[gov2001-l] tolerance in R - Gov2001 - lists.gking.harvard.edu

List overview All Threads
Download

[gov2001-l] tolerance in R

[gov2001-l] Replication:...

[gov2001-l] Spell Checking LaTeX...

sandip.sukhtankar＠gmail.com

2 Apr 2006 2 Apr '06

7:54 a.m.

Does anyone know how to set tolerance in R? The help files and a web search have been fairly unhelpful. We have been having trouble with constrOptim returning local optima unless you feed it initial parameters close to the answers, and wanted to change the tolerance so that R finds it harder to return an optimum. Thanks Sandip & Dan

Reply

Show replies by date

ghumphr＠fas.harvard.edu

2 Apr 2 Apr

8:33 a.m.

Have you considered simulated annealing? Quoting Sandip Sukhtankar <sandip.sukhtankar(a)gmail.com>om>:

Does anyone know how to set tolerance in R? The help files and a web search have been fairly unhelpful. We have been having trouble with constrOptim returning local optima unless you feed it initial parameters close to the answers, and wanted to change the tolerance so that R finds it harder to return an optimum. Thanks Sandip & Dan

Reply

dhopkins＠fas.harvard.edu

11:12 a.m.

Hi Sandip and Dan, These two optim arguments are passed from constrOptim(), and could be helpful--though perhaps you have already seen them. Both are arguments passed to the control list (e.g. control=list(abstol=XXX)). 'abstol' The absolute convergence tolerance. Only useful for non-negative functions, as a tolerance for reaching zero. 'reltol' Relative convergence tolerance. The algorithm stops if it is unable to reduce the value by a factor of 'reltol * (abs(val) + reltol)' at a step. Defaults to 'sqrt(.Machine$double.eps)', typically about '1e-8'. Best, Dan ---- Ph.D. Student Department of Government Harvard University Tutor, Currier House dhopkins(a)fas.harvard.edu http://www.danhopkins.org On Sun, 2 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

Have you considered simulated annealing? Quoting Sandip Sukhtankar <sandip.sukhtankar(a)gmail.com>om>:

Does anyone know how to set tolerance in R? The help files and a web search have been fairly unhelpful. We have been having trouble with constrOptim returning local optima unless you feed it initial parameters close to the answers, and wanted to change the tolerance so that R finds it harder to return an optimum. Thanks Sandip & Dan

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

Reply

a.harrow＠bris.ac.uk

5:30 p.m.

You might also use the maxit option. I was having problems with optim stopping too early until I found it. From the help file: maxit: The maximum number of iterations. Defaults to 100 for the derivative-based methods, and 500 for "Nelder-Mead". For "SANN" maxit gives the total number of function evaluations. There is no other stopping criterion. Defaults to 10000. Also, you can check to see if optim has found a maximum by checking to see if the gradient at that point is zero. If you don't want to write a function for the gradient, then you could sample randomly near the point that optim is telling you is a maximum. so if that point is p, then pick a bunch of random (but very small) x_i and test for linear correlation between log.likelihood(p + x_i) and x_i. -aram On 4/2/06, Dan Hopkins <dhopkins(a)fas.harvard.edu> wrote:

Hi Sandip and Dan, These two optim arguments are passed from constrOptim(), and could be helpful--though perhaps you have already seen them. Both are arguments passed to the control list (e.g. control=list(abstol=XXX)). 'abstol' The absolute convergence tolerance. Only useful for non-negative functions, as a tolerance for reaching zero. 'reltol' Relative convergence tolerance. The algorithm stops if it is unable to reduce the value by a factor of 'reltol * (abs(val) + reltol)' at a step. Defaults to 'sqrt(.Machine$double.eps)', typically about '1e-8'. Best, Dan ---- Ph.D. Student Department of Government Harvard University Tutor, Currier House dhopkins(a)fas.harvard.edu http://www.danhopkins.org On Sun, 2 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

Have you considered simulated annealing? Quoting Sandip Sukhtankar <sandip.sukhtankar(a)gmail.com>om>:

Does anyone know how to set tolerance in R? The help files and a web search have been fairly unhelpful. We have been having trouble with constrOptim returning local optima unless you feed it initial parameters close to the answers, and wanted to change the tolerance so that R finds it harder to return an optimum. Thanks Sandip & Dan

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

Reply

6618

days inactive

6618

days old

gov2001@lists.gking.harvard.edu

Manage subscription

3 comments

4 participants

tags (0)

participants (4)

a.harrow＠bris.ac.uk
dhopkins＠fas.harvard.edu
ghumphr＠fas.harvard.edu
sandip.sukhtankar＠gmail.com