Comparing reliability of training error and bootstrap error for least square fitting on noisy data
Nur Soffiah Sahubar Ali, Ahmad Lutfi Amri Ramli, Adila Aida Azahar, Nuzlinda Abdul Rahman · AIP conference proceedings · 2017
In least square fitting (LSF), increasing the degree of fitting reduces the training error but may lead to overfitting. Therefore, we cannot rely on training error as visual evaluation may be difficult in some cases. Bootstrap error estimation method on 2 dimensional data is applied for LSF. Simulated data is generated with some added noise. In this paper we compare the training error and bootstrap error for different degree of LSF on selected data. We observe that the best polynomial fitting is based on the smallest bootstrap error obtained and the result agree with the original ground truth data.