I am using the caret package with the nnet method in a classification problem, and i would like to know what is the stop criterion used in nnet to stop the training to avoid over fitting.
I have done some research and i found that in the RSNNS package there is no stop criterion implemented, but it is considered for future work, see it here: https://www.jstatsoft.org/index.php/jss/article/view/v046i07/v46i07.pdf
But for the nnet package i didn't find any information about the stop criterion used or if it does use one. So, my question is. What is the stop criterion used in nnet package?
Thanks!!!
The stop criterion used in the nnet package is defined by the
abstol
andreltol
parameters of thennet()
function.(copied from https://cran.r-project.org/web/packages/nnet/nnet.pdf)
Their default values are: abstol = 1.0e-4; reltol = 1.0e-8
The fit criterion is some kind of error between ground truth and the network output. It is part of the objective function for iterative minimization.
There is a question regarding the fit criterion with a partly answer: https://stats.stackexchange.com/questions/156117/what-is-the-value-of-fitting-criterion-on-the-nnet-package-in-r