[Ipopt] Stochastic lbfgs

[Stefan Vigerske]

Hi,

the silence may mean that there are no plans.

There is an implementation of L-BFGS for Hessian approximation in the 
class LimMemQuasiNewtonUpdater: 
https://coin-or.github.io/Ipopt/classIpopt_1_1LimMemQuasiNewtonUpdater.html
This is derived from HessianUpdater 
(https://coin-or.github.io/Ipopt/classIpopt_1_1HessianUpdater.html), so 
if one were to implement another Hessian approximation, then this should 
be derived from HessianUpdater, too, and 
AlgorithmBuilder::BuildHessianUpdater 
(https://coin-or.github.io/Ipopt/classIpopt_1_1AlgorithmBuilder.html#a4646cc682cbb462505b2dafd3971ef2d) 
should be extended to possibly instantiate the new class.

Stefan


On 8/29/20 9:35 PM, Panos Lambrianides wrote:
> Hi Everyone,
> 
> I wanted to ask if there were any plans to expand ipopt to include
> stochastic lbfgs variants such as these
> https://arxiv.org/pdf/1802.05374.pdf
> 
> How much work do you think it would be to do so? It seems to me that
> writing an optimizer from scratch that incorporates these new advances into
> interior point methods is not the best approach.
> 
> 
> Best,
> -Panos Lambrianides
> Department of Applied Mathematics UC Santa Cruz
> 
> 
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> https://list.coin-or.org/mailman/listinfo/ipopt
>