Artificial Intelligence and Neural Networks

Answer: A. Turing test

Turing test evaluates machine intelligence by checking if distinguishable from human.

Answer: A. DBMS

Database Management Systems are not AI applications.

Answer: C. Building systems that perceive, reason, learn, and act to achieve goals in their environment

Artificial Intelligence is not simply about fast computation or blind automation. Classical definitions from Russell & Norvig describe AI as the study and design of intelligent agents: systems that perceive their environment through sensors and act upon that environment through actuators in a way that maximizes their chances of achieving goals. Four classical perspectives help clarify this: (1) Systems that think like humans (cognitive modeling), (2) Systems that act like humans (Turing test), (3) Systems that think rationally (laws of thought), and (4) Systems that act rationally (rational agents). Modern AI largely follows the rational agent perspective. An intelligent system should: (a) Perceive: interpret inputs such as images, speech, text, or sensor readings, (b) Reason: build internal representations and draw inferences from them, (c) Learn: improve its performance from experience instead of relying only on fixed rules, and (d) Act: choose actions that are appropriate for its goals and environment dynamics. Fast computation, automation, or mimicry of human behavior are sometimes side-effects of AI techniques, but they do not fully define AI. For instance, a simple script that clicks a button automatically is not intelligent; it does not perceive changing context, adapt, or optimize towards a goal. In contrast, a self-driving car that observes the road, predicts other vehicles’ behavior, plans routes, and continuously improves via data is operating as an intelligent agent.

Answer: A. Thinking humanly, acting humanly, thinking rationally, acting rationally

Foundations of AI textbooks describe four high-level perspectives on AI, based on two axes: human vs rational and thought vs behavior. This yields four quadrants: (1) Thinking humanly: Trying to model how humans actually think internally (cognitive science, cognitive modeling). This requires psychological experiments and possibly brain imaging to check whether the model matches human reasoning patterns. (2) Acting humanly: Trying to make machines behave indistinguishably from humans, regardless of how they internally work. The Turing test is the classical example: if a judge cannot reliably distinguish the machine from a human in conversation, the machine is said to exhibit human-level intelligence in that domain. (3) Thinking rationally: Emphasizing correct reasoning according to formal logic. Early AI research tried to encode knowledge and reasoning using logical systems, proving theorems and deriving conclusions that are always sound and logically valid. (4) Acting rationally: Focusing on rational agents that choose actions that maximize expected goal achievement given their beliefs and constraints. This perspective is more flexible than strict logic, because it can handle uncertainty, limited computation, and noisy sensors. Modern AI tends to emphasize rational agency: design agents that do the “right” thing given their objectives, knowledge, and computational resources. The other options in the question either mix unrelated terms or miss this canonical 2×2 classification.

Answer: B. An entity that perceives its environment and takes actions that maximize its expected performance measure

An intelligent agent is formally defined as an entity that perceives its environment via sensors and acts upon that environment via actuators, choosing actions so as to maximize its expected performance measure over time. Several components are important here: (1) Percepts: At each time step, the agent receives percepts (observations) from the environment, which might be partial or noisy. For example, a robot receives camera images, distance sensor readings, and encoder values. (2) Actions: The agent has a set of possible actions it can execute, such as moving forward, turning, speaking, or firing a control signal. (3) Performance measure: Rather than simply “doing something”, the agent is evaluated according to a performance measure defined by the system designer (for example, safety, speed, fuel efficiency, or game score). The agent’s goal is to maximize this measure, not necessarily to mimic human behavior. (4) Agent function vs agent program: The agent function maps percept histories to actions. The agent program is a concrete implementation of that function on some hardware. An autonomous script that just runs without user input but does not sense the environment or optimize a performance measure is not necessarily an intelligent agent. Similarly, a static rule-based system that never adapts can still be an agent, but whether it is intelligent depends on whether it chooses actions that reasonably maximize performance in its environment. Intelligence is judged relative to what was feasible for an ideal agent with the same information. This definition is general enough to cover simple reflex agents, model-based agents, goal-based planners, and utility-based agents.

Answer: B. Performance measure, Environment, Actuators, Sensors

PEAS is a useful checklist for specifying a task environment for an intelligent agent before designing the agent itself. It stands for: (1) Performance measure: The quantitative or qualitative criteria used to evaluate the agent’s success. For a self-driving car, this might include safety, travel time, comfort, and legal compliance. For a chess agent, it is usually winning the game. Defining the performance measure carefully is critical; a poorly chosen measure can lead to undesirable behavior. (2) Environment: The external world with which the agent interacts. This includes other agents, physical constraints, rules, and dynamics. For a vacuum-cleaning robot, the environment might be rooms, dirt distribution, furniture, and humans walking around. Specifying the environment helps determine its properties (deterministic vs stochastic, static vs dynamic, etc.). (3) Actuators: The mechanisms through which the agent affects the environment. Examples include wheels and motors for a robot, keystrokes or API calls for a software agent, arms and grippers in industrial robotics, or network packets for a network management agent. (4) Sensors: The input channels through which the agent perceives the environment. These could be cameras, microphones, IR sensors, keyboard/mouse input, or system logs. Distinguishing sensors from actuators clarifies what information the agent receives and what it can change. A good PEAS specification provides a structured way to think about the design problem and directly influences the choice of agent architecture (simple reflex, model-based, goal-based, or utility-based) and the algorithms used (search, planning, learning, etc.).

Answer: D. Utility-based agent – optimizes a numeric preference function over states or outcomes

Agent architectures in AI can be arranged from simple to sophisticated: (1) Simple reflex agents act only on the basis of the current percept, ignoring history. They apply condition–action rules like “if dirty then suck, else move right”. They have no internal model of the world’s dynamics. This makes them easy to implement but fragile in partially observable or changing environments. (2) Model-based reflex agents maintain some internal state to track aspects of the world that are not directly observed in the current percept. They use a model of how the world evolves and how actions affect the world, allowing them to handle partial observability and remember past information. They are still driven by condition–action rules but with state as input. (3) Goal-based agents introduce explicit representation of goals (desired states). Rather than simply reacting, they plan: they search for action sequences that lead from the current state (or belief state) to a goal state. They may consider multiple paths and choose one that reaches a goal, but they treat all goal states as equally acceptable without measuring degrees of preference. (4) Utility-based agents extend goal-based behavior by associating numeric utilities (preferences) with states or outcomes. They choose actions that maximize expected utility, taking into account trade-offs among conflicting objectives (such as speed vs safety vs comfort). This is especially important in stochastic or multi-objective environments where simply reaching a goal is not enough. Therefore, only the utility-based agent explicitly optimizes a numeric preference function. The other options in the question mix characteristics: simple reflex agents do not maintain internal models, and model-based agents do not act solely on current percept; they use state. Goal-based agents use goals, not utilities, as their primary evaluation criterion.

Answer: B. Deterministic, fully observable, sequential, multi-agent

To classify an environment, several dimensions are considered: (1) Deterministic vs stochastic: Chess is deterministic because the result of an action (moving a piece) is completely determined by the rules and current board state; there is no randomness in transitions. Stochastic environments involve random effects (for example, dice rolls in some board games). (2) Fully observable vs partially observable: In chess, both players can see the entire board configuration at all times; there is no hidden information. Therefore, it is fully observable. Games like poker are partially observable because players cannot see opponents’ cards. (3) Static vs dynamic: A static environment does not change while the agent is deliberating; chess is essentially static in this sense because the environment changes only when a player makes a move, not spontaneously. In contrast, driving a car is dynamic: the world changes continuously. (4) Episodic vs sequential: In episodic environments, the agent’s current action does not depend on previous actions. Chess is sequential: the current decision depends heavily on the sequence of prior moves and has long-term consequences. (5) Single-agent vs multi-agent: Chess clearly involves two agents (our agent and the opponent) whose goals conflict, so it is a competitive multi-agent environment (adversarial). Thus, the best classification among the options is deterministic, fully observable, sequential, multi-agent. This classification guides the choice of solution methods: adversarial search algorithms like minimax and alpha–beta pruning are appropriate rather than single-agent shortest-path search.

Answer: B. Deterministic environment

In a deterministic environment, the next state is completely determined by the current state and the action taken by the agent. Environment types in AI: (1) Deterministic - Next state fully determined by current state and action, (2) Stochastic/Non-deterministic - Action produces multiple possible outcomes with probabilities, (3) Observable - Agent perceives all relevant information, (4) Partially observable - Agent perceives only partial information. Deterministic environment characteristics: (1) Predictable - Same action always produces same result, (2) No randomness - No chance events affecting outcome, (3) Complete causality - Action directly causes specific outcome, (4) Reversible - Can always reverse to previous state (usually). Deterministic examples: (1) Chess - Move produces exact board state, (2) Tic-tac-toe - Same move always same result, (3) Scripted simulations - Predetermined outcomes, (4) Mathematical calculations - Deterministic formula. Non-deterministic examples: (1) Card games - Action outcome depends on random card draw, (2) Real world - Weather, traffic unpredictable, (3) Game with dice - Multiple outcomes possible. Combined characteristics: (1) Fully observable + Deterministic - Complete, predictable information, (2) Partially observable + Stochastic - Incomplete, unpredictable - most real-world, (3) Observable + Stochastic - Complete info but random outcomes. Related concepts: (1) Episodic - Task independent episodes vs sequential, (2) Static - Environment doesn't change while agent decides, (3) Discrete - Finite states vs continuous. Planning implications: (1) Deterministic - One valid plan works, (2) Stochastic - Need contingency plans, (3) Requires different algorithms - Search vs policy. Real-world AI: (1) Most real environments are stochastic, (2) Some specialized systems can be deterministic, (3) Hybrid approaches handle uncertainty. This distinction is important for algorithm selection.

Answer: C. Actuators

Actions on the environment are done by Actuators. AI Agent Architecture: (1) Agent - Perceives environment, takes actions, (2) Sensor - Perceives environment (input), (3) Actuator - Performs actions (output), (4) Controller - Decides what to do. Actuator Functions: (1) Mechanical movement - Motors, joints, (2) Environmental change - Heating, cooling, opening, (3) Output signals - Controlling other systems, (4) Physical interaction - Manipulating objects. Examples of Actuators: (1) Robot arm - Movement actuators, (2) Motor - Rotating or linear motion, (3) Valve - Controlling flow, (4) Speaker - Sound generation, (5) Light - Visual output. Sensor vs Actuator: (1) Sensor - Input device, (2) Actuator - Output device, (3) Together form agent interaction, (4) Feedback loop - Sensor→Control→Actuator. Control Loop: (1) Sense current state (Sensor), (2) Decide action (Controller), (3) Perform action (Actuator), (4) Observe result (Sensor), (5) Repeat. Performance Consideration: (1) Agent performance - How well objectives met, (2) Not an action mechanism, (3) Measure of success, (4) Evaluation metric. Types of Actuators: (1) Electric - Most common, fast, (2) Hydraulic - High power, (3) Pneumatic - Quick response, (4) Thermal - Heat generation. Robotics: (1) Actuators power robots, (2) Multiple actuators for complex tasks, (3) Coordination is key, (4) Precision and control important. Smart Systems: (1) Home automation - Lights, locks, thermostats, (2) Industrial control - Manufacturing equipment, (3) Autonomous vehicles - Steering, acceleration, braking, (4) Medical - Surgical robots. Real-world Applications: (1) Self-driving cars - Steering/braking actuators, (2) Drones - Motor actuators, (3) Smart homes - Light and door actuators, (4) Manufacturing - Robot arms. This demonstrates understanding of agent-environment interaction in AI systems.

Answer: B. To create machines that can perform tasks requiring human intelligence

Answer: B. An entity that perceives its environment and takes actions to achieve goals

Answer: B. Maintaining internal state

Answer: D. Acting emotionally

Answer: B. The next state is completely determined by the current state and agent's action

Answer: B. Acts based on current percepts only

Answer: A. Performance, Environment, Actuators, Sensors

Answer: C. Multiple agents operate in the same environment

Answer: D. Knowledge-based agent

Answer: B. Maximizes its expected utility

Answer: B. The environment does not change while the agent is deliberating

Answer: B. The next state depends partly on the agent's action and partly on random factors

Answer: B. The agent can see the complete state of the environment

Answer: C. O(d)

IDDFS space complexity is O(d) because only the current search path is maintained in memory.

Answer: C. O(d)

IDDFS space is O(d) storing only current search path.

Answer: D. All

CSP can be solved using search, heuristic search, or greedy approaches.

Answer: None of the options

The source’s answer, “State, function, goal, path, test”, is not one of the options.

Key components: initial state, successor function, goal path, goal test.

Answer: D. All

CSP can be solved using search, heuristic search, or greedy approaches.

Answer: B. Modeling the problem as states, actions, transition model, start state, and goal test, then searching for a path from start to goal

In classical AI planning and search, a problem is typically formulated as a state-space search. This means: (1) States: Abstract representations of the world at a given time. For instance, the board configuration in 8-puzzle or location of a robot in a grid. (2) Initial state: The state from which the agent starts. (3) Actions: The set of actions available in each state. Actions have preconditions (when they can be applied) and effects (how they change the state). (4) Transition model: A function that, given a state and an action, returns the resulting state (or distribution over states if stochastic). (5) Goal test: A procedure that checks whether a given state satisfies the goal condition (e.g., puzzle solved, destination reached). (6) Path cost function: Optional but often included to evaluate the cost of sequences of actions. Once the problem has been formulated in this way, it can be represented as a graph or tree, where nodes are states and edges are actions. Search algorithms like Breadth-First Search (BFS), Depth-First Search (DFS), Uniform Cost Search, A*, etc., then explore this graph to find a path from the initial state to some goal state. This abstraction separates problem description (what is the problem) from problem-solving strategy (how we search). The other options in the question refer to algorithm analysis, knowledge representation, or probabilistic reasoning, which are related AI topics but do not describe the standard state-space search formulation.

Answer: B. It provides an initial state, a set of actions, a transition model, a goal test, and optionally a path cost function

A well-defined search problem must specify enough information for a search algorithm to systematically explore possible solutions. The key components are: (1) Initial state: The starting point of the agent in the state space. Without this, the algorithm would not know where to begin. (2) Actions (successor function): For each state, we must know what actions are available and what successor states each action leads to. This is sometimes called the successor function. (3) Transition model: A description of the effect of each action, i.e., a function Result(s, a) that returns the state reached from state s after action a. In deterministic domains, this is a single next state; in stochastic domains, it might be a probability distribution over possible next states. (4) Goal test: A function that checks whether a given state satisfies the goal condition. This may be a simple predicate (“is the robot at target cell?”) or more complex. (5) Path cost function (optional but very common): A function that assigns a numeric cost to each path. Commonly, each action has a step cost, and path cost is the sum. This allows algorithms to distinguish between cheaper and more expensive solutions and define optimality. Once these are specified, a search algorithm can conceptually explore the state space and find solutions. There is no requirement that the problem be solvable in polynomial time or that it be represented in logic. In fact, many interesting search problems (like optimal route planning with complex constraints) are NP-hard or worse, but they are still well-defined search problems.

Answer: C. Informed search uses additional problem-specific knowledge (heuristics) to guide exploration toward the goal

Uninformed (or blind) search algorithms know only the problem definition in terms of state transitions and goal test; they do not have any domain-specific hints about where the goal is located in the state space. Examples include Depth-First Search (DFS), Breadth-First Search (BFS), Uniform Cost Search (if considered without heuristics), Depth-Limited Search, Iterative Deepening Search, and Bidirectional Search. They are systematic but can explore huge regions of the state space that are irrelevant to the goal, particularly in large or infinite spaces. Informed (heuristic) search algorithms, in contrast, exploit problem-specific knowledge to estimate how promising each state is with respect to reaching the goal. This knowledge is encoded in a heuristic function h(n), which approximates the cost (or distance) from node n to the nearest goal. Examples of informed search include: (1) Greedy Best-First Search, which selects nodes with smallest h(n) (most promising according to heuristic) regardless of cost so far; (2) A* Search, which selects nodes with minimal f(n) = g(n) + h(n), balancing cost so far (g(n)) and estimated remaining cost (h(n)); (3) Hill Climbing, which moves in the direction of decreasing heuristic cost (like gradient-based local search); (4) Simulated Annealing, which uses heuristic costs but occasionally allows uphill moves to escape local minima. Informed search does not guarantee optimality by default; only under conditions like an admissible and consistent heuristic does A* guarantee optimal solutions. So it is incorrect to claim that informed search always finds an optimal solution or that uninformed search never does. For example, BFS finds optimal solutions in unweighted graphs, and Uniform Cost Search finds optimal paths with non-negative step costs.

Answer: B. Breadth-First Search (BFS)

Breadth-First Search (BFS) explores the search tree level by level: it visits the start node, then all nodes at depth 1, then all nodes at depth 2, and so on. In a tree (no repeated states) with uniform step costs (e.g., cost = 1 per action), the first time BFS reaches a goal node, that node is at the shallowest depth among all goal nodes. Since path cost is proportional to depth when costs are uniform, the shallowest goal also has minimal path cost, so BFS is guaranteed to find an optimal solution under these conditions. Depth-First Search (DFS), on the other hand, dives as deep as possible along one path before backtracking. It may find a deep goal even if a shallower (cheaper) goal exists, so it is not optimal. Depth-Limited Search behaves like DFS but stops after a fixed depth limit; it can miss shallow solutions if the limit is poorly chosen and is not inherently optimal. Greedy Best-First Search uses a heuristic h(n) to choose the node that appears closest to the goal, ignoring path cost so far g(n); it often finds a solution quickly, but has no guarantee of optimality unless the heuristic happens to preserve ordering of true costs, which is rarely proven. For non-uniform step costs, Uniform Cost Search or A* with a consistent heuristic are used instead of BFS for optimality.

Answer: B. Breadth-First Search – complete and optimal when step costs are equal

Uninformed search algorithms have well-studied properties: (1) Depth-First Search (DFS) is neither complete in infinite-depth or cyclic state spaces nor optimal. DFS may get stuck going down an infinite path and never reach shallow solutions elsewhere. It also does not guarantee minimal path cost. (2) Breadth-First Search (BFS) is complete when the branching factor is finite and a solution exists. With uniform step costs, BFS is also optimal because it finds the shallowest goal. However, BFS has high memory cost because it stores all nodes at the current frontier. (3) Iterative Deepening Search (IDS) combines benefits of BFS and DFS. It performs depth-limited DFS repeatedly with increasing depth limits (0, 1, 2, …). IDS is complete and optimal under uniform step costs like BFS, but uses memory comparable to DFS (O(bd) time and O(bd) space instead of O(b^d) space for BFS). It does not require storing all frontier nodes simultaneously; instead, it repeatedly re-explores upper levels, which is surprisingly not too expensive because most nodes are at the deepest level. (4) Bidirectional Search simultaneously searches forward from the initial state and backward from the goal, aiming to meet in the middle. If both directions can be efficiently implemented and a predecessor function is available, bidirectional search can dramatically reduce time complexity from O(b^d) to roughly O(b^{d/2}) in each direction, but it is more complex and requires storing both frontiers. Therefore, the only correct pairing among the options is that BFS is complete and optimal when all step costs are equal.

Answer: C. A* is complete and optimal, expanding no node with f(n) greater than the cost of an optimal solution

A* search is an informed search algorithm that uses a combined evaluation function f(n) = g(n) + h(n), where g(n) is the cost from the start to node n, and h(n) is a heuristic estimate of the cost from n to a goal. Under certain conditions on the heuristic, A* enjoys strong guarantees: (1) Admissible heuristic: h(n) is admissible if it never overestimates the true minimal cost h*(n) to reach a goal from n (i.e., 0 ≤ h(n) ≤ h*(n) for all n). Under admissibility alone, A* is optimal in tree search (no repeated states), but in graph search, it may still re-expand nodes unless consistency is also satisfied. (2) Consistent (monotone) heuristic: h is consistent if for every edge (n, n') with cost c, h(n) ≤ c + h(n'). Intuitively, this enforces a triangle inequality: estimated cost from n to goal is no more than cost of going to successor n' plus estimated cost from n' to goal. Consistency implies admissibility and also ensures that f(n) values along any path are non-decreasing. For graph search, with a consistent heuristic, A* is complete and optimally efficient among all optimal algorithms using the same heuristic information: it is guaranteed to find an optimal solution (minimum path cost) and will not expand any node whose f(n) exceeds the optimal solution cost. It also ensures that each node needs to be expanded at most once. The algorithm terminates on finite graphs because: (a) only finitely many nodes have f(n) less than or equal to the optimal cost, and (b) A* expands them in increasing f(n) order. So the correct statement is that A* is complete and optimal and expands no node with f(n) greater than the optimal solution cost when h is both admissible and consistent.

Answer: C. Getting stuck in local maxima, plateaus, or ridges on the search landscape

Local search algorithms like hill climbing, steepest-ascent hill climbing, and sometimes simulated annealing operate directly on complete states and usually keep only a small number of states in memory. Their goal is not to discover an explicit path but to find a good solution state by iteratively improving the current state according to some objective or heuristic function. In hill climbing, from a current state, the algorithm evaluates neighboring states and moves to the neighbor that most improves the objective. This greedy local move can cause characteristic problems: (1) Local maxima: A state whose objective value is better than that of all its neighbors, but not the best possible globally. The algorithm has no incentive to move away because all neighbors are worse, so it becomes stuck. (2) Plateaus: A flat region where neighboring states have the same objective value (no ascent). The algorithm may wander randomly or terminate, failing to find a direction that leads upward. (3) Ridges: Regions where the optimal path to higher values is not aligned with any single coordinate direction, so naive local steps cannot easily ascend without coordinated moves. These issues mean hill climbing can fail to locate global optima even when they exist and be extremely sensitive to the starting state. Simulated annealing addresses these issues by sometimes accepting worse states with a probability controlled by a “temperature” parameter, allowing it to escape local maxima early in the search. Stochastic hill climbing and random restarts are additional strategies to reduce the chance of getting stuck, but they do not eliminate it entirely. The primary limitation is not memory or absence of goal tests but the local, greedy nature of the improvement process.

Answer: B. The best guaranteed outcome for the maximizing player assuming optimal play from both players

In adversarial search, particularly in deterministic two-player zero-sum games with perfect information (like chess, checkers, and tic-tac-toe), the minimax algorithm assigns a value to each game state. This minimax value is defined recursively: (1) For terminal states (game over), the value is given by the utility function (for example, +1 for a win, 0 for a draw, −1 for a loss from the perspective of the maximizing player). (2) For non-terminal states where it is the maximizing player’s turn, the minimax value is the maximum of the minimax values of its successor states (because the maximizing player chooses the move that maximizes the final outcome). (3) For non-terminal states where it is the minimizing player’s turn, the minimax value is the minimum of the minimax values of its successor states (because the opponent chooses the move that minimizes the maximizing player’s outcome). Thus, the minimax value of a node from the maximizing player’s perspective represents the utility of the game outcome under optimal play from both players: the maximizing player chooses moves to maximize this value, while the minimizing (opponent) player chooses moves to minimize it. It is a worst-case guarantee for the maximizing player: even if the opponent plays perfectly adversarially, the maximizing player can secure at least that utility by following the corresponding minimax strategy. It is not simply the maximum heuristic value among children (that would ignore the opponent’s choices) and not a probability of winning unless utilities are chosen in that way. It also does not represent a cost in the path-planning sense, but rather game-theoretic utility.

Answer: B. To reduce the number of nodes evaluated by pruning branches that cannot influence the final decision

Alpha–beta pruning is an optimization of the minimax algorithm that uses bounds to avoid exploring parts of the game tree that cannot possibly affect the final decision. It does not change the result of minimax; it only reduces the number of nodes evaluated. Two values are maintained during depth-first traversal: (1) alpha (α): The best value (highest lower bound) found so far along any path for the maximizing player. It is a lower bound on the final outcome that the maximizing player can guarantee given choices already considered. (2) beta (β): The best value (lowest upper bound) found so far along any path for the minimizing player. It is an upper bound on the outcome from the maximizing player’s perspective because the minimizing player will try to reduce the value. When exploring a node: (a) For a MAX node: if its current value becomes greater than or equal to β, further exploration of its remaining children is unnecessary, because the minimizing player (ancestor MIN node) will never allow this branch to be chosen (it has a better or equal alternative). This is called a beta cut-off. (b) For a MIN node: if its current value becomes less than or equal to α, further exploration is unnecessary because the maximizing player has a better or equal alternative already. This is called an alpha cut-off. With optimal move ordering, alpha–beta pruning can reduce the effective branching factor dramatically. In the best case, it prunes enough branches to make search complexity roughly O(b^{d/2}) instead of O(b^d), where b is branching factor and d is depth. This makes deeper lookahead feasible in practice. However, alpha–beta does not approximate minimax with heuristics (that would be evaluation functions at cut-off depth), nor does it guarantee reaching terminal states. Its purpose is purely to prune branches that cannot affect the final choice.

Answer: B. Breadth-first search (BFS)

Breadth-first search (BFS) is guaranteed to find the optimal solution in a state space search problem when there are no path costs. Why BFS is optimal without path costs: (1) Explores level by level, (2) Finds solution with minimum number of steps, (3) With equal cost (1) per step, minimum steps = minimum cost, (4) First solution found is optimal. BFS characteristics: (1) Complete - Always finds solution if exists, (2) Optimal - Finds best solution without path costs, (3) Systematic - Explores all possibilities at each depth, (4) Memory intensive - Stores all nodes at current level. Search algorithm comparison: (1) DFS - Not optimal, can go deep in wrong direction, (2) BFS - Optimal without path costs, (3) Uniform Cost Search - Optimal with path costs, (4) A* - Optimal with admissible heuristic, (5) Hill climbing - Local search, greedy, not optimal. When BFS is optimal: (1) Uniform step costs - Each step costs the same (1 unit), (2) Shortest path problem - Minimize number of steps, (3) Unweighted graph - All edges equal weight. Implementation: (1) Use queue (FIFO) for frontier, (2) Explore all neighbors at current level, (3) Move to next level when level exhausted. Time/space complexity: (1) Time: O(b^d) where b = branching factor, d = depth, (2) Space: O(b^d) - stores all frontier nodes. When not to use BFS: (1) With path costs - Use Uniform Cost Search, (2) With heuristic info - Use A*, (3) Very deep solutions - Memory problems, (4) When first solution acceptable - DFS faster (doesn't expand all). Real-world applications: (1) Shortest path in unweighted graphs, (2) Social networks - degrees of separation, (3) Puzzle solving - minimum moves, (4) Level-by-level exploration. This is fundamental to search algorithm theory.

Answer: B. b) Dijkstra's Algorithm and Greedy Best-First Search

A* search combines features of Dijkstra's Algorithm and Greedy Best-First Search. A* uses f(n) = g(n) + h(n), where g(n) is actual cost from start (Dijkstra's approach) and h(n) is heuristic estimate to goal (Greedy approach). This combination provides optimal pathfinding while being more efficient than Dijkstra alone. A* is widely used in game AI, robotics, and navigation systems.

Answer: C. Best-first search

The A* search algorithm uses best-first search strategy. A* evaluates nodes using f(n) = g(n) + h(n), where g(n) is the actual cost from start node and h(n) is the heuristic estimate to goal. This hybrid approach combines the optimality of Dijkstra's algorithm with the speed of greedy best-first search. A* always expands the most promising node first (best-first strategy), leading to optimal pathfinding when the heuristic is admissible. A* is widely used in game AI, robotics, and navigation systems.

Answer: D. All of these

Answer: A. Initial state, actions, transition model, goal test, path cost

Answer: B. A search through a space of possible states to find a goal state

Answer: A. The deepest unexpanded node

Answer: B. The shallowest unexpanded node

Answer: A. Variables, domains, and constraints

Answer: B. DFS and BFS

Answer: B. Heuristic information to guide the search

Answer: B. Reduces the number of nodes evaluated by minimax

Answer: A. f(n) = g(n) + h(n)

Answer: C. Adversarial search problems

Answer: C. A local search algorithm

Answer: B. The node closest to the goal according to the heuristic

Answer: B. Is a local search algorithm that can escape local optima

Answer: B. Searches from both the initial state and the goal state

Answer: D. Representational verification

Representational verification is not a standard property. The main properties are representational adequacy, inferential adequacy, and inferential efficiency.

Answer: A. Knowledge representation

Semantic networks are method for knowledge representation.

Answer: A. Knowledge representation

Semantic networks are method for knowledge representation in AI.

Answer: C. Data-driven

Forward chaining is data-driven inference starting from facts to reach conclusions.

Answer: B. Top-down

Backward chaining is top-down inference starting from goal to find supporting facts.

Answer: B. To encode information in a form that a computer can reason with effectively and efficiently

Knowledge Representation (KR) in AI is not merely about storing data in some structured form like a database. The primary objective is to encode knowledge about the world (objects, properties, relations, rules, and uncertainty) in a symbolic or probabilistic format that supports efficient reasoning and decision-making. An effective KR scheme should satisfy several requirements: (1) Representational adequacy: It must be expressive enough to capture the kinds of knowledge needed for the target domain: facts, taxonomies (is-a relations), part–whole relations, temporal and causal relations, procedural knowledge, default assumptions, and uncertainty. (2) Inferential adequacy: It must support sound and (ideally) complete inference mechanisms to derive new knowledge from existing knowledge. For example, resolution in propositional or first-order logic, or Bayesian inference in probabilistic models. (3) Inferential efficiency: Reasoning should not only be sound but also computationally feasible. Some logics are very expressive but make inference undecidable or intractable. KR often balances expressiveness against tractability by restricting the language (e.g., Horn clauses). (4) Acquisitional efficiency: It should be relatively easy for humans (or automated tools) to add, maintain, and modify knowledge. This is crucial in expert systems and large-scale knowledge bases. Examples of KR formalisms include propositional logic, first-order predicate logic, semantic networks, frames, ontologies, production rules, probabilistic graphical models (like Bayesian networks), and fuzzy logic systems. Each comes with its own syntax, semantics, and inference methods. The correct answer emphasizes that KR is about representing information in a form that supports effective and efficient reasoning, not just storing raw data or focusing solely on numbers.

Answer: B. Choosing a representation that balances expressiveness, efficiency, and ease of use

Several important issues arise when designing a knowledge representation for an AI system: (1) Expressiveness vs. efficiency: A very expressive language (such as full first-order logic with unrestricted quantification, higher-order predicates, or temporal operators) can represent complex facts and relationships, but inference may become undecidable or computationally intractable. More restricted languages (like Horn clauses or certain description logics) may be less expressive but allow efficient reasoning (polynomial or at least semi-decidable). Designers must choose a level of expressiveness that is “just enough” for the domain. (2) Reasoning support: The representation should support sound and preferably complete inference methods. It should be clear what kinds of questions can be answered (entailment, consistency, explanation, planning) and what algorithms are available (resolution, forward/backward chaining, belief propagation, etc.). (3) Handling incomplete and uncertain knowledge: Real-world domains rarely allow complete or certain information. KR must address defaults, exceptions, uncertainty, and inconsistency (through default logic, Bayesian networks, fuzzy logic, or non-monotonic reasoning). (4) Modularity and maintainability: Knowledge bases inevitably evolve. The representation should allow incremental additions and modifications without breaking the entire system. Good modularity (through frames, objects, or ontologies) helps maintain and scale knowledge. (5) Mapping to the real world: Symbols must be grounded in perception and action; otherwise, the system risks the “symbol grounding problem,” where internal symbols are manipulated without clear relation to external reality. (6) Computational properties: It is important to consider whether inference will run in acceptable time and memory limits for realistic problem sizes. The answer options that talk about purely numerical representation, avoiding inference, or arbitrary size limits do not capture these core issues. The central design problem is finding a good balance between expressiveness, efficiency, and usability.

Answer: B. P ∨ ¬P

In propositional logic, a tautology is a formula that evaluates to true under every possible truth assignment of its atomic propositions. Consider each candidate: (1) P ∧ ¬P: This formula is a contradiction. If P is true, then ¬P is false, and the conjunction is false. If P is false, ¬P is true, but the conjunction is still false. So P ∧ ¬P is false in all cases and is the opposite of a tautology. (2) P ∨ ¬P: This is an instance of the Law of the Excluded Middle. If P is true, then the disjunction is true; if P is false, then ¬P is true, so the disjunction is again true. There is no assignment of P that makes P ∨ ¬P false, so it is a tautology. (3) P → ¬P: This is not a tautology. If P is true, then ¬P is false, so the implication P → ¬P is false in that case. If P is false, the implication is true (because an implication with false antecedent is true), so the formula is true in some models and false in others, not a tautology. (4) ¬(P ∨ P): This simplifies to ¬P, which is clearly not always true because for P = false, ¬P is true, but for P = true, ¬P is false. Therefore, only P ∨ ¬P is a tautology. Understanding tautologies is important for logical equivalence, proof transformations, and simplifying formulas in automated reasoning systems.

Answer: A. Resolution is a sound and complete inference rule for propositional logic when applied to clauses in CNF

Resolution is a single inference rule that, when applied to a set of clauses in Conjunctive Normal Form (CNF), is both sound and complete for propositional logic. (1) Soundness means that any clause derived by resolution is logically entailed by the original set of clauses. That is, resolution never produces a false conclusion from true premises. The rule works as follows: from (A ∨ L) and (¬L ∨ B), resolution derives (A ∨ B), where L is a literal and ¬L is its negation. This new clause is a logical consequence of the premises. (2) Completeness means that if a formula logically entails a conclusion, then resolution will eventually be able to derive a contradiction when attempting to refute its negation (provided we systematically apply it in all possible ways). The standard use is refutation: to show that a set of clauses entails some sentence S, one adds ¬S to the knowledge base, converts everything to CNF, and repeatedly applies resolution. If the empty clause (contradiction) is derived, the original set entails S. For propositional logic, this approach is complete, meaning that if S is indeed entailed, resolution will eventually find a proof. Resolution becomes the basis for many SAT solvers and theorem provers in propositional logic. In first-order logic, resolution with unification can also be made complete (for refutation) but requires additional machinery such as Skolemization and handling of variables. Therefore, the correct statement is that resolution is sound and complete for propositional logic when used on CNF. The other options misunderstand its role or properties.

Answer: C. To find a substitution that makes different logical expressions identical so that literals can be resolved

In first-order predicate logic, resolution extends the propositional resolution rule by allowing variables and quantified formulas. To resolve two clauses that contain complementary literals (for example, P(t1, …, tk) and ¬P(s1, …, sk)), we must ensure that the predicates and their arguments match. However, because arguments can contain variables, we need a way to find a substitution for these variables that makes the two literals syntactically identical. Unification is the algorithmic process that finds such a substitution (if one exists). Specifically: (1) A substitution θ is a mapping from variables to terms (constants, variables, or function applications). (2) Applying θ to an expression E, written Eθ, replaces every variable in E according to θ. (3) Two expressions E1 and E2 are said to unify if there exists a substitution θ such that E1θ = E2θ (syntactically equal after substitution). (4) A unifier is such a substitution; a most general unifier (MGU) is a unifier that makes the fewest commitments, from which all other unifiers can be obtained by further substitutions. During resolution, when we want to resolve clauses C1 and C2 containing complementary literals L and ¬L', we first attempt to unify L and L'. If unification succeeds with substitution θ, we then apply θ to all literals in both clauses and form a new resolvent clause containing all remaining literals. For example, from clause1: P(x, f(y)) ∨ A and clause2: ¬P(a, f(z)) ∨ B, unification finds θ = {x/a, y/z}. Applying θ yields P(a, f(z)) and ¬P(a, f(z)), which can be resolved, producing Aθ ∨ Bθ. Unification is therefore central to first-order resolution because it allows matching of patterns with variables. The other options describe related preprocessing steps: conversion to CNF includes Skolemization and moving quantifiers, and variable renaming avoids confusion between bound variables, but unification specifically performs pattern matching for resolution.

Answer: A. It relates prior, likelihood, and posterior probabilities to update beliefs given evidence

Bayes' rule is fundamental in probabilistic reasoning and learning. It expresses how to update the probability of a hypothesis given observed evidence. In the formula P(A|B) = [P(B|A) P(A)] / P(B): (1) P(A) is the prior probability of hypothesis A before observing any evidence B. It encodes our initial belief. (2) P(B|A) is the likelihood: the probability of observing evidence B assuming that hypothesis A is true. It describes how compatible the evidence is with the hypothesis. (3) P(B) is the marginal probability of the evidence B under all possible hypotheses. It acts as a normalizing constant to ensure that posterior probabilities sum to 1. It can be computed as P(B) = Σ_i P(B|Ai) P(Ai), summing over mutually exclusive hypotheses Ai. (4) P(A|B) is the posterior probability: our updated belief in hypothesis A after we have seen evidence B. Bayes' rule says that posterior ∝ likelihood × prior. In AI, this formula is used in many contexts: (a) Naive Bayes classifiers for text classification and spam detection, (b) Bayesian networks for reasoning under uncertainty, (c) Bayesian parameter estimation in machine learning, (d) probabilistic robotics (localization, mapping), (e) medical diagnosis, and more. It does not express logical entailment; probabilities are degrees of belief, not truth values. It also does not correspond to maximum-likelihood estimation (which ignores priors) but rather to Bayesian inference that combines prior beliefs with evidence. Decision trees are based on information gain and splitting criteria, not directly on Bayes' rule.

Answer: B. A directed acyclic graph (DAG) where nodes are random variables and edges encode conditional dependencies

A Bayesian network, also called a belief network, is a compact representation of a joint probability distribution over a set of random variables using a directed acyclic graph (DAG). Its key components are: (1) Graph structure: Nodes correspond to random variables (e.g., Disease, Symptom1, Symptom2, Weather). Directed edges represent direct probabilistic influences, typically from cause to effect. The graph is acyclic, meaning there are no directed cycles. (2) Local conditional probability distributions (CPDs): For each node X with parents Pa(X), the network specifies P(X | Pa(X)). For root nodes (nodes without parents), these are just prior probabilities P(X). (3) Markov condition: The graph encodes conditional independence assumptions: each node is independent of its non-descendants given its parents. This drastically reduces the number of parameters required compared to a full joint distribution. For n binary variables, a full joint would need 2^n − 1 probabilities; a sparse Bayesian network needs far fewer. (4) Inference: Given evidence (observed variables), Bayesian networks can be used to compute posterior probabilities of query variables. Inference can be done exactly (variable elimination, belief propagation in polytrees) or approximately (sampling methods, loopy belief propagation) when the graph is complex. Bayesian networks are used for diagnostic reasoning (effect to cause), predictive reasoning (cause to effect), intercausal reasoning (explaining away), and mixed inference. They do not require full connectivity; in fact, sparse connectivity is desirable. They are not neural networks nor pure rule systems, although they can sometimes be compiled from conditional rules. They provide a principled way to model uncertainty and perform reasoning in domains such as medical diagnosis, fault detection, spam filtering, and robotics.

Answer: C. To prove the satisfiability or unsatisfiability of a given set of logical statements

The main goal of the resolution algorithm in inference is to prove the satisfiability or unsatisfiability of a given set of logical statements. Resolution mechanism: (1) Converts statements to CNF (Conjunctive Normal Form), (2) Applies resolution rule: if (A ∨ B) and (¬B ∨ C), then (A ∨ C), (3) Continues until proof found or contradiction reached, (4) Proves goal by contradiction - assumes negation and derives contradiction. Satisfiability: (1) Satisfiable - Exists assignment making all statements true, (2) Unsatisfiable - No assignment makes all true, always false (contradiction). Resolution process: (1) Start with knowledge base + negated goal, (2) Apply resolution rule repeatedly, (3) If empty clause (⊥) derived → unsatisfiable, (4) If no new clauses → satisfiable. Advantages: (1) Sound and complete for propositional logic, (2) Systematic approach, (3) Automated reasoning possible, (4) Basis for many theorem provers. Limitations: (1) Requires CNF conversion - can exponentially increase formulas, (2) Can be slow for large problems, (3) Not practical for very large knowledge bases. Applications: (1) Theorem proving - Automated proof generation, (2) SAT solving - SAT solver algorithms, (3) Logic programming - Prolog uses resolution, (4) AI reasoning - Automated reasoning systems. Related concepts: (1) Unification - Substitution making formulas identical, (2) Refutation - Proof by contradiction, (3) Inference rules - General logical deduction. Modern improvements: (1) DPLL algorithm - More efficient, (2) SMT solvers - Handle constraints, (3) CDCL - Conflict-driven clause learning. This is fundamental to automated reasoning and AI systems.

Answer: C. c) ¬

In FOPL (First Order Predicate Logic), ¬ (negation symbol) is used for negation/NOT operation. ∧ represents AND, ∨ represents OR, → represents implication. ¬P means 'not P'. Negation is fundamental to logic for expressing negative statements and negation normal form. Combined with other operators, negation allows expressing complex logical formulas.

Answer: B. The way knowledge is encoded for use by a computer system

Answer: B. Nodes and links

Answer: C. Data structures that represent stereotyped situations

Answer: A. The process of finding values for variables that make two expressions identical

Answer: A. Statements that are either true or false

Answer: B. A statement that is always true

Answer: B. Probabilistic inference

Answer: B. Variables, quantifiers, and predicates

Answer: B. A graphical model that represents probabilistic relationships

Answer: C. Both universal and existential quantification

Answer: A. A method for proving theorems

Answer: D. A and B

Inference engines use forward chaining and backward chaining.

Answer: C. Both A and B

Inference engines use forward chaining and backward chaining strategies.

Answer: C. Analyzing and interpreting the meaning of natural language text

The main goal of natural language understanding (NLU) is analyzing and interpreting the meaning of natural language text. NLU vs related fields: (1) NLP (Natural Language Processing) - Broader field including NLU, (2) NLU - Focused on semantic understanding, (3) Named Entity Recognition - Specific subtask of NLU, (4) Machine Translation - Different task (translation, not understanding). NLU tasks: (1) Semantic parsing - Converting text to logical form, (2) Intent recognition - Understanding user intention, (3) Sentiment analysis - Understanding emotional tone, (4) Question answering - Understanding question and finding answer, (5) Paraphrase detection - Recognizing equivalent meanings. How NLU works: (1) Tokenization - Break into words, (2) POS tagging - Mark word types, (3) Syntactic parsing - Build parse tree, (4) Semantic analysis - Extract meaning, (5) Pragmatic analysis - Use context. Challenges: (1) Ambiguity - Words with multiple meanings, (2) Idioms - Non-literal meanings, (3) Context - Same sentence different meanings in different contexts, (4) Negation and modality - Handle 'not', 'should', 'could'. Different from: (1) Translation - Understanding different, translation different problem, (2) Response generation - Separate NLG task, (3) Entity extraction - Part of understanding but not complete, (4) Text classification - Simpler than understanding. Modern NLU: (1) Deep learning models - BERT, GPT-based, (2) Transformer models - Attention-based, (3) Pre-trained models - Transfer learning, (4) End-to-end learning - Learn end-to-end instead of pipeline. Applications: (1) Chatbots - Understand user queries, (2) Virtual assistants - Alexa, Siri, Google Assistant, (3) Information extraction - Pull facts from text, (4) Document understanding - Analyze documents. This is key to AI systems that interact with humans.

Answer: C. c) Computer Vision

Computer Vision is the AI application that helps self-driving cars. It enables cars to perceive the environment through cameras, recognize objects, lanes, traffic signals, and obstacles. Computer vision combined with machine learning allows vehicles to make driving decisions. NLP handles text/speech, Expert Systems use rules, and general Machine Learning supports vision, but computer vision specifically addresses autonomous vehicle perception.

Answer: D. d) The emotion or opinion expressed in text

Sentiment analysis determines the emotion or opinion expressed in text. It classifies text as positive, negative, or neutral. Applications include social media monitoring, customer feedback analysis, and product reviews. Sentiment analysis uses NLP techniques and machine learning to extract subjective information from unstructured text. It's widely used in marketing, reputation management, and customer service.

Answer: B. A system that uses knowledge from human experts to solve problems

Answer: B. Facts and relationships

Answer: A. Knowledge base and inference engine

Answer: B. Understanding the meaning of human language

Answer: B. How to do something

Answer: B. Processing human language

Answer: B. Understanding images and visual information

Answer: C. AI, mechanical engineering, and electronics

Answer: C. Multiple methods including interviews, observation, and document analysis

Answer: D. All

HMM used in supervised, unsupervised, and reinforcement learning.

Answer: B. Hierarchical clustering

Hierarchical clustering permits convenient graphical display through dendrograms. A dendrogram is a tree-like diagram that shows the hierarchical clustering results clearly. Hierarchical clustering approaches: (1) Agglomerative (bottom-up) - Starts with individual points, merges closest pairs iteratively, (2) Divisive (top-down) - Starts with one cluster, recursively splits. Dendrogram visualization: (1) Horizontal axis - Data points or clusters, (2) Vertical axis - Distance/linkage criterion at which merging occurs, (3) Cutting horizontal line - Determines final number of clusters. Advantages of hierarchical clustering: (1) Produces dendrograms for visual analysis, (2) Reveals hierarchical relationships, (3) No need to specify cluster count beforehand, (4) Multiple granularity levels, (5) Can extract any number of clusters by cutting at different heights. Other clustering methods' visualization: (1) Partition-based (K-means) - Requires 2D/3D reduction for visualization, (2) Probabilistic (EM) - Requires separate visualization techniques, (3) Agglomerative - Same as hierarchical (generates dendrograms). Practical applications: (1) Gene expression analysis - Biological sample relationships, (2) Customer segmentation - Market analysis, (3) Species phylogeny - Evolutionary relationships. The dendrogram is one of the most intuitive visualizations in data analysis, making hierarchical clustering popular for exploratory data analysis.

Answer: A. Variance^0.5 or C. Variance^0.5

Two options have the same text; both match the source’s answer.

The standard deviation is the square root of variance, or Variance^0.5. Statistical Measures: (1) Variance - Average squared deviation from mean, (2) Standard Deviation - Square root of variance, (3) Both measure data spread, (4) Different units and scales. Mathematical Definition: (1) Variance σ² = Σ(xi - mean)² / n, (2) Standard Deviation σ = √Variance = √(σ²), (3) For sample: divide by (n-1) not n, (4) Result in same units as data. Why Square Root: (1) Variance uses squared deviations (quadratic units), (2) Square root brings back to original units, (3) Standard deviation more interpretable, (4) Standard for statistical reporting. Numerical Example: (1) Data: 2, 4, 6, 8, 10, (2) Mean = 6, (3) Deviations: -4, -2, 0, 2, 4, (4) Squared: 16, 4, 0, 4, 16, (5) Variance = 40/5 = 8, (6) Standard Deviation = √8 ≈ 2.83. Properties: (1) Always non-negative, (2) Same units as data, (3) Larger spread → larger standard deviation, (4) Can be compared across datasets. Uses: (1) Confidence intervals - ±1 SD ≈ 68% data, (2) Anomaly detection - Values > 3 SD are outliers, (3) Quality control - Process variation, (4) Finance - Risk measurement (volatility). Why Not Other Options: (1) Variance^2 - Would be variance squared, (2) Variance^0.25 - Fourth root, no statistical meaning, (3) Wrong formula doesn't match relationship. Normal Distribution: (1) 68% within ±1 SD, (2) 95% within ±2 SD, (3) 99.7% within ±3 SD, (4) SD defines distribution shape. This is fundamental statistical concept used widely.

Answer: B. Supervised

Naive Bayes is supervised learning algorithm using labeled training data.

Answer: A. Classification

Speech recognition is classification task predicting word categories.

Answer: B. Regression

Movie rating prediction is regression predicting continuous values.

Answer: B. Good training, bad new

Overfitting occurs when model is too complex, memorizing training data including noise.

Answer: A. A type of learning algorithm that uses fuzzy logic to handle uncertain or imprecise data

Fuzzy learning is a type of learning algorithm that uses fuzzy logic to handle uncertain or imprecise data. Fuzzy logic vs classical logic: (1) Classical - True or false (binary), (2) Fuzzy - Degrees of truth (0 to 1 continuous), (3) Allows partial truth - Something 70% true, 30% false. Fuzzy sets: (1) Members have degree of membership (0 to 1), (2) Unlike classical sets where membership is binary, (3) Example: temperature 'warm' - could be 25° (0.6 membership), 28° (0.9 membership). Fuzzy learning techniques: (1) Fuzzy rule-based systems - Rules with fuzzy conditions/conclusions, (2) Fuzzy decision trees - Decision trees with fuzzy splits, (3) Fuzzy neural networks - Combine neural nets with fuzzy logic, (4) Fuzzy c-means - Clustering with fuzzy membership. Why fuzzy learning needed: (1) Real-world data is imprecise, (2) Human reasoning is approximate, (3) Some concepts inherently vague, (4) Transitions are gradual not sudden. Advantages: (1) Handles uncertainty naturally, (2) Mimics human reasoning, (3) Works with imprecise data, (4) Interpretable rules. Applications: (1) Control systems - Temperature control, washing machine, (2) Medical diagnosis - Symptoms have fuzzy relationships, (3) Financial - Risk assessment with uncertainty, (4) Image recognition - Fuzzy boundaries between objects. Different from: (1) Probabilistic learning - Uses probability, not fuzzy membership, (2) Bayesian learning - Probability-based inference, (3) Crisp neural networks - Binary True/False. Fuzzy vs probabilistic: (1) Fuzzy represents vagueness - inherent imprecision, (2) Probability represents uncertainty - lack of information. Modern usage: (1) Often combined with other techniques, (2) Explicit fuzzy learning less common, (3) Used in specialized domains (control, decision support), (4) Integrated into neural networks as fuzzy layers. This addresses real-world imprecision.

Answer: A. Supervised learning

Learning using labelled data is called Supervised learning. Supervised Learning Concept: (1) Training data has input-output pairs, (2) Model learns to map inputs to correct outputs, (3) Teacher provides correct answers (labels), (4) Goal - Minimize prediction error. Data Structure: (1) Input features (X) - Independent variables, (2) Output labels (Y) - Correct answers, (3) Training set - Data for learning, (4) Test set - Data for evaluation. Learning Process: (1) Present input-output pair, (2) Model predicts output, (3) Compare with actual output, (4) Adjust model to reduce error. Types of Supervised Learning: (1) Regression - Continuous output (house price), (2) Classification - Discrete output (spam/not spam), (3) Multi-class - Multiple categories (image labels), (4) Multi-label - Multiple outputs per input. Algorithms: (1) Decision Trees - Hierarchical decisions, (2) Support Vector Machines - Optimal boundary finding, (3) Neural Networks - Pattern learning, (4) Regression - Linear/polynomial fitting. Example: (1) Email filter - Labeled spam/not spam emails, (2) Image recognition - Labeled image categories, (3) Medical diagnosis - Labeled patient outcomes, (4) Stock prediction - Historical prices and trends. Advantages: (1) Generally more accurate, (2) Well-defined success metric, (3) Faster convergence, (4) Better generalization. Disadvantages: (1) Requires labeled data, (2) Expensive to obtain labels, (3) Limited to available labels, (4) Errors in labels propagate. vs Other Learning Types: (1) Unsupervised - No labels, find patterns, (2) Reinforcement - Agent learns through rewards, (3) Semi-supervised - Mix of labeled and unlabeled. Data Requirements: (1) Quality labels crucial, (2) Representative sampling needed, (3) Class balance important, (4) Label consistency required. This is the most common machine learning paradigm.

Answer: A. To convert crisp input values into fuzzy sets

The function of a fuzzifier in fuzzy logic systems is to convert crisp (precise, numerical) input values into fuzzy sets. A crisp input like '25 degrees' is converted to fuzzy sets like 'warm' and 'cool' with membership values (e.g., 0.6 warm, 0.4 cool). The fuzzifier bridges the gap between the real world (precise measurements) and fuzzy logic (linguistic variables). After fuzzy inference, a defuzzifier converts fuzzy output back to crisp values for system control. Fuzzification is the first step in fuzzy logic processing.

Answer: C. K-means clustering

K-means clustering is an example of unsupervised learning. K-means finds natural groupings (clusters) in unlabeled data by minimizing within-cluster variance. It requires no labeled training data, only input data points. Decision trees, SVMs, and neural networks for classification are supervised learning methods that require labeled training data. K-means is used for customer segmentation, image compression, and pattern discovery. Other unsupervised learning examples include hierarchical clustering, DBSCAN, and principal component analysis (PCA).

Answer: B. Labeled data

Answer: B. The field of study that gives computers the ability to learn without being explicitly programmed

Answer: B. Unlabeled data

Answer: B. Classification and regression

Answer: C. Learning from rewards and punishments

Answer: B. Features are independent given the class

Answer: B. Degrees of truth

Answer: C. Biological evolution

Answer: B. Maps inputs to outputs using fuzzy logic

Answer: A. Selection, crossover, mutation

Answer: B. Evaluates how good a solution is

Answer: C. Cannot explain results

ANNs cannot explain their decision-making process, acting as black boxes despite good predictions.

Answer: B. Minimize by descent

Gradient descent minimizes function iteratively moving in steepest descent direction.

Answer: A. σ(x) = 1/(1+e^-x)

Sigmoid σ(x) = 1/(1+e^-x) maps inputs to (0,1) creating S-curve.

Answer: C. Softmax

Softmax function classifies into 3 or more classes.

Answer: B. max(0,x)

ReLU (Rectified Linear Unit) outputs max(0,x) for activation.

Answer: B. Backward sink to source

Backpropagation propagates errors backward from output to input layers.

Answer: C. Cannot explain

ANNs cannot explain their decision-making process (black box problem).

Answer: C. Input, hidden, output

Neural networks have input layer, hidden layers, and output layer.

Answer: B. Constant weight

Bias is constant weight added to neuron output for shifting activation.

Answer: B. Connection strength

Weights represent connection strength between neurons in network.

Answer: D. All

Number of layers, neurons, and weights determine NN model complexity.

Answer: B. Complete dataset pass

Epoch is one complete pass through entire training dataset.

Answer: B. Controls step size

Learning rate controls step size in gradient descent optimization.

Answer: D. All

Dropout removes neurons randomly, reducing overfitting through regularization.

Answer: D. All

Batch size affects training stability, memory, and convergence rate.

Answer: A. Classification loss

Cross-entropy is loss function for classification tasks.

Answer: D. Recurrent neural network (RNN)

Recurrent neural network (RNN) is commonly used for processing sequential data like time series and natural language. RNN characteristics: (1) Recurrent connections - Output fed back as input, (2) Memory - Can remember previous inputs, (3) Processes sequences - Handles variable-length inputs, (4) Hidden state - Maintains context. RNN mechanics: (1) At each step t, takes input x(t) and previous hidden state h(t-1), (2) Computes new hidden state h(t) based on both, (3) Produces output y(t), (4) Hidden state acts as memory of sequence. RNN variants: (1) LSTM (Long Short-Term Memory) - Handles long-term dependencies, (2) GRU (Gated Recurrent Unit) - Simplified LSTM, (3) Bidirectional RNN - Processes forward and backward. Why RNN for sequences: (1) Maintains sequence history, (2) Can model dependencies, (3) Variable-length input handling, (4) Weight sharing across time steps. Applications: (1) Machine translation - Sequence-to-sequence, (2) Speech recognition - Temporal acoustic features, (3) Time series prediction - Stock prices, weather, (4) Language modeling - Next word prediction, (5) Text generation. Advantages: (1) Handles variable-length sequences, (2) Captures temporal dependencies, (3) Shared weights reduce parameters. Limitations: (1) Vanishing/exploding gradient problem - Hard to train, (2) LSTM/GRU more complex, (3) Slower than feedforward. Other architectures: (1) FNN - No memory, only static inputs, (2) SOM - Unsupervised clustering, not sequential, (3) RBFN - Local basis functions, not sequential. Modern developments: (1) Transformer architecture - Attention-based, (2) BERT, GPT - Pre-trained language models, (3) Attention mechanisms - Replace RNN in many applications. This is fundamental to natural language and time series processing.

Answer: C. c) Model converges very slowly

If the learning rate is too low, the model converges very slowly. Gradient descent takes tiny steps toward optimal weights, requiring many iterations. While low learning rates avoid overshooting, they waste computational resources and time. Conversely, too high a learning rate causes the model to diverge (jump over optimal values). Optimal learning rate balances convergence speed with stability. Learning rate scheduling adaptively adjusts this value during training.

Answer: D. Sign

The Sign activation function is used in Hopfield networks. Hopfield Network Characteristics: (1) Recurrent neural network, (2) Auto-associative memory, (3) Binary outputs (0 and 1, or -1 and +1), (4) Uses Sign function for binary classification. Sign Function: (1) Output = -1 if input < 0, (2) Output = +1 if input > 0, (3) Produces binary classification, (4) Threshold at zero. Why Sign for Hopfield: (1) Binary memory storage, (2) Convergence to stable states, (3) Implements attractors, (4) Pattern recognition capability. Hopfield Network Operation: (1) Update rule: If sum of weighted inputs > threshold, fire 1, (2) Energy function minimization, (3) Converges to stable patterns, (4) Used for pattern completion. Comparison with Other Functions: (1) ReLU - f(x) = max(0,x), continuous, (2) Sigmoid - f(x) = 1/(1+e^-x), smooth 0-1 output, (3) Softmax - Probabilistic multi-class output, (4) Sign - Binary discrete output. Network Properties: (1) Bidirectional connections (recurrent), (2) Symmetric weights, (3) Diagonal zeros (no self-connections), (4) Can store multiple patterns. Learning Process: (1) Hebb rule for weight learning, (2) Can memorize patterns, (3) Noise tolerance, (4) Limited storage capacity. Applications: (1) Pattern recognition, (2) Content-addressable memory, (3) Optimization problems, (4) Image restoration. Limitations: (1) Local minima (spurious patterns), (2) Limited storage capacity, (3) Slow convergence, (4) Binary restriction. This demonstrates activation function selection for specific network architectures.

Answer: C. Sign function

The sign function is the activation function commonly used in Hopfield neural networks. The sign function outputs +1 or -1 (binary values) based on whether the input is positive or negative. Hopfield networks are recurrent neural networks used for pattern recognition and associative memory. The sign function (also called threshold or step function) enables Hopfield networks to converge to stable states that represent stored patterns. ReLU and Sigmoid produce continuous values. Softmax produces probability distributions. The binary nature of the sign function is essential for Hopfield network operation.

Answer: B. Biological neural networks

Answer: B. A mathematical model of a biological neuron

Answer: B. Determines the output of a neuron given its inputs

Answer: B. A single-layer neural network

Answer: B. Minimize the error function

Answer: A. Determines how quickly the network learns

Answer: B. Training single-layer networks

Answer: B. A single-layer neural network with a linear activation function

Answer: A. Neurons that fire together, wire together

Answer: C. Input, hidden, and output layers

Answer: B. Train multi-layer networks

Answer: B. A recurrent network