Re: [gnugo-devel] trevor_1_20.5

[Gunnar Farneback]

Trevor wrote:
> http://www.public32.com/games/go/trevor_1_20.5
>  - tuning
>
> As reported before, this patch gives 41 passes.
> Here's my analisys of the failures:

Actually trevord.tst and global.tst weren't updated with the 3.1.19
results (these have been added in CVS afterwards) so 11 passes and 1
failure are old results. Still good of course. :-)

> global:16 FAILED this one keeps coming back.

This was the non-updated failure in global.tst.

> strategy3:147 FAILED   really was already broken.

This can be fixed again with some simple tuning. I have a patch for
this in preparation.

> trevorc:1430 FAILED Didn't used to call endgame module - way
>                     overvalued - don't know why.

Seems to be some problem with the influence function, which should be
possible to fix by tuning. Still this move is double sente and should
be valued fairly high.

Comments to the patch:
> +Pattern EB25
> +# tm New Pattern (3.1.20)
> +
> +O!.  cutable kiema on edge.
> +.!O
> +.!.
> +---
> +
> +:8,B
> +
> +Obc
> +daf
> +.!.
> +---
> +
> +; xplay_defend_both(a,b,c,a,c) ||
> +; xplay_defend_both(b,a,d,b,d)

This constraint looks wrong. Surely the xplay_defend_both calls should
be xplay_attack_either since O is in turn to play after the 3 moves?

> +# tm modified (3.1.20)
> +#   FIXME: Should dragons be connectable via ko?
> +#   see trevord:260

Probably not. It seems likely that having them separate will give the
best valuation.

> -Obc
> -daf
> +Gbc
> +daF
>  ooo
>  
> -;xplay_attack_either(a,b,c,a,c) && xplay_attack_either(b,a,d,b,d)
> +; xplay_attack_either(a,b,c,a,c) == WIN && 
> +; xplay_attack_either(b,a,d,b,d) == WIN

Why the new label G which isn't used? Why upper case F and G labels?
The usual convention is that X stones in the pattern get upper case
labels while O stones and empty vertices get lower case labels.

/Gunnar