[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

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

From: Misja Alma
Subject: RE: [Bug-gnubg] gnubg defends poorly against outer primes
Date: Wed, 26 Nov 2003 20:50:25 +0100

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
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?


-----Oorspronkelijk bericht-----
Van: address@hidden
[mailto:address@hidden Øystein
Verzonden: Wednesday, November 26, 2003 8:25 PM
Aan: Misja Alma; gnubg (E-mail)
Onderwerp: Re: [Bug-gnubg] gnubg defends poorly against outer primes

Misja Alma wrote:
> I have a few questions about the training of gnubg. Perhaps one of you can
> help me out?
> In reply to my mail about gnubg's poor handling of outer primes, Øystein
> wrote that it was extremely hard to train those positions. Why is that?
> I have read the interesting article 'the gnubg training program' on the
> internet, and I know some about neural nets myself, but I still don't
> understand? If the problem is in the evaluating of benchmarks because
> gnubg's rollouts are biased, why not let them be rolled out by Snowie?
> seems to handle them pretty well.

Hi again!

I'm sending you a link a mail from Joseph on this issue.

The post is about containment positions, but the same applies to outside
prime positions.

Using Snowie to roll out positions have some major drawbacks. First of
all the platform it's running on. Then it's not possible to automate the
rollout process, since the rollouts must be performed from the GUI. It
will be a hell of a work to make a program for doing these automatically
from a program, and it will be a lot of work to set up all the positions
manually since it will be an enourmous amount of positions. Then: How do
we know that Snowie plays this correct. What if we train gnubg to make
the same mistakes as Shitie^H^H^H^H^Hnowie? Then again I don't think
many of the developers own a copy of Snowie so the availability for
doing rollouts with Snowie won't be that high. I also think it will have
a tast of shame if we use this approach.

I rather think an expert system trained network can work as a starting
point. This will also be _very_ slow process, but I think it will be
considerably better than than using Snowie.

> My other question is about the fact that gnubg uses three different neural
> nets; the crashed net, the race net and the contact net. I can see how
> knows when to use the race net, but how exactly does it know when to use
> crashed net and when the contact net? I could imagine there are some
> borderline positions there.

It's a CRASHED position if a player have only 6 or less loose checkers
to play, and the remaining checkers are crashed on the acepoint,
duecepoint or borne off. From eval.c line 1925 ff.

     if( nBack + nOppBack > 22 ) {

       /* contact position */

       unsigned int const N = 6;
       unsigned int i;
       unsigned int side;

       for(side = 0; side < 2; ++side) {
         unsigned int tot = 0;

         const int* board = anBoard[side];

         for(i = 0;  i < 25; ++i) {
           tot += board[i];

         if( tot <= N ) {
           return CLASS_CRASHED;
         } else {
           if( board[0] > 1 ) {
             if( (tot - board[0]) <= N ) {
               return CLASS_CRASHED;
             } else {
               if( board[1] > 1 && (1 + tot - (board[0] + board[1])) <=
N ) {
                 return CLASS_CRASHED;
           } else {
             if( ((int)tot - (board[1] - 1)) <= (int)N ) {
               return CLASS_CRASHED;

       return CLASS_CONTACT;
     else {
       /* Check if it's race or bearoff */


Bug-gnubg mailing list

reply via email to

[Prev in Thread] Current Thread [Next in Thread]