UGC NET Computer Science Artificial Intelligence (AI) Syllabus

Distinguish forward chaining from backward chaining inference.

Forward chaining is data-driven: it starts from known facts and applies rules to derive new facts until the goal is reached. Backward chaining is goal-driven: it starts from a goal/hypothesis and works backward to find supporting facts.

How does Bayes' theorem handle uncertainty in knowledge, and what is its formula?

It updates the probability of a hypothesis given evidence: P(H|E) = [P(E|H)·P(H)] / P(E). It combines prior probability with likelihood to compute the posterior, forming the basis of probabilistic reasoning and Bayesian networks.

What are certainty factors (CF) and where were they used?

A non-probabilistic measure of belief in a hypothesis given evidence, ranging from -1 (certainly false) to +1 (certainly true), used in MYCIN to handle uncertainty without requiring full probability distributions.

What are the key components of a planning system?

A representation of states, goals, and actions (operators with preconditions and effects); a way to choose/order actions; and a method to detect and resolve interactions among subgoals. The output is a plan (sequence or partial order of actions) achieving the goal.

Differentiate linear (total-order) planning from non-linear (partial-order) planning.

Linear planning solves subgoals one after another in a strict sequence and may suffer from goal interactions (Sussman anomaly). Non-linear planning allows interleaving of steps and represents plans as a partial order, committing to ordering only when necessary.

What is Goal Stack Planning?

A planning method (used in STRIPS) that uses a stack to hold goals and operators. It pushes the goal, then unsatisfied preconditions and operators, popping and applying operators as preconditions become satisfied until the stack is empty and all goals hold.

What is Hierarchical (HTN) planning?

Planning that decomposes high-level abstract tasks into lower-level subtasks using methods, refining abstract actions into primitive executable actions. It reduces search by reasoning at multiple levels of abstraction (e.g., ABSTRIPS).

What is STRIPS and how is an action represented in it?

STRIPS (Stanford Research Institute Problem Solver) is a classic planner. Each operator is defined by preconditions (must hold before), an add list (literals made true), and a delete list (literals made false). States are sets of literals.

What is Partial Order Planning (POP)?

A least-commitment planning approach that produces a partially ordered set of actions with ordering constraints and causal links, leaving unordered steps unordered unless a constraint requires sequencing. It resolves threats by promotion or demotion.

In partial-order planning, what is a causal link and a threat?

A causal link A →(p) B means action A achieves precondition p for action B. A threat is an action C that could delete p; it is resolved by promotion (place C after B) or demotion (place C before A).

In NLP, what is a (context-free) grammar and what defines a language?

A grammar is a set of production rules (G = {N, T, P, S}: non-terminals, terminals, productions, start symbol) that generate strings. The language is the set of all strings derivable from the start symbol. CFGs (Type-2) are common in NLP.

Compare top-down and bottom-up parsing.

Top-down parsing starts from the start symbol and rewrites it toward the input (goal-driven, e.g., recursive descent). Bottom-up parsing starts from the input tokens and combines them up to the start symbol (data-driven, e.g., shift-reduce, CYK).

What is the difference between semantic analysis and pragmatics in NLP?

Semantic analysis derives the literal, context-independent meaning of words and sentences (word sense, semantic roles). Pragmatics interprets meaning in context, accounting for speaker intent, references, and real-world knowledge (e.g., resolving 'it', indirect requests).

What are the typical stages of natural language processing?

Lexical/morphological analysis, syntactic analysis (parsing), semantic analysis, discourse integration, and pragmatic analysis. They move from word-level structure to context-dependent meaning.

How do agents differ from objects (in object-oriented terms)?

Objects encapsulate state and expose methods but execute when invoked (no control over their behaviour). Agents are autonomous: they decide whether to act on a request, are reactive/proactive/social, and have their own thread of control and goals.

How does an agent differ from an expert system?

Expert systems are typically passive, disembodied reasoners answering queries in a single domain via user interaction. Agents are situated in an environment, perceive and act autonomously, are proactive and social, and often operate continuously without direct prompting.

What is the generic structure of a multiagent system (MAS)?

A collection of autonomous agents, each with limited local knowledge/viewpoint, situated in a shared environment, interacting via communication and coordination mechanisms. There is no global control; agents cooperate, negotiate, or compete to achieve individual/collective goals.

What is the Semantic Web?

An extension of the Web (Tim Berners-Lee) where information is given well-defined, machine-understandable meaning using standards like RDF, RDFS, and OWL, enabling agents to process and integrate data automatically through ontologies and linked data.

What is an Agent Communication Language (ACL), and name two examples.

A standardized language enabling agents to exchange knowledge and coordinate, based on speech-act theory (performatives like inform, request, query). Examples: KQML (Knowledge Query and Manipulation Language) and FIPA-ACL.

How are ontologies used for knowledge sharing among agents?

A shared ontology provides a common vocabulary and agreed meanings for terms, so that messages exchanged between heterogeneous agents are interpreted consistently. This ensures semantic interoperability and reuse of knowledge across agents and systems.

Name some agent development tools/platforms.

JADE (Java Agent Development Framework, FIPA-compliant), JACK, ZEUS, AgentBuilder, MadKit, and FIPA-OS. JADE is the most widely used, providing middleware and a runtime for building distributed multi-agent systems.

What is the notion of fuzziness, and how do fuzzy sets differ from classical sets?

Fuzziness models vagueness where membership is a matter of degree. In a classical (crisp) set membership is binary (0 or 1); in a fuzzy set (Zadeh) an element has a membership degree μ(x) in the continuous interval [0,1].

What is a membership function in fuzzy logic, and name common shapes.

A function μ_A(x): X → [0,1] that assigns each element a degree of membership in fuzzy set A. Common shapes: triangular, trapezoidal, Gaussian, sigmoidal, and singleton.

Define fuzzification and defuzzification.

Fuzzification converts crisp input values into fuzzy membership degrees over linguistic terms. Defuzzification converts the aggregated fuzzy output back into a single crisp value (e.g., via centroid/center-of-gravity, mean-of-maxima, or weighted-average methods).

State the standard fuzzy set operations for union, intersection, and complement.

For fuzzy sets A and B: Union μ_{A∪B}(x) = max(μ_A(x), μ_B(x)); Intersection μ_{A∩B}(x) = min(μ_A(x), μ_B(x)); Complement μ_{Ā}(x) = 1 − μ_A(x).

What is a linguistic variable in fuzzy systems?

A variable whose values are words/terms (linguistic terms) rather than numbers, e.g., 'temperature' with values {cold, warm, hot}, each defined by a fuzzy set. It is characterized by a name, set of linguistic terms, universe of discourse, and membership functions.

What is a fuzzy relation, and how are fuzzy relations composed?

A fuzzy relation R(X,Y) assigns a membership degree to each pair (x,y), generalizing crisp relations. Fuzzy relations are combined using max-min composition: μ_{R∘S}(x,z) = max_y [min(μ_R(x,y), μ_S(y,z))].

What is a fuzzy rule, and what are the steps of fuzzy inference?

A fuzzy rule has the form IF x is A THEN y is B (antecedent and consequent are fuzzy sets). Inference (e.g., Mamdani): fuzzify inputs, apply rules (evaluate antecedents, implication), aggregate rule outputs, then defuzzify to a crisp output.

What are the main components of a fuzzy (rule-based) control system?

A fuzzifier (crisp→fuzzy), a knowledge base/rule base (fuzzy IF-THEN rules and membership functions), an inference engine (applies rules), and a defuzzifier (fuzzy→crisp). Mamdani and Sugeno are the two common fuzzy inference models.

What are common encoding (representation) strategies in Genetic Algorithms?

Binary encoding (bit strings), value/real-valued encoding, permutation encoding (for ordering problems like TSP), and tree encoding (for genetic programming). The encoding defines how candidate solutions are represented as chromosomes.

What are the three main genetic operators in a GA?

Selection (choosing fitter individuals to reproduce, e.g., roulette-wheel, tournament, rank), Crossover/recombination (combining parent chromosomes, e.g., one-point, two-point, uniform), and Mutation (random small changes to maintain diversity).

What is a fitness function, and what are the steps of the GA cycle?

A fitness function evaluates how good a candidate solution is (the objective to optimize). GA cycle: initialize population → evaluate fitness → selection → crossover → mutation → form new generation → repeat until a termination criterion is met.

How is a problem solved using a Genetic Algorithm?

Encode candidate solutions as chromosomes, define a fitness function, initialize a random population, then iteratively select parents and apply crossover and mutation to evolve better solutions across generations, stopping when convergence or a fitness threshold is reached.

What is supervised learning? Give example tasks and algorithms.

Learning a mapping from inputs to outputs using labelled training data (input-output pairs). Tasks: classification and regression. Algorithms: decision trees, SVM, k-NN, naive Bayes, linear/logistic regression, neural networks.

What is unsupervised learning? Give example tasks.

Learning patterns/structure from unlabelled data without target outputs. Tasks: clustering (e.g., k-means, hierarchical), association rule mining, and dimensionality reduction (e.g., PCA). The system discovers hidden groupings or relationships.

What is reinforcement learning, and what are its key elements?

Learning by interacting with an environment to maximize cumulative reward through trial and error. Key elements: agent, environment, states, actions, reward signal, policy, and value function. Examples include Q-learning and SARSA.

What is a single perceptron, and what is its key limitation?

A single-layer neural unit computing output = activation(Σ wᵢxᵢ + b), used as a linear binary classifier trained by the perceptron learning rule. Limitation: it can only separate linearly separable data, so it cannot solve the XOR problem.

What is a Multi-Layer Perceptron (MLP) and how is it trained?

A feedforward neural network with one or more hidden layers of nonlinear units between input and output, able to learn non-linearly separable functions (e.g., XOR). It is trained by the backpropagation algorithm using gradient descent to minimize error.

What is a Self-Organizing Map (SOM)?

An unsupervised competitive-learning neural network (Kohonen) that maps high-dimensional input onto a low-dimensional (usually 2D) grid while preserving topology. The Best Matching Unit and its neighbours have weights adjusted toward each input, forming clusters.

What is a Hopfield network and what is it used for?

A recurrent, fully connected single-layer network with symmetric weights and no self-connections, acting as associative (content-addressable) memory. It stores patterns as stable states and converges to the nearest stored pattern by minimizing an energy function.