Nov 7, 2003
-------------

- Type C learning
	- inductive logic programming (ILP)
		- a form of type C

- Simple case
        - background knowledge is empty
        - Q: how can we learn concept of grandparent(x,y) using
          a decision tree or neural network?
        - A: you can't
        - Reason: need to capture relations between instances, and
          not just a mapping from one instance to one target value

- Essentials of ILP
        - reverses deduction process seen earlier

- Consider
        - 1: All CS courses are easy.
        - 2: CS 4804 is a CS course.
        - 3: CS 4804 is easy.

- Deduction is when
        - 1 and 2 are given
        - we infer 3

- Induction is when
        - 2 and 3 are given
        - we infer 1

- Deduction vs induction
        - logically sound vs. logically unsound
        - (general->specific) vs. (specific->general)

- More examples
	- learning who is a credit risk
	  (through spousal relationship)

- Example ILP systems
	- PROGOL
	- CIGOL