Computer Go Olympiad 2016

[Detlef Schmicker]

OK,

we had some nice game, but won none :)

The game*.sgf are autosave files, with some output from oakfoam

Detlef

[Steve Kroon]

Hey Detlef.

Did you identify anything that seems like a systematic weakness in Oakfoam's play based on discussions of the games with others at the Olympiad?

Steve

--
You received this message because you are subscribed to the Google Groups "Oakfoam" group.
To unsubscribe from this group and stop receiving emails from it, send an email to oakfoam+u...@googlegroups.com.
To post to this group, send email to oak...@googlegroups.com.
Visit this group at https://groups.google.com/group/oakfoam.
For more options, visit https://groups.google.com/d/optout.

[Detlef Schmicker]

Hey Steve,

I agreed with Tobias, that we both should enlarge the value network,
as it does not understand large groups.

When playing zen, it is seen, that our playouts do get life and death
much worse.

Others do use smaller prediction networks down the tree, mainly
because of performance I think. But it might make sense, as it plays
moves which help the playout understand life and death better. (a stong network might not play moves if it knows life and death of a group, which the playouts do not understand) I tried
that (I have weak gamma network), but I could not get improvenments...

Make it short, move prediction seems compatible, rest needs
improvenment 

Detlef