review by Lincoln Ritter

From: Lincoln Ritter (lritter_at_cs.washington.edu)
Date: Wed Nov 05 2003 - 09:09:44 PST

  • Next message: Karthik Gopalratnam: "Review for Paper 3"

    Automatic SAT-Compilation of Planning Problems
    Ernst, et al.
    Reviewed by Lincoln Ritter

    Summary:
    The authors develop a system for evaluating the relative performance
    of various planning encoding schemes and use this tool to discover
    that Regular and Split-fluent explanatory encodings dominate all
    others tested.

    Probably the most important idea present in this paper from a computer
    *science* perspective is the emphasis on experimental testing. It is
    ironic that computer scientists seem prone to argue with religious
    fervor about that which has undergone no experimental evaluation.
    This paper presents a system by which various encoding strategies can
    be compared by translation to a common form and subsequent
    evaluation. In this way, substantive arguments can be made about the
    effectiveness of each method.

    Another striking point of this paper was the discovery that, against
    expectation, bit-wise encodings and regular encodings have the
    smallest and largest number of variables respectively, and that in
    fact regular and simply split explanatory encodings are the smallest
    and fastest. While these results are useful and enlightening in
    their own right, it also emphasizes the point made above about the
    importance of experimental verification. Even though our intuition
    leads us to believe something (say, that bit-wise encodings should be
    smallest), it is not necessarily so.

    While the paper is laudable for bringing experimental verification to
    the forefront, it mentions that WalkSAT based methods with the
    presented compiler might result in the "world's fastest STRIPS-style
    planner" but does not compare the results of this work with other
    STRIPS-style planners. This may be admissible due to the fact that
    the paper focuses on encodings, yet a claim this large needs to be
    backed up.

    To take the experimental methods of this paper forward, methods for
    comparing planners using divers styles need to be developed. For
    example, can some non-STRIPS-style planner compete with the fastest,
    best encoded planner from this paper? Is this an apple vs oranges
    debate? If so, can we develop tools that determine the be style of
    planner to use for a particular problem?


  • Next message: Karthik Gopalratnam: "Review for Paper 3"

    This archive was generated by hypermail 2.1.6 : Wed Nov 05 2003 - 09:09:45 PST