I'm not so knowledgeable about the ELO system and had a few questions
about how it's used by the CGOS server.
Firstly, on the CGOS server page there's an explanation of how it
works with a table of expected winning percentages vs. difference in
ELO ratings:
http://cgos.boardspace.net/ (towards the bottom)
This table is rather different from another table I found on the
GoBase page about the ELO rating system:
http://gobase.org/studying/articles/elo/
Is there a reason why they are so different?
Secondly, I have let my MCTS reference implementation (with a few
modifications) play on CGOS with different numbers of playouts. What
I'm noticing is that initially the rating changes rapidly, slowing
down over time. This is explained on the CGOS page that it uses a K-
factor that slowly goes down to 3.0 over 200-300 games. What I tend to
see however is that the rating after some 200 games is very much
determined by how well it did early on. After these 200-300 games I
still see a drift towards its actual rating (I assume) for a very long
time afterwards.
Since one of the purposes of CGOS is for a large part so that people
can verify the strength of their program against others, it's
desirable to converge on the actual rating as fast as possible. What
I'm seeing suggest that maybe the K-factor reaches 3.0 too fast. Is
that a reasonable conclusion? I'd be interested to hear opinions about
that.
Mark
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
