SIMS%PLU@AMES-IO.ARPA.UUCP (03/27/87)
Title: LEARNING CONCEPTS TO IMPROVE PERFORMANCE:
The Role of Context
By: Dr. Richard Keller
(KELLER@RED.RUTGERS.EDU)
Computer Science Department
Rutgers University
Where: NASA AMES
When: Monday, April 6
Concept learning, like most intelligent behavior, should be
influenced by the context in which the behavior takes place. If
concept learning occurs in the context of improving the performance
of a problem solving system, then the type of concept learned and
the form of its description should depend on the goals and the
capabilities of the problem solver. Unfortunately, most current
inductive learning systems incorporate a set of fixed, implicit
assumptions about the problem solver being improved by learning.
This causes problems when the original problem solver changes over
time, and also makes it difficult to reuse the same inductive system
to improve a different problem solver.
As an alternative to the inductive framework, I describe a new
concept learning framework -- the concept operationalization
framework -- which makes contextual assumptions more explicit and
easier to change. To illustrate the new framework, I discuss how an
existing inductive system (the LEX system [Mitchell et al. 1981])
was converted to a concept operationalization system (the MetaLEX
system). In contrast with LEX, MetaLEX adapts more successfully to
certain changes in its learning context, learns contextually
suitable approximations of its target concept as necessary or
expedient, and has the potential to automatically generate its own
concept learning tasks to improve its problem solver.