[UAI] New Article, Parameter Learning of Logic...

From: Steve Minton (Steve.Minton@fetch.com)
Date: Wed Dec 19 2001 - 19:44:27 PST

  • Next message: Colette Faucher: "[UAI] Invitation to organise and chair a session at KES 2002"

    JAIR is pleased to announce the publication of the following article, which
    readers of this newsgroup may find relevant:

    Sato, T. and Kameya, Y. (2001)
      "Parameter Learning of Logic Programs for Symbolic-Statistical Modeling",
       Volume 15, pages 391-454.

       Available in PDF, PostScript and compressed PostScript.
       For quick access via your WWW browser, use this URL:
         http://www.jair.org/abstracts/sato01a.html
       More detailed instructions are below.

       Abstract: We propose a logical/mathematical framework for statistical
       parameter learning of parameterized logic programs, i.e. definite
       clause programs containing probabilistic facts with a parameterized
       distribution. It extends the traditional least Herbrand model
       semantics in logic programming to distribution semantics, possible
       world semantics with a probability distribution which is
       unconditionally applicable to arbitrary logic programs including ones
       for HMMs, PCFGs and Bayesian networks.

       We also propose a new EM algorithm, the graphical EM algorithm, that
       runs for a class of parameterized logic programs representing
       sequential decision processes where each decision is exclusive and
       independent. It runs on a new data structure called support graphs
       describing the logical relationship between observations and their
       explanations, and learns parameters by computing inside and outside
       probability generalized for logic programs.

       The complexity analysis shows that when combined with OLDT search for
       all explanations for observations, the graphical EM algorithm, despite
       its generality, has the same time complexity as existing EM
       algorithms, i.e. the Baum-Welch algorithm for HMMs, the Inside-Outside
       algorithm for PCFGs, and the one for singly connected Bayesian
       networks that have been developed independently in each research
       field. Learning experiments with PCFGs using two corpora of moderate
       size indicate that the graphical EM algorithm can significantly
       outperform the Inside-Outside algorithm.

    The article is available via:
       
     -- comp.ai.jair.papers (also see comp.ai.jair.announce)

     -- World Wide Web: The URL for our World Wide Web server is
           http://www.jair.org/
        For direct access to this article and related files try:
           http://www.jair.org/abstracts/sato01a.html

     -- Anonymous FTP from Carnegie-Mellon University (USA):
            ftp://ftp.cs.cmu.edu/project/jair/volume15/sato01a.ps
        The compressed PostScript file is named sato01a.ps.Z (323K)

    For more information about JAIR, visit our WWW or FTP sites, or
    contact jair-ed@isi.edu

    Steven Minton
    CTO, Fetch Technologies



    This archive was generated by hypermail 2b29 : Wed Dec 19 2001 - 19:48:50 PST