These are chat archives for ujh/iomrascalai

6th
Apr 2015
iopq
@iopq
Apr 06 2015 04:59
note that the light playouts are still default
iopq
@iopq
Apr 06 2015 06:33
ah, you changed it
iopq
@iopq
Apr 06 2015 12:34
I have a new pull request upgrading from old_io
iopq
@iopq
Apr 06 2015 13:52
You lost on time vs. AmigoGTP
Urban Hafner
@ujh
Apr 06 2015 14:08
Yes, I noticed that, too. I'm not near my computer so I can't check but I believe it's due to my internet connection. Sometimes it gets severed and when the CGOS client tries to reconnect it can't because the server thinks that a second client is trying to connect.
iopq
@iopq
Apr 06 2015 14:10
I checked the game
Iomrascalai was winning, but ran out of time while filling its own liberties
Urban Hafner
@ujh
Apr 06 2015 14:17
Yes, that an obvious missing feature. Especially when playing against humans it would be nice to pass early.
iopq
@iopq
Apr 06 2015 14:18
well also I have some tests failing from time to time related to timer
I added a comment about it
Urban Hafner
@ujh
Apr 06 2015 14:20
One way to do it seems to be to keep track of who owned each point at the end of each playout. Apparently if a point is owned 90% or more by one color it's safe to assume that the ownership won't change anymore. This could then be used to calculate a "real" score. I have no idea however how many playouts need to be run for this to work. But maybe we could give it a try some time.
I saw that comment, yes. The Windows version seems to be more prone to those things, it seems.
iopq
@iopq
Apr 06 2015 14:22
I'm not sure if that is related, though
threading code is always hard to debug or even understand
Urban Hafner
@ujh
Apr 06 2015 15:08
The comment made on Github was a misunderstanding. Somehow I thought these tests were for the chain code and not only for your new methods. Obviously that wasn't the case.
iopq
@iopq
Apr 06 2015 15:09
Ah, gotcha
Urban Hafner
@ujh
Apr 06 2015 15:09
As for the loss to AmiGo. I checked and the CGOS client crashed. The code I have in place to restart the client only runs once a minute and at that point the game was already lost.
iopq
@iopq
Apr 06 2015 15:11
oh, I see, I checked the wrong game
wait, the sgf said we only used 1:33 for our first moves
maybe the restarting doesn't work
iopq
@iopq
Apr 06 2015 16:11
seems that we do badly with the no self atari playouts because of the slow down, even though in my tests (over ~50 games) we did OK
I'll make that playout type faster
iopq
@iopq
Apr 06 2015 17:18
I found a small bug
    if game.winner() == color {
        sender.send(Pass(color)).unwrap();
this makes us pass when we're losing but have dead stones on the board