r6144 [Fri, 15 Jul 2011 08:07:31 +0000 (15 16:07 +0800)]
More notes on the possible min/max method.
r6144 [Fri, 15 Jul 2011 07:27:34 +0000 (15 15:27 +0800)]
Now prints last_urgency to help debugging. Seems that the problematic moves in 110713-85x is indeed due to the average being biased by the not-so-good follow-up moves, whose weight has increased when the algorithm does more exploring. Consequently, even when some opponent moves have a poor outcome for us, the value of our preceding move may still be quite high.
r6144 [Fri, 15 Jul 2011 05:55:22 +0000 (15 13:55 +0800)]
Some tests regarding the 110713 game. Turns out that pachi-my made fewer mistakes than I had thought, and the mistakes (e.g. in 85x) are frequently due to the bias introduced by the early non-optimal followup moves. A large expand_p for nodes that the playout engine has trouble dealing with may also be a problem, since it tends to bias the result for many playouts.
r6144 [Fri, 15 Jul 2011 03:30:18 +0000 (15 11:30 +0800)]
More notes on the possible improvements to UCT.
r6144 [Thu, 14 Jul 2011 15:50:37 +0000 (14 23:50 +0800)]
Added 110714-73.gtp. The original move 73 is apparently no longer generated after more playouts.
r6144 [Thu, 14 Jul 2011 15:45:11 +0000 (14 23:45 +0800)]
Added some investigations of another game. Move 61 seems to yield rather random results, suggesting that the UCT method does not explore enough, but then it would be a good idea to study carefully the consequences of a move that we are actually going to take. In any case, 110714-61x.gtp shows that 61 at J16 won't work.
r6144 [Thu, 14 Jul 2011 11:05:00 +0000 (14 19:05 +0800)]
Added promote-move2.gtp, which is a shorter test case for the promote-move problem, but since this is a random problem, it is not often triggered by this test case.
r6144 [Thu, 14 Jul 2011 07:07:10 +0000 (14 15:07 +0800)]
Updated NOTES after the last commit.
r6144 [Thu, 14 Jul 2011 07:02:31 +0000 (14 15:02 +0800)]
Now preserves extra_komi after tree resets causing by the opponent making a move not in the tree.
r6144 [Thu, 14 Jul 2011 06:12:51 +0000 (14 14:12 +0800)]
Some more notes on pruning and the cause of the cannot-promote-move-node problem.
r6144 [Thu, 14 Jul 2011 04:28:28 +0000 (14 12:28 +0800)]
Yet more notes on memory management.
r6144 [Thu, 14 Jul 2011 03:25:00 +0000 (14 11:25 +0800)]
More notes. Also investigating the unable-to-promote-move-node problem.
r6144 [Thu, 14 Jul 2011 02:55:49 +0000 (14 10:55 +0800)]
More notes.
r6144 [Thu, 14 Jul 2011 02:44:36 +0000 (14 10:44 +0800)]
Now uct1amaf generates a pass if it is indeed the best.
r6144 [Wed, 13 Jul 2011 15:46:33 +0000 (13 23:46 +0800)]
Need to investigate why a pass isn't generated even though it is the winner of the tree.
r6144 [Wed, 13 Jul 2011 15:31:16 +0000 (13 23:31 +0800)]
Reduced the annoying verbosity in best2_ratio stuff.
r6144 [Wed, 13 Jul 2011 15:29:01 +0000 (13 23:29 +0800)]
Updated the notes after the twogtp game.
r6144 [Wed, 13 Jul 2011 15:19:39 +0000 (13 23:19 +0800)]
Added some twogtp stuff for testing. pachi-my loses by 1.5 points, and does not know how to pass as well.
r6144 [Wed, 13 Jul 2011 14:12:04 +0000 (13 22:12 +0800)]
Added tests for moves 55 and 56 in the 110709 game. The results are satisfactory.
r6144 [Wed, 13 Jul 2011 14:03:59 +0000 (13 22:03 +0800)]
Now expand_p is set automatically so that memory won't be filled during the entire simulation.
r6144 [Wed, 13 Jul 2011 13:46:27 +0000 (13 21:46 +0800)]
Disabled resigning for now due to potential dynkomi issues.
r6144 [Wed, 13 Jul 2011 13:31:09 +0000 (13 21:31 +0800)]
Changed some verbosity settings for -d 3.
r6144 [Wed, 13 Jul 2011 13:24:58 +0000 (13 21:24 +0800)]
Now move58 works much of the time.
r6144 [Wed, 13 Jul 2011 13:14:40 +0000 (13 21:14 +0800)]
Now makes ucb1amaf do a bit more exploration. However, move58 is not solved yet.
r6144 [Wed, 13 Jul 2011 11:57:51 +0000 (13 19:57 +0800)]
Now shows memory consumption in the progress information.
r6144 [Wed, 13 Jul 2011 11:43:28 +0000 (13 19:43 +0800)]
Added some notes.
r6144 [Wed, 13 Jul 2011 09:57:21 +0000 (13 17:57 +0800)]
Disabled early break due to sure_win_threshold since it is no longer that meaningful under dynkomi, and tends to lead to slack moves near the endgame. In particular it isn't necessary under TD_GAMES.
r6144 [Wed, 13 Jul 2011 09:48:34 +0000 (13 17:48 +0800)]
Now allows stopping early in TD_GAMES mode, since occasionally more simulations might make things worse.
r6144 [Wed, 13 Jul 2011 07:24:09 +0000 (13 15:24 +0800)]
Fixed potential buffer overflow in the spaces array. Untested.
r6144 [Wed, 13 Jul 2011 06:28:09 +0000 (13 14:28 +0800)]
Fixed dynkomi and progress notification problems under TD_GAMES time control.
r6144 [Wed, 13 Jul 2011 05:24:55 +0000 (13 13:24 +0800)]
Added some debugging code regarding the problem that the dynkomi is sometimes too small. Turns out that searching stops too early for dynkomi to reach the correct value. However, uct_search_stop() is waiting for a long time. Investigate.
r6144 [Wed, 13 Jul 2011 04:30:19 +0000 (13 12:30 +0800)]
Fixed the comments in the last patch.
r6144 [Tue, 12 Jul 2011 15:08:07 +0000 (12 23:08 +0800)]
No longer disables use_extra_komi near the end of the game, since the heuristic is too rough.
r6144 [Tue, 12 Jul 2011 13:09:22 +0000 (12 21:09 +0800)]
Added the unlimited_tree_dump option so that reducing dumpthres actually becomes useful.
r6144 [Tue, 12 Jul 2011 12:43:41 +0000 (12 20:43 +0800)]
Now no longer rounds the result when converting it back to a score (which may contain .5's due to komi).
r6144 [Tue, 12 Jul 2011 12:40:47 +0000 (12 20:40 +0800)]
Added default size for sgf files without an explicit SZ statement.
Jean-loup Gailly [Thu, 16 Jun 2011 20:12:08 +0000 (16 22:12 +0200)]
Distributed engine: better default value for stats_delay
Jean-loup Gailly [Tue, 14 Jun 2011 13:43:06 +0000 (14 15:43 +0200)]
Distributed engine: replace MIN_STATS_INTERVAL constant with stats_delay parameter
Matthew Woodcraft [Mon, 30 May 2011 14:13:31 +0000 (30 16:13 +0200)]
[PATCH] board_handicap(): Correct placement for three-stone handicaps
Petr Baudis [Mon, 30 May 2011 13:57:17 +0000 (30 15:57 +0200)]
Merge branch 'master' of ssh://repo.or.cz/srv/git/pachi
Petr Baudis [Mon, 30 May 2011 13:57:05 +0000 (30 15:57 +0200)]
ucb1amaf_update criticality stats: Fix invalid memory access on pass
Petr Baudis [Sun, 29 May 2011 19:11:39 +0000 (29 21:11 +0200)]
board_clear(): Cache set up board structures - major speedup for loading fbook
Petr Baudis [Sun, 29 May 2011 18:38:19 +0000 (29 20:38 +0200)]
board_clear(): Factor out board_init_data()
Petr Baudis [Sun, 29 May 2011 18:29:23 +0000 (29 20:29 +0200)]
fbook: Enable debug prints
Petr Baudis [Sun, 29 May 2011 18:21:22 +0000 (29 20:21 +0200)]
fbook_init(): Return NULL if the book turned out to stay empty
This also prevents cache reset on such occassion.
Petr Baudis [Sun, 29 May 2011 18:16:27 +0000 (29 20:16 +0200)]
fbook caching: Store last loaded fbook and reuse it if it matches current board
Petr Baudis [Sun, 29 May 2011 18:09:30 +0000 (29 20:09 +0200)]
fbook: Increase fbook_hash_bits to 20 to accomodate even large fbooks like ra6
Petr Baudis [Sun, 29 May 2011 15:49:16 +0000 (29 17:49 +0200)]
Merge branch '2lib'
Petr Baudis [Sat, 28 May 2011 21:21:20 +0000 (28 23:21 +0200)]
can_atari_group(): When switching to selfatari cousin, check defense_is_hopeless()
Petr Baudis [Sat, 28 May 2011 19:37:48 +0000 (28 21:37 +0200)]
selfatari_cousin(): Make info on by-group available
Petr Baudis [Sat, 28 May 2011 19:34:46 +0000 (28 21:34 +0200)]
can_atari_group(): Factor out hope check to defense_is_hopeless()
Petr Baudis [Sat, 28 May 2011 18:58:51 +0000 (28 20:58 +0200)]
can_atari_group(): Add massive amount of debugging prints
Petr Baudis [Sat, 28 May 2011 16:56:51 +0000 (28 18:56 +0200)]
TODO: Liberty maps (developed in libmap branch), reverse status learning
Petr Baudis [Wed, 25 May 2011 22:24:41 +0000 (26 00:24 +0200)]
tools/sgf2gtp.pl: Support HA[] sgf tag
Petr Baudis [Wed, 25 May 2011 22:24:27 +0000 (26 00:24 +0200)]
tools/sgf2gtp.pl: Accept filename as optional argument
Petr Baudis [Wed, 25 May 2011 21:53:22 +0000 (25 23:53 +0200)]
gtp: Add fixed_handicap to list of supported commands
Petr Baudis [Sun, 22 May 2011 17:27:03 +0000 (22 19:27 +0200)]
Pachi Hakugen-devel 8.99
Petr Baudis [Sun, 22 May 2011 17:24:00 +0000 (22 19:24 +0200)]
Pachi Shuhaku 8.00
Petr Baudis [Sun, 22 May 2011 17:23:31 +0000 (22 19:23 +0200)]
Parallel make: Fix pachi being built before some libraries
Petr Baudis [Sun, 22 May 2011 08:05:10 +0000 (22 10:05 +0200)]
fbook: Share no-handicap fbook even for low komi values
Suggested by jlg.
Jean-loup Gailly [Sat, 14 May 2011 20:40:42 +0000 (14 22:40 +0200)]
fbook_init(): do not force h1 for large negative komi
Jean-loup Gailly [Fri, 13 May 2011 17:38:24 +0000 (13 19:38 +0200)]
Dynamic komi: fix linear_permove() when pondering
Petr Baudis [Mon, 16 May 2011 14:36:44 +0000 (16 16:36 +0200)]
Criticality: New defaults based on tuning by jlg
Petr Baudis [Mon, 16 May 2011 14:35:13 +0000 (16 16:35 +0200)]
Moggy ladderrate: Default to 80, confirmed by jlg
Jean-loup Gailly [Fri, 13 May 2011 15:12:58 +0000 (13 17:12 +0200)]
Fix parsing of mqprob, moves & handicap_value parameters
Petr Baudis [Fri, 13 May 2011 10:33:23 +0000 (13 12:33 +0200)]
Merge branch 'master' of ssh://repo.or.cz/srv/git/pachi
Matthew Woodcraft [Sun, 8 May 2011 14:44:07 +0000 (8 15:44 +0100)]
Add final_score to known_commands
Matthew Woodcraft [Sun, 8 May 2011 14:42:52 +0000 (8 15:42 +0100)]
Implement known_command; make gtp_is_valid() stricter
Petr Baudis [Mon, 9 May 2011 20:17:44 +0000 (9 22:17 +0200)]
Makefile.lib: Allow parallel build
Petr Baudis [Sun, 8 May 2011 17:38:38 +0000 (8 19:38 +0200)]
fbook: Add support for handicap lines; do not use normal lines in handicap games
Petr Baudis [Mon, 25 Apr 2011 20:16:03 +0000 (25 22:16 +0200)]
Moggy: cap_stone_denom=200, ladderrate=40
This turns out the capture stone weighting and preferring ataris that
walk the ladder. I have measured both to be gains with reasonably high
probability, but more tuning is needed to find the optimum.
Petr Baudis [Mon, 25 Apr 2011 19:19:08 +0000 (25 21:19 +0200)]
UCB1AMAF crit_*: New defaults based on very rough tuning
On 19x19, this may be as much as 50 ELO gain vs. crit_rave=0. More
tuning still going on.
Petr Baudis [Mon, 25 Apr 2011 10:48:21 +0000 (25 12:48 +0200)]
Merge branch 'master' into criticality
Conflicts:
uct/tree.c
Petr Baudis [Mon, 25 Apr 2011 10:46:36 +0000 (25 12:46 +0200)]
tree_node_dump(): Much more condensed output
Petr Baudis [Mon, 25 Apr 2011 00:22:44 +0000 (25 02:22 +0200)]
UCB1AMAF: Add criticality support, with many tuning options
crit_rave=2, crit_min_playouts=32, crit_negative=1, crit_amaf=1
Petr Baudis [Sun, 24 Apr 2011 23:50:38 +0000 (25 01:50 +0200)]
tree_node_criticality(): Add further explanation for drunk readers like me
Petr Baudis [Sun, 24 Apr 2011 22:01:40 +0000 (25 00:01 +0200)]
Merge branch 'master' into criticality
Petr Baudis [Sun, 24 Apr 2011 22:00:54 +0000 (25 00:00 +0200)]
tree_node_criticality(): Return floating_t instead of double
Petr Baudis [Sat, 23 Apr 2011 20:39:41 +0000 (23 22:39 +0200)]
Moggy: Turn off laderrate for now
Optimum value not definitely determined yet, but it is likely that the
gain is negative for -1.
Petr Baudis [Sat, 23 Apr 2011 20:34:44 +0000 (23 22:34 +0200)]
Merge branch 'ladders'
Petr Baudis [Sat, 16 Apr 2011 21:39:12 +0000 (16 23:39 +0200)]
README: Wording tweak
Petr Baudis [Sat, 16 Apr 2011 02:15:30 +0000 (16 04:15 +0200)]
media: Add two pictures of Pachi; let's not neglect its personality!
Petr Baudis [Sat, 16 Apr 2011 02:13:18 +0000 (16 04:13 +0200)]
TODO: Remove two stale items
Petr Baudis [Sat, 16 Apr 2011 02:10:15 +0000 (16 04:10 +0200)]
Rename the main program binary from zzgo to pachi
'zzgo' is still residuum from the earliest times when I was not decided
on the program name. Good bye!
Petr Baudis [Sat, 16 Apr 2011 02:05:22 +0000 (16 04:05 +0200)]
Header guards: Rename ZZGO_* to PACHI_*
Petr Baudis [Sat, 16 Apr 2011 02:03:43 +0000 (16 04:03 +0200)]
UCT max_tree_size: Lower the default to 1.4GiB; adjust and polish README
Many notebooks still sold have just 2GiB of memory, plus some memory may
be taken by OS and other apps. Overally, the 1.4GiB limit is IMHO a much
safer default, especially given there is no pondering and just single
core used by default too.
Petr Baudis [Sat, 16 Apr 2011 01:58:28 +0000 (16 03:58 +0200)]
UCT fast_alloc: Make default
Petr Baudis [Sat, 16 Apr 2011 01:56:49 +0000 (16 03:56 +0200)]
README: Move the installation instructions to the top
Petr Baudis [Sat, 16 Apr 2011 01:56:31 +0000 (16 03:56 +0200)]
Deprecate the zzgo-profiled make target
Petr Baudis [Sat, 16 Apr 2011 01:53:52 +0000 (16 03:53 +0200)]
README: Grammar polishing, rewording and extending the text
Petr Baudis [Sat, 16 Apr 2011 01:29:22 +0000 (16 03:29 +0200)]
More user friendly build-time configuration
The Makefile now contains descriptions for all available build time
options.
Petr Baudis [Thu, 14 Apr 2011 22:08:22 +0000 (15 00:08 +0200)]
Moggy stone weighing: Fix case when cap_stone_denom == 0
Petr Baudis [Thu, 14 Apr 2011 17:03:28 +0000 (14 19:03 +0200)]
Moggy 1lib prior: Add support for stone weighing, turned off by default
New options cap_stone_min, cap_stone_max, cap_stone_denom. Needs a lot of tuning.
Petr Baudis [Thu, 14 Apr 2011 16:34:33 +0000 (14 18:34 +0200)]
Merge branch 'master' into ladders
Petr Baudis [Thu, 14 Apr 2011 16:31:42 +0000 (14 18:31 +0200)]
Moggy pattern3, group_atari_check(): Ignore ladder extensions in case counter-capture is possible
Petr Baudis [Thu, 14 Apr 2011 16:31:12 +0000 (14 18:31 +0200)]
can_countercapture(): Export interface
Petr Baudis [Wed, 13 Apr 2011 20:08:23 +0000 (13 22:08 +0200)]
Petr Baudis [Sun, 20 Feb 2011 22:36:46 +0000 (20 23:36 +0100)]
Merge branch 'master' into ladders
Petr Baudis [Sun, 20 Feb 2011 22:36:23 +0000 (20 23:36 +0100)]
Merge branch 'master' of git+ssh://repo.or.cz/srv/git/pachi
Petr Baudis [Sun, 20 Feb 2011 19:48:23 +0000 (20 20:48 +0100)]
Moggy: Play the proper atari in case we can catch a group in ladder