From: Lucas Kreger-Stickles (lucasks_at_cs.washington.edu)
Date: Mon Dec 08 2003 - 07:52:14 PST
The authors present PROVERB, a system which uses a number of component
modules working on individual tasks which are then combined together to
solve crossword puzzles.
One of the ideas that I really like from this paper is the concept of
having specific modules that are good at certain tasks and which are
combined together to solve the larger, harder problem. This process
reminds me in many ways of the ensemble classifiers discussed earlier in
the course and of 'blackboard' architectures I have seen before.
This approach was interesting in particular for how it brought these
component parts together, relying on probability models gleaned from a
huge corpus of solved crossword puzzles. It is this combination of
modularization with the idea of training data that I found most
interesting and exciting.
While by and large I found the paper quite good and very readable, I had
a hard time following the exact mechanisms that they used to construct
the various weightings and probabilities for the various modules and for
the grid filler.
For future research I would like to see what other sorts of problems can
be tackled using this technique (expert modules combined using
probability modules gleaned from a corpus of examples)
Brief Review:
<i>Automated Theory Formation in Mathematics</i>
Douglas Lenat
This seems like an almost intractable problem and it occurs to me that
much of what the author was doing is having the system return bad 'proof
by example' solutions which were then confirmed by the user. That
aside, the one suggestion I would have for how to improve such a
situation given todays hardware and AI would be the use of an
explanation based learning system. The author points out that the
system often chokes once it starts to operate on higher level concepts
which it has learned and it must do so using the primitives originally
given to it. This would be akin to defining fibonaci numbers
recursively and then trying to perform mathematical functions by using
all the ones that lie at the leaves of the tree. If instead the system
could generalize the concepts that it learned and 'chunk' them together
it would have a better chance of manipulating higher order concepts.
This archive was generated by hypermail 2.1.6 : Mon Dec 08 2003 - 09:49:58 PST