urquhart@utcs.UUCP (Prof. A. Urquhart) (08/04/85)
Since there seems to be some interest in the idea of inductive
learning schemes for expert systems, let me give some precision
to the general question I posted a while ago.
I wanted know if there has been any research done to make an
expert system learn rules from given 'facts' using an inductive
inference system. The rule thus learned would either prompt a
question to the human 'teacher' or else have a corresponding
probability, which would be used in drawing inferences (using
certainty factors as in MYCIN or Baysian probability measures as
in PROSPECTOR). For example, if a system was given a population
of 15 individuals ten of which bore the relation of 'child-of' to
some other individual, the inductive inference system would
notice a statistically relavent fact, namely that 10 out of 15
individuals are 'child-of' someone. It might, then, either
"experiment" with the world and ask "is everyone someone else's
child?" or else, perhaps, make the assumption that this rule is
true with probability 2/3.
It is clear that in a large database some clever 'learning
strategies' would have to be used to prevent a meaningless
accumulation of rules, but what about the general idea? Has it
been done? Is there any literature?
Will mail replies to interested parties.
--
----
Andre Vellino
University of Toronto
{cbosgd,decvax,harpo,ihnp4,utcsri,{allegra,linus}!utzoo}!utcs!urquhart