[mlpack] GSoC Proposal

[Ryan Curtin]

On Thu, Apr 04, 2019 at 06:34:17PM +0300, Дворкин Игорь wrote:
>    Hello!
>     
>    My name is Igor, I'm a student at Moscow State University. It's basically
>    my first time trying to work on an open-source project and I might be
>    asking some dumb questions. I'm very interested in the idea of
>    implementing NEAT algorithm, so I wrote a proposal and I ask you to give
>    some feedback on it
>    ([1]https://docs.google.com/document/d/13O_bMCSO1UhqN8kl415pKAu28h5YtolxjKu6PbGxlLQ/edit?usp=sharing).
>    But at the same time a have a concern about whole project.  

Hey Igor,

Thanks for getting in touch.  I took a quick look at the proposal; I
can't leave comments directly on it, but I might suggest spending some
more time thinking about the testing.  You're right that MarI/O would be
a nice test to have, but from the perspective of unit testing it may be
good to focus on simpler situations that we can run very quickly to
hopefully help ensure that NEAT is working right.  This also means
structuring the NEAT code so that it is easy to test.

>    I have found a little problem. I was looking through the optimization API
>    and it doesn't really add up with how I understand that NEAT works. As I
>    understand, NEAT only optimizes object functions which take Neural
>    Networks as inputs (and finds the Network, which gives best result on that
>    function). This doesn't correspond with optimization API, as it doesn't
>    seem to make any limitations on function input. Wouldn't this issue make
>    it impossible to implement NEAT with those requirements?

We talked about this in IRC too but just for the sake of the rest of the
list, the outcome of the discussion was basically that NEAT may not fit
nicely into ensmallen because it is specific to neural networks and thus
we may be best off keeping it in mlpack separately.

However it may be interesting to think about how it might fit into
ensmallen without changing any of the existing FunctionTypes; but I am
not sure how feasible that is.

Hope this helps.

Thanks!

Ryan

-- 
Ryan Curtin    | "I know... but I really liked those ones."
ryan at ratml.org |   - Vincent