An Information-Based Approach to Learning for Rule-Based Expert Systems
Dr. Michael J. Pazzani
Department of Information and Computer Science
University of California, Irvine
pazzani@ics.uci.edu
4:00 pm Thur. Oct. 13 in 2310 Computer Sciences and Statistics Bldg.
A recent advance in concept formation, FOIL, The First Order Inductive
Learner (Quinlan, 1990) learns constant-free Horn clauses, a useful
subset of first-order predicate calculus. We describe the learning
system, FOCL, the First Order Combined Learner that adds a compatible
knowledge-intensive learning program to FOIL. We describe how FOCL
use a variety of types of background knowledge to increase the class
of problems that can be solved, to decrease the hypothesis space
explored, and to increase the accuracy of learned rules. In
particular, we concentrate on the problem when the background
knowledge consists of an existing set of classification rules that are
not entirely accurate on a database of classified examples. In this
case, the learning process is focused on acquiring rules that improve
the accuracy of the existing knowledge. We will report results on
several small problems (chess end-games and loan processing) and on a
larger problem involving the troubleshooting of a telephone network.