Re: computer-go: Results on self play vs recorded games

[Erik van der Werf <E.vanderWerf@xxxxxxxxxxxxx>]

Nicol N. Schraudolph wrote:
> > > Interesting results, I'm curious, how good are the recorded games?
> > > Thinks like average rank of say the weakest 10% of the players would 
> > > be interesting. (I assume these were human players, right?).
> > I got them from Nici Schraudolph, who would have to make any such 
> > comments.
>
> The games were not selected according to player strength, so you get
> whatever IGS (the source) had to offer.  The selection criteria were 9x9,
> no handicap, and recorded territory.  To select only games between good
> players you'd need a larger pool of games, which means going to 19x19 and
> "bootstrapping" final territory, e.g. via gnugo as suggested here.
Thanks for the info!

The games provided here recently by Guido easily contain over 3000 9x9 
games of much higher average rank. I would be interested to see a 
comparison with statistics on such a training set.


> I'd consider these ranks a lower bound on actual player strength, since
> IGS regulars are bound to create an unranked/low-ranked login on the
> side for things such as playing a 9x9, in order to avoid messing up
> their main rating.
That's strange, normally 9x9 games are not used to set the main rating. 
Was this any different in the time that IGS was non-commercial?

Best,
Erik