From: Sandra B Fan (sbfan_at_cs.washington.edu)
Date: Wed Oct 22 2003 - 04:09:59 PDT
"Two Theses of Knowledge Representation"
Jon Doyle & Ramesh S. Patil
Doyle & Patil claim that Levesque and Brachman's restricted
classification and restricted languages theses are actually
counterproductive, and argue instead for a broader view of KR.
One of the main points of the paper is that restricting a language in the
interest of maintaining polynomial runtime may hurt the system in the long
run because the language becomes less general and less expressive.
Levesque and Brachman figure that by making the language more sound and
complete, they enable it to be more useful in general, but Doyle & Patil
believe it is less so, because you end up creating more problems.
They believe that merely measuring worst-case efficiency is not the way to
go when considering the design of a system, because the yardstick by which
to measure efficiency is ill-defined, and because the worst case isn't
necessarily representative of the general case.
One of the flaws of the paper is that it seems to be one long, overly
detailed run of KL bashing. I think it's good that they're arguing a
different side, but a more concrete, detailed account of their solution
would have made the paper stronger.
Possible directions for research include trying to find a way take into
account both Levesque and Brachman's view, as well as Doyle and Patil's.
Maybe there's some happy medium we can come to. Perhaps the solution is to
keep Doyle & Patil's less restricted solution, but use Levesque &
Brachman's more restricted one when needed.
This archive was generated by hypermail 2.1.6 : Wed Oct 22 2003 - 04:10:21 PDT