Artificial Intelligence - Quick Revision

Module 1: Introduction to AI

• AI studies systems that perceive, reason, learn, and act.
• Rational means maximizing expected performance with available information.
• Specify a task with PEAS and classify its environment before choosing an agent.
• Agent progression: reflex -> model -> goal -> utility -> learning.

Module 2: Problem Spaces and Search

• Problem = initial state, actions, transition model, goal test, path cost.
• BFS uses depth; UCS uses g; greedy uses h; A* uses g + h.
• CSP = variables, domains, constraints.
• Minimax handles adversarial choices; alpha-beta avoids irrelevant work.

Module 3: Knowledge Representation

• Logic gives formal semantics and proof.
• Semantic networks and frames support structured relationships and inheritance.
• Ontologies define a domain schema; knowledge graphs store linked facts.
• Resolution proves a query by deriving a contradiction from its negation.

Module 4: Uncertain Knowledge and Reasoning

• Conditional probability changes after evidence is known.
• Bayes' rule combines prior and likelihood to find a posterior.
• Naive Bayes simplifies classification with a conditional-independence assumption.
• HMMs model hidden states changing over time.

Module 5: Artificial Neural Networks

• A neuron computes weighted sum + bias + activation.
• A perceptron learns a linear boundary; hidden layers model nonlinear patterns.
• Gradient descent reduces loss; backpropagation computes the gradients.
• Use validation and regularization to detect and reduce overfitting.

Ten memory lines

1. Rational is not omniscient.
2. BFS is not cheapest-path search when costs differ.
3. A* balances cost paid and estimated cost remaining.
4. A heuristic is knowledge that guides search.
5. Representation determines which inferences are easy.
6. Conditional probability is directional.
7. Priors matter when evidence is interpreted.
8. HMM states are hidden; emissions are observed.
9. Bias shifts a neural decision boundary.
10. Backpropagation computes gradients; an optimizer applies them.

PDF-derived high-yield additions

Module 1

• Narrow/weak AI solves a limited task; general/strong AI is a hypothetical broad capability.
• AI combines representation, reasoning/computation, and learning.
• Environment assumptions determine whether a task is observable, deterministic, static, discrete, single-agent, and known.
• A responsible AI answer may mention privacy, bias, explainability, safety, and human oversight.

Module 2

• CSP notation is (V, D, C); constraints can be unary, binary, higher-order, or global such as AllDifferent.
• Forward checking removes inconsistent neighbour values after an assignment.
• AC-3 repeatedly removes values with no supporting neighbour value.
• Minimax assumes a finite, deterministic, two-player, zero-sum, perfect-information game. Alpha-beta prunes when alpha >= beta without changing the answer.
• Beam search keeps only a fixed number of frontier nodes and may lose the best solution.

Module 3

• Explicit knowledge is stored; implicit knowledge is inferred.
• Domain knowledge is specialized; common-sense knowledge is general everyday knowledge.
• Good representations handle ambiguity, change, defaults, and exceptions.
• Resolution: implication elimination -> CNF -> negate query -> resolve -> empty clause means contradiction and entailment.

Module 4

• Time may be a point, interval, duration, sequence, or relation such as before, after, during, or overlaps.
• Probability expresses uncertainty; fuzzy logic expresses degree of membership.
• Bayes: posterior = likelihood x prior / evidence.
• HMM: filtering = forward, smoothing = forward-backward, decoding = Viterbi.
• Expected utility is more appropriate than most-likely outcome when consequences have different costs.

Module 5

• Learning types: supervised uses labels, unsupervised finds structure, and reinforcement learning uses rewards or penalties.
• An algorithm trains; a model stores the learned parameters.
• A perceptron converges only for linearly separable data.
• For backpropagation, show forward pass, loss, chain-rule gradients, and the learning-rate update.