Mihyun,
I don't know what Hugin does, but Netica's default routine is not obvious,
though it is described in the manual. It is NOT a max. likelihood fill:
* Netica starts all CPTs with a uniform distribution,
and gives that a weight of 1 observation
* Each cell in each node's CPT keeps a record of how many
times it has been seen in the data. (Starting with 1 before
looking at the data, as noted above.)
* So parent-state/child-state combinations which are never
seen will remain at starting state. Others move away
in proportion to the number of observations
* When you learn from observations, you can set a weight.
Setting a weight of say 1000 will swamp the prior and
give you something more like a max-likelihood, except for
combinations you never see.
Netica also has an EM algorithm that fills in missing values.
-C
--
Charles R. Twardy, Res.Fellow, Monash University, School of CSSE
ctwardy at alumni indiana edu +61(3) 9905 5823 (w) 5146 (fax)
