Selecting model complexity in learning problems

To learn (or generalize) from noisy data, one must resist the temptation to pick a model for the underlying process that overfits the data. Many existing techniques solve this problem at the expense of requiring the evaluation of an absolute, a priori measure of each model´s complexity. The authors present a method that does not. Instead, it uses a natural, relative measure of each model´s complexity. This method first creates a pool of “simple” candidate models using part of the data and then selects from among these by using the rest of the data