Re: [Bug-gnubg] Fwd: Long GNU rollout results

[Joern Thyssen]

On Sat, Jan 04, 2003 at 05:23:08PM +0100, Øystein Johansen wrote
> 
> Overnight I did 6480 games 1ply/cubeful with GNU which will play 
> according to score. It shows that Douglas is correct that this just 
> isn't good enough to cube, even at this score.
> 
> Anyhow..I remain confused about a couple of things with GNU rollouts.
> 
> 1) Can someone, in detail. please explain how the numbers in the first 
> line (cube analysis) are obtained.

The +0.490 is from the Centered 1-cube ([... CF +0.490]). That is, the
equity for no double.

The +0.470 is from the "Player gnubg owns 2-cube) ([...CF +0.470]). The
equity for "double, pass" is trivially +1. 

> 2) Why are there always very slight differences in the rollout details 
> between centered 1 cube and opp owns 2 cube ? This makes absolutely no 
> sence at all. 

There a differences because the two (independent) rollouts are truncated
during the live-cube rollout at different plies following a "double,
pass". Also, since the two rollouts are independent they may very well
lead to different positions due to different gammon prices during play.
However, the do use the same dice in order to reduce the variances
between the two rollouts.

The cubeless equity is calculated as the sum of:

(a) the cubeless percentages at the truncation point
    ("game over" for full rollouts, or at intermediate plies for
     truncated rollouts)
(b) the cubeless percentages for cube decisions that lead to
    "double, pass"
    (for cubeful rollouts)

Analogous for the cubeful equity.

Jørn