Re: [gnugo-devel] Olympiad games

[Arend Bayer]

Gunnar wrote:

> I wrote:
> > Before I add them to CVS I need to strip out the move value markings
> > and fix a couple of undos in one of the games.
>
> regression/games/olympiad2002. This gives us a wealth of games to
> discuss. I'll start with some comments about the first Go Intellect
> game, game1-19-goint-gnugo-1-0.sgf.
>
> Inge wrote:
> > The game against Go Intellect was lost with 19.5 points and was very
> > exiting.  I will try to send the game record later today. GNU Go
> > started by defending somewhat badly against an invasion.

This game is indeed exciting, with lots of places for owl
tuning/improvements.

> All moves in the combination B11 (17), B13 (19), D11 (21) are at best
> questionable. Incidentally they are all thought to be owl attacks on
> C12.
>
> A relevant question is whether we should even consult the owl code for
> a light stone like C12. It would make more sense to strengthen the
> surrounding neighbors and aim to capture it on a large scale. Any
> suggestions for how to
> (1) identify light stones?
> (2) identify and value moves capturing on a large scale?

I have seen similar problems, and I suggest starting at another end. C12
looks worth owl attacking here, if it escapes, the surrounding groups will
get weak. However, of course one shouldn't start doing this with B11 or B13
(D11 looks acceptable in my opinion). I think it would make a lot of sense
to do some serious effort on implementing a different move order for
owl at stackp == 0. Further down the tree, we are only interested
_whether_ we can kill/live. At stackp==0, where the move found by owl
will often be the move played, it is very important to use the _right
move_ to kill/live.

Hence, at stackp>0, the move ordering should be "most likely move to
kill first", whereas at stackp==0 we want "the move that would be the
best way to kill first".

I assume the different policy about vital moves (see comments in owl.c around
1285) according to stackp == 0 resp. stackp > 0 is motivated by the same
reasoning.
One way to go further than that would be to order the moves according to
their territorial effect computed via the "dragons_unknown"-influence.

I am sure that at move 17, owl would find that D11 kills C12. This would
be a perfectly fine move, punishing C12. So D11 should be tried first,
and only if this does not kill, owl should try like B11.

(I have run into this when looking at PASSes/FAILs resulting from the
owl indeterminacy. A couple of those were caused by owl returning a
different move to kill/live for an obviously critical dragon.)

> > Then the invading group got into trouble, but managed to escape.
>
> That trouble seemed to be kind of accidental. GNU Go didn't fight very
> well. H12 (47) should have been G12, G11 (51) should have been F11.

And R11 (29) was a bad tenuki. Should be at H14 although the shape there is
already less than perfect. We do need to find a way to award a good
strategical bonus to moves like H14 here.

G14 (57) probably has to be at G16, but may be hard to see without
lookahead.

B8 (69) looses a lot of points. The strategic (over-)valuation of B8
is bogus, whereas Q16 is enormously undervalued here (is 16.5; 30 looks
closer to the truth).

Similar problem with A9 (81).

> > Gnugo had to run with a weak group, managed to save it and then
> > abandoned it!
>
> The major problem was J15 (91) which must be at R12. There's no
> discussion that the owl reading made a major mistake, thinking J15 was
> an owl defense for L15. This was partly caused by, unusually enough,
> an underamalgamation. With --experimental-connections L15 and Q14 are
> correctly amalgamated, but then it instead fails to find an owl attack
> at all.

Which looks correct; it does find the owl defense at S13 in the next
move. Of course still better to play R12 in the first place, but S13
would reduce the damage quite a lot.
>
> > After that it came back pretty strongly with a huge moyo which
> > solidified into territory.
>
> S7 (95) and N11 (97) weren't all that great, but after that it did
> better than Go Intellect at least.

S7 could also be seen as a an owl-move-order-at-stackp==0-problem
(Q7 still wouldn't be the move to play here, but at least it would be
clean capture).

O9 (119) is a bit strange. There is s.th. weird with the moyo size of
L11 here.

Altogether Go Intellect did better in the fights, but strategically
GNU Go was a little better I would say. A close game I think.

Arend