[Bug-gnubg] ratings formula, checker play vs. cube

[Robert-Jan Veldhuizen]

[copy from GOL]

One factor that is also complicating things here, is that player's actual 
(FIBS/ELO) ratings fluctuate a lot even regardless of their skill. I think 
it's pretty common to find players that might be >100 rating points off 
their "real" rating f.i.

Apart from that, some single match that is not really long (say 15pts or 
more) will not always give good data on both player's abilities; you'd need 
more matches for that.

As to GNUBG's ratings, I think the main problem is (still) the fact that 
the total cube error is divided by the number of "actual or close decions", 
while checker play errors get divided by all unforced moves. This causes 
the cube error rate per decision to be a lot higher than the checker play 
error per decision, which in itself causes cube errors to be overvalued.

I think a good rating formula should use the cube error rate per (every 
unforced) decision as long as the same is done with checker play errors.


--
Robert-Jan (Zorba on FIBS)