We plan to submit the paper to CG2012, though there haven't been any
official announcements. It would be good for us if you could contribute the
results of your 'untuned' program before mid-June.

The following explanation provides more details of the regression test.

1. The 20 seki cases are for testing White's 'correct answer'. A case may
have more than one answer. Correct answers lead to White's win while other
wrong answers result in Black's win. Note that in some positions komi is
set to 7.0.

2. The 15 two-safe-groups cases are for testing Black's 'evaluation', all
should be 0% because White wins by living two groups. Every case has
another two relaxed versions, each makes one of White's two groups 100%
alive (assume the program doesn't fill a real eye in the playout). Take for
instance case1. case1-1 makes White's bottom group 100% alive; case1-2.sgf
secures the top group. So case1-1 and case1-2 are supposed to be easier
than case1 itself. To assess program's evaluation, 'sg_compare_float' is
used. For example,

sg_compare_float 0.4 uct_value

If uct_value (the winning rate of the search) is smaller than 0.4, return
-1; 1 otherwise.

Please run your program over the test in terms of number of simulations:1k,
2k, 4k, 8k until 128k simulations. For convenience, you can write a simple
script like

run.sh -p "PATH_TO_PROGRAM -playouts 1000" -t g_seki_moves.tst
rename html\g_seki_moves.tst\index.html g_seki_moves_1k.html

run.sh -p "PATH_TO_PROGRAM -playouts 2000" -t g_seki_moves.tst
rename html\g_seki_moves.tst\index.html g_seki_moves_2k.html

run.sh -p "PATH_TO_PROGRAM -playouts 4000" -t g_seki_moves.tst
rename html\g_seki_moves.tst\index.html g_seki_moves_4k.html

...

run.sh -p "PATH_TO_PROGRAM -playouts 128000" -t g_seki_moves.tst
rename html\g_seki_moves.tst\index.html g_seki_moves_128k.html


Best regards,
Aja
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to