Known bugs and things that needs to be done:
----------------------------------------------------------------------

Not directly Deal-related:

The irritating sound when drawing graphs can be deactivated in Windows
as follows:
	- Open Control Panel and choose "Sounds and Multimedia"
	- Choose "asterisk" and set sound to "none".

It would be nice, if C-W (not buffered output) was default in R.

In genpicfiles, the pictex() does not use colors of nodes.

----------------------------------------------------------------------

When creating a network from a dataframe and arrows are specified so
that the data actually is sparse in the configuration of the discrete
parents, then an error like 
"Error in lsfit(X, y) : 2 responses, but only 3 variables"
or 
"Error in cbind(...) : cannot create a matrix from these types"
occur. (eg. in ksl data with arrows from all nodes pointing to FEV)

The log-score may become positive!  Apparently, it is due to numeric
instability and may be handled by scaling the observations.

The score may breakdown if the img.data base size is not large enough
or if the s2 in the prob-attribute is too small.

print.node: Conditionals are printed as indices of parents and not
	    names. This is not very pretty... 

learn.node: Check if deal.dll is loaded (No! it has to be loaded,
	    otherwise the First.lib fails). If not, switch to the
            (slower) R versions of postcc and postc0c. Or make an
	    option to the deal procedures.

Latent variables: cannot be handled...

Missing observations: cannot be handled...

network: Should include an option 'saturated' that inserts as many
	 arrows as possible. 

drawnetwork, etc: nw need to be learned before calling these
                  procedures. Instead, do a check and learn it, if it
	          is needed. 

learning: We do not need to learn parameters to calculate
          log-lik. Perhaps this may speed up search. So we need to separate
          parameter learning and likelihood-calculation. 

makesimprob: no dependency in discrete variables

jointprior: when N is too small, we get an error for numerical
            reasons. Find out the minimum N that does not cause an error.

learning: names on parameter matrices on cont. nodes.

prior: remove irrelevant parameters that currently are returned as NA.

Memory problems: Apparently the memory used in the C-programs are not
	         released. This is possibly the issue that makes
		 especially heuristic() slow.


