ROSIE@sri-unix.UUCP (08/14/84)
[Forwarded from the MIT bboard by SASW@MIT-MC.]
DATE: Thursday, August 16, 1984
PLACE: NE43-7th Floor Playroom
MECHANISMS FOR LEARNING
Kamesh Ramakrishna
Ohio State University, Columbus, Ohio
A number of different mechanisms for machine learning have been
proposed, though the definition of "learning" itself has not been
particularly clear. We show that many proposed learning mechanisms
can be placed into two classes and that mechanisms within each class
are reducible to each other. These two classes correspond roughly to
the "knowledge acquisition" and "skill refinement" classes proposed by
Mitchell, Carbonell, and Michalski; however, (and more interestingly)
they correspond to the two different levels of knowledge-based
processor architecture proposed by Newell in "The Knowledge Level".
The knowledge acquisition type learners appear to be at the
Symbol/Program level. This observation lets us integrate this
approach to machine learning with the taxonomy of problem-solving
types proposed by Chandrasekaran et al., leading to the hope of an
integrated knowledge-level approach to both problem-solving and
learning.
With appropriate restrictions placed on the functioning of the
learning mechanisms, we show that the two classes also differ in the
fundamental learning problem that they solve. We identify some
learning problems that are not solved by either class -- identifying
some possible future directions for research.
HOST: Prof. Ramesh Patil