[Bug-gnubg] Bug(?!): Rollouts involving noise

[mailinglists]

Hi folks,

I'm forwarding an e-mail by Bob Koca (he also tried to show me this  
behavior of gnubg whwn we met in Monte Calro):

***************************************************
 Here is a summary of some of the weird behavior I?ve noticed when  
doing rollouts involving noise. I suspect it may just be an issue of  
something being done in a wrong order. I used a 1 point match, zero  
skill position so the weakening of either player and the clicking of  
cubeful on or off should not matter but it does. It is the latest  
build which I just downloaded today.

In particular:

 Comparing 1A to 1C shows that if cubeful is on then weakening one  
player can influence the result.

 Comparing 2C to 2D shows that if cubeful is off then determinsitc or  
nondeterministic noise makes a difference. However this is not the  
case if cubeful is on as can be seen from comparing 1C to 1D.

 On a related issue it would be nice for the rollout results to give  
more data. For example

 Whether the databases were being used, what type of noise was used,  
what MEQ was used. Basically every setting that is alterable seems  
like it should be included in the summary.

Thanks for any help, Bob Koca

Player 1 = bobk Orange checkers on bottom
Player 0 = gnubg White checkers on top

Rollout of 15 on ace vs. 15 on ace 1 point match with various settings:

All use no variance reduction,  1296 trials, quasi random dice,  no  
bearoff database, player 0 on roll,  and seed = 1,000,000

1)  Using CUBEFUL ON
A)  expert  with cubeful eval checked vs  expert with cubeful eval checked
    win% of  .658951   and s.d of .013173
B)  expert with cubeful eval off vs expert with cubeful eval off

same as A

C)  advanced (.015 det noise and cubeful eval)  vs expert
    win % of .592607 and s.d of .011854
***************************************************

Ciao

Achim