RE: [Bug-gnubg] gnubg defends poorly against outer primes

[Misja Alma]

There is of course the possibility of having Snowie play a position against
itself many times... Or you have to find some experts wanting to play such a
position and annotate it.
I think the goal should be to only achieve some minimal level of play, so
that gnubg can then learn from its own rollouts. Then there is less risk
that gnubg will learn to make the same mistakes as its teacher.
Of course it would be nicer to let gnubg learn to play these positions well
all the way by itself. But I think that would probably take ages?

Misja

-----Oorspronkelijk bericht-----
Van: Øystein Johansen [mailto:address@hidden
Verzonden: Wednesday, November 26, 2003 9:10 PM
Aan: Misja Alma; gnubg (E-mail)
Onderwerp: Re: [Bug-gnubg] gnubg defends poorly against outer primes


Misja Alma wrote:
> Thanks for the link!
> I see that there might be a possibility that gnubg would not be able to
> learn this class of positions due to the topology of its neural net. That
> would be pretty depressing, so let's assume that gnubg is able to learn it
> ;-)

Yes, let's make this assumption!

> How many positions would gnubg then need to train on to achieve some
minimal
> level of play, that would enable it to learn the rest from its own
rollouts?
> If you are talking about hundreds then it would be feasible I think .. Or
> are you talking about thousands?

I can't make an estimate, but we're definatly talking thousands!

An other aproach would be to collect positions from two experts playing
such a psoition many times. The expert player makes a move, and the
system will estimate the probability of winning based on the fact that
the move she must have a higher winning probability than the other
possible moves in position. In that way we may or may not be able to
estimate good enough data to make reliable rollout results.

I have not thought of an algorithm for generating a database like this,
I'm just thinking (brainstoming) while I type.

-Øystein