MIX-EXTENSIONS:  Possible extensions to the mixture modeling software.

The mixture modeling software presently provides the core inferential
facilities for simple models, but does not do many things that one
would often need to use these models in practice.  Here are some
extensions that would be feasible to add, and which may be added one
day (though I make no guarantees).  They are presented in roughly
increasing order of difficulty of implementation.

1) The concentration parameter of the Dirichlet prior for component
   mixing proportions could be a variable hyperparameter, rather than
   being constant as as present.

2) A program could be written for giving the predictive density at
   a given point (ie, for a given set of target values).

3) Models for problems in which some attributes are binary and some
   are real could be supported.  Attributes taking on values from some
   finite set with more than two elements could also be supported.

4) Models in which the distribution of the target variables depends
   on a set of input variables (eg, via a linear regression model)
   could be supported.

5) Support for missing targets in the training cases (assumed to
   be missing at random) could be provided.  A program could also
   be written that fills in missing targets in test cases, based 
   on the targets that are not missing.

6) The shape parameters in the various priors could be made variable
   hyperparameters.  This may be more important for mixture models
   than for neural network and Gaussian process models, because
   of the effects these shape parameters have on how many mixture 
   components get used (due to "Occam" effects).