While discretization is doable for known distributions, it is problematic for variables whose posterior distributions is not known and may depend on the evidence. One alternative is not to discretize continuous variables, but to approximate continuous densities with mixtures of truncated exponentials (MTE). One can get good approximations with few components and one can propagate MTE potentials exactly using the Shenoy-Shafer architecture. See working paper below for use of MTE potentials in modeling conditional Gaussian distributions.
Cobb, B. R. and P. P. Shenoy, "Inference in Hybrid Bayesian Networks with Mixtures of Truncated Exponentials," Working Paper No. 294, June 2003, School of Business, University of Kansas. Can download pdf version from: <http://lark.cc.ku.edu/~pshenoy/Papers/WP294.pdf> Prakash Shenoy - -- > From: Mi Hyun Park <[EMAIL PROTECTED]> > Date: Thu, 04 Sep 2003 11:27:23 -0700 > To: [EMAIL PROTECTED] > Subject: [UAI] Discretization (or quantization) methods of data > > Dear UAI Collegues, > > I am working with Bayesian networks in image processing. Would you give me > some help on data discretization (or quantization) methods? The methods > I've used were equal interval, equal frequency, and standard deviation for > discretization. It would be very helpful if somebody gives me references > on this matter and other proposed methodologies (or free software for > test). If there is an outstanding or recommended method, please let me > know. > > Thank you, > > Mi-Hyun
