These are chat archives for ujh/iomrascalai

9th
Apr 2015
Urban Hafner
@ujh
Apr 09 2015 07:25
I haven’t check, but probably. I had to move the computer around a lot yesterday and I wasn’t able to wait for the games to finish every time.
Urban Hafner
@ujh
Apr 09 2015 14:31
The new no-self-atari playouts are actually doing quite well. A few more games against the various Aya variations would be nice though to confirm that it’s as good as the light playouts from 0.2.0.
iopq
@iopq
Apr 09 2015 15:16
we won vs. Aya on time, we were actually losing in that game, needs more testing
Urban Hafner
@ujh
Apr 09 2015 15:32
I guess it’s hard to tell what the best way to use this is, anyway. Have you looked at the Michi source code? I think Petr uses the self atari check in a different way.
iopq
@iopq
Apr 09 2015 15:37
I did, he actually does check for self-atari in playouts (and also checks if it's a two stone liberty group at the edge of the board and prohibits that too)
and he also checks for it in the tree but does a full 2 liberty solver
Urban Hafner
@ujh
Apr 09 2015 15:51
I see. So you actually know the Michi code better than me. :)
Urban Hafner
@ujh
Apr 09 2015 16:07
For 0.3.0 I was thinking of seeing how far we could get with UCT with some simple priors.
iopq
@iopq
Apr 09 2015 16:36
here's a prior: if the ladder solver says this move will get us captured then it's bad