Re: [gnugo-devel] Recent games

[Daniel Bump]

Arend wrote:

> If abandoning a group is done by a move that creates a new group, in the
> opponent's sphere of influence, then this is an experimental influence
> fault. I am now back from traveling, and I also have rough idea how to
> resolve this problem. (I will try to filter out such "invasion" moves with
> a few heuristics, and assign them a negative strategical value in
> value_moves. That value will depend on our and opponent's influence at the
> point where we want to play. Any suggestions welcome.)
> 
> This patch might take a few days, as it obviously needs quite some tuning.

I didn't think this was what was happening in the GoFun games. 
Look at moves 28 and 34 in gnugo-3.1.27-GoFuN-200203100711.sgf.
W tenukis twice, leaving the group at the bottom to die. In
the first case, it thinks the group is alive, and in the second
case, it thinks it's dead. The tenuki moves are not overvalued,
rather the urgency of defense is not recognized.

This is why I increased the value of uncertain defense move
reasons.

Could you give some examples of the phenomenon you are describing?
I think they may be frequent in games where GNU Go gives a 9 stone
handicap. But generally I think GNU Go is better now at giving
big handicaps than 3.0.0.

Bear in mind that we're approaching the late stage of the
3.2 release. We can continue to tune for a while but at
some point we must stop and freeze the code, fixing only
bugs and portability issues. So it's important to prioritize now.
Especially anything that might make it weaker by accident should be
approached with caution.

About a timetable, my thinking is that the code freeze
would come no earlier than April 1 and more realistically
maybe around April 15. We will want to play a lot of
games before and after that looking for crashes and
build on different platforms.

I don't think there will be a 3.1.28 release until
early next week but there are some pre release tags
in the CVS.

The hashing patch, BTW seems to give about a 5% speedup
in the testing I did. I think you intended to revise that
again, per your comments in
http://mail.gnu.org/pipermail/gnugo-devel/2002-March/001596.html
and a previous message of Gunnar's. I brought it current with the patch
hash_1_28.3 once but it may clash with komaster_1_28.1 which is now in the
CVS. (I haven't checked.)

Dan