Reference topics: 6.034 ยท independent study notes, not affiliated

Artificial Intelligence

Introduces the core methods of AI: search (A*, minimax, alpha-beta), knowledge representation, reasoning under uncertainty (Bayesian networks, HMMs), machine learning, and natural language processing.

8Lectures
2Assignments
2Exams
1hDuration
1

Introduction to AI and Search

Overview Symbolic reasoning and search. Key ideas Uninformed: BFS, DFS. Informed: greedy, A*. Admissibility guarantees A* optimality. Worked example Map with straight-line heuristic: A* expands fewer...

โฑ 1 min read

Overview

Symbolic reasoning and search.

Key ideas

  • Uninformed: BFS, DFS.
  • Informed: greedy, A*.
  • Admissibility guarantees A* optimality.

Worked example

Map with straight-line heuristic: A* expands fewer nodes than BFS.

Common pitfalls

  • Using inadmissible heuristic with A*.

Why it matters

Search is the foundation of planning and optimisation.

2

Games: Minimax and Alpha-Beta

Overview Two-player deterministic games. Key ideas Model as game tree. Minimax assumes optimal play. Alpha-beta prunes branches. Worked example Tic-tac-toe minimax: 255,168 nodes; alpha-beta halves....

โฑ 1 min read

Overview

Two-player deterministic games.

Key ideas

  • Model as game tree.
  • Minimax assumes optimal play.
  • Alpha-beta prunes branches.

Worked example

Tic-tac-toe minimax: 255,168 nodes; alpha-beta halves.

Common pitfalls

  • Forgetting terminal states.

Why it matters

Real engines add evaluation and deepening.

3

Constraint Satisfaction

Overview Assign values to variables subject to constraints. Key ideas Backtracking with propagation. Forward checking. Arc consistency. Worked example Sudoku: 27 constraints. Propagation solves many...

โฑ 1 min read

Overview

Assign values to variables subject to constraints.

Key ideas

  • Backtracking with propagation.
  • Forward checking.
  • Arc consistency.

Worked example

Sudoku: 27 constraints. Propagation solves many puzzles.

Common pitfalls

  • Ignoring propagation power.

Why it matters

Propagation removes whole subtrees.

4

Knowledge Representation and Inference

Overview Telling a machine what we know. Key ideas Propositional and first-order logic. Forward and backward chaining. Resolution. Worked example "All humans are mortal" + "Socrates is...

โฑ 1 min read

Overview

Telling a machine what we know.

Key ideas

  • Propositional and first-order logic.
  • Forward and backward chaining.
  • Resolution.

Worked example

"All humans are mortal" + "Socrates is human" -> mortal.

Common pitfalls

  • Ignoring computational cost.

Why it matters

Real systems combine rules with search.

5

Bayesian Networks and HMMs

Overview Graphical models for probability. Key ideas DAG encodes conditional independence. Inference is NP-hard generally. HMM: latent state chain. Worked example Medical diagnosis: symptoms ->...

โฑ 1 min read

Overview

Graphical models for probability.

Key ideas

  • DAG encodes conditional independence.
  • Inference is NP-hard generally.
  • HMM: latent state chain.

Worked example

Medical diagnosis: symptoms -> diseases.

Common pitfalls

  • Ignoring arc direction.

Why it matters

HMMs powered speech recognition for decades.

6

Machine Learning: Decision Trees and Nearest Neighbours

Overview Two foundational supervised learners. Key ideas Trees: split on best feature. k-NN: majority vote of neighbours. Trees: low bias high variance. Worked example Classify iris from petal...

โฑ 1 min read

Overview

Two foundational supervised learners.

Key ideas

  • Trees: split on best feature.
  • k-NN: majority vote of neighbours.
  • Trees: low bias high variance.

Worked example

Classify iris from petal length.

Common pitfalls

  • Growing trees until pure.

Why it matters

Understanding bias/variance is key.

7

Neural Networks and Deep Learning

Overview From perceptron to multi-layer networks. Key ideas Perceptron: weighted sum + activation. Stacking gives universal approximator. Backprop via chain rule. Worked example 2-layer network...

โฑ 1 min read

Overview

From perceptron to multi-layer networks.

Key ideas

  • Perceptron: weighted sum + activation.
  • Stacking gives universal approximator.
  • Backprop via chain rule.

Worked example

2-layer network approximates XOR.

Common pitfalls

  • Linear activation in hidden layers.

Why it matters

ReLU, dropout, batch norm, Adam made it practical.

8

Natural Language Processing

Overview Language is sequential and ambiguous. Key ideas Tokenisation, stemming, POS. Constituency and dependency parsing. Embeddings and transformers. Worked example Sentiment: "not bad"...

โฑ 1 min read

Overview

Language is sequential and ambiguous.

Key ideas

  • Tokenisation, stemming, POS.
  • Constituency and dependency parsing.
  • Embeddings and transformers.

Worked example

Sentiment: "not bad" positive, "not good" negative.

Common pitfalls

  • Ignoring context window.

Why it matters

Pre-trained LMs dominate NLP benchmarks.

๐Ÿ“ Search

Q1. A* on grid vs BFS.

Show Solution

Heuristic reduces expansions.

๐Ÿ“ Bayesian Inference

Q1. P(disease | positive test).

Show Solution

Bayes rule.

๐ŸŽฏ Midterm Exam

Q1. Alpha-beta on given tree.

Show Solution

Prune when alpha >= beta.

๐ŸŽฏ Final Exam

Q1. First two splits of decision tree.

Show Solution

Highest information gain.