[Pachi] Regression test suite
jloup at gailly.net
Wed Jun 8 00:45:13 CEST 2011
Here is my current list of good candidates to put in a regression suite.
move 133 P1 forces opponent to live, should be R1 to kill
move 135 S1 plays the ko threat before the ko at L12
(this occurs frequently)
pachi tries to run away from a losing ladder against a 7k in even
game. The 7k, disgusted, passed then resigned.
pachi shouldn't play the broken ladders up to move 53
* http://files.gokgs.com/games/2011/6/5/Zen19-pachi2.sgf should have won
. move 185 S15 very dubious
. move 211 T18 pachi still thinks the semeai is undecided.
but white has a lot more liberties. How can the playouts
not see this?
. move 227 S3: here are suggestions by Aja
If a false will become a real eye by capturing a diagonal group,
then capture the group instead of filling the false eye. So if
pachi wants to fill s3 it will capture at T1 instead. Then S3
will become an illegal move so pachi will evaluate this group
100% alive and won't play at T2 such stupid move.
pachi doesn't see even one move ahead in a semeai, it expects
an answer outside the semeai. The score drops from 0.625 to 0.337:
 best 0.624794 komi 0.4 | seq L19 G12 F11 P11 | can L19(0.625)
O19(0.623) N19(0.469) P11(0.523)
FINEST: Got successful response to command "genmove b": = L19
FINEST: Got successful response to command "play w j16": =
FINEST: Queued command sent to engine: genmove b
 best 0.336825 komi 0.2 | seq P11 T8 R3 M9 | can P11(0.337)
O11(0.336) O19(0.334) N19(0.340)
FINEST: Got successful response to command "genmove b": = P11
The game was lost already but pachi should have known this before.
move 244: T13 doesn't make sense. This is due to wrong estimate of
white's right group:
 best 0.492722 | seq T13 T4 D5 M6 | can T13(0.493)
S18(0.493) L4(0.493) E6(0.491)
FINEST: Got successful response to command "genmove b": = T13
At move 282 pachi still thinks white's group is undecided.
We must find a way of not filling the S7 eye in the playouts.
169 B5 seems to be the losing move. This move doesn't make sense.
Pachi has been wrongly evaluating both bottom group around S4 (it's just
hairy for simulations) and top group around A19 - that one is a bit more
interesting, the problem is that black is double likely to play E19
(since it's C19 selfatari cousin), but after E19, white will not notice
the need to connect at C19. Not sure how to solve this yet, perhaps some
way to add cousins to last_move logic...
won game lost: 212 H13 should be at J12.
had won game, move 135 F1 very bad. at end doesn't see ko at t7. (b t10
leading up to 90% then lost.
bad evaluation of 3 areas according to pasky.
lost but was leading according to ajahuang and Zen itself
good example of losing a won game.
terrible ladder bug
big mistake at F9 didn't see snapback. Pasky comments:
The "does not seen snapback" bug is nasty. The problem is that
there is also a "global capture heuristic" if there're no fun
local moves, it looks for groups to capture and if white tenukis
and does not capture immediately, this heuristic kicks in. But
there is no "global atari heuristic" that would drive black
similarly to play at F1 and create the snapback if I test with
playout=moggy:capturerate=0, it reliably destroys the snapback
i.e. with capturerate=90, white is much more likely to come back
to the situation later even when playing tenuki first.
big semeai with easy counting of liberties misjuged
move 169 0.7322 -> move 171 0.2600
lost because of unrecognized seki
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Pachi