Previous empirical work in using multiple models (e.g. Buntine, 1990; Kononenko &
Kovacic, 1992) has mainly focused on demonstrating error reduction through using mul - tiple models and exploration of novel methods of generating models and combining their
classifications. The work can be characterized along three dimensions: the kind of model
being learned (tree, rule etc.), the method of generating multiple models, and the method
of combining classifications of the models to produce an overall classification. The work