Re: [gnugo-devel] nngs tests

[Gunnar Farneback]

Paul wrote:
> here are some tests from a recent gnugo game on nngs. they may need a
> revision. 3-3-10-pre-3 passes tests 560, 570 and 580.

The reason why the results differ is that --experimental-connections
are used on NNGS.

> in test 580, it also liked move H19, which works. maybe H19 and
> several other moves on 19th line should be added to the list of
> correct answers.

N19 is better since it also captures the two M18 stones but K19 is
acceptable (very solid). H19 is clearly inferior, both with regard to
endgame considerations and aji (leaves a big ko threat behind).

> +# nailer-gnugo-3.3.9-200210191421 problems.

Please use the filenames at
http://www.lysator.liu.se/~gunnar/gnugo/nngs/ rather than anything you
can find on the server. This game is called
gnugo-3.3.9-nailer-200210192319.sgf.

> +# The black dragon is still dead. No need eating tails.
> +loadsgf games/nngs/nailer-gnugo-3.3.9-200210191421.sgf 146
> +560 gg_genmove white
> +#? [!L16]

Without experimental-connections G9 is badly overvalued at 29.06
points and is the only reason why the test passes. L16 is still valued
too high with 16.3 points. This is mostly due to the
"cautious_impact_values", which give value to cutting dead dragons
apart if we are ahead. This example clearly shows a need to revise
that policy somehow.

> +# This move is even worse.
> +loadsgf games/nngs/nailer-gnugo-3.3.9-200210191421.sgf 148
> +570 gg_genmove white
> +#? [!F19]

The same thing goes for this move, where the cutting apart value is
clearly completely inappropriate.

> +# Killing black dragon is clearly better than an opportunity
> +# to eat some stones.
> +loadsgf games/nngs/nailer-gnugo-3.3.9-200210191421.sgf 152
> +580 gg_genmove white
> +#? [K19]

This is exactly the kind of problem with experimental connections I
hoped to find by running them on NNGS. The problem here is that J19 is
considered a separate dragon (worms which can be tactically captured
are never amalgamated with experimental connections) with the effect
that the owl reading for F18 never considers the K19 move. This needs
to be solved.

> +# Not even sure if it is needed, but it is much safer than D2
> +# and white is enough ahead to play conservatively.
> +loadsgf games/nngs/nailer-gnugo-3.3.9-200210191421.sgf 220
> +590 gg_genmove white
> +#? [B2]

This move does not look necessary. If anything, the result should be
[!D2].

/Gunnar