Re: [gnugo-devel] arend_3_20.8 and EB1401

[Arend Bayer]

On Sun, 25 May 2003, Paul Pogonyshev wrote:

> some changes went in with this patch are bogus:
>
> Pattern EB1403
> # gf Revised pattern. (3.1.3)
> # gf Revised constraint. (3.3.13)
> # ab Removed followup value (handled by influence code). (3.3.20) <<<
>
> ?oX??          sente hane
> ..XO?
> ?.*.o
> -----
>
> :8,Xe
>
> ?oD??
> b.Da?
> ?.*..
> -----
>
> see move valuation in the attached game (current cvs - gnugo 3.2) at move 91:
>
>   C1: 2.00 - change in territory
>   C1:   0.00 - total followup value, added 0.00 as territorial followup
>   C1: 0.08 - connects strings (connect value 2, shape factor 0.040)
> Move generation values C1 to 2.08
>   T17: 2.00 - change in territory
>   T17:   0.00 - total followup value, added 0.00 as territorial followup
>   T17: 0.08 - connects strings (connect value 2, shape factor 0.040)
> Move generation values T17 to 2.08
>
> i can't say for sure if the move played by both current cvs and gnugo 3.2 
> (B19)
> is worse than C1 or T17, but the latter are clearly undervalued. is B19 really
> worth 5 points like EB1401 claims? to me it rather looks like 3 points (plus a
> vague followup, maybe). and the 'o' in the pattern seems like influencing move
> valuation.

That's right. There definitely were a lot of cases where EB1403 produced
a follow-up value that was duplicated with the follow-up value given by
the influence code, making the hane overvalued. (It was also usually
played too early by GNU Go.)

I will try to find out why the influence code doesn't finde the followup
value in these cases.

Arend