Re: Re: [Bug-gnubg] Removal of absolute fibs ratings

[Albert Silver]

Sure, but I think that 0.160 is to tight a window. Another example,
which came to mind as soon as I sent the last note, was on holding
games, or 3-pt anchors. Often the defending side is down by quite a
bit in the race (10-20 pips) and the other side has escaped the
checkers. It can be a No Double well outside of that 0.160 range, and
yet I have seen many Intermediate/Advanced (GNU's rating) players drop
them!

I think the terminology of 'close decisions' is the choking point
here. The point is to determine the range of equity it will use to
consider a cube decision as a genuine decision. The only point where
I'd consider tightening up, or leaving it at 0.160 would be in
non-contact races where the chance of such large cube mistakes is
extremely low. However, for contact positions, I see (and make) so
many blunders (WT, WP, MD, etc.) that to skip over them because they
are larger than 0.160 is a genuine mistake IMHO.

Albert

On 9/5/06, address@hidden
<address@hidden> wrote:
> >I didn't mean it was close with my example. Perhaps I've
> >misunderstood, but I was under the impression that these 'close calls'
> >really meant what range of equity loss it would use to consider a
> >double into the number of decisions it judged a human made for the
> >rating. In other words it would count how many cube decisions the
> >human had to make, and the ones it counted were the 'close decisions'.
> >My point was to show that even if that blunder in my example hadn't
> >been made, I'd prefer it if GNU considered it as one of the decisions
> >made, and not skip over it simply because the size of the equity loss
> >made it not worth considering. It *is* worth considering because a
> >human would likely debate it, and doesn't matter whether GNU thinks it
> >a big mistake.
>
> Close or actual cube decisions = decisions where the equity swing (between
> the right one and the wrong one) is smaller that the threshold.
>
> There's no easy and realiable way to estabilish with an algorithm if
> a cube decision, even if "easy" in terms of equity/error, could be a
> tough one for a human player. Hence two approaches: either you count
> everything (Snowie) either you count only the "close ones" (gnubg).
>
>
> MaX.