🚀 go-pugleaf

Can anyone think of any domains/tasks that a learning algorithm with the following
properties could be applied to?

1.	Learns DNF concepts in 0-order logic (i.e. attribute/value).
2.	Learns "long" conjuncts (relatively specific conjuncts which cover only
	a small percentage of the examples in the Universe).
3.	Needs some negative examples, but only a few hundred at most.
4.	Needs 10 or more (on average) positive examples of each conjunct.

General-to-specific learning algorithms usually start with the most general concept
possible (i.e. assumes that everything in the Universe is a positive example), and restrict it
while known negative examples are covered.  If the target concept has long conjuncts in it,
this tends to result in guessed concepts with conjuncts that are more general (and thus cover
many unknown negative examples) than the target concept.  This is because we need huge numbers
of random negative examples to be persuaded that we have found all the predicates in a
conjunct of the concept  (Of course, a teacher would supply a small number of
critical negative examples - I am assuming that the examples are all randomly selected from
the Universe of possible examples).

An algorithm which I have been investigating (called "Weka" after a New Zealand bird)
does not stop restricting a conjunct when it covers no negative examples, but continues
to make guesses about which predicates are likely to be necessary.

		Regards,

			Paul Brebner.

Softway Pty Ltd, P.O. Box 305, Strawberry Hills, NSW 2012, AUSTRALIA
Phone: +61 2 698 2322;       Fax: +61 2 699 9174;     Telex: AA27987
Internet: brebs@softway.oz.au	    UUCP: ...!uunet!softway.oz!brebs