Abstract
We have investigated the performance of a generalisation error predictor, Gest, in the context of error correcting output coding ensembles based on multi-layer perceptrons. An experimental evaluation on benchmark datasets with added classification noise shows that over-fitting can be detected and a comparison is made with the Q measure of ensemble diversity. Each dichotomy associated with a column of an ECOC code matrix is presented with a bootstrap sample of the training set. Gest uses the out-of-bootstrap samples to efficiently estimate the mean column error for the independent test set and hence the test error. This estimate can then be used select a suitable complexity for the base classifiers in the ensemble.