The concept of measure functions for generalization performance is
suggested. This concept provides an alternative way of selecting and
evaluating learned classifiers, and it allows us to define the learning
problem as a computational problem.
Preprint (postscript)
Preprint (pdf)