An embedded and charset-unspecified text was scrubbed... Name: not available Url: https://stat.ethz.ch/pipermail/r-help/attachments/20050527/bed8bd9a/attachment.pl
nlminb to optmin
3 messages · Stefan Pohl, Sundar Dorai-Raj, Douglas Bates
Stefan Pohl wrote:
Hi! I want to convert S-Plus 6.2 code to R 2.1.0. Instead of the function nlminb I use the function optmin optmin(start,fn,gr,method="L-BFGS-B", lower, upper, hess,...) But then I get the Error in optmin ...: L-BFGS-B needs finite values of fn Then I used optmin(start,fn,gr,method="BFGS", hess, ...) But then I get the Error in optmin ...: initial value in vmmin is not finite I know the final parameter estimates from S-Plus which I use as starting values in R. The upper and lower bounds are close around the final estimates. So there is not much to maximize. What can I do? Thank you for help, Peter
What is "optmin"? Do you mean "optim"? Either way, you can always try your function at the initial values outside of optim. If it returns Inf, you have a problem with your objective function. Have you considered that your objective function in S-PLUS didn't port the way you expected to R? I.e. if you call your objective function in S-PLUS and then in R, using the same inputs, do you get identical outputs? --sundar
Stefan Pohl wrote:
Hi! I want to convert S-Plus 6.2 code to R 2.1.0. Instead of the function nlminb I use the function optmin optmin(start,fn,gr,method="L-BFGS-B", lower, upper, hess,...) But then I get the Error in optmin ...: L-BFGS-B needs finite values of fn Then I used optmin(start,fn,gr,method="BFGS", hess, ...) But then I get the Error in optmin ...: initial value in vmmin is not finite I know the final parameter estimates from S-Plus which I use as starting values in R. The upper and lower bounds are close around the final estimates. So there is not much to maximize. What can I do? Thank you for help,
I have a test version of a package available as http://www.stat.wisc.edu/~bates/port_0.1-1.tar.gz that provides nlminb for R. If you can install packages from source code then you may want to try that. If you are running under Windows and only install binary packages then we will need to ask for a volunteer to create a Windows binary from the source package. I do not plan to upload this package to CRAN. Instead I plan to incorporate nlminb into r-devel in time to have it become part of R-2.2.0