Re: [UAI] K2 learning

From: Denver Dash (ddash@sis.pitt.edu)
Date: Fri May 25 2001 - 09:17:29 PDT

  • Next message: simon gay: "[UAI] CHARME 2001 Call For Participation"

    Hi.

    Your best bet is to take the log of the marginal likelihood so you are
    dealing with smaller numbers, and use the gamma function instead of
    factorials. See Heckerman's tutorial
    (http://research.microsoft.com/~heckerman/) for the gamma-function version
    of the Marginal likelihood. A constant-time series solution can be used to
    calculate the log-gamma functions, see Numerical Recipes for details:
    http://www.ulib.org/webRoot/Books/Numerical_Recipes/bookcpdf/c6-1.pdf

    Best,
    Denver.
    -----
    Denver Dash http://www.sis.pitt.edu/~ddash

    ----- Original Message -----
    From: "Estevam Rafael Hruschka Junior" <estevamr@terra.com.br>
    To: <uai@cs.orst.edu>
    Sent: Monday, May 21, 2001 6:33 PM
    Subject: [UAI] K2 learning

    > Hi all,
    >
    > I'm using the K2 algorithm (Cooper & Herskovitz, 1992) to learn a bayesian
    > model from data. The point is that in this algorithm I need to compute the
    > fatorial of Nijk (the number of cases in which a variable xi has the value
    > vik, and the parent of x is instantiated as wij). When the number of cases
    > (in the database) is big, I can't compute this fatorial.
    > I was wondering if anybody have already faced this problem and have any
    > sugestion.
    >
    > Thank you in advance,
    >
    > Estevam.
    >
    > - -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_
    > Estevam Rafael Hruschka Junior
    > Curitiba - PR - Brazil
    > e-mail: estevamr@terra.com.br
    >
    >



    This archive was generated by hypermail 2b29 : Fri May 25 2001 - 09:25:39 PDT