Re: Some references about ontology and analogy: SUO redux
Pat,
DARPA's "Grand Challenge" is intended to be a challenge.
Its goal is to push technology for the next 10 to 20 years
with a range of challenging tests. Some of the simpler
examples could be handled with existing technology, while
others would be sufficiently difficult that they might
take years to solve.
PC> If you used instead a geometric representation of scenes
> (as in a flight simulator) and then asked the computer
> questions about its internal representation, that would
> still be a challenge, and if successful could be extended
> to the scene recognition problem.
That is one of the options. Following is an excerpt from
page 2 of http://www.jfsowa.com/ai/gcprop.pdf
Since very few groups have expertise in the entire range
of technologies that are needed, the overall task could be
divided into four steps:
1. Analyze an image and generate a geometrical model of the scene.
2. Map the geometrical model to a description in some knowledge
representation language.
3. Map the English question a logical form in some knowledge
representation language, which could be the same language as
in step #2 or some superset of it.
4. Answer the question from step #3 in terms of the description
from step #2.
What you're suggesting is to assume that the steps #1 and #2 have
already been done and to focus on steps #3 and #4. That is what we
were proposing: allow groups to address one or more of the four
steps or to form partnerships with other groups that are doing some
of the other steps. We suggested that winners could be announced for
each of the four steps as well as an overall winner (which could be
some partnership of two or more groups).
In their solicitation, DARPA asked for suggestions that defined
what to do, but not how to do it. In our proposal, we made no
assumption about the knowledge representation languages, the
linguistic theories, or the theories of geometrical modeling and
representation. It's possible or even likely that the challenge
of integrating the end-to-end process from perception to language
will stimulate the development of theories, representations, and
computational methods that are very different from any we have today.
Re Cyc: There are many very complex issues involved, but Cyc has
definitely evolved very far from where it started in 1984, and Lenat
has consistently said that every design change was motivated by the
requirements of dealing with very serious problems they faced. Their
original system, as documented in their first 5-year report (published
in book form by Lenat & Guha in 1990) was a frame-based system with
no microtheories -- in fact, it was very similar to a typical DL
such as OWL. But they changed it because they had to.
PC> The extensive use of microtheories may well have allowed them
> to ignore fundamental problems by pretending that they could cut
> off parts of the ontology from each other.
Their original hierarchy was monolithic, and it became impossible
to extend it consistently. Anybody who argues for a single
monolithic ontology will have to explain why they believe they
can do something that the Cyc group tried and was forced to abandon.
PC> I would worry about the loss of time developing agreement on
> general principles of ontology development...
Please read (or reread) Fred Brooks' _Mythical Man Month_. The
desire to plunge in and start coding is always very hard to resist.
There is nothing wrong with giving in to that desire as a learning
exercise, but it must be recognized that the results must be thrown
away before the real work can begin. Trying to preserve that
throw-away inevitably leads to disaster.
John Sowa