June,
On reflection, the assignment problem is not
appropriate to your question, because the end result of an
AP is more like a marriage between row and column index
pairs.
I think a better approach is to use hierarchical
clustering but instead of supplying quantitative
characteristics for each entity to be clustered, supplying a
square matrix of pairwise distances or dissimilarities
between entity-pairs to be minimized. In your case, the
dissimilarities would produce a nonnegative, symmetric
square array of subjectively assessed dissimilarity values
with zeros on the diagonal. The clustering algorithm would
suggest clusters (groups, in your case) of various sizes.
The cluster algorithm I placed on the jwiki could be adapted
to accept a square dissimilarity array as input instead of
the distinct data vectors for each entity that it now
expects.
This seems like a better approach than the one I
first recommended, but you will have to judge it.
(B=)
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
