Quoting Brian Sheppard :
What komi did you use? It is nice to have the sgf in addition to the
position.
It is 7.5, and I do not have the SGF. I will try to create SGF for future
posts, to make reproduction easier for all.
Could it be that Pebbles have trouble seeing that the semeai is won
after white C9.
Yes, exactly. Pebbles has no code (in either search or playouts) for winning
semeais of more than one move. No pattern that identifies C9 as a good
move for O.
There is code that finds C9 for X: it is a winning snapback. There is
no code that tries to play the opponent's move, so the knowledge does not
transfer from one color to the other.
What rule proposes C9 in Many Faces or Valkyria?
C9 is not treated in a special way. I guess in this case it is AMAF
that finds it and play at te root level. I use the same code to bias
the tree part as I use for the playouts. In the playouts a lot of
tactical code for pruning bad suicidal moves and playing forced moves
(when the number of liberties in the semeai is 2 or less) are used.
I can also be that Valkyria finds C9 for X quickly if O does not play it.
1 2 3 4 5 6 7 8 9
A - - - - - - - - -
B - O O O O X X X O
C - X X X O O X O -
D - O - X O X O O -
E O - O O X X O X X
F - O X X X X X O -
G - X X - - O X O -
H X O O O - - X X O
J - - - - - - - O -
On the face of it, C9 doesn't put O ahead in any semeai. After C9, O is
behind X's B8 string by 3 liberties to 2, and O is behind X's E8 string
by 2 liberties to 2 with X to move.
Anyway I entered the position manually with komi 7.5, and Valkyria
plays C9 right away winning the semeai in the upper right corner and
after that white wins 0.5 even if black gets everything else.
C9 is winning, but it isn't so obvious as Magnus suggests. There are
several complexities to see through.
Your analysis is correct. And to me it just appears quite obvious. And
as David point out it is not physical liberties that are counted but
the move necessary to play.
1) O wins the semeai on top only if O moves first. After X's A8, O must
find A6, A5, and A7. Those moves must be played in that order, because if
X plays A8 and A6 then X wins the semeai because O cannot play A9, which
is self-atari.
2) If O tenukis at any time, then X wins by playing F9 or G9, followed
by capturing O's F8/G8 string, and atari with D9.
3) In a random-play game, the fact that X has 2 sequences versus 1 (i.e.,
F9 or G9 for X versus only A6 for O) makes up for the fact that O gets to
move first. So it is vitally important to have code in the playouts that
handle the semeai more accurately for O than purely random.
4) Magnus says "wins by 0.5 even if Black gets everything else" but that's
not right. O must also win the semeai at left. In a random-play game, O
would lose that battle fairly often. The principal way to lose is X C1,
O tenuki, X D3 (atari), O E2, X F1 (atari), O G1 and a Ko fight for life.
The O must win the semeai is also obvious. What I meant is that O can
let X start ko fights and win them as long O just captures the large X
block.
5) The Ko fight there is a picnic for X, since X was counting on losing
anyway. It follows that X will gain a move in the battle of his choice,
so O will only win if his tenuki was played in the semeai at upper right.
I think that to play this situation completely right you must have playout
policies that specifically drive success in semeais and/or ko.
Valkyria does nothing intelligent when it comes to ko and wins always
in this position. But this is because white can let black win all
small ko fight as long as all important semeais is won. So in this
position only semeai knowledge is necessary.
But you might try to decrease the komi, because then white will have
to win with a larger margin! That could make the position much harder
to read out correctrly.
Those successful policies do not have to be right for the right reason.
For example, if you play C9 because you believe that it is the only move
that has a chance of winning the semeai against E8/E9, then you are right
for the wrong reason, because there is no way to win against E8/E9;
C9 just happens to win against another string.
Without heuristics that specifically drive success, the combination of
multiple battles make matters combinatorially worse.
For example, the dynamic in Pebbles is for the moves that win the semeais to
be
the second, third, fourth or higher move generated. This is not so bad
if there is only one battle. But when multiple battles are joined, X can
play a forcing move for a turn in another battles, and then return
to the other side. It is a dynamic version of the horizon effect, where bad
effects are first pushed out, and then delayed by bad move ordering.
I think that working on this position will yield several advances in
move ordering. I have already found the snapbacks, and I think there
will be others.
Are you talking about move ordering in the tree or the playouts