Unit 9: Artificial Intelligence (AI)

ARTIFICIAL INTELLIGENCE

UGC NET / JRF — Computer Science & Applications

High-Yield Study Notes & PYQ-Pattern Workbook (Unit 9)

Contents

Ch.1 — Introduction to AI

Ch.2 — Intelligent Agent

Ch.3 — Problem Solving

Ch.4 — Knowledge Representation & Expert System

Ch.5 — NLP & Planning in AI

Ch.6 — Genetic Algorithms & ANN

Ch.7 — Fuzzy Set

One-Shot Revision

Mini Mock Test (Answer Key)

Beginner → Concept → NET-level → JRF-level. Compact by design — exam value over page count.

Chapter 1 — Introduction to AI

1.1 What is AI?

SimpleArtificial Intelligence is the field of building machines/programs that can perform tasks normally requiring human intelligence — reasoning, learning, perception, language understanding.

Approach	Idea
Thinking humanly	Mimic actual human cognitive processes (cognitive modeling)
Thinking rationally	Follow correct LOGICAL reasoning ("laws of thought")
Acting humanly	Behave indistinguishably from a human (the Turing Test approach)
Acting rationally	Do whatever action best achieves the agent's goals (the RATIONAL AGENT approach — most widely adopted in modern AI)

JRF trap"Acting rationally" is the dominant modern approach (used as the foundation of most textbooks/AI systems) BECAUSE it doesn't require the AI to mimic human thought exactly (unlike "thinking/acting humanly") — a rational agent just needs to make the BEST decision given its knowledge, even if a human would decide differently. Confusing "acting humanly" (Turing Test) with "acting rationally" (goal-optimal behaviour) as if they were the same approach is a common JRF trap.

1.2 The Turing Test

IdeaProposed by Alan Turing (1950) as an OPERATIONAL test for machine intelligence: a human judge has text-only conversations with both a human and a machine (without knowing which is which); if the judge cannot reliably tell them apart, the machine is said to have passed.

JRF trapThe Turing Test evaluates "acting humanly" — it does NOT require the machine to actually think, reason logically, or be self-aware, only to convincingly IMITATE human conversational behaviour. This distinction (imitation vs genuine reasoning) is frequently used to critique the Turing Test's validity as a true measure of intelligence — a common essay/conceptual JRF angle.

1.3 AI Application Areas (Overview)

• Expert Systems: mimic decision-making of a human domain expert (e.g., medical diagnosis).
• NLP: understanding/generating human language.
• Computer Vision: interpreting visual data (images/video).
• Robotics: physical agents perceiving and acting in the real world.
• Machine Learning: systems that improve performance through experience/data (a major sub-field, not synonymous with all of AI).

NET pointMachine Learning is a SUBSET of AI (a specific approach where the system learns patterns from data), NOT the same thing as AI itself — AI is the broader goal (intelligent behaviour), ML is one (very popular, modern) technique to achieve it. Older AI systems (e.g., rule-based expert systems) used NO machine learning at all.

MUST REMEMBER — Chapter 1

• 4 AI approaches: Thinking humanly, Thinking rationally, Acting humanly (Turing Test), Acting rationally (dominant modern approach).
• Turing Test tests IMITATION of human behaviour, not genuine reasoning/self-awareness.
• Machine Learning ⊂ AI — ML is one technique to achieve AI, not synonymous with it.

DON'T CONFUSE

• Acting humanly (imitate humans, Turing Test) vs Acting rationally (optimize toward goals — the dominant modern framing).
• AI (the broad field/goal) vs Machine Learning (one specific technique within AI).

JRF CHALLENGE ZONE — Chapter 1

1. The Turing Test primarily evaluates: (a) Logical reasoning ability (b) Whether a machine can imitate human conversational behaviour (c) Machine learning accuracy (d) Self-awareness
Answer: (b)

2. Which statement is TRUE? (a) AI and Machine Learning are exactly the same thing (b) Machine Learning is a subset of AI (c) AI is a subset of Machine Learning (d) They are unrelated fields
Answer: (b)

Practice Questions — Chapter 1 (6)

1. List the four classical approaches/definitions of AI.
   Ans: Thinking humanly, thinking rationally, acting humanly, acting rationally
2. Which approach does the Turing Test correspond to?
   Ans: Acting humanly
3. Why is "acting rationally" considered the dominant modern approach to AI?
   Ans: Because it only requires optimal goal-directed behaviour, not mimicking human thought processes exactly
4. Who proposed the Turing Test, and in what year?
   Ans: Alan Turing, in 1950
5. Is Machine Learning the same as Artificial Intelligence?
   Ans: No — Machine Learning is a subset/technique within the broader field of AI
6. Give one example of an AI system that uses no machine learning at all.
   Ans: A rule-based expert system

---

Chapter 2 — Intelligent Agent

2.1 Agent, Rationality & PEAS

SimpleAn agent perceives its ENVIRONMENT through sensors and acts upon it through actuators. A RATIONAL agent selects the action expected to MAXIMIZE its performance measure, given what it has perceived so far and its built-in knowledge.

PEAS framework (used to specify a task environment) Performance measure — how success is judged Environment — the world the agent operates in Actuators — how the agent acts Sensors — how the agent perceives

JRF trapRationality is NOT the same as omniscience or perfection — a rational agent can still make a "wrong" decision if its available information was incomplete or misleading; rationality only requires the BEST possible decision GIVEN what it has perceived, not the objectively best decision with perfect hindsight. This "rational ≠ omniscient/perfect" distinction is frequently tested.

2.2 Types of Agents (by Architecture)

Agent Type	Idea
Simple Reflex	Acts using ONLY the current percept (condition-action rules) — no memory of the past
Model-Based Reflex	Maintains an INTERNAL STATE/model of the world to handle PARTIALLY observable environments
Goal-Based	Chooses actions based on whether they help achieve an explicit GOAL (considers the future, not just the present)
Utility-Based	Chooses actions that maximize a UTILITY function — handles tradeoffs between MULTIPLE competing/conflicting goals
Learning Agent	Improves its own performance over time via a "critic", "learning element", and "performance element" working together

JRF trap — goal-based vs utility-basedA Goal-based agent only knows whether a state IS or ISN'T a goal (binary success/failure). A Utility-based agent can compare DIFFERENT successful states and prefer the BETTER one (e.g., a FASTER or SAFER route, not just "any route that reaches the destination"). This "binary goal vs graded utility" distinction is a very common JRF trap.

2.3 Environment Types

Property	Distinction
Fully vs Partially Observable	Can sensors access the COMPLETE state at each point? (Chess=fully; Poker=partially, opponent's cards hidden)
Deterministic vs Stochastic	Does the next state depend ENTIRELY on current state+action? (Chess=deterministic; Dice game=stochastic)
Episodic vs Sequential	Is each action independent of previous ones? (Image classification=episodic; Chess=sequential, moves depend on history)
Static vs Dynamic	Does the environment change WHILE the agent is deciding? (Crossword=static; Self-driving car=dynamic)
Discrete vs Continuous	Finite, distinct states/actions, or infinitely many? (Chess=discrete; Robot navigation=continuous)
Single vs Multi-agent	Does the agent operate alone, or with/against other agents? (Solving a puzzle=single; Chess=multi (competitive))

JRF trapThe HARDEST/most complex environment type combination is: Partially observable + Stochastic + Sequential + Dynamic + Continuous + Multi-agent (e.g., real-world autonomous driving). NET/JRF frequently gives a SCENARIO (e.g., "an autonomous taxi in city traffic") and asks you to classify it correctly across ALL SIX dimensions simultaneously.

MUST REMEMBER — Chapter 2

• PEAS: Performance measure, Environment, Actuators, Sensors.
• Rational = best decision given available info, NOT omniscient/perfect.
• Agent hierarchy: Simple Reflex → Model-Based Reflex (adds internal state) → Goal-Based (adds future/goals) → Utility-Based (adds preference/tradeoffs) → Learning (adds self-improvement).
• Goal-based = binary (goal or not); Utility-based = graded comparison between successful outcomes.
• 6 environment dimensions: Observable(full/partial), Deterministic/Stochastic, Episodic/Sequential, Static/Dynamic, Discrete/Continuous, Single/Multi-agent.

DON'T CONFUSE

• Goal-based agent (binary success) vs Utility-based agent (graded preference between successes).
• Episodic (independent actions) vs Sequential (actions depend on history) — easy to conflate with static/dynamic.

JRF CHALLENGE ZONE — Chapter 2

1. An agent that can compare two different successful outcomes and prefer the better one is: (a) Simple reflex (b) Goal-based (c) Utility-based (d) Model-based only
Answer: (c)

2. In Poker, the environment (with respect to opponents' cards) is: (a) Fully observable (b) Partially observable (c) Deterministic (d) Episodic
Answer: (b)

3. Is a rational agent guaranteed to always make the objectively correct decision? (a) Yes, always (b) No — only the best decision given its available information (c) Only in deterministic environments (d) Only if it's a learning agent
Answer: (b)

Practice Questions — Chapter 2 (7)

1. What does PEAS stand for?
   Ans: Performance measure, Environment, Actuators, Sensors
2. Does rationality require perfect/omniscient decision-making?
   Ans: No — only the best decision given the information actually available to the agent
3. What additional capability does a Model-Based Reflex agent have over a Simple Reflex agent?
   Ans: It maintains an internal state/model of the world, allowing it to handle partially observable environments
4. Differentiate a Goal-based and a Utility-based agent.
   Ans: Goal-based only checks if a state is a goal (binary); utility-based can compare and prefer better successful outcomes
5. Is Chess fully or partially observable? Is Poker?
   Ans: Chess: fully observable; Poker: partially observable (opponents' cards hidden)
6. Differentiate an episodic and a sequential environment.
   Ans: In episodic environments each action is independent of previous ones; in sequential environments actions depend on history
7. What three components work together in a Learning Agent?
   Ans: Critic, learning element, and performance element

---

Chapter 3 — Problem Solving

3.1 State Space Search — Formulation

IdeaA search problem is defined by: Initial State, Actions (available at each state), Transition Model (result of an action), Goal Test, and Path Cost (numeric cost of a path).

3.2 Uninformed (Blind) Search

Algorithm	Idea	Complete?	Optimal?
BFS	Expands shallowest node first (Queue)	Yes	Yes (if uniform step cost)
DFS	Expands deepest node first (Stack)	No (infinite paths)	No
Uniform Cost Search (UCS)	Expands LOWEST path-cost node first (priority queue)	Yes	Yes
Iterative Deepening DFS (IDDFS)	Repeats DFS with increasing depth limits	Yes	Yes (if uniform cost)

JRF trapIDDFS combines DFS's LOW MEMORY usage (O(bd), linear in depth) with BFS's COMPLETENESS/OPTIMALITY guarantee — it seems wasteful (re-exploring shallow levels repeatedly) but the overhead is actually SMALL in a big-O sense, since the bottom level dominates total node count in an exponentially-branching tree. "Why isn't IDDFS's repeated work actually a big problem" is a frequently tested JRF conceptual point.

3.3 Informed (Heuristic) Search

Greedy Best-First Search: expands node with lowest h(n) [estimated cost to goal] — fast but NOT optimal A* Search: expands node with lowest f(n) = g(n) + h(n) g(n) = actual cost from start to n h(n) = estimated (heuristic) cost from n to goal

JRF trap — admissibilityA* is GUARANTEED OPTIMAL only if the heuristic h(n) is ADMISSIBLE — meaning h(n) NEVER OVERESTIMATES the true remaining cost to the goal (h(n) ≤ actual cost). An inadmissible (overestimating) heuristic can make A* return a SUBOPTIMAL solution. This "admissibility is what guarantees A*'s optimality" fact is THE most tested A* concept.

JRF-level numerical — A* traceStart node S, goal G. g(S)=0, h(S)=10 (estimate). S has neighbours A(edge cost 2, h(A)=6) and B(edge cost 4, h(B)=5). f(A) = g(A)+h(A) = 2+6 = 8. f(B) = g(B)+h(B) = 4+5 = 9. A* expands A first (lower f). This "compute f(n)=g(n)+h(n) for each frontier node, expand the lowest" trace is the standard A* JRF numerical.

NET pointA CONSISTENT (monotonic) heuristic satisfies h(n) ≤ cost(n,n') + h(n') for every neighbour n' — every consistent heuristic is automatically admissible, but not vice versa. Consistency guarantees A* never needs to re-expand a node (found path to any node is already optimal the first time it's expanded).

3.4 Adversarial Search — Minimax & Alpha-Beta Pruning

IdeaMinimax: used for two-player zero-sum games. MAX player tries to MAXIMIZE the score; MIN player tries to MINIMIZE it. The algorithm recursively backs up values from the leaves (terminal states) to the root, assuming BOTH players play optimally.

JRF-level numerical — minimax treeA game tree: MAX node has two MIN children; each MIN child has two leaf children. Left MIN child's leaves: 3, 5 → MIN picks 3. Right MIN child's leaves: 6, 2 → MIN picks 2. MAX node picks the BETTER (higher) of its two MIN children's results: max(3,2) = 3. This "propagate min at MIN levels, max at MAX levels, bottom-up" technique is the standard minimax JRF numerical.

JRF trap — Alpha-Beta pruningAlpha-Beta pruning gives the EXACT SAME result as plain minimax (it's an OPTIMIZATION, not an approximation) — it just skips evaluating branches that CANNOT possibly influence the final decision (using α=best value MAX can guarantee so far, β=best value MIN can guarantee so far; prune when α≥β). "Alpha-beta doesn't change the answer, only the efficiency" is a very frequently tested clarification.

MUST REMEMBER — Chapter 3

• BFS: complete+optimal(uniform cost), high memory. DFS: low memory, not complete/optimal. UCS: expands lowest path-cost. IDDFS: DFS memory + BFS completeness.
• Greedy best-first: uses only h(n), fast but not optimal. A*: uses f(n)=g(n)+h(n).
• A* is optimal ONLY with an admissible heuristic (h(n) never overestimates true cost).
• Consistent heuristic ⟹ admissible (not vice versa); consistency avoids re-expansion.
• Minimax: MAX maximizes, MIN minimizes, backed up bottom-to-top, both play optimally.
• Alpha-Beta pruning gives the IDENTICAL result to minimax, just more efficiently (prunes irrelevant branches).

DON'T CONFUSE

• Admissible heuristic (never overestimates, guarantees A* optimality) vs Consistent heuristic (stronger condition, also avoids re-expansion).
• Greedy best-first (uses only h(n), not optimal) vs A* (uses g(n)+h(n), optimal if admissible).
• Alpha-Beta pruning changes EFFICIENCY, not the final answer/result.

JRF CHALLENGE ZONE — Chapter 3

1. A* search with an INADMISSIBLE (overestimating) heuristic is: (a) Still guaranteed optimal (b) Not guaranteed optimal (c) Guaranteed to fail completely (d) The same as BFS
Answer: (b)

2. Does Alpha-Beta pruning ever change the FINAL chosen move compared to plain Minimax? (a) Yes, it's an approximation (b) No, it gives the identical result, just faster (c) Only in games with more than 2 players (d) Only for very deep trees
Answer: (b)

3. Node n has g(n)=5. Two heuristic estimates are offered: h1(n)=3 (true remaining cost is 4) and h2(n)=6 (true remaining cost is 4). Which heuristic is admissible?
Answer: h1(n)=3 — it does not overestimate (3≤4); h2 overestimates (6>4) and is inadmissible.

Practice Questions — Chapter 3 (8)

1. Which uninformed search algorithm uses a Queue? Which uses a Stack?
   Ans: BFS uses a Queue; DFS uses a Stack
2. Why is IDDFS considered memory-efficient despite repeating work?
   Ans: Its memory usage stays linear in depth (like DFS), and the repeated shallow-level work is small relative to the dominant bottom level in an exponential tree
3. Write the A* evaluation function f(n).
   Ans: f(n) = g(n) + h(n)
4. What condition must a heuristic satisfy for A* to be guaranteed optimal?
   Ans: It must be admissible — never overestimate the true remaining cost to the goal
5. Differentiate an admissible and a consistent heuristic.
   Ans: Admissible never overestimates the true cost; consistent additionally satisfies h(n) ≤ cost(n,n')+h(n') for every neighbour, and also avoids node re-expansion
6. In Minimax, does the MAX player minimize or maximize the score?
   Ans: Maximize
7. Does Alpha-Beta pruning change the final answer compared to plain Minimax?
   Ans: No — it gives the exact same result, only more efficiently
8. A MIN node has two leaf children with values 7 and 4. What value does the MIN node return?
   Ans: 4 (the minimum)

---

Chapter 4 — Knowledge Representation & Expert System

4.1 Knowledge Representation Techniques

Technique	Idea
Propositional Logic	Statements are TRUE/FALSE, combined with connectives — cannot express relationships between objects or quantify over them
Predicate (First-Order) Logic	Adds OBJECTS, PREDICATES/relations, and QUANTIFIERS (∀, ∃) — far more expressive than propositional logic
Semantic Networks	Graph-based: nodes=concepts/objects, edges=relationships (e.g., "IS-A", "HAS-A")
Frames	Structured record-like representation — a frame has "slots" (attributes) and "fillers" (values), good for representing stereotypical situations/objects
Production Rules	IF-THEN rules (condition→action), the basis of rule-based Expert Systems

JRF trapPropositional Logic CANNOT express "All humans are mortal" precisely as a general rule (it can only encode specific, separately-stated facts like "Socrates is mortal") — this requires Predicate Logic's UNIVERSAL QUANTIFIER (∀x Human(x)→Mortal(x)). "Why do we need predicate logic over propositional logic" is a classic JRF conceptual question.

4.2 Semantic Networks — Inheritance

IdeaSemantic networks support INHERITANCE via "IS-A" links — a property attached to a general/parent node (e.g., "Bird can fly") is automatically inherited by more specific/child nodes (e.g., "Sparrow IS-A Bird" → Sparrow can fly), UNLESS an exception is explicitly overridden at the child level (e.g., "Penguin IS-A Bird" but "Penguin cannot fly" explicitly stated).

4.3 Expert System Architecture

Component	Role
Knowledge Base	Stores domain-specific facts and rules (the "expertise")
Inference Engine	Applies logical rules to the knowledge base to derive new facts/conclusions
User Interface	Allows the user to interact with the system (input queries, receive explanations)
Explanation Facility	Explains WHY/HOW a conclusion was reached — a distinguishing feature of expert systems vs plain programs

JRF trap — forward vs backward chainingFORWARD chaining starts from KNOWN FACTS and applies rules to derive NEW facts, moving TOWARD a goal ("data-driven") — good when there are many possible conclusions and you want to see what follows from the data. BACKWARD chaining starts from a HYPOTHESIS/GOAL and works BACKWARD to check if it's supported by known facts ("goal-driven") — good when you have a SPECIFIC question to answer (e.g., medical diagnosis: "does this patient have disease X?"). This forward(data→conclusion) vs backward(goal→data) direction distinction is THE most tested expert-system inference concept.

MUST REMEMBER — Chapter 4

• Propositional logic: no quantifiers, only specific facts. Predicate logic: adds objects, predicates, quantifiers (∀,∃) — needed for general rules.
• Semantic networks: IS-A links enable inheritance (overridable by exceptions at child nodes).
• Frames: slot-filler structure for stereotypical situations.
• Expert system components: Knowledge Base, Inference Engine, User Interface, Explanation Facility.
• Forward chaining: data→conclusion (data-driven). Backward chaining: goal→data (goal-driven, verifies a hypothesis).

DON'T CONFUSE

• Propositional logic (no quantifiers) vs Predicate logic (has quantifiers, objects, relations).
• Forward chaining (data-driven, toward a goal) vs Backward chaining (goal-driven, toward supporting data).

JRF CHALLENGE ZONE — Chapter 4

1. Which logic type is needed to express "All humans are mortal" as a single general rule? (a) Propositional logic (b) Predicate logic (c) Neither can express it (d) Semantic networks only
Answer: (b)

2. A medical diagnosis system starts from the hypothesis "patient has disease X" and checks if symptoms support it. This is: (a) Forward chaining (b) Backward chaining (c) Semantic network traversal (d) Frame instantiation
Answer: (b)

Practice Questions — Chapter 4 (6)

1. What can Predicate Logic express that Propositional Logic cannot?
   Ans: General rules using quantifiers over objects (e.g., "for all x, if x is human then x is mortal")
2. What does an "IS-A" link represent in a semantic network, and what capability does it enable?
   Ans: A type/category relationship; it enables inheritance of properties from parent to child nodes
3. What are "slots" and "fillers" in a Frame representation?
   Ans: Slots are attributes/properties; fillers are the values assigned to those attributes
4. List the four main components of an Expert System architecture.
   Ans: Knowledge Base, Inference Engine, User Interface, Explanation Facility
5. Differentiate forward chaining and backward chaining.
   Ans: Forward chaining starts from known facts and derives new conclusions (data-driven); backward chaining starts from a goal/hypothesis and checks if facts support it (goal-driven)
6. Which chaining direction is better suited to answering a specific diagnostic question?
   Ans: Backward chaining

---

Chapter 5 — NLP & Planning in AI

5.1 Stages/Levels of NLP

Stage	Idea
Morphological	Analyzes WORD STRUCTURE (roots, prefixes, suffixes — e.g., "running" = "run"+"ning")
Lexical	Analyzes individual words, resolves word-level meaning (part of speech)
Syntactic (Parsing)	Analyzes GRAMMATICAL structure — checks if a sentence follows valid grammar rules
Semantic	Analyzes literal MEANING of words/sentences (independent of context)
Discourse	Analyzes meaning ACROSS multiple connected sentences (e.g., what a pronoun refers to)
Pragmatic	Analyzes meaning in CONTEXT — what the speaker actually INTENDS, beyond literal words

JRF trap — semantic vs pragmaticSemantic analysis handles LITERAL meaning ("Can you pass the salt?" = a literal question about ABILITY). Pragmatic analysis handles the INTENDED meaning IN CONTEXT (the same sentence is actually a POLITE REQUEST to pass the salt, not a genuine question about capability). This "literal vs contextual/intended meaning" distinction (semantic vs pragmatic) is one of the most frequently tested NLP-stage questions.

5.2 NLP Ambiguity Types

Ambiguity Type	Example
Lexical	"Bank" (river bank vs financial bank) — one word, multiple meanings
Syntactic	"I saw the man with a telescope" — unclear who has the telescope (structural/parsing ambiguity)
Semantic	Multiple valid literal interpretations even after parsing is resolved
Pragmatic/Referential	Unclear what a pronoun refers to across sentences (e.g., "John told Bill he was late" — who was late?)

5.3 Planning in AI — STRIPS

IdeaSTRIPS (STanford Research Institute Problem Solver) represents planning actions using: PRECONDITIONS (what must be true before the action can be applied), an ADD LIST (facts that become true after the action), and a DELETE LIST (facts that become false after the action).

JRF-level exampleAction "Pickup(Block)": Precondition = HandEmpty ∧ OnTable(Block) ∧ Clear(Block). Add list = {Holding(Block)}. Delete list = {HandEmpty, OnTable(Block)}. After applying this action, HandEmpty and OnTable(Block) are REMOVED from the world state, and Holding(Block) is ADDED. This "trace the world-state change through precondition→add→delete" technique is the standard STRIPS planning JRF numerical.

MUST REMEMBER — Chapter 5

• NLP stages (roughly in order): Morphological → Lexical → Syntactic → Semantic → Discourse → Pragmatic.
• Semantic = literal meaning; Pragmatic = intended/contextual meaning (e.g., indirect requests).
• Ambiguity types: Lexical(word), Syntactic(structure), Semantic(literal meaning), Pragmatic/Referential(context/reference).
• STRIPS action = Precondition + Add list + Delete list.

DON'T CONFUSE

• Semantic analysis (literal meaning) vs Pragmatic analysis (contextual/intended meaning).
• Lexical ambiguity (word has multiple meanings) vs Syntactic ambiguity (sentence structure is unclear).

JRF CHALLENGE ZONE — Chapter 5

1. "Can you pass the salt?" being understood as a polite REQUEST rather than a literal question about ability is resolved at which NLP stage? (a) Syntactic (b) Semantic (c) Pragmatic (d) Morphological
Answer: (c)

2. "I saw the man with a telescope" (unclear who holds the telescope) is an example of: (a) Lexical ambiguity (b) Syntactic ambiguity (c) Morphological ambiguity (d) No ambiguity
Answer: (b)

Practice Questions — Chapter 5 (6)

1. List the six stages/levels of NLP.
   Ans: Morphological, Lexical, Syntactic, Semantic, Discourse, Pragmatic
2. Differentiate semantic and pragmatic analysis.
   Ans: Semantic analysis handles literal meaning; pragmatic analysis handles meaning in context, including the speaker's actual intent
3. What type of ambiguity does the word "bank" (river vs financial) illustrate?
   Ans: Lexical ambiguity
4. What does STRIPS stand for?
   Ans: STanford Research Institute Problem Solver
5. What three components define a STRIPS action?
   Ans: Precondition, Add list, Delete list
6. What does the "Delete list" of a STRIPS action represent?
   Ans: The facts that become false after the action is performed

---

Chapter 6 — Genetic Algorithms & Artificial Neural Networks

6.1 Genetic Algorithm (GA) — Overview

IdeaA GA is a search/optimization technique inspired by natural evolution — maintains a POPULATION of candidate solutions ("chromosomes"), evaluates each with a FITNESS function, and evolves the population over GENERATIONS using Selection, Crossover, and Mutation.

Operator	Idea
Selection	Chooses FITTER individuals to become parents (e.g., roulette-wheel selection, weighted by fitness)
Crossover	Combines TWO parent chromosomes to produce offspring (e.g., single-point crossover: swap a segment)
Mutation	Randomly FLIPS/alters a small part of a chromosome — maintains genetic DIVERSITY, avoids premature convergence

JRF-level numerical — single-point crossoverParent1 = 11001011, Parent2 = 11011111, crossover point after position 5. Offspring1 = first 5 bits of Parent1 + last 3 bits of Parent2 = 11001 + 111 = 11001111. Offspring2 = first 5 bits of Parent2 + last 3 bits of Parent1 = 11011 + 011 = 11011011. This "cut both parents at the SAME point, swap the tails" trace is the standard single-point crossover JRF numerical.

NET pointMutation's role is specifically to prevent PREMATURE CONVERGENCE (getting stuck at a local optimum because the population has become too similar/lost diversity) — without mutation, crossover alone can only RECOMBINE existing genetic material, never introduce a value that wasn't already present somewhere in the population.

6.2 Artificial Neural Network (ANN) — Basics

SimpleAn ANN is a computing system loosely inspired by biological neurons — layers of interconnected "neurons" (nodes) process input via WEIGHTED connections, aiming to learn patterns from data.

Neuron output = ActivationFunction( Σ(input_i × weight_i) + bias )

JRF-level numerical — perceptron computationInputs x1=1, x2=0, x3=1; weights w1=0.5, w2=−0.3, w3=0.8; bias=−0.2. Using a step activation function (output=1 if net≥0, else 0): net = (1×0.5)+(0×−0.3)+(1×0.8)+(−0.2) = 0.5+0+0.8−0.2 = 1.1. Since 1.1≥0 → output = 1. This "compute the weighted sum plus bias, then apply the activation function" is the fundamental perceptron JRF numerical.

6.3 Perceptron & Limitations

JRF trap — the XOR problemA SINGLE-LAYER Perceptron can only learn LINEARLY SEPARABLE functions (e.g., AND, OR) — it CANNOT learn XOR, because XOR's outputs cannot be separated by a single straight line/hyperplane. This limitation was historically significant (contributed to the first "AI winter") and is solved by using a MULTI-LAYER perceptron (MLP) with at least one hidden layer, which CAN represent non-linearly-separable functions like XOR. "Why can't a single perceptron learn XOR" is one of the most classic, frequently tested ANN facts.

6.4 Common Activation Functions

Function	Range	Idea
Step (Threshold)	{0,1}	Binary output; not differentiable (unsuitable for gradient-based learning)
Sigmoid	(0,1)	Smooth S-curve; differentiable, but suffers from "vanishing gradient" for very large/small inputs
Tanh	(−1,1)	Zero-centred version of sigmoid
ReLU	[0,∞)	max(0,x); simple, fast, avoids vanishing gradient for positive inputs — widely used in deep networks

6.5 Backpropagation (Concept)

IdeaBackpropagation trains a multi-layer network by: (1) forward pass — compute the network's output, (2) compute the ERROR (difference from the desired output), (3) propagate that error BACKWARD through the layers, adjusting WEIGHTS via gradient descent to reduce the error, repeating over many iterations/epochs.

MUST REMEMBER — Chapter 6

• GA operators: Selection(pick fitter parents), Crossover(combine parents), Mutation(random change, maintains diversity, prevents premature convergence).
• Single-point crossover: cut both parents at the same point, swap tails.
• Perceptron output = ActivationFunction(Σ(input×weight)+bias).
• Single-layer perceptron CANNOT learn XOR (not linearly separable); Multi-layer perceptron CAN (with a hidden layer).
• ReLU=max(0,x), avoids vanishing gradient, widely used in deep learning. Sigmoid/Tanh can suffer vanishing gradient.
• Backpropagation: forward pass → compute error → propagate error backward → adjust weights (gradient descent).

DON'T CONFUSE

• Crossover (combines existing genetic material) vs Mutation (introduces new/random variation, prevents stagnation).
• Single-layer perceptron (linearly separable only, can't do XOR) vs Multi-layer perceptron (can learn XOR via hidden layers).

JRF CHALLENGE ZONE — Chapter 6

1. Which GA operator is primarily responsible for maintaining population diversity and avoiding premature convergence? (a) Selection (b) Crossover (c) Mutation (d) Fitness evaluation
Answer: (c)

2. Can a single-layer Perceptron learn the XOR function? (a) Yes, always (b) No, XOR is not linearly separable (c) Only with a step activation function (d) Only with more training data
Answer: (b)

3. Inputs x1=1,x2=1; weights w1=0.4,w2=0.4; bias=−0.5. Using step activation (≥0→1), find the output.
Answer: net=0.4+0.4−0.5=0.3≥0 → output=1

Practice Questions — Chapter 6 (8)

1. Name the three main operators of a Genetic Algorithm.
   Ans: Selection, Crossover, Mutation
2. Parent1=1010, Parent2=1101, crossover point after position 2. Find Offspring1.
   Ans: First 2 bits of Parent1 + last 2 bits of Parent2 = 10+01 = 1001
3. Why is Mutation necessary in a Genetic Algorithm, given Crossover already recombines genes?
   Ans: Crossover can only recombine EXISTING genetic material; mutation introduces new variation, preventing premature convergence/stagnation
4. Write the general formula for a neuron's output in an ANN.
   Ans: ActivationFunction(Σ(input×weight) + bias)
5. Why can't a single-layer Perceptron learn the XOR function?
   Ans: XOR's outputs are not linearly separable, and a single perceptron can only represent a linear decision boundary
6. What architecture change allows a neural network to learn XOR?
   Ans: Using a multi-layer perceptron with at least one hidden layer
7. Name an activation function that helps avoid the vanishing gradient problem for positive inputs.
   Ans: ReLU
8. Describe the three main steps of backpropagation.
   Ans: Forward pass to compute output, compute the error, propagate the error backward while adjusting weights via gradient descent

---

Chapter 7 — Fuzzy Set

7.1 Fuzzy Set — Basics

SimpleIn a CRISP (classical) set, an element either FULLY belongs (membership=1) or FULLY does not belong (membership=0). In a FUZZY set, an element can have a PARTIAL degree of membership — any value between 0 and 1 (inclusive) — representing "how much" it belongs.

Worked exampleFuzzy set "Tall" for heights, with membership function μ_Tall(h). μ_Tall(150cm)=0.1, μ_Tall(170cm)=0.6, μ_Tall(190cm)=1.0. A person of 170cm is "somewhat tall" (60% degree of membership) — NOT strictly tall or not-tall as a crisp set would force.

7.2 Fuzzy Set Operations

Operation	Formula
Union (A∪B)	μ_A∪B(x) = max(μ_A(x), μ_B(x))
Intersection (A∩B)	μ_A∩B(x) = min(μ_A(x), μ_B(x))
Complement (A')	μ_A'(x) = 1 − μ_A(x)

JRF-level numericalμ_A(x)=0.7, μ_B(x)=0.4 for some element x. Union: max(0.7,0.4) = 0.7. Intersection: min(0.7,0.4) = 0.4. Complement of A: 1−0.7 = 0.3. This direct max/min/complement calculation is THE standard fuzzy-set-operations JRF numerical — remember union uses MAX (OR-like, take the more generous membership) and intersection uses MIN (AND-like, take the more restrictive membership).

JRF trap — Law of Excluded Middle does NOT holdIn CRISP sets, A∪A' always equals the universal set, and A∩A' always equals ∅ (Law of Excluded Middle / Law of Contradiction). In FUZZY sets, these laws generally do NOT hold: e.g., if μ_A(x)=0.6, then μ_A'(x)=0.4, so μ_A∪A'(x)=max(0.6,0.4)=0.6 ≠ 1, and μ_A∩A'(x)=min(0.6,0.4)=0.4 ≠ 0. This "classical laws break down in fuzzy logic" fact is a very frequently tested JRF distinction.

7.3 Fuzzy Logic vs Crisp (Boolean) Logic

	Crisp (Boolean) Logic	Fuzzy Logic
Truth values	Exactly {0,1} (True/False)	Any real number in [0,1] (degree of truth)
Handles	Precise, unambiguous statements	Vague/imprecise concepts (e.g., "hot", "tall", "fast")
Used in	Traditional digital logic/programming	Control systems (e.g., washing machines, AC thermostats), approximate reasoning

7.4 Membership Functions

IdeaCommon shapes used to define fuzzy membership: Triangular, Trapezoidal, Gaussian — each maps an input value to a membership degree in [0,1], chosen based on how the "vagueness" should be modeled for that particular variable.

MUST REMEMBER — Chapter 7

• Crisp set: membership ∈ {0,1} only. Fuzzy set: membership ∈ [0,1], any real value (partial membership).
• Fuzzy Union = MAX; Fuzzy Intersection = MIN; Fuzzy Complement = 1−μ.
• Law of Excluded Middle (A∪A'=Universe) and Law of Contradiction (A∩A'=∅) do NOT generally hold for fuzzy sets.
• Fuzzy logic handles vague/imprecise concepts; crisp/Boolean logic handles only precise true/false statements.
• Common membership function shapes: Triangular, Trapezoidal, Gaussian.

DON'T CONFUSE

• Fuzzy Union (MAX) vs Fuzzy Intersection (MIN) — easy to swap.
• Crisp set laws (Excluded Middle/Contradiction always hold) vs Fuzzy set laws (generally do NOT hold).

JRF CHALLENGE ZONE — Chapter 7

1. μ_A(x)=0.8, μ_B(x)=0.3. Find μ_A∩B(x).
Answer: min(0.8,0.3) = 0.3

2. In fuzzy sets, does A∪A' always equal the universal set (membership 1) for every element? (a) Yes, always (b) No, generally not (c) Only for crisp sets (d) Only when μ_A(x)=0.5
Answer: (b)

3. μ_A(x)=0.45. Find μ_A'(x) (complement).
Answer: 1−0.45 = 0.55

Practice Questions — Chapter 7 (6)

1. What is the key difference between a crisp set and a fuzzy set?
   Ans: Crisp set membership is only 0 or 1; fuzzy set membership can be any value in [0,1] (partial membership)
2. Write the formulas for fuzzy union and fuzzy intersection.
   Ans: Union: max(μ_A(x), μ_B(x)); Intersection: min(μ_A(x), μ_B(x))
3. μ_A(x)=0.55, μ_B(x)=0.9. Find the fuzzy union and intersection at x.
   Ans: Union: max(0.55,0.9)=0.9; Intersection: min(0.55,0.9)=0.55
4. Does the Law of Excluded Middle (A∪A' = universal set) hold in fuzzy set theory?
   Ans: No, generally not
5. Give two examples of common membership function shapes.
   Ans: Any two of: Triangular, Trapezoidal, Gaussian
6. Name a practical application area that commonly uses fuzzy logic.
   Ans: Control systems, e.g., washing machines or AC thermostats

---

One-Shot Revision — Unit 9

Key facts across all chapters

• 4 AI approaches: think humanly, think rationally, act humanly(Turing Test), act rationally(dominant modern approach). ML ⊂ AI.
• PEAS=Performance,Environment,Actuators,Sensors. Rational≠omniscient — best decision given available info.
• Agent hierarchy: Simple Reflex→Model-Based(+state)→Goal-Based(+future)→Utility-Based(+graded preference)→Learning(+self-improve).
• 6 environment dims: Observable(full/partial), Deterministic/Stochastic, Episodic/Sequential, Static/Dynamic, Discrete/Continuous, Single/Multi-agent.
• BFS=complete+optimal(uniform cost),high memory. DFS=low memory,not complete. UCS=lowest path-cost. IDDFS=DFS memory+BFS completeness.
• A*: f(n)=g(n)+h(n); optimal ONLY if h admissible(never overestimates). Consistent⟹admissible. Alpha-Beta=same result as Minimax, just faster.
• Propositional logic: no quantifiers. Predicate logic: adds ∀,∃, objects, relations. Semantic networks: IS-A→inheritance.
• Expert system: Knowledge Base+Inference Engine+UI+Explanation Facility. Forward chaining=data→conclusion; Backward=goal→data.
• NLP stages: Morphological→Lexical→Syntactic→Semantic(literal)→Discourse→Pragmatic(contextual/intended).
• STRIPS action = Precondition+Add list+Delete list.
• GA: Selection+Crossover+Mutation(prevents premature convergence, adds new variation). Single-point crossover: cut both parents same point, swap tails.
• Perceptron output=Activation(Σ(input×weight)+bias). Single-layer CANNOT learn XOR(not linearly separable); Multi-layer CAN.
• ReLU=max(0,x), avoids vanishing gradient. Backprop: forward pass→error→backward propagate→adjust weights.
• Fuzzy: membership∈[0,1]. Union=MAX; Intersection=MIN; Complement=1−μ. Excluded Middle/Contradiction laws do NOT hold for fuzzy sets.

Potential future exam areasPotential high-value exam area based on syllabus importance and historical question patterns: environment-classification scenario questions (classify a given scenario across all 6 dimensions); A* admissibility numericals; minimax/alpha-beta tree-evaluation numericals; forward-vs-backward chaining scenario questions; GA crossover-trace numericals; perceptron weighted-sum computations (especially the XOR limitation); and fuzzy set union/intersection/complement numericals.

---

Unit 9 — UGC NET/JRF Mini Mock Test

40 questions across all 7 chapters. NTA/UGC NET-style question patterns — mixed NET/JRF difficulty, numerical, statement-based, matching and scenario-based. Answer key with brief explanations follows each question.

Q1. The Turing Test corresponds to which AI approach? (a) Thinking humanly (b) Thinking rationally (c) Acting humanly (d) Acting rationally
Ans: (c) [Ch1 | NET]

Q2. Is Machine Learning the same as Artificial Intelligence? (a) Yes, identical (b) No, ML is a subset of AI (c) No, AI is a subset of ML (d) They are unrelated
Ans: (b) [Ch1 | NET]

Q3. What does PEAS stand for?
Ans: Performance measure, Environment, Actuators, Sensors [Ch2 | NET]

Q4. Which agent type can compare and prefer between multiple successful outcomes? (a) Simple reflex (b) Goal-based (c) Utility-based (d) Model-based only
Ans: (c) [Ch2 | NET]

Q5. Is Poker fully or partially observable (regarding opponents' cards)?
Ans: Partially observable [Ch2 | NET]

Q6. Which uninformed search combines DFS's memory efficiency with BFS's completeness? (a) UCS (b) IDDFS (c) Greedy search (d) A*
Ans: (b) [Ch3 | NET]

Q7. A* search is optimal only if the heuristic is: (a) Consistent only (b) Admissible (c) Always zero (d) Overestimating
Ans: (b) [Ch3 | NET]

Q8. Node n: g(n)=4, h(n)=6. Find f(n).
Ans: f(n)=g(n)+h(n)=10 [Ch3 | NET numerical]

Q9. Does Alpha-Beta pruning change the final result compared to Minimax? (a) Yes (b) No, only improves efficiency (c) Sometimes (d) Only in 3+ player games
Ans: (b) [Ch3 | NET]

Q10. A MAX node has two MIN children returning 5 and 8. What value does the MAX node return?
Ans: 8 (the maximum) [Ch3 | NET numerical]

Q11. Which logic type includes quantifiers like ∀ and ∃? (a) Propositional logic (b) Predicate logic (c) Fuzzy logic (d) None
Ans: (b) [Ch4 | NET]

Q12. Which chaining approach is goal-driven, starting from a hypothesis? (a) Forward chaining (b) Backward chaining (c) Semantic chaining (d) Frame chaining
Ans: (b) [Ch4 | NET]

Q13. What structure in an Expert System explains WHY a conclusion was reached?
Ans: The Explanation Facility [Ch4 | NET]

Q14. Which NLP stage handles the LITERAL meaning of a sentence?
Ans: Semantic analysis [Ch5 | NET]

Q15. "Bank" (river vs financial) is an example of which ambiguity type? (a) Syntactic (b) Lexical (c) Pragmatic (d) Morphological
Ans: (b) [Ch5 | NET]

Q16. In STRIPS, what does the "Add list" represent?
Ans: Facts that become true after the action [Ch5 | NET]

Q17. Which GA operator maintains population diversity, preventing premature convergence? (a) Selection (b) Crossover (c) Mutation (d) Fitness scaling
Ans: (c) [Ch6 | NET]

Q18. Parent1=1100, Parent2=1011, crossover after position 2. Find Offspring1 (first 2 of P1 + last 2 of P2).
Ans: 11+11=1111 [Ch6 | JRF numerical]

Q19. Can a single-layer Perceptron learn the XOR function? (a) Yes (b) No, not linearly separable (c) Only with more epochs (d) Only with a step function
Ans: (b) [Ch6 | NET]

Q20. Inputs x1=1,x2=0; weights w1=0.6,w2=0.9; bias=−0.3. Using step activation, find net and output.
Ans: net=0.6+0−0.3=0.3≥0 → output=1 [Ch6 | JRF numerical]

Q21. Which activation function is defined as max(0,x)?
Ans: ReLU [Ch6 | NET]

Q22. In a crisp set, what are the only possible membership values?
Ans: 0 or 1 [Ch7 | NET]

Q23. μ_A(x)=0.6, μ_B(x)=0.2. Find the fuzzy union.
Ans: max(0.6,0.2)=0.6 [Ch7 | NET numerical]

Q24. μ_A(x)=0.35. Find the fuzzy complement.
Ans: 1−0.35=0.65 [Ch7 | NET numerical]

Q25. Does the Law of Excluded Middle (A∪A'=universal set) hold for fuzzy sets in general? (a) Yes, always (b) No, generally not (c) Only for crisp sets (d) Only when membership=1
Ans: (b) [Ch7 | JRF]

Q26. Which environment property describes Chess (fixed rules, no randomness)? (a) Stochastic (b) Deterministic (c) Partially observable (d) Continuous
Ans: (b) [Ch2 | NET]

Q27. Which search algorithm expands the node with the lowest PATH COST so far (not heuristic)?
Ans: Uniform Cost Search (UCS) [Ch3 | NET]

Q28. A heuristic h1(n)=5 when the true remaining cost is 3. Is h1 admissible?
Ans: No — it overestimates (5>3). [Ch3 | JRF]

Q29. Which knowledge representation uses "slots" and "fillers"?
Ans: Frames [Ch4 | NET]

Q30. Which NLP stage analyzes word structure (roots, prefixes, suffixes)?
Ans: Morphological analysis [Ch5 | NET]

Q31. What does a fitness function do in a Genetic Algorithm?
Ans: Evaluates how good/fit each candidate solution (chromosome) is [Ch6 | NET]

Q32. Which activation function outputs values in the range (0,1)?
Ans: Sigmoid [Ch6 | NET]

Q33. μ_A(x)=0.7, μ_B(x)=0.7. Find the fuzzy intersection.
Ans: min(0.7,0.7)=0.7 [Ch7 | NET numerical]

Q34. Which agent type has NO memory of the past, acting only on the current percept? (a) Model-based (b) Simple reflex (c) Goal-based (d) Learning agent
Ans: (b) [Ch2 | NET]

Q35. Which search strategy is NOT guaranteed to be complete (may loop forever on infinite paths)? (a) BFS (b) UCS (c) DFS (d) IDDFS
Ans: (c) [Ch3 | NET]

Q36. What is the key difference between forward and backward chaining?
Ans: Forward chaining moves from known facts toward conclusions (data-driven); backward chaining moves from a goal toward supporting facts (goal-driven) [Ch4 | NET]

Q37. Which NLP ambiguity type concerns unclear pronoun reference across sentences?
Ans: Pragmatic/Referential ambiguity [Ch5 | NET]

Q38. What architecture allows a neural network to represent the XOR function?
Ans: A multi-layer perceptron with at least one hidden layer [Ch6 | NET]

Q39. Which is TRUE about a consistent heuristic? (a) It is never admissible (b) Every consistent heuristic is admissible (c) Consistency is unrelated to admissibility (d) Consistent heuristics are always inadmissible
Ans: (b) [Ch3 | JRF]

Q40. Which membership function shapes are commonly used in fuzzy logic? Name any two.
Ans: Any two of: Triangular, Trapezoidal, Gaussian [Ch7 | NET]

— End of Mock Test — Cross-check your score, revisit the "Don't Confuse" and "JRF Challenge Zone" boxes for any topic you missed, then re-attempt after 48 hours. —

Share: