Jean-loup Gailly [Sat, 22 Jan 2011 11:00:43 +0000 (22 12:00 +0100)]
Moggy: Better nlibrate default for 19x19
Jean-loup Gailly [Sat, 22 Jan 2011 10:14:37 +0000 (22 11:14 +0100)]
Merge branch 'master' into derm6
Jean-loup Gailly [Sat, 22 Jan 2011 10:07:53 +0000 (22 11:07 +0100)]
Better defaults for atari_def_no_hopeless, cfgd & local_tree_depth_decay
Jean-loup Gailly [Sat, 22 Jan 2011 08:33:32 +0000 (22 09:33 +0100)]
Distributed engine: Wait more for late slaves at the beginning
Jean-loup Gailly [Fri, 21 Jan 2011 21:54:24 +0000 (21 22:54 +0100)]
N-liberty semeai defensive tactics: Introduce, add Moggy nlibrate, nlib_count
I am very hopeful. With nlib_count==4, it already repairs (at least to some
degree) evaluation of most situations misevaluated by pachi2 (64x20 core
cluster) in recent KGS games.
It also includes currently disabled code for making eyespace from
row-of-three; needs to be tested, esp. if it makes sense with nakaderate
from nakade-vs-seki.
Conflicts:
playout/moggy.c
Jean-loup Gailly [Fri, 21 Jan 2011 21:52:49 +0000 (21 22:52 +0100)]
2lib check_group_atari() -> can_atari_group(), export
Conflicts:
tactics/2lib.c
Petr Baudis [Sat, 15 Jan 2011 16:51:43 +0000 (15 17:51 +0100)]
Selfatari three_liberty_suicide(): Fix d5c9b, do not trigger if the group has an eye
Petr Baudis [Thu, 13 Jan 2011 04:00:04 +0000 (13 05:00 +0100)]
Selfatari: Try to avoid bad moves of three-liberty groups too
Petr Baudis [Thu, 13 Jan 2011 02:38:11 +0000 (13 03:38 +0100)]
TODO: Playouts - remove two implemented ideas, add 2-lib ladder check
Petr Baudis [Thu, 13 Jan 2011 00:20:47 +0000 (13 01:20 +0100)]
check_group_atari(): Use lumpiness guard only as preference, not for total pruning
We would miss many crucial counter-atari simply because the only
possible liberty to play was too lumpy.
Jean-loup Gailly [Sat, 15 Jan 2011 23:25:19 +0000 (16 00:25 +0100)]
Revert "Moggy ko check: Aside of connecting the ko, also consider countercapturing a neighbor"
This reverts commit
323eb0da8866b4835d34fef7a4cc3b5c52a381fb.
Jean-loup Gailly [Sun, 16 Jan 2011 23:30:07 +0000 (17 00:30 +0100)]
TODO: Support for different float types already done
Jean-loup Gailly [Sun, 16 Jan 2011 23:26:48 +0000 (17 00:26 +0100)]
gtp_parse(): Do not clear the board on kgs-game_over
Jean-loup Gailly [Sun, 16 Jan 2011 23:16:27 +0000 (17 00:16 +0100)]
Distributed engine: Do not forward unsupported gtp commands to slaves
Jean-loup Gailly [Sun, 16 Jan 2011 21:01:00 +0000 (16 22:01 +0100)]
Merge branch 'master' into derm6
Petr Baudis [Sat, 15 Jan 2011 12:55:02 +0000 (15 13:55 +0100)]
Moggy defaults: Pattern2 on large board, turn off capturerate even on small ones
Petr Baudis [Sat, 15 Jan 2011 12:49:38 +0000 (15 13:49 +0100)]
Moggy ko check: Aside of connecting the ko, also consider countercapturing a neighbor
Petr Baudis [Fri, 14 Jan 2011 12:50:49 +0000 (14 13:50 +0100)]
UCT max_maintime_ratio default 3.0 -> 8.0
Found by Jean-Loup Gailly to be few tens ELO improvement.
Petr Baudis [Fri, 14 Jan 2011 12:48:28 +0000 (14 13:48 +0100)]
New Moggy, UCT prior defaults tuned for 19x19 by Jean-Loup Gailly
Local climbing has been used. More tuning to come. Few (hopefully
independent) new default values have not been committed yet.
Petr Baudis [Fri, 14 Jan 2011 12:42:33 +0000 (14 13:42 +0100)]
board_large(): Introduce, use instead of current 9-or-19 tests
Instead, new threshold is 15x15. It is just my feeling that from this
point on, the board behaves more like 19x19. 13x13 is kind of
borderline, but I do not think anyone really cares much about this
board size anyway; and I do like to run Pachi on 15x15 occassionally.
Petr Baudis [Sun, 9 Jan 2011 04:46:41 +0000 (9 05:46 +0100)]
uct_leaf_node(): Take *descent in a way so that we can append to it
The argument is not used in mainline but it is in side branches.
Petr Baudis [Sun, 9 Jan 2011 04:12:43 +0000 (9 05:12 +0100)]
play_random_move(): Split out of play_random_game() for manual invocation
Petr Baudis [Sat, 8 Jan 2011 18:20:59 +0000 (8 19:20 +0100)]
HACKING: Describe local trees
Petr Baudis [Sat, 8 Jan 2011 16:02:02 +0000 (8 17:02 +0100)]
Remove the General Pattern Matcher
This has been a costly failed experiment. I spent many weeks of hard
work trying to make it work, unfortunately it seems unlikely that it
will ever attain the required performance. Therefore, let's remove it
from the current version so that it does not clutter up the source.
It is still available in the history and I expect to yet dig bits out
of there.
This leaves the board traits support orphaned, and for future probdist
experiments we will need to re-add the board gamma support. But I have
not abandoned probability distribution support entirely; in the moggypd
branch I'm trying a different approach, using direct Moggy policy hints
for pattern matching, instead of building up an independent matcher.
Petr Baudis [Sat, 8 Jan 2011 13:47:20 +0000 (8 14:47 +0100)]
Moggy: Add eye-piercing side pattern - commented out
My tests are inconclusive, but I still believe that perhaps if it is
tweaked, it can be made to work. So save for later.
Jean-loup Gailly [Sat, 8 Jan 2011 02:21:22 +0000 (8 03:21 +0100)]
Distributed engine: avoid playout overflow with many slaves.
Petr Baudis [Sat, 8 Jan 2011 01:36:05 +0000 (8 02:36 +0100)]
UCT Local Trees: New default values, tuned for local_tree=1
Should give about 20 ELO boost (increasing with playout count) over
normal; you still need to enable local trees manually to make use of
this.
Petr Baudis [Thu, 6 Jan 2011 00:53:05 +0000 (6 01:53 +0100)]
Merge branch 'master' of ssh://repo.or.cz/srv/git/pachi
Petr Baudis [Thu, 6 Jan 2011 00:52:51 +0000 (6 01:52 +0100)]
2lib: Introduce miaisafe, def_no_hopeless knobs, twiddle by respective Moggy flags
It seems that the benefit of these is questionable on some board sizes.
Petr Baudis [Tue, 17 Aug 2010 20:47:47 +0000 (17 22:47 +0200)]
UCT local_tree_rootseqval: Implement
Jean-loup Gailly [Wed, 5 Jan 2011 18:25:09 +0000 (5 19:25 +0100)]
Distributed engine: ensure that pass & resign have more weight than normal moves
Petr Baudis [Sat, 1 Jan 2011 13:43:06 +0000 (1 14:43 +0100)]
Merge branch 'master' of git+ssh://repo.or.cz/srv/git/pachi
Conflicts:
uct/policy/ucb1amaf.c
Jean-loup Gailly [Thu, 30 Dec 2010 19:02:06 +0000 (30 20:02 +0100)]
UCT engine: change virtual_loss from bool to int
Jean-loup Gailly [Sat, 25 Dec 2010 11:57:13 +0000 (25 12:57 +0100)]
UCT engine: Allow spending more time on a single move
Jean-loup Gailly [Sat, 25 Dec 2010 11:27:55 +0000 (25 12:27 +0100)]
Distributed engine: Use larger receive queue to allow longer time settings
Jean-loup Gailly [Sat, 25 Dec 2010 10:54:40 +0000 (25 11:54 +0100)]
Merge branch 'master' into test3
Conflicts:
uct/internal.h
uct/walk.c
Jean-loup Gailly [Fri, 24 Dec 2010 10:40:52 +0000 (24 11:40 +0100)]
Distributed engine: Transmit values in double precision
Petr Baudis [Thu, 23 Dec 2010 17:38:17 +0000 (23 18:38 +0100)]
Debugging: Introduce -D (debug_boardprint) for more gogui friendliness
Petr Baudis [Wed, 22 Dec 2010 18:20:41 +0000 (22 19:20 +0100)]
UCB1AMAF local tree: Add missing LTREE_PLAYOUTS_MULTIPLIER division
Petr Baudis [Wed, 22 Dec 2010 17:53:37 +0000 (22 18:53 +0100)]
local_tree_depth_decay: Introduce
Petr Baudis [Wed, 22 Dec 2010 17:52:41 +0000 (22 18:52 +0100)]
ucb1rave_evaluate() ltree_rave: Fix float multiplication
Jean-loup Gailly [Wed, 22 Dec 2010 17:31:33 +0000 (22 18:31 +0100)]
Use correct format when scanning double floating point values.
Jean-loup Gailly [Tue, 21 Dec 2010 09:16:41 +0000 (21 10:16 +0100)]
Allow double for all floating point values in large configurations.
Petr Baudis [Wed, 15 Dec 2010 17:55:39 +0000 (15 18:55 +0100)]
Merge remote branch 'origin/master'
Petr Baudis [Wed, 15 Dec 2010 17:55:06 +0000 (15 18:55 +0100)]
UCT treepool_pickfactor: Introduce for non-uniform treepool selection
Petr Baudis [Wed, 15 Dec 2010 11:11:36 +0000 (15 12:11 +0100)]
UCT significant: Move out of descent structure, track separately
Petr Baudis [Wed, 15 Dec 2010 11:11:01 +0000 (15 12:11 +0100)]
UCT treepool_setup(): Fix finding best move
Petr Baudis [Wed, 15 Dec 2010 11:09:12 +0000 (15 12:09 +0100)]
Debug prints improvements
Petr Baudis [Wed, 15 Dec 2010 11:08:07 +0000 (15 12:08 +0100)]
uct_playout_hook(): Fix board_is_valid_play() parameter order
Petr Baudis [Sun, 12 Dec 2010 23:06:36 +0000 (13 00:06 +0100)]
TD_GAMES: Fix new deadlock when stop-signal and finish-signal would get coalesced
Petr Baudis [Sun, 12 Dec 2010 00:57:09 +0000 (12 01:57 +0100)]
uct_search_stop_early(): Consider cannot-change break only after estimate gets sensible
Frequently, scheduler would actually make us enter this check when
played is still == 0, immediately triggering it.
Petr Baudis [Sat, 11 Dec 2010 22:14:36 +0000 (11 23:14 +0100)]
Merge commit '460bf'
Conflicts:
playout/moggy.c
Petr Baudis [Sat, 11 Dec 2010 20:59:58 +0000 (11 21:59 +0100)]
Autotest show: Color lines based on their status
Petr Baudis [Sat, 11 Dec 2010 20:59:31 +0000 (11 21:59 +0100)]
Autotest worker: Print twogtp invocation for debugging
Petr Baudis [Sat, 11 Dec 2010 19:44:27 +0000 (11 20:44 +0100)]
Merge branch 'master' of git+ssh://repo.or.cz/srv/git/pachi
Petr Baudis [Sat, 11 Dec 2010 19:44:23 +0000 (11 20:44 +0100)]
UCT TreePool: Introduce; replays best significant children during playout
Petr Baudis [Sat, 11 Dec 2010 19:43:16 +0000 (11 20:43 +0100)]
UCT significant node: Make per-color, drop redundant color info
Petr Baudis [Wed, 8 Dec 2010 10:28:27 +0000 (8 11:28 +0100)]
uct_leaf_node(): Add scaffolding for prepolicy/postpolicy hooks
Petr Baudis [Wed, 8 Dec 2010 10:27:28 +0000 (8 11:27 +0100)]
Compilation fix. Oops.
Jean-loup Gailly [Tue, 7 Dec 2010 20:16:38 +0000 (7 21:16 +0100)]
TD_GAMES: Do not crash when pondering is enabled (fix
95e09d4)
Petr Baudis [Tue, 7 Dec 2010 09:52:02 +0000 (7 10:52 +0100)]
Merge branch 'master' of git+ssh://repo.or.cz/srv/git/pachi
Petr Baudis [Tue, 7 Dec 2010 01:57:35 +0000 (7 02:57 +0100)]
README: Add a note on board symmetry, suggested by Francois
Francois van Niekerk [Tue, 7 Dec 2010 01:53:37 +0000 (7 02:53 +0100)]
With TD_GAMES, perform an exact number of playouts
Patch by Francois, some enhancements by Pasky.
Petr Baudis [Mon, 6 Dec 2010 17:36:43 +0000 (6 18:36 +0100)]
UCT descent: Introduce information on last significant node
This will be useful e.g. for pool-rave. The threshold of significancy is
configurable. (The default value comes from the pool RAVE paper.)
Petr Baudis [Mon, 6 Dec 2010 16:48:13 +0000 (6 17:48 +0100)]
Playout: Support for engine hooks altering policy decisions
Petr Baudis [Mon, 6 Dec 2010 12:36:54 +0000 (6 13:36 +0100)]
Playout choose: Take also playout_setup * parameter
Petr Baudis [Sun, 5 Dec 2010 11:47:39 +0000 (5 12:47 +0100)]
joseki19.pdict: Dictionary with lines auto-extracted from Kogo
Petr Baudis [Sun, 5 Dec 2010 11:40:52 +0000 (5 12:40 +0100)]
joseki*.dict: Skip lines starting with #
Jean-loup Gailly [Sun, 5 Dec 2010 01:19:29 +0000 (5 02:19 +0100)]
Merge branch 'derm6'
Jean-loup Gailly [Sun, 5 Dec 2010 00:46:06 +0000 (5 01:46 +0100)]
fbook: Fix for distributed mode
Petr Baudis [Sat, 4 Dec 2010 16:20:53 +0000 (4 17:20 +0100)]
book.dat: Remove, it turns out it is GPLv3 while Pachi is GPLv2
Petr Baudis [Sat, 4 Dec 2010 01:56:48 +0000 (4 02:56 +0100)]
2lib solver: Improve the no-lib-gain test
Petr Baudis [Sat, 4 Dec 2010 00:23:04 +0000 (4 01:23 +0100)]
Merge branch 'master' of git+ssh://repo.or.cz/srv/git/pachi
Petr Baudis [Fri, 3 Dec 2010 16:51:41 +0000 (3 17:51 +0100)]
2lib solver: Do not try hopeless escapes
Petr Baudis [Fri, 3 Dec 2010 03:41:13 +0000 (3 04:41 +0100)]
fbook: In case of multiple candidates, pick a semi-random one instead of the first one
Petr Baudis [Fri, 3 Dec 2010 03:32:23 +0000 (3 04:32 +0100)]
fbook: Properly flip not just sequences, but also the final coordinate
Petr Baudis [Fri, 3 Dec 2010 03:31:38 +0000 (3 04:31 +0100)]
fbook genmove: Add debug print for out-of-fbook too
Petr Baudis [Fri, 3 Dec 2010 03:19:05 +0000 (3 04:19 +0100)]
fbook: Deal with conflicts properly, make the hash size smaller
Petr Baudis [Fri, 3 Dec 2010 02:57:50 +0000 (3 03:57 +0100)]
genmove: Always make sure the timer is running
Doing this in timer_stop_conditions() was just wrong. And in case of
fbook, that simply never gets done.
Petr Baudis [Fri, 3 Dec 2010 01:07:07 +0000 (3 02:07 +0100)]
fbook: Use only at the game beginning (until first mismatch)
Petr Baudis [Fri, 3 Dec 2010 01:00:54 +0000 (3 02:00 +0100)]
fbook: Fix double free of fbook structure
Petr Baudis [Wed, 1 Dec 2010 02:21:30 +0000 (1 03:21 +0100)]
book.dat: Introduce fbook data file, from Fuego r1174
Petr Baudis [Wed, 1 Dec 2010 02:20:03 +0000 (1 03:20 +0100)]
Introduce Fuego-compatible forcing opening book (fbook)
Petr Baudis [Mon, 29 Nov 2010 18:26:39 +0000 (29 19:26 +0100)]
UCT book -> tbook (tree book), added short explanation to HACKING
Petr Baudis [Mon, 29 Nov 2010 17:54:11 +0000 (29 18:54 +0100)]
README: Elaborate command-line usage
Petr Baudis [Mon, 29 Nov 2010 17:45:09 +0000 (29 18:45 +0100)]
uct_state_init(): Major rearrangement of command-line options to topical sections
Petr Baudis [Thu, 25 Nov 2010 02:01:11 +0000 (25 03:01 +0100)]
GTP uct_evaluate: Add a hacker-friendly description
Petr Baudis [Thu, 25 Nov 2010 01:56:45 +0000 (25 02:56 +0100)]
Add support for situation analysis: uct_evaluate
Inspired by Magnus Javerberg.
Jean-loup Gailly [Sun, 21 Nov 2010 12:53:55 +0000 (21 13:53 +0100)]
Uct prior: use cfgd=3%40%20%20 as default, better for usual time settings
Jean-loup Gailly [Fri, 19 Nov 2010 19:02:10 +0000 (19 20:02 +0100)]
Distributed engine: allow shared_levels=0
Petr Baudis [Thu, 18 Nov 2010 00:43:09 +0000 (18 01:43 +0100)]
Merge branch 'master' of ssh://repo.or.cz/srv/git/pachi
Petr Baudis [Thu, 18 Nov 2010 00:42:50 +0000 (18 01:42 +0100)]
UCT prior: Fix cfgd levels parsing
Jean-loup Gailly [Sun, 14 Nov 2010 18:36:35 +0000 (14 19:36 +0100)]
Distributed engine: wait for enough slave replies before early stop.
Petr Baudis [Tue, 9 Nov 2010 17:42:38 +0000 (9 18:42 +0100)]
Moggy mq_tag: Base at 0, not 1 (one extra bit available)
Petr Baudis [Tue, 9 Nov 2010 17:42:06 +0000 (9 18:42 +0100)]
Moggy: More detailed initial description
Petr Baudis [Tue, 9 Nov 2010 11:56:58 +0000 (9 12:56 +0100)]
playout_moggy_fullchoose(): Remove rates, always consider all candidates
Petr Baudis [Mon, 8 Nov 2010 19:37:44 +0000 (8 20:37 +0100)]
pattern.h: Warn that FEAT_PATTERN3 broke
Petr Baudis [Mon, 8 Nov 2010 19:24:58 +0000 (8 20:24 +0100)]
playout_moggy_partchoose() -> playout_moggy_seqchoose()
As in, we sequentially perform various tests.
Petr Baudis [Mon, 25 Oct 2010 22:11:59 +0000 (26 00:11 +0200)]
pat_90rot(): Fix atari info rotation
Spotted by jlg.
Petr Baudis [Mon, 25 Oct 2010 22:11:29 +0000 (26 00:11 +0200)]
Moggy: Add atari information to couple of patterns
I tried adding atari info to all of them, but it is not always beneficial, surprisingly.