech sounds, while the stated stress and intonation patterns are specifically prosodic.

B · incorrect — Phonemic contrasts distinguish sound categories; they are not the specific stress-and-intonation rule category requested.

C · correct — Prosodic rules address stress, prominence, rhythm and intonation across syllables and phrases.

D · incorrect — Morphemes are meaningful word parts, not the rule category for sentence intonation.

Question 32

A system is evaluated by whether a judge can distinguish its natural-language conversation from a human’s. Which of the four AI approaches supplies this evaluation criterion?

  • A. Thinking humanly
  • B. Acting Humanly
  • C. Thinking rationally
  • D. Acting rationally

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI02 · Four approaches and the Turing test

Study-guide references

Correct answer: B — Acting Humanly

The criterion is observable humanlike behavior: a judge compares the system’s conversation with a human’s. That is the acting-humanly approach underlying the conversational Turing test. Natural-language ability alone would not establish this classification; the evaluation criterion does. Thinking approaches concern internal processes, while rational approaches judge valid conclusions or effective actions rather than resemblance to humans.

Evidence in the complete study note

Thinking humanly models human mental processes; thinking rationally derives conclusions by valid logic. Acting humanly seeks humanlike behavior; acting rationally selects actions expected to achieve objectives. Thought processes and observed behavior are different evaluation targets.AI02
The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02

Choice by choice

A · incorrect — Thinking humanly concerns internal mental processes, not only observed conversation.

B · correct — Acting humanly evaluates humanlike observable behavior, as in the conversational Turing test.

C · incorrect — Thinking rationally evaluates valid reasoning, not indistinguishability from human dialogue.

D · incorrect — Acting rationally evaluates expected objective achievement, which need not resemble a human.

Question 33

From which components Intelligence is composed of

  • A. Reasoning
  • B. Problem solving
  • C. Linguistic perception
  • D. All

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI01 · AI capabilities and scope

AI37 · Language processing tasks

Study-guide references

Correct answer: D — All

Reasoning, learning, language understanding, problem solving and perception are all listed aspects of intelligent behavior (PDF p. 6). D aggregates the capabilities named by the question.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37

Choice by choice

A · incorrect — One listed capability. See PDF p. 6.

B · incorrect — Another listed capability. See PDF p. 6.

C · incorrect — Interpreted as language understanding/perception, both listed. See PDF p. 6.

D · correct — Includes all three capabilities. See PDF p. 6.

Question 34

An individual who involve in solving problems regarding a certain profession and are considered as knowledge providers in a certain field of study known as ---------------------

  • A. Domain Expert
  • B. Knowledge Engineer
  • C. User
  • D. All

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: A — Domain Expert

The knowledge engineer acquires information from a subject expert and organizes it into rules (PDF p. 63). The professional who supplies specialist problem-solving knowledge is therefore the domain expert.

Evidence in the complete study note

An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · correct — Supplies the subject/domain expertise. See PDF p. 63.

B · incorrect — Elicits and organizes knowledge from the expert. See PDF p. 63.

C · incorrect — Uses the system rather than necessarily providing expert knowledge. See PDF p. 63.

D · incorrect — Does not preserve the distinct roles. See PDF p. 63.

Question 35

Which statement best describes a rational artificial agent?

  • A. An artificial agent that selects actions for the best expected performance given its percept history and knowledge.
  • B. It used only sensors and effectors to be effective
  • C. It is a combination of software and hardware agents
  • D. All

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

AI04 · Rationality and performance

Study-guide references

Correct answer: A — An artificial agent that selects actions for the best expected performance given its percept history and knowledge.

A rational artificial agent senses its environment and chooses actions expected to maximize its performance measure given its information. It need not know actual outcomes in advance. Sensors and effectors alone do not supply an action-selection policy, and combining two vaguely named agent categories is not the architecture-plus-program definition. The choice gives the intended rational-agent criterion without imposing omniscience.

Evidence in the complete study note

An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03
A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03
A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04

Choice by choice

A · correct — This states the rational action-selection criterion without requiring omniscience.

B · incorrect — Sensors and effectors provide interaction, but alone do not establish intelligent action selection.

C · incorrect — Architecture plus program is not a combination of a software agent and a hardware agent.

D · incorrect — The component statements are defective, so their aggregate is not a valid definition.

Question 36

Which option gives the complete set of correct descriptions of robotics?

  • A. It is application area of AI emphasis of creation of intelligent and Efficient robot
  • B. It is a multi disciplinary area of AI
  • C. It focus on Electrical, Mechanical and software part of robot
  • D. All

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI39 · Perception, vision and robotics

AI01 · AI capabilities and scope

Study-guide references

Correct answer: D — All

Robotics integrates mechanical, electrical, control and computing systems to sense, plan and act physically. Building intelligent robots is an AI application, and the required disciplines extend across hardware and software. Thus the three individual descriptions are true. Because this is a single-choice question requesting the complete set, the aggregate is selected and each individual description is an incomplete selection.

Evidence in the complete study note

Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39
Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01

Choice by choice

A · incorrect — Creating intelligent, efficient robots is a valid application description, but this alone omits the other true descriptions.

B · incorrect — Robotics is multidisciplinary, but this alone omits the other true descriptions.

C · incorrect — Electrical, mechanical and software systems are part of robotics, but this alone omits the other true descriptions.

D · correct — All three descriptions are compatible with robotics as an integrated discipline that builds physical sensing and acting systems.

Question 37

Which option describes a conventional computer’s instruction-driven storage and retrieval of information, without making a universal comparison with human intelligence?

  • A. Machine can figure out complete objects even some part is missed
  • B. Machines are Naturally intelligent
  • C. Store and Recall information with set of instruction
  • D. Take precepts from environment with pattern

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI-S01 · Programmed computation and learned patterns

AI34 · Neural networks and deep learning

Study-guide references

Correct answer: C — Store and Recall information with set of instruction

A conventional computer manipulates information according to program instructions, including instructions for storage and retrieval. This describes a mode of computation rather than a universal difference in intelligence between humans and machines. Programs can also implement pattern learning and object recognition. Those capabilities do not change which option names the particular storage-and-retrieval operation requested.

Evidence in the complete study note

A conventional computer manipulates digital information according to a program or sequence of instructions. Programs can implement storage, retrieval and learning operations. Neural networks can learn pattern recognition from examples, so executing instructions and learning patterns are not mutually exclusive descriptions. This provides no universal claim that every machine is better or worse than every human at completing missing information.AI-S01
Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34

Choice by choice

A · incorrect — Recognizing an occluded object is a perception capability, not the definition of instruction-driven storage and retrieval.

B · incorrect — A computer’s programmed operations do not establish natural biological intelligence.

C · correct — Storing and recalling information by executing instructions matches the operation described.

D · incorrect — Processing perceptual patterns is a sensing or recognition task, not specifically storage and retrieval.

Question 38

Which types of searching strategies is not grouped under blind search?

  • A. Breadth first search
  • B. Uniform cost search
  • C. Constraint satisfaction search
  • D. Depth limited search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI21 · Constraint satisfaction

AI17 · Heuristics and greedy best-first search

AI13 · Breadth-first search

AI16 · Depth limits, iterative deepening and two directions

Study-guide references

Correct answer: C — Constraint satisfaction search

BFS, UCS and depth-limited search appear in the uninformed-search sequence (PDF pp. 18–22). Constraint satisfaction is treated separately, so C is the item not grouped under blind search.

Evidence in the complete study note

Backtracking assigns a variable, tries a legal value, and undoes the choice on failure. Choose the variable with fewest remaining values; break ties by degree; try values that rule out fewest neighbor choices. Forward checking removes incompatible neighbor values. Arc consistency removes a value when a binary constraint offers it no support in the other domain. CSP is a problem framework using search and propagation, not a universal fourth uninformed traversal.AI21
A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
DLS runs DFS to depth ℓ and reports cutoff separately from failure. It misses goals deeper than ℓ. Iterative deepening runs DLS with limits 0,1,2,…; for finite b it finds the shallowest solution with O(b^d) time and O(bd) frontier space, optimal for equal step costs. Repeated shallow work is small compared with the deepest level.AI16

Choice by choice

A · incorrect — BFS is in the uninformed sequence. See PDF p. 18, 21, 36.

B · incorrect — UCS is in the uninformed sequence. See PDF p. 18, 21, 36.

C · correct — A separately presented constraint-based problem formulation/search topic. See PDF p. 18, 21, 36.

D · incorrect — Depth-limited search is in the uninformed sequence. See PDF p. 18, 21, 36.

Question 39

Which statement correctly distinguishes propositional logic from first-order logic?

  • A. It evaluates whole propositions using a truth assignment, without quantifying over objects in a domain.
  • B. It is predicate logic
  • C. It assigns truth to open predicates without assigning values to their object variables.
  • D. It contains quantifier

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI26 · First-order logic and quantifier scope

AI25 · Propositional logic and models

Study-guide references

Correct answer: A — It evaluates whole propositions using a truth assignment, without quantifying over objects in a domain.

Propositional logic treats atoms as whole true-or-false claims and evaluates compound formulas from a truth assignment. It does not quantify over individual objects. First-order logic adds predicates, object variables, interpretations and quantifiers. The distinction does not mean that propositional truth needs no assignment: the truth values of its atoms still matter.

Evidence in the complete study note

First-order logic adds objects, predicates, functions and quantified object variables. Human(a) is a property; Likes(a,b) a relation; father(a) a function term. Propositional logic has no object quantifiers, but its truth still depends on the assigned atom values.AI26
Propositional atoms represent whole true-or-false claims. ¬P negates; P∧Q requires both; P∨Q requires at least one; P→Q is false only when P is true and Q false; P↔Q requires matching truth values. Thus P→Q ≡ ¬P∨Q. De Morgan gives ¬(P∧Q) ≡ ¬P∨¬Q.AI25

Choice by choice

A · correct — Propositional evaluation uses atom truth values rather than object-domain quantification.

B · incorrect — Propositional logic is not identical to first-order predicate logic.

C · incorrect — Open predicates need interpretation and variable assignments; this is not a propositional truth rule.

D · incorrect — Object quantifiers belong to first-order logic, not propositional syntax.

Question 40

Which option names the complete set of dataset roles used across model fitting, model selection and final held-out evaluation?

  • A. Validation dataset
  • B. Train dataset
  • C. Test dataset
  • D. All

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI35 · Generalization and validation

Study-guide references

Correct answer: D — All

Training fits parameters, validation selects models or settings, and a held-out test set estimates final performance. The overall workflow uses all three roles, with the test set kept separate from fitting and repeated tuning. Each individual option is true about one role, but only the aggregate gives the complete set requested by this single-choice question.

Evidence in the complete study note

Train data fits parameters; validation data selects models or settings; a held-out test set estimates final performance. In k-fold validation, train k models, each on k−1 folds and evaluate on the remaining fold, then average scores. Stratification approximately preserves label proportions. Fit preprocessing within each training fold. Time-ordered data needs splits respecting time; repeated tuning on the final test set leaks information.AI35

Choice by choice

A · incorrect — Validation supports model selection, but alone does not cover fitting and final held-out evaluation.

B · incorrect — Training fits model parameters, but alone does not cover model selection and final evaluation.

C · incorrect — Held-out test data estimates final performance, but alone omits fitting and model selection.

D · correct — Training, validation and test sets together cover the three distinct workflow roles.

Question 41

algorithm traverses a graph in a Breadth-ward motion and uses a queue to remember to get the next vertex to start a search, when a dead end occurs in any iteration.

  • A. Breadth first search
  • B. Binary search
  • C. Depth first search
  • D. A&C

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI13 · Breadth-first search

AI14 · Depth-first search

AI-S02 · Ordered lookup versus state-space traversal

Study-guide references

Correct answer: A — Breadth first search

Expanding a graph breadthwise with a FIFO queue is precisely BFS (PDF pp. 19–20). DFS follows a path deeply and uses a stack; binary search is not this traversal.

Evidence in the complete study note

BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14
Binary search exploits the order of a sorted sequence: a comparison narrows the remaining candidate interval instead of exploring arbitrary successor states. Array bisection therefore differs from breadth-first or depth-first traversal of a general state-space graph. A storage container such as an array does not by itself specify FIFO or LIFO frontier behavior.AI-S02

Choice by choice

A · correct — Matches breadthwise expansion and FIFO frontier. See PDF p. 19, 20.

B · incorrect — Not the stated graph traversal. See PDF p. 19, 20.

C · incorrect — DFS uses a stack and goes deeply first. See PDF p. 19, 20.

D · incorrect — DFS does not meet the queue/breadthwise description. See PDF p. 19, 20.

Question 42

Which Al approach most directly supports understanding biological intelligence behavior as a scientific goal?

  • A. Weak AI
  • B. Applied AI
  • C. . Cognitive Al
  • D. Strong AI

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI01 · AI capabilities and scope

Study-guide references

Correct answer: C — . Cognitive Al

Cognitive AI tests theories of how minds recognize objects and solve problems; the science-based goal is understanding biological intelligence (PDF pp. 7–8). C links these two descriptions.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
Narrow or weak AI addresses bounded tasks, such as face recognition, spam filtering or a voice assistant. General AI denotes broad competence transferable across domains; strong AI also has a philosophical meaning involving actual understanding. Do not equate a humanoid body, autonomous vehicle, or many features with general intelligence.AI01

Choice by choice

A · incorrect — Simulating cognition does not by itself specify the scientific-theory-testing aim. See PDF p. 7, 8.

B · incorrect — Commercially viable systems are the stated focus of applied AI. See PDF p. 7, 8.

C · correct — Directly tests theories of mind and cognition. See PDF p. 7, 8.

D · incorrect — Aims for genuine human-level reasoning rather than specifically theory testing. See PDF p. 7, 8.

Question 43

Why is Greedy Best-First Search not guaranteed to be optimal or complete?

  • A. It ignores path cost g( n) and relies only on h(n)
  • B. It always chooses the shallowest node
  • C. . It expands nodes in FIFO order
  • D. It requires full knowledge of the goal state in advance

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI18 · Admissibility, consistency and A*

Study-guide references

Correct answer: A — It ignores path cost g( n) and relies only on h(n)

Greedy chooses by estimated remaining cost h(n), ignoring accumulated path cost g(n), so a promising estimate can hide an expensive route. The module also warns of loops and nonoptimality (PDF pp. 25–26). Completeness depends on implementation and state-space assumptions; the general guarantee is absent.

Evidence in the complete study note

Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18

Choice by choice

A · correct — Its h-only choice gives the stated lack of a general guarantee. See PDF p. 25, 26.

B · incorrect — Shallowest-first describes BFS. See PDF p. 25, 26.

C · incorrect — FIFO order describes BFS, not heuristic priority. See PDF p. 25, 26.

D · incorrect — Not the reason the module gives for greedy failures. See PDF p. 25, 26.

Question 44

Which search strategy tries to expand the node that is closest to the goal, on the grounds that this is likely to lead to a solution quickly; Thus, it evaluates nodes by using just the heuristic function: f (n)=h (n).

  • A. Uniform-cost search
  • B. Greedy best-first search
  • C. ¢. Depth-first search
  • D. A* search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI15 · Uniform-cost search and Dijkstra

Study-guide references

Correct answer: B — Greedy best-first search

Selecting the node estimated closest to the goal uses h(n) alone, as the pure-heuristic and lowest-h OPEN-list rules state (PDF p. 25). The same page prints f(n)=g(n), an internal typo contradicted by those rules; visual inspection confirms the typo is in the PDF.

Evidence in the complete study note

Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15

Choice by choice

A · incorrect — Uses g(n), accumulated cost. See PDF p. 25, 25.

B · correct — Matches h-only priority. See PDF p. 25, 25.

C · incorrect — Uses depth priority. See PDF p. 25, 25.

D · incorrect — Uses g(n)+h(n) rather than h alone. See PDF p. 25, 25.

Question 45

Which of the following best defines a heuristic in the context of AI search?

  • A. A function that estimates how close a state is to the goal
  • B. A guaranteed optimal path to the goal
  • C. . A method that always avoids revisiting states
  • D. A brute-force enumeration of all possible states

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: A — A function that estimates how close a state is to the goal

The module says a heuristic estimates how close a state is to the goal (PDF p. 25). An estimate guides node ordering but is neither a path nor a guarantee against revisiting states.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17

Choice by choice

A · correct — Direct definition. See PDF p. 25.

B · incorrect — An estimate does not itself specify a guaranteed optimal path. See PDF p. 25.

C · incorrect — Repeated-state management is a separate search concern. See PDF p. 25.

D · incorrect — Brute force is the uninformed contrast. See PDF p. 25.

Question 46

Which type of heuristic search strategy that evaluates nodes by combining g (n), the cost to reach the node, and h ( n.), the cost to get from the node to the goal: f (n) = g (n) +h ( n) Since g( n) gives the path cost from the start node to node n, and h(n) is the estimated cost of the cheapest path from n to the goal, we have f ( n) = estimated cost of the cheapest solution through n ?

  • A. A* search
  • B. Greedy Best-First Search
  • C. Uniform-cost search
  • D. Depth-first search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI18 · Admissibility, consistency and A*

AI17 · Heuristics and greedy best-first search

AI15 · Uniform-cost search and Dijkstra

AI14 · Depth-first search

Study-guide references

Correct answer: A — A* search

A* uses f(n)=g(n)+h(n), combining the cost already paid with an estimate of the remaining cost. Greedy best-first uses only h; uniform-cost search uses only g; depth-first search follows its stack discipline. The sum is the distinguishing rule. The stray answer annotation attached to the depth-first option has been removed from the practice choices.

Evidence in the complete study note

A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14

Choice by choice

A · correct — A* ranks by accumulated path cost plus estimated remaining cost.

B · incorrect — Greedy best-first ranks by h alone and omits accumulated cost.

C · incorrect — Uniform-cost search ranks by g alone and omits the heuristic estimate.

D · incorrect — DFS follows a stack order rather than minimum g+h.

Question 47

Why is perception still a challenge for AI—even though systems can recognize faces?

  • A. Because perception is unrelated to reasoning
  • B. Because AI lacks actuators
  • C. Because open-world inputs may contain occlusion, noise, unfamiliar conditions and incomplete context.
  • D. Because sensors are too expensive

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI39 · Perception, vision and robotics

AI01 · AI capabilities and scope

Study-guide references

Correct answer: C — Because open-world inputs may contain occlusion, noise, unfamiliar conditions and incomplete context.

Recognizing faces in some settings does not solve all perception problems. Occlusion can hide relevant features, noise can corrupt observations, and unfamiliar conditions or missing context can make an interpretation unreliable. These are difficulties in interpreting sensory information. They do not require every perception task to use natural-language understanding, and they are distinct from actuator availability or sensor price.

Evidence in the complete study note

Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39
Narrow or weak AI addresses bounded tasks, such as face recognition, spam filtering or a voice assistant. General AI denotes broad competence transferable across domains; strong AI also has a philosophical meaning involving actual understanding. Do not equate a humanoid body, autonomous vehicle, or many features with general intelligence.AI01

Choice by choice

A · incorrect — Perception can use context and support reasoning; it is not unrelated to it.

B · incorrect — Actuators implement actions, while the question concerns difficulty interpreting sensory input.

C · correct — Occlusion, noise and unfamiliar or incomplete context can defeat otherwise successful recognition systems.

D · incorrect — Cost is not the conceptual reason that open-world interpretation remains difficult.

Question 48

In which of the following Learning methods an agent receives feedback in the form of rewards or penalties to maximize cumulative reward?

  • A. Supervised learning
  • B. Reinforcement learning
  • C. . Semi-supervised learning
  • D. Unsupervised learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI33 · Reinforcement learning

AI32 · Unsupervised, semi-supervised and active learning

AI31 · Supervised learning and prediction

Study-guide references

Correct answer: B — Reinforcement learning

Reinforcement learning gives rewards for right actions and penalties for wrong ones, with the goal of obtaining the most reward points (PDF p. 69). B matches both interaction and feedback.

Evidence in the complete study note

An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Semi-supervised learning combines some labels with additional unlabeled data. Active learning chooses informative cases for an oracle to label; it still uses labels. “Exploratory learning” is not an exclusive standard synonym: it can mean unsupervised exploration or active data acquisition, so define the feedback. A task named summarization or time-series analysis is not automatically supervised; its training targets determine that.AI32
Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31

Choice by choice

A · incorrect — Supervised learning uses input–target labels rather than action-reward feedback.

B · correct — Reinforcement learning improves decisions to maximize cumulative reward from interaction.

C · incorrect — Semi-supervised learning mixes labeled and unlabeled examples; that mixture does not define cumulative-reward learning.

D · incorrect — Unsupervised learning discovers unlabeled structure rather than defining an action-reward objective.

Question 49

Which one of the following is wrongly matched as AI environment property pairs?

  • A. Fully observable <-> Partially observable
  • B. Discrete <-> Continuous
  • C. Deterministic <-> Stochastic
  • D. Static <-> Stochastic

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: D — Static <-> Stochastic

The module pairs static with dynamic, and deterministic with nondeterministic (PDF p. 13). Thus static versus stochastic mixes two independent dimensions; D is the wrong pairing.

Evidence in the complete study note

Fully observable means sensors reveal relevant state; partially observable means relevant information is hidden or noisy. Deterministic means state and action determine the next state; stochastic means outcomes involve randomness. Strategic uncertainty about an opponent is distinct from dice-like randomness.AI09
Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · incorrect — Opposite observability properties. See PDF p. 13, 13.

B · incorrect — Opposite state/action granularity properties. See PDF p. 13, 13.

C · incorrect — Stochastic is the question’s term for nondeterministic transition uncertainty. See PDF p. 13, 13.

D · correct — Incorrect pair: static’s counterpart is dynamic. See PDF p. 13, 13.

Question 50

Which option describes general data processing without AI learning or inference, rather than naming an AI-related field?

  • A. Machine learning
  • B. Robotics
  • C. Data analysis limited to fixed arithmetic summaries, without learning or inference.
  • D. Natural language processing

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI01 · AI capabilities and scope

AI39 · Perception, vision and robotics

AI37 · Language processing tasks

Study-guide references

Correct answer: C — Data analysis limited to fixed arithmetic summaries, without learning or inference.

Machine learning develops systems that improve from experience, natural language processing interprets or generates language, and robotics combines sensing, planning and physical action. Data analysis is broader: it may use AI, but fixed arithmetic summaries alone need not involve learning or inference. The option restricts that broad category to ordinary non-AI processing, avoiding the incorrect claim that all data analysis lies outside AI.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37

Choice by choice

A · incorrect — Machine learning is part of AI, and AI43 explicitly places learning architectures and regimes within it.

B · incorrect — Robotics combines sensing, planning and physical action and is treated as an AI-related field in AI39.

C · correct — The restricted fixed-summary task need not use AI; this does not exclude learning-based data analysis.

D · incorrect — Natural language processing interprets and generates human language and is covered as an AI field in AI37.

Question 51

Which techniques can improve learned parameters, candidate solutions or policies using training experience, fitness feedback or reward? Select all that apply. Treat D as a fixed knowledge-based expert system.

  • A. Artificial neural networks
  • B. Genetic algorithms
  • C. Reinforcement learning
  • D. Expert systems

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI43 · Clarify overlapping learning and AI labels

AI40 · Expert systems and choosing methods

Study-guide references

Correct answers (select all): A, B, C — Artificial neural networks; Genetic algorithms; Reinforcement learning

Neural-network training adjusts weights from experience, genetic algorithms evolve candidates using fitness and variation, and reinforcement learning improves policies using rewards from interaction. These are different improvement mechanisms that satisfy the broader capability requested. A fixed expert system applies encoded domain knowledge and does not inherently revise itself through experience. The fixed-system assumption distinguishes using knowledge from learning or optimizing with feedback.

Evidence in the complete study note

ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · correct — Training adjusts neural weights from examples or feedback and can improve performance.

B · correct — Genetic search improves a candidate population through fitness-based experience and variation.

C · correct — RL improves policies from action outcomes and rewards.

D · incorrect — A conventional fixed expert system applies encoded knowledge without inherently learning from experience.

Question 52

Which listed programming languages can implement AI or machine-learning algorithms? Select all that apply; this asks about implementation capability, not an exclusive popularity ranking.

  • A. C++
  • B. Java
  • C. Python
  • D. Ruby

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI43 · Clarify overlapping learning and AI labels

AI-S01 · Programmed computation and learned patterns

Additional reading: Rumale — official project README

Study-guide references

Correct answers (select all): A, B, C, D — C++; Java; Python; Ruby

All four languages can implement AI or machine-learning algorithms. The note explicitly identifies Python, C++ and Java as implementation languages. The supplementary primary Rumale project documentation supplies the missing Ruby fact by documenting a Ruby machine-learning library with classifiers and clustering algorithms. This establishes capability rather than a popularity ranking, and avoids treating one widely used language as the only valid implementation choice.

Evidence in the complete study note

Face recognition is both an applied task and a narrow capability. Broad machine-intelligence goals can include emulating human intelligence and building systems that learn. Humanlike problem-solving alone does not uniquely select expert systems. Python is widely used, but C++ and Java also implement AI; “commonly used” does not isolate one language. McCarthy is associated with naming AI, not its sole invention.AI43
AI algorithms are not exclusive to one programming language. Rumale is a Ruby machine-learning library with classification and clustering implementations, demonstrating that Ruby can implement learning algorithms as well as the general-purpose languages already discussed. Language choice and learning regime are different decisions.AI-S01

Additional authoritative reference

Rumale is a Ruby machine-learning library implementing classifiers and clustering methods, demonstrating that Ruby can implement AI/ML algorithms.

Rumale — official project README

Choice by choice

A · correct — The note explicitly identifies C++ as an AI implementation language.

B · correct — The note explicitly identifies Java as an AI implementation language.

C · correct — The note identifies Python as widely used for AI implementation.

D · correct — Rumale’s primary project documentation demonstrates machine-learning algorithms implemented for Ruby.

Question 53

Which listed goal specifically expresses the humanlike approaches to AI, rather than autonomy, general automation or replacement of workers?

  • A. To replace humans with machines
  • B. To develop autonomous systems
  • C. To simulate human intelligence
  • D. To automate tasks and improve efficiency

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI01 · AI capabilities and scope

AI02 · Four approaches and the Turing test

AI43 · Clarify overlapping learning and AI labels

Study-guide references

Correct answer: C — To simulate human intelligence

Humanlike AI approaches aim to model human thought or produce humanlike behavior. Simulating aspects of human intelligence expresses that goal. Autonomous action, automation and efficiency can also be useful AI objectives, but none requires human imitation. Replacing workers is a social or deployment choice rather than the field’s defining aim. This classification does not impose one primary goal on every AI system.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
Thinking humanly models human mental processes; thinking rationally derives conclusions by valid logic. Acting humanly seeks humanlike behavior; acting rationally selects actions expected to achieve objectives. Thought processes and observed behavior are different evaluation targets.AI02
Face recognition is both an applied task and a narrow capability. Broad machine-intelligence goals can include emulating human intelligence and building systems that learn. Humanlike problem-solving alone does not uniquely select expert systems. Python is widely used, but C++ and Java also implement AI; “commonly used” does not isolate one language. McCarthy is associated with naming AI, not its sole invention.AI43

Choice by choice

A · incorrect — Replacing humans is a deployment choice, not the defining goal of humanlike AI.

B · incorrect — Autonomy concerns independent action and need not emulate human intelligence.

C · correct — Simulating human intelligence expresses the humanlike approach to thought or behavior.

D · incorrect — Automation and efficiency can be achieved without modeling human intelligence.

Question 54

Which AI application is used to understand and interpret human language?

  • A. Computer vision
  • B. Speech recognition
  • C. Natural language processing
  • D. Expert systems

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI37 · Language processing tasks

AI39 · Perception, vision and robotics

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: C — Natural language processing

NLP is technology for machines to understand, analyze, manipulate and interpret human language (PDF p. 79). Speech recognition is narrower: converting spoken words into text (p. 82).

Evidence in the complete study note

NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · incorrect — Works with image information. See PDF p. 79.

B · incorrect — Transcribes spoken words; narrower than language interpretation. See PDF p. 79.

C · correct — Matches the language-understanding definition. See PDF p. 79.

D · incorrect — Uses domain expertise, not specifically language interpretation. See PDF p. 79.

Question 55

Which technique is defined by a specialized knowledge base of conventional domain facts and rules, with an inference engine applying them to solve domain problems?

  • A. Genetic algorithms
  • B. Fuzzy logic
  • C. Expert systems
  • D. Reinforcement learning

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI40 · Expert systems and choosing methods

AI30 · Fuzzy logic and vague categories

Study-guide references

Correct answer: C — Expert systems

An expert system is characterized by specialized domain knowledge and an inference engine that reasons over its knowledge base and working facts. Merely saying that a method uses rules would not distinguish it from fuzzy systems, which may also use rules. Genetic algorithms instead evolve populations, and reinforcement learning uses reward feedback. The architecture in the question supplies the specific distinction.

Evidence in the complete study note

An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40
A fuzzy controller fuzzifies measurements, applies graded IF–THEN rules, combines outputs and often defuzzifies to a numerical action. Fuzzy control suits linguistic rules about degrees; reinforcement learning suits reward-based adaptation. Merely saying an environment is complex or uncertain does not uniquely select either method.AI30

Choice by choice

A · incorrect — Genetic algorithms evolve candidate populations through selection and variation rather than defining a knowledge-base inference architecture.

B · incorrect — Fuzzy logic represents graded membership and can use rules, but rules alone do not make it the specialized architecture described.

C · correct — Expert systems apply domain knowledge through an inference engine operating on a knowledge base and working facts.

D · incorrect — Reinforcement learning improves decisions from rewards rather than defining the described knowledge-base architecture.

Question 56

Which learning regime discovers structure in unlabeled data without action-reward interaction? Select the regime rather than a neural architecture family.

  • A. Supervised learning
  • B. Unsupervised learning
  • C. Reinforcement learning
  • D. Deep learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI32 · Unsupervised, semi-supervised and active learning

AI33 · Reinforcement learning

AI43 · Clarify overlapping learning and AI labels

Study-guide references

Correct answer: B — Unsupervised learning

Unsupervised learning seeks clusters, associations or other structure without target labels. Reinforcement learning also need not receive correct-action labels, but it learns through actions and rewards, which this scenario excludes. Deep learning describes an architecture family that can use several learning regimes. Supervised learning requires labeled targets. Naming both the feedback and the requested category therefore isolates the intended regime.

Evidence in the complete study note

Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43

Choice by choice

A · incorrect — Supervised learning requires labeled input-target examples.

B · correct — Unsupervised learning discovers structure without target labels.

C · incorrect — RL learns from action rewards, excluded by the stem.

D · incorrect — Deep learning is an architecture family, not the specific unlabeled structure-discovery regime requested.

Question 57

What does the Turing test assess?

  • A. Machine learning capabilities
  • B. Expert system knowledge
  • C. Natural language processing proficiency
  • D. Artificial intelligence in simulating human conversation

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI02 · Four approaches and the Turing test

AI37 · Language processing tasks

Study-guide references

Correct answer: D — Artificial intelligence in simulating human conversation

Turing’s test asks whether an interrogator can distinguish a computer from a human using textual questions (PDF p. 5). D captures the overall conversational behavioral test; NLP proficiency is a component rather than the full criterion.

Evidence in the complete study note

The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37

Choice by choice

A · incorrect — Not a test specifically of learning. See PDF p. 5.

B · incorrect — Not restricted to expert-system domain knowledge. See PDF p. 5.

C · incorrect — A required conversational ability, not the complete assessment. See PDF p. 5.

D · correct — Matches indistinguishability in textual human conversation. See PDF p. 5.

Question 58

Which AI technique aims to mimic the structure and function of the human brain?

  • A. Genetic algorithms
  • B. Neural networks
  • C. Expert systems
  • D. Reinforcement learning

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI34 · Neural networks and deep learning

AI20 · Local optimization and evolutionary search

AI40 · Expert systems and choosing methods

AI33 · Reinforcement learning

Study-guide references

Correct answer: B — Neural networks

Artificial neural networks are brain-inspired abstractions built from connected processing units with adjustable weights. This is inspiration rather than a literal recreation of the brain. Genetic algorithms use evolutionary population mechanisms, expert systems use encoded domain knowledge, and reinforcement learning defines a feedback regime that can be implemented with many model types, including neural networks.

Evidence in the complete study note

A neuron computes z=Σiwixi+b and output a=φ(z). For x=(2,1), w=(0.5,−1), b=0.2, z=0.2. Linear activation returns z; sigmoid returns 1/(1+e^(−z)); tanh is bipolar; ReLU returns max(0,z). Nonlinearities let multiple layers represent nonlinear relationships.AI34
A genetic algorithm maintains a population, evaluates fitness, selects parents, applies crossover and mutation, and repeats replacement. It searches without guaranteeing global optimality. Means–ends analysis selects a difference between current and goal states, chooses an operator reducing it, and recursively solves its preconditions. Neither method is simply A* running backward.AI20
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34

Choice by choice

A · incorrect — Genetic algorithms are inspired by population selection and genetic variation, not a network of neural processing units.

B · correct — Artificial neural networks are brain-inspired abstractions of interconnected weighted processing units.

C · incorrect — Expert systems encode domain knowledge for inference rather than defining a brain-inspired network structure.

D · incorrect — Reinforcement learning defines reward-based interaction and can use neural or non-neural implementations.

Question 59

Which AI application is used to make predictions based on historical data?

  • A. Data mining
  • B. Pattern recognition
  • C. Predictive analytics
  • D. Genetic algorithms

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

AI20 · Local optimization and evolutionary search

Study-guide references

Correct answer: C — Predictive analytics.

AI31 directly defines predictive analytics as using historical patterns to forecast outcomes. This names the application requested by the question. Data mining and pattern recognition are broader related activities that can contribute, while genetic algorithms perform evolutionary optimization. Their possible supporting roles do not make them the precise application term defined in the note.

Evidence in the complete study note

Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
A genetic algorithm maintains a population, evaluates fitness, selects parents, applies crossover and mutation, and repeats replacement. It searches without guaranteeing global optimality. Means–ends analysis selects a difference between current and goal states, chooses an operator reducing it, and recursively solves its preconditions. Neither method is simply A* running backward.AI20

Choice by choice

A · incorrect — Data mining is a broader discovery activity; the note names historical-outcome forecasting specifically as predictive analytics.

B · incorrect — Recognizing patterns can support prediction, but does not name the defined historical-forecasting application as precisely.

C · correct — Predictive analytics is explicitly defined as forecasting outcomes using historical patterns.

D · incorrect — Genetic algorithms evolve candidate solutions through fitness and variation, rather than naming historical forecasting itself.

Question 60

Which descriptions explicitly establish a bounded task-specific AI capability, rather than only a body shape or an unspecified “general-purpose” label? Select all that apply.

  • A. Autonomous vehicles limited to driving tasks.
  • B. Virtual assistants like Siri or Alexa performing bounded assistant tasks.
  • C. Humanoid robots
  • D. Robots described only as general-purpose, with no capability details.

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI01 · AI capabilities and scope

AI39 · Perception, vision and robotics

Study-guide references

Correct answers (select all): A, B — Autonomous vehicles limited to driving tasks.; Virtual assistants like Siri or Alexa performing bounded assistant tasks.

Narrow AI addresses bounded tasks rather than broad competence transferable across domains. Driving-only automation and a bounded voice assistant therefore establish task-specific capabilities. A humanoid body reveals shape, not intellectual scope, while an unexplained general-purpose label supplies no demonstrated cross-domain competence. The question asks which descriptions establish narrowness, so the two insufficient descriptions are excluded without claiming that their particular robots could never be narrow.

Evidence in the complete study note

Narrow or weak AI addresses bounded tasks, such as face recognition, spam filtering or a voice assistant. General AI denotes broad competence transferable across domains; strong AI also has a philosophical meaning involving actual understanding. Do not equate a humanoid body, autonomous vehicle, or many features with general intelligence.AI01
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39

Choice by choice

A · correct — Autonomous driving is a bounded capability and does not by itself establish broad cross-domain general intelligence.

B · correct — Virtual voice assistants are the note’s direct examples of bounded narrow-AI capability.

C · incorrect — Humanoid shape alone does not establish a bounded task-specific capability.

D · incorrect — An undefined general-purpose label does not establish a specific bounded capability.

Question 61

Which learning regime is defined by an agent choosing actions and improving its policy from environmental rewards or penalties, rather than by a model architecture?

  • A. Reinforcement learning
  • B. Supervised learning
  • C. Unsupervised learning
  • D. Deep learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI33 · Reinforcement learning

AI43 · Clarify overlapping learning and AI labels

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

Study-guide references

Correct answer: A — Reinforcement learning

Reinforcement learning is defined by interaction: the agent acts, receives reward feedback and improves its policy to increase expected cumulative return. Correct-action target labels are not required. Supervised and unsupervised learning use different feedback definitions. Deep learning describes a model architecture and may support reinforcement learning, but it does not name the particular feedback regime asked for.

Evidence in the complete study note

An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43
Supervised learning fits a mapping from labeled input–target examples.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features.AI32

Choice by choice

A · correct — Reinforcement learning uses action–reward interaction to improve expected return.

B · incorrect — Supervised learning uses supplied input–target pairs rather than this reward-feedback definition.

C · incorrect — Unsupervised learning seeks unlabeled structure rather than an action-reward policy.

D · incorrect — Deep learning specifies multilayer neural representations and can implement several feedback regimes.

Question 62

Which AI application is used to identify and classify objects in images?

  • A. Natural language processing
  • B. Speech recognition
  • C. Computer vision
  • D. Expert systems

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI39 · Perception, vision and robotics

AI37 · Language processing tasks

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: C — Computer vision

Computer vision automatically analyzes image information and includes object recognition (PDF p. 84), directly matching identifying/classifying image objects.

Evidence in the complete study note

Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · incorrect — Processes language rather than image objects. See PDF p. 84, 84.

B · incorrect — Transcribes speech. See PDF p. 84, 84.

C · correct — The explicitly visual recognition application. See PDF p. 84, 84.

D · incorrect — Domain expertise alone does not specify vision. See PDF p. 84, 84.

Question 63

Which approach represents imprecise linguistic categories such as warm by graded membership between zero and one, rather than treating membership as strictly true or false?

  • A. Genetic algorithms
  • B. Fuzzy logic
  • C. Neural networks
  • D. Reinforcement learning

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI30 · Fuzzy logic and vague categories

AI20 · Local optimization and evolutionary search

AI33 · Reinforcement learning

AI34 · Neural networks and deep learning

Study-guide references

Correct answer: B — Fuzzy logic

Fuzzy logic represents imprecise categories through graded membership, so an input can belong to a category to a degree rather than only completely or not at all. This differs from uncertainty about whether an event occurs. Genetic search, neural networks and reinforcement learning can be combined with fuzzy methods, but their defining mechanisms do not supply this membership interpretation.

Evidence in the complete study note

Fuzzy sets assign graded membership μ(x) between 0 and 1 to concepts such as warm or tall. A temperature can belong to warm with degree 0.7. This is not a 70% probability that an otherwise crisp warm proposition is true. Boolean logic uses only true and false; probability describes uncertainty about events.AI30
A genetic algorithm maintains a population, evaluates fitness, selects parents, applies crossover and mutation, and repeats replacement. It searches without guaranteeing global optimality. Means–ends analysis selects a difference between current and goal states, chooses an operator reducing it, and recursively solves its preconditions. Neither method is simply A* running backward.AI20
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Training adjusts weights, often through gradients and backpropagation.AI34

Choice by choice

A · incorrect — Genetic algorithms evolve populations; graded category membership is not their defining mechanism.

B · correct — Fuzzy logic represents degrees of membership in vague categories.

C · incorrect — Neural networks adjust weighted representations and do not inherently define fuzzy membership.

D · incorrect — Reinforcement learning uses rewards and policy improvement rather than defining graded category membership.

Question 64

Which AI application is used to generate human-like text or speech?

  • A. Natural language processing
  • B. Speech recognition
  • C. Computer vision
  • D. Expert systems

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI37 · Language processing tasks

AI39 · Perception, vision and robotics

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: A — Natural language processing

NLG is the NLP component for generating language from computerized representations, including text and sentence planning (PDF pp. 79–80). Thus NLP is the offered application matching generation; the source does not provide a full speech-synthesis treatment.

Evidence in the complete study note

NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · correct — Includes the explicit language-generation component. See PDF p. 79, 80.

B · incorrect — Recognizes spoken input rather than generating output. See PDF p. 79, 80.

C · incorrect — Works with image understanding. See PDF p. 79, 80.

D · incorrect — Generates domain conclusions, not necessarily natural-language output. See PDF p. 79, 80.

Question 65

Which technique involves evolving solutions using principles from genetics and natural selection?

  • A. Genetic algorithms
  • B. Neural networks
  • C. Reinforcement learning
  • D. Expert systems

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI20 · Local optimization and evolutionary search

AI34 · Neural networks and deep learning

AI33 · Reinforcement learning

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: A — Genetic algorithms

AI20 describes a genetic algorithm as maintaining a population, evaluating fitness, selecting parents and applying crossover and mutation. Those operations embody the genetics and natural-selection principles in the stem. Neural networks, reward-based learning and rule-based expert systems use different organizing mechanisms, even where hybrid systems combine them with evolutionary search.

Evidence in the complete study note

A genetic algorithm maintains a population, evaluates fitness, selects parents, applies crossover and mutation, and repeats replacement. It searches without guaranteeing global optimality. Means–ends analysis selects a difference between current and goal states, chooses an operator reducing it, and recursively solves its preconditions. Neither method is simply A* running backward.AI20
A neuron computes z=Σiwixi+b and output a=φ(z). For x=(2,1), w=(0.5,−1), b=0.2, z=0.2. Linear activation returns z; sigmoid returns 1/(1+e^(−z)); tanh is bipolar; ReLU returns max(0,z). Nonlinearities let multiple layers represent nonlinear relationships.AI34
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · correct — Genetic algorithms use fitness, parent selection, crossover and mutation to evolve candidate populations.

B · incorrect — Neural networks learn weighted representations; evolutionary population replacement is not required.

C · incorrect — Reinforcement learning improves policies from environmental rewards rather than requiring genetic variation.

D · incorrect — Expert systems reason with specialized encoded knowledge rather than requiring evolutionary populations.

Question 66

Which technique directly represents degrees of linguistic concepts such as warm and applies graded IF–THEN rules to choose an action?

  • A. Fuzzy logic
  • B. Reinforcement learning
  • C. Expert systems
  • D. Genetic algorithms

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI30 · Fuzzy logic and vague categories

AI33 · Reinforcement learning

Study-guide references

Correct answer: A — Fuzzy logic

Fuzzy control represents degrees of membership in concepts such as warm and combines graded rules to produce an action. This handles vague category boundaries rather than requiring probabilistic uncertainty or reward-based adaptation. Reinforcement learning improves policies through rewards, expert systems apply specialized knowledge, and genetic algorithms evolve populations. The added representation and rule requirements distinguish fuzzy control from other methods that may also address complex decisions.

Evidence in the complete study note

Fuzzy sets assign graded membership μ(x) between 0 and 1 to concepts such as warm or tall. A temperature can belong to warm with degree 0.7. This is not a 70% probability that an otherwise crisp warm proposition is true. Boolean logic uses only true and false; probability describes uncertainty about events.AI30
A fuzzy controller fuzzifies measurements, applies graded IF–THEN rules, combines outputs and often defuzzifies to a numerical action. Fuzzy control suits linguistic rules about degrees; reinforcement learning suits reward-based adaptation. Merely saying an environment is complex or uncertain does not uniquely select either method.AI30
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33

Choice by choice

A · correct — Fuzzy logic uses graded membership and fuzzy IF–THEN rules.

B · incorrect — RL uses reward-driven adaptation rather than defining graded linguistic membership.

C · incorrect — An ordinary expert system need not represent degrees of membership in linguistic categories.

D · incorrect — Genetic algorithms evolve candidates through fitness and variation rather than fuzzy rule evaluation.

Question 67

Which AI application is used to simulate human conversation?

  • A. Natural language processing
  • B. Speech recognition
  • C. Computer vision
  • D. Expert systems

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI37 · Language processing tasks

AI02 · Four approaches and the Turing test

Study-guide references

Correct answer: A — Natural language processing

The module lists chatbots as an important NLP application used for chat services (PDF p. 82). That directly supports NLP for simulated conversation.

Evidence in the complete study note

NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02

Choice by choice

A · correct — Includes the chatbot application. See PDF p. 82.

B · incorrect — Transcription alone is not conversation. See PDF p. 82.

C · incorrect — Image comprehension does not define conversation. See PDF p. 82.

D · incorrect — Expert conclusions may be communicated, but chat is specifically listed under NLP. See PDF p. 82.

Question 68

A collection of documents has no target labels. The task is to group similar documents, with no action–reward interaction. Which learning regime describes this objective, regardless of the architecture used?

  • A. Supervised learning
  • B. Unsupervised learning
  • C. Reinforcement learning
  • D. Deep learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI32 · Unsupervised, semi-supervised and active learning

AI43 · Clarify overlapping learning and AI labels

AI31 · Supervised learning and prediction

AI33 · Reinforcement learning

Study-guide references

Correct answer: B — Unsupervised learning

Grouping similar documents without target labels is clustering, an unsupervised learning objective. The data modality does not determine the feedback regime, and a neural architecture could still be used for this task. Supervised learning requires targets; reinforcement learning requires an action-and-reward objective. The question asks for the learning regime rather than which architecture might process documents.

Evidence in the complete study note

Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43
Supervised learning fits a mapping from labeled input–target examples.AI31
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward.AI33

Choice by choice

A · incorrect — Supervised learning requires target labels, absent from the described task.

B · correct — Unsupervised learning discovers structure, including clusters, without supplied targets.

C · incorrect — No actions or reward feedback are present, so this is not the reinforcement objective.

D · incorrect — Deep learning describes a multilayer architecture that might implement clustering, not the learning regime.

Question 69

Which AI technique involves training a model on a large dataset and multiple layers of artificial neurons?

  • A. Genetic algorithms
  • B. Neural networks
  • C. Expert systems
  • D. Reinforcement learning

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI34 · Neural networks and deep learning

AI20 · Local optimization and evolutionary search

AI40 · Expert systems and choosing methods

AI33 · Reinforcement learning

Study-guide references

Correct answer: B — Neural networks

AI34 describes artificial neurons with learned weights and nonlinear outputs, then defines deep learning through multiple representation layers. This matches the large-data, multilayer neural architecture in the stem. Reinforcement learning defines reward-based feedback rather than a specific architecture; genetic algorithms and expert systems do not inherently consist of layered artificial neurons.

Evidence in the complete study note

A neuron computes z=Σiwixi+b and output a=φ(z). For x=(2,1), w=(0.5,−1), b=0.2, z=0.2. Linear activation returns z; sigmoid returns 1/(1+e^(−z)); tanh is bipolar; ReLU returns max(0,z). Nonlinearities let multiple layers represent nonlinear relationships.AI34
A genetic algorithm maintains a population, evaluates fitness, selects parents, applies crossover and mutation, and repeats replacement. It searches without guaranteeing global optimality. Means–ends analysis selects a difference between current and goal states, chooses an operator reducing it, and recursively solves its preconditions. Neither method is simply A* running backward.AI20
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34

Choice by choice

A · incorrect — Genetic algorithms evolve populations and are not defined by layered artificial neurons.

B · correct — A layered collection of artificial neurons is a neural-network architecture whose weights are adjusted during training.

C · incorrect — Expert systems define knowledge-base reasoning rather than layers of trainable neurons.

D · incorrect — Reinforcement learning specifies the reward objective, not a required layered-neuron architecture.

Question 70

What is Artificial Intelligence?

  • A. Artificial Intelligence is a field that aims to make humans more intelligent
  • B. Artificial Intelligence is a field that aims to improve the security
  • C. Artificial Intelligence is a field that aims to develop intelligent machines
  • D. Artificial Intelligence is a field that aims to mine the data

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI01 · AI capabilities and scope

Study-guide references

Correct answer: C — Artificial Intelligence is a field that aims to develop intelligent machines

The module defines AI as designing intelligence in an artificial device (PDF p. 4). Developing intelligent machines is the matching general description, rather than an individual application.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01

Choice by choice

A · incorrect — Making people smarter is not the stated field definition. See PDF p. 4.

B · incorrect — Security is a possible application, not the field definition. See PDF p. 4.

C · correct — Matches intelligence in an artificial device. See PDF p. 4.

D · incorrect — Data mining is one learning application, not the field’s definition. See PDF p. 4.

Question 71

Who is credited with coining the term artificial intelligence in the proposal for the Dartmouth summer research project?

  • A. Geoffrey Hinton
  • B. Andrew Ng
  • C. John McCarthy
  • D. Jürgen Schmidhuber

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI-S04 · Attributing the name of artificial intelligence

AI43 · Clarify overlapping learning and AI labels

Study-guide references

Correct answer: C — John McCarthy

John McCarthy is credited with coining the term in the 1955 proposal for the Dartmouth summer research project. The workshop itself took place in 1956 and helped establish the research field. This naming attribution does not make McCarthy the sole inventor of every AI idea or method. The question now asks for the documented attribution rather than an oversimplified invention claim.

Evidence in the complete study note

John McCarthy is credited with coining the term artificial intelligence in the 1955 proposal for the Dartmouth summer research project. The Dartmouth workshop took place in 1956 and helped establish the field. Naming the field does not mean that one person invented all its ideas or methods.AI-S04
Face recognition is both an applied task and a narrow capability. Broad machine-intelligence goals can include emulating human intelligence and building systems that learn. Humanlike problem-solving alone does not uniquely select expert systems. Python is widely used, but C++ and Java also implement AI; “commonly used” does not isolate one language. McCarthy is associated with naming AI, not its sole invention.AI43

Choice by choice

A · incorrect — The historical naming attribution in the note is to McCarthy, not Hinton.

B · incorrect — The historical naming attribution in the note is to McCarthy, not Ng.

C · correct — McCarthy is credited with introducing the term in the 1955 Dartmouth proposal.

D · incorrect — The historical naming attribution in the note is to McCarthy, not Schmidhuber.

Question 72

Which activity, as stated, directly requires an intelligence-related capability, rather than merely naming a computing operation that may or may not use AI?

  • A. It helps to exploit vulnerabilities to secure the firm
  • B. Language understanding and problem-solving (Text analytics and NLP)
  • C. Easy to create a website
  • D. It helps to deploy applications on the cloud

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI01 · AI capabilities and scope

AI37 · Language processing tasks

Study-guide references

Correct answer: B — Language understanding and problem-solving (Text analytics and NLP)

AI concerns capabilities such as reasoning, learning and language understanding. The language-understanding and problem-solving option explicitly names such capabilities. Website creation, cloud deployment and security operations can incorporate AI, but the activity names alone do not establish its use. The distinction is between an explicit intelligence-related task and a computing context that could host many kinds of software.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37

Choice by choice

A · incorrect — Security testing or exploitation can use ordinary programmed tools; the activity alone does not require intelligent learning or reasoning.

B · correct — Understanding human language and solving problems directly name intelligence-related capabilities.

C · incorrect — Creating a website can use conventional software tools without intelligent behavior.

D · incorrect — Deploying an application to a cloud platform does not itself require an intelligent application.

Question 73

Which of the following is a component of Artificial Intelligence?

  • A. Learning
  • B. Training
  • C. Designing
  • D. Puzzling

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI08 · Learning-agent components

Study-guide references

Correct answer: A — Learning

Learning appears explicitly among intelligent behaviors (PDF p. 6), and is defined as improving performance from experience on p. 67. The other words are not offered as parallel components in the module.

Evidence in the complete study note

The performance element selects actions. The learning element changes the agent to improve future performance. The critic evaluates behavior against a performance standard and supplies feedback. The problem generator suggests exploratory actions likely to yield useful experience.AI08
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08

Choice by choice

A · correct — Explicitly named intelligent capability. See PDF p. 6.

B · incorrect — Training is a model-building activity, not the named component here. See PDF p. 6.

C · incorrect — Designing is not presented as the matching intelligent capability. See PDF p. 6.

D · incorrect — Puzzling is not a defined component. See PDF p. 6.

Question 74

Which of the following is not a type of Artificial Intelligence agent?

  • A. Learning AI agent
  • B. Goal-based AI agent
  • C. Simple reflex AI agent
  • D. Unity-based AI agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI42 · Distinguish agent categories and components

AI06 · Simple and model-based reflex agents

AI08 · Learning-agent components

Study-guide references

Correct answer: D — Unity-based AI agent

The module lists simple reflex, model-based reflex, goal-based, utility-based and learning agents (PDF pp. 14–15). “Unity-based” is not “utility-based” and is not one of those classes.

Evidence in the complete study note

The common teaching list contains five agent classes when learning agents are included, or four decision architectures before learning is added. State the convention when counting agent types. “Target-based” can mean goal-based; “unity-based” is not utility-based.AI42
A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean. It has no persistent world-state memory. This is adequate when the current observation contains everything needed for its rule, but partial observability can make the same percept require different actions.AI06
The performance element selects actions. The learning element changes the agent to improve future performance. The critic evaluates behavior against a performance standard and supplies feedback. The problem generator suggests exploratory actions likely to yield useful experience.AI08

Choice by choice

A · incorrect — Learning is a listed class. See PDF p. 14, 15.

B · incorrect — Goal-based is listed. See PDF p. 14, 15.

C · incorrect — Simple reflex is listed. See PDF p. 14, 15.

D · correct — Unity-based is not a listed class. See PDF p. 14, 15.

Question 75

Which agent architecture applies fixed condition–action rules directly to the current percept, without retaining an internal world-state model?

  • A. Simple reflex agent
  • B. Model-based agent
  • C. Goal-based agent
  • D. Utility-based agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI06 · Simple and model-based reflex agents

AI07 · Goals, utility and model updates

Study-guide references

Correct answer: A — Simple reflex agent

The absence of persistent internal state is the key distinction. A simple reflex agent maps the current percept to an action using condition–action rules. Model-based reflex agents may also use rules, but they maintain a world-state representation. Goal-based and utility-based agents use goals or outcome preferences to guide selection, rather than being defined by current-percept-only rule matching.

Evidence in the complete study note

A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean. It has no persistent world-state memory. This is adequate when the current observation contains everything needed for its rule, but partial observability can make the same percept require different actions.AI06
A model-based reflex agent maintains internal state using the previous state, previous action, current percept, and models of world change and action effects. It then applies rules to that state. Both kinds can use condition–action rules; memory and a world model distinguish them.AI06
A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07

Choice by choice

A · correct — A simple reflex agent chooses from the current percept without persistent world-state memory.

B · incorrect — A model-based reflex agent adds internal state and applies rules to that modeled state.

C · incorrect — Goal-based selection evaluates actions or future sequences against a desired state.

D · incorrect — Utility-based selection ranks outcomes by desirability or expected utility.

Question 76

Which type of agent maintains an internal model of the world to make decisions?

  • A. Simple reflex agent
  • B. Model-based agent
  • C. Goal-based agent
  • D. Utility-based agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI06 · Simple and model-based reflex agents

AI07 · Goals, utility and model updates

Study-guide references

Correct answer: B — Model-based agent

An internal state based on percept history and a model of how the world changes define the model-based category. Goal and utility agents can extend that architecture, so the requested defining type is B. Evidence: PDF 14.

Evidence in the complete study note

A model-based reflex agent maintains internal state using the previous state, previous action, current percept, and models of world change and action effects. It then applies rules to that state. Both kinds can use condition–action rules; memory and a world model distinguish them.AI06
A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07

Choice by choice

A · incorrect — Current-percept rules omit history. See PDF p. 14.

B · correct — Matches the internal model definition. See PDF p. 14.

C · incorrect — Adds explicit goal information to a model. See PDF p. 14.

D · incorrect — Adds a numerical utility criterion. See PDF p. 14.

Question 77

Which type of agent selects actions based on their expected utility or desirability?

  • A. Simple reflex agent
  • B. Model-based agent
  • C. Goal-based agent
  • D. Utility-based agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI07 · Goals, utility and model updates

Study-guide references

Correct answer: D — Utility-based agent

A utility function assigns real numbers to states to compare how efficiently actions achieve goals; this is precisely desirability-based selection. Evidence: PDF 15.

Evidence in the complete study note

A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07

Choice by choice

A · incorrect — Uses condition-action rules. See PDF p. 15.

B · incorrect — A world model alone supplies no preference scale. See PDF p. 15.

C · incorrect — Goal achievement alone does not rank desirable outcomes. See PDF p. 15.

D · correct — Numerical utility ranks alternatives. See PDF p. 15.

Question 78

Which type of agent represents goals and selects actions that maximize the achievement of those goals?

  • A. Simple reflex agent
  • B. Model-based agent
  • C. Goal-based agent
  • D. Utility-based agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI07 · Goals, utility and model updates

Study-guide references

Correct answer: C — Goal-based agent

Explicit desirable goal situations and choosing actions to reach them characterize goal-based agents. Evidence: PDF 15.

Evidence in the complete study note

A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07
Updating an internal state after selecting an action predicts the action’s effect; it does not itself modify the external environment. A stored variable named goal is insufficient to prove a goal-based policy if BEST_ACTION never uses it. Model-based, goal-based, utility-based and learning capabilities can coexist.AI07

Choice by choice

A · incorrect — Responds to current percept. See PDF p. 15.

B · incorrect — Tracks state without necessarily representing a goal. See PDF p. 15.

C · correct — Explicit goal information supplies the requested criterion. See PDF p. 15.

D · incorrect — Adds utility rankings beyond mere goal achievement. See PDF p. 15.

Question 79

Which type of agent improves its performance over time through learning from experience?

  • A. Simple reflex agent
  • B. Model-based agent
  • C. Goal-based agent
  • D. Learning agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI08 · Learning-agent components

Study-guide references

Correct answer: D — Learning agent

Learning agents adapt using past experience and a learning element that changes performance. Evidence: PDF 15.

Evidence in the complete study note

The performance element selects actions. The learning element changes the agent to improve future performance. The critic evaluates behavior against a performance standard and supplies feedback. The problem generator suggests exploratory actions likely to yield useful experience.AI08
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08

Choice by choice

A · incorrect — Current rules do not themselves explain improvement. See PDF p. 15.

B · incorrect — Tracking state is not itself learning. See PDF p. 15.

C · incorrect — Goal pursuit need not change performance over time. See PDF p. 15.

D · correct — Explicitly learns from experience. See PDF p. 15.

Question 80

A face-recognition system is built to control entry to a building, without trying to explain biological recognition processes. Which approach classifies this practical engineering purpose, rather than its breadth of capability?

  • A. Weak AI approach
  • B. Applied AI approach
  • C. Cognitive AI approach
  • D. Strong AI approach

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI01 · AI capabilities and scope

AI43 · Clarify overlapping learning and AI labels

Study-guide references

Correct answer: B — Applied AI approach

Applied AI builds useful solutions to practical tasks, so building-access recognition fits that purpose. The same system is also narrow or weak AI in capability scope. These labels answer different classification questions and are compatible. Cognitive AI would investigate biological intelligence, while strong-AI claims concern broader competence or actual understanding. The explicit purpose axis isolates the intended selection.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
Narrow or weak AI addresses bounded tasks, such as face recognition, spam filtering or a voice assistant. General AI denotes broad competence transferable across domains; strong AI also has a philosophical meaning involving actual understanding. Do not equate a humanoid body, autonomous vehicle, or many features with general intelligence.AI01
Face recognition is both an applied task and a narrow capability. Broad machine-intelligence goals can include emulating human intelligence and building systems that learn. Humanlike problem-solving alone does not uniquely select expert systems. Python is widely used, but C++ and Java also implement AI; “commonly used” does not isolate one language. McCarthy is associated with naming AI, not its sole invention.AI43

Choice by choice

A · incorrect — Weak AI describes bounded capability, which also fits face recognition, but does not classify the purpose requested.

B · correct — Applied AI builds a useful practical solution, matching the entry-control purpose.

C · incorrect — Cognitive AI investigates biological intelligence; that is not the stated purpose.

D · incorrect — Strong AI concerns actual understanding or broad intelligence claims, not the application purpose described.

Question 81

Which single option gives the complete set of possible AI benefits in suitable, evaluated applications? Interpret the choices operationally, without assuming consciousness or guaranteed improvement.

  • A. Can reduce search work and time with an effective heuristic.
  • B. Can support protective filtering by detecting unwanted spam messages.
  • C. Can automate reasoning and learning to help solve tasks.
  • D. All of the above

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI08 · Learning-agent components

AI01 · AI capabilities and scope

Study-guide references

Correct answer: D — All of the above

Useful AI methods can reduce search work when a heuristic guides exploration effectively. Learned detection of unwanted spam can support protective filtering, and computational reasoning or learning can automate parts of a task. These are possible operational benefits rather than guarantees for every deployment or claims of consciousness. Each individual choice gives only one benefit, so the inclusive response supplies the complete set.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08
Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01

Choice by choice

A · incorrect — Effective heuristics can reduce explored work, but this is only one of the listed possible benefits.

B · incorrect — Spam classification can support protective filtering, but this alone is incomplete.

C · incorrect — Automating computational reasoning and learning can help tasks, but this alone is incomplete.

D · correct — All three are possible operational benefits under the stated suitable-application scope.

Question 82

Which of the following is/are the composition for AI agents?

  • A. Program only
  • B. Architecture only
  • C. Both Program and Architecture
  • D. None of the mentioned

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

AI-S03 · Agent task models and practical planning

Study-guide references

Correct answer: C — Both Program and Architecture

The structural equation adds architecture and agent program; neither component alone is the complete agent. Evidence: PDF 12.

Evidence in the complete study note

A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03
An AI task is modeled as an agent interacting with its environment: observations enter through sensors and selected actions affect the environment through actuators. This task-level description differs from an agent implementation, which combines a program with a computing architecture.AI-S03

Choice by choice

A · incorrect — Omits executing machinery. See PDF p. 12.

B · incorrect — Omits implementing program. See PDF p. 12.

C · correct — Contains both terms in the equation. See PDF p. 12.

D · incorrect — Contradicts the explicit equation. See PDF p. 12.

Question 83

Which are examples of artificial agents, assuming the spacecraft and robot use constructed controllers that sense and select actions autonomously? Select all that apply.

  • A. Autonomous Spacecraft
  • B. Human
  • C. Robot
  • D. All of the mentioned

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

AI39 · Perception, vision and robotics

Study-guide references

Correct answers (select all): A, C — Autonomous Spacecraft; Robot

An artificial agent uses a constructed sensing and action-selection system to interact with its environment. Under the explicit control assumption, both the autonomous spacecraft and robot qualify. Humans also perceive and act, so they are agents, but they are not artificial agents. The inclusive choice fails because it includes humans. This classification concerns artificial agency and does not assert that a robot has general intelligence.

Evidence in the complete study note

An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39

Choice by choice

A · correct — An autonomous spacecraft uses constructed sensing and action-selection mechanisms to interact with its environment.

B · incorrect — A human is an agent but is not an artificial agent.

C · correct — An AI-controlled robot senses and acts physically and is an artificial-agent example.

D · incorrect — Including humans makes this aggregate false under the question’s artificial-agent wording.

Question 84

Which of the following is an expansion of Artificial Intelligence application?

  • A. Game Playing
  • B. Planning and Scheduling
  • C. Diagnosis
  • D. All of the mentioned

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI-S03 · Agent task models and practical planning

AI22 · Games and minimax

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: D — All of the mentioned

The application list includes medical diagnosis, telescope planning/scheduling, and grand-master game play. All three categories are present. Evidence: PDF 7, 7.

Evidence in the complete study note

Planning constructs actions to reach a goal. Scheduling additionally assigns activities to times and resources while respecting constraints such as resource conflicts and precedence. It is an AI application when search or constraint reasoning chooses feasible or useful assignments; automated telescope experiment scheduling is an example.AI-S03
Adversarial search models agents with competing objectives. A game specifies states, player to move, legal actions, successor function, terminal test and utility. Perfect information means previous moves and current state are visible; it does not reveal future choices or imply absence of chance. Zero-sum utility makes one player’s gain the other’s loss.AI22
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · incorrect — Games are explicitly listed. See PDF p. 7, 7.

B · incorrect — Planning and scheduling are explicitly listed. See PDF p. 7, 7.

C · incorrect — Medical diagnosis is explicitly listed. See PDF p. 7, 7.

D · correct — Combines all three documented applications. See PDF p. 7, 7.

Question 85

Which option gives the complete set of valid general descriptions or examples of artificial agents?

  • A. Receives input from its surroundings and selects operations that act on those surroundings.
  • B. A line-following robot controller that reads sensors and commands the robot’s motors.
  • C. Perceives its environment through sensors and acting upon that environment through actuators
  • D. All of the mentioned

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

Study-guide references

Correct answer: D — All of the mentioned

An agent perceives its environment and acts on it. The input-and-operation paraphrase describes that interaction, the sensor-driven line-following controller is an implementation example, and the sensor/actuator statement supplies the general definition. All are valid under the stated descriptions. Since the question requests the complete set in one selection, the aggregate is selected; each individual statement is true but incomplete.

Evidence in the complete study note

An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03
A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03

Choice by choice

A · incorrect — This is a valid input-and-action paraphrase, but it alone omits the other valid descriptions.

B · incorrect — The controller senses and acts through the robot, so it is a valid example, but alone is not the complete set.

C · incorrect — This gives the general sensing-and-acting definition, but alone omits the valid paraphrase and example.

D · correct — The paraphrase, robot-controller example and sensor/actuator definition all fit artificial agency.

Question 86

What is the name of Artificial Intelligence which allows machines to handle vague information with a deftness that mimics human intuition?

  • A. Human intelligence
  • B. Boolean logic
  • C. Functional logic
  • D. Fuzzy logic

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI30 · Fuzzy logic and vague categories

AI25 · Propositional logic and models

Study-guide references

Correct answer: D — Fuzzy logic.

AI30 represents vague categories with degrees between zero and one and explains fuzzy control using graded rules. This supplies the concrete mechanism behind the stem’s loose phrase about human intuition. Boolean logic uses only true or false. Human intelligence is not the named machine-reasoning formalism here, and functional logic is not the provided graded-membership method.

Evidence in the complete study note

Fuzzy sets assign graded membership μ(x) between 0 and 1 to concepts such as warm or tall. A temperature can belong to warm with degree 0.7. This is not a 70% probability that an otherwise crisp warm proposition is true. Boolean logic uses only true and false; probability describes uncertainty about events.AI30
Propositional atoms represent whole true-or-false claims. ¬P negates; P∧Q requires both; P∨Q requires at least one; P→Q is false only when P is true and Q false; P↔Q requires matching truth values. Thus P→Q ≡ ¬P∨Q. De Morgan gives ¬(P∧Q) ≡ ¬P∨¬Q.AI25

Choice by choice

A · incorrect — Human intelligence names biological capability, not the specified artificial graded-membership technique.

B · incorrect — Boolean logic uses crisp true/false values rather than degrees of membership in vague categories.

C · incorrect — Functional logic is not the graded-membership and fuzzy-rule formalism described in AI30.

D · correct — Fuzzy logic handles vague linguistic categories through graded membership and graded rules.

Question 87

An AI system is composed of?

  • A. agent
  • B. environment
  • C. Both A and B
  • D. None of the Above

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI-S03 · Agent task models and practical planning

AI03 · Agents, percepts and action

Study-guide references

Correct answer: C — Both A and B

The explicit system composition is agent plus environment. Evidence: PDF 11.

Evidence in the complete study note

An AI task is modeled as an agent interacting with its environment: observations enter through sensors and selected actions affect the environment through actuators. This task-level description differs from an agent implementation, which combines a program with a computing architecture.AI-S03
A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03

Choice by choice

A · incorrect — Only one component. See PDF p. 11.

B · incorrect — Only the other component. See PDF p. 11.

C · correct — Exactly both documented components. See PDF p. 11.

D · incorrect — Conflicts with the explicit composition. See PDF p. 11.

Question 88

What are Agent?

  • A. An agent perceives its environment through sensors and acts on it through actuators.
  • B. An agent is anything that can change its environment through sensors
  • C. An agent is anything that can control its environment through sensors
  • D. None of the Above

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

Study-guide references

Correct answer: A — An agent perceives its environment through sensors and acts on it through actuators.

An agent is an entity that perceives an environment through sensors and acts on it through actuators or effectors. Sensing alone is only part of this definition, so the first option now includes acting. The other descriptions incorrectly give sensors the action or control role. An agent need not control everything in its environment or guarantee a desired outcome.

Evidence in the complete study note

An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03

Choice by choice

A · correct — This states both parts of the agent interaction: sensing and acting.

B · incorrect — Sensors supply observations; they are not the general components that execute environmental changes.

C · incorrect — Sensors do not generally control the environment; action uses actuators or effectors.

D · incorrect — A complete agent definition is supplied by the sensing-and-acting statement.

Question 89

Which instruments are used for perceiving and acting upon the environment?

  • A. Sensors and Actuators
  • B. Sensors
  • C. Perceiver
  • D. Perceiver and Sensor

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

AI05 · Specify the task with PEAS

Study-guide references

Correct answer: A — Sensors and Actuators

Sensors receive environmental input and actuators deliver output, so both are needed. Evidence: PDF 13.

Evidence in the complete study note

An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03
PEAS means Performance measure, Environment, Actuators, Sensors. Specify all four before designing an agent. For a taxi: performance includes safety and travel time; environment includes roads, traffic and passengers; actuators include steering, brakes and displays; sensors include cameras, positioning and speed readings.AI05

Choice by choice

A · correct — Includes perception and action interfaces. See PDF p. 13.

B · incorrect — Omits action interface. See PDF p. 13.

C · incorrect — Does not identify the two instruments. See PDF p. 13.

D · incorrect — Two perception terms omit actuators. See PDF p. 13.

Question 90

In the teaching taxonomy of decision architectures before learning agents are added as a separate class, how many agent types are counted?

  • A. 1
  • B. 2
  • C. 3
  • D. 4

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI42 · Distinguish agent categories and components

AI06 · Simple and model-based reflex agents

AI07 · Goals, utility and model updates

Study-guide references

Correct answer: D — 4

The note distinguishes a four-architecture convention from a five-class list that also includes learning agents. The specified convention counts simple reflex, model-based reflex, goal-based and utility-based architectures. Learning can improve several of those architectures, but it is excluded as a separate class for this count. The explicit taxonomy therefore determines the number without claiming a universal count for every possible classification.

Evidence in the complete study note

The common teaching list contains five agent classes when learning agents are included, or four decision architectures before learning is added. State the convention when counting agent types. “Target-based” can mean goal-based; “unity-based” is not utility-based.AI42
A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean. It has no persistent world-state memory. This is adequate when the current observation contains everything needed for its rule, but partial observability can make the same percept require different actions.AI06
A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07

Choice by choice

A · incorrect — One does not match either teaching convention described in AI42.

B · incorrect — Two does not match the stated architecture or learning-inclusive list.

C · incorrect — Three is not either count given here.

D · correct — The stated pre-learning architecture taxonomy contains four types.

Question 91

Which of the following is not an agent decision architecture?

  • A. Model based
  • B. Utility based
  • C. Simple reflex
  • D. Search frontier

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI42 · Distinguish agent categories and components

AI12 · Frontiers and evaluation criteria

AI06 · Simple and model-based reflex agents

AI07 · Goals, utility and model updates

Study-guide references

Correct answer: D — Search frontier

Model-based, utility-based and simple reflex describe ways to organize agent decision making. A search frontier instead stores generated nodes awaiting selection or expansion during a search. It is an algorithm component rather than a decision architecture. The overlapping target-based distractor has been replaced because target-based can reasonably mean goal-based, which is a valid agent category.

Evidence in the complete study note

The common teaching list contains five agent classes when learning agents are included, or four decision architectures before learning is added. State the convention when counting agent types. “Target-based” can mean goal-based; “unity-based” is not utility-based.AI42
A search node records a state plus parent, action, depth and path cost. Expanding a node generates its successors. OPEN or frontier stores discovered candidates awaiting expansion; CLOSED records expanded states. Reaching the same state through different paths does not imply the same node or cost.AI12
A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean.AI06
A model-based reflex agent maintains internal state using the previous state, previous action, current percept, and models of world change and action effects.AI06
A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07

Choice by choice

A · incorrect — A model-based architecture maintains an internal representation used in action selection.

B · incorrect — A utility-based architecture selects using outcome desirability.

C · incorrect — Simple reflex is a current-percept condition–action architecture.

D · correct — A search frontier is the collection of generated candidates awaiting selection, not an agent decision architecture.

Question 92

Which is used to improve the agents performance?

  • A. Perceiving
  • B. Observing
  • C. Learning
  • D. Sequence

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI08 · Learning-agent components

Study-guide references

Correct answer: C — Learning

The learning element makes improvements from the environment; perception and observation supply data but are not the improvement mechanism. Evidence: PDF 15.

Evidence in the complete study note

The performance element selects actions. The learning element changes the agent to improve future performance. The critic evaluates behavior against a performance standard and supplies feedback. The problem generator suggests exploratory actions likely to yield useful experience.AI08
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08

Choice by choice

A · incorrect — Supplies input without necessarily changing policy. See PDF p. 15.

B · incorrect — Observation alone need not improve performance. See PDF p. 15.

C · correct — Named improvement mechanism. See PDF p. 15.

D · incorrect — A history is data, not the learning operation. See PDF p. 15.

Question 93

Rationality is?

  • A. being reasonable
  • B. being sensible
  • C. having good sense of judgment.
  • D. All of the above

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI-S03 · Agent task models and practical planning

AI04 · Rationality and performance

Study-guide references

Correct answer: D — All of the above

The module defines rationality with all three phrases: reasonable, sensible, and good judgment. Evidence: PDF 11.

Evidence in the complete study note

In ordinary usage, rationality means reasonable, sensible judgment. For an AI agent, the operational criterion is the action with the best expected performance given its percept history and available knowledge, rather than guaranteed success or knowledge of the actual future.AI-S03
A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04

Choice by choice

A · incorrect — Part of the definition. See PDF p. 11.

B · incorrect — Part of the definition. See PDF p. 11.

C · incorrect — Part of the definition. See PDF p. 11.

D · correct — Includes every listed phrase. See PDF p. 11.

Question 94

Rationality of an agent does not depends on?

  • A. performance measures
  • B. Percept Sequence
  • C. reaction
  • D. actions

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI04 · Rationality and performance

Study-guide references

Correct answer: C — reaction

The explicit determinants are performance measures, percept history, environmental knowledge, and available actions. Reaction is not a separate determinant in that list. Evidence: PDF 12, 12.

Evidence in the complete study note

A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04

Choice by choice

A · incorrect — Explicit determinant. See PDF p. 12, 12.

B · incorrect — Explicit determinant. See PDF p. 12, 12.

C · correct — Not a separate listed factor; may colloquially refer to action. See PDF p. 12, 12.

D · incorrect — Available actions explicitly matter. See PDF p. 12, 12.

Question 95

What is the full form of PEAS?

  • A. Perceptual Measure, Environment, Actuators, and Sensors
  • B. Performance Measure, Environment, Actuators, and Sensors
  • C. Performance Measure, Entity, Actuators, and Sensors
  • D. Performance Measure, Environment, Agent Function, and Sensors

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI05 · Specify the task with PEAS

Study-guide references

Correct answer: B — Performance Measure, Environment, Actuators, and Sensors

PEAS comprises Performance measure, Environment, Actuators, Sensors, spanning the page boundary. Evidence: PDF 13, 14.

Evidence in the complete study note

PEAS means Performance measure, Environment, Actuators, Sensors. Specify all four before designing an agent. For a taxi: performance includes safety and travel time; environment includes roads, traffic and passengers; actuators include steering, brakes and displays; sensors include cameras, positioning and speed readings.AI05

Choice by choice

A · incorrect — Perceptual replaces Performance incorrectly. See PDF p. 13, 14.

B · correct — All four components match. See PDF p. 13, 14.

C · incorrect — Entity replaces Environment. See PDF p. 13, 14.

D · incorrect — Agent Function replaces Actuators. See PDF p. 13, 14.

Question 96

Chess is example of which properties?

  • A. Discrete
  • B. Continuous
  • C. Episodic
  • D. Non-deterministic

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: A — Discrete

Chess is the explicit discrete-environment example: finitely many distinct states/actions. Evidence: PDF 13, 13.

Evidence in the complete study note

Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · correct — Explicit chess classification. See PDF p. 13, 13.

B · incorrect — Driving, not chess, is the continuous example. See PDF p. 13, 13.

C · incorrect — Moves affect future positions rather than independent episodes. See PDF p. 13, 13.

D · incorrect — Board transitions are determined by state and selected move. See PDF p. 13, 13.

Question 97

For an untimed crossword puzzle whose clues and grid do not change while the solver deliberates, which environment property follows from that lack of change?

  • A. Dynamic
  • B. Static
  • C. Semi Dynamic
  • D. Continuous

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: B — Static

A static environment remains unchanged while the agent deliberates. The untimed, fixed-grid puzzle meets that condition. A running clock that affects performance could make an otherwise fixed task semidynamic, but the stem excludes it. Dynamic change and continuous-valued state or action descriptions are separate properties and do not follow from the stated puzzle setting.

Evidence in the complete study note

Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · incorrect — Dynamic means the environment can change during deliberation, excluded by the premise.

B · correct — Static means the relevant environment does not change while the agent deliberates.

C · incorrect — Semidynamic describes changing performance, such as a running clock, despite an otherwise fixed environment; none is present.

D · incorrect — Continuity concerns real-valued states or actions, not whether the puzzle changes during deliberation.

Question 98

Which learning regime improves an agent’s action choices through interaction with an environment that supplies rewards or penalties?

  • A. Supervised learning
  • B. Reinforcement learning
  • C. Semi-supervised learning
  • D. Unsupervised learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI33 · Reinforcement learning

AI32 · Unsupervised, semi-supervised and active learning

AI31 · Supervised learning and prediction

Study-guide references

Correct answer: B — Reinforcement learning

Interaction alone is too broad to identify a learning regime. Here the environment supplies rewards or penalties for the agent’s actions, and the agent improves its choices from that feedback. This defines reinforcement learning. Supervised and semi-supervised methods depend on target labels; unsupervised methods discover unlabeled structure. None of those label-based descriptions specifies this action-and-reward loop.

Evidence in the complete study note

An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Semi-supervised learning combines some labels with additional unlabeled data. Active learning chooses informative cases for an oracle to label; it still uses labels. “Exploratory learning” is not an exclusive standard synonym: it can mean unsupervised exploration or active data acquisition, so define the feedback. A task named summarization or time-series analysis is not automatically supervised; its training targets determine that.AI32
Supervised learning fits a mapping from labeled input–target examples.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features.AI32

Choice by choice

A · incorrect — Supervised learning supplies correct target outputs rather than defining action-reward interaction.

B · correct — Reinforcement learning improves decisions through environmental reward feedback.

C · incorrect — Semi-supervised learning mixes labeled and unlabeled examples, not a defining action-reward objective.

D · incorrect — Unsupervised learning seeks unlabeled structure without the stated reward-feedback objective.

Question 99

Which is reinforcement learning’s immediate parent field in the AI → machine learning → reinforcement learning hierarchy?

  • A. Artifical Intelligence
  • B. Data Analytics
  • C. Machine Learning
  • D. None of the above

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI33 · Reinforcement learning

AI43 · Clarify overlapping learning and AI labels

Study-guide references

Correct answer: C — Machine Learning

Reinforcement learning is a branch of machine learning, and machine learning is within artificial intelligence. Artificial intelligence therefore also contains reinforcement learning, but it is the broader ancestor rather than the immediate parent requested. Data analytics is an application context rather than the stated parent field. The explicit hierarchy makes the nearer containing field the unique taxonomic selection.

Evidence in the complete study note

An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43

Choice by choice

A · incorrect — AI contains RL indirectly through machine learning, but is not the immediate parent requested.

B · incorrect — Data analytics is not the stated parent-field classification of RL in the note.

C · correct — Machine learning is the immediate parent of reinforcement learning in the stated hierarchy.

D · incorrect — The immediate parent is present among the choices.

Question 100

Which type of feedback does an agent in reinforcement learning receive?

  • A. Predictions
  • B. Labels
  • C. Clusters
  • D. Rewards or penalties

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI33 · Reinforcement learning

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

Study-guide references

Correct answer: D — Rewards or penalties

Rewards and penalties are the stated reinforcement feedback; supervised targets and unsupervised grouping differ. Evidence: PDF 69.

Evidence in the complete study note

An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32

Choice by choice

A · incorrect — A prediction is output, not the reward signal. See PDF p. 69.

B · incorrect — Labels belong to supervised examples. See PDF p. 69.

C · incorrect — Clusters are an unsupervised result. See PDF p. 69.

D · correct — Matches the explicit feedback description. See PDF p. 69.

Question 101

What is the main task of a problem-solving agent?

  • A. Solve the given problem and reach to goal
  • B. To find out which sequence of action will get it to the goal state
  • C. Both A and B
  • D. None of the Above

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI11 · Formulate before searching

Study-guide references

Correct answer: C — Both A and B

A solution is an action sequence leading from start to goal; this supports both general goal achievement and finding that sequence. Evidence: PDF 18.

Evidence in the complete study note

First formulate the goal; then formulate the problem by specifying initial state, legal actions, transition model, goal test, and action costs. Search selects a sequence; execution applies it. A state abstracts away irrelevant world details.AI11

Choice by choice

A · incorrect — Describes the overall objective. See PDF p. 18.

B · incorrect — Gives the operational search task. See PDF p. 18.

C · correct — Both are supported by the solution definition. See PDF p. 18.

D · incorrect — Contradicts the defined objective. See PDF p. 18.

Question 102

For DFS on an implicit tree with branching factor b and maximum depth m, storing pending siblings, which expression gives the standard worst-case space bound?

  • A. b
  • B. b^2
  • C. b^b
  • D. O(bm)

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI14 · Depth-first search

Study-guide references

Correct answer: D — O(bm)

Depth-first search retains the active branch and pending alternatives rather than an entire breadth frontier. With up to b pending entries per level over depth m, the standard stored-frontier space bound is O(bm). Path-only on-demand generation can use O(m), but that is a different implementation convention. The expression gives the depth-dependent storage bound rather than an exponential time bound.

Evidence in the complete study note

DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14
Implicit-tree worst-case time is O(b^m), and space is O(bm) when pending siblings are stored, or O(m) for path-only recursive generation. DFS is not generally optimal and can fail on infinite paths or cycles. On a finite graph, visited-state tracking makes traversal complete but needs O(V) storage. Do not confuse the O(b^m) worst-case time bound with the DFS space bounds above.AI14

Choice by choice

A · incorrect — b omits growth with search depth m in the general DFS storage bound.

B · incorrect — b² is not the stated depth-dependent worst-case DFS storage expression.

C · incorrect — b^b is neither of the note’s DFS space bounds.

D · correct — With pending siblings, at most a branching-factor-sized set per depth gives the standard O(bm) space bound.

Question 103

For BFS on an implicit tree with branching factor b>1 and shallowest goal depth d, using generation-time goal testing, which expression gives the standard worst-case generated-node order?

  • A. b
  • B. b^d
  • C. b^2
  • D. b^b

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI13 · Breadth-first search

Study-guide references

Correct answer: B — b^d

BFS generates levels in order. Under generation-time goal testing, the generated nodes through the shallowest goal depth d are bounded by the geometric sum 1+b+…+b^d, whose dominant term is b^d when b>1. Testing only after expanding deeper nodes can change the conventional bound by another factor of b, so the question states the goal-test convention.

Evidence in the complete study note

The tree contains 1+b+…+b^d nodes through depth d, giving O(b^d) under generation-time goal testing. Expanding depth-d nodes before testing can generate depth d+1 and yields O(b^(d+1)); specify the convention. Explicit finite graph traversal with visited-state tracking takes O(V+E) time and O(V) space.AI13

Choice by choice

A · incorrect — b counts one level and does not account for arbitrary goal depth.

B · correct — The sum 1+b+…+b^d has order b^d for b>1.

C · incorrect — b² only describes a fixed depth-two case rather than general depth d.

D · incorrect — b^b wrongly replaces solution depth d with the branching factor.

Question 104

Which labels describe the recursive DFS stack’s removal order? Select all that apply.

  • A. LIFO
  • B. LILO
  • C. FIFO
  • D. FILO

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI14 · Depth-first search

AI13 · Breadth-first search

Study-guide references

Correct answers (select all): A, D — LIFO; FILO

Recursive depth-first search returns first from the most recently entered unfinished call. That is last-in-first-out stack order, also called first-in-last-out. The two labels are synonyms rather than distinct data structures. FIFO or LILO describes the opposite end-to-end order and does not characterize the DFS call stack. The multiple-selection instruction therefore retains both correct names for the same discipline.

Evidence in the complete study note

DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13

Choice by choice

A · correct — LIFO is the stack order used by recursion and depth-first exploration.

B · incorrect — LILO does not express the required last-in-first-out removal order.

C · incorrect — FIFO is queue order, associated with breadth-first rather than depth-first selection.

D · correct — FILO is another name for LIFO and therefore also describes the DFS stack.

Question 105

Which of the following search algorithm searches forward from initial state and backward from goal state till both meet to identify a common state?

  • A. Uniform Cost Search
  • B. Iterative Deepening Depth-First Search
  • C. Bidirectional Search
  • D. None of the Above

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI16 · Depth limits, iterative deepening and two directions

AI15 · Uniform-cost search and Dijkstra

Study-guide references

Correct answer: C — Bidirectional Search

The algorithm starts one search at the initial state and another at the goal and stops at their intersection. Evidence: PDF 23.

Evidence in the complete study note

Bidirectional search grows from start and goal until frontiers connect. With suitable reverse actions, meeting tests and equal-cost BFS, time and space may fall to O(b^(d/2)). Weighted variants require a sound stopping bound; the first intersection need not be cheapest. IDA* deepens an f=g+h threshold rather than only a depth limit.AI16
UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15
DLS runs DFS to depth ℓ and reports cutoff separately from failure. It misses goals deeper than ℓ. Iterative deepening runs DLS with limits 0,1,2,…; for finite b it finds the shallowest solution with O(b^d) time and O(bd) frontier space, optimal for equal step costs. Repeated shallow work is small compared with the deepest level.AI16

Choice by choice

A · incorrect — Orders by path cost rather than two directions. See PDF p. 23.

B · incorrect — Repeats depth-limited search from the start. See PDF p. 23.

C · correct — Exactly the two-front definition. See PDF p. 23.

D · incorrect — C explicitly fits. See PDF p. 23.

Question 106

When is breadth-first search is optimal?

  • A. When there is less number of nodes
  • B. When all step costs are equal
  • C. When all step costs are unequal
  • D. All of the above

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI13 · Breadth-first search

Study-guide references

Correct answer: B — When all step costs are equal

Equal positive step costs make total path cost proportional to depth; BFS first reaches a shallowest solution and hence a cheapest one. Evidence: PDF 20.

Evidence in the complete study note

BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
The tree contains 1+b+…+b^d nodes through depth d, giving O(b^d) under generation-time goal testing. Expanding depth-d nodes before testing can generate depth d+1 and yields O(b^(d+1)); specify the convention. Explicit finite graph traversal with visited-state tracking takes O(V+E) time and O(V) space.AI13

Choice by choice

A · incorrect — Node count alone says nothing about cost. See PDF p. 20.

B · correct — Equal step costs ensure depth order is cost order. See PDF p. 20.

C · incorrect — Unequal costs can make a deeper path cheaper. See PDF p. 20.

D · incorrect — A and C are not sufficient conditions. See PDF p. 20.

Question 107

What is the heuristic function of greedy best-first search?

  • A. f(n) != h(n)
  • B. f(n) < h(n)
  • C. f(n) = h(n)
  • D. f(n) > h(n)

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: C — f(n) = h(n)

The operational rule removes the node with lowest h(n), so the greedy evaluation is f(n)=h(n). The printed f(n)=g(n) on this same page conflicts with its own algorithm and nearby h definition; the operational evidence supports C. Evidence: PDF 25.

Evidence in the complete study note

Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17

Choice by choice

A · incorrect — Inequality would not express that rule. See PDF p. 25.

B · incorrect — Strictly smaller is not the definition. See PDF p. 25.

C · correct — Matches ranking exclusively by h. See PDF p. 25.

D · incorrect — Strictly greater is not the definition. See PDF p. 25.

Question 108

Greedy best-first search selects which descriptions of the next node? Interpret “closest” as estimated remaining cost h, not actual distance. Select all that apply.

  • A. Shallowest
  • B. Deepest
  • C. The one closest to the goal node
  • D. Minimum heuristic cost

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI13 · Breadth-first search

AI14 · Depth-first search

Study-guide references

Correct answers (select all): C, D — The one closest to the goal node; Minimum heuristic cost

Greedy best-first search orders the frontier by h alone, choosing the smallest estimated remaining cost and ignoring cost already paid. Under the explicit interpretation, estimated closest to the goal and minimum heuristic cost express that same rule. Shallowest-first uses depth as in breadth-first search, while deepest-first follows depth-first ordering. Both heuristic descriptions are therefore valid and neither depth description is selected.

Evidence in the complete study note

Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14

Choice by choice

A · incorrect — Shallowest-first is BFS-style depth ordering rather than greedy heuristic selection.

B · incorrect — Deepest-first is DFS-style selection rather than minimum estimated remaining cost.

C · correct — Read closest as estimated by h; it is the same greedy criterion as minimum heuristic cost.

D · correct — Greedy best-first selects a node with minimum h, ignoring already-paid path cost.

Question 109

Which general search family contains A* because it selects frontier nodes by a minimum evaluation value?

  • A. Breadth-First-Search
  • B. Depth-First-Search
  • C. Uniform Cost Search
  • D. Best-First-Search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI18 · Admissibility, consistency and A*

AI15 · Uniform-cost search and Dijkstra

AI17 · Heuristics and greedy best-first search

AI13 · Breadth-first search

AI14 · Depth-first search

Study-guide references

Correct answer: D — Best-First-Search

Best-first search is the general family that selects a frontier node according to an evaluation value. A* is one member, using f=g+h. Uniform-cost search is another member, using g alone. Breadth-first and depth-first names describe depth-oriented selection disciplines. The question asks for the containing family rather than claiming A* is identical to one simpler algorithm.

Evidence in the complete study note

A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15
Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
BFS removes the oldest frontier entry using a FIFO queue.AI13
DFS follows the deepest pending branch using a LIFO stack or recursion.AI14

Choice by choice

A · incorrect — BFS selects by depth/FIFO order rather than the general evaluation-priority family.

B · incorrect — DFS follows stack/deepest-branch order rather than minimum evaluation priority.

C · incorrect — UCS is another best-first method using g alone; it is not the general family containing A*.

D · correct — Best-first search is the family that selects by an evaluation, with A* using g+h.

Question 110

Function MY_AGENT (PERCEPT) return action static state, a description of the current world state goal, a description of the goal to achieve may be in terms of state state ← UPDATE_STATE(state, percept) actionSet ← POSSIBLE_ACTIONS(state) action ← BEST_ACTION(actionSet) state ← UPDATE_STATE(state, action) return action What is the purpose of the final instruction state ← UPDATE_STATE(state, action), after BEST_ACTION has selected an action?

  • A. Modify the state of the environment if necessary
  • B. Update the state information to be like the one after executing the best action chosen
  • C. Update the environment so that the agent can perceive the next percept
  • D. Modify the state of the environment once the decision of which action to choose is made.

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI07 · Goals, utility and model updates

AI06 · Simple and model-based reflex agents

Study-guide references

Correct answer: B — Update the state information to be like the one after executing the best action chosen

The final update uses the chosen action to predict what the agent’s internal state representation should look like after that action. It is distinct from the earlier update using the current percept. Updating memory does not execute an actuator or directly modify the external environment. Naming the exact instruction resolves the missing referent while preserving the model-update interpretation of the original prototype.

Evidence in the complete study note

Updating an internal state after selecting an action predicts the action’s effect; it does not itself modify the external environment. A stored variable named goal is insufficient to prove a goal-based policy if BEST_ACTION never uses it. Model-based, goal-based, utility-based and learning capabilities can coexist.AI07
A model-based reflex agent maintains internal state using the previous state, previous action, current percept, and models of world change and action effects. It then applies rules to that state. Both kinds can use condition–action rules; memory and a world model distinguish them.AI06

Choice by choice

A · incorrect — Updating an internal state representation does not itself modify the external environment.

B · correct — The final call predicts and records the internal state after the selected action.

C · incorrect — Neither internal update executes an environmental action or produces the next percept.

D · incorrect — Selecting an action and updating memory are distinct from changing the external world through execution.

Question 111

To implement greedy best-first search, let the evaluation be the estimated remaining cost h(n). Which OPEN-list insertion policy supplies its node-selection order?

  • A. Append newly generated nodes at the beginning of the open list
  • B. Append newly generated nodes at the end of the open list
  • C. Insert newly generated nodes in the open list according to their path cost (lowest values first)
  • D. Insert newly generated nodes in the open list according to their evaluation function (lowest values first)

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI15 · Uniform-cost search and Dijkstra

AI13 · Breadth-first search

AI14 · Depth-first search

Study-guide references

Correct answer: D — Insert newly generated nodes in the open list according to their evaluation function (lowest values first)

Greedy best-first search orders its frontier by h alone, the estimate of remaining cost. A priority-ordered OPEN list with the smallest evaluation first therefore implements the rule when that evaluation is h. Ordering by g would instead implement uniform-cost search. Unconditionally prepending or appending nodes gives stack-like or FIFO behavior without the required heuristic ranking.

Evidence in the complete study note

Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14

Choice by choice

A · incorrect — Prepending every new node gives a stack-like policy rather than ordering by remaining-cost estimates.

B · incorrect — Appending every new node gives FIFO breadth order rather than heuristic priority.

C · incorrect — Ordering by accumulated path cost g implements uniform-cost selection, not greedy h-only selection.

D · correct — Ordering by the stated evaluation h puts the smallest estimated remaining cost first.

Question 112

In classical propositional logic over a nonempty set of interpretations, which statements are false? In D, “models” means all interpretations. Select all that apply.

  • A. A sentence is said to be invalid iff it is unsatisfiable.
  • B. A sentence is valid if it is satisfiable.
  • C. A sentence which is not invalid is satisfiable.
  • D. A sentence is unsatisfiable if it is false in all models.

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI25 · Propositional logic and models

Study-guide references

Correct answers (select all): A, B — A sentence is said to be invalid iff it is unsatisfiable.; A sentence is valid if it is satisfiable.

Validity requires truth in every interpretation, while satisfiability requires truth in at least one. A contingent atom can therefore be satisfiable and invalid, refuting both the proposed invalid-unsatisfiable equivalence and the claim that satisfiability implies validity. Not invalid means valid and hence satisfiable in the stated nonempty setting. Unsatisfiable means false in every interpretation, so the remaining two statements are true.

Evidence in the complete study note

A model is an interpretation making a sentence true. Valid means true in every interpretation; satisfiable means true in at least one; unsatisfiable means true in none. Invalid means not valid, and may still be satisfiable. P∨¬P is valid; P is satisfiable but not valid; P∧¬P is unsatisfiable. Entailment KB⊨α means every model of KB also satisfies α.AI25

Choice by choice

A · correct — False: a satisfiable contingent sentence can be invalid without being unsatisfiable.

B · correct — False: truth in one interpretation does not imply truth in every interpretation.

C · incorrect — True: not invalid means valid, which entails satisfiability in the ordinary model setting.

D · incorrect — True: unsatisfiable means there is no interpretation making the sentence true.

Question 113

Consider this agent prototype: Function MY_AGENT(percept) returns action static state, goal state ← UPDATE_STATE(state, percept) actionSet ← POSSIBLE_ACTIONS(state) action ← BEST_ACTION(actionSet, state, goal) state ← UPDATE_STATE(state, action) return action BEST_ACTION uses lookahead over future action sequences and chooses by whether the specified desired state is reached; it does not rank successful outcomes by utility or learn from experience. Which category best describes this action-selection criterion?

  • A. Goal based
  • B. Model based reflex
  • C. Simple reflex
  • D. Utility based

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI07 · Goals, utility and model updates

AI06 · Simple and model-based reflex agents

Study-guide references

Correct answer: A — Goal based

The action selector explicitly consults the desired state and looks ahead over sequences that can reach it. That is the defining goal-based decision criterion. Persistent state also gives the agent a model-based capability, but memory alone does not identify its selection policy. A reflex policy would apply condition-action rules instead of goal-directed lookahead, and utility-based choice would rank degrees of desirability among outcomes.

Evidence in the complete study note

A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07
Updating an internal state after selecting an action predicts the action’s effect; it does not itself modify the external environment. A stored variable named goal is insufficient to prove a goal-based policy if BEST_ACTION never uses it. Model-based, goal-based, utility-based and learning capabilities can coexist.AI07
A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean. It has no persistent world-state memory. This is adequate when the current observation contains everything needed for its rule, but partial observability can make the same percept require different actions.AI06

Choice by choice

A · correct — Goal-directed lookahead chooses sequences reaching a desired state, matching goal-based action selection.

B · incorrect — The agent maintains a model, but the stated policy uses goal-directed lookahead rather than reflex rules.

C · incorrect — Persistent state and future-sequence evaluation contradict a current-percept-only reflex policy.

D · incorrect — No utility ranking is used; the selector evaluates attainment of the desired state.

Question 114

Which are general terms for the components through which an arbitrary AI agent acts on its environment, rather than sensing components or one particular robot subtype? Select all that apply.

  • A. Sensors
  • B. Effectors
  • C. Actuators
  • D. Manipulators

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

AI42 · Distinguish agent categories and components

AI39 · Perception, vision and robotics

Study-guide references

Correct answers (select all): B, C — Effectors; Actuators

Actuators and effectors are equivalent general names for the agent’s acting interface here. Sensors instead supply observations. A manipulator can be a particular physical acting device, but it is not the general term covering every agent, including software agents. The scope allows the two synonymous general terms without incorrectly denying that robotic manipulators can carry out actions.

Evidence in the complete study note

An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03
A generic rule does not distinguish simple from model-based reflex without specifying memory. A goal declaration unused by the action policy does not establish goal-based behavior. “Sensors and effectors” describes interaction, not omniscience or intelligence by itself. Define the mechanical distinction when comparing effectors and actuators as acting components.AI42
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39

Choice by choice

A · incorrect — Sensors receive observations rather than enact the agent’s selected environmental changes.

B · correct — Effectors are an accepted name for the components through which the agent acts.

C · correct — Actuators are the note’s equivalent general term for acting components.

D · incorrect — A manipulator is a particular physical device, not the general acting-interface term requested.

Question 115

A control agent has incomplete observations, uncertain transition outcomes, actions that affect later decisions, a world that changes during deliberation, and real-valued measurements and controls. Which list matches all these properties?

  • A. Fully observable, discrete, episodic, static.
  • B. Static, episodic, continuous, partially observable.
  • C. Partially observable, non-deterministic, non-episodic, dynamic, continuous.
  • D. Deterministic, sequential (non-episodic), dynamic, continuous.

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: C — Partially observable, non-deterministic, non-episodic, dynamic, continuous.

The scenario establishes each environment dimension directly. Incomplete observations give partial observability; uncertain outcomes give non-determinism; effects on later decisions give sequentiality; ongoing change gives dynamics; and real-valued quantities give continuity. Only one list matches the whole combination. This classification avoids claiming that these dimensions establish a universal numerical ranking of task difficulty.

Evidence in the complete study note

Fully observable means sensors reveal relevant state; partially observable means relevant information is hidden or noisy. Deterministic means state and action determine the next state; stochastic means outcomes involve randomness. Strategic uncertainty about an opponent is distinct from dice-like randomness.AI09
Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · incorrect — Full observability, episodic independence and a static discrete setting contradict the scenario.

B · incorrect — Partial observability and continuity fit, but static and episodic do not.

C · correct — Partial observability, non-determinism, sequentiality, dynamics and continuity all match.

D · incorrect — The other stated properties fit, but determinism contradicts uncertain transition outcomes.

Question 116

Which described training tasks are supervised learning? Select all that apply.

  • A. Summarization trained on document/reference-summary pairs.
  • B. Time-series forecasting trained on history/next-value target pairs.
  • C. Sequence discovery without target labels.
  • D. Clustering unlabeled examples.

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

Study-guide references

Correct answers (select all): A, B — Summarization trained on document/reference-summary pairs.; Time-series forecasting trained on history/next-value target pairs.

Supervised training uses paired inputs and target outputs. Document/reference-summary pairs therefore make the summarization task supervised, and history/next-value pairs make the forecasting task supervised. Sequence discovery and clustering without targets instead seek structure in unlabeled data. The labels in the descriptions determine the regime; task names alone would not establish whether summarization or time-series analysis is supervised.

Evidence in the complete study note

Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32

Choice by choice

A · correct — The reference summary is a target output paired with each input document.

B · correct — The next value is a numeric target paired with each input history.

C · incorrect — Without target labels, sequence discovery seeks unlabeled structure.

D · incorrect — Clustering unlabeled examples is unsupervised structure discovery.

Question 117

A common name used to describe search-based optimization technique based on the principles of Genetics and Natural Selection is ____.

  • A. Genetic algorithms
  • B. Heuristic algorithm
  • C. Hill climbing
  • D. Simulated annealing

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI20 · Local optimization and evolutionary search

AI-S05 · Related search and game-evaluation methods

Study-guide references

Correct answer: A — Genetic algorithms.

AI20 names the search method with population, fitness, parent selection, crossover and mutation as a genetic algorithm. That is the specific common name requested by the genetics/natural-selection description. Heuristic algorithm is a broad umbrella rather than this name; hill climbing follows improved neighbors, and simulated annealing is not the population-selection mechanism described.

Evidence in the complete study note

A genetic algorithm maintains a population, evaluates fitness, selects parents, applies crossover and mutation, and repeats replacement. It searches without guaranteeing global optimality. Means–ends analysis selects a difference between current and goal states, chooses an operator reducing it, and recursively solves its preconditions. Neither method is simply A* running backward.AI20
Local search keeps a complete candidate configuration and improves its objective; the path may be irrelevant. Hill climbing chooses better neighbors. First-improvement accepts the first improvement; steepest-ascent examines all neighbors; stochastic variants choose among improvements. Local maxima, plateaus and ridges can stop progress. Random restarts or limited sideways steps can help without guaranteeing success.AI20
Simulated annealing follows a current candidate and randomly samples a neighboring move. It accepts improvements and sometimes accepts worse moves according to a temperature that decreases over time. Unlike genetic search, its defining mechanism is not a population undergoing parent selection, crossover and mutation.AI-S05

Choice by choice

A · correct — This is the specific name of the evolutionary search technique based on fitness, selection and genetic variation.

B · incorrect — Heuristic algorithm is a broad category and does not specifically name the genetics-based population method.

C · incorrect — Hill climbing improves a candidate through neighbors rather than maintaining the described genetic population.

D · incorrect — Simulated annealing is not the named selection/crossover/mutation method defined in the note.

Question 118

A complete history of everything that the agent has ever perceived is

  • A. Agent program
  • B. Agent function
  • C. Percept
  • D. Percept sequence

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

Study-guide references

Correct answer: D — Percept sequence

Percept sequence is explicitly the history of all the agent has perceived, unlike a single current percept. Evidence: PDF 10.

Evidence in the complete study note

A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03

Choice by choice

A · incorrect — Implementation of agent function, not history. See PDF p. 10.

B · incorrect — Maps history to actions, not history itself. See PDF p. 10.

C · incorrect — One perceptual instant. See PDF p. 10.

D · correct — Exact history definition. See PDF p. 10.

Question 119

A good cluster is characterized by

  • A. Equal number of instances
  • B. Equal distance among cluster members
  • C. High inter-class similarity
  • D. High intra-class similarity

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI32 · Unsupervised, semi-supervised and active learning

Study-guide references

Correct answer: D — High intra-class similarity

The module groups similar input vectors into clusters. Good within-cluster (intra-class) similarity follows from that grouping principle, with no requirement of equal sizes or equal pairwise distances. Evidence: PDF 75.

Evidence in the complete study note

Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32

Choice by choice

A · incorrect — No equal-size condition given. See PDF p. 75.

B · incorrect — Similarity does not require identical distances. See PDF p. 75.

C · incorrect — Between-cluster similarity is not the grouping objective. See PDF p. 75.

D · correct — Matches similarity inside a group. See PDF p. 75.

Question 120

A problem in supervised machine learning model due to too much attention to the particular data set it is trained on, causing it to perform poorly on unseen data is ____

  • A. Under fitting
  • B. Over fitting
  • C. Learning curve
  • D. Variance

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI35 · Generalization and validation

AI-S06 · Prediction, transfer and explicit representations

Study-guide references

Correct answer: B — Overfitting.

AI35 directly defines overfitting as fitting training-specific details so performance drops on unseen data. This is the complete failure pattern described in the stem. Underfitting is insufficient fit even on training data; a learning curve is a plot used to inspect performance. Variance may describe sensitivity but does not name the training-versus-generalization failure asked for.

Evidence in the complete study note

Overfitting means fitting training-specific details so performance drops on unseen data; underfitting means the model is too limited or poorly fitted even on training data. A learning curve plots performance against training size or training progress. Complexity control, regularization, pruning and more representative data can help.AI35
Support-vector machines include supervised classification and regression methods. They use support vectors in their decision function rather than being defined by a next-state conditional-dependence assumption. In estimation, variance measures how predictions change across different training samples; overfitting instead names poor generalization caused by fitting training-specific details.AI-S06

Choice by choice

A · incorrect — Underfitting means the model is too limited or poorly fitted even on the training data.

B · correct — Overfitting captures training-specific details at the expense of performance on unseen examples.

C · incorrect — A learning curve plots performance against training size or progress; it is not the failure itself.

D · incorrect — Variance is related to sensitivity, but overfitting is the precise failure pattern defined by the stem.

Question 121

Which general AI search category uses the problem definition and known costs without a problem-specific estimate of remaining cost to a goal?

  • A. Informed search
  • B. Uninformed search
  • C. Greedy best-first search
  • D. Binary search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI-S02 · Ordered lookup versus state-space traversal

Study-guide references

Correct answer: B — Uninformed search

Uninformed search uses the problem definition, including legal actions, goal tests and known action costs, without a domain-specific remaining-cost estimate. Informed methods add such guidance, and greedy best-first uses it directly as h. Binary search is a different ordered-lookup algorithm for sorted sequences. The formerly undefined circular-search distractor is replaced with a meaningful search method.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
Binary search exploits the order of a sorted sequence: a comparison narrows the remaining candidate interval instead of exploring arbitrary successor states. Array bisection therefore differs from breadth-first or depth-first traversal of a general state-space graph. A storage container such as an array does not by itself specify FIFO or LIFO frontier behavior.AI-S02
Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17

Choice by choice

A · incorrect — Informed search adds problem-specific guidance beyond the basic problem definition.

B · correct — Uninformed search uses the problem definition without a heuristic estimate of remaining cost.

C · incorrect — Greedy best-first explicitly uses h, an estimate of remaining cost.

D · incorrect — Binary search narrows a sorted sequence by comparisons; it is a specific ordered-lookup algorithm, not the broad AI search category requested.

Question 122

Which named test uses a judge’s inability to distinguish a machine from a human through conversation as its operational criterion?

  • A. The Chinese room
  • B. A psychological response-time experiment
  • C. Turing test
  • D. Black box testing

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI02 · Four approaches and the Turing test

AI-S07 · Behavior tests, understanding and software tests

Study-guide references

Correct answer: C — Turing test

The conversational Turing test evaluates whether a judge can reliably distinguish a machine’s dialogue from a human’s. It supplies an observable behavior criterion rather than proof of understanding. The Chinese-room argument addresses understanding philosophically, response-time experiments test cognitive models empirically, and black-box testing concerns externally visible software functionality. Those are different questions or procedures.

Evidence in the complete study note

The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02
Searle’s Chinese-room argument challenges whether executing formal symbol-manipulation rules is sufficient for understanding. It is a philosophical argument about understanding, not the conversational judge-and-hidden-participant procedure of the Turing test.AI-S07
Black-box software testing examines externally visible functionality without inspecting implementation internals. A specification may concern any software behavior; this testing method does not by itself require conversational indistinguishability from a human.AI-S07

Choice by choice

A · incorrect — The Chinese-room argument concerns whether formal symbol manipulation suffices for understanding; it is not the judge-and-conversation test.

B · incorrect — A response-time experiment evaluates a cognitive model against human measurements rather than applying conversational indistinguishability.

C · correct — The conversational Turing test uses this judge-and-hidden-participant criterion.

D · incorrect — Black-box testing checks external software functionality without inspecting internals; it does not inherently require humanlike conversation.

Question 123

A learning agent has mainly four conceptual components, which component is responsible for suggesting actions that will lead to new and informative experiences.

  • A. Problem generator
  • B. Performance element
  • C. Learning element
  • D. Critic

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI08 · Learning-agent components

Study-guide references

Correct answer: A — Problem generator

Problem generator proposes actions producing new informative experience, while the performance element chooses external action, learning element improves it, and critic evaluates it. Evidence: PDF 15.

Evidence in the complete study note

The performance element selects actions. The learning element changes the agent to improve future performance. The critic evaluates behavior against a performance standard and supplies feedback. The problem generator suggests exploratory actions likely to yield useful experience.AI08

Choice by choice

A · correct — Matches the explicit exploration role. See PDF p. 15.

B · incorrect — Selects external action, different component role. See PDF p. 15.

C · incorrect — Changes performance based on feedback. See PDF p. 15.

D · incorrect — Evaluates against a performance standard. See PDF p. 15.

Question 124

A learning scheme in which agent learns from a series of rewards & punishments is ____

  • A. Opinion mining
  • B. Unsupervised learning
  • C. Supervised learning
  • D. Reinforcement learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI33 · Reinforcement learning

AI37 · Language processing tasks

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

Study-guide references

Correct answer: D — Reinforcement learning

Rewards and penalties define the module reinforcement scheme. Evidence: PDF 69.

Evidence in the complete study note

An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Tokenization divides text into processing units; tokens may be words, punctuation or subwords. POS tagging assigns grammatical roles; named-entity recognition identifies people, organizations and locations. Parsing identifies grammatical structure; semantics concerns meaning; discourse links sentences; pragmatics uses context and speaker intent. Sentiment analysis predicts evaluative tone. Ambiguity can occur at every level: “bank” has multiple senses, and “I saw the man with a telescope” has multiple attachments.AI37
Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32

Choice by choice

A · incorrect — Opinion mining is a language application. See PDF p. 69.

B · incorrect — Input-only pattern discovery lacks reward feedback. See PDF p. 69.

C · incorrect — Uses paired desired outputs. See PDF p. 69.

D · correct — Exactly reward/penalty feedback. See PDF p. 69.

Question 125

A learning scheme in which agent learns patterns in the input without any explicit feedback is _____________.

  • A. Supervised learning
  • B. Unsupervised learning
  • C. Opinion mining
  • D. Reinforcement learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI32 · Unsupervised, semi-supervised and active learning

AI31 · Supervised learning and prediction

AI33 · Reinforcement learning

AI37 · Language processing tasks

Study-guide references

Correct answer: B — Unsupervised learning

The unsupervised network must discover features without feedback about correct outputs. Evidence: PDF 75.

Evidence in the complete study note

Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Tokenization divides text into processing units; tokens may be words, punctuation or subwords. POS tagging assigns grammatical roles; named-entity recognition identifies people, organizations and locations. Parsing identifies grammatical structure; semantics concerns meaning; discourse links sentences; pragmatics uses context and speaker intent. Sentiment analysis predicts evaluative tone. Ambiguity can occur at every level: “bank” has multiple senses, and “I saw the man with a telescope” has multiple attachments.AI37

Choice by choice

A · incorrect — Has desired outputs supplied. See PDF p. 75.

B · correct — Discovers patterns without teacher feedback. See PDF p. 75.

C · incorrect — Opinion mining is an application, not this learning scheme. See PDF p. 75.

D · incorrect — Uses evaluative environmental feedback. See PDF p. 75.

Question 126

Suppose you have developed a self-driving car equipped with relevant sensors and cameras. Its sensors are able to detect the environment but not all the aspects. Hence, in which environment is your car acting?

  • A. Stochastic
  • B. Discrete
  • C. Fully observable
  • D. Partially observable

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: D — Partially observable

The sensors cannot determine complete state; this is precisely partial observability. It says nothing about transition randomness. Evidence: PDF 13.

Evidence in the complete study note

Fully observable means sensors reveal relevant state; partially observable means relevant information is hidden or noisy. Deterministic means state and action determine the next state; stochastic means outcomes involve randomness. Strategic uncertainty about an opponent is distinct from dice-like randomness.AI09
Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · incorrect — Stochastic describes transitions, not hidden aspects. See PDF p. 13.

B · incorrect — Discrete describes state structure, not visibility. See PDF p. 13.

C · incorrect — Contradicted by incomplete detection. See PDF p. 13.

D · correct — Matches missing state information. See PDF p. 13.

Question 127

Which quantifier can be used for some portion of the universe?

  • A. For all (universal)
  • B. Implies
  • C. There exists (existential)
  • D. And

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI26 · First-order logic and quantifier scope

AI25 · Propositional logic and models

Study-guide references

Correct answer: C — There exists (existential)

Some or at least one is existential quantification; universal applies to all. Evidence: PDF 49.

Evidence in the complete study note

First-order logic adds objects, predicates, functions and quantified object variables. Human(a) is a property; Likes(a,b) a relation; father(a) a function term. Propositional logic has no object quantifiers, but its truth still depends on the assigned atom values.AI26
Propositional atoms represent whole true-or-false claims. ¬P negates; P∧Q requires both; P∨Q requires at least one; P→Q is false only when P is true and Q false; P↔Q requires matching truth values. Thus P→Q ≡ ¬P∨Q. De Morgan gives ¬(P∧Q) ≡ ¬P∨¬Q.AI25

Choice by choice

A · incorrect — All instances rather than some. See PDF p. 49.

B · incorrect — A connective, not a quantifier. See PDF p. 49.

C · correct — Matches some/at least one. See PDF p. 49.

D · incorrect — A connective, not a quantifier. See PDF p. 49.

Question 128

Which one of the following is a step in problem solving that lists the actions and states that the agent has to consider provided a goal?

  • A. Problem formulation
  • B. Execute
  • C. Search
  • D. Goal formulation

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI11 · Formulate before searching

Study-guide references

Correct answer: A — Problem formulation.

AI11 says the goal is formulated first, then the problem specifies initial state, legal actions, transitions, goal test and action costs. Given a goal, listing the relevant actions and states belongs to that second step. Search chooses a sequence within the formulated problem, and execution applies it; neither defines the alternatives initially.

Evidence in the complete study note

First formulate the goal; then formulate the problem by specifying initial state, legal actions, transition model, goal test, and action costs. Search selects a sequence; execution applies it. A state abstracts away irrelevant world details.AI11

Choice by choice

A · correct — Problem formulation specifies the states, available actions and model the agent must consider for its goal.

B · incorrect — Execution applies a selected action sequence after formulation and search.

C · incorrect — Search selects a solution using the already specified states, actions, model and goal test.

D · incorrect — Goal formulation chooses the objective; the stem already provides a goal before listing actions and states.

Question 129

Which term names the entire ordered history of observations an agent has received so far?

  • A. Performance measure
  • B. Percept sequence
  • C. Knowledge
  • D. Action

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

AI04 · Rationality and performance

Study-guide references

Correct answer: B — Percept sequence

A percept is one observation, while a percept sequence is the entire observation history. A performance measure defines success instead of storing observations. Knowledge may include prior facts and inferences beyond the observation history, and an action is selected behavior rather than observed input. Replacing the overly broad perception option supplies the precise established term without changing the agent-history concept.

Evidence in the complete study note

A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03
A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04

Choice by choice

A · incorrect — A performance measure defines success, not what the agent has observed.

B · correct — A percept sequence is the entire observation history.

C · incorrect — Knowledge includes prior and inferred information, not precisely the observation history.

D · incorrect — An action is selected behavior rather than the agent’s accumulated observations.

Question 130

Every training example has an input and its desired target output; there is no extra unlabeled training pool or interaction reward. Which learning regime fits this setup?

  • A. Semi-supervised learning
  • B. Unsupervised learning
  • C. Supervised learning
  • D. Reinforcement learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

AI33 · Reinforcement learning

Study-guide references

Correct answer: C — Supervised learning

The training data consist entirely of input–target pairs, so the task is supervised learning. The desired target provides the learning signal. Semi-supervised learning would add an unlabeled pool, unsupervised learning would seek structure without targets, and reinforcement learning would use rewards from interaction. Stating the complete data setup distinguishes the regimes without implying that any use of labels excludes semi-supervised methods.

Evidence in the complete study note

Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Semi-supervised learning combines some labels with additional unlabeled data. Active learning chooses informative cases for an oracle to label; it still uses labels. “Exploratory learning” is not an exclusive standard synonym: it can mean unsupervised exploration or active data acquisition, so define the feedback. A task named summarization or time-series analysis is not automatically supervised; its training targets determine that.AI32
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33

Choice by choice

A · incorrect — Semi-supervised learning requires additional unlabeled training data alongside labeled examples, absent here.

B · incorrect — Unsupervised learning discovers structure without the supplied target labels.

C · correct — Supervised learning fits the mapping from the labeled input–target pairs described.

D · incorrect — Reinforcement learning uses evaluative interaction rewards, not the described desired-output pairs.

Question 131

The simple strategy in which the root node is expanded first, then all successors are expanded next, layer by layer, refers to:

  • A. Greedy best-first search
  • B. Depth-first search (DFS)
  • C. Depth limited search (DLS)
  • D. Breadth-first search (BFS)

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI13 · Breadth-first search

AI14 · Depth-first search

AI16 · Depth limits, iterative deepening and two directions

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: D — Breadth-first search (BFS)

Expanding every node at one level before the next is the stated BFS order. Evidence: PDF 19.

Evidence in the complete study note

BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14
DLS runs DFS to depth ℓ and reports cutoff separately from failure. It misses goals deeper than ℓ. Iterative deepening runs DLS with limits 0,1,2,…; for finite b it finds the shallowest solution with O(b^d) time and O(bd) frontier space, optimal for equal step costs. Repeated shallow work is small compared with the deepest level.AI16
Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17

Choice by choice

A · incorrect — Ranks by estimated goal distance. See PDF p. 19.

B · incorrect — Follows a path deeply before siblings. See PDF p. 19.

C · incorrect — Restricts DFS to a bound. See PDF p. 19.

D · correct — Exact layer-order expansion. See PDF p. 19.

Question 132

Which statement is not a correct characterization of reasoning?

  • A. Using knowledge to derive conclusions that can guide what a system does.
  • B. A process of thinking, logically arguing, and drawing inference
  • C. Theory of blind searching
  • D. The act of deriving a conclusion from certain premises using a given methodology

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI24 · Represent knowledge for inference

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: C — Theory of blind searching

Reasoning uses knowledge and a specified methodology to derive conclusions from premises. Logical argument and inference are therefore appropriate descriptions. Blind search is instead a search strategy that explores alternatives without heuristic guidance; it does not define reasoning as a whole. Clarifying the incomplete knowledge-system option into a claim about deriving conclusions leaves the intended search-versus-inference distinction as the unique exception.

Evidence in the complete study note

Knowledge representation encodes facts, objects, relations, events and rules so a machine can use them. Syntax defines legal expressions; semantics defines their meaning. A knowledge base stores assertions; an inference system derives consequences. Declarative knowledge says what is true; procedural knowledge specifies how to act.AI24
A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17

Choice by choice

A · incorrect — Using knowledge to derive conclusions is a valid characterization, so it is not the exception.

B · incorrect — Logical argument and drawing inferences describe reasoning, so this is not a false characterization for the negative question.

C · correct — Blind search is one search strategy rather than a definition of reasoning.

D · incorrect — Deriving a conclusion from premises using a methodology describes reasoning and does not satisfy the negative question.

Question 133

Which reasoning technique applies valid inference rules so that true premises guarantee a true conclusion?

  • A. Abductive reasoning
  • B. Analogical reasoning
  • C. Deductive reasoning
  • D. Inductive reasoning

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI28 · Deduction, induction and abduction

Study-guide references

Correct answer: C — Deductive reasoning

Deductive reasoning uses truth-preserving inference: when the premises are true and the rule is valid, the conclusion must be true. Rules in a broad sense are not sufficient to identify deduction, since many reasoning methods use procedures. Induction generalizes from cases, abduction proposes explanations, and analogy transfers conclusions across similarities; these do not supply the stated logical guarantee.

Evidence in the complete study note

Deduction applies truth-preserving rules: all people are mortal; Ada is a person; therefore Ada is mortal. Induction generalizes observations to a broader rule and can be overturned by new examples. Abduction proposes a plausible explanation: wet ground suggests rain, although sprinklers could explain it. Analogy transfers a pattern between similar situations.AI28
Classical logical entailment is monotonic: adding premises preserves previous consequences. Default reasoning can be nonmonotonic: infer a bird usually flies, then withdraw this when told it is a penguin. Common-sense defaults are useful with incomplete knowledge, but are not guaranteed truths. Blind search is a search strategy, not a definition of reasoning.AI28

Choice by choice

A · incorrect — Abduction proposes plausible explanatory hypotheses that can still be false.

B · incorrect — Analogy transfers a conclusion from a similar case without a truth-preservation guarantee.

C · correct — Valid deductive inference preserves truth from premises to conclusion.

D · incorrect — Induction generalizes beyond the observed cases and remains revisable.

Question 134

For A* tree search with an admissible heuristic and positive step costs, what is the intended benefit of using the heuristic alongside accumulated edge costs?

  • A. To seek a least-cost path while potentially reducing expansions through heuristic guidance.
  • B. To find the path with the maximum distance from the start to the goal node
  • C. To find paths with non-optimal searches from node to the goal node
  • D. To find the path with the maximum cost from the start to the goal node

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI18 · Admissibility, consistency and A*

Study-guide references

Correct answer: A — To seek a least-cost path while potentially reducing expansions through heuristic guidance.

A* adds accumulated cost g to an admissible remaining-cost estimate h. Under the stated tree-search conditions, this supports least-cost solutions, while useful heuristic guidance can reduce expansions. It does not guarantee a speed improvement for every problem or heuristic. Maximizing distance or cost is a different objective, and deliberately seeking nonoptimal paths is not the benefit described.

Evidence in the complete study note

A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
Admissibility suffices for optimal A* tree search; graph search that never reopens CLOSED states normally requires consistency. With an admissible inconsistent heuristic, reopen states when g improves. Test a goal when removed at minimum f. Completeness also needs finite branching and positive lower-bounded costs in infinite spaces. Worst-case time and memory can be exponential; no heuristic universally guarantees speed.AI18

Choice by choice

A · correct — An admissible remaining-cost estimate supports optimal tree search and can guide exploration toward useful candidates.

B · incorrect — A* minimizes total path cost; it does not seek maximum path length.

C · incorrect — Nonoptimality is not the purpose of admissible A* tree search.

D · incorrect — A* seeks minimum, not maximum, total cost.

Question 135

What role does machine learning play in understanding human perception?

  • A. Machine learning helps simulate human perception through neural networks
  • B. Machine learning has no relevance to understanding human perception
  • C. Machine learning is not capable of analyzing perceptual data
  • D. Machine learning only focuses on logical reasoning and decision-making SYSTEMS

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI34 · Neural networks and deep learning

AI39 · Perception, vision and robotics

Study-guide references

Correct answer: A — Machine learning helps simulate human perception through neural networks

ANNs model brain-inspired processing and learn pattern recognition; combined with p67 recognition applications this supports the perception-modeling role. Evidence: PDF 71, 71, 67.

Evidence in the complete study note

Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34
Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39

Choice by choice

A · correct — Matches brain-inspired models applied to recognition. See PDF p. 71, 71, 67.

B · incorrect — Contradicts documented pattern-recognition applications. See PDF p. 71, 71, 67.

C · incorrect — Contradicts recognition/classification capabilities. See PDF p. 71, 71, 67.

D · incorrect — Machine learning also handles sensory patterns, not only logic. See PDF p. 71, 71, 67.

Question 136

An informed search keeps one priority frontier ordered by g(n)+h(n), permanently closes states, and tests goals on removal. With a consistent heuristic, finite branching and positive lower-bounded step costs, which algorithm is this?

  • A. Iterative Deepening A* Search
  • B. Greedy Best-First Search
  • C. Uniform Cost Search
  • D. A* Search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI18 · Admissibility, consistency and A*

AI16 · Depth limits, iterative deepening and two directions

AI17 · Heuristics and greedy best-first search

AI15 · Uniform-cost search and Dijkstra

Study-guide references

Correct answer: D — A* Search

A* uses a best-first priority frontier ordered by g+h. Consistency provides the standard guarantee for graph search that permanently closes states, with the stated cost and goal-test conditions. IDA* also uses f values but organizes repeated depth-first threshold searches instead. Greedy search omits g, and UCS omits h. The frontier organization distinguishes the algorithms rather than a worst-case complexity shared by several methods.

Evidence in the complete study note

A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
Admissibility suffices for optimal A* tree search; graph search that never reopens CLOSED states normally requires consistency. With an admissible inconsistent heuristic, reopen states when g improves. Test a goal when removed at minimum f. Completeness also needs finite branching and positive lower-bounded costs in infinite spaces. Worst-case time and memory can be exponential; no heuristic universally guarantees speed.AI18
Bidirectional search grows from start and goal until frontiers connect. With suitable reverse actions, meeting tests and equal-cost BFS, time and space may fall to O(b^(d/2)). Weighted variants require a sound stopping bound; the first intersection need not be cheapest. IDA* deepens an f=g+h threshold rather than only a depth limit.AI16
Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
UCS uses a priority queue ordered by actual accumulated path cost g(n).AI15

Choice by choice

A · incorrect — IDA* repeats depth-first searches under increasing f thresholds rather than retaining this single priority frontier.

B · incorrect — Greedy best-first uses h alone and does not have the stated cost-aware priority.

C · incorrect — UCS uses g alone, without the heuristic term.

D · correct — A* uses g+h; consistency supports optimality when graph states are closed without reopening.

Question 137

Which general term denotes an entity that perceives its environment through sensors and acts on it through effectors?

  • A. API
  • B. Expert System
  • C. Intelligence
  • D. Agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

AI40 · Expert systems and choosing methods

AI-S06 · Prediction, transfer and explicit representations

Study-guide references

Correct answer: D — Agent

An agent is defined by perceiving an environment and acting on it through sensors and effectors. The acting mechanism may be physical or software-based. An API can expose operations used by an agent, and an expert system can contribute domain reasoning, but neither names the general entity defined here. Intelligence is a capability rather than the entity itself.

Evidence in the complete study note

An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40
An application programming interface (API) exposes defined operations through which software requests functionality. It can be an interface used by an agent, but the interface alone is not the general entity that perceives an environment and selects actions.AI-S06

Choice by choice

A · incorrect — An API exposes software operations; it may support an agent’s actions but is not the general perceiving-and-acting entity.

B · incorrect — An expert system is a specialized knowledge-based application, not the general term for every sensing-and-acting entity.

C · incorrect — Intelligence is a capability, not the name of the defined entity.

D · correct — Agent is the general term for an entity with this perception-and-action interaction.

Question 138

A new dataset has no target labels. The objective is to discover its groups or structure, with neither interaction rewards nor reuse of a pretrained model. Which learning regime applies?

  • A. Reinforcement learning
  • B. Unsupervised learning
  • C. Transfer learning
  • D. Supervised learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI32 · Unsupervised, semi-supervised and active learning

AI31 · Supervised learning and prediction

AI33 · Reinforcement learning

AI-S06 · Prediction, transfer and explicit representations

Study-guide references

Correct answer: B — Unsupervised learning

The objective is structure discovery without target labels, which defines unsupervised learning. The stem separately excludes action-reward interaction and reuse of a pretrained model. Those exclusions distinguish reinforcement feedback and transfer learning from this task. Transfer learning describes a reuse strategy rather than a label-availability regime and can be combined with different learning objectives.

Evidence in the complete study note

Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Transfer learning reuses representations or a model learned on an earlier task for a new task, for example adapting pretrained image features to new image classes. This is a reuse strategy, not a specification that the new task has no labels or rewards.AI-S06

Choice by choice

A · incorrect — Reinforcement learning requires a reward-based interaction objective, which is excluded.

B · correct — Unsupervised learning discovers structure in data without supplied target labels.

C · incorrect — Transfer learning concerns reuse of prior learned representations or models; no such reuse is present.

D · incorrect — Supervised learning requires target outputs paired with the inputs.

Question 139

One of the following statement is correct about Artificial Intelligence (Al)?

  • A. It refers to the development of machines that can perform tasks only in a controlled laboratory setting.
  • B. It refers to the development of machines that can only perform simple, repetitive tasks.
  • C. Itrefers to the development of machines that can only perform physical tasks.
  • D. It refers to the development of machines that can perform tasks that typically require human intelligence.

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI01 · AI capabilities and scope

AI39 · Perception, vision and robotics

Study-guide references

Correct answer: D — It refers to the development of machines that can perform tasks that typically require human intelligence.

AI designs intelligence in artificial devices and includes language, learning, reasoning and real-world action, not the restrictive only-claims. Evidence: PDF 4, 6.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39

Choice by choice

A · incorrect — Real-world robots contradict laboratory-only. See PDF p. 4, 6.

B · incorrect — Reasoning and complex problem solving exceed repetition. See PDF p. 4, 6.

C · incorrect — Language/reasoning are not merely physical. See PDF p. 4, 6.

D · correct — Matches design of artificial intelligence. See PDF p. 4, 6.

Question 140

A techniques for generating plans with conditionals and loops are almost identical to those for generating programs from logical specifications is called__

  • A. Automatic learning
  • B. Automatic goal testing
  • C. Automatic monitoring
  • D. Automatic programming

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI11 · Formulate before searching

AI08 · Learning-agent components

Study-guide references

Correct answer: D — Automatic programming

Automatic programming, also called program synthesis, constructs a program that satisfies a specification. Generating conditional plans or loops is analogous because the output must implement the required behavior. Learning changes performance through experience, goal testing checks whether a candidate succeeds, and monitoring checks execution. None of those by itself names the specification-to-program construction task.

Evidence in the complete study note

In route finding, states are cities, actions are roads, successors are neighboring cities, and path cost is accumulated distance or time. A solution is a legal path ending at a goal; an optimal solution has minimum cost. Automatic programming or program synthesis searches for a program satisfying a specification; conditional plans and loops require richer representations than a fixed action list.AI11
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08
First formulate the goal; then formulate the problem by specifying initial state, legal actions, transition model, goal test, and action costs. Search selects a sequence; execution applies it. A state abstracts away irrelevant world details.AI11

Choice by choice

A · incorrect — Learning improves performance through experience rather than specifically synthesizing a program from logical requirements.

B · incorrect — A goal test checks whether a candidate meets the goal; it does not generate a program from its specification.

C · incorrect — Monitoring checks an executing system rather than generating a program from a specification.

D · correct — Automatic programming or program synthesis constructs a program meeting a specification, including conditional or looping behavior.

Question 141

Which choices describe broad intelligence-related goals of AI rather than implementation means or hardware capacity? Select all that apply.

  • A. Emulating human intelligence
  • B. Creating machines that can think and learn
  • C. Developing advanced algorithms
  • D. Enhancing computer processing power

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI01 · AI capabilities and scope

AI43 · Clarify overlapping learning and AI labels

Study-guide references

Correct answers (select all): A, B — Emulating human intelligence; Creating machines that can think and learn

AI can aim to emulate aspects of human intelligence or to build systems that reason and learn. Those are compatible broad goals, although different AI approaches evaluate humanlike behavior and rational performance differently. Advanced algorithms are means that may serve many fields, while additional processing power is hardware capacity rather than intelligence itself. The distinction therefore permits both intelligence-related objectives.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
Face recognition is both an applied task and a narrow capability. Broad machine-intelligence goals can include emulating human intelligence and building systems that learn. Humanlike problem-solving alone does not uniquely select expert systems. Python is widely used, but C++ and Java also implement AI; “commonly used” does not isolate one language. McCarthy is associated with naming AI, not its sole invention.AI43

Choice by choice

A · correct — Humanlike thought and behavior are established AI approaches and fit this broad goal.

B · correct — Building systems with intelligent reasoning and learning is another supported broad AI objective.

C · incorrect — Advanced algorithms are implementation means, not the defining intelligence goal.

D · incorrect — More processing power alone does not establish intelligent behavior or learning.

Question 142

What is knowledge representation in artificial intelligence?

  • A. The process of storing information in a computer
  • B. The ability of machines to reason and make decisions
  • C. The study of how humans acquire knowledge
  • D. The process of encoding knowledge for machine use

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI24 · Represent knowledge for inference

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: D — The process of encoding knowledge for machine use

Knowledge representation organizes and formalizes knowledge in a machine-usable knowledge base. Storage alone omits the encoding/formalization purpose. Evidence: PDF 62.

Evidence in the complete study note

Knowledge representation encodes facts, objects, relations, events and rules so a machine can use them. Syntax defines legal expressions; semantics defines their meaning. A knowledge base stores assertions; an inference system derives consequences. Declarative knowledge says what is true; procedural knowledge specifies how to act.AI24
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · incorrect — Too broad: mere storage need not represent usable knowledge. See PDF p. 62.

B · incorrect — Reasoning uses representations but is not representation itself. See PDF p. 62.

C · incorrect — Human acquisition is a different subject. See PDF p. 62.

D · correct — Best match to formal machine-usable encoding. See PDF p. 62.

Question 143

Which representation system matches IF-THEN production rules against working memory and applies a selected rule to derive a conclusion or action?

  • A. Semantic networks
  • B. Neural networks
  • C. Production systems
  • D. Bayesian networks

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI24 · Represent knowledge for inference

AI29 · Probability and uncertain models

AI34 · Neural networks and deep learning

Study-guide references

Correct answer: C — Production systems

A production system uses IF-THEN rules whose conditions are matched against working memory; an applicable rule is selected and applied. Semantic networks encode concept relationships, neural networks learn weighted representations, and Bayesian networks encode probabilistic dependencies. All can support reasoning in broader systems, so the distinctive feature here is the explicit production-rule cycle rather than inference in general.

Evidence in the complete study note

A semantic network connects concept nodes by labeled relations such as is-a. Frames describe stereotyped entities through slots, values and defaults. Production systems use IF conditions THEN conclusions or actions, matching rules against working memory and selecting among applicable rules. A graph suits relationships such as frequently bought together; frames suit attributes. No representation is always best for recommendation without task requirements.AI24
A Bayesian network is a directed acyclic graph with conditional probability tables; its joint distribution factors as the product of each node’s probability given its parents. A Markov model assumes the next state depends on the current state rather than the entire past. These represent uncertainty; fuzzy membership represents a different issue.AI29
Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34

Choice by choice

A · incorrect — Semantic networks represent concepts and labeled relationships, not this working-memory production-rule cycle.

B · incorrect — Neural networks compute through weighted units and learned representations rather than the specified rule-matching architecture.

C · correct — Production systems match condition–action or condition–conclusion rules against working memory.

D · incorrect — Bayesian networks represent conditional probabilistic dependencies rather than the stated production-rule mechanism.

Question 144

Which reasoning technique involves drawing conclusions from specific observations or examples?

  • A. Deductive reasoning
  • B. Inductive reasoning
  • C. Abductive reasoning
  • D. Probabilistic reasoning

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI28 · Deduction, induction and abduction

AI29 · Probability and uncertain models

Study-guide references

Correct answer: B — Inductive reasoning

Specific facts generalized into a conclusion define induction. The conclusion is supported rather than logically guaranteed by the finite examples. Evidence: PDF 53.

Evidence in the complete study note

Deduction applies truth-preserving rules: all people are mortal; Ada is a person; therefore Ada is mortal. Induction generalizes observations to a broader rule and can be overturned by new examples. Abduction proposes a plausible explanation: wet ground suggests rain, although sprinklers could explain it. Analogy transfers a pattern between similar situations.AI28
Probability describes uncertainty about events. P(¬A)=1−P(A); P(A|B)=P(A∩B)/P(B), for P(B)>0. Bayes gives P(H|E)=P(E|H)P(H)/P(E): posterior equals likelihood times prior divided by evidence probability. For mutually exclusive exhaustive hypotheses, P(E)=ΣiP(E|Hi)P(Hi).AI29

Choice by choice

A · incorrect — Typically general premises to specific conclusion. See PDF p. 53.

B · correct — Exact specific-to-general pattern. See PDF p. 53.

C · incorrect — Seeks most likely explanation for observations. See PDF p. 53.

D · incorrect — Uses probability to express uncertainty, not this defining direction. See PDF p. 53.

Question 145

A classifier is trained on labeled examples and predicts by following attribute tests from a root to a class-labeled leaf. Which learning technique matches this representation?

  • A. Reinforcement learning
  • B. Unsupervised learning
  • C. Deep learning
  • D. Decision tree learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI31 · Supervised learning and prediction

AI43 · Clarify overlapping learning and AI labels

AI34 · Neural networks and deep learning

Study-guide references

Correct answer: D — Decision tree learning

A decision tree predicts by testing an attribute at each internal node, following a branch and returning the leaf prediction. Labeled training makes this a supervised classification task. Deep learning can also be supervised, so supervision alone would not distinguish it. The representation in the stem provides the distinction; reinforcement and unsupervised learning describe different feedback regimes.

Evidence in the complete study note

Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
A decision tree tests an attribute at each internal node, follows the corresponding branch and predicts at a leaf. ID3 chooses high information gain: H(S)=−Σpᵢlog₂pᵢ; gain is parent entropy minus weighted child entropy. Gini impurity is 1−Σpᵢ², zero for a pure class. Pruning limits overfitting. Neural networks, support-vector machines and trees can all be supervised.AI31
ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43
Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34

Choice by choice

A · incorrect — Reinforcement learning specifies environmental reward feedback rather than this attribute-test representation.

B · incorrect — Unsupervised learning seeks structure without supplied target labels.

C · incorrect — Deep learning uses multiple neural representation layers; it can be supervised but is not the attribute-test-and-leaf model described.

D · correct — Decision tree learning produces attribute-test branches with predictions at leaves.

Question 146

What is the primary objective of natural language processing (NLP) in artificial intelligence?

  • A. Understanding and generating human language by machines
  • B. Analyzing and processing large datasets
  • C. Enhancing computer hardware capabilities
  • D. Simulating human-like decision-making

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI37 · Language processing tasks

Study-guide references

Correct answer: A — Understanding and generating human language by machines

Understanding/interpretation is defined in NLP, and NLG turns data into language, supporting both understanding and generation. Evidence: PDF 79, 79.

Evidence in the complete study note

NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37

Choice by choice

A · correct — Matches understanding plus NLG. See PDF p. 79, 79.

B · incorrect — Large-data processing is broader than language. See PDF p. 79, 79.

C · incorrect — Not a hardware discipline. See PDF p. 79, 79.

D · incorrect — Too broad; omits the defining language objective. See PDF p. 79, 79.

Question 147

Which model family assumes that the distribution of the next state depends on the current state rather than the entire earlier history?

  • A. Decision trees
  • B. Markov models
  • C. Support vector machines
  • D. Reinforcement learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI29 · Probability and uncertain models

AI31 · Supervised learning and prediction

AI33 · Reinforcement learning

AI-S06 · Prediction, transfer and explicit representations

Study-guide references

Correct answer: B — Markov models

The Markov assumption makes the current state sufficient for the next-state distribution, rather than requiring the full preceding history. That identifies the model family. Decision trees and support-vector machines are prediction methods with other defining structures. Reinforcement learning is a reward-based learning framework that can use a Markov model, but it is not itself the particular conditional-dependence assumption.

Evidence in the complete study note

A Bayesian network is a directed acyclic graph with conditional probability tables; its joint distribution factors as the product of each node’s probability given its parents. A Markov model assumes the next state depends on the current state rather than the entire past. These represent uncertainty; fuzzy membership represents a different issue.AI29
A decision tree tests an attribute at each internal node, follows the corresponding branch and predicts at a leaf. ID3 chooses high information gain: H(S)=−Σpᵢlog₂pᵢ; gain is parent entropy minus weighted child entropy. Gini impurity is 1−Σpᵢ², zero for a pure class. Pruning limits overfitting. Neural networks, support-vector machines and trees can all be supervised.AI31
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Support-vector machines include supervised classification and regression methods. They use support vectors in their decision function rather than being defined by a next-state conditional-dependence assumption. In estimation, variance measures how predictions change across different training samples; overfitting instead names poor generalization caused by fitting training-specific details.AI-S06

Choice by choice

A · incorrect — Decision trees are defined by attribute tests and leaf predictions rather than this conditional transition assumption.

B · correct — A Markov model conditions the next state on the current state instead of the full previous history.

C · incorrect — Support-vector machines use support vectors for prediction; a next-state conditional-dependence assumption does not define the family.

D · incorrect — Reinforcement learning specifies improvement from rewards and may use different model families.

Question 148

Which learning regime is defined by an agent trying actions, receiving environmental rewards and improving a policy to maximize expected cumulative return?

  • A. Reinforcement learning
  • B. Unsupervised learning
  • C. Deep learning
  • D. Genetic algorithms

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI33 · Reinforcement learning

AI43 · Clarify overlapping learning and AI labels

AI20 · Local optimization and evolutionary search

AI32 · Unsupervised, semi-supervised and active learning

Study-guide references

Correct answer: A — Reinforcement learning

Reinforcement learning improves an action policy through rewards received during interaction, with expected cumulative return as the objective. Generic trial and error is too broad because evolutionary search also evaluates and improves candidates. Genetic algorithms are identified by population selection and variation, while deep learning identifies architecture. Unsupervised learning concerns unlabeled structure rather than this sequential reward objective.

Evidence in the complete study note

An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43
A genetic algorithm maintains a population, evaluates fitness, selects parents, applies crossover and mutation, and repeats replacement. It searches without guaranteeing global optimality. Means–ends analysis selects a difference between current and goal states, chooses an operator reducing it, and recursively solves its preconditions. Neither method is simply A* running backward.AI20
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features.AI32

Choice by choice

A · correct — Reinforcement learning is defined by this action–reward policy-improvement objective.

B · incorrect — Unsupervised learning discovers unlabeled structure rather than defining cumulative-return optimization.

C · incorrect — Deep learning specifies multilayer neural representations and can be used within several learning regimes.

D · incorrect — Genetic algorithms improve candidate populations using selection and genetic variation; trial and error alone does not make them the stated action–reward learning regime.

Question 149

Which type of learning algorithm aims to mimic the structure and function of the human brain?

  • A. Supervised learning
  • B. Unsupervised learning
  • C. Deep learning
  • D. Reinforcement learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI34 · Neural networks and deep learning

AI43 · Clarify overlapping learning and AI labels

Study-guide references

Correct answer: C — Deep learning, as a multilayer brain-inspired neural architecture.

AI34 calls ANNs brain-inspired abstractions and describes deep learning as using multiple representation layers. AI43 stresses that deep models can use supervised, unsupervised or reinforcement objectives. Thus deep learning is the offered architecture-based answer, while the other choices classify feedback regimes rather than requiring brain-inspired structure. This is inspiration, not a claim of literal brain simulation.

Evidence in the complete study note

Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34
ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43

Choice by choice

A · incorrect — Supervised learning specifies labeled targets and can use non-neural models such as trees or SVMs.

B · incorrect — Unsupervised learning specifies unlabeled structure discovery and does not require a neural architecture.

C · correct — Deep learning uses multilayer neural representations within the brain-inspired ANN family.

D · incorrect — Reinforcement learning specifies reward-based interaction and need not use brain-inspired neural structure.

Question 150

What is the primary purpose of expert systems in artificial intelligence?

  • A. Emulate human expertise in a specific domain
  • B. Perform complex mathematical calculations
  • C. Process and analyze large datasets
  • D. Optimize computer hardware performance

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: A — Emulate human expertise in a specific domain

Expert systems solve complex problems in a particular domain at the level of human expertise. Evidence: PDF 60.

Evidence in the complete study note

An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · correct — Exact domain-expertise purpose. See PDF p. 60.

B · incorrect — May be a tool, not defining purpose. See PDF p. 60.

C · incorrect — May be a workload, not defining purpose. See PDF p. 60.

D · incorrect — Not the defining expert-system purpose. See PDF p. 60.

Question 151

Within the cognitive approach to AI, which listed objective expresses the use of computational models to investigate human intelligence and perception?

  • A. Replicate human intelligence and perception
  • B. Enhance computer processing power
  • C. Analyze large datasets efficiently
  • D. Improve natural language processing

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI01 · AI capabilities and scope

AI02 · Four approaches and the Turing test

Study-guide references

Correct answer: A — Replicate human intelligence and perception

Cognitive AI uses computational models to investigate biological intelligence, including human perception and thought. Replicating relevant aspects of these processes expresses that modeling objective, provided the model is compared with observations. Greater processing power and data efficiency are resources or methods. Language processing is a narrower task and need not aim to explain biological intelligence.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
Thinking humanly models human mental processes; thinking rationally derives conclusions by valid logic. Acting humanly seeks humanlike behavior; acting rationally selects actions expected to achieve objectives. Thought processes and observed behavior are different evaluation targets.AI02

Choice by choice

A · correct — Modeling or replicating aspects of intelligence and perception is the cognitive investigation objective.

B · incorrect — Faster hardware enables computation but does not itself supply a model of intelligence.

C · incorrect — Efficient dataset analysis is a method or workload, not the specific cognitive-modeling objective.

D · incorrect — NLP is one application area and does not by itself express the broader intelligence-and-perception investigation.

Question 152

Which AI approach uses computational models to investigate human thought processes and compares those models with human behavior?

  • A. Machine learning,
  • B. Natural language processing
  • C. Cognitive AI (cognitive modeling)
  • D. Robotics

Study-guide references

Correct answer: C — Cognitive AI (cognitive modeling)

Cognitive AI investigates biological intelligence through computational models, testing those models against human behavior or experimental observations. This is a scientific modeling purpose rather than simply useful automation. Machine learning, NLP and robotics may contribute tools to cognitive research, but none is defined by that purpose. The option now uses the explicit approach name taught in the note.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
Thinking humanly models human mental processes; thinking rationally derives conclusions by valid logic. Acting humanly seeks humanlike behavior; acting rationally selects actions expected to achieve objectives. Thought processes and observed behavior are different evaluation targets.AI02
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39
Learning is improvement on task T, measured by P, through experience E.AI08

Choice by choice

A · incorrect — Machine learning improves performance from experience and need not model human thought processes.

B · incorrect — NLP focuses on language interpretation and generation rather than necessarily explaining human cognition.

C · correct — Cognitive AI uses computational models to investigate biological intelligence and human thought.

D · incorrect — Robotics integrates physical sensing and action; it need not aim to explain human thought.

Question 153

Which area of AI research focuses on understanding and replicating human vision?

  • A. Natural language processing
  • B. Computer vision
  • C. Robotics
  • D. Expert systems

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI39 · Perception, vision and robotics

AI37 · Language processing tasks

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: B — Computer vision

Computer vision extracts and comprehends image information, making it the specific vision field. Evidence: PDF 84.

Evidence in the complete study note

Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · incorrect — Language rather than image interpretation. See PDF p. 84.

B · correct — Exact visual-comprehension field. See PDF p. 84.

C · incorrect — Robots may use vision but robotics is broader. See PDF p. 84.

D · incorrect — Domain expertise rather than vision specifically. See PDF p. 84.

Question 154

Which AI technique is used to extract insights and patterns from large datasets?

  • A. Natural language processing
  • B. Machine learning
  • C. Expert systems
  • D. Robotics

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI08 · Learning-agent components

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: B — Machine learning

Machine learning applications include data mining/knowledge discovery and learning patterns; ANN example p71 specifically extracts complex patterns and trends. Evidence: PDF 67, 71.

Evidence in the complete study note

Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08
Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · incorrect — Language-specific tools do not define general data learning. See PDF p. 67, 71.

B · correct — Matches learning/discovery from data. See PDF p. 67, 71.

C · incorrect — Knowledge-based domain expertise rather than general learned patterns. See PDF p. 67, 71.

D · incorrect — Physical agents rather than the broad data-analysis technique. See PDF p. 67, 71.

Question 155

Which concept in AI refers to the ability of machines to understand and interpret human language?

  • A. Natural language processing
  • B. Machine learning
  • C. Expert systems
  • D. Robotics

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI37 · Language processing tasks

AI08 · Learning-agent components

AI40 · Expert systems and choosing methods

AI39 · Perception, vision and robotics

Study-guide references

Correct answer: A — Natural language processing

NLP is expressly machine understanding, analysis and interpretation of human languages. Evidence: PDF 79.

Evidence in the complete study note

NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39

Choice by choice

A · correct — Exact language-understanding concept. See PDF p. 79.

B · incorrect — Can support NLP but is broader learning from data. See PDF p. 79.

C · incorrect — Domain expertise is not language understanding generally. See PDF p. 79.

D · incorrect — Physical embodied action is a different focus. See PDF p. 79.

Question 156

Which AI technique is used to simulate human-like conversation and interaction?

  • A. Machine learning
  • B. Natural language processing
  • C. Robotics
  • D. Expert systems

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI37 · Language processing tasks

AI08 · Learning-agent components

AI40 · Expert systems and choosing methods

AI39 · Perception, vision and robotics

Study-guide references

Correct answer: B — Natural language processing

The module explicitly identifies chatbots providing customer chat as an NLP application. Evidence: PDF 82.

Evidence in the complete study note

NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39

Choice by choice

A · incorrect — May power a chatbot, but not the specific application field. See PDF p. 82.

B · correct — Explicit conversation application field. See PDF p. 82.

C · incorrect — Embodiment not required for conversation. See PDF p. 82.

D · incorrect — Expert advice is not the defining conversation technique. See PDF p. 82.

Question 157

What is the specifically computational-modeling contribution of cognitive AI to studying human thought, as distinct from collecting measurements or running psychological experiments alone?

  • A. By analyzing human brain activity
  • B. By replicating human thought processes
  • C. By conducting psychological experiments
  • D. By improving computer hardware capabilities

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI01 · AI capabilities and scope

AI02 · Four approaches and the Turing test

Study-guide references

Correct answer: B — By replicating human thought processes

Cognitive AI contributes computational models that implement proposed mechanisms of human thought. Their predictions can then be compared with behavior, response times, errors or other observations. Brain measurements and psychological experiments can provide valuable evidence, but collecting evidence alone is distinct from constructing the computational model. Faster hardware similarly enables work without supplying a cognitive explanation.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
Thinking humanly models human mental processes; thinking rationally derives conclusions by valid logic. Acting humanly seeks humanlike behavior; acting rationally selects actions expected to achieve objectives. Thought processes and observed behavior are different evaluation targets.AI02
The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02

Choice by choice

A · incorrect — Brain measurements can provide evidence, but measurement alone does not construct the computational thought model requested.

B · correct — Replicating aspects of thought in a computational model provides a theory whose behavior can be tested.

C · incorrect — Psychological experiments can test a model, but experiments alone do not supply the computational representation asked for.

D · incorrect — More capable hardware supplies resources without itself representing a theory of human thought.

Question 158

Which field builds integrated physical agents that sense their surroundings and act autonomously through motors or other actuators?

  • A. Machine learning
  • B. Robotics
  • C. Natural language processing
  • D. Expert systems

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI39 · Perception, vision and robotics

AI08 · Learning-agent components

AI37 · Language processing tasks

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: B — Robotics

Robotics builds physical systems that sense, plan and act, combining mechanical, electrical, control and computing components. The requirement for integrated physical sensing and actuation distinguishes it from software autonomy in general. Machine learning, NLP and expert systems can be components of a robot, but none by itself names the field responsible for the complete physical agent.

Evidence in the complete study note

Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · incorrect — Machine learning provides experience-based improvement but does not itself define construction of physical sensing-and-acting systems.

B · correct — Robotics integrates the mechanical, electrical, control and computing systems of physical agents.

C · incorrect — NLP processes human language and does not require a physical body.

D · incorrect — Expert systems supply specialized knowledge-based reasoning and need not control physical actuators.

Question 159

When studying how people recognize objects, what is the computational-modeling contribution of cognitive AI, rather than building a sensor or gathering biological and behavioral measurements alone?

  • A. By simulating human sensory organs
  • B. By analyzing human brain waves
  • C. By replicating human perceptual processes
  • D. By conducting psychological experiments

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI01 · AI capabilities and scope

AI02 · Four approaches and the Turing test

AI39 · Perception, vision and robotics

Study-guide references

Correct answer: C — By replicating human perceptual processes

Cognitive AI can implement proposed perceptual processes and compare the resulting recognition behavior with human observations. This is distinct from only building an input sensor or collecting brain-wave and behavioral measurements. Those activities can inform or test the model, but they do not alone provide its computational account of perception. The scope is modeling the recognition process, not claiming that other research methods lack value.

Evidence in the complete study note

Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02
Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39

Choice by choice

A · incorrect — A sensory-organ simulation models an input mechanism, but does not by itself explain how observed patterns are recognized.

B · incorrect — Brain-wave measurements provide data rather than the computational recognition model requested.

C · correct — Replicating aspects of perceptual processing supplies a model whose recognition behavior can be tested.

D · incorrect — Psychological experiments provide tests or observations; by themselves they do not construct the computational perception model.

Question 160

Which AI technique solves specialized domain problems by applying expert knowledge encoded in a knowledge base through an inference engine?

  • A. Machine learning
  • B. Expert systems
  • C. Natural language processing
  • D. Robotics

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI40 · Expert systems and choosing methods

AI08 · Learning-agent components

AI37 · Language processing tasks

AI39 · Perception, vision and robotics

Study-guide references

Correct answer: B — Expert systems

An expert system applies specialized domain knowledge through an inference engine operating on a knowledge base and working facts. This is more specific than merely exhibiting intelligent problem solving. Machine learning emphasizes improvement from experience, NLP concerns language, and robotics combines sensing, planning and physical action. Those methods may be combined with expert systems, but they are not the technique defined by the stated knowledge-based architecture.

Evidence in the complete study note

An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39

Choice by choice

A · incorrect — Machine learning fits behavior from experience rather than defining this expert-knowledge architecture.

B · correct — Expert systems use a knowledge base and inference engine for specialized domain reasoning.

C · incorrect — NLP concerns human language rather than this domain-expertise architecture.

D · incorrect — Robotics concerns physical sensing and action rather than the stated knowledge-base technique.

Question 161

What is a heuristic in the context of artificial intelligence (AI)?

  • A. A predefined set of rules for decision making
  • B. A mathematical model used to solve search problems
  • C. An intelligent agent capable of learning from experience
  • D. A strategy or rule of thumb used to guide problem-solving

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: D — A strategy or rule of thumb used to guide problem-solving

A heuristic guides search toward promising states using estimated closeness; it is a problem-solving guide rather than a complete learning agent or guaranteed exact solver. Evidence: PDF 24, 24.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17

Choice by choice

A · incorrect — A predefined rule could implement a heuristic, but not all decision rules are heuristics. See PDF p. 24, 24.

B · incorrect — A heuristic need not be a complete mathematical search model. See PDF p. 24, 24.

C · incorrect — A learning agent is an entity, not the guiding strategy. See PDF p. 24, 24.

D · correct — Closest to the module’s guiding-way definition. See PDF p. 24, 24.

Question 162

How are heuristics useful in search problems?

  • A. They guarantee finding the optimal solution.
  • B. They reduce the search space and speed up the search process.
  • C. They eliminate the need for search algorithms.
  • D. They ensure a systematic exploration of the search space.

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI18 · Admissibility, consistency and A*

Study-guide references

Correct answer: B — They reduce the search space and speed up the search process.

Problem knowledge guides search toward promising states, often reducing explored nodes and time. It does not universally guarantee optimality or replace a search algorithm. Evidence: PDF 24.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
Admissibility suffices for optimal A* tree search; graph search that never reopens CLOSED states normally requires consistency. With an admissible inconsistent heuristic, reopen states when g improves. Test a goal when removed at minimum f. Completeness also needs finite branching and positive lower-bounded costs in infinite spaces. Worst-case time and memory can be exponential; no heuristic universally guarantees speed.AI18

Choice by choice

A · incorrect — Module explicitly warns heuristic may not give best solution. See PDF p. 24.

B · correct — Matches guided efficient exploration. See PDF p. 24.

C · incorrect — A heuristic is used within informed search. See PDF p. 24.

D · incorrect — Systematic exploration alone does not distinguish heuristic guidance. See PDF p. 24.

Question 163

Which general characterization says that search uses problem-specific knowledge to estimate remaining cost, without itself requiring either non-overestimation or an edge-by-edge consistency inequality?

  • A. Admissible heuristic
  • B. Consistent heuristic
  • C. Informed heuristic
  • D. Dominance comparison between heuristics

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI18 · Admissibility, consistency and A*

AI-S05 · Related search and game-evaluation methods

Study-guide references

Correct answer: C — Informed heuristic

Informed search adds problem-specific knowledge to guide exploration, often through a remaining-cost estimate. That general characterization does not guarantee admissibility or consistency, which impose additional mathematical conditions. Dominance instead compares two estimates pointwise. The question asks for the broad knowledge-use property rather than one of those stronger conditions or comparisons.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
For two admissible heuristics, h₂ dominates h₁ when h₂(n)≥h₁(n) for every node. Dominance compares estimates; it is not itself a condition ensuring that an arbitrary heuristic never overestimates. Both estimates must separately satisfy the relevant admissibility assumptions.AI-S05

Choice by choice

A · incorrect — Admissibility adds a non-overestimation bound, stronger than the general knowledge-use characterization requested.

B · incorrect — Consistency adds an edge-by-edge inequality, stronger than the general property requested.

C · correct — Informed heuristic guidance uses problem-specific knowledge to estimate remaining cost or promise.

D · incorrect — Dominance compares the numerical estimates of two heuristics rather than naming general use of problem knowledge.

Question 164

Which term names the heuristic property defined by 0≤h(n)≤h*(n), where h*(n) is the true cheapest remaining cost?

  • A. Admissible heuristic
  • B. Consistent heuristic
  • C. Informed heuristic
  • D. Dominance comparison between heuristics

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI18 · Admissibility, consistency and A*

AI-S05 · Related search and game-evaluation methods

Study-guide references

Correct answer: A — Admissible heuristic

The displayed non-overestimation inequality defines admissibility. Consistency is a different edge condition; with the usual nonnegative heuristic estimates and a zero goal estimate, it implies admissibility for states with a goal path. Thus a consistent heuristic can also satisfy the bound, but consistency is not the name of the property being defined. Informedness and dominance do not independently establish the bound.

Evidence in the complete study note

A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
For two admissible heuristics, h₂ dominates h₁ when h₂(n)≥h₁(n) for every node. Dominance compares estimates; it is not itself a condition ensuring that an arbitrary heuristic never overestimates. Both estimates must separately satisfy the relevant admissibility assumptions.AI-S05

Choice by choice

A · correct — Admissibility is exactly the non-overestimation property stated by this inequality.

B · incorrect — Consistency is an edge-by-edge inequality with goal normalization, rather than the name of the displayed non-overestimation bound.

C · incorrect — Being informed only means using problem knowledge and does not impose this bound.

D · incorrect — Dominance compares two heuristic estimates and does not by itself bound either one by the true remaining cost.

Question 165

Which forward best-first algorithm ranks frontier nodes by accumulated path cost plus a heuristic estimate of remaining cost?

  • A. Depth-first search
  • B. Breadth-first search
  • C. A* search algorithm
  • D. Hill climbing algorithm

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI18 · Admissibility, consistency and A*

AI13 · Breadth-first search

AI14 · Depth-first search

AI20 · Local optimization and evolutionary search

Study-guide references

Correct answer: C — A* search algorithm

A* orders frontier nodes by f(n) = g(n) + h(n), combining the cost already paid with an estimate of the cost still needed. Estimating future cost is not a backward search frontier. DFS and BFS use depth-related removal orders, while hill climbing improves neighboring complete candidates. Replacing the mistaken backward-search phrase preserves the intended heuristic-and-cost algorithm identification.

Evidence in the complete study note

A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14
Local search keeps a complete candidate configuration and improves its objective; the path may be irrelevant. Hill climbing chooses better neighbors. First-improvement accepts the first improvement; steepest-ascent examines all neighbors; stochastic variants choose among improvements. Local maxima, plateaus and ridges can stop progress. Random restarts or limited sideways steps can help without guaranteeing success.AI20

Choice by choice

A · incorrect — DFS follows a stack order and does not inherently combine a heuristic with backward search.

B · incorrect — BFS uses breadth order without requiring heuristics or a backward frontier.

C · correct — A* is forward best-first search using f = g + h.

D · incorrect — Hill climbing selects improving neighbors rather than ordering a frontier by accumulated plus remaining path cost.

Question 166

What is the main advantage of using heuristics in game playing AI systems?

  • A. They guarantee victory in every game.
  • B. They make the game unpredictable and exciting.
  • C. They allow the AI to learn from its opponents.
  • D. They improve the AI's decision-making and gameplay.

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI23 · Alpha–beta and game evaluation

AI22 · Games and minimax

Study-guide references

Correct answer: D — They improve the AI's decision-making and gameplay.

Game search computes candidate moves with a heuristic and chooses the best backed-up outcome, supporting improved decision quality rather than automatic victory or learning. Evidence: PDF 33, 34.

Evidence in the complete study note

At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23
Adversarial search models agents with competing objectives. A game specifies states, player to move, legal actions, successor function, terminal test and utility. Perfect information means previous moves and current state are visible; it does not reveal future choices or imply absence of chance. Zero-sum utility makes one player’s gain the other’s loss.AI22

Choice by choice

A · incorrect — Optimal opponent may prevent a win. See PDF p. 33, 34.

B · incorrect — Unpredictability is not the stated purpose. See PDF p. 33, 34.

C · incorrect — Heuristic evaluation is not necessarily learning. See PDF p. 33, 34.

D · correct — Matches heuristic move evaluation and selection. See PDF p. 33, 34.

Question 167

Which component is the scoring mapping that assigns a heuristic desirability value to a game state, rather than a search, pruning or sampling procedure?

  • A. Minimax algorithm
  • B. Alpha-beta pruning
  • C. Monte Carlo simulation
  • D. Evaluation function

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI23 · Alpha–beta and game evaluation

AI22 · Games and minimax

AI-S05 · Related search and game-evaluation methods

Study-guide references

Correct answer: D — Evaluation function

An evaluation function assigns a desirability score to a game state, especially when search stops at a depth cutoff. Minimax uses scores to choose moves, and alpha–beta avoids branches that cannot alter the backed-up result. Monte Carlo sampling can estimate values from possible continuations, but the sampling procedure is distinct from the mapping whose output is a state score.

Evidence in the complete study note

At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23
Alpha is MAX’s best lower bound; beta is MIN’s best upper bound. Pass bounds down, update alpha at MAX and beta at MIN, and prune when alpha≥beta. Pruning preserves the minimax value for the same searched tree. Good move ordering increases pruning: worst-case O(b^m), ideal ordering about O(b^(m/2)).AI23
Adversarial search models agents with competing objectives. A game specifies states, player to move, legal actions, successor function, terminal test and utility. Perfect information means previous moves and current state are visible; it does not reveal future choices or imply absence of chance. Zero-sum utility makes one player’s gain the other’s loss.AI22
A Monte Carlo game simulation samples possible continuations and uses their outcomes to estimate position value. An evaluation function is the scoring mapping itself; sampling is one way of obtaining an estimate, while minimax backs values up and alpha–beta removes branches that cannot affect that backed-up decision.AI-S05

Choice by choice

A · incorrect — Minimax backs values up through competing choices; it uses state scores rather than being the score mapping itself.

B · incorrect — Alpha–beta pruning skips branches that cannot affect a minimax decision; it is not the scoring mapping.

C · incorrect — Monte Carlo simulation samples continuations whose outcomes may estimate value; sampling is a procedure, not the score mapping itself.

D · correct — An evaluation function maps a state to its estimated desirability score.

Question 168

What is the purpose of using heuristics in game tree search algorithms?

  • A. To guarantee a win in every game.
  • B. To eliminate the need for search algorithms.
  • C. To reduce the search space and improve efficiency.
  • D. To ensure a systematic exploration of the search tree.

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI23 · Alpha–beta and game evaluation

Study-guide references

Correct answer: C — To reduce the search space and improve efficiency.

Pruning removes nodes that cannot affect the decision and speeds game search; heuristic domain-informed ordering improves pruning effectiveness. This supports efficiency, not guaranteed wins or eliminating search. Evidence: PDF 35, 35.

Evidence in the complete study note

Alpha is MAX’s best lower bound; beta is MIN’s best upper bound. Pass bounds down, update alpha at MAX and beta at MIN, and prune when alpha≥beta. Pruning preserves the minimax value for the same searched tree. Good move ordering increases pruning: worst-case O(b^m), ideal ordering about O(b^(m/2)).AI23
At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23

Choice by choice

A · incorrect — A better search can still lose to stronger positions. See PDF p. 35, 35.

B · incorrect — Pruning/search guidance still requires search. See PDF p. 35, 35.

C · correct — Matches reducing irrelevant exploration. See PDF p. 35, 35.

D · incorrect — Systematic traversal does not describe the efficiency purpose. See PDF p. 35, 35.

Question 169

Which game-state heuristic evaluates the weighted number and value of pieces held by each player, rather than legal move count or controlled territory?

  • A. Material heuristic
  • B. Mobility heuristic
  • C. Distance heuristic
  • D. Tactical heuristic

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI23 · Alpha–beta and game evaluation

Study-guide references

Correct answer: A — Material heuristic

Material evaluation counts pieces with weights reflecting their values. Mobility instead measures legal options, while territory evaluates controlled regions. Restricting the stem to weighted pieces removes the original conflation of material and territory. Distance and tactical features may inform other evaluations, but they do not name the weighted piece-count feature requested. The distinction concerns what the evaluator measures.

Evidence in the complete study note

At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23

Choice by choice

A · correct — Material is the weighted piece-count/value feature.

B · incorrect — Mobility measures available legal moves rather than the amount of material or territory controlled.

C · incorrect — Distance evaluates spatial separation, not the piece-count or controlled-region feature stated here.

D · incorrect — This does not name weighted piece holdings; the feature described is material evaluation.

Question 170

Which heuristic evaluates a game state by considering the available moves and options for each player?

  • A. Material heuristic
  • B. Mobility heuristic
  • C. Distance heuristic
  • D. Tactical heuristic

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI23 · Alpha–beta and game evaluation

Study-guide references

Correct answer: B — The mobility heuristic.

AI23 explicitly says mobility counts legal options. Available moves and options for each player are exactly that feature, making mobility the supported answer. Material measures weighted pieces, while distance and tactics refer to different aspects of a position. The heuristic may be combined with other features, but the specific feature requested is unambiguous.

Evidence in the complete study note

At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23

Choice by choice

A · incorrect — Material evaluates the amount or value of pieces, not the number of available moves.

B · correct — Mobility directly counts or scores the legal move options available to a player.

C · incorrect — Distance concerns separation or proximity, not the legal-option count described in the question.

D · incorrect — Tactical evaluation concerns concrete threats or sequences rather than the specific mobility feature requested.

Question 171

________ is a set of procedures that uses representational language to conclude new facts from existing ones.

  • A. Inference Engine
  • B. Explanatory facility
  • C. Interface
  • D. Knowledge Base

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI40 · Expert systems and choosing methods

AI24 · Represent knowledge for inference

Study-guide references

Correct answer: A — Inference Engine

Inference engine manipulates knowledge and repeatedly applies rules to facts to derive new knowledge. The knowledge base stores facts, while the user interface handles interaction/explanation. Evidence: PDF 63, 63, 64.

Evidence in the complete study note

An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40
Knowledge representation encodes facts, objects, relations, events and rules so a machine can use them. Syntax defines legal expressions; semantics defines their meaning. A knowledge base stores assertions; an inference system derives consequences. Declarative knowledge says what is true; procedural knowledge specifies how to act.AI24

Choice by choice

A · correct — Matches procedural derivation of new facts. See PDF p. 63, 63, 64.

B · incorrect — Explaining an outcome differs from deriving facts. See PDF p. 63, 63, 64.

C · incorrect — Interaction layer differs from inference procedures. See PDF p. 63, 63, 64.

D · incorrect — Stores knowledge rather than the inference procedure. See PDF p. 63, 63, 64.

Question 172

Which one is a space where an agent perceives and acts, so that an intelligent agent designer has to understand the type of it?

  • A. Environment
  • B. Actuator
  • C. Sensor
  • D. Search space

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI03 · Agents, percepts and action

AI12 · Frontiers and evaluation criteria

Study-guide references

Correct answer: A — Environment

Environment is the surrounding space in which an agent operates and senses/acts; sensors and actuators are interfaces to it. Evidence: PDF 13.

Evidence in the complete study note

An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03
A percept is one observation; a percept sequence is the entire observation history. An agent function maps percept sequences to actions: f:P*→A. An agent program implements that mapping on an architecture; agent = architecture + program. Environment means the setting in which it senses and acts, not merely the search space. Autonomy means relying substantially on experience rather than only a designer’s prior instructions.AI03
A search node records a state plus parent, action, depth and path cost. Expanding a node generates its successors. OPEN or frontier stores discovered candidates awaiting expansion; CLOSED records expanded states. Reaching the same state through different paths does not imply the same node or cost.AI12

Choice by choice

A · correct — Exact surrounding operating context. See PDF p. 13.

B · incorrect — Action-output component. See PDF p. 13.

C · incorrect — Perceptual-input component. See PDF p. 13.

D · incorrect — Abstract candidate-solution set, not necessarily the operating surroundings. See PDF p. 13.

Question 173

Which one of the following is an uninformed searching method?

  • A. A* search
  • B. Iterative improvement
  • C. Breadth first search
  • D. Greedy search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI13 · Breadth-first search

AI17 · Heuristics and greedy best-first search

AI20 · Local optimization and evolutionary search

Study-guide references

Correct answer: C — Breadth first search

The listed uninformed strategies include BFS. A* and best-first are separately listed informed strategies; local iterative improvement is not the cited uninformed breadthwise procedure. Evidence: PDF 17, 17.

Evidence in the complete study note

BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
Local search keeps a complete candidate configuration and improves its objective; the path may be irrelevant. Hill climbing chooses better neighbors. First-improvement accepts the first improvement; steepest-ascent examines all neighbors; stochastic variants choose among improvements. Local maxima, plateaus and ridges can stop progress. Random restarts or limited sideways steps can help without guaranteeing success.AI20

Choice by choice

A · incorrect — Informed heuristic method. See PDF p. 17, 17.

B · incorrect — Local improvement is not the listed uninformed method. See PDF p. 17, 17.

C · correct — Explicit uninformed method. See PDF p. 17, 17.

D · incorrect — Informed heuristic best-first method. See PDF p. 17, 17.

Question 174

Which one of the following searching strategies best explains: 'The searching agent is based on problem-specific knowledge and the agent knows how far away a different state is from the goal'?

  • A. Blind search
  • B. Informed search
  • C. Uninformed search
  • D. Not informed search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: B — Informed search

Problem-specific knowledge of estimated distance to the goal is the module’s informed/heuristic-search distinction. Blind search lacks such guidance. Evidence: PDF 24, 18.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17

Choice by choice

A · incorrect — Blind search omits goal-closeness knowledge. See PDF p. 24, 18.

B · correct — Exact informed-search distinction. See PDF p. 24, 18.

C · incorrect — Lacks the stated domain guidance. See PDF p. 24, 18.

D · incorrect — Synonymous with uninformed/blind search. See PDF p. 24, 18.

Question 175

Which type of AI is capable of performing any intellectual task that a human can do?

  • A. Narrow AI
  • B. General AI
  • C. Machine Learning
  • D. Expert Systems

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI01 · AI capabilities and scope

AI08 · Learning-agent components

AI40 · Expert systems and choosing methods

Study-guide references

Correct answer: B — General AI.

AI01 defines general AI by broad competence transferable across domains, matching the hypothetical ability to perform any human intellectual task. Narrow AI is bounded; machine learning names a family of methods and expert systems supply specialized domain knowledge. Neither automatically gives universal intellectual capability. The statement describes a category, not a claim that such a system presently exists.

Evidence in the complete study note

Narrow or weak AI addresses bounded tasks, such as face recognition, spam filtering or a voice assistant. General AI denotes broad competence transferable across domains; strong AI also has a philosophical meaning involving actual understanding. Do not equate a humanoid body, autonomous vehicle, or many features with general intelligence.AI01
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08
An expert system applies specialized knowledge to a domain such as equipment diagnosis. The domain expert supplies expertise; the knowledge engineer elicits and encodes it; the inference engine reasons over a knowledge base and working facts. The user interface handles interaction; an explanation facility presents why or how a conclusion was reached.AI40

Choice by choice

A · incorrect — Narrow AI is limited to bounded tasks rather than broad intellectual transfer.

B · correct — General AI is the broad cross-domain competence category described.

C · incorrect — Machine learning is a method family and does not itself imply general intellectual competence.

D · incorrect — Expert systems apply specialized domain knowledge, not arbitrary human intellectual capability.

Question 176

What is the primary goal of Machine Learning?

  • A. To hard-code rules into a machine
  • B. To allow systems to learn and improve from experience without being explicitly programmed
  • C. To simulate physical environments
  • D. To build robotic hardware

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI08 · Learning-agent components

AI-S01 · Programmed computation and learned patterns

Study-guide references

Correct answer: B — Improve task performance through experience.

AI08 defines learning as improvement on a task, measured by performance, through experience. B captures that adaptive objective; without explicitly programmed means the learned behavior is not simply an exhaustive hand-coded rule list, not that no algorithm is programmed. Simulation and robot construction can use ML but do not define its primary goal.

Evidence in the complete study note

Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08
A conventional computer manipulates digital information according to a program or sequence of instructions. Programs can implement storage, retrieval and learning operations. Neural networks can learn pattern recognition from examples, so executing instructions and learning patterns are not mutually exclusive descriptions. This provides no universal claim that every machine is better or worse than every human at completing missing information.AI-S01

Choice by choice

A · incorrect — Hard-coding all task rules is not the experience-based improvement defining machine learning.

B · correct — The system improves measured task performance from experience rather than only fixed task rules.

C · incorrect — Physical simulation is one application context, not the definition or primary learning objective.

D · incorrect — Building hardware is not required for machine learning, which also operates in software.

Question 177

Which reasoning type guarantees that if premises are true, the conclusion is also true?

  • A. Inductive Reasoning
  • B. Abductive Reasoning
  • C. Deductive Reasoning
  • D. Probabilistic Reasoning

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI28 · Deduction, induction and abduction

AI29 · Probability and uncertain models

Study-guide references

Correct answer: C — Deductive reasoning.

AI28 defines deduction by truth-preserving rules and contrasts induction’s revisable generalizations and abduction’s plausible explanations. With valid deductive inference and true premises, the conclusion must be true. Probabilistic support expresses uncertainty rather than this logical guarantee, so deductive is the relevant reasoning type.

Evidence in the complete study note

Deduction applies truth-preserving rules: all people are mortal; Ada is a person; therefore Ada is mortal. Induction generalizes observations to a broader rule and can be overturned by new examples. Abduction proposes a plausible explanation: wet ground suggests rain, although sprinklers could explain it. Analogy transfers a pattern between similar situations.AI28
Probability describes uncertainty about events. P(¬A)=1−P(A); P(A|B)=P(A∩B)/P(B), for P(B)>0. Bayes gives P(H|E)=P(E|H)P(H)/P(E): posterior equals likelihood times prior divided by evidence probability. For mutually exclusive exhaustive hypotheses, P(E)=ΣiP(E|Hi)P(Hi).AI29

Choice by choice

A · incorrect — Induction generalizes beyond observations and can be overturned by new examples.

B · incorrect — Abduction proposes plausible explanations without guaranteeing that they are true.

C · correct — Valid deductive rules preserve truth from true premises to the conclusion.

D · incorrect — Probability can quantify uncertain conclusions rather than provide the stated deductive guarantee.

Question 178

A self-driving car's environment is best classified as:

  • A. Fully Observable, Deterministic, Static
  • B. Partially Observable, Stochastic, Dynamic
  • C. Fully Observable, Stochastic, Static
  • D. Partially Observable, Deterministic, Dynamic

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: B — Partially observable, stochastic and dynamic.

AI09 describes driving as partially observable, sequential and dynamic. Traffic participants and noisy observations make outcomes uncertain rather than fully determined by the car’s chosen action. B matches that ordinary driving setting. Neither full visibility nor a static world fits roads whose relevant conditions change while the agent deliberates.

Evidence in the complete study note

Fully observable means sensors reveal relevant state; partially observable means relevant information is hidden or noisy. Deterministic means state and action determine the next state; stochastic means outcomes involve randomness. Strategic uncertainty about an opponent is distinct from dice-like randomness.AI09
Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · incorrect — Driving has hidden/noisy information and a changing world rather than full observability and static conditions.

B · correct — Hidden state, uncertain outcomes and ongoing traffic changes fit the ordinary driving task.

C · incorrect — Full observability and static conditions do not fit the stated driving environment.

D · incorrect — Partial observation and dynamics fit, but deterministic outcomes ignore the uncertainty in ordinary driving.

Question 179

In the A* algorithm, what does the evaluation function f(n) represent?

  • A. Only the heuristic cost to goal
  • B. Only the cost from start to node n
  • C. The sum of cost from start to n (g(n)) and estimated cost from n to goal (h(n))
  • D. The product of g(n) and h(n)

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI18 · Admissibility, consistency and A*

AI15 · Uniform-cost search and Dijkstra

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: C — f(n)=g(n)+h(n).

AI18 directly defines A* evaluation as actual start-to-node cost g plus estimated node-to-goal cost h. This estimates a solution cost through the node. Using only h is greedy best-first, using only g is uniform-cost selection, and multiplication is not the A* rule.

Evidence in the complete study note

A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15
Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17

Choice by choice

A · incorrect — Using h alone is greedy best-first rather than A* evaluation.

B · incorrect — Using g alone is uniform-cost ordering and omits the remaining-cost estimate.

C · correct — A* adds the accumulated path cost and estimated remaining cost.

D · incorrect — The defined evaluation is a sum, not a product.

Question 180

Which reflex architecture maintains an internal world state and applies condition–action rules to that modeled state, rather than using only the current percept?

  • A. Simple Reflex Agent
  • B. Goal-Based Agent
  • C. Model-Based Reflex Agent
  • D. Learning Agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI06 · Simple and model-based reflex agents

AI07 · Goals, utility and model updates

AI08 · Learning-agent components

Study-guide references

Correct answer: C — Model-Based Reflex Agent

Model-based reflex agents update internal state from previous state, action and current observation, using models of world change and action effects. They then apply rules to that modeled state. Simple reflex agents lack persistent world-state memory. Goal-based and learning capabilities may coexist with a model, but the question specifically asks for the reflex architecture defined by state maintenance and rule application.

Evidence in the complete study note

A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean. It has no persistent world-state memory. This is adequate when the current observation contains everything needed for its rule, but partial observability can make the same percept require different actions.AI06
A model-based reflex agent maintains internal state using the previous state, previous action, current percept, and models of world change and action effects. It then applies rules to that state. Both kinds can use condition–action rules; memory and a world model distinguish them.AI06
A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07
The performance element selects actions. The learning element changes the agent to improve future performance. The critic evaluates behavior against a performance standard and supplies feedback. The problem generator suggests exploratory actions likely to yield useful experience.AI08

Choice by choice

A · incorrect — Simple reflex uses only the current percept without persistent world-state memory.

B · incorrect — Goal-based choice uses future sequences reaching desired states, whereas the requested architecture applies reflex rules to modeled state.

C · correct — Model-based reflex maintains internal state and applies rules to that state.

D · incorrect — Learning concerns improving the agent from experience rather than naming this reflex architecture.

Question 181

What distinguishes a Rational Agent from an Omniscient Agent?

  • A. Rational agents always make perfect decisions
  • B. Rational agents act on best available information but can make mistakes; omniscient agents know all outcomes
  • C. Omniscient agents are a common AI type
  • D. Rational agents do not use information

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI04 · Rationality and performance

Study-guide references

Correct answer: B — Rational choice uses available information; omniscience knows actual outcomes.

AI04 defines rationality by best expected performance using information available, while omniscience means knowing actual outcomes in advance. Rational actions can fail and do not require perfect predictions. The contrast is epistemic, not whether an agent uses any information or whether omniscient systems are common implementations.

Evidence in the complete study note

A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04
Omniscience would mean knowing actual action outcomes in advance. Rationality does not require omniscience or guaranteed success. Even complete outcome knowledge cannot create a successful action when every available action fails. Information gathering and exploration may therefore be rational actions.AI04

Choice by choice

A · incorrect — Rationality does not guarantee perfect outcomes or foreknowledge.

B · correct — This distinguishes information-conditioned expected performance from knowing actual outcomes.

C · incorrect — The note treats omniscience as an idealized outcome-knowledge condition, not a common AI implementation.

D · incorrect — Rationality explicitly depends on percepts, prior knowledge and available actions.

Question 182

Which search algorithm is guaranteed to find the shortest path in an unweighted graph?

  • A. Depth-First Search
  • B. Depth-Limited Search
  • C. Breadth-First Search
  • D. Best-First Search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI13 · Breadth-first search

AI14 · Depth-first search

AI16 · Depth limits, iterative deepening and two directions

Study-guide references

Correct answer: C — Breadth-first search.

AI13 says BFS explores levels in increasing depth and minimizes the number of edges. An unweighted graph gives equal step costs, making that shortest-path property applicable with visited-state tracking. DFS or a fixed depth limit does not guarantee minimum depth, and generic best-first needs its evaluation rule specified.

Evidence in the complete study note

BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14
DLS runs DFS to depth ℓ and reports cutoff separately from failure. It misses goals deeper than ℓ. Iterative deepening runs DLS with limits 0,1,2,…; for finite b it finds the shallowest solution with O(b^d) time and O(bd) frontier space, optimal for equal step costs. Repeated shallow work is small compared with the deepest level.AI16

Choice by choice

A · incorrect — DFS can reach a longer path first and is not generally optimal.

B · incorrect — A depth limit can miss a solution or return a nonshortest path within its bound.

C · correct — BFS explores increasing edge counts and finds shortest paths in an unweighted graph.

D · incorrect — Best-first is a family; an unspecified evaluation rule does not guarantee shortest paths.

Question 183

Why is Uniform Cost Search (UCS) NOT considered a heuristic approach?

  • A. It doesn't consider costs
  • B. It uses cost as a known factor but employs no domain-specific knowledge or estimates
  • C. It uses heuristic functions
  • D. It only works on trees, not graphs

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI15 · Uniform-cost search and Dijkstra

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: B — UCS uses known accumulated cost without a heuristic estimate.

AI15 orders UCS by actual path cost g, and AI17 explicitly calls it uninformed despite using g. Known action costs come from the problem definition; a heuristic adds an estimate of future cost. UCS can operate on graphs with duplicate handling, so neither ignoring costs nor tree-only operation explains the distinction.

Evidence in the complete study note

UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15
A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17

Choice by choice

A · incorrect — UCS explicitly orders its frontier by accumulated path cost.

B · correct — Known costs are used, but no problem-specific estimate of remaining cost is required.

C · incorrect — UCS does not require an h estimate; g is actual accumulated cost.

D · incorrect — UCS works on graphs with proper relaxation and duplicate handling.

Question 184

Which data structure supplies the defining next-node removal order of a depth-first search frontier?

  • A. Queue (FIFO)
  • B. Stack (LIFO)
  • C. Priority Queue
  • D. Hash Table

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI14 · Depth-first search

AI13 · Breadth-first search

AI15 · Uniform-cost search and Dijkstra

Study-guide references

Correct answer: B — Stack (LIFO)

DFS follows the deepest pending branch using a LIFO stack or recursion. That removal discipline distinguishes it from breadth-first FIFO selection and cost- or heuristic-priority selection. A hash table can still be used alongside DFS to record visited states. The question therefore specifies frontier removal rather than broadly asking which structures may track nodes.

Evidence in the complete study note

DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15

Choice by choice

A · incorrect — A FIFO queue removes the oldest entry, giving breadth-first rather than depth-first frontier order.

B · correct — A LIFO stack processes the most recently pending branch first, implementing DFS.

C · incorrect — A priority queue removes according to a key rather than the defining DFS LIFO order.

D · incorrect — A hash table may store visited states, but does not supply the frontier’s next-node removal order.

Question 185

What is the main limitation of Depth-First Search regarding completeness?

  • A. It uses too much memory
  • B. It can get stuck in cycles or infinite paths and fail to find a solution
  • C. It always finds suboptimal solutions
  • D. It only works for trees

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI14 · Depth-first search

Study-guide references

Correct answer: B — Cycles or infinite branches can prevent finding an existing solution.

AI14 says DFS can fail on infinite paths or cycles and is not generally optimal. Without suitable restrictions or duplicate handling, it may never return to a pending branch containing a goal. Finite visited-state graph traversal changes that result; the limitation is therefore not that DFS only works on trees or always returns a bad path.

Evidence in the complete study note

DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14
Implicit-tree worst-case time is O(b^m), and space is O(bm) when pending siblings are stored, or O(m) for path-only recursive generation. DFS is not generally optimal and can fail on infinite paths or cycles. On a finite graph, visited-state tracking makes traversal complete but needs O(V) storage. Do not confuse the O(b^m) worst-case time bound with the DFS space bounds above.AI14

Choice by choice

A · incorrect — DFS is usually memory-efficient; excess storage is not its defining completeness failure.

B · correct — It may follow an infinite branch or cycle indefinitely while a solution waits elsewhere.

C · incorrect — A returned path can be optimal by coincidence; always suboptimal is too strong.

D · incorrect — DFS also traverses graphs, especially with visited-state tracking.

Question 186

Which options can be used without a training set of labeled input–target examples? Select all that apply. For D, consider whether some deep-learning objectives permit unlabeled or reward feedback.

  • A. Supervised Learning
  • B. Reinforcement Learning
  • C. Unsupervised Learning
  • D. Deep Learning

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

AI33 · Reinforcement learning

AI43 · Clarify overlapping learning and AI labels

Study-guide references

Correct answers (select all): B, C, D — Reinforcement Learning; Unsupervised Learning; Deep Learning

Unsupervised learning discovers structure without target labels. Reinforcement learning can learn from rewards rather than correct-action target examples. Deep learning is an architecture family and can be trained with unsupervised or reinforcement objectives, so it also need not require labeled input-target data. Supervised learning is specifically defined by those labeled pairs. The capability wording permits both regimes and architecture families to be judged accurately.

Evidence in the complete study note

Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43

Choice by choice

A · incorrect — Supervised learning uses labeled examples by definition.

B · correct — RL learns through rewards and interaction, without requiring labeled correct actions.

C · correct — Unsupervised learning explicitly seeks structure without target labels.

D · correct — Deep learning can use unlabeled or reward objectives, so labels are not inherently required.

Question 187

For A* tree search with finite branching, step costs bounded below by a positive constant, goal testing on removal and h(goal)=0, which listed heuristic condition is the weakest sufficient guarantee of optimality? Treat monotone as synonymous with consistent.

  • A. Consistent
  • B. Admissible (never overestimates the actual cost)
  • C. Monotone
  • D. Domain-independent

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI18 · Admissibility, consistency and A*

Study-guide references

Correct answer: B — Admissible (never overestimates the actual cost)

Admissibility requires that a heuristic never overestimate the true cheapest remaining cost. It suffices for optimal A* tree search under the stated cost and goal-testing assumptions. Consistency, also called monotonicity here, is stronger and is particularly important when graph search permanently closes states. Domain independence is unrelated to the optimality condition. Asking for the weakest sufficient listed condition distinguishes admissibility from the stronger alternatives.

Evidence in the complete study note

A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
Admissibility suffices for optimal A* tree search; graph search that never reopens CLOSED states normally requires consistency. With an admissible inconsistent heuristic, reopen states when g improves. Test a goal when removed at minimum f. Completeness also needs finite branching and positive lower-bounded costs in infinite spaces. Worst-case time and memory can be exponential; no heuristic universally guarantees speed.AI18

Choice by choice

A · incorrect — Consistency is sufficient but stronger than the weakest listed condition for tree A*.

B · correct — Admissibility is sufficient for optimal tree A* under the stated assumptions.

C · incorrect — Monotonicity is consistency here and is stronger than necessary for the tree-search guarantee.

D · incorrect — Domain independence is not a heuristic optimality condition.

Question 188

Iterative Deepening Search (IDS) combines the advantages of which two algorithms?

  • A. UCS and A*
  • B. BFS and DFS
  • C. Greedy Search and BFS
  • D. DLS and UCS

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI16 · Depth limits, iterative deepening and two directions

AI13 · Breadth-first search

AI14 · Depth-first search

Study-guide references

Correct answer: B — Breadth-first solution depth with depth-first memory use.

AI16 defines IDS as DLS runs at limits 0,1,2,…, finding the shallowest solution under finite branching with O(bd) frontier space. It obtains BFS’s shallow-solution behavior and DFS-style memory use. It does not combine UCS priorities, greedy ordering or A* cost thresholds; IDA* is a separate threshold-based variant.

Evidence in the complete study note

DLS runs DFS to depth ℓ and reports cutoff separately from failure. It misses goals deeper than ℓ. Iterative deepening runs DLS with limits 0,1,2,…; for finite b it finds the shallowest solution with O(b^d) time and O(bd) frontier space, optimal for equal step costs. Repeated shallow work is small compared with the deepest level.AI16
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14

Choice by choice

A · incorrect — IDS does not combine UCS costs with A* evaluation.

B · correct — It obtains breadth-first shallow-solution behavior through repeated depth-first limited searches.

C · incorrect — No greedy heuristic ordering is required for IDS.

D · incorrect — DLS is used, but UCS ordering is not part of standard iterative deepening.

Question 189

A zero-sum game evaluator measures a position from the designated MAX player’s perspective. What does a positive evaluation value indicate?

  • A. A draw or neutral position.
  • B. An estimated advantage for the designated MAX player.
  • C. An estimated advantage for the opposing MIN player.
  • D. An invalid game state.

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI23 · Alpha–beta and game evaluation

Study-guide references

Correct answer: B — An estimated advantage for the designated MAX player.

Under this convention, positive values estimate an advantage for MAX, negative values favor MIN, and zero is neutral. A positive heuristic score does not guarantee an eventual win, because it may be only an estimate at a search cutoff. The reference player must be specified before a sign can be interpreted; player names alone do not determine the convention.

Evidence in the complete study note

At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23

Choice by choice

A · incorrect — A neutral result is represented by zero under the stated sign convention.

B · correct — A positive value favors the designated MAX player under this convention.

C · incorrect — An advantage for MIN would have a negative sign under this convention.

D · incorrect — The sign estimates desirability, not whether the position is legally valid.

Question 190

Which knowledge representation technique uses interconnected nodes and arcs to show concept relationships?

  • A. Frames
  • B. Logic
  • C. Semantic Networks
  • D. Probabilistic Models

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI24 · Represent knowledge for inference

AI29 · Probability and uncertain models

Study-guide references

Correct answer: C — Semantic networks.

AI24 defines semantic networks using concept nodes and labeled relations such as is-a. That matches the node/arc representation in the stem. Frames organize slots, logic uses formal sentences and probabilistic models attach uncertainty; those may have related graph forms, but the specific concept-relationship network term is semantic network.

Evidence in the complete study note

A semantic network connects concept nodes by labeled relations such as is-a. Frames describe stereotyped entities through slots, values and defaults. Production systems use IF conditions THEN conclusions or actions, matching rules against working memory and selecting among applicable rules. A graph suits relationships such as frequently bought together; frames suit attributes. No representation is always best for recommendation without task requirements.AI24
Knowledge representation encodes facts, objects, relations, events and rules so a machine can use them. Syntax defines legal expressions; semantics defines their meaning. A knowledge base stores assertions; an inference system derives consequences. Declarative knowledge says what is true; procedural knowledge specifies how to act.AI24
A Bayesian network is a directed acyclic graph with conditional probability tables; its joint distribution factors as the product of each node’s probability given its parents. A Markov model assumes the next state depends on the current state rather than the entire past. These represent uncertainty; fuzzy membership represents a different issue.AI29

Choice by choice

A · incorrect — Frames represent stereotyped entities through slots and values rather than this named concept-link network.

B · incorrect — Logic represents claims with formal expressions, not specifically the node/arc scheme described.

C · correct — Semantic networks connect concept nodes with labeled relationship arcs.

D · incorrect — Probabilistic models focus on uncertainty rather than the specific semantic-relation representation requested.

Question 191

Which listed group comprises unstructured input modalities from which deep neural models can learn hierarchical features?

  • A. Tabular numerical data only
  • B. Structured relational data
  • C. Unstructured data like images, sound, and text
  • D. Small labeled datasets

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI34 · Neural networks and deep learning

AI35 · Generalization and validation

Study-guide references

Correct answer: C — Unstructured data like images, sound, and text

Deep learning uses multiple representation layers and can learn features from images, sound and text. Those are the unstructured modalities identified here. Deep models can also process other data, so this is not a claim of universal superiority over every alternative model. Structured relational data is a different organization, and small labeled datasets describe size and supervision rather than modality.

Evidence in the complete study note

Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34
Overfitting means fitting training-specific details so performance drops on unseen data; underfitting means the model is too limited or poorly fitted even on training data. A learning curve plots performance against training size or training progress. Complexity control, regularization, pruning and more representative data can help.AI35

Choice by choice

A · incorrect — Tabular numerical data only excludes the non-tabular modalities requested and is not an inherent deep-learning restriction.

B · incorrect — Structured relational data is not the unstructured modality group asked for.

C · correct — Images, sound and text are explicit modalities for learned multilayer representations.

D · incorrect — A small labeled dataset describes sample size and feedback, not an input modality group.

Question 192

Which agent type improves its performance over time through experience and feedback?

  • A. Simple Reflex Agent
  • B. Model-Based Reflex Agent
  • C. Goal-Based Agent
  • D. Learning Agent

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI08 · Learning-agent components

Study-guide references

Correct answer: D — A learning agent.

AI08 defines the learning element as changing the agent to improve future performance from experience and feedback. This identifies the learning-agent category. Reflex, model-based and goal-based describe decision architectures that may be enhanced by learning, but do not by themselves require improvement over time; D explicitly names that capability.

Evidence in the complete study note

The performance element selects actions. The learning element changes the agent to improve future performance. The critic evaluates behavior against a performance standard and supplies feedback. The problem generator suggests exploratory actions likely to yield useful experience.AI08
Learning is improvement on task T, measured by P, through experience E. For spam filtering, T is classifying email, E is past labeled messages, and P might be recall at a required precision. Exploration trades immediate reward for information; exploitation uses the best action currently known. Learning can improve reflex, goal-based or utility-based agents.AI08

Choice by choice

A · incorrect — Simple reflex rules alone do not require adaptation from experience.

B · incorrect — Maintaining a model does not by itself require learning or improving it over time.

C · incorrect — Planning toward a goal does not itself establish experience-based policy improvement.

D · correct — Learning agents explicitly use experience and evaluative feedback to improve future behavior.

Question 193

What is abductive reasoning best described as?

  • A. Drawing certain conclusions from known facts
  • B. Making educated guesses from patterns
  • C. Finding the best explanation for a set of observations
  • D. Using probability to decide outcomes

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI28 · Deduction, induction and abduction

AI29 · Probability and uncertain models

Study-guide references

Correct answer: C — Finding a plausible best explanation for observations.

AI28 illustrates abduction with wet ground suggesting rain while allowing sprinklers as another cause. It proposes an explanatory hypothesis rather than guaranteeing truth or merely generalizing a pattern. C gives that explanatory role. Probability may rank hypotheses, but numerical probability alone is not the definition of abductive reasoning.

Evidence in the complete study note

Deduction applies truth-preserving rules: all people are mortal; Ada is a person; therefore Ada is mortal. Induction generalizes observations to a broader rule and can be overturned by new examples. Abduction proposes a plausible explanation: wet ground suggests rain, although sprinklers could explain it. Analogy transfers a pattern between similar situations.AI28
Probability describes uncertainty about events. P(¬A)=1−P(A); P(A|B)=P(A∩B)/P(B), for P(B)>0. Bayes gives P(H|E)=P(E|H)P(H)/P(E): posterior equals likelihood times prior divided by evidence probability. For mutually exclusive exhaustive hypotheses, P(E)=ΣiP(E|Hi)P(Hi).AI29

Choice by choice

A · incorrect — Certain truth-preserving conclusions describe valid deduction rather than explanatory abduction.

B · incorrect — Pattern-based generalization is closer to induction and omits the explanatory-hypothesis requirement.

C · correct — Abduction seeks a plausible explanation for observed evidence without guaranteeing that explanation.

D · incorrect — Probability can support a diagnosis but does not itself define the explanatory reasoning form.

Question 194

How is ordinary untimed chess classified with respect to transition outcomes, change during deliberation and state/action values? Assume both players’ moves are specified when evaluating a transition.

  • A. Stochastic, Dynamic, Continuous
  • B. Deterministic, Static, Discrete
  • C. Stochastic, Static, Discrete
  • D. Deterministic, Dynamic, Continuous

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: B — Deterministic, Static, Discrete

Chess has discrete positions and legal moves. Given the selected move, the resulting position is determined by the rules rather than random outcome sampling. In untimed play, the environment does not change while the player deliberates, so the relevant classification is static. Opponent choice is strategic uncertainty rather than dice-like stochasticity. Specifying untimed play excludes the semidynamic effect of a score-affecting clock.

Evidence in the complete study note

Fully observable means sensors reveal relevant state; partially observable means relevant information is hidden or noisy. Deterministic means state and action determine the next state; stochastic means outcomes involve randomness. Strategic uncertainty about an opponent is distinct from dice-like randomness.AI09
Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · incorrect — Chess moves are discrete, and opponent strategy is not dice-like stochasticity.

B · correct — Untimed chess is deterministic given chosen moves, static while deliberating and discrete.

C · incorrect — Ordinary chess transition outcomes are deterministic given chosen moves rather than stochastic.

D · incorrect — Chess is discrete rather than continuous, and a clock gives semidynamic rather than this classification.

Question 195

What does the 'deep' in Deep Learning refer to?

  • A. The depth of the training dataset
  • B. The depth of layers in a neural network
  • C. The complexity of the problem being solved
  • D. The deep analysis of heuristic functions

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI34 · Neural networks and deep learning

Study-guide references

Correct answer: B — Multiple neural representation layers.

AI34 defines deep learning through multiple representation layers in a neural model. Dataset size and problem difficulty can affect training, but neither is what the word deep denotes. Heuristic analysis belongs to search methods rather than the architecture depth of a neural network.

Evidence in the complete study note

Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34

Choice by choice

A · incorrect — Dataset size or depth does not define the deep part of deep learning.

B · correct — Deep refers to multiple layers of learned representations in the neural architecture.

C · incorrect — Problem difficulty does not determine whether a model architecture is deep.

D · incorrect — Heuristic analysis is not the neural-layer definition of deep learning.

Question 196

Which search strategy is called 'blind search' because it uses no extra problem knowledge?

  • A. Informed Search
  • B. Heuristic Search
  • C. Uninformed Search
  • D. Best-First Search

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI15 · Uniform-cost search and Dijkstra

Study-guide references

Correct answer: C — Uninformed search.

AI17 says uninformed methods use the problem definition and known costs without a problem-specific remaining-cost estimate. This is the usual blind-search meaning. Informed or heuristic search explicitly adds such estimates, and best-first is a broad ordering family that may use heuristic information rather than naming the blind category.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15

Choice by choice

A · incorrect — Informed search adds problem-specific guidance beyond the basic problem definition.

B · incorrect — Heuristic search uses the extra estimate excluded by the stem.

C · correct — Uninformed search is the category using only the problem definition and known costs.

D · incorrect — Best-first specifies evaluation-based ordering and is not generally synonymous with blind search.

Question 197

In Reinforcement Learning, what guides an agent to improve its behavior?

  • A. Labeled training examples
  • B. Hard-coded rules
  • C. Rewards and penalties from environment interactions
  • D. Pre-trained heuristic functions

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI33 · Reinforcement learning

AI31 · Supervised learning and prediction

Study-guide references

Correct answer: C — Reward feedback from interactions.

AI33 describes an agent acting, receiving reward and a new observation, and improving its policy to maximize expected cumulative reward. Rewards may be positive or negative and delayed; they need not name the correct action. Labeled examples, fixed rules and predefined heuristics do not provide this distinctive reinforcement feedback mechanism.

Evidence in the complete study note

An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31

Choice by choice

A · incorrect — Correct target labels characterize supervised learning rather than reinforcement feedback.

B · incorrect — Fixed rules alone do not provide the reward-based adaptation defining RL.

C · correct — Rewards and penalties evaluate outcomes and guide improvement of expected cumulative return.

D · incorrect — A pretrained heuristic may help implementation but is not the defining RL feedback signal.

Question 198

For greedy best-first search on an implicit finite tree with branching factor b and maximum depth m, what is the standard worst-case node-count time bound?

  • A. O(m)
  • B. O(b*m)
  • C. O(b^m)
  • D. O(log m)

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: C — O(b^m)

Greedy best-first ordering can be misled by a heuristic and may need to explore an exponentially large portion of the finite search tree. With branching factor b and depth m, the standard worst-case generated-node bound is O(b^m). This is a node-count convention rather than a claim about every priority-queue implementation or all best-first algorithms. The explicit algorithm and tree model determine the bound.

Evidence in the complete study note

Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17

Choice by choice

A · incorrect — A misleading heuristic can force exponentially many nodes, so a bound merely linear in maximum depth m is insufficient.

B · incorrect — O(bm) is a familiar DFS storage form, not this generic search-time guarantee.

C · correct — The specified implicit finite-tree worst case can require O(b^m) nodes.

D · incorrect — A logarithmic-in-depth bound is insufficient when the search can explore an exponential number of tree nodes.

Question 199

Which knowledge representation uses templates with slots to describe stereotyped situations?

  • A. Semantic Networks
  • B. Frames
  • C. Logic
  • D. Bayesian Networks

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI24 · Represent knowledge for inference

AI29 · Probability and uncertain models

Study-guide references

Correct answer: B — Frames.

AI24 defines frames as stereotyped entity descriptions with slots, values and defaults. This is the template-and-slot representation in the question. Semantic networks focus on relationships among concept nodes, formal logic expresses sentences and Bayesian networks represent probabilistic dependencies; none is the specific slot-template term requested.

Evidence in the complete study note

A semantic network connects concept nodes by labeled relations such as is-a. Frames describe stereotyped entities through slots, values and defaults. Production systems use IF conditions THEN conclusions or actions, matching rules against working memory and selecting among applicable rules. A graph suits relationships such as frequently bought together; frames suit attributes. No representation is always best for recommendation without task requirements.AI24
A Bayesian network is a directed acyclic graph with conditional probability tables; its joint distribution factors as the product of each node’s probability given its parents. A Markov model assumes the next state depends on the current state rather than the entire past. These represent uncertainty; fuzzy membership represents a different issue.AI29

Choice by choice

A · incorrect — Semantic networks emphasize concept nodes and labeled relations rather than the slot-template representation.

B · correct — Frames organize stereotyped descriptions into slots with values and defaults.

C · incorrect — Logic uses formal sentences rather than the specifically described frame slots.

D · incorrect — Bayesian networks represent conditional probabilities, not the named stereotyped slot template.

Question 200

What is Natural Language Processing (NLP) primarily focused on?

  • A. Visual image interpretation
  • B. Physical robot movement
  • C. Enabling machines to understand, interpret, and generate human language
  • D. Optimizing database queries

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI37 · Language processing tasks

AI39 · Perception, vision and robotics

Study-guide references

Correct answer: C — Understanding, interpreting and generating human language.

AI37 directly defines NLP as interpreting and generating human language and lists dialogue, translation and summarization. Visual interpretation belongs to computer vision, physical movement to robotics and query optimization to database processing. NLP can interact with those systems, but its primary language-processing domain remains distinct.

Evidence in the complete study note

NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39

Choice by choice

A · incorrect — Interpreting visual images is primarily computer vision rather than language processing.

B · incorrect — Physical robot movement concerns control and robotics, not the definition of NLP.

C · correct — Human-language interpretation and generation are the primary NLP tasks named in AI37.

D · incorrect — Database-query optimization is not the human-language-processing focus described.

Question 201

Which listed term is a reasoning method rather than one of the common learning regimes classified by availability of target labels?

  • A. Supervised Learning
  • B. Unsupervised Learning
  • C. Deductive reasoning
  • D. Semi-supervised Learning

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI31 · Supervised learning and prediction

AI32 · Unsupervised, semi-supervised and active learning

AI28 · Deduction, induction and abduction

Study-guide references

Correct answer: C — Deductive reasoning

Supervised, unsupervised and semi-supervised learning differ in the target labels available during training. Deductive reasoning instead derives consequences from premises using truth-preserving rules. Deduction can be used inside a learning system, so the distinction is not that the two can never coexist. The question separates a reasoning method from the conventional label-availability taxonomy.

Evidence in the complete study note

Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
Semi-supervised learning combines some labels with additional unlabeled data. Active learning chooses informative cases for an oracle to label; it still uses labels. “Exploratory learning” is not an exclusive standard synonym: it can mean unsupervised exploration or active data acquisition, so define the feedback. A task named summarization or time-series analysis is not automatically supervised; its training targets determine that.AI32
Deduction applies truth-preserving rules: all people are mortal; Ada is a person; therefore Ada is mortal. Induction generalizes observations to a broader rule and can be overturned by new examples. Abduction proposes a plausible explanation: wet ground suggests rain, although sprinklers could explain it. Analogy transfers a pattern between similar situations.AI28

Choice by choice

A · incorrect — Supervised learning uses labeled input–target pairs, so it is a label-availability learning regime.

B · incorrect — Unsupervised learning has no supplied targets and discovers structure, so it is a label-availability regime.

C · correct — Deduction derives truth-preserving consequences from premises rather than naming a label-availability regime.

D · incorrect — Semi-supervised learning combines labeled and unlabeled data, so it is a label-availability regime.

Question 202

In adversarial search, what makes it different from traditional search?

  • A. It only uses one agent
  • B. Multiple agents with conflicting goals compete in the same search space
  • C. It guarantees the optimal solution always
  • D. It does not use evaluation functions

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI22 · Games and minimax

AI23 · Alpha–beta and game evaluation

Study-guide references

Correct answer: B — Competing agents with conflicting objectives.

AI22 defines adversarial search by agents with competing objectives and gives minimax for opposition. That distinguishes it from a single-agent path search. It can use evaluation functions at cutoffs, and its best-outcome guarantee is conditional on the modeled game and opponent, not an unconditional guarantee that every implementation finds a globally optimal solution.

Evidence in the complete study note

Adversarial search models agents with competing objectives. A game specifies states, player to move, legal actions, successor function, terminal test and utility. Perfect information means previous moves and current state are visible; it does not reveal future choices or imply absence of chance. Zero-sum utility makes one player’s gain the other’s loss.AI22
At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23

Choice by choice

A · incorrect — A single decision-making agent does not capture the competing-objective feature of adversarial search.

B · correct — Competing agents and conflicting objectives are its defining distinction.

C · incorrect — Guarantees depend on the algorithm, model and search depth, not merely the adversarial label.

D · incorrect — Depth-limited adversarial search commonly uses heuristic evaluation functions.

Question 203

A Depth-Limited Search (DLS) is NOT complete in itself because:

  • A. It uses too much memory
  • B. Solutions beyond the preset depth limit will be missed
  • C. It explores nodes in random order
  • D. It requires a heuristic function

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI16 · Depth limits, iterative deepening and two directions

AI14 · Depth-first search

Study-guide references

Correct answer: B — A solution beyond the fixed depth limit is missed.

AI16 defines DLS as DFS stopped at depth ℓ and explicitly says it misses goals deeper than that. A fixed limit therefore does not establish completeness for arbitrary problem depths. This is independent of randomized ordering, heuristics or memory use; increasing limits through iterative deepening addresses the depth restriction under its stated assumptions.

Evidence in the complete study note

DLS runs DFS to depth ℓ and reports cutoff separately from failure. It misses goals deeper than ℓ. Iterative deepening runs DLS with limits 0,1,2,…; for finite b it finds the shallowest solution with O(b^d) time and O(bd) frontier space, optimal for equal step costs. Repeated shallow work is small compared with the deepest level.AI16
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14

Choice by choice

A · incorrect — Depth-limited DFS is memory-efficient; excessive memory is not the reason goals are missed.

B · correct — A preset cutoff prevents exploring a solution whose depth exceeds that limit.

C · incorrect — DLS need not use random order; the depth cutoff alone causes the limitation.

D · incorrect — No heuristic is required for depth-limited DFS.

Question 204

Which probabilistic model represents knowledge in terms of conditional probabilities among variables?

  • A. Semantic Networks
  • B. Frames
  • C. Bayesian Networks
  • D. Ontologies

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI29 · Probability and uncertain models

AI24 · Represent knowledge for inference

AI-S06 · Prediction, transfer and explicit representations

Study-guide references

Correct answer: C — Bayesian networks.

AI29 defines a Bayesian network as a DAG with conditional probability tables and joint factorization by parents. That is the probabilistic knowledge representation in the stem. Semantic networks and frames encode relations or slots without inherently supplying conditional distributions, and ontologies specify conceptual structure rather than this particular probabilistic factorization.

Evidence in the complete study note

A Bayesian network is a directed acyclic graph with conditional probability tables; its joint distribution factors as the product of each node’s probability given its parents. A Markov model assumes the next state depends on the current state rather than the entire past. These represent uncertainty; fuzzy membership represents a different issue.AI29
A semantic network connects concept nodes by labeled relations such as is-a. Frames describe stereotyped entities through slots, values and defaults. Production systems use IF conditions THEN conclusions or actions, matching rules against working memory and selecting among applicable rules. A graph suits relationships such as frequently bought together; frames suit attributes. No representation is always best for recommendation without task requirements.AI24
An ontology describes entities or individuals, classes and properties and may state logical relationships and restrictions. It organizes explicit conceptual knowledge. Conditional probability tables and a probabilistic factorization are additional modeling commitments, not inherent in the word ontology.AI-S06

Choice by choice

A · incorrect — Semantic networks encode labeled concept relations rather than inherently conditional probability distributions.

B · incorrect — Frames encode slots, values and defaults rather than the named probabilistic dependency structure.

C · correct — Bayesian networks explicitly store conditional probability relationships among variables.

D · incorrect — An ontology’s conceptual organization is not itself the conditional-probability model specified.

Question 205

What is an 'expanded node' in a search algorithm?

  • A. A node that has been visited but not explored further
  • B. A node selected for further exploration whose neighboring nodes have been generated
  • C. A node removed from the open list
  • D. A node that has reached the goal state

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI12 · Frontiers and evaluation criteria

AI13 · Breadth-first search

Study-guide references

Correct answer: B — A node whose successors have been generated.

AI12 defines expansion as generating successors. B states that action, unlike mere visitation, removal from OPEN or reaching a goal. A node can be removed and goal-tested without expanding it, and discovered frontier nodes await expansion. The criterion is what was generated, not whether the node was seen or selected.

Evidence in the complete study note

A search node records a state plus parent, action, depth and path cost. Expanding a node generates its successors. OPEN or frontier stores discovered candidates awaiting expansion; CLOSED records expanded states. Reaching the same state through different paths does not imply the same node or cost.AI12
BFS removes the oldest frontier entry using a FIFO queue. Start with the root, remove the head, test it, append unseen successors at the tail, and repeat. It explores depths 0,1,2,… and is complete for finite b and a finite-depth goal. It minimizes number of edges, so minimum cost requires equal positive step costs.AI13

Choice by choice

A · incorrect — Visiting without generating successors does not meet the definition of expansion.

B · correct — Generating the selected node’s neighboring successor nodes is expansion.

C · incorrect — Removal alone may precede a goal test without generating any successors.

D · incorrect — Goal status is not the definition of expansion; a goal may be accepted without expanding it.

Question 206

For which type of problem would Supervised Learning with CNNs be most appropriate?

  • A. Customer segmentation from unlabeled data
  • B. Classifying a large labeled image dataset
  • C. Learning chess actions from rewards obtained through trial-and-error play.
  • D. Finding hidden patterns in transaction data

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI31 · Supervised learning and prediction

AI34 · Neural networks and deep learning

AI32 · Unsupervised, semi-supervised and active learning

AI33 · Reinforcement learning

Study-guide references

Correct answer: B — Classifying a large labeled image dataset

The labeled images provide input–target pairs for supervised classification, and CNN filters exploit spatial image structure. Unlabeled customer segmentation and transaction-pattern discovery seek structure without class targets. Learning chess actions from rewards is a reinforcement objective.

Evidence in the complete study note

Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34
Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33

Choice by choice

A · incorrect — Unlabeled segmentation seeks groups rather than supervised image classes.

B · correct — Labeled images supply class targets, and CNN filters learn spatial image features.

C · incorrect — Trial-and-error chess with rewards defines a reinforcement objective rather than supplied image-class targets.

D · incorrect — Hidden transaction-pattern discovery seeks unlabeled structure rather than labeled image classification.

Question 207

What does 'perfect information' mean in the context of game theory?

  • A. The agent knows the future
  • B. Both players have complete and accurate knowledge of the entire game state
  • C. The game has no randomness
  • D. Each player can see only their own pieces

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI22 · Games and minimax

AI04 · Rationality and performance

Study-guide references

Correct answer: B — The current game state and previous moves are visible to the players.

AI22 defines perfect information as visibility of previous moves and the current state. This makes B the correct state-knowledge description. It does not reveal future choices and does not by itself remove chance from a game. Seeing only one’s own pieces instead describes hidden relevant state, inconsistent with perfect information.

Evidence in the complete study note

Adversarial search models agents with competing objectives. A game specifies states, player to move, legal actions, successor function, terminal test and utility. Perfect information means previous moves and current state are visible; it does not reveal future choices or imply absence of chance. Zero-sum utility makes one player’s gain the other’s loss.AI22
Omniscience would mean knowing actual action outcomes in advance. Rationality does not require omniscience or guaranteed success. Even complete outcome knowledge cannot create a successful action when every available action fails. Information gathering and exploration may therefore be rational actions.AI04

Choice by choice

A · incorrect — Future choices are not known merely because the present game state is visible.

B · correct — Complete visibility of current state and relevant move history is the intended information condition.

C · incorrect — A game can include chance while still revealing the state and prior moves.

D · incorrect — Seeing only one’s own pieces leaves relevant state hidden and is not perfect information.

Question 208

Semi-supervised Learning is typically used when:

  • A. All data is labeled
  • B. No data is labeled
  • C. Obtaining a fully labeled dataset is time-consuming or costly
  • D. The problem requires reinforcement feedback

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI32 · Unsupervised, semi-supervised and active learning

AI31 · Supervised learning and prediction

AI33 · Reinforcement learning

Study-guide references

Correct answer: C — When labels are costly and some labeled data can be supplemented by unlabeled data.

AI32 defines semi-supervised learning as combining some labels with additional unlabeled data. Costly full labeling provides the motivation described in C, provided at least some labeled targets exist. Fully labeled data is ordinary supervised learning; wholly unlabeled structure fits unsupervised learning, and environmental rewards specify reinforcement feedback rather than this mixed dataset regime.

Evidence in the complete study note

Semi-supervised learning combines some labels with additional unlabeled data. Active learning chooses informative cases for an oracle to label; it still uses labels. “Exploratory learning” is not an exclusive standard synonym: it can mean unsupervised exploration or active data acquisition, so define the feedback. A task named summarization or time-series analysis is not automatically supervised; its training targets determine that.AI32
Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33

Choice by choice

A · incorrect — A fully labeled dataset does not give the defining labeled-plus-unlabeled mixture.

B · incorrect — With no labels at all, the described regime is unsupervised rather than semi-supervised.

C · correct — Expensive labeling motivates using a limited labeled subset together with plentiful unlabeled examples.

D · incorrect — Reward feedback defines reinforcement learning rather than a mixture of labeled and unlabeled examples.

Question 209

In the context of AI agents, what are 'effectors'?

  • A. Sensors that gather environmental data
  • B. Components that allow an agent to take actions in its environment
  • C. Internal memory modules
  • D. Heuristic evaluation functions

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI03 · Agents, percepts and action

AI06 · Simple and model-based reflex agents

AI17 · Heuristics and greedy best-first search

Study-guide references

Correct answer: B — Components through which an agent acts on its environment.

AI03 uses actuators or effectors for acting components and sensors for perception. Therefore B gives the correct role. Memory stores internal information and heuristic functions estimate promise; neither is the acting interface. An effector may be physical hardware or a software action mechanism depending on the agent’s architecture.

Evidence in the complete study note

An agent perceives an environment through sensors and acts through actuators or effectors. Cameras sense; motors move. Software sensors may be incoming messages and software actions may be API requests. Humans are agents but are not artificial agents.AI03
A model-based reflex agent maintains internal state using the previous state, previous action, current percept, and models of world change and action effects. It then applies rules to that state. Both kinds can use condition–action rules; memory and a world model distinguish them.AI06
A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17

Choice by choice

A · incorrect — Sensors gather percepts; effectors implement actions rather than observations.

B · correct — Effectors are the components through which selected actions affect the environment.

C · incorrect — Internal memory maintains information but is not the general acting interface.

D · incorrect — Heuristic functions evaluate promise and do not themselves execute environmental actions.

Question 210

Which of the following best describes a Utility-Based Agent?

  • A. It only reacts to current perceptions using condition-action rules
  • B. It selects actions that maximize expected utility, handling conflicting goals through trade-offs
  • C. It plans sequences of actions toward a single fixed goal
  • D. It learns entirely from environmental feedback

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI07 · Goals, utility and model updates

AI06 · Simple and model-based reflex agents

AI08 · Learning-agent components

Study-guide references

Correct answer: B — Choose actions with maximum expected utility, balancing outcome trade-offs.

AI07 defines utility-based agents by comparing desirability and trade-offs and maximizing expected utility over uncertain outcomes. This is broader than merely satisfying a fixed goal. Current-percept rule matching describes simple reflex behavior, and experience-based improvement describes learning, which may coexist with utility choice but does not define it.

Evidence in the complete study note

A goal-based agent considers future action sequences that reach a desired state. A utility-based agent compares degrees of desirability and trade-offs, choosing maximum expected utility: EU(a)=Σs P(s|a,evidence)U(s). Reaching the destination is a goal; balancing arrival time, safety and cost requires preferences among outcomes.AI07
A simple reflex agent matches condition–action rules against the current percept only: if dirty, clean. It has no persistent world-state memory. This is adequate when the current observation contains everything needed for its rule, but partial observability can make the same percept require different actions.AI06
The performance element selects actions. The learning element changes the agent to improve future performance. The critic evaluates behavior against a performance standard and supplies feedback. The problem generator suggests exploratory actions likely to yield useful experience.AI08

Choice by choice

A · incorrect — Current-percept condition-action matching describes a simple reflex architecture.

B · correct — Expected-utility maximization explicitly compares trade-offs among uncertain outcomes.

C · incorrect — Goal-reaching planning alone does not rank degrees of desirability among outcomes.

D · incorrect — Learning describes adaptation from feedback, not the utility-based decision criterion itself.

Question 211

What is Computer Vision in AI primarily concerned with?

  • A. Processing and generating spoken language
  • B. Enabling machines to understand and interpret visual information
  • C. Designing physical robotic systems
  • D. Managing database indexing

Match the input, output and purpose to the appropriate language, perception or physical-action task.

Study-guide sections

AI39 · Perception, vision and robotics

AI37 · Language processing tasks

Study-guide references

Correct answer: B — Interpreting visual information from images or video.

AI39 defines computer vision as interpreting images or video and includes classification, detection, segmentation and recognition. Language generation, robot mechanics and database indexing are separate task areas, although vision may be one component of a robot. B states the visual information-processing purpose directly.

Evidence in the complete study note

Computer vision interprets images or video: classification names image content, detection locates objects, segmentation labels pixels, and recognition identifies objects or people. Occlusion, lighting change, noise, unfamiliar conditions and incomplete context make open-world perception difficult even when a face detector works well.AI39
NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
Robotics combines mechanical, electrical, control and computing systems to sense, plan and act physically. Components include sensors, controller, power supply, actuators and an end effector. Locomotion may be legged, wheeled or tracked, with trade-offs in terrain, stability and efficiency. Software can act autonomously without being a robot. Humanoid shape is not evidence of general intelligence.AI39

Choice by choice

A · incorrect — Speech and language generation belong to language/speech processing rather than visual interpretation.

B · correct — Computer vision extracts and interprets information from visual inputs such as images and video.

C · incorrect — Robotic system design may use vision but is broader than the visual-processing field.

D · incorrect — Database indexing is not the interpretation of visual scenes or objects.

Question 212

In a zero-sum Tic-Tac-Toe terminal utility convention, MAX wins score +1 and MIN wins score −1. Which listed value represents a draw giving neither player an advantage?

  • A. +1
  • B. -1
  • C. +Infinity
  • D. 0

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI23 · Alpha–beta and game evaluation

Study-guide references

Correct answer: D — 0

The convention assigns equal and opposite utilities to the two decisive outcomes. A draw gives neither side an advantage, so its neutral utility is zero under that symmetric scale. Positive one favors MAX and negative one favors MIN; positive infinity would strongly favor MAX rather than indicate neutrality. This is an explicit utility convention, not an intrinsic numerical property of the game outcome.

Evidence in the complete study note

At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23

Choice by choice

A · incorrect — Under the common stated scale, +1 is a win rather than a draw.

B · incorrect — Under the common stated scale, −1 is a loss rather than a draw.

C · incorrect — Positive infinity is not the draw value in the note’s common scale.

D · correct — Zero is the neutral utility between +1 and −1 under this convention.

Question 213

Treat a whole independently reset chess match as one episode with no effects on later matches; treat the other options as ongoing tasks at the action level described. Which task is episodic under those units?

  • A. A self-driving car navigating city traffic
  • B. A chess game (each individual match)
  • C. A robot cleaning a multi-room house continuously
  • D. A stock trading algorithm operating over months

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: B — A chess game (each individual match)

An episodic task has independent decision units. Under the stated convention, each whole chess match is reset and cannot affect later matches, so matches can be treated as episodes. Individual moves within a match remain sequential. Driving, continuous cleaning and continuing stock trading all have actions that affect later situations within the ongoing task. The classification therefore follows the explicit unit and independence assumptions.

Evidence in the complete study note

Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · incorrect — Driving decisions affect later positions and traffic interactions, making the task sequential.

B · correct — Whole independently reset matches are independent episodes under the stated convention.

C · incorrect — Cleaning actions change later room states and decisions in the continuing task.

D · incorrect — Trading decisions affect later holdings and opportunities, so the continuing task is sequential.

Question 214

The heuristic function h(n) in A* represents:

  • A. The exact cost from start to node n
  • B. The total path cost already traveled
  • C. The estimated cost from node n to the goal state
  • D. The branching factor at node n

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI18 · Admissibility, consistency and A*

AI12 · Frontiers and evaluation criteria

Study-guide references

Correct answer: C — Estimated remaining cost from n to a goal.

AI17 defines h as an estimate of remaining cost or promise, and AI18 adds it to accumulated g in A*. Thus C is the heuristic component. Exact start-to-node cost belongs to g, and branching factor b is a search-size parameter rather than a cost-to-go estimate.

Evidence in the complete study note

A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
Completeness means finding a solution if one exists under stated assumptions; optimality means returning the best solution. Time counts work, often generated nodes; space counts simultaneous storage. Let b be branching factor, d shallowest goal depth, m maximum search depth, and ℓ a depth limit. Graph-search bounds in vertices V and edges E differ from implicit-tree bounds.AI12

Choice by choice

A · incorrect — Exact cost from the start belongs to g(n), not the heuristic h(n).

B · incorrect — Already-traveled path cost is g(n), not estimated future cost.

C · correct — h estimates the remaining cost from the current node to a goal.

D · incorrect — Branching factor counts successor alternatives and is not the heuristic cost estimate.

Question 215

When a doctor proposes a disease as a plausible cause explaining observed symptoms, which term names this explanatory direction of reasoning, rather than a method for numerically ranking hypotheses?

  • A. Deductive Reasoning
  • B. Inductive Reasoning
  • C. Abductive Reasoning
  • D. Probabilistic Reasoning

Translate the statement carefully, then check the scope of its quantifiers, connectives or inference rule. Distinguish guaranteed consequences from plausible generalizations or explanations.

Study-guide sections

AI28 · Deduction, induction and abduction

AI29 · Probability and uncertain models

Study-guide references

Correct answer: C — Abductive Reasoning

Abduction proposes an explanation for observed evidence, such as a disease that could account for symptoms. Other causes may also fit, so the conclusion is not logically guaranteed. Probability can help rank the possible explanations, but that numerical method is distinct from the explanatory direction. Deduction preserves truth from premises, and induction generalizes patterns across cases.

Evidence in the complete study note

Deduction applies truth-preserving rules: all people are mortal; Ada is a person; therefore Ada is mortal. Induction generalizes observations to a broader rule and can be overturned by new examples. Abduction proposes a plausible explanation: wet ground suggests rain, although sprinklers could explain it. Analogy transfers a pattern between similar situations.AI28
Probability describes uncertainty about events. P(¬A)=1−P(A); P(A|B)=P(A∩B)/P(B), for P(B)>0. Bayes gives P(H|E)=P(E|H)P(H)/P(E): posterior equals likelihood times prior divided by evidence probability. For mutually exclusive exhaustive hypotheses, P(E)=ΣiP(E|Hi)P(Hi).AI29

Choice by choice

A · incorrect — Deduction follows truth-preserving implications from premises, not a merely plausible cause inferred from its effects.

B · incorrect — Induction generalizes across observations rather than specifically proposing a cause that explains an observation.

C · correct — Abduction proposes a plausible explanation for observed evidence.

D · incorrect — Probability can rank disease hypotheses, but does not itself name the observation-to-explanation direction.

Question 216

What major ethical concern does AI raise related to decision-making systems?

  • A. AI systems use too little data
  • B. Bias in algorithms leading to unfair or discriminatory outcomes
  • C. AI always makes perfectly rational decisions
  • D. AI cannot process unstructured data

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI40 · Expert systems and choosing methods

AI34 · Neural networks and deep learning

AI04 · Rationality and performance

Study-guide references

Correct answer: B — Biased decisions causing unfair outcomes.

AI40 says bias in data or evaluation can create unfair outcomes and calls for checking errors across groups, robustness and oversight. B states the ethical concern directly. Data scarcity may be a technical issue, but little data is not the stated general ethical principle; perfect rationality and inability to process unstructured inputs are false generalizations.

Evidence in the complete study note

Rules offer inspectable domain reasoning but can be brittle and costly to maintain. Learned models handle data-rich patterns but depend on representative data and validation. Search suits explicit action alternatives; probabilistic models suit uncertain evidence. Bias in data or evaluation can create unfair outcomes. Evaluate errors across relevant groups, robustness, privacy and human oversight rather than assuming automation is always beneficial.AI40
Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34
A performance measure defines success externally: safety, useful deliveries, elapsed time, or cost. Rationality depends on that measure, prior knowledge, percept history, and available actions. A rational agent selects the action with the best expected performance given its information.AI04

Choice by choice

A · incorrect — Using too little data is not the specific fairness concern described here.

B · correct — Biased data or evaluation can produce unfair or discriminatory decisions.

C · incorrect — AI is not guaranteed perfect rational decisions or error-free outcomes.

D · incorrect — The note explicitly includes models that process images, sound and text.

Question 217

In a continuous environment like self-driving cars, what challenge does 'continuous' specifically pose?

  • A. A finite set of states makes decisions trivial
  • B. Infinite or large state/action spaces require more complex computations
  • C. The environment never changes
  • D. All outcomes are deterministic

Read the relevant agent component, action-selection criterion or environment dimension. Use only properties established by the description.

Study-guide sections

AI09 · Environment dimensions

Study-guide references

Correct answer: B — Real-valued state/action variables create large or infinite spaces to represent and search.

AI09 defines continuous quantities as varying along real-valued ranges, unlike distinct discrete alternatives. Such ranges can create unboundedly many state/action possibilities, motivating approximation or richer computation rather than exhaustive finite enumeration. Continuity is separate from whether a world changes over time or whether outcomes are stochastic.

Evidence in the complete study note

Episodic decisions are independent; sequential decisions affect later situations. Static worlds do not change while deliberating; dynamic ones do. A clock affecting score makes an otherwise static game semidynamic. Discrete quantities have distinct alternatives; continuous quantities vary along real-valued ranges. Known versus unknown concerns knowledge of the transition model, not observability.AI09

Choice by choice

A · incorrect — Continuous ranges do not reduce the task to a trivial finite set of alternatives.

B · correct — Real-valued ranges can produce very large or infinite possibilities needing suitable representation and computation.

C · incorrect — No world change defines static rather than continuous.

D · incorrect — Determinism is a separate transition-property dimension and does not follow from continuity.

Question 218

The successor function in game search is responsible for:

  • A. Evaluating terminal game states
  • B. Generating new states by applying legal moves to the current game state
  • C. Pruning unpromising branches
  • D. Calculating the heuristic value of each state

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI22 · Games and minimax

AI11 · Formulate before searching

AI23 · Alpha–beta and game evaluation

Study-guide references

Correct answer: B — Generate successor states by applying legal moves.

AI22 lists legal actions and a successor function as game components; AI11 identifies successors as states reached by actions. In game search, applying each legal move yields its successor position. Terminal utilities and heuristic scores evaluate positions, while alpha-beta pruning avoids exploring branches; neither is the transition-generation function.

Evidence in the complete study note

Adversarial search models agents with competing objectives. A game specifies states, player to move, legal actions, successor function, terminal test and utility. Perfect information means previous moves and current state are visible; it does not reveal future choices or imply absence of chance. Zero-sum utility makes one player’s gain the other’s loss.AI22
First formulate the goal; then formulate the problem by specifying initial state, legal actions, transition model, goal test, and action costs. Search selects a sequence; execution applies it. A state abstracts away irrelevant world details.AI11
At a depth cutoff use an evaluation function estimating position desirability. Positive favors the designated MAX player, negative MIN; the scale is a convention, with +1/0/−1 often win/draw/loss. Material counts weighted pieces; mobility counts legal options; territory scores controlled regions. These are different features, often combined. Heuristics improve limited computation, not guarantee victory.AI23

Choice by choice

A · incorrect — Terminal utility evaluation scores a completed outcome rather than generating the next position.

B · correct — The successor function maps legal moves at the current position to resulting game states.

C · incorrect — Pruning skips exploration under a bound; it is not state-transition generation.

D · incorrect — A heuristic evaluation scores a state’s promise rather than producing legal successor states.

Question 219

For implicit-tree search with branching factor b and shallowest goal depth d, which is a key memory advantage of iterative deepening over BFS when each depth-limited pass stores pending siblings?

  • A. It finds deeper solutions faster
  • B. It requires storing all nodes at each level in memory
  • C. It uses O(bd) frontier space rather than storing an entire BFS level.
  • D. It does not guarantee completeness

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI16 · Depth limits, iterative deepening and two directions

AI13 · Breadth-first search

AI14 · Depth-first search

Study-guide references

Correct answer: C — It uses O(bd) frontier space rather than storing an entire BFS level.

Iterative deepening repeats depth-limited DFS while increasing the limit. A pass retains a depth-first path and pending siblings, giving O(bd) frontier space rather than BFS’s potentially exponential level frontier. It remains complete for finite branching and a finite-depth solution, although it repeats some shallow work. The option states the correct memory advantage without claiming that every implementation stores only the current path.

Evidence in the complete study note

DLS runs DFS to depth ℓ and reports cutoff separately from failure. It misses goals deeper than ℓ. Iterative deepening runs DLS with limits 0,1,2,…; for finite b it finds the shallowest solution with O(b^d) time and O(bd) frontier space, optimal for equal step costs. Repeated shallow work is small compared with the deepest level.AI16
The tree contains 1+b+…+b^d nodes through depth d, giving O(b^d) under generation-time goal testing. Expanding depth-d nodes before testing can generate depth d+1 and yields O(b^(d+1)); specify the convention. Explicit finite graph traversal with visited-state tracking takes O(V+E) time and O(V) space.AI13
DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14

Choice by choice

A · incorrect — IDS repeats shallow work and does not universally find deeper solutions faster than BFS.

B · incorrect — Storing all nodes at a level is BFS-style memory behavior, not the IDS advantage.

C · correct — The stated implementation keeps only a depth-first frontier, requiring O(bd) rather than the BFS level frontier.

D · incorrect — IDS is complete under finite branching and a finite-depth solution, so loss of completeness is not its benefit.

Question 220

Representation learning in Deep Learning refers to:

  • A. Manually engineering input features
  • B. The ability to automatically discover and learn relevant features from raw data
  • C. Storing knowledge in frames and ontologies
  • D. Using labeled data with a teacher signal

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI34 · Neural networks and deep learning

AI43 · Clarify overlapping learning and AI labels

AI24 · Represent knowledge for inference

AI-S06 · Prediction, transfer and explicit representations

Study-guide references

Correct answer: B — Automatically learning useful features from data.

AI34 says deep models learn features from images, sound and text through multiple representation layers. That is representation learning rather than manually specifying every input feature. A label signal may guide training but is not required for every deep objective, and frames/ontologies are separate explicit knowledge structures rather than the learned representations meant here.

Evidence in the complete study note

Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34
ANNs, genetic search and reinforcement learning all improve candidate behavior or performance through experience in different ways. Deep learning is an architecture family usable with supervised, unsupervised or reinforcement objectives. Decision trees and deep models may both be supervised. To distinguish learning regimes, specify labels, reward or unlabeled structure.AI43
A semantic network connects concept nodes by labeled relations such as is-a. Frames describe stereotyped entities through slots, values and defaults. Production systems use IF conditions THEN conclusions or actions, matching rules against working memory and selecting among applicable rules. A graph suits relationships such as frequently bought together; frames suit attributes. No representation is always best for recommendation without task requirements.AI24
An ontology describes entities or individuals, classes and properties and may state logical relationships and restrictions. It organizes explicit conceptual knowledge. Conditional probability tables and a probabilistic factorization are additional modeling commitments, not inherent in the word ontology.AI-S06

Choice by choice

A · incorrect — Manual feature engineering is contrasted with learning the representations automatically.

B · correct — The model learns task-useful features or representations from the supplied data.

C · incorrect — Frames and ontologies are explicit knowledge representations, not the deep feature-learning mechanism described.

D · incorrect — Teacher labels define supervision; representation learning can also use unsupervised or reinforcement objectives.

Question 221

What makes Best-First (Greedy) Search non-optimal?

  • A. It always explores all possible paths
  • B. It relies entirely on heuristics and can overlook cheaper paths, making suboptimal choices
  • C. It uses actual path costs rather than estimates
  • D. It explores nodes in breadth-first order

Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

Study-guide sections

AI17 · Heuristics and greedy best-first search

AI19 · Work a heuristic-search example

Study-guide references

Correct answer: B — It uses only h and can ignore a cheaper route with a larger estimate.

AI17 defines greedy best-first as f=h, ignoring cost already paid. AI19 demonstrates a route of cost 11 selected greedily while another route costs 4. The heuristic ranking can therefore prefer a worse total solution. Greedy is neither exhaustive nor breadth-first and does not order primarily by actual path cost.

Evidence in the complete study note

Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
For S→A cost 2, S→B cost 1, A→G cost 2, B→G cost 10, set h(A)=2, h(B)=1, h(G)=0. After expanding S, greedy chooses B because 1<2 and can then choose G, returning cost 11. A* compares A:f=2+2=4 and B:f=1+1=2; after B, its G entry costs 11. It next expands A and replaces G with cost 4, which is removed and accepted.AI19

Choice by choice

A · incorrect — Greedy search may stop at an early found goal rather than exploring every path.

B · correct — Ignoring accumulated cost can favor a seemingly close but expensive route over a cheaper solution.

C · incorrect — Using actual g alone describes UCS, whereas greedy uses h.

D · incorrect — Breadth-first depth order is not the greedy minimum-heuristic policy.

Question 222

In AI, what does 'Acting Humanly' (Turing Test concept) primarily evaluate?

  • A. Whether a machine can solve math problems
  • B. Whether a machine's behavior is indistinguishable from a human's in conversation
  • C. Whether a machine can perceive visual stimuli
  • D. Whether a machine follows logical rules perfectly

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI02 · Four approaches and the Turing test

Study-guide references

Correct answer: B — Humanlike conversational behavior judged by indistinguishability.

AI02 asks whether a judge can reliably distinguish a machine from a human through dialogue. This operational test concerns observed conversation, not a proof of consciousness or perfect logical reasoning. Mathematical success and perception may contribute capabilities but are not the primary distinguishing conversational criterion. B states that criterion directly.

Evidence in the complete study note

Thinking humanly models human mental processes; thinking rationally derives conclusions by valid logic. Acting humanly seeks humanlike behavior; acting rationally selects actions expected to achieve objectives. Thought processes and observed behavior are different evaluation targets.AI02
The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02

Choice by choice

A · incorrect — Mathematical ability alone does not supply the test’s conversational indistinguishability criterion.

B · correct — The test evaluates whether conversational behavior can be reliably distinguished from a human’s.

C · incorrect — Visual perception alone is not the conversational test described in AI02.

D · incorrect — Perfect logic is a rationality criterion, not the test’s humanlike observed-behavior requirement.

Question 223

Which of the following is the correct relationship between AI, ML, and DL?

  • A. DL ⊂ AI ⊂ ML
  • B. AI ⊂ ML ⊂ DL
  • C. ML ⊂ DL ⊂ AI
  • D. DL ⊂ ML ⊂ AI

Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

Study-guide sections

AI34 · Neural networks and deep learning

AI33 · Reinforcement learning

Study-guide references

Correct answer: D — DL⊂ML⊂AI.

AI34 explicitly gives deep learning inside machine learning inside AI. Deep neural architectures are a subset of learning methods, and learning is one component of the broader intelligence field, which also includes planning and reasoning. Every other ordering reverses at least one of these stated inclusions.

Evidence in the complete study note

Feedforward networks pass signals toward outputs without recurrent state loops; recurrent networks carry state across sequence steps. Training adjusts weights, often through gradients and backpropagation. Backpropagation is not the same as a recurrent connection. Deep learning uses multiple representation layers and can learn features from images, sound and text. CNNs share filters across image locations. ANN is a brain-inspired abstraction, not a literal brain simulation; DL⊂ML⊂AI.AI34
An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33

Choice by choice

A · incorrect — This wrongly puts AI inside machine learning rather than the reverse.

B · incorrect — Both inclusions are reversed relative to the note’s hierarchy.

C · incorrect — Machine learning contains deep learning, not the reverse.

D · correct — Deep learning is contained in machine learning, which is contained in AI.

Question 224

A recommendation system must directly represent products and labeled “frequently bought together” relationships, then traverse those relationships to find connected products. Which listed representation most directly fits that requirement?

  • A. Rule-based IF-THEN logic
  • B. Frames describing product attributes
  • C. Graph-based model with products as nodes and relationships as edges
  • D. Probabilistic Bayesian network of user demographics

Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

Study-guide sections

AI24 · Represent knowledge for inference

AI29 · Probability and uncertain models

Study-guide references

Correct answer: C — Graph-based model with products as nodes and relationships as edges

A semantic graph represents entities as nodes and their labeled relationships as edges, making connected products directly traversable. Frames primarily organize attributes, rules express conditions and conclusions, and a Bayesian demographic model represents uncertain dependencies rather than the requested product-edge structure. The answer follows the explicit relational representation requirement; it does not claim that graphs are universally best for every recommendation system.

Evidence in the complete study note

A semantic network connects concept nodes by labeled relations such as is-a. Frames describe stereotyped entities through slots, values and defaults. Production systems use IF conditions THEN conclusions or actions, matching rules against working memory and selecting among applicable rules. A graph suits relationships such as frequently bought together; frames suit attributes. No representation is always best for recommendation without task requirements.AI24
A Bayesian network is a directed acyclic graph with conditional probability tables; its joint distribution factors as the product of each node’s probability given its parents. A Markov model assumes the next state depends on the current state rather than the entire past. These represent uncertainty; fuzzy membership represents a different issue.AI29

Choice by choice

A · incorrect — Rules can encode recommendations, but do not directly provide the requested product-node relationship structure.

B · incorrect — Frames describe product attributes rather than making product-to-product links the primary structure.

C · correct — Products as nodes and labeled relationships as edges directly support the required traversal.

D · incorrect — A demographic Bayesian network models uncertain dependencies rather than the specified product-edge structure.

Question 225

What is the difference between supervised and unsupervised learning?

    Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

    Study-guide sections

    AI31 · Supervised learning and prediction

    AI32 · Unsupervised, semi-supervised and active learning

    Study-guide references

    Answer: Supervised learning fits labeled input–target examples; unsupervised learning discovers structure without target labels.

    AI31 defines supervised learning through labels, with classification for discrete targets and regression for numeric targets. AI32 defines unsupervised discovery, including clustering, association and dimensionality reduction. The distinction is the supplied training feedback, not whether the data are images, text or numbers; the same modality can support either regime.

    Evidence in the complete study note

    Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
    Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32

    Question 226

    Under what search and heuristic conditions does A* find a least-cost path, and why?

      Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

      Study-guide sections

      AI18 · Admissibility, consistency and A*

      Study-guide references

      Answer: A* is optimal under its heuristic and search assumptions: admissible tree search, or consistent graph search without reopening; otherwise reopen improved states.

      With admissible h, A* tree search uses f=g+h as a lower bound on solution cost through each frontier node. Removing a minimum-f goal cannot skip a cheaper represented solution. Graph search that never reopens closed states normally requires consistency; with an admissible inconsistent heuristic, improved states must be reopened. Goal testing is on removal. Finite branching and positive lower-bounded step costs supply the usual infinite-space completeness conditions.

      Evidence in the complete study note

      A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
      Admissibility suffices for optimal A* tree search; graph search that never reopens CLOSED states normally requires consistency. With an admissible inconsistent heuristic, reopen states when g improves. Test a goal when removed at minimum f. Completeness also needs finite branching and positive lower-bounded costs in infinite spaces. Worst-case time and memory can be exponential; no heuristic universally guarantees speed.AI18

      Question 227

      What is the Turing Test and what does it measure?

        Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

        Study-guide sections

        AI02 · Four approaches and the Turing test

        Study-guide references

        Answer: The conversational Turing test asks a judge to distinguish machine from human through dialogue; it measures humanlike behavior, not consciousness.

        AI02 sets the observable criterion: whether a judge can reliably tell a machine’s conversation from a human’s. Language use, knowledge, reasoning and often learning support the performance. Passing would meet this operational behavioral standard, but it does not prove consciousness, actual understanding or that the machine internally thinks in the same way as a human.

        Evidence in the complete study note

        The conversational Turing test asks whether a judge can reliably distinguish a machine from a human through their dialogue. It requires language use, knowledge, reasoning and often learning. It tests an operational behavior criterion, not proof of consciousness. Cognitive models can be tested against response times, errors and psychological experiments.AI02

        Question 228

        What is the difference between precision and recall?

          Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

          Study-guide sections

          AI36 · Measure the right errors

          Study-guide references

          Answer: Precision=TP/(TP+FP): how many predicted positives are correct. Recall=TP/(TP+FN): how many actual positives are found.

          AI36 distinguishes precision’s predicted-positive denominator from recall’s actual-positive denominator. False positives reduce precision; false negatives reduce recall. With TP=30, FP=10 and FN=20, precision is 30/40=75% and recall is 30/50=60%. The positive class and zero-denominator convention should be fixed before comparing scores.

          Evidence in the complete study note

          With a fixed positive class, TP means true positives, FP false positive predictions, FN missed positives, TN true negatives. Accuracy=(TP+TN)/N; precision=TP/(TP+FP); recall=TP/(TP+FN); F1=2PR/(P+R). State the averaging and zero-denominator convention when relevant.AI36

          Question 229

          Why is tokenization important in natural language processing?

            Match the input, output and purpose to the appropriate language, perception or physical-action task.

            Study-guide sections

            AI37 · Language processing tasks

            Study-guide references

            Answer: Tokenization divides text into processing units such as words, punctuation or subwords so later NLP stages can operate on a consistent input sequence.

            AI37 defines tokenization as dividing text into processing units and lists several permitted granularities. Establishing that sequence lets later tagging, recognition and parsing refer to units consistently. Tokens need not equal whitespace-separated words, and tokenization by itself does not determine grammatical structure or resolve word meaning; those belong to later analyses.

            Evidence in the complete study note

            NLP interprets and generates human language: translation, dialogue, summarization, question answering and text generation. Speech recognition converts audio to text; speech synthesis generates speech. They overlap language processing but solve different input/output tasks.AI37
            Tokenization divides text into processing units; tokens may be words, punctuation or subwords. POS tagging assigns grammatical roles; named-entity recognition identifies people, organizations and locations. Parsing identifies grammatical structure; semantics concerns meaning; discourse links sentences; pragmatics uses context and speaker intent. Sentiment analysis predicts evaluative tone. Ambiguity can occur at every level: “bank” has multiple senses, and “I saw the man with a telescope” has multiple attachments.AI37

            Question 230

            Which AI approach most directly supports understanding biological intelligence behavior as a scientific goal?

            • A. Weak AI
            • B. Applied AI
            • C. Cognitive AI
            • D. Strong AI

            Read the definition and the classification axis requested; distinguish the defining purpose from supporting techniques or implementation resources.

            Study-guide sections

            AI01 · AI capabilities and scope

            Study-guide references

            Correct answer: C — Cognitive AI.

            AI01 explicitly describes cognitive AI as using computational models to investigate biological intelligence, making it the direct scientific-goal match. Applied AI emphasizes useful solutions. Narrow AI describes capability scope, while strong AI has a philosophical understanding sense; neither is the named research approach for testing models against biological behavior.

            Evidence in the complete study note

            Artificial intelligence studies systems that perform tasks requiring intelligence: reasoning, learning, planning, perception and language. Applied AI builds useful solutions; cognitive AI uses computational models to investigate biological intelligence. A model should be compared with human behavior or experimental observations before claiming it explains human thought.AI01
            Narrow or weak AI addresses bounded tasks, such as face recognition, spam filtering or a voice assistant. General AI denotes broad competence transferable across domains; strong AI also has a philosophical meaning involving actual understanding. Do not equate a humanoid body, autonomous vehicle, or many features with general intelligence.AI01

            Choice by choice

            A · incorrect — Weak AI denotes bounded capability rather than the specific scientific modeling goal.

            B · incorrect — Applied AI builds useful solutions rather than primarily explaining biological intelligence.

            C · correct — Cognitive AI uses computational models to investigate biological intelligence and behavior.

            D · incorrect — Strong AI concerns broad or actual-understanding claims, not this specific empirical modeling approach.

            Question 231

            Why is Greedy Best-First Search not guaranteed to be optimal or complete?

            • A. It ignores path cost g( n) and relies only on h( n)
            • B. It always chooses the shallowest node
            • C. It expands nodes in FIFO order
            • D. It requires full knowledge of the goal state in advance

            Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

            Study-guide sections

            AI17 · Heuristics and greedy best-first search

            AI19 · Work a heuristic-search example

            Study-guide references

            Correct answer: A — It uses h alone, ignoring accumulated g.

            AI17 says greedy best-first chooses f=h and can take an expensive route or follow an infinite misleading branch. Ignoring g explains nonoptimality, while cycles or infinite paths without appropriate handling explain failure of completeness. On a finite graph with full duplicate handling it is complete, so the lack of a general guarantee does not mean every implementation fails.

            Evidence in the complete study note

            Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
            For S→A cost 2, S→B cost 1, A→G cost 2, B→G cost 10, set h(A)=2, h(B)=1, h(G)=0. After expanding S, greedy chooses B because 1<2 and can then choose G, returning cost 11. A* compares A:f=2+2=4 and B:f=1+1=2; after B, its G entry costs 11. It next expands A and replaces G with cost 4, which is removed and accepted.AI19

            Choice by choice

            A · correct — Using only h ignores total path cost and can follow misleading or unbounded branches.

            B · incorrect — Shallowest-first describes breadth-first order rather than greedy heuristic order.

            C · incorrect — FIFO is breadth-first queue behavior, not minimum-h priority selection.

            D · incorrect — Heuristics estimate promise; exact full goal-state knowledge is not required by the algorithm.

            Question 232

            Which search strategy tries to expand the node that is closest to the goal, on the grounds that this is likely to lead to a solution quickly; Thus, it evaluates nodes by using just the heuristic function: f ( n) = h ( n).

            • A. Uniform-cost search
            • B. Greedy best-first search
            • C. Depth-first search
            • D. A* search

            Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

            Study-guide sections

            AI17 · Heuristics and greedy best-first search

            AI15 · Uniform-cost search and Dijkstra

            AI18 · Admissibility, consistency and A*

            AI14 · Depth-first search

            Study-guide references

            Correct answer: B — Greedy best-first search.

            AI17 explicitly defines greedy best-first by removing the smallest remaining-cost estimate h while ignoring paid cost g. This is exactly the stem’s equation. UCS uses g, A* combines g and h, and DFS uses depth/stack discipline rather than a minimum heuristic evaluation.

            Evidence in the complete study note

            Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
            UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15
            A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
            DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14

            Choice by choice

            A · incorrect — UCS orders by accumulated g rather than h alone.

            B · correct — Greedy best-first selects the node with the smallest heuristic estimate h.

            C · incorrect — DFS follows a LIFO/deepest-branch order rather than minimum h.

            D · incorrect — A* uses g+h, while the stem explicitly uses only h.

            Question 233

            Which of the following best defines a heuristic in the context of AI search?

            • A. A function that estimates how close a state is to the goal
            • B. A guaranteed optimal path to the goal
            • C. A method that always avoids revisiting states
            • D. A brute-force enumeration of all possible states

            Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

            Study-guide sections

            AI17 · Heuristics and greedy best-first search

            AI12 · Frontiers and evaluation criteria

            Study-guide references

            Correct answer: A — A problem-specific estimate of remaining cost or promise.

            AI17 defines a heuristic h(n) as estimating remaining cost or evaluating promise. A’s closeness wording captures that estimate rather than an exact path guarantee. A heuristic does not inherently prevent repeated states and is not exhaustive enumeration; those are separate search-policy or bookkeeping choices. Its quality and computation cost affect practical search.

            Evidence in the complete study note

            A heuristic h(n) estimates remaining cost or evaluates promise using problem-specific knowledge. It can reduce explored work, but computing it also costs time. Uninformed algorithms use only the problem definition and known costs; UCS remains uninformed despite using g.AI17
            A search node records a state plus parent, action, depth and path cost. Expanding a node generates its successors. OPEN or frontier stores discovered candidates awaiting expansion; CLOSED records expanded states. Reaching the same state through different paths does not imply the same node or cost.AI12

            Choice by choice

            A · correct — Estimating closeness or remaining cost is the heuristic role defined in AI17.

            B · incorrect — An estimate is not itself a guaranteed optimal path.

            C · incorrect — Repeated-state avoidance requires duplicate handling rather than following from heuristic estimation.

            D · incorrect — Exhaustive enumeration is a search method, not the definition of a heuristic estimate.

            Question 234

            Which type of heuristic search strategy that evaluates nodes by combining g ( n), the cost to reach the node, and h ( n.), the cost to get from the node to the goal: f ( n) = g ( n) + h ( n) Since g( n) gives the path cost from the start node to node n, and h( n) is the estimated cost of the cheapest path from n to the goal, we have f ( n) = estimated cost of the cheapest solution through n ?

            • A. A* search
            • B. Greedy Best-First Search
            • C. Uniform-cost search
            • D. Depth-first search

            Check how the method selects its next candidate, what information it uses, and which cost, depth or heuristic assumptions the question supplies.

            Study-guide sections

            AI18 · Admissibility, consistency and A*

            AI17 · Heuristics and greedy best-first search

            AI15 · Uniform-cost search and Dijkstra

            AI14 · Depth-first search

            Study-guide references

            Correct answer: A — A* search.

            AI18 defines A* by f(n)=g(n)+h(n), exactly as the long stem specifies. The sum estimates the cheapest solution through the node when h estimates remaining cost. Greedy ignores g, UCS ignores h, and DFS does not prioritize this sum, so none of those matches the stated evaluation.

            Evidence in the complete study note

            A* orders by f(n)=g(n)+h(n). Admissibility means 0≤h(n)≤h*(n), the true cheapest remaining cost. Consistency means h(n)≤c(n,n′)+h(n′) on every edge, with h(goal)=0; it implies admissibility on states with goal paths. Consistency makes f nondecreasing along a path.AI18
            Best-first is a family that orders OPEN by an evaluation function. Greedy best-first uses f=h and removes the smallest estimate, ignoring the cost already paid. It can choose an expensive route, loop without duplicate checks, or follow an infinite misleading branch. On a finite graph with full duplicate handling it is complete, but not generally optimal. Its implicit finite-tree worst case can be O(b^m).AI17
            UCS uses a priority queue ordered by actual accumulated path cost g(n). Remove a minimum-g node; accept a goal when removed, not when first generated. Relax each edge: if a cheaper path is found, update its frontier cost and parent.AI15
            DFS follows the deepest pending branch using a LIFO stack or recursion. FILO means the same order as LIFO. With the same children as BFS, a left-first DFS visits S,A,C before B; push rightmost children first to pop the leftmost first.AI14

            Choice by choice

            A · correct — A* adds actual cost-to-node g and estimated remaining cost h.

            B · incorrect — Greedy best-first uses only h and omits accumulated path cost.

            C · incorrect — UCS uses only g and does not add a remaining-cost heuristic.

            D · incorrect — DFS uses stack/depth ordering rather than minimum g+h.

            Question 235

            In which of the following Learning methods an agent receives feedback in the form of rewards or penalties to maximize cumulative reward?

            • A. Supervised learning
            • B. Reinforcement learning
            • C. Semi-supervised learning
            • D. Unsupervised learning

            Identify the training feedback, prediction target or measured error. Keep learning objectives separate from model architecture and data modality.

            Study-guide sections

            AI33 · Reinforcement learning

            AI31 · Supervised learning and prediction

            AI32 · Unsupervised, semi-supervised and active learning

            Study-guide references

            Correct answer: B — Reinforcement learning.

            AI33 defines RL through interaction, rewards and improvement of a policy to maximize expected cumulative reward. The reward may be delayed and does not provide a correct-action label. Supervised and semi-supervised learning instead concern target labels, while unsupervised discovery lacks the defining action-reward interaction described here.

            Evidence in the complete study note

            An RL agent observes state information, takes an action, receives reward and a new observation, and improves its policy to maximize expected cumulative reward. Rewards evaluate outcomes rather than supply the correct action label. Rewards may be delayed; not every correct-looking action receives immediate reward.AI33
            Supervised learning fits a mapping from labeled input–target examples. Classification predicts discrete labels, such as institution generations or spam/nonspam; regression predicts numeric quantities, such as prices. Predictive analytics uses historical patterns to forecast outcomes and can use either. Specify the target before choosing a method.AI31
            Semi-supervised learning combines some labels with additional unlabeled data. Active learning chooses informative cases for an oracle to label; it still uses labels. “Exploratory learning” is not an exclusive standard synonym: it can mean unsupervised exploration or active data acquisition, so define the feedback. A task named summarization or time-series analysis is not automatically supervised; its training targets determine that.AI32
            Unsupervised learning seeks structure without target labels: clustering groups similar examples; association finds co-occurrence; dimensionality reduction compresses features. Good clusters have high within-cluster similarity and low between-cluster similarity under an appropriate distance. K-means alternates assignment to nearest centroids and centroid updates.AI32

            Choice by choice

            A · incorrect — Supervised learning fits labeled targets rather than defining behavior through cumulative interaction rewards.

            B · correct — Reward or penalty feedback guiding cumulative-return improvement is the defining RL mechanism.

            C · incorrect — Semi-supervised learning mixes labeled and unlabeled examples, not necessarily action rewards.

            D · incorrect — Unsupervised learning discovers unlabeled structure rather than optimizing an interaction reward policy.