Review of Two Theses of Knowledge Representation

From: Lucas Kreger-Stickles (lucasks_at_cs.washington.edu)
Date: Wed Oct 22 2003 - 10:58:53 PDT

  • Next message: Bhushan Mandhani: ""Two These" Review"

    Lucas Kreger-Stickles

    Paper: Two Theses of Knowledge Representation / Doyle and Patil

    Summary: Doyle and Patil explore the trade-off between efficiency and
    expressibility in knowledge based systems, arguing that languages should
    not be restricted in the quest for guaranteed polynomial execution but
    rather a general purpose system should allow the user to selectively
    request information within certain constraints of completeness and runtime.

    Most Important Ideas:
    *There is a trade off between efficiency and expressibility and this
    trade off should be left for the user (not the system designer ) to
    decide.
    The authors argue that when the system designer determines this trade
    off it makes so many ideas inexpressible that it defeats the purpose of
    the system. Essentially, the restrictions don't force the programmer to
    say what they were going to say in a better way, but instead make it so
    they can't say it at all.

    *Classification and worst case run time are not the most accurate
    measures of the viability of a system.
    The authors argue that classification is only one operation performed by
    a system and that system performance depends on "the complete mix of
    operations performed". Furthermore they suggest that certain systems do
    not need 100% accuracy nor do they need worst case bounded performance.
    They argue that the success of a system should be determined by its
    utility given the function it must perform and this is the true behavior
    of a 'rational' agent or system.

    Largest Flaws:
    *The authors indicate that for knowledge bases, the decisions about
    which constructs to use should be left to the programmer, arguing that
    "what works best is for programmers to know when they are using a
    dangerous construct." However, they make no attempt to address how
    difficult it is to know when a construct is 'dangerous'. Furthermore,
    they spend so much of their paper arguing that 'dangerous' constructs
    are the most intuitive (or only) way of describing certain information
    that it seems clear that a programmer given the choice could abuse their
    freedom.

    *They don't in anyway address the feasibility of their 'solution' to the
    problem of general purpose knowledge base systems. This is
    particularily troublesome in that they indicate that a systems
    performance depends on all the operations that it performs. Clearly such
    an augmentation would be difficult to encode and would incur significant
    runtime overhead, therefore making it an impractical solution for
    'critical' applications.

    Open Research Questions:
    *Could a system such as the one which the authors propose (a 'librarian'
    type knowledge base) be constructed, which is actually helpful and for
    which the overhead is not prohibitive?

    *What are some other solutions to the trade off between efficiency and
    expressibility?


  • Next message: Bhushan Mandhani: ""Two These" Review"

    This archive was generated by hypermail 2.1.6 : Wed Oct 22 2003 - 10:58:53 PDT