Re: [Bug-gnubg] more about cubefull on/off and wierd results using noise in rollouts

[Christian Anthon]

Hi Bob,

the rollout is truncated in two bearoff situations

a) when cubeful only for moneygame and only when the positon is in gnubg_ts
b) when cubeles and the position is in either gnubg_os or gnubg_ts. If the positon is in gnubg_ts the actual lookup is done there.

You might argue that positions with a dead cube should have turned the cubeful option turned off automatically, but it doesn't matter much as you can see and probably makes the interface even more confusing.

So for expert/expert you get the result of a real rollout in the first three columns and the result of a lookup in gnubg_ts in the last two columns in general. The data in gnubg_ts has two byte precision and therefore shouldn't be trusted beyond an accuracy of 
0.0001.

As for the remaining data I don't know and I don't care too much. I can see how the noise might enter the winning chances: if noise is added to the evaluation of game over positions that is if 1.00 is sometimes (for example) 
0.91.

Personally I think that adding noise is a bit of an abomination, which has only small justification in human vs. computer play, and no justification whatsoever when doing rollouts. The data you obtain might be right, but I fail to see what they can be used for even when they are right. Setting GNU backgammon up to play against itself with different settings for the two players makes more sense to me.

Christian.