From: Lincoln Ritter (lritter_at_cs.washington.edu)
Date: Wed Oct 22 2003 - 08:56:30 PDT
Two Theses of Knowledge Representation
Jon Doyle, Ramesh S. Patil
Reviewed by Lincoln Ritter
Summary: This paper posits that restricting languages used for
knowledge representation to avoid high computational costs leads to a
high loss of expresivity and forces non-domain independence, but that
rational management of inference tools provide a reasonable compromise
between full expresivity and computational cost.
The most important point this paper offers is that there is a
trade-off between computational efficiency and expresivity when
designing and using a particular model for KR. Furthermore that, as a
result of restricting languages to garner efficiency, generality is
lost in these languages is significant as it implies that the exact
constructs needed for general purpose knowledge representation systems
are the ones that make such systems impractical.
Not to be left in despair, the authors try and illustrate that by
judiciously using expensive constructs, loss of expressivity can be
minimized or eliminated, while keeping complexity to a minimum. This
is important as it implies there is hope for general (or nearly
general) KR systems.
While the the bredth of examples given by the authors illustrating the
loss of generality in restricted systems was impressive, it is not
fully convincing as they present no formal verification of their
assertions. While most of the examples are straightforward, perhaps
there is a really, really clever encoding that is being missed
somewhere. This would obviously undermine their argument.
Additionally, many of the criticisms offered by the authors are bit
obvious. Moreover, the effects of the language restrictions they
present were largely know to those using the restricted languages.
So, really, the authors are not presenting any new information.
Looking to the future, it seems that determining how much expresivity
is needed in a given situation, or accross a set of situations, would
be useful. In other words, is it possible to formally study the
exactly how expressivity decreases with restriction and how
computational efficiency increases. Is there optimal point where
these two "curves" intersect?
This archive was generated by hypermail 2.1.6 : Wed Oct 22 2003 - 08:56:30 PDT