"Symbolic Heuristic Search for Factored Markov Decision Processes"

From: Daniel Lowd (lowd_at_cs.washington.edu)
Date: Wed Nov 19 2003 - 02:38:33 PST

  • Next message: Sandra B Fan: "Symbolic Heuristic Search for Factored Markov Decision Processes"

    "Symbolic Heuristic Search for Factored Markov Decision Processes"
    [Feng & Hansen, 2003]

    This paper presented a new algorithm for solving factored MDPs that
    combined the heuristic approach of LAO* with the symbolic gains of SPUDD.

    The main contribution of this paper is an algorithm that may have all the
    advantages of two MDP solution approaches: heuristic search and factoring.
    The authors present in moderate detail how these approaches can be
    combined and present empirical evidence suggesting that it really is a
    good idea. As a second contribution, they provide some interesting
    insight into the factory examples developed for testing SPUDD.

    The weaknesses of this paper lie in the complexity of the algorithm
    explanation and the limitations of the empirical results. In the
    algorithm explanation, a few comments in the pseudocode or more helpful
    variable names (E? F?) might have clarified it some. The empirical
    results suggested that in some cases LAO* works better than SPUDD, in
    other cases SPUDD works better than LAO*, but in all cases Symbolic LAO*
    is competitive with or much better than the both of them. This is a very
    nice result, but the datasets seem somewhat limited -- they admit to
    discovering a lot of structure in the factory dataset, and the new dataset
    they propose is entirely artificial. (Actually, the factory dataset was
    made up, too...) Are there any other factored MDP problems appropriate
    for these algorithms? I'd like to know if these results really hold
    across a variety of types of data.

    Open questions include the possibility of integrating other MDP techniques
    into a still more complicated (and hard to explain in 6 pages) algorithm,
    as well as in what domains various algorithms do best. We've already seen
    from the factory examples vs. the a examples that the dataset structure
    matters a great deal; what other dataset structures are there in the real
    world, and do certain algorithms take advantage of them?


  • Next message: Sandra B Fan: "Symbolic Heuristic Search for Factored Markov Decision Processes"

    This archive was generated by hypermail 2.1.6 : Wed Nov 19 2003 - 02:38:35 PST