[mlpack] Conjugate Gradient and GSoC Proposal Feedback

[Ryan Curtin]

On Mon, Apr 05, 2021 at 01:28:44PM +0700, Wahyu Guntara wrote:
> Hi everybody,
> 
> I already submitted my GSoC proposal draft through GSoC's website. I was
> proposing policy gradient methods. Upon reviewing further the literature
> for my proposed project, I realized that it employs the conjugate gradient
> optimization method. Just to be sure, I was wondering if such a method was
> already implemented. Upon reading ensmallen documentation, I hadn't
> stumbled upon such a method. If it is indeed not yet implemented, may I
> just put the conjugate gradient method local to my project first?
> Additionally, I was wondering whether all submitted drafts or just a select
> few will get feedback?

Hey there Wahyu,

The conjugate gradient method is indeed not implemented in ensmallen.
But, I am not sure it is appropriate---to my understanding, the CG
method is specific to solving a system of linear equations.  So, I am
not sure how it could be implemented as, e.g., a generic optimizer for
differentiable functions.  But, maybe there is something I overlooked?

Also, if you are indeed just aiming to use CG to solve a system of
linear equations, you may be able to rely on Armadillo support instead
(which may or may not use the conjugate gradient method under the
hood---I'm not sure).

If you like, I'd be happy to try and take a look at your proposal in the
next day or two and leave some comments.

Thanks!

Ryan

-- 
Ryan Curtin    | "So, it's just you 57 punks against KUNG FU JOE?"
ryan at ratml.org |   - Kung Fu Joe