Re: computer-go: Recorded games with dead stones marked

["Nicol N. Schraudolph" <compgo@xxxxxxxxxxxxxxx>]

> We ran across one paper which claimed there are recorded IGS games with 
> the dead stones marked.  This would obviously make point-by-point 
> scoring much easier.  Alas, we can't seem to find any such files.  Any 
> clues?

Back in late 1994 I convinced Tim Casey to record dead stones or territory
(forgot which) in IGS' SGF records; what you're looking for is IGS games
from '95 or '96 (before it went commercial).  I think I have about 75'000
of those, which I'd be happy to share.

I have tried repeatedly to get NNGS to do the same, going as far as
sending in patches against current server versions, twice -- alas
nobody ever seems to have bothered incorporating them into the running
server code.  If any of you know any go server admins, *please* get
them to record dead stones or territory!  It's trivial to do (one line
of server code would suffice) and would make the game records so much
more useful for machine learning.

A word of warning though: a lot of go server games (I guess about 10-15%)
are actually mis-scored.  To get good training data, you have to run the
scored game records through a scoring program, then throw out the ones
where program and game record disagree, say, by more than a single point.

Best wishes,

-- 
    Dr. Nicol N. Schraudolph                 http://n.schraudolph.org/
    Steinwiesstr. 32                         mobile:  +41-76-585-3877
    CH-8032 Zuerich, Switzerland                tel:      -1-251-3661