From: Danny Wyatt (danny_at_cs.washington.edu)
Date: Mon Dec 08 2003 - 11:39:40 PST
PROVERB: The Probabilistic Cruciverbalist
Greg A. Keim, Noam M. Shazeer, Michael L. Littman, et al.
Summary
The authors present a system for solving crossword puzzles that
integrates solution suggestions from several separate solution modules.
The integration is based on the module's own rating of it probability of
success and the module's rating of each candidate it suggests.
Important Ideas
What seems like an overwhelming problem with many possible approaches
can be tackled by following all approaches and merging them. This kind
of voting and merging system seems to me to approximate one human way of
solving crosswords.
Probability is a good lingua franca. Each module need not justify its
candidates or their probabilities. I was unclear, though, on whether a
module's previous successes increased its future weight.
Flaws
There were no obvious flaws other than the brief descriptions of some of
the expert modules. I would like to know where to get more information
on the differences between there 4 Dijkstra/Mutual Information modules,
and how those in turn differ from their Encyclopedia module (which also
sounds like it uses MI).
Open Questions
This might be a topic of discussion elsewhere, but I'm curious about
these kinds of modular systems. How can they be extended in a more open
framework while remaining safe from poor or even hostile modules?
All of their information sources are static. I'm curious to see how
some of the IR modules work using the web as a corpus.
---------------------------------------------------
Automated Theory Formation in Mathematics
Doug Lenat
I'm not well-acquainted with the past 25 years of theorem-proving
research, but I don't think that improvements in computing power would
improve AM today. It would be able to reach the same conclusions
faster, and then venture out beyond them---but that would probably only
serve to produce such a glut of output that it would be useless without
improved heuristics to find the "interesting" output. As for
improvements in A.I. techniques, I'm sure there are more refined
mathematical models for deriving and using the kinds of scores that
guide AM's search, but they must still be tied to real insight into
what's interesting output and what isn't.
This archive was generated by hypermail 2.1.6 : Mon Dec 08 2003 - 11:39:03 PST