Re: [gnugo-devel] Pattern tuning and regression tests

[Arend Bayer]

> One could do a run of 100 games using twogtp and record
> which patterns contributed some move, and which patterns
> contributed the move that was actually played. This might
> be useful information. If the database was considerably
> larger than 100 games it might help find patterns that
> are candidates for deletion.
Hmm, I would be VERY careful in trusting such a test set as well. E.g.,
in a recent NNGS game, I saw GNU Go play (GNU is O)

|.......
|..X1...
|..2....
|...O...
|.....X.
|.......
|.......
+-------

and tenuki. You will hardly see the urgency of a pattern treating this
situation from GNU Go vs GNU Go games, as GNU Go would never play 2.
I could imagine there are tons of such situations. (In fact, I once tried
to play lots of unusual moves in a high handicap game against GNU Go, and
easily succeeded in provoking non-sense move. This will always happen
against humans who have only a vague idea about the weaknesses of go
programs.) That's why I thought about systematically using "real life"
games.

Happy Xmas
Arend