Hi,
On Sat, Mar 08, 2008 at 10:18:34AM +0100, Petr Baudis wrote:
> (By the way, pachi1-*-light are UCT bots with completely light
> playouts with various UCB1 c values, if anyone wants to use that as
> reference. Surprisingly, it seems that my heavy playouts do not make big
> difference so far, though the rating is still very unstable.)
after two days of play, it seems the ratings are fairly settled now.
For clarity, here is the UCB1 formula I use:
UCB1 = X_i + sqrt(log(N) * c / n)
Specifically, the c is withing the sqrt(); some of the papers put it in
front of the sqrt.
Also, I expand UCT leaves at the second hit. This retains conservative
memory usage but it is important for strength - I saw huge strength
increase when I lowered this to 2 from the original value of 5.
With 110k playouts per move and no domain knowledge in the playouts,
the ratings are now:
c=0.2 (pachi1-p0.2-light) ELO 1627 (285 games)
c=1.0 (pachi1-p1.0-light) ELO 1590 (120 games)
c=0.05 (pachi1-p0.05-light) ELO 1531 (286 games)
c=2.0 (pachi1-p2.0-light) ELO 1511 (118 games)
The main two messages of this post are: If you are developing own UCT
bot, with this number of playouts you should be aiming at least at 1600
ELO on CGOS. And choosing the right c can easily make a 100 ELO
difference! In particular, the "default" UCB1 c=2.0 appears to be very
unsuitable choice.
I'm pretty sure my code is fairly well debugged now, but of course
there may be still bugs lurking; when I have put my bots on CGOS for the
first time it was awfully bug-ridden (and about 800 ELO worse ;-). What
ELO rating did pure UCT bots get historically with how many playouts?
P.S.: Looks like the heavy playouts I described in my other mail bring
no improvement to the bot strength at all, and mostly make it few ELO
weaker. :-( I'm rethinking my approaches now.
--
Petr "Pasky" Baudis
Whatever you can do, or dream you can, begin it.
Boldness has genius, power, and magic in it. -- J. W. von Goethe
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
