CS540: Topics for Final Exam. May 12, 1:30-3:30 PM, regular classroom

• Search
  • Hill-climbing with restarts
  • Simulated annealing
    • probability of taking a step downhill
    • annealing the temperature parameter in the probability distribution
  • Genetic algorithms
    • representation with bit strings of parameter values to be optimized
    • selection of individuals to reproduce
    • cross-over
    • mutation
    • hyperplanes
    • schema theory - do not need to know the formulas, but do understand them
• Belief Networks
  • conditional probability tables (CPT) - be able to calculate any joint probability given a belief network and the CPTs
  • Dempster-Schafer - adds explicit representation of ignorance
  • fuzzy logic - adds explicit vagueness
• Utility Theory
  • utility function
  • rational agents maximize expected utility (MEU) - given utilities and probabilities, be able to calculate the rational action
• Decision Networks
  • add decision nodes and utility nodes to belief networks
  • value of knowing the value of a variable with certainty, value of perfect information - be able to calculate for simple example
• Planning
  • Partial-order planning
  • Hierarchical planning
  • complex conditions, resource contraints (adding measures that change value)
  • conditional planning, inserting a test that is yet to be performed
  • defining a planning problem as a reinforcement learning problem (Zhang and Dietterich)
• Machine Learning
  • Ockham's razor
  • hypothesis space
  • decision trees - using information measure to determine which attribute to test at each node as decision tree is constructed
  • version space method
  • PAC-learning, just remember what "probably" and "approximately" refer to and how the final inequality is used.
• Computer Vision
  • Pinhole camera, derivation of perspective foreshortening
  • Binocular stereopsis
  • Finding edges
    • convolving with gaussian
    • convolving with derivative of gaussian
    • discrete equivalents - masks
    • Laplacian and zero-crossings
    • clustering points with similar gradients and fitting lines to them
    • Hough transform, just the basics