[R] Optim in R

Z P nusbj at hotmail.com
Wed Sep 1 12:13:24 CEST 2004

Dear all,

I recently use the optim function to find the maxima for some likelihood 
function. I have many parameters, more than 12. As the help file mention, 
the default method does not do well in univariate case. How about the 
different method in optim?

I notice the nlm function uses newton method. I have also done the newton 
method by myself, the iteration does not converge due to the difficulty in 
solve the hessian matrix for this large dimension matrix. Is the newton 
method worse than the method supplied in optim? In the mle package, optim is 
used instead of nlm.

For the different methods in optim, they give somewhat quite different 
estimate for some of the parameters. And the most annoying one is that the 
hessian matrix is sometimes not positive definite after the optim gives 
output using method (the default one and BFGS). Shall we admit that we have 
arrived at the maximum point?

For CG and SANN method, they are both very very slow, however, their maximum 
values are usually bigger than the default method.

Is there any suggestion in choosing these methods? I know it is a very hard 
optimazation question, however, I can not find the reference book in optim 
help at the moment. Thank you.



More information about the R-help mailing list