RE: Re: [Bug-gnubg] small suggestion and questions

[Ian Shaw]

From: Christian Anthon on 24 May 2006 12:33

        On 5/24/06, address@hidden <
address@hidden
<mailto:address@hidden> > wrote: 


                the problem in gnubg's severe cube-rating may be related
to the 
                computation of the actual/close number of cube
decisions. If the
                error is above very bad (something like 0.160 I think,
it's
                hard-coded somewhere in the code source) the cube
decision is not
                considered actual/close (hence the error per
actual/close decision 
                rate will be higher)... maybe the threshold should be
higher ...


        Correct, but adjusting the threshold because you don't like the
size of the resulting error-rate makes no sense to me.
        
In Snowie, all cube decisions are considered, no matter how trivial.
Therefore its threshold is infinite (or perhaps 3.0). Gnubg counts only
cube decisions which it thinks are close, so the error rate per cube
decision is higher, because the divisor is lower. The only way to make
the cube stats relate to Snowie's is to do it as per Snowie.

There is a similar effect on the chequer plays, where gnubg considers
only unforced plays, while Snowie counts all plays in the divisor.

If it's distressing enough people, perhaps gnubg should acknowledge that
Snowie is the de facto standard and go with that. (I consider it
unlikely that Snowie group will be persuaded that gnubg's method is
better, and people have already got a good "feel" for the Snowie
ratings.)

I'm uncomfortable with applying some empirically derived fudge-factor to
make things look similar, when there is a proper mathematical approach
that can be adopted.

-- Ian