[comp.ai.digest] Reasoning with Defaults - Hector Geffner

dlm@allegra.att.COM (05/12/89)

Date:    Monday, 5/22/89 
Time:    10:00 am
Place:   AT&T Bell Laboratories - Murray Hill 3D436


Title:   Reasoning with Defaults

Speaker: Hector Geffner 
	 Cognitive Systems Laboratory - Department of Computer Science, UCLA


Defaults play a central role in commonsense reasoning, permitting the
generation of useful predictions in the absence of complete information.
These predictions are nonmonotonic, in the sense that they often need to
be revised in light of new information.  A number of extensions to classical 
logics have been proposed which successfully accommodate this non-monotonic 
behavior.  Recent work in defeasible inheritance, however, has shown that 
there are additional issues, beyond non-monotonicity, which also need to 
be addressed in order to capture the defaults intended meaning.

I will present two alternative formalizations which address these issues.
In the first part I will discuss a qualitative inference system that
results from interpreting defaults as high conditional probability
statements.  The system is characterized by a core of five rules of
inference which permit derivations to be constructed in the style of
natural deduction systems and which capture the context-sensitivity of
defaults.  A sixth rule is then introduced which extends the core with
assumptions about conditional independence.

In the second part of the talk, I will present a model theoretic account
which provides an alternative validation of both the core rules and the
conditional independence assumptions.  This account appeals to a preference 
relation among models hinted by the probabilistic interpretation.  We will 
then present some examples and discuss ideas about implementation.

Sponsor: David Etherington  - ether@research.att.com