XPost: comp.ai.neural-nets, comp.theory, sci.stat.math   
      
   > Bayesian networks use the same inference tool   
   > of Bayesian methods, so why use a network to   
   > model? Any NEW properties or advantage   
   > could be generated?   
      
   A Bayesian network is just a visualization of a   
   joint probability distribution -- it is a means of   
   focusing on dependencies among variables   
   and ignoring all other details.   
      
   My experience has been that dependencies   
   have more effect on results than choosing   
   one distribution or another. Reasonable choices   
   of distributions tend to yield similar results,   
   while making two variables dependent or   
   independent may have a very broad (and   
   interesting) qualitative effect.   
      
   Unfortunately, changing the dependencies   
   may easily change a tractable problem into   
   an intractable problem; the relevant characteristic   
   is the presence or absence of loops (undirected   
   cycles) in the network. I'll mention that models   
   for temporal problems usually have loops.   
      
   Exact methods for loopy networks are known   
   for a rather limited set of distributions; this is   
   the major limitation of existing Bayesian network   
   software.   
      
   My advice to the original poster is to formulate   
   one or more Bayesian networks of interest,   
   and then look for software which can handle   
   them; don't let the capabilities of a software   
   package steer your project.   
      
   For what it's worth,   
   Robert Dodier   
      
   [ comp.ai is moderated.  To submit, just post and be patient, or if ]   
   [ that fails mail your article to , and ]   
   [ ask your news administrator to fix the problems with your system. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)