Re: [UAI] how to evaluate approximate algorithms when exact solution is not available?

From: Marek J. Druzdzel (marek@andrew.cmu.edu)
Date: Tue May 08 2001 - 11:32:57 PDT

  • Next message: Nikitas Assimakopoulos: "Special Issue of JASS journal"

    - --On Monday, May 07, 2001, 11:56 AM -0700 Rina Dechter
    <dechter@ics.uci.edu> wrote:

    > The question of evaluating the quality of an approximation to
    > probabilistic inference is really a very important one and a difficult
    > one, with which we also struggle with in Irvine.
    >
    > Sometimes it is possible to generate by an approximation an upper and
    > lower bounds of the exact probability which can bound provide some
    > estimate of the accuracy.
    > However, such methods are often very inaccurate.
    > Therefore one produces arbitrary approximations (no guarantees).
    > In that case, I dont see any other way but to test your approximation
    > algorithm on relatively small or moderate networks and compare against
    > the exact figure (computed by an exact algorithm) with the
    > hope that the information gained from such comparison scales up.
    > You can then try too show superiority relative to other
    > competing approximation algorithms.

    I agree with Rina here. I suggest that you find some networks that are
    large or complex enough for your algorithm to be challenging and yet
    solvable using exact methods. This is the approach that my doctoral
    student, Jian Cheng, and I followed when testing stochastic sampling
    algorithms (e.g., http://www.jair.org/abstracts/cheng00a.html). Unless you
    have a good theory for putting bounds on your posteriors (we have a
    forthcoming UAI paper along these lines; still we test this approach by
    comparing the results to exact answers!), only when you have exact answers
    can you saying anything meaningful about your approximate results.

    > I have seen some people comparing approximation reults to each other
    > (comparing to Gibbs sampling for instance) however such comparisons are
    > meaningless, I think.

    I agree here.

    Marek
    --------------------------------------------------------------------------
    Marek J. Druzdzel http://www.pitt.edu/~druzdzel



    This archive was generated by hypermail 2b29 : Tue May 08 2001 - 11:38:19 PDT