Nov 3, 2003
-------------

- Flavor 1 of Knowledge in Learning
	- EBG or EBL
	- explanation-based generalization, or
	  explanation-based learning

- General schema of EBL
	- (new knowledge ^ descriptions) entails classifications
	- (old knowledge) entails new knowledge
	- also known as "Type A" learning

- Two examples where you can recognize EBL
	- tying grocery bags so that contents don't
	  fall during a car ride
	- asking a polite request such as "I was wondering
	  if ..."

- More examples of EBG
	- derivative formulas
	- prearing cheatsheet for exams

- Output for EBG is always
	- a general rule, to apply in future situations

- Input for EBG is
	- an example (successful)

- Process of EBG
	- find a good example to learn from
	- "explain" it (derive a proof tree)
	- learn a generalized rule from proof tree

- What do we need to explain an example?
	- ans: a domain theory
		- another input to EBG

- Two ways of thinking of EBG
	- generalizing example -> rule
	- specializing domain theory -> rule

- Explaining and learning a rule involves
	- deriving the proof tree
	- chopping it off, with a cutting plane

- What is the role of data (examples) in EBG?
        - only serves to highlight the relevant parts of
          the domain theory
        - domain theory is sufficient to entail the
          new knowledge, but needs data to "guide it"

- Learning politeness
        - old knowledge: English
        - data: "I was wondering if you could point me
          in the direction of Central Park."
        - new knowledge: "I was wondering if X"

- Couldn't have done it with data alone
        - need to know rules of English

- Couldn't have done it with English alone
        - too many possibilities for specialization

- Slicing the proof tree too close to bottom
	- gives a rule that works only in few cases
	- but doesn't require too much effort to "apply"

- Slicing the proof tree too close to top
	- gives a rule that works in many cases
	- but requires some effort in filling in details

- Options
	- one (or few) generally applicable rule versus
	- many specially applicable rules