Packages for Learning Algorithm Independent Branch and Bound for Feature Selection

I am looking for packages that can run a branch-and-bound algorithm to
maximize a distance measure (such as Bhattacharyya or Mahalanobis) on a set
of features.

I would like this to be learning algorithm independent, so that the method
just looks at the features, and selects the subset of a user-defined size
that maximizes a distance criteria such as those stated above.

Can anyone give some suggestions?

Alex Byrley
Graduate Student
Department of Electrical Engineering
235 Davis Hall
(716) 341-1802