CS533 Final Exam Study Guide

The Final will cover everything in the course, but at least two-thirds of the exam will cover material that we have discussed since the midterm. This includes:

• Machine Learning and Pattern Recognition
  • Definition of classification learning problem.
  • Neural networks:
    • weights, bias values, sigmoid functions.
    • error function J(S,W), gradient descent search
    • stochastic gradient descent, conjugate gradient descent
    • measuring error using holdout data.
    • the problem of overfitting; solving via early stopping.
  • Decision trees:
    • structure and how they are executed to classify data points.
    • top-down divide-and-conquer method for growing trees.
    • Scoring a proposed splitting test using mutual information.
    • Pruning using a validation set.
    • Converting trees to rules.
  • Nearest neighbor method:
    • finding the k nearest neighbors.
    • importance of the distance metric.
    • problems with noisy or irrelevant features.
    • finding the nearest neighbor using a kd tree.
• Reinforcement Learning:
  • Definition of a reinforcement learning task: states, actions, immediate rewards.
  • Definition of policy, optimal policy, and value function.
  • Computing the optimal policy from the optimal value function.
  • The value iteration algorithm.
  • Temporal difference learning using a neural network.
• Diagnosis with Belief Networks
  • Diagnosis problem: repair the device while minimizing total average cost of repair.
  • Computing optimal policies in the repair-only case with the single-fault assumption (ratio of probability of failure divided by cost of observation).
  • Computing optimal policies in the general case by complete analysis of the decision tree (working backwards taking expected values and max's).
  • Computing approximately-optimal policies for the case where we have a mix of repairable and purely observable components.
• Probabilistic Reasoning:
  • Random variables, expected values, joint distribution, marginal probability, conditional probability.
  • Computing marginal and conditional probabilities from the joint distribution.
  • Algebra of probability distributions: chain rule, independence, conditional independence.
  • Belief networks. What probability distribution is stored at each node. Computing the joint distribution by taking the conformal product of all of the individual distributions.
  • Computing probabilities for diagnosis. Implementing the single fault assumption.