a (broader) note about EBL

[Subbarao Kambhampati <rao@parichaalak.eas.asu.edu>]

So we talked about the value of "explanation" of why an
scenario/instance satisfies a concept (failure/success/cupness etc) 
in "learning" and "generalizing" from that experience. 

This looks like a pretty powerful idea on the face of it--we used it
to "generalize" the order in a given plan and, and we used it to
improve the effectiveness of "memoization". 

However, there is an important caveat--EBL's effectiveness is pretty
much dependent on having a good "causal theory" of the concept under
consideration (e.g. the "theory of when a plan is correct", the theory
of "when a search branch fails). 

Now where does this causal theory come from originally? We would  have
to have "learned" it and then hopefully validated it. The problem is
that this learning process is frought with errors--leading often to
incorrect causal theories. 

  When we are not sure of the correctness of the causal theories, or
if our explanations are "vacuous" (e.g. your accident hapened
because of everything that was part of the scene of and the scenes
leading to the accident" or "this particular partial plan failed
because it crossed some artificially imposed depthlimit for the
search.) then we will not be able to take advantage of EBL. 

 So, there is a lot of temptation to provide a "plausible sounding"
"circumstantial" explanation for the concept and use it to predict
from it latter. Sometimes, this may well lead to improvements in
performance. Afterall, most interesting learning is "inductive" and
thus fallible (not deductively provable)--so why not give a plausible
sounding explanation for depth-limit failures just so that you can try
to avoid fruitless search branches latter on? Well, it may improve
efficiency, but then it may also lead to loss of completeness (and
sometimes soundness). After all, one can view the Salem witch hunts
and the HIV-induced homophobia as EBL applied with faulty causal
theories..

regards
Rao
[Feb 20, 2003]

ps: some gratuitous quotes from my quote file.. ;-)
--------------
  "Induction extends our expectation, not our experience.."

   "All models are wrong. Some models are useful, - George Box 

  "Life is really simple. You do somethings. Many fail. Some succeed.
  You do some more of the things that succeed. If they succeed big,
   others will copy you. Then you do something else. 
  
  The trick is in doing something else."
                       Tome Peters 
                       Chicago Tribune