Hi, as usual Valkyria seems to handle this position well at the price
of being a super slow program in general.
This is just one example of how it reacts.
After 100 simulations it treats F1 as the best almost always, having
searched 30 to 100 times. Perahps 50-70 times is the most common result.
Valkyria handles this position well for two reasons. As soon as X
plays F1, the semeai is defended properly, so the Winrate for O is
just 5% and then drops down to 0%. The few wins for O comes from X
randomly removing one of its own liberties in a dumb way at J5.
Removing the liberty at G5 that makes an empty triangle is safely
pruned since it is dominated by G6 which removes a liberty from the
same block of O but gain a liberty. So one of the two gamelosing
mistakes are pruned.
The hard part for Valkyria is to search F1 at all, and there is a
special mechanism (or hack really...) to ensure that.
Valkyria computes AMAF win rates for all moves including those that
are pruned or illegal in the position. What I noticed is that in cases
of critical semeais the AMAF values of moves that are for example
illegal can get very high since they only get legal when you win the
semeai and thus win the game
Therefore Valkyria takes the AMAF values of moves that cannot be
played yet, and try to find approach moves that will enable playing
them and replaces the AMAF values of the approach move with the AMAF
value of the illegal move if it is higher.
This is a costly computation so it is only done if the position has
been visited several times.
Without this AMAF-hack Valkyria sometimes has a problem finding F1. It
also seems to take a longer time to find F1 in all cases where does
find it.
I have not yet tested the effect on the playing strength from this.
Best
Magnus
Quoting Brian Sheppard <[email protected]>:
The simplest problems give me new appreciation for the difficulties we face
in programming this
maddening game. Here is an example, with X to play:
1 2 3 4 5 6 7 8 9
A - X - - - - X - -
B - - - - X X - X X
C X - - - X - X O O
D X X X O X X O O O
E O O O X X X X O O
F - X X O O O O - O
G - X O O - - - O O
H O O X X X X O - -
J - O O X - O O - -
heuristics often produce the *wrong* move ordering, too. In that case there
is a loss of efficiency.
Yes this is the hard part. Note for example how Valkyria in this
position will prune X at G5 becuase G6 must always be a better move.
However, X J5 is not pruned because at the moment I have no simple way
of proving that removing a liberty from the opponent is not a good
move. The easiest way to prune that move would be to read the ladder
it creates, still it might be confused with a nakade that sacrifices
stones to kill and capture the surrounding stones so one need to make
sure this is not the case.... so I will not do that in Valkyria right
now it is just too complicated.
Best
Magnus
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
