Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!rutgers!ames!ucbcad!ucbvax!CAD.CS.CMU.EDU!Steven.Minton
From: Steven.Minton@CAD.CS.CMU.EDU
Newsgroups: mod.ai
Subject: Seminar - Learning Decomposition Methods (CMU)
Message-ID: <8703190539.AA21991@ucbvax.Berkeley.EDU>
Date: Wed, 18-Mar-87 01:11:28 EST
Article-I.D.: ucbvax.8703190539.AA21991
Posted: Wed Mar 18 01:11:28 1987
Date-Received: Fri, 20-Mar-87 05:28:06 EST
Sender: daemon@ucbvax.BERKELEY.EDU
Distribution: world
Organization: The ARPA Internet
Lines: 22
Approved: ailist@sri-stripe.arpa

This week's speaker is Sridhar Mahadevan. As usual, the seminar is
in 7220 Wean on Friday at 3:15. Come one, come all.

	LEARNING DECOMPOSITION METHODS TO IMPROVE HIERARCHICAL 
		PROBLEM-SOLVING PERFORMANCE

Previous work in machine learning on improving problem-solving
performance has usually assumed a @i(state-space) or "flat"
problem-solving model.  However, problem-solvers in complex domains,
such as design, usually employ a hierarchical or problem-reduction
strategy to avoid the combinatorial explosion of possible operator
sequences. Consequently, in order to apply machine learning to
complex domains, hierarchical problem-solvers that automatically
improve their performance need to designed.  One general approach is
to design an @i(interactive) problem-solver -- a @i(learning
apprentice) -- that learns from the problem-solving activity of expert
users. In this talk we propose a technique, VBL, by which such a
system can learn new problem-reduction operators, or @i(decomposition
methods), based on a verification of the correctness of example
decompositions. We also discuss two important limitations of the VBL
technique -- intractability of verification and specificity of
generalization -- and propose solutions to them.