Review: Two Theses of KR

From: Julie Letchner (letchner_at_cs.washington.edu)
Date: Tue Oct 21 2003 - 16:03:56 PDT

  • Next message: Russell Power: "Review: Two Theses..."

    Authors: Jon Doyle, Ramesh Patil

    The fundamental theme of the paper is the tradeoff between the expressivity
    of a language, and the performance-in terms of efficiency, soundness, and
    completeness-that it can guarantee. The authors explore this theme in the
    context of an argument favoring expressivity over guarantees.

    The first part of the argument is that KL-ONE-derived systems are too
    restrictive to be general-purpose. The authors present various examples of
    situations in which this is the case. While none of their examples are
    surprises, an encompassing list of the limitations of contemporary systems
    was probably valuable in 1991 as a tool for motivating a change in the
    philosophy underlying the development of these systems.

    The second part of the argument is that a new philosophy emphasizing
    generality over performance guarantees is both justified and beneficial.
    The author's arguments here are pretty common-sense, and while this may not
    have been the first time that such a design philosophy had been proposed, I
    think that the direct feature-by-feature comparison of system was clear and
    powerful (though a bit lengthy!), and probably reached more KR people than a
    stand-alone paper on the virtues of generality-based systems would have.
    It seems that the KL-ONE-derived systems were the standard when this paper
    was published, which makes this presentation of an alternate approach
    valuable.

    I think the biggest flaw in the paper is its relative one-sidedness. The
    authors explain the limitations of one system and propose their own to cope
    with these limitations, but they don't address the obvious fact that the
    system they propose has obvious limits. The paper would have been more
    powerful as a presentation of two ends of a spectrum, instead of an argument
    for the basic superiority of one end over the other.

    Another major problem I had with the paper was its organization; it was long
    and repetitive, and included more detail than was necessary to get the point
    across. I think it could have been more appropriately divided into two
    papers: one discussing the basic tradeoffs of KR system design, and the
    other exploring this tradeoff in a neutral comparison of the two systems
    described.

    Finally, the paper made it clear that we need a way to formalize the
    effectiveness of a KR system. It could be either a single effectiveness
    measurement, or a multi-faceted evaluation that determines effectiveness
    over several axes such as computational efficiency, expressivity, cost to
    the user/developer, etc. Without a concrete method of evaluation, the
    debate in this paper will see no resolution. More importantly, the reality
    is that both extremes of the spectrum have potential applications in the
    real world, and a concrete evaluation method will allow users to choose the
    KR system most appropriate to their situation.


  • Next message: Russell Power: "Review: Two Theses..."

    This archive was generated by hypermail 2.1.6 : Tue Oct 21 2003 - 16:03:37 PDT