Parent Directory | Revision Log
Links to HEAD: | (view) (download) (annotate) |
Sticky Tag: |
Fix error in the implementation of the PIPLOSS input (bug #65035).
Tune cube efficiency ratio for 0- and 1-ply evaluations. Using 0.76 (instead of 0.68) gives a smaller cube error rate at these plies, while 0.68 remains better for deeper evaluations. Moreover, the repartition of errors (premature doubles vs. missed doubles vs. take or pass errors) is now similar to higher plies instead of being mostly premature doubles.
Use g_memdup instead of g_malloc+memcpy
Fix dead stores
Don't offer an optional double if we are certain to lose the game. This can happen is some match situations, if we are certain to lose a gammon and the match. Issue mentionned in rec.games.backgammon by Murat.
Fix crash in evaluations deeper that 4ply
Initialize some variables in a way that allows them to be declared as const
Add missing header file that was previously included indirectly
Remove redundant condition
Tidy up and translate error messages
Expose cache statistics functions only when the are actually implemented
Mark previously missed strings as translatable
Fix [-Wanalyzer-possible-null-dereference] warning
Silence most cppcheck warnings
Remove now unneeded workaround for a bug in early versions of glib2 on Windows
Fix and simplify tie-breaking code. Use it for hints as well as for computer play. Fix related comment in play.c
Use the number of men off as a tie breaker when sorting moves. This avoids some potentially puzzling (but correct) moves in the bearoff, such as the one reported in bug #56316.
Add some branching hints to the inputs calculation code
Remove useless test : j cannot be null here.
Silence compiler warnings
Make check in ApplySubMove stricter.
Check that bearoff databases are of the right type when opening them
Fix memory leaks in some error cases
Tune the number of candidate moves selected by the pruning nets depending on the number of legal moves in the position.
Fix numerical errors in raceBGprob()
Clarify the warning message issued at startup if the 2-sided db is missing.
Declare static functions as such
Declare static functions as such
Minor speed improvement
Small speed improvement
Add Isight count to the Race Theory features
Evaluation outputs outside the [0,1] interval are not inaccuracies to be silently "fixed" by SanityCheck(), they are bugs. Catch them in debugging builds with an assert. Fix noisy evaluations earlier.
Silence compiler warnings
Fix various typos found by codespell
Clarify error messages if weights file cannot be read. Exit instead of continuing without them.
Fix "Invalid argument" problem with g_fopen on Win32. Release 1.05.001
I_TIMING as computed cannot be negative. The final check is unnecessary.
Avoid under-resigning if gammon chances are very small and rounded down to zero
Fix deadlock when analysing hypergammon matches with an executable with multithread support enabled. This (the fix, not the bug :-) applies only if glib version is 2.8 or higher.
Silence compiler warnings
Silence compiler warnings
Silence compiler warnings
Display more detailed evaluations cache statistics
Fix delaying of a sure win by not bearing off as many chequers as possible
Fix wrong test when choosing the best looking among equal moves
Remove dead store
Silence compiler warnings (cast drops const qualifier)
Avoid most of the board swapping overhead during evaluations
Put doublets first in aaRoll array. It makes a difference in evaluation speed.
Put components of the aRoll structure in the order they tend to be used
Use a faster method to find the most significant bit of an integer. This substantially improves the speed of calculation of I_PIPLOSS and related inputs.
Minor speed improvement
Always use SanityCheck on evaluations with noise
Remove some redundant comparisons in live cube computations
Simplify SanityCheck for backgammons in non-contact positions. If there are men in opponent's inner board, the already applied EvalRaceBG is more accurate than what used to be done in SanityCheck. Just ensure there are no backgammons in the other cases.
Don't use SanityCheck() on positions evaluated with the one-sided database
Remove unreachable statements
Silence compiler warnings
Remove unused code
Silence compiler warnings
Remove unused function
Use __builtin_expect hints for a small speed improvement
Fix outdated error message
Fix format string
Displays the weights file and bearoff databases installation directory
Use defined constants instead of numerals
Silence compiler warnings
Small speed improvement
Avoid superfluous copying when evaluating moves selected by the pruning nets
Remove unused function
Fix stack issues that cause AVX Win32 builds to segfault when optimizations are used
Move static .bss thread local data from eval.c to the heap in multithread.c
evallock.c now treated like eval.c when building
AVX bug fix in baseinputs, and general cleanup
Silence compiler warnings
Cleanup SIMD and cputest autoconf processing
If SIMD instructions are compiled in and they are not supported by the processor, output an error message and then terminate immediately.
If SIMD instructions are compiled in and they are not supported by the processor, output an error message and then terminate immediately.
Where appropriate use the term SIMD instead of SSE
Don't use double precision library function to compute power of 2. Part forgotten in r1.417.
Add support for the new AVX instructions introduced by plm
Don't use double precision library function to compute power of 2
Standardized the code formatting with indent -kr -l120 -fc1 -sc -nut -psl
Use 32 bits cache keys instead of long integers
Put default movefilters' definitions in movefilters.inc instead of duplicating them
Use SSE instructions for the pruning nets
Minor speed improvement
Remove superfluous break statements
Remove unused constants
Use named constants as indices instead of magic numbers
Silence compiler warnings
Typo fixed
Fix compiler warning
Improve backgammons estimation in race pruning
Don't display number of training positions. This must be a relic from when nets could be trained from gnubg itself.
Avoid some implicit type conversions (many remain)
Wrong test in evalcontexts comparison
Improve evaluation of late bearoff positions in match play
Add 4ply to predefined levels
Reset locale before reading weights from text file
Make changes to compile nogui/nosse on Win32 platform
Computation of baseInputs() with SSE
erf.inc was removed years ago
Use a defined type for the position key.
Small speed improvement
Remove unused NeuralNetResize()
Small speed improvement
Small speed improvement
Small speed improvement
Handles a different number of hidden nodes with SSE
Bump up evaluation cache size
Minor fixes from Debian and FreeBSD ports.
Fix a counting error in import match code
use rollout and analysis code from mt with --disable-threads
Properly initialize all thread specific data to NULL
Increase score bits in key to allow for 64away-64away (Coincides with max match length of 64)
silence prototype warnings
Remove noise entry from cache key and increase plies size
Remove the (currently unused) reduction evaluation code
Try to speed up single threaded code in multi-threaded build.
declare pm volatile to avoid compiler optimizations
clean up FindBestMovesInEval and fix a prune eval problem caused by bad gcc compiler
Move some functions around (no actual code changes)
Simplify SSE, now it's either supported or it isn't
Fix pointer type warning
Fix compiler warnings
add command to clear cache
remove a debug message and reformat a bit
cmark/hint issues
sort moves according to "back" chequer as a last resort
minor compiler warnings
minor warnings and inconsistencies
Fuzzy compiler warnings
Change cache structure to better reflect the 2-tier lookup being used
minor compiler issues
Minor cache changes
fix the legal destpoints function
add functionality to mark moves and cubes for rollout
allow meaningful rollouts with noise
fix eval problem in hypergammon db
Evaluate accept resign at 2ply
remove unused mman.h
update strings and po files
silence some compiler warnings
replace 6, 15 with nPoints, nChequers
replace fopen with g_fopen
fix some printf format compiler warnings
Minor fixes for MSVC compiler happinnes (mainly adding "f" suffix to decimal number literals etc.) My build environment is still a bit unstable, so I hope I didn't break anything; if you have any problems with this, don't hesitate to revert (or fix, of course).
fix minor translation problems and update po files. remove gnubgi18n.[ch]
First part of translation string cleanup
always mmap database files using g_mapped_file
Small changes for visual studio 2008
#if -> #ifdef
>= should be < in last checkin
fix no, double beaver // double beaver mistake
Tidy up some common gtk functions and remove SIG_IO signal
Change uint to unsigned int
replace uint with unsigned int
*** empty log message ***
Minor misc changes
minor compiler quibbles
Fixes for recent uint changes.
change board type to unsigned int
fix minor uint//int warnings
Minor changes
Small fixes for GLIB_THREADS
Fix const board arrays (for gcc)
Code changes so -ansi -pedantic (just about) works for gcc on windows.
make compilation of utility programs work again(without threads)
rename sgf.l and sgf.y and fix a bunch of minor compiler warnings
Fix build when multithreaded not defined
Misc changes
remove pub_eval
fix messed up evaluation of hypergammon positions
cleanup neuralnet.[ch]
compiler warnings
remove unused training functions.
make sure that files are either static or declared in some header
fix a startup bug when gnubg.wd was missing
*** empty log message ***
win32 fixes after removal of path.c
win32 fixes after removal of path.c
remove path.[ch]
removed egyptian rule
race cleanup
icc compiler fixes
race widget update and analyse menu simplification
fix bug caused by signed/unsigned changes
fix icc compiler warnings
small fixes to compile tools with threads
minor change
changes to make gnubg work with threads on linux
Fix some more concurrency problems
Fix some concurrency problems.
Add multi-thread compatiblity for game analysis.
Made minor changes to code after run through lint code checker
mmap confusion
mingw fix after autools changes
mingw fixes of autotool changes
major autotools cleanup
alloca cleanup
remove meaningless net data from eval window
*** empty log message ***
Fix type in previous check in
Remove gcc compiler warnings (ansi + pedantic)
Remove comipler warnings (mainly typecasts)
keith count
Moved some 3d settings into 3d lib and several compiler fixes
linux cleanup after i18n changes
remove i18n.[ch] and remove gdbm database training
autoconf magick and move of SSE_Support()
Tidy up min/max macros
Fix isCloseCube
3 small fixes from Philippe Michel
Commit patch from Russ Allbery
Fix assembler labels in eval.c to be locals for gcc 3.42
Whoops....
Pacify fussy compilers
Minor changes and new files to make windows building easier
Fix a couple of egyptian rule bugs (14532 + 14427)
Add runtime SSE check
avoid mixing sse and mmap loading of weights
align arrays for sse
Minor speed improvment
Vectorized neural nets
Increase poll time
Remove compiler warning
Hypergammon fixes
MSVC compatibility fix
*** empty log message ***
*** empty log message ***
Fixes based on Neds report
Add evaluation setting interface to prune neural nets
add new sgf versions for pruning, compiler warning fixes, bad bmp file
Minor (non gcc) compiler fixes
add prunning net
Fix bug (navigate to last match crash) - hopefully
* eval.c (EvalInitialise): Suppress compiler warnings.
fix bug leading to wrong classification of wrong doubles. (reported by Hans-JüSchär)
fix bug #9211
Fixes for non-gcc code
Remove compiler warnings
initialise table needed for neural net evaluation
fix bug in calculation of W and L. (reported by Massimiliano.Maini)
fix bugs for "strange" match play scenarios (see 2-ply and 3-ply eval on cAmgACAAGAAA/4HPkAUgzW8EBMA) (reported by Ege Lundgren)
add position ID and match ID to the "eval" output. (suggested by Hugh Sconyers)
ensure "eval" reports cubeless money equituies rather than cubeless equity. (reported by Chris D. Yep)
more constify. move badSkill to format.h
fix resignation bug
Merged 3d board branch
Merged latest head changes
Clear cache when changing MET
Merged latest changes
Initial python match command
Merged latest code
move all Output* functions to new file format.c (change all callers)
fix bug in calculation of cubeless equity. (reported by Ned Cross)
Extendable rollout changes
Initial experimental support of Hugh Sconyers' full 15x15 bearoff database.
do not include cubeful flags, match scores, etc. in calculation of hash key for 0-ply evaluations unless called from EvaluatePositionCubeful.
clean up some compiler warnings add a cast in gnubg.c for iconv() call add (unused) returns in eval.c to satisfy reaching end of non-void function remove unused local variables, comment out unused routines add config.h and backgammon.h in openurl.c, fix bad call to outputerr() Currently builds without warnings using gcc -Wall
fix DumpContact
In EvalInitialise(), corrected error messages when nn weights file version is incorrect (fprintf % arguments were in incorrect order)
bug in DumpContact
New commands to allow modification of cube efficiency (note that the values are *NOT* saved with "save settings").
evaluate top move and actual move on the highest ply if their difference in equity exceeds the given threshold. (suggested by Michael Depreli)
New net
Change how cubeful equities are calculated for money play to match match play. (reported by Robert Eberlein and Michael Depreli)
ensure nReductionGroup is reset to zero. (reported by Michael Depreli)
* Fixed problems which prevented from compiling with PROCESSING_UNITS defined to 0. * Added GDK/GTK thread support; problems still remaining with gtk_main() when not called from main thread. * Redesigned RPU messaging mechanism -- now struct alignment safe and big/little-endian safe. * Started implementing analysis distributed processing. * Added support for hostnames:port syntax when creating remote processing units, eg "myhost.gn u.org:1000"
*** empty log message ***
fix bugs in temp map reported by Ned Cross
more autoconf magic.
Add autoconf magic for multi-processor code.
*** empty log message ***
fix logic for evaluating actual move on nMaxPly (reported by Ian Shaw).
force evaluation for accept=1 and cMoves=1. (reported by Michael Depreli)
Restore incremental NN evaluation
keep track of the max ply used. (reported by Michael Depreli)
fTop *must* be a part of EvalKey... unfortunately we're out of bits... This is just a quick'n'dirty fix! (bug reported by Achim)
fTop *must* be a part of EvalKey... unfortunately we're out of bits... This is just a quick'n'dirty fix! (bug reported by Achim)
rollback
free all bearoff databases, neural nets, and evaluation cache.
small update to code comment (reported by Marek Turski)
use money game cube efficiency for match play for bearoffs.
preparing for hypergammon
Try to stop the never-ending discussion regarding cubeless equities for Jacoby rule and money games with centered cubes.
Better support for variations in SGF, SGG, and TMG.
Rename fNackgammon into a enum bgvariation
Perform the incremental evaluation decision per-net.
Temporarily disable incremental evaluations.
Disable incremental evaluations.
Minor nit-picking changes (mainly making functions static unless they need to be extern).
Bug fixes for cubeful equity caching.
Bug fixes for cubeful equity caching.
New parameter fCubefulEquity, so that cache entries which set OUTPUT_CUBEFUL_EQUITY will not collide with those which do not.
pacify fussy compilers.
change algorithm for accept=0 according to Michael Depreli's suggestions.
cache cubeful evaluations going through EvaluatePositionCubeful3.
Use Michael Depreli's set of predefined filters, and introduce 0 as an acceptable "accept" value.
weights version bumped to 13
Avoid wrong analysis of beavers and raccoons
Fix small bug in yesterday's code.
Generalise reduced evaluations to other than 2-ply, and introduce 25% and 50% reductions.
First iteration of predefined movefilters.
First iteration of predefined movefilters.
Movefilters for rollouts
Separate movefilters for evaluation, analysis, rollouts, and players.
* eval.c: Fix ply number when doing the filters.
Convert code in FindnSaveBestMoves to use movefilters.
change prototype of BearoffCubeful
Add --enable-bearoff configure option to control the 1-sided database.
Remove constructor type initialisation, as it's a gcc specific extension
fix bug leading to wrong equities in bearoffs.
split Path functions from eval.[ch] to separate files path.[ch]
Use cubeful bearoff equities for truncation of rollouts.
Use cubeful bearoff equities in bearoff evaluations.
Don't say "trained on 1 positions".
fix typo. (reported by Paul Selick)
Cosmetics to suppress compiler warnings
pacify fussy compilers.
New race widget with misc counts, one chequer bearoff, and one sided rollout.
Enhanced support for bearoff databases.
* eval.c (raceBGprob): fix a bug in race bg code.
* bearoffgammon.c (getRaceBGprobs): added exact BG race probs for 6 checkers or less inside opponent home.
* eval.c: Use move filters instead of previous hacks.
Implement exact one sided bearoff gammon rate
fix typo. Replace "variance" with "std dev".
fixed incorrect detection of crashed position
New rollout parameters: truncate at bearoff databases.
pacify fussy compilers
Make OSR code compile.
Gammon percentages working with huge OS database.
fix typo
pacify fussy compilers
Initial implementation of huge OSR database.
Initial implementation of huge OSR database.
Rename some of the ratings
Pacify fussy compilers.
Highlight selected move in GTKHint.
adjust threshold for when a cube decision is considered close.
Fix bug that give "Optional double, pass" when it's not.
fix bug introduced in 1.165.
ahem, fix typo
Implement "unforced cube decisions" -- should give more realistic cube and overall ratings.
Implement optional (re)doubles, {pass,take}.
Correct 'show engine' output.
RefreshMoveList: return new ordering.
return "Cube not available" for cube decision NOT_AVAILABLE.
general sort function for moves, changes to ScoreMove, and more
new cache code
remove dead code.
Mark strings for i18n.
locale independency.
change the prototype of eq2mwc and mwc2eq.
changes to evalwidget. new prefined play/search space settings.
new functions. Translates between standard errors in MWC and equity.
Fix bug in bearoff cube evaluations.
* eval.c (NNevalAction): Complete small neural net speedup via an ugly hack.
Add neural net speedup from fibs2html
Make FindBestMovePlied pay attention to fNoOnePlyPrune
Initalise all evalcontexts for new member. Make ShowEvaluation say something about the 1-ply pruning.
Add option to avoid 1-ply pruning of candidate moves.
calculate how many cubes that should be dropped to make incorrect double correct.
support new cube decisions.
new function getCubeDecisionOrdering
remove unused variable.
Cache gammon prices for performance.
fix bug in calculation of gammon prices.
Fix bug in SanityCheck.
remove unused variables in eval.c
Fix typo.
#include "matchid.h"
support for skewed METs.
use post-Crawford table for player 0.
Added PathSearch() function.
Pacify fussy compilers.
move LogCube to positionid.c
Get current gammon rates, and calculate various points (take, double, too good etc) for money and match play.
bugfixes in cmp_evalcontext; define predefined evalcontexts
make GetCubeRecommendation and FindBestCubeDecision external
Reintroduce code which coped if bearoff databases were not present (lost in the merge of the crashed net code).
Show progress indicators during slow operations.
Add a progress bar when generating bearoff database.
Use mmap() on the neural net weights, if possible.
remove REDOUBLE_BEAVER, since it is not possible.
Added StatusCrashed().
don't calculate cubeful equities in Crawford game.
Handle the crashed net in obscure places.
* eval.c: Added new crashed net.
fix logic in FindBestCubeDecision.
Don't use MaxTurns() to calculate the number of turns to bear off if no position database is available.
Detect some certain or impossible wins, gammons and backgammons.
only beaver if equity > -2
Handle interruptions/errors.
Cache evaluations above 0-ply in matches.
Fix typo.
Oops, missing !
Only call Cl2CfMatch* routines when cube is not dead.
New functions. Compare evalsetups, evalcontexts and rolloutcontexts. To be used in analysis to avoid redoing the same analysis.
`eval' and `hint' used different cubeful evaluators. Bug reported by spurs.
Only beavers for beaver-enabled money sessions.
ScoreMove: Convert MWC to EMG for match play.
New parameter to InvertEvaluationR which depends on nMatchTo.
Improvements in DumpPosition and FindCubeDecision.
Cosmetic change: no '+' in output for match equity.
'eval' and 'hint' now handles beavers and distinguishes between initial doubles and redoubles.
Avoid infinite recursion in EvaluatePositionCubeful3().
New improved cubeful code.
Complain if no weights are loaded and -n not specified.
Eliminate fixed PATH_MAX sized arrays.
Optionally disallow more than 5 chequers per point (contributed by Tristan Colgate).
Bugfix: GetDPEq returns MWC for match play. Convert to equity.
Fix stupid bug.
Lots of changes and bugfixes. See ChangeLog for details.
Moved GeneralEvaluation, GeneralCubeDecision, etc. from eval.c to rollout.c.
- use the new better/faster EvaluatePositionCubeful2. - new functions GeneralEvaluation, GeneralEvaluationE, GeneralEvaluationR, GeneralCubeDecision, GeneralCubeDecisionE, GeneralCubeDecisionR - evaltype moved into evalsetup
- EvaluatePositionCubeful2: new function (initial implementation of cubeful projections as suggested by David). The function is not used anywhere yet. - FindCubeDecision: new function (basically a wrapper for EvaluatePositionCubeful). - define rolloutcontext and redefine evalsetup
Verify the checksum of the bearoff database.
Make sure we used unsigned ints for the bearoff probabilities.
Report weights version.
More work on analysis.
fix bug in the calculation of gammon chances and fix stupid copy'n'paste error (reported by David Montgomery)
Pacify fussy compilers.
Added optional noise in evaluations.
Add David's reduced 2-ply search code.
Include rScore2 when comparing moves (for cubeful evaluations).
Pacify fussy compilers.
reinstate accidentally deleted code
net version 11
Added "show engine" and "show version" commands.
Use binary mode on non-POSIX systems.
Minor changes to pacify gcc -pedantic.
Allow annealing rate to be specified.
Added "set training alpha" and "set training anneal" commands.
Pay attention to the user-specified net size.
"new weights" command added.
Check fAction and fInterrupt in FindBestMovePlied.
Removed unused definitions.
Check fAction in the cubeful evaluation functions.
Give better error messages about bad weights files.
EvaluatePositionCubeful1: Fixed typo (fix of bug noted by Steve Baedke).
Added fCheckLegal parameter to ApplyMove().
Added remaining partial derivative outputs.
Report evaluations specified with =n notation from player's perspective.
Global game state (e.g. nCube, fCubeOwner, anScore) removed from eval.c; this information must now be given as part of the cubeinfo parameter. All callers changed.
Move random number generators into lib/, and configure everything from a single configure script at the top level.
- EvaluatePositionCubeful now returns cubeless evaluations as well. - ScoreMoves: save cubeless evaluation - FindnSaveBestMoves: sort moves for 0-ply evaluation as well.
Correct small problem that could lead to numerical problems.
New functions SetCubeInfoMoney and SetCubeInfoMatch.
Remove the fRelativeAccuracy code.
Handle hints when there are no legal moves.
Change code so it compiles without warnings.
Removed debug printf.
SetCubeInfo: bugfix: reset negative(!) gammon price to zero. Gammon prices can be negative due to numerical instabilities (FIXME: think of a stable way to calculate them)
- FormatEval: FormatEval5 renamed to FormatEval; now returns a variable string describing the evaluation type - and not some junk text :-) - EvaluatePositionCubeful1: add new parameter that controls whether we should check for automatic redoubles or not (this is a bugfix for `cubeful' evaluations in the crawford game) - ScoresMoves: Implement cubeful checkerplay - ScoresMoves: invert fMove in cubeinfo. Before this bugfix GNU used the wrong gammon prices, which gives wrong equities for all scores where anScore[ 1 ] != anScore[ 2 ]. ----------------------------------------------------------------------
Show progress indicator while building database.
- FindnSaveBestMoves: new function that find best moves, but save evaluation for all moves (used for analysis). - Cl2CfMatchCentered: new function to calculate cube centered cubeful equities for match play. - FormatEval5: return a 5 character string of the evaluation type, for example: `2-ply' or 'R'.
If the bearoff database isn't found, generate an approximate one on the fly using Eric Groleau's method.
Bugfix in Cl2CfMoney: too good to double with centered cube did not work.
- corrected bugs introduced in eval.c 1.49 - renamed match equity table (MET) from aafA1 to aafMET. - shamelessly ripped METs from Joseph's fibs2html.
Don't calculate cubeful equity when cube is dead (or score is -2, -2).
Record moves made by pubeval.
Minor changes to avoid compiler warnings.
redo changes lost since 1.42
Bugfix: missing 'pci ->' s.
- remove unused debug code in DumpPosition. - only cache 0-ply evaluations for match play, as `deeper' moves might have been played wrong at current matchscore. - add basic code for calculating cube efficiency.
Implemented cubeless to cubeful formulae for match play. Doubling strategy is now working for match play.
* eval.c (EvalBearoff1Full): Fix evaluation of bearoff positions where gammons are possible. The former use of the race net is not appropriate, since the race net has not been trained on such positions.
Bugfixes: (1) Zero *prOutput (2) Don't recurse in EvaluatePositionCubeful1 if cube action is pass (3) Recurse with the correct cubeinfo struct.
Bugfix: corrected some typos, and initialize fDoubleBranch. Using an un-initialized value often led to wrong evaluations.
Fixed bug in GetDoublePointDeadCube.
CommandHint gives hints for cube action.
Removed debug printf's.
Implemented double strategy for money game. Cubeful equities are based on the simple version of Rick Janowski's formula.
* neuralnet.c: Improve sigmoid accuracy by using a table lookup for the first two digits and using (exp(x) = 1 + x) for x < 0.1. * eval.c: Create new race net. See race inputs in file.
Fix positionclass bug.
Improve DumpBearoff1 output.
Add --no-bearoff (-b) option.
Show output from commands in windows when using GTK.
Added missing member to ecBasic.
Add more information to the hash key.
Added "set (eval) consistency" command, to force the same evaluator to be used for all moves.
Improve doubling code. Deduct some equity if we don't double. It is possible to play around with rDelta rDelta = 0.9 gives very small errors for the Robertie 18-27 test set.
Fixed typo.
Tidying code. Removing debug printf's.
Implemented new (better) double heuristics: Don't take double branch if our winning chance is below the dead cube doubling point.
- Corrected some silly copy'n'paste bug in eval.c - Implemented some basic heuristics - EvalCubeful now takes a flag that indicated which paths to take: No-double and/or Double. - Don't call EvalCubeful unless we're in upper 1/3 of market window. - Don't take no-double path when deciding whether to take a cube or not.
A few bugfixes.
cubeful code upgraded.
New doubling code. Still some problems to solve.
new inputs to net
Implementation of reduced 2+ ply searches.
Use binary mode for gnubg.wd, to avoid problems on non-POSIX systems.
Add partial derivative code.
Fix stupid syntax error.
Calculate alpha correctly.
new inputs. disable BPG
Change net inputs slightly.
More comments on net inputs.
Add comments about net inputs.
Bugfix. Corrected typo.
Add --datadir option.
Fix bugs bearing off in board window.
set a minimum tolerance for ply 0
Better model for cube owner ship implemented.
Bug fix in barPrimeBackGame
Fixed the post-Crawford bug (I think).
"Finished" version. Still a problem with 2-away, n-away scores. Also check post-Crawford doubles.
Better code for match play. Stills some errors/flaws, especially at 2-away, n-away.
Minor updates.
Updated all files to newest files on main trunk.
New faster code (I hope) for EvaluatePositionCubeful. Still a few FIXMEs for match play.
New cubeful code used for doubling strategy. Still missing is: * automatic redoubles, dead cubes etc in Cubeful evaluations. * correct take points in Cubeful evaluations. However, the code should be correct for money game.
Handle X events asynchronously.
Reorganise user settings for evaluations and rollouts.
Add a third net for special cases.
Allow the user to modify search candidates and tolerance, and add Nackgammon option.
Added code for cubeful evaluations. So far only called from CommandEval.
Changed code allowed different moves in the no-double and double cases.
Better error handling reading weights files, backgammons can\'t exceed gammons, and more detailed DumpBearoff1 output.
Doubling strategy for money games.
Call Utility for calculation of equity.
Update moving code in X board window
Added support for binary weight files
Small fix in SanityCheck
Update X position ID after every move. Add "set cache". Calculate rollout standard deviations correctly. Fix freeing bug in "new game".
Allow saving matches.
Added "set prompt" command.
Added Joseph Heled's C++ race evaluation code.
Imported old source from gnubg.sourceforge.net
Initial revision
This form allows you to request diffs between any two revisions of this file. For each of the two "sides" of the diff, select a symbolic revision name using the selection box, or choose 'Use Text Field' and enter a numeric revision.
savannah-hackers-public@gnu.org | ViewVC Help |
Powered by ViewVC 1.1.26 |