Conclusions

PSU CS441/541 Lecture 9
November 29, 2001

• Topics:
  • Bayes' Nets: George Borden
  • Genetic Algorithms and Genetic Programming
  • Project (due date change?)
  • Exam Review (Final 7:30-9:00 Thursday)
  • Course Evals
• Genetic Algorithms and Genetic Programming
  • Idea: Nature is smart (Sphex wasp again)
    • Perform local-search-style optimization
    • Work on population of solutions at once
    • Replace heuristic repair with simple keep-or-discard
  • Questions:
    • How smart can something that takes 50M yrs be?
    • Interaction between learning, tropism, etc.?
    • How seriously do we take natural model of evolution?
      • genome -> organism map
      • crossover, mutation
      • interacting populations
  • GA model
    • Create random population
    • Discard bad instances
    • Create new instances to replace discards
    • Repeat last three steps until satisfied
  • Genetic Programming: use GA to evolve algorithm
  • Nilsson example: reactive wall follower
    • Represent reaction as prop-logic tree
    • Create 5000 random reactions
    • Run each for 60 steps in grid world
    • Pick groups of 7 and select best until 500 chosen
    • Fill in remaining 4500 by exchanging subtrees of pairs out of 500 survivors
    • (Could mutate 500 or so during creation)
    • Repeat loop with new 5000
  • More Questions:
    • How to pick magic numbers, algorithm details?
    • Note redundancies in good programs (e.g. NOT(NOT(NOT(NOT(NOT(X)))))). Why? How to cope?
  • The cool thing about GAs/GPs: no heuristics necessary
  • The AI "Big Tent": consider
    • STRIPS Planning
    • Bayes' Nets
    • Neural Nets
    • GAs
• Final Exam Review
  • Games
    • Adversary search
    • Minimax
    • Depth-limits and heuristic evaluators
  • KR: Horn clauses, Prolog, and rule-based systems
  • Truth Maintenance Systems and assumptions
  • Probabilistic inference
    • Likelihood v. probability
    • Bayes' rule, Bayes Nets
  • Neural Nets
    • neurons, weights, nonlinearity
    • layers, evaluation, back-propagation
    • transparency, under- or over-training, performance
  • Planning
    • initial state, goal, actions, plan
    • Reasoning about actions using situation calculus / FOL
    • STRIPS
    • Blocks World
    • forward v. backward chaining
  • GAs
    • Concept
    • Advantages
    • Problems