From: Danny Wyatt (danny_at_cs.washington.edu)
Date: Wed Oct 22 2003 - 08:09:09 PDT
"Two Theses of Knowledge Representation" by Jon Doyle & Ramesh S. Patil
Summary:
The authors detail the impoverishments that result when a KR language is
restricted in accordance with Brachman & Levesque's objections to
systems that perform classification-based reasoning in greater than
polynomial time. They argue that this is an unreasonable sacrifice of
expressiveness in favor of efficiency, and that even that type of
efficiency may be of limited use and in limited demand.
Two Most Important Ideas:
Rigidly separating the upper and lower ontologies doesn't make much
sense. Without having read the Brachman and Levesque paper, I'm
inclined to agree. Restricting certain classes of inference to only one
or the other seems arbitrary since humans don't appear to reason
differently about essentials than incidentals. Indeed, (as the authors
point out) essentials may not be fixed: they may need to be as mutable
and non-monotonic as the incidentals.
There's more than one measure for utility. Proscribing everything at
tractable running times, completeness, or soundness is just a way to
carve a system to fit a certain use. Other uses may not have the same
restrictions.
Two Flaws:
There's more than one measure for utility. Some applications may need
tractability, soundness, and completeness. The authors explain why (a
specific fruit of) B&L's restrictions is unacceptable for some cases,
and assert that their solution is "more general". Deciding what is
"more general" and what is "too niche" also seems arbitrary.
OK, you don't like the KL-ONE kids, specifically that NIKL boy. Much of
the paper seems predicated on two research projects that found NIKL too
restrictive, but there's no control case of trying to implement these
projects in another, contemporary KR language.
Two Open Questions:
Overly tight restrictions in formal languages can cause knowledge
engineers to "overload" other features of the language with workaround
semantics. Does this happen only in tractable classification systems,
or do other KR systems have hacks that effect meanings other than their
explicit ones?
"The moral is that the tradeoff between expressiveness and complexity
means that there is no general purpose language." Must all KR systems
occupy specific points on the expressive-efficient spectrum? Are there
no better inference (classification based or otherwise) algorithms to be
found? The authors use the complexity and ambiguity of natural language
as an example: the (relatively) slow human brain comprehends natural
language in real time. Advances are still being made in Natural
Language Processing, should we expect the same in formal-logic language
processing?
This archive was generated by hypermail 2.1.6 : Wed Oct 22 2003 - 08:09:54 PDT