Re: [computer-go] GnuGo self testing results

[Gunnar Farnebäck <gunnar@xxxxxxxxxxxxxx>]

Don wrote:
> So  I  have been  self-play  testing  gnugo  3.3.21 at  various  level
> settings with my autotester to make sure it was bug-free at 19x19.

Maybe not the ideal choice of version (included some fresh
experimental code which was heavily revised in subsequent development
versions). I would recommend 3.4 or the latest development version
(currently 3.5.4) for experiments.

> Strangely, the default  level 10 seems to play  a slightly weaker game
> than some  of the lower levels,  but the others are  almost in perfect
> order.

Surprising. I have no good explanation for this.

> This of course could change  when more games are played and we
> have a larger and probably more representative sample.  There does not
> appear to be  a huge difference in levels,  the difference between the
> lowest and highest level is approx 200 ELO rating points.

I might comment that the levels below 10 are primarily intended to get
higher speed. The decrease in strength is just a secondary effect. I
think it's also somewhat likely that lower levels lose more in
strength against human opponents than in selfplay. Generally speaking
selfplay tends to be relatively peaceful for GNU Go.

> Apparently, I  set the  komi too low  as black  is winning 57%  of the
> games.   I'm using 5.5 Komi

This is consistent with other observations. I don't know what komi is
needed to get even results though.

/Gunnar
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go