You are on page 1of 6

Daniel

Bauer
December 6, 2017
COMS W4701 – Artificial Intelligence
Final Exam, Topic Overview

General Concepts (Lecture 1 – R&N Ch. 1)

• Definitions of AI: Thinking Humanly, Thinking Rationally, Acting Humanly, Acting
Rationally
• Turing test.

Agents (Lecture 2 – R&N Ch. 2)
• Agent function vs. Agent program
• Rationality.
• Performance measure.
• Task environment descriptions using PEAS (Performance, Environment, Actuator,
Sensors).
• Properties of task environments: Observable?, Single vs. multi-agent, Deterministic?,
Episodic vs. sequential, Static vs. dynamic, discrete vs. continuous.
• Skills: be able to write down a task definition in a PEAS table. Recognize properties of a
task environment.

Search (Lecture 3&4 – R&N Ch. 3.1-3.4)
• Goal-based agents (subtype of Model-based agents).
• Formulating search problems: states, initial states, actions, transition model, goal test,
path costs.
• State space.
• Search tree.
• State space regions (explored, frontier, unexplored).
• Generic tree search and graph search (using an explored set).
• Properties of search strategies: completeness, optimality, time/Space complexity.
• Using backpointers to retrieve solutions.
• Examples: Route finding on a map / on a grid. 8-queens problem. 8-puzzle.
• Skills: Write down a formal definition for a search problem.

Uninformed Search (Lecture 3&4 – R&N Ch. 3.1-3.4)
• BFS (algorithm and properties)
• DFS (algorithm and properties)
• Depth-limited search.
• Iterative-deepening search.
• Uniform-cost search (UCS, algorithm and properties).
• Skills: Simulate the steps of BFS, DFS, UCS on paper.


Informed/Heuristic Search (Lecture 5, R&N Ch. 3.5, 3.6)
• Greedy best-first search.
• A* Search (algorithm and properties)
• Heuristic functions:
o Admissibility
o Consistency
o Why these properties matter for A*.
• Straight line distances and Manhattan distance.
• Skills: Simulate the steps of Greedy best-first search and A* on paper. Decide if a
heuristic is admissible.

Games / Adversarial Search (Lecture 6 & 7, R&N Ch. 5.1-5.5)

• Zero-sum games
• Formulating adversarial search problems: states, initial state, player function, actions,
transition model, terminal test, utility function.
• Minimax value, minimax trees, Max vs. Min Node. Minimax algorithm and properties.
• α-β Pruning (Algorithm and properties). Effect of node ordering (worst case, best case).
• Making decisions under limited time:
o Cutting off search (depth limit). Can also use iterative deepening.
o Evaluation functions: linear combination of features of the state.
• Stochastic games:
o Chance nodes.
o Stochastic Minimax.
o Evaluation function in stochastic games.
o Monte-Carlo simulations (basic idea).
• Examples: Tic-tac-toe. Othello. Chess. Go.
• Skills: Simulate Minimax and α-β on paper.

Constraint Satisfaction Problems (CSPs), (Lecture 8 & 9, R&N Ch. 6)
• CSPs vs. search problems.
• Defining CSPs: variables, domains, constraints
• Types of variables:
o Discrete, finite infinite, continuous.
• Types of constriants:
o Unary, binary, global (e.g. Alldiff)
• Constraint graphs. Hyperedges.
• Search vs. Inferences.
• Backtracking search (vs. regular search).
o Minimum Remaining Value (MRV) heuristic to select next variable.
o Least Constraining Value (LCV) heuristic to select next value assignment.

• Constraint propagation:
o Local consistency: Node consistency, arc consistency.
o Forward checking during backtracking search.
o AC-3 algorithm.
• Benefits of tree-structured CSPs.
• Examples: Map Coloring. 8-queens as CSP. Cryptarithmetic puzzles. Course assignment
(homework)
• Skills: Formulate a CSP problem. Simulate backtracking search on paper, including
heuristics. Apply AC-3 to a CSP (make the problem arc consistent).

Logical Agents (Lecture 10 & 11 & 12, R&N Ch. 7.1-7.5, 8.1-8.2)
• Knowledge-based agents, problem solving using inference.
• Syntax vs. semantics of formal logics. Model-based semantics (“possible worlds”).
• Syntax of propositional logic: recursive definition as a CFG, atomic and compositional
propositions, logical connectives. Order of precedence.
• Semantics of propositional logic: Truth tables.
• Tautology, contradiction, contingency.
• Entailment, logical equivalence.
• Validity, satisfiability.
• Two ways of determining entailment:
o Model checking
o Theorem proving using inference rules (deduction).
• Inference rules:
o Soundness, completeness.
o Validity and Satisfiability.
• Automatic theorem proving:
o Resolution:
§ Conjunctive Normal Form (CNF).
§ Unit resolution and full resolution.
§ Resolution algorithm (proof by contradiction)
o Forward/Backward chaining:
§ Horn clauses
• Basic syntax of first order logic, universal and existential quantification.
• Examples: Wumpus World.
• Skills:
o Translate natural language sentences into propositional and first order logic.
o Determine Entailment, Equivalence, Validity, Satisfiability using model checking.
o Basic deductive proofs (using the inference rules mentioned in the slides).
o Convert propositional logic sentence into CNF and proof by resolution.

Probabilistic Reasoning (Lecture 13 & 14 – R&N Ch. 13, 14.1-14.4)
• Uncertainty in decision making.
• Probability basics: Sample space, probability model, events.
• Random variables (Boolean, discrete, continuous).
• Prior probability.
• Joint probability.
• Conditional (posterior) probability. Product rule. Chain rule.
• Probability tables.
• Inference by enumeration. Marginalization. Conditioning.
• Normalizing in case of unknown prior.
• Independence and conditional independence.
• Bayes’ rule.
• Simple Bayesian Networks.
• Examples:
o Deciding when to leave for the airport.
o Determining whether a patient has a cavity given they have a toothache and/or
the dentist’s probe catches.
o Burglary or Earthquake and Alarm example.
o Wumpus world (using independence assumptions).
• Skills:
o Apply Bayes’ rule.
o Estimate probabilities from observations by counting.
o Given a Bayesian network, compute the joint probability and answer inference
questions by enumeration.

Machine Learning (Lecture 15,16,17,18,19 – R&N Ch. 18.1-18.4, 18.7 Coverage in the textbook
is slightly different)
• Concepts: Supervised, unsupervised, semi-supervised, reinforcement learning.
• Inductive learning (vs. deductive learning). Ockham’s Razor.
• Supervised learning:
o Input/output samples.
o Hypothesis function.
o Classification vs. regression.
o Loss function (least squares error, classification error).
o Overfitting/underfitting. Complexity vs. goodness of fit.
o Structural risk minimization.
o Evaluating ML models:
§ Training/dev/test set.
§ K-fold cross-validation.
§ True/false positives/negatives. Precision, Recall, F-measure.
• Probabilistic models:
o Discriminative vs. generative models.
o Naïve Bayes’ Classifier.
§ Estimating prior and joint probabilities from data, including smoothing.
• Examples:
o Deciding whether to hire a software engineer.
o Spam filtering.
• Skills:
o Estimate probabilities for a Naïve Bayes classifier from data and perform the
calculations for a naïve Bayes classifier on paper.

• Decision Trees
• Decision tree learning:
o Entropy.
o Information gain.
• Skills:
o Evaluate a given decision tree given input.
o Construct decision trees by selecting the attribute with the highest
information gain.
• Linear Models
• Linear regression.
o Least squares loss.
o Closed form solution for convex loss functions.
o Gradient descent.
o Skills: Compute loss on some data given a linear regressor.
• Perceptron (linear classification):
o Vs. biological neurons
o Activation function (step/threshold function).
o Perceptron learning algorithm
o Convergence theorem
o Linear separability
o Examples: representing Boolean functions. XOR is not linearly separable.
o Skills:
§ perform perceptron learning on paper.
§ create a multi-layer network of perceptrons for Boolean functions.
• Neural networks:
o Non-linear activation functions (sigmoid, rectifier function).
o Two-hidden layers can represent any function.
o Multi-layer neural networks.
o Basic idea of backpropagation.
• Deep learning (not in the book):
o Deep neural networks (more than one hidden layer).
o Deep learning = automatic feature extraction + classification.
o Feature hierarchies.
o Convolution neural nets:
§ Convolution: filter, feature map, depth, stride, padding
§ Pooling
o Examples: Image classification, Handwritten Character recognition.
o Skills: Design a convolution neural net.
• Reinforcement Learning (not on exam, Lecture 20, R&N Ch. 17.1-17.3, 21.1-21.3):
o Markov Decision Processes
o Reward function and utility function
o Policy function
o Bellman Equations
o Value iteration, policy iteration.
o Q-Learning
o Exploitation vs. exploration
• Natural Language Processing (not on exam, Lecture 21, R&N Ch. 22.1-22.2, 23.1, 23.2,
book differs significantly)
o Applications: Text classification, sentiment analysis, summarization, question
answering, machine translation, spoken language dialog systems, virtual
assistants, image captioning.
o Levels of representation (phonetics, phonology, morphology, syntax,
semantics, pragmatics).
o Ambiguity.
o Coreference.
o Low-resource languages.
o Sequence tagging (Part of Speech tagging, Named Entity Recognition).
§ Hidden Markov models: Markov chains, transition probabilities,
emission probabilities.
§ Viterbi algorithm.
o Context free grammars.
o Constituent parsing with CFGs:
§ Bottom up vs. top-down.
§ Relation to search problems / Dijkstra’s .

You might also like