Re: [Bug-gnubg] Dice Rolls

[Michael Petch]

On 16/08/2013 7:35 PM, JD wrote:
> Played5 games (me vs gnu-bg).
> Of course, I played the best I know how for both
> myself and gnu-bg.
> I only won 1 game.
> Sounds funny doesn't it?
> I played BG against myself and I lost!!
>
> Well,the fact of the matter is that I was
> astounded by the amazing dice rolls for gnu-bg
> when I had the advantage on the board. Gnu-bg
> got itself oout of multiple jams by dice rolls favorable
> to itself.
>
> Because of what was clear to me as a preferential
> dice rolls for gnu-bgin all 5 games, I decided that
> the dice areskewed in favor of gnu-bg.
>
> Best way to test it is to have actually equally matched
> individuals play 10 or even 100 games (or of course,
> some might say thousands or millions of games) of Human vs Human,
> always making note of the dice rolls for gnu-bg.
>
> As of right now, I am not impressed with the designers of the dice
> roll algorithm.
>

The default dice roll algorithm is actually based on Mersenne Twister.
The source code is available and to date (After 15 years) no one has
found the code that gives preferential treatment to the bot unless you
set that option. GNUBG supports a number of alternative dice generation
algorithms including manual dice (you can roll them with your own dice
and enter manually). There is also dice generated from www.random.org
which provides randomness through real world entropy (and isn't based on
a pseudo random number generator. That can be set via
Settings/Options/Dice . If you want the bot to give itself better rolls
(or yourself) there is a mode for that. No one has found code that makes
the bot look ahead to future rolls to make its decisions.

The neural net plays at a world class level and is better (IMHO) than a
significant majority of the players in the world. Many players also
overestimate their own skill and don't easily accept that the bots play
significantly better. The better someone plays the more lucky they may
appear.

I also did a statistical study (I'll see if I can dig it up) of 100
million rolls and 10's of thousands of matches the bot played against
itself. The results were that the rolls produced were indistinguishable
from what you'd get with properly thrown precision dice in the real world.

If GNUBG outplays a particular human consistently then my guess would be
that the bot is better than that human.

-- 
Michael Petch
GNU Backgammon Maintainer / Developer
OpenPGP FingerPrint=D81C 6A0D 987E 7DA5 3219 6715 466A 2ACE 5CAE 3304