Re: [Bug-gnubg] Double Decisions

[Achim Mueller]

* Peter Carlson <address@hidden> [070822 15:23]:
> I dont really have any precise data.  It's more of a feeling.  Obviously if 
> gnu thinks it is in a losing position it wont offer a 
> double.  I guess what my main point was that once gnu offers, if you press 
> hint, the hint will > 75% of the time - based upon actual 

One of the weird things a neural net like gnubg is that these bots offer
nearly perfect doubles. If you are offered a double and the bot didn't
roll a marketloser before you can almost always be sure it's a close
take, otherwise a pass.

> game playing - advise you to take the double.  Then almost 100% of the time 
> I'll lose, following the hints for every play.  For 
> example:
> 
> gnu offers double
> hint says take, so I take
> gnu rolls and plays
> I roll and press hint and I have *usually* less then 20% chance of winning.  
> I follow each hint for every roll and I end up losing. 
> So again the real question is would gnu offer a double if it didn't think it 
> was in a significantly better position to win.  And if 
> it is in a better position to win, which I have to believe it is, why would 
> the hint offer to take?

If you use hint all the time this is a good way to see how a certain 
position develops. This will help your understanding of the game. That
was the good news. Now to the bad news:

Better do a rollout and let the computer play both sides. If you really
want meaningful numbers you need _at least_ 108 games (from the good old
days where rollouts were performed on a real backgammon board), better
1296. So play a certain position 1296 times, record every result and
then come back.

Ciao

Achim