Hi Remi,
What komi did you use for 5x5 and 6x6 ?
I used 7.5 komi for both board sizes.
I find it strange that you get only 70 Elo points from supervised
learning over uniform random. Don't you have any feature for atari
extension ? This one alone should improve strength immensely (extend
string in atari because of the previous move).
Actually no. The features are very simple, and know how to capture but
not how to defend ataris. I'm sure that a better set of features could
improve by more than 70 Elo, but I expect we would still see a benefit
to balancing the weights correctly. For example, the Fuego policy
defends ataris and follows several other common-sense rules, but the
results in 5x5 and 6x6 show that it is not well balanced on small
boards.
Let us extrapolate: I got more than 200 Elo points of improvements
from
my patterns in 9x9 over uniform random (I never really measured
that, it
may be even more than 200).
I guess you got more than 200 Elo on 9x9, in Mogo (Gelly et al. 2006)
the improvement from uniform random was at least 400 Elo and I think
your simulation policy is probably at least as strong.
By the way I was sad to hear you're not working on Crazystone any
more. Is it because you are you busy with other projects?
So maybe I could get 600 more Elo points
with your method. And even more on 19x19.
I noticed that, in general, changes in the playout policy have a much
bigger impact on larger boards than on smaller boards.
As to whether we can extrapolate, I hope so :-)
I share the same feeling that improving the simulation policy will be
more impactful on bigger boards with longer simulations.
On the other hand I've been surprised by many things in MC Go, so
nothing is certain until we try it!
-Dave
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
