ylfink@water.waterloo.edu (ylfink) (12/13/88)
DEPARTMENT OF COMPUTER SCIENCE
UNIVERSITY OF WATERLOO
SEMINAR ACTIVITIES
ARTIFICIAL INTELLIGENCE SEMINAR
- Friday, December 16, 1988
Professor Larry Rendell, Department of Computer
Science, University of Illinois at Urbana-Champaign,
will speak on ``Learning Hard Concepts''.
TIME: 11:30 AM
ROOM: DC 1304
ABSTRACT
Hard boolean concepts have arbitrary compositions of
conjuncts and disjuncts. Hard graded concepts are
arbitrary functions having any composition of
disjuncts, modes, or peaks. If instances are k-tuples
of attribute values labeled with (binary or graded)
class membership values, learning hard concepts is
uniformly difficult for a class of methods that do
``similarity based learning'' or ``selective
induction''. The problem has been named differently in
different areas of research: ``curve fitting'' and
``density function estimation'' (statistics),
``discriminant boundary'' detection (pattern
recognition), ``finding weights'' (neural systems), and
``concept learning'' (artificial intelligence). The
poor behavior of these basic methods leads to
techniques for the creation of ``better'' attributes
through ``constructive induction.'' This talk
characterizes the problem and presents an algorithm
schema designed to use domain knowledge when available.
The analysis shows how constructive induction can
alleviate difficulties of selective induction.