Re: [Bug-gnubg] Neural nets current state

[Joseph Heled]

On 2/21/07, George Stephanopoulos <address@hidden> wrote:
> Hi all , just subscribed.
> I would like to contribute to the development of gnubg. I am
> particularly interested in the
> neural net engine. I would like to know what is the current state of
> affairs regarding the neural nets. How can we improve on the current
> nets? (0.15, i am running the very latest 0.16-mingw-20070219 build
> recently posted)
> I have googled and found an old page by Joseph describing some of the
> techniques he used to train gnubg. But what is the current state? Do
> we have a data set of rollouts or 2-ply evaluations from the 0.15
> weights version?. In general i would like to start the subject on this
> mailing list, i presume there are a lot of ideas but i don't know
> where we are currently and what work has been done already.

I don't know of any significant advancements. The methods I developed
seemed to have reached their limit of usefulness, as more iterations
don't produce a better net. (typically worse).

There is one known weakness of gnubg when it plays (or defends) deep
backgames. This turns to be a very hard problem. I tried several times
to attack this with no real success.

> George
>
>
> _______________________________________________
> Bug-gnubg mailing list
> address@hidden
> http://lists.gnu.org/mailman/listinfo/bug-gnubg