Re: [Bug-gnubg] User training of the Neural Nets

[Øystein Johansen]

Albert Silver wrote:
> I've been trying to get information on this, by looking up the archives, but
> with limited luck, so I'll just ask the experts:
> 
> 1) Can a user do their own neural net training without necessarily being a
> full-fledged programmer? I'm presuming/hoping they can.

The simple answer is unfortunately "no". However, if you want to train
GNU Backgammon without much programming experience, you should get the
gnubg-nn code from the repository. (I remember I once compiled this for
Windows, and it worked.) However, training with these tools alone will
not make you able to alter the input calculations.

Don't rely on the training algorithms in the main program. It's based on
TD training, and that was abandoned many years ago. Also, I don't think
that training even gets the contact/crash split.

> 2) Presuming it is possible, what are the limitations? Meaning what would
> they be able to do, and not be able to do.

Without any programming experience at all there isn't much you can do.
Let me guess that a _good_ computer user (but w/o programming
experience) can fetch the gnubg-nn code from the repository and compile
it and maybe use it. However, you _must_ read the code to understand the
output of those tools.

> 3) Can one train via both standard TD method and providing one's own
> training positions? If so, how? 

Hmmmm.... I guess you can.... You have to hack some code, but I wouldn't
have used TD at all.

> I realize that even if one is able to, there will always be the question of
> whether the result is worse or not, but I'd still be quite curious.

There is no problem breeding a new neural net. That's really simple, I
guess I can do it in 15 minutes or so. However, will it be better?
Compared to the reference databases, I've not been able to make anything
better.

-Øystein

signature.asc
Description: OpenPGP digital signature