[R] Strating value problem for CG in optim()

[Ruben]

Hi,

It isn't surprising that convergence results depend on initial par 
values in a nonlinear optimization problem. There might be local maxima 
in the landscape defined by your likelihood function and you have to 
explore it for the proper starting places.

One suggestion is to try many different sets of initial values and 
select the best ones by, for each convergence result, examining the 
value of the maximized function, the gradients of the parameters at the 
maximized value, the standard errors of the parameters, the correlation 
matrix between parameter estimates.

Also, if you use optimx (from package optimx) as your wrapper, you can 
check the Karush-Kuhn-Tucker (KKT) conditions, and try several numerical 
methods (not just CG) in one go.

HTH

Ruben

On 6/28/2012 12:54 PM, nataraj at orchidpharma.com wrote:
>
> Dear list-members,
>
> I have done optimization of 3 parameters by maximum likelihood method using conjugate gradient as optimizer.
Since I have the reported value of the parameters from an article, I can 
validate the result of the optimized
parameters. The problem is that optimizer converges to the desirable 
value. (as reported in the article) only
for a certain starting value.
>
> If the staring value of the parameters are like (2,1,1) then the value of parameters the function converges is
> $par
> [1] -0.4169408 -0.2800828  2.9614670
>
> And the value is close to the value of article reported in the article
> Vs = 0.4861 ; Vn = 0.1478 and m is some positive value (no value mentioned in the article)
>
> And If I fine tune the starting value from (2,1,1) to (1.949,1.13,1) then I get the value of the parameters
very close to the one reported.
>
> $par
> [1] -0.4700892 -0.1428245  2.9614670
>
> Now the question is how I can find the starting values for the test experiment for which I am going to implement
this optimization procedure.
> Is there any function to be wrapped with optim() to find the right starting value.
>
>
>
> Regards,
> B.Nataraj
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>

-- 
Ruben H. Roa-Ureta, Ph. D.
Senior Scientist
Marine Studies Section, Center for Environment and Water,
Research Institute, King Fahd University of Petroleum and Minerals,
KFUPM Box 1927, Dhahran 31261, Saudi Arabia
Office Phone : 966-3-860-7850
Cellular Phone : 966-5-61151014

Save a tree. Don't print this e-mail unless it's really necessary