Dynamic selection of classifiers: when few are better than all