From: Karthik Gopalratnam (karthikg_at_cs.washington.edu)
Date: Wed Oct 22 2003 - 01:51:20 PDT
"Two Theses of Knowledge Representation: Language Restrictions, Taxonomic
Classification and the Utility of Representation Services"
- Jon Doyle & Ramesh s Patil
This paper presents an argument against restricting the constructs of a
knowledge representation system in order to gain speedup to
polynomial-time classsification and completeness, while sacrificing the
range of problems that can be expressed by the system.
The two most important ideas in the paper are:
1) The most general purpose knowledge representation systems must not
restrict their expressive power by omitting the constructs that require
non-polynomial time to classify concepts - *limiting* as opposed to
omitting these constructs, can lead to the system being capable
of expressing a significantly larger set of real world problems, thereby
enabling these systems to be more like the general-purpose representation
systems that they purport to be.
2) The measure of efficiency for the classifying power of a representation
system cannot be viewed in the simplistic / reductionist terms of
runtime-efficiency (often ill-defined) or completeness, but in terms of
its better quantified decision-theoretic utility, which allows for a more
versatile representation system that is better tuned to the requirements
imposed by the 'real' world.
The authors forward a compelling argument against the simplistic /
reductionist view taken by earlier representation systems by giving
examples of how the space of problems that can be expressed by a system
conforming to the Restricted Language thesis and the Restricted
Classification thesis is drastically reduced and they make a strong case
for a new paradigm of knowledge representation system, as well as the
need for a more real-world performance metric for such systems.
However, apart from describing the properties that such new systems
should possess in abstract terms, the authors do not provide rigorous
constructs for these more expressive representation systems. Such a
rigorous description, either of a complete system, or of the constructs
necessary to a pre-existing system such as KL-ONE or its descendants would
have lent much more credence to the properties described, as well as a
benchmark for the decision-theoretic performance measures proposed in the
paper.
Unless the authors have already done so, it would be interesting to see
mathematically quantified systems incorporating the features that the
authors deem necessary for representation systems to be effective. Also,
in light of the advances in dealing with uncertain knowledge, and the
burgeoning field of Statistical Relational AI, it might be enlightening to
reinterpret the authors' ideas, and evaluate them in a more contemporary
research setting.
This archive was generated by hypermail 2.1.6 : Wed Oct 22 2003 - 01:51:20 PDT