Re: [Bug-gnubg] Snowie error rate bug?

[Joern Thyssen]

On Thu, May 15, 2003 at 02:07:05PM +0200, Scott Steiner wrote
> Joern Thyssen wrote:
[snip]
> > 
> > gnubg error rate : -3,658/(139+64) = -0.018
> > snowie error rate: -3,658/(152+160)= -0.0117
> > 
> > You
> > 
> > gnubg error rate : -2,907/(134+53) = -0.016
> > snowie error rate: -2.907/(148+88) = -0,0123
> > 
> > Your opponent had the cube ownership a lot with 160 cube decisions.
> > However, gnubg estimates than 100 of these were trivial (e.g.,
> > scrambling home to save a gammon or trailing a gazillion pips in the
> > race), so even though your opponent gave up more equity than you his
> > snowie error rate is still lower due to 100 trivial cube decisions.
> 
> I understand.  So basically we have here a clash of 2 different
> philosophies here.  Now the inevitable question: Which would you say is
> the more significant and meaningful method?  I, as a layman on this
> topic, find without giving it too much thought that gnubg's method is
> better, but I might be biased considering the above stats ;-)

Both methods have their merits.

The advantage of the formluae used by Snowie is that most people know
the scale. For example, if somebody says that "oh, he played 3" the most
people know that he played very well. If you say "oh, he played 5 on
gnubg", most people will reply: "that does that correspond to in
snowie?" :-)

Snowie's formulae also have the nice feature that the overall error rate
is the sum of the chequer play and cube decision error rates.


gnubg's definition seems to be most "fair" in the sense that only
non-trivial decisions are counted.

Henrik Bukkjær also pointed out that you can actually lower you gnubg
cube decision error rate by making small cube errors.

Jørn