From: ted.dunning@gmail.com   
      
   Michael wrote:   
      
   >   
   > ... Could the author at Microsoft Research   
   > you refer to be David Heckerman?  ...   
      
   Yes.  David Heckerman's tutorial is the one I was thinking of.   
      
   >   
   > Most of the model utilities I've seen have been along the traditional   
   > BIC/AIC lines; reward the goodness of fit and penalize the complexity.   
   > Of course, this requires you to select a penalty parameter for model   
   > complexity which can be arbitrary.  ...   
      
   I think what you want is something like Mackay's work on the evidence   
   principle.  You can use the Bayesian framework to find the optimal   
   value of the penalty parameters.   
      
   Essentially AIC and BIC are approximations of the posterior probability   
   with all parameters marginalized out.  All that is left are the   
   parameters of the prior distributions (the so-called hyper-parameters)   
   and these correspond to the penalty parameters.   
      
   The big difference between the approaches is that the Bayesian   
   framework gives you two things:   
      
   a) many cases that were previously approximated using normal   
   distributions can be done without the approximations.  This is   
   important for small count situations.   
      
   b) you get a real probability out of the far end which makes   
   comparisons more principled.   
      
   [ comp.ai is moderated ... your article may take a while to appear. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)