[Ipopt] Sparse Hessian approximation

[Andreas Waechter]

Hi Sebastian,

Please reply to the Ipopt mailing list - you get answers that are better 
than mine...

I don't know of any optimization algorithm in C++ that can take do what 
you want.

As far as I know, Lancelot, using partial separability, can do 
structured quasi-Newton updates that makes use of sparsity.  But it is 
written in Fortran, requires a different interface, and uses quite a 
different algorithm.

Regards,

Andreas

On 09/12/2011 07:03 PM, lengagne sebastien wrote:
> Dear Andreas,
>
> Thank you for your answer.
> In fact the hessian has half of the value non zero is the woth cases 
> (when I consider only 100 or 200 variables).
> When I use 3000 variables, I think there is only 10000 or 100000 
> values that are non null.
>
> So I understand that Ipopt does not allow me to do that.
> Maybe you know another algorithm in C++ that can do it.
>
> Best regards
>
> sebastien
>
> Le 13/09/2011 00:30, Andreas Waechter a écrit :
>> Hi Sébastien,
>>
>> Sorry, there is not way to specify a sparsity structure for the 
>> Hessian approximation.  By the way, Ipopt uses the limited-memory 
>> BFGS approximation, i.e., it never stores the full dense Hessian 
>> anyway.  And if half of the entries in the Hessian are zero, this is 
>> not particularly sparse at all!  Usually, you would call a matrix 
>> sparse if it is has only a few nonzeros per row/column.
>>
>> Hope this helps,
>>
>> Andreas
>>
>> On 09/12/2011 04:06 AM, lengagne sebastien wrote:
>>> Hi every body,
>>>
>>> I am using IPOPT with more than 3000 variables and more than 5000 
>>> constraints.
>>>
>>> I cannot compute the Hessian (too long computation time), 
>>> nevertheless I know its structure. It is very sparse (more than the 
>>> halk is set with zero).
>>>
>>> Is it possible to use BFGS method and to specify the non null value, 
>>> that way it could be faster and more efficient
>>>
>>> thank you
>>>
>>> Sébastien
>>> _______________________________________________
>>> Ipopt mailing list
>>> Ipopt at list.coin-or.org
>>> http://list.coin-or.org/mailman/listinfo/ipopt
>>
>> _______________________________________________
>> Ipopt mailing list
>> Ipopt at list.coin-or.org
>> http://list.coin-or.org/mailman/listinfo/ipopt
>
>