David Poole wrote:
> I have seen this restriction that no discrete child can be a child of a
> continuous variable written down as if it can't be done. Why? A
> disceteization is exactly what a discrete child of a continuous variable
> is. I would expect that making the discretization explicit would be
> advantageous, as, for example, different discretizations may be
> appropriate for different purposes.
The problem is a practial one, rather than theoretical. If you have a
normal parent and a discrete child, then marginalization in some cases
is an integral which can't be written in closed form. As a consequence,
none
of the off the shelf uncertainy propagation software supports such
models. Lauritzen's book has the details
There are, however, approaches to approximate inference. I wouldn't be
surprized
to see them start turning up in UAI papers in the next couple of years.
--Russell Almond
