User control of gradtol argument
Posted: Mon Feb 28, 2011 4:11 pm
Hi,
I notice that when I fit secr models, the last few hundered iterations of the minimization procedure seem to have all the same LL and parameter values (within the # of digits displayed). I was wondering if I could reduce computation time without sacrificing reliability of the results by increasing gradtol (argument of function nlm) by an order of magnitude or so. I might try fitting the same model with gradtol at the default (1e-6), and at higher values (1e-4 or 5), and checking for differences in the results. Can I simply include the gradtol argument in the call to secr.fit, similarly to stepmax? And, would you recommend against increasing gradtol?
Thanks,
Eric
I notice that when I fit secr models, the last few hundered iterations of the minimization procedure seem to have all the same LL and parameter values (within the # of digits displayed). I was wondering if I could reduce computation time without sacrificing reliability of the results by increasing gradtol (argument of function nlm) by an order of magnitude or so. I might try fitting the same model with gradtol at the default (1e-6), and at higher values (1e-4 or 5), and checking for differences in the results. Can I simply include the gradtol argument in the call to secr.fit, similarly to stepmax? And, would you recommend against increasing gradtol?
Thanks,
Eric