Le mercredi 12 décembre 2007, Gunnar Farnebäck a écrit : > Heikki Levanto wrote: > > On Mon, Dec 10, 2007 at 04:08:48PM -0500, Don Dailey wrote: > >> Would you rather be 95% confident of a win or 90% confident? There is > >> only 1 correct answer to that question. > > > > Yes, if you can offer me reliable confidence numbers. We all (should) > know > > that MC evaluations suffer from systematic problems that can not just be > > averaged away statistically. > > > > Compare these two positions: > > > > playout_benchmark 10000 > > = Initial board: > > komi 7.5 > > A B C D E F G H J > > 9 . . . . . O O O O 9 > > 8 O O O O O O O O O 8 > > 7 O O O O O O O O O 7 > > 6 O O O O O O O O O 6 > > 5 # # # # # # # # # 5 > > 4 O O O # # # # # # 4 > > 3 O O O O . # # # # 3 > > 2 . O O O . # # # . 2 > > 1 # . O O . # # . # 1 > > A B C D E F G H J > > Performance: > > 10000 playouts > > 0.032002 seconds > > 312.481 kpps > > Black wins = 1937 > > White wins = 8063 > > P(black win) = 0.1937 > > > > > > playout_benchmark 10000 > > = Initial board: > > komi 7.5 > > A B C D E F G H J > > 9 . # . . . O O O O 9 > > 8 O O O O O O O O O 8 > > 7 O O O O O O O O O 7 > > 6 O O O O O O # # # 6 > > 5 # # # # # # # # # 5 > > 4 O O O # # # # # # 4 > > 3 O O O O . # # # # 3 > > 2 . O O O . # # # . 2 > > 1 . . O O . # # . # 1 > > A B C D E F G H J > > Performance: > > 10000 playouts > > 0.084006 seconds > > 119.039 kpps > > Black wins = 7746 > > White wins = 2254 > > P(black win) = 0.7746 > > > > > > Which one is better, 77% or 19%? > > This reminds me of the first testcase I wrote when I started with > MonteGNU. Black to play, no komi. > > A B C D E F G H J > 9 . . O O X . X . X 9 > 8 . . . O X . X O X 8 > 7 O . O O X X O O X 7 > 6 O O O . X . X O O 6 > 5 X X X X X O O O . 5 > 4 . . X . O O . O . 4 > 3 X X O X O . + O . 3 > 2 X X O X O . . O . 2 > 1 . O O O O . . . . 1 > A B C D E F G H J > > Naturally B has to play B8, or white plays there and wins big. This is > trivial to find for a classic program and easy enough for a Monte > Carlo program. What's interesting is that it takes some work to make > black think that it has better than even winning chances after B8. The > Monte Carlo code in GNU Go CVS version gets 0.079 with 10k, 0.387 with > 100k, and 0.475 with 1M simulations. I suspect that stronger programs > tend to be more optimistic about winning chances here. So please fill > in this table if you have an MC program: > > 10k 100k 1M > -------------------------------- > GNU Go CVS 0.079 0.387 0.475 > > The sgf file is attached, load it before the first move. The positions > before move 3 and 5 are also relevant tests. > > /Gunnar >
Can't this test be used to find a threshold for not playing in opponent_territory_with_likehood_greater_than_N% I mean H1 is W territory with very high probability, B8 is W with some rather high probability (even if B can kill) so cutting search to prevent (like standard gnugo) B trying moves in H1 territory might help to find b8 quickly ? Or maybe i misunderstood something ? Alain _______________________________________________ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel