[R] constrOptim with method SANN

[Jonas Rumpf]

Hi,

thanks again for the clarification.
However, then the documentation of constrOptim is
somehow misleading, because it says

"Any optim method that permits infinite values for the
objective function may be used (currently all but "L-BFGS-B")."

But I assume that this (documentation issues) is not something
that is supposed to be discussed on this particular mailing list (or is 
it, I'm new here?).

In addition, it appears strange to me that constrOptim doesn't issue any 
warning
or error when called with method="SANN", (except when the grad argument 
is missing) ?!

Thanks again,
Best regards,
Jonas

Thomas Lumley schrieb:
> On Fri, 18 Jan 2008, Jonas Rumpf wrote:
>
>> Hi,
>>
>> thank you for your help, Thomas,
>> (and my apologies to everyone for the double-post earlier),
>> however I'm not sure if I understand you correctly:
>>
>> Are you basically saying that it doesn't really make sense
>> to use SANN with constrOptim and I should use another
>> algorithm (like CG etc.) ?
>
>
> Well, you should use some other algorithm because constrOptim doesn't 
> support SANN, as you already had found out.  One reason it doesn't 
> support SANN is that it doesn't really make sense to.
>
>       -thomas
>
>
>
>> Best regards,
>> Jonas
>>
>>> On Fri, 18 Jan 2008, Jonas Rumpf wrote:
>>>
>>>  
>>>> Hi Everyone,
>>>>
>>>> I'm trying to minimize a function using constrOptim with
>>>> the simulated annealing method SANN.
>>>>
>>>> If I understand constrOptim well, it basically passes most
>>>> of its arguments to optim while somehow enforcing the constraints.
>>>>
>>>> My problem is, that since SANN does not need gradients,
>>>> when using optim with SANN, the gr argument of optim is
>>>> used to specify a function to create the next candidate point
>>>> for the annealing algorithm. If it is left NULL, a default Gaussian
>>>> Markov kernel is used - which is fine for my purposes.
>>>>
>>>>     
>>>
>>>   
>>
>>> SANN doesn't need gradients, but constrOptim does, and if you have 
>>> the sort of function for which simulated annealing is useful it is 
>>> unlikely to be the sort of function where the adaptive log-barrier 
>>> method works well.
>>>   
>>
>>
>>
>>>      -thomas
>>>
>>> Thomas Lumley            Assoc. Professor, Biostatistics
>>> tlumley at u.washington.edu    University of Washington, Seattle
>>>
>>>   
>>
>>
>
> Thomas Lumley            Assoc. Professor, Biostatistics
> tlumley at u.washington.edu    University of Washington, Seattle
>
>