Christian Anthon wrote:
> Hi Jon,
>
> I did a bit of investigation. I think the problem is in the multi
> thread loop. The number of games isn't always correct (off by 1) and
> that hits both the average and the std err in the beginning of a
> rollout, but only the std. err. is affected in the long run because it
> is calculated recursively.
It's possible - working out how many games have been played is difficult (when
there are multiple threads working). What seems strange is that it only effects
cube decision rollouts, or is that not the case?
Jon