Intro to AI

What this covers

• Formulating a problem as a state space and choosing between BFS, DFS, uniform-cost, greedy best-first and A*; proving a heuristic admissible or consistent and comparing heuristic dominance
• Minimax and expectimax on game trees, hand-tracing alpha-beta pruning, and reasoning about how move ordering changes the number of nodes pruned
• Constraint satisfaction: backtracking search with forward checking and arc consistency (AC-3), plus MRV, degree and least-constraining-value heuristics
• Propositional and first-order logic: converting English to sentences, CNF conversion, resolution refutation, unification, and the difference between entailment, validity and satisfiability
• Probabilistic reasoning: joint distributions, Bayes' rule, building and reading a Bayesian network, d-separation, and variable elimination by hand
• Sequential decision making and learning: MDPs, Bellman equations, value and policy iteration, Q-learning versus value iteration, and supervised learning basics including overfitting, train/test splits, perceptrons and gradient descent

Where learners get stuck

Treating 'admissible' and 'consistent' as the same property, and assuming any heuristic that 'looks reasonable' is admissible.

Courses introduce both terms in one lecture and most textbook heuristics happen to satisfy both. Students rarely see the counterexample where an admissible-but-inconsistent heuristic makes A* re-expand a node, so the distinction never gets grounded in a concrete failure case.

Reading a Bayesian network as a causal flowchart and concluding that two nodes with no edge between them are independent — missing 'explaining away' at a collider.

The graph looks like a dependency diagram from a data structures course, so students apply reachability intuition. The rule that conditioning on a common effect creates dependence runs against everyday intuition about causes being unrelated.

Believing alpha-beta pruning can change the value minimax returns, or that pruning is a heuristic approximation.

Every other speed-up students have met (greedy search, sampling, early stopping) trades accuracy for time, so they generalise the pattern. Tracing a tree by hand and seeing the root value stay identical is what fixes it.

Confusing the Q-learning update with the Bellman equation for value iteration, and assuming Q-learning needs the transition model.

The two equations are written with nearly the same symbols on the board, and the model-free versus model-based distinction is stated once in passing rather than demonstrated on an example where the transition probabilities are genuinely unknown.

What a session looks like

You talk through problems with the AI tutor by voice. A typical session might be: state a search problem out loud and defend your choice of algorithm, then hand-trace A* on a small graph node by node while the tutor checks your frontier and explored set at each step. For probability topics, you'll be asked to state which independence assumptions a network encodes before computing anything. The tutor pushes back when an answer is right for the wrong reason — for example, when a correct minimax value comes from an incorrect pruning trace — and will ask you to re-derive rather than supplying the next line. You can bring a lecture slide, problem set question or exam past paper and work from it.

Helpful to know first

• Programming in Python at the level of loops, functions, classes and recursion (most intro AI assignments are Python-based)
• Basic data structures: stacks, queues, priority queues, hash sets, and graph representation — enough to know why a frontier is a priority queue in UCS and a stack in DFS
• Discrete mathematics: propositional logic, quantifiers, proof by induction, and Big-O notation for stating branching-factor complexity
• Introductory probability: conditional probability, independence, Bayes' rule, expectation over a discrete distribution
• Enough calculus and linear algebra to follow a gradient and a dot product; vector notation for the machine learning weeks

Questions

Is Intro to AI the same thing as a machine learning course?

No. Machine learning is usually the last three or four weeks of an intro AI course. The bulk of the syllabus is search, games, constraint satisfaction, logic and probabilistic reasoning — topics with no training data involved. If your course is entirely neural networks and model fitting, it is a machine learning course and the emphasis here on search and inference will only partly match.

How much math do I need for intro AI?

Less than for a machine learning course, but more than for data structures. You need conditional probability fluently, comfort with summations, and the ability to follow a proof by contradiction (used for A* optimality and resolution refutation). Calculus appears mainly in the gradient descent section. If your probability is shaky, that is the single highest-value gap to close first.

Can you help with my Pacman or game-agent programming project?

The tutor can help you understand the algorithm the project asks you to implement, trace it on a small example, and reason about why your output differs from the expected one. It works by voice, so it is not writing or debugging code line by line for you, and you should follow your institution's academic integrity policy on outside help.

What is usually on the intro AI midterm?

Most midterms cover search through probability: hand-tracing a search algorithm, proving or disproving admissibility, an alpha-beta pruning trace, an arc-consistency or backtracking trace, a resolution proof, and a Bayes net question on independence or variable elimination. These are traceable-by-hand problems, which is why speaking through the steps is effective preparation.