These are chat archives for ujh/iomrascalai

28th
Apr 2015
iopq
@iopq
Apr 28 2015 05:45
You refactored the command line options before merging pull request
Urban Hafner
@ujh
Apr 28 2015 06:09
I’m sorry, I don’t quite understand what you mean.
iopq
@iopq
Apr 28 2015 06:10
Can't automatically merge my pull request now
Urban Hafner
@ujh
Apr 28 2015 06:15
Yes, sorry about that. You’ll have to rebase on the current master or merge it in. It only touches the config stuff so it shouldn’t be too complicated. And that has the added benefit that you can fix the bug in the command line parsing for your new option. I just saw that you’re setting use_empty instead of playout_aftermath when —playout-aftermath is provided.
iopq
@iopq
Apr 28 2015 06:16
oh yeah
Urban Hafner
@ujh
Apr 28 2015 06:19
And sorry for being such a PITA. I just don’t seem to understand why we need the ability to turn that off for CGOS.
iopq
@iopq
Apr 28 2015 06:27
let's say we're playing white
we'd pass when there's a dead black group on the board
because of how CGOS works
we'd lose
in fact, our own scoring algorithm would have us losing
we'd pass anyway because a pass move gives us 100% win rate in playouts (actually for some reason we always get 8 less wins than plays, but that's a bug)
Urban Hafner
@ujh
Apr 28 2015 06:29
But I thought you only add the pass move when the position is a win for the player? Then we shouldn’t have pass moves for positions that would be a loss.
There’s a detail I’m just missing, it seems.
iopq
@iopq
Apr 28 2015 06:49
I also add the pass move when the config option is turned on
Urban Hafner
@ujh
Apr 28 2015 06:51
Ah, yes. You used and or instead of an and. Where’s the head smack emoji when you need it? ;)
iopq
@iopq
Apr 28 2015 06:52
I'll fix the tree reuse thing
Urban Hafner
@ujh
Apr 28 2015 06:52
Great!
So to recap: We get shorter games even with CGOS/Tromp-Taylor rules when we’re ahead as we pass early even then. But with other rules (e.g. chinese) we also pass when we’re clearly ahead.
Unfortunately we can’t turn that on for KGS yet, as we don’t support final_status_list. But that’s something for later.
iopq
@iopq
Apr 28 2015 07:14
well it shows us that we could implement final_status_list by seeing when groups are verifiably dead in playouts
we could also disable the annoying "pass when behind" feature because this makes seki detection more accurate
Urban Hafner
@ujh
Apr 28 2015 07:18
Yes, that would be more reliable. A while ago I even implemented what Petr Baudis suggested. Keep track of which player owned which point after each playout. A point that belongs to one color for >= 90% is considered his territory. If there are opponent stones on points that belong >= 90% to the opponent they are considered dead.
This would also give use “correct” scoring under chinese rules, wouldn’t it?
For some reason (I don’t remember why) I deleted that code again. But it wasn’t that much work.