Re: [UAI] Definition of Bayesian network

From: David Poole (poole@cs.ubc.ca)
Date: Sat Jul 21 2001 - 11:21:57 PDT

  • Next message: Richard Zemel: "[UAI] NIPS*2001 web-site back online"

    Seeing no one else has replied, I thought I would put in my 2 cents
    worth.

    > My reason for defining it the way I did is that I feel `causal' networks
    > exist in nature without anyone specifying conditional probability
    > distributions. We identify them by noting that the conditional
    > independencies exist, not by seeing if the joint is the product of the
    > conditionals. So to me the conditional independencies are the more basic
    > concept.

    A belief network (Bayesian network) is *model*. What dependencies exist
    in a model may be different from the dependencies in the world. The
    simplest example is the Markov chain which is a simple belief network.
    It is arguable that the world is Markovian; the past can only influence
    the future by affecting the present. (Whether this is true or not is
    presumably an empirical fact of physics; I have no idea what the current
    state of thinking of this is in physics. But it is definitely true in
    the common sense world). But this does not mean that our models of the
    world have to be Markovian. We may not be modelling all the variables
    that make the other variables Markovian; thus k-Markovian models are
    often useful and appropriate. At the other extremes if we are predicting
    the outcome of two-up (or matching pennies; where someone wins if two
    coins match their heads) if we are only modeling one of the pennies, who
    wins is independent of the the value of the coin toss we are modelling.

    In response to your query, to specify the model you have to specify the
    variables, the domains, the parents (i.e., the DAG structure) and the
    conditional probabilities. That's all. So I would say that these form
    the belief network.

    But a diversity of ways of explaining the same concepts is definitely
    useful to the community. If you want to know what you should do, I
    suggest that you run some pilot studies testing different explanations
    and seeing what explanations are the most natural to the people who are
    the target audience for your book. I, for one, would be interested in
    the results of such experiments.

    David

    -- 
    David Poole,                      Office: +1 (604) 822-6254
    Professor,                        Fax:    +1 (604) 822-5485
    Department of Computer Science,   poole@cs.ubc.ca
    University of British Columbia,   http://www.cs.ubc.ca/spider/poole
    Vancouver, B.C., Canada V6T 1Z4   ftp://ftp.cs.ubc.ca/ftp/local/poole
    



    This archive was generated by hypermail 2b29 : Sat Jul 21 2001 - 11:23:18 PDT