Re: [UAI] how to evaluate approximate algorithms

From: Bruce D'Ambrosio (dambrosi@cs.orst.edu)
Date: Fri May 11 2001 - 19:14:43 PDT

  • Next message: Bob Welch: "Re: [UAI] how to evaluate approximate algorithms when exact solution is not available?"

    Suggestion:

    Don't be to quick to assume exact solutions aren't available. bucket
    elimination or SPI can often find single posterior marginals when join
    tree methods fail.

    For larger networks, good implementations of bucket or spi that handle
    things like noisy-or well (static parent divorcing has limitations)
    can handle nets much more complex than one might imagine.

    Of course, I don't know of any implementations meeting the above
    criteria besides IET's :)

    Mail me if you have a net in some std format that you would like a
    query run on. Include evidence and a node you want the posterior for.
    No promises, but I'll take a look. I presume we are talking about
    "real" networks. it is trivial, and uninteresting I believe, to
    randomly generate intractable nets.

    Finally, there is always Adnan Darwiche's space-time tradeoff algorithm
    that will compute an exact answer to almost anything if you have
    enough time - although I don't know if there are good implementations (Adnan?),
    and no-one has enough time to compensate for a bad factoring
    (elimination order)

    Bruce
    dambrosi@cs.orst.edu
    dambrosi@iet.com



    This archive was generated by hypermail 2b29 : Fri May 11 2001 - 19:17:31 PDT