[mod.ai] Seminar - An EBG system which Learns from Failures

FAWCETT@RED.RUTGERS.EDU (Tom Fawcett) (12/08/86)

On Thursday, December 11th in Hill-250 at 10 AM, Neeraj Bhatnagar will
present a talk on learning from failures.  The abstract follows.
PLEASE BE PROMPT; we only have the room until 11:10.

          AN EBG SYSTEM THAT LEARNS FROM ITS FAILURES


I shall discuss my implementation of a design system that learns from its
failures. The learning technique used is the explanation based
generalization widely reported in the literature with the modification that
our system tries to explain the failures that it encounters in its search
for solution. These explanations give necessary conditions for success which
are used for pruning out the unacceptable solutions.  The implemented system
reported here acts as a Generate and Test (GT) problem solver in its general
problem solver mode. In its learning mode it tries to explain the reason why
a generated solution turned out to be unacceptable and generalizes this
explanation to prune out the failure paths in future.

The test bed for experimenting with the suggested technique is a restricted
version of the floor planning domain. Due to the restrictions imposed by us
on the operators used for planning the failures that can occur while
planning have become monotonic in nature which has facilitated their
detection, explanation and recovery from them.

Time permitting, I shall also discuss some of the future directions of my
research which include detection, proof and recovery from non-monotonic
failures, defining new terms and new operators in the context of explanation
based learning and a suggested method for making more effective use of the
knowledge learned by explanation based generalization.
-------