[Top][All Lists]

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

[gnugo-devel] edmondton_3_6.1

From: bump
Subject: [gnugo-devel] edmondton_3_6.1
Date: Sat, 3 Aug 2002 10:12:11 -0700

I've put up a patch called edmondton_3_6.1 at:

After this patch, the state of the engine should be
identical to the last four rounds of the tournament.

This patch contains changes which I made before and during the
Edmondton tournament.

In addition to these changes, I compiled with assertions turned
off, probably a good idea for tournaments. I just added a line
by hand to config.h:

#define NDEBUG 1

I increased the value of j patterns because move valuations
seem to be higher in the opening these days.

The change in value_moves is in response to the problem in
century21-2002:30, which as I noted in my e-mail may be
related to the problem reported by Gunnar in
This is a quick fix for this bug but as Gunnar notes this
position probably requires more careful investigation. I
don't have the game gnugo-3.3.2-Programmer-200206021522.sgf
and the nngs game page is not responding at the moment. This
change was added between rounds.

The gtp function experimental_score is related to a 
metamachine experiment.

The splitting of EB4 fixes the bug in century-2002:20
and was added between rounds.

The new patterns J42 and J43 in fuseki.db are in response
to strategy5:232, a position that was discussed in:

Of course adding a J pattern to fix this regression does not address the GNU
Go's tendency to play clumsily when owl reasons are involved. The second
pattern makes sure that GNU Go will not lose the connection. I tested
followup play after these two patterns and they seem to be acceptable.

The pattern CB329a was added between rounds. Before move 132
in the game manyfaces-gnugo.sgf B can make an atari at L16,
reducing white's moyo in sente. At one time this move would
have been contributed by CB329, which was correctly taken
out by Trevor. This pattern might be OK.

The pattern EJ97 is relevant to strategy5:229.

The patterns RE8 and RE9 were added between rounds. The
pattern RE8 is relevant to century-2002:20 and the
pattern RE9 is relevant to the position in manyfaces-gnugo.sgf
at mvoe 27. I didn't add a test at this point but generally
the keimi suberi is better than the kosumi.


reply via email to

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