Re: [gnugo-devel] owl goal dragon gets cut

[Arend Bayer]

Gunnar wrote:
> Arend wrote:
> > Say there were two strings, and they can be cut now. Then we compute the
> > distance map from both strings (using the readconnect code), and build two
> > components of the goal by assigning each stone to the string to which it is
> > closer. The new goal is then the bigger of the two components.
> 
> It should probably not only take size into account but also where the
> stones of the original dragon are.

That sounds reasonable, but I haven't hit this as a problem yet, so I'd
prefer to leave that as a later improvement. 

> > I would very much appreciate comments/ideas on this approach.
> 
> The trickiest part is of course to do this with reasonable
> performance. Do you know how the node counters compare to CVS?

Reading +1%, owl +2%, connection +5%. Overall performance penalty <3%.
A detailed analysis of the PASSes/FAILs showed that in the overwhelming
majority of the cases, owl reading was clearly improved, so I think it
is well worth the performance cost.

Only problem I have encountered are that sometimes gnugo now generates
sensesless "escape moves" towards cut off former goal stones, which have
no escape value at all (and then running out of depth limit etc.).

I have put it in CVS now.

Arend


I wrote:
> 16 PASS, 9 FAIL (not much analyzed yet):
> 
> blunder:21      PASS L9 [L9]
Very good: understands M7 can be attacked.
> strategy:8      FAIL F1 [N2|N1]
Reading worse
> arb:210         PASS D8 [D8]
Completely lucky.
> strategy2:73    PASS P15 [F7|R17|P15]
P15 immediately threatens cut, and gnugo understands this now.
> trevor:630      PASS 1 H13 [1 H13]
What it was designed for.
> nngs:510        PASS G14 [G14]
A bit lucky. (Goal dragon will be cut immediately once I add cuts for
saving lunches.)
> nngs:875        FAIL E7 [C5]
Now gnugo thinks white can still be captured after white C5. Of course
it has to try, nevertheless. Reading change more or less random.
> global:34       FAIL F3 [N6]
Unlucky: gnugo thinks white can capture the whole thing if it defends at
N6. A bit pessimistic, but the actual difference in the variation tree
is a clear improvement.
> vie:8           PASS 1 S9 [1 (R8|S9)]
Very good: gnugo understands now that q8 is a problem after black R7.
> arend:9         FAIL S2 [S17]
Reading improved.
> 13x13:36        PASS C5 [C5]
Improved.
> strategy4:182   FAIL S11 [E7|F8]
Owl reading worse.
> strategy4:188   PASS C4 [B6|C4]
The owl reading is improved. However, A6 is still valued too high.
> strategy4:206   FAIL K6 [L8]
This is mostly a semeai problem, with a random result coming from
running out of nodes. (Owl now thinks K6 kills the big white dragon,
which seems reasonable when ignoring the semeai.)
> owl1:311        PASS 1 E2 [1 E2]
A prime example why the patch is necessary. (CVS thinks W B2 defends
after B E2.)
> owl1:324        PASS 0 [0]
Reading very random but improved.
> nngs3:230       PASS E12 [!L12]
I don't understand why L12 is bad. Owl reading is slightly worse.
> nngs3:250       FAIL G13 [G11]
G13 is played as an owl attack against L17, which seems reasonable. The
owl reading is not very convincing, but slightly improved.
> nngs3:350       PASS E12 [E12]
Owl reading clearly improved.
> nngs4:230       PASS Q18 [Q18]
A bit lucky, but reading improved.
> nngs4:650       FAIL G12 [Q2]
Acceptable, owl reading improved.
> ninestones:630  PASS F5 [F5|H7]
Just lucky.
> ninestones:650  PASS Q5 [Q5]
Very important.
> gunnar:39       FAIL N1 [N4]
Reading improved (gnugo now understands n4 is necessary to save dragon,
but unfortunately misreads that n1 is good enough as well)
> kgs:320         PASS K12 [K12]
Very important.