[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[gnugo-devel] 9x9 openings

From: Douglas Ridgway
Subject: [gnugo-devel] 9x9 openings
Date: Wed, 5 May 2004 16:10:30 -0600 (MDT)

I've been looking at 9x9 openings. Using extract_fuseki, I built an
alternative pattern db from a ~300 game set of pro-pro 9x9 games, and
tried it against the usual one.  It looks like it might be slightly
stronger as white (winning 32/50 with 6.5 komi, mean score -3.4 +- 1.4),
and is about the same as black (winning 26/50 giving 6.5, mean score 0.0
+- 1.4). I haven't tossed out repeats (~20%), so there could be bias. Both
are otherwise 3.5.6+recent CVS.

A few questions. Could someone explain the weighting of random moves?  
It seems to me that if two moves are equally good, then they should be
played equally often, to be least predictable. If one move is worse than
another, either objectively or because GnuGo can't exploit it properly,
then GnuGo will be stronger if it never gets played at all. This seems to
argue for equal weighting of indistingushable moves. Or am I missing

Is there an easy way to merge patterns? That way we could find the novel
ones from this db and add them. And are there any tools for collecting 
statistics on pattern use or results? 



PS. Did my regressions get rejected or forgotten?

reply via email to

[Prev in Thread] Current Thread [Next in Thread]