nov 8
finish bayes discussion
- assignment for Monday
- Jim's sample files and perl code in spam
chap 18 - learning
Three types:
- supervised (i.e. training sets)
- unsupervised (i.e. recognizing groups)
- rewards (explore; get feedback)
supervised learning basics:
- collection of examples (training set)
- model of machine / space of possible solutions to search
- test cases to evaluate model
Supervised learning is basically a form of "induction" :
deduce general principle from examples.
- Tradeoff between model complexity and search complexity:
- a simple model (i.e. line to describe data) is easer to fit ... but not as expressive
- Typical models favor simplicity : easier to find, easier to implement once found.
"decision tree" restaurant problem, pg 654
- explain algorithm for finding best decision true
- good variable for decision: best information content
- discuss information theory basic idea
- bad variables: chisq statistics test to avoid "overfitting"
- note that this is all basically "propositional logic" (at least if variables are boolean)
- ... but is in the same sort of problem as curve fitting
neural nets
human brain
AI
General discussion
Examples from the text
Software
Course at Willamette
