GREKO CHESS ENGINE (c) 2002-2021 Vladimir Medvedev GreKo 2021.12 31-Dec-2021 ------------- - More parameters per each piece-square table. - Trained on 100 K games of version 2021.08 at 7-ply depth. Test tournament, time control 60+1: Engine Score Gr Gr Gr Gr S-B 1: GreKo 2021.12 34.0/60 ···················· 10110000=0=0=0100111 ==0111========01=1=1 =110101=11=1101=101= 969.25 2: GreKo 2018.08 29.5/60 01001111=1=1=1011000 ···················· 00110=1=010=011=00== 0==0=100=0111==10001 899.50 3: GreKo 2021.08 29.5/60 ==1000========10=0=0 11001=0=101=100=11== ···················· ===11=110=011==000== 880.00 4: GreKo 2021.10 27.0/60 =001010=00=0010=010= 1==1=011=1000==01110 ===00=001=100==111== ···················· 825.75 Estimated playing strength: +30...40 Elo points over 2021.08. GreKo 2021.10 22-Oct-2021 ------------- - Fixed "go nodes" and "go depth" UCI commands. - Diminished number of weights in eval from 1912 to 760. - New format of intermediate learning files. - Added temporal-difference error discount in learning formula. - Evaluation tuning. GreKo 2021.08 NNUE 15-Aug-2021 ------------------ Experimental version with NNUE evaluation code from Stockfish. Using Daniel Shawul's nnue-probe code from https://github.com/dshawul/nnue-probe. Network file: nn-04cf2b4ed1da.nnue from https://tests.stockfishchess.org/nns. Estimated playing strength: 3000 CCRL in 1-CPU mode, 3050-3100 in 4-CPU mode. GreKo 2021.08 01-Aug-2021 ------------- - Improved stability in SMP search. - Maximum number of threads increased to 16. - Added "seldepth" and "hashfull" UCI info. - Fixed bug under Chessbase GUI with reject to start search in some theoretically drawn positions. - Added MultiPV UCI mode. - Big code cleanup and refactoring. GreKo PSQ 2.0 MP 30-June-2021 ---------------- - Merged Lazy SMP changes from GreKo 2021.06 MP. - Weight file tuned. GreKo 2021.06 MP 30-June-2021 ---------------- - Refactoring of search data structures. - Added Lazy SMP parallel search in a very basic form. Max. number of threads = 4, default value = 1. Number of threads can be configured with UCI option "Threads" or with command line parameter: GreKo.exe -threads 4 GreKo.exe -t 4 Match of 4-CPU vs. 1-CPU version, time control 60+1: 67-33, +99 Elo points, LOS = 99% GreKo PSQ 2.0 24-June-2021 ------------- Test match with PSQ 1.2, time control 3+.2: 161:139, +23 Elo points - Merged changes from full eval version GreKo 2021.06. - Weights file updated. GreKo 2021.06 01-June-2021 ------------- Test match with 2021.05, time control 3+.2: +338-260=402, +25 Elo points - Use compiler intrinsics for bitboard operations. - Use 64-bit int internal representation of time. - Use 16-bit float internal representation of eval. - Windows compiler update: TDM/MinGW GCC 9.2.0. - Added solution and project files for MSVC 2019. - More precise time measurement in Linux. - Tuned parameter for null-move reduction. - Re-implemented easy levels with fixed search speed. Easy levels can be turned on by UCI option "Strength" or by command line parameter: GreKo.exe -s GreKo.exe -strength where level is [0...100]. Estimated speed and strength: s knps rating 0 0.1 1500 10 0.25 1600 20 0.63 1700 30 1.58 1800 40 3.98 1900 50 10 2200 60 25 2400 70 63 2500 80 158 2600 90 398 2675 99 912 2700 100 full strength GreKo 2021.05 15-May-2021 ------------- Numerous changes in search and evaluation. Main differences with 2020.03: - Refactored and simplified code for search. Use the same function for all node types. - More parameters for evaluation weights. Use tabulated values instead of quadratic approximations. - Optimization in hash key calculation. Pawn hash key now is a part of main hash key. - Use simpler time management algorithm. - Use stochastic gradient descent as a single training method, instead of mix with coordinate descent before. - Trained from pure material weights in 1.6 M positions from self-play games, fixed depth = 5. Estimated playing strength is +20...+50 Elo points over 2020.03, depending on time settings. The slower time control the stronger. GreKo PSQ 1.1 13-Apr-2021 ------------- - Fixed bug in formula for gradient in SGD learning algorithm. - Trained on 200 K self-play games at fixed depth = 5. Playing strength is ~50...100 Elo stronger than PSQ 1.0. GreKo PSQ 1.0 01-Apr-2021 ------------- Version with evaluation function limited to material + piece/square tables. Playing strength is 2500-2550 CCRL (200 points weaker than 2020.03). Trained almost from zero (pure material weights 100-300-300-500-900) in self-play: fixed depth = 3 ply, ~700 K games in 8 hours. Also contains a lot of refactoring changes in search, learning and time management. GreKo 2020.03 31-Mar-2020 ------------- This is mostly a refactoring version. No significant changes in strength. - Fixed Rook mobility code in features generator for eval tuner. - Small optimization in Knight mobility code (use table instead of counting bits). - Refactoring in eval tuner. GreKo 2020.01 ------------- 31-Jan-2020 A lot of changes in search, evaluation and built-in evaluation tuner. Most significant are: - added stochastic gradient descent algorithm for learning - more aggressive pruning: dynamic null move and LMR reduction values - separate search functions for PV and non-PV nodes - search extensions modified - new evaluation terms for piece combinations and attacks - use quadratic approximation for almost all evaluation features Estimated playing strength: +20 Elo points to version 2018.08. GreKo 2018.08 ------------- 22-Aug-2018 (fix) * Endgame code fixed and modified for KNBK, KNNK cases. (chg) * Search modifications for nullmove hashing, IID, checks in q-search. Blitz, 60+1 Score vs. GreKo 2018.07: +39-28=53 (+29 Elo). GreKo 2018.07 ------------- 21-July-2018 (new) * New terms in evaluation. (chg) * More frequent LMR. Blitz, 60+1 Score vs. GreKo 2018.02: +47-20=33 (+89 Elo). GreKo 2018.06 ------------- 30-June-2018 (fix) * Fixed bug in checks generator for qsearch. (fix) * Fixed bug with incorrect weights file format. (chg) * Small changes in selfplay and learning built-in utils. GreKo 2018.05 ------------- 25-May-2018 (fix) * Fixed bug in selfplay mode. (new) * More evaluation terms: material, piece-squares, passed pawns, king safety. (new) * All evaluation weights bootstrapped from random values with selfplay. GreKo 2018.02 ------------- 27-Feb-2018 (fix) * Do not crash when running for FEN with incorrect castling rights. (chg) * Time management improved, especially for "N moves in M minutes" mode. (chg) * Evaluation tuning. GreKo 2018.01 ------------- 31-Jan-2018 (fix) * Hashtable fix: do not record score returned from unfinished iteration. (chg) * Evaluation terms for passed pawns modified. (chg) * Time management formula modified. Better for very short time controls. (old) * Variable strength is back. This version supports reduction of strength by adjustement of search speed. You can use UCI parameter "Strength" or run GreKo with a command line parameter: GreKo.exe -s where level is an integer number from 0 to 100: Level Knps Rating ---------------------------- 0 0.1 1600 20 1 1800 40 10 2000 60 100 2200 80 1000 2400 100 10000 2600 You can use any intermediate value for level, speed of search will be interpolated. GreKo 2017 ----------- 08-Jan-2018 *** HAPPY NEW YEAR!!! *** (new) * Rewritten almost from scratch. Code cleanup. (new) * Evaluation function redesigned. (new) * Trained on CCRL games collection (700K+ games) and self-play matches. (chg) * Time management formula changed. This version is mostly for experiments with evaluation function tuning. It doesn't have some features from previous versions (opening book, multipv, variable strength). Estimated rating 2700-2750 on CCRL scale. To set hash size in WB mode use command line argument: GreKo.exe -h GreKo 2016 ----------- 31-Dec-2016 *** HAPPY NEW YEAR!!! *** (chg) * Search changes: tuning LMR and Nullmove reduction formulas. (chg) * New terms in evaluation: piece-to-King tropism, strong and weak fields, etc. (chg) * UCI handles for board control, mobility, pawns etc. replaced with single parameter: PositionalKnowledge. (chg) * Evaluation weights auto-tuned in self-play matches. (chg) * Source code clean-up: removed unnecessary pseudo-OOP classes Search, Bitboard and Engine. More procedural style of code: less verbose, easier to read and type. (chg) * Fixed compiler warnings in gcc build. GreKo 2015 ML ------------- (new) * Added build-it command "learn" for learning from PGN file. (chg) * Evaluation weights tuned, based on self-play. This is yet another final release. All search and evaluation algorithms are the same as in GreKo 2015. ML stands for "machine learning". This is the first time automated evaluation tuning is used in GreKo. GreKo 2015 ML implements algorithm which is very similar to Texel's Tuning Method (https://chessprogramming.wikispaces.com/Texel's+Tuning+Method). Main difference is using fast eval() function instead of expensive qsearch() for making predictions of game results. GreKo's evaluation is parameterized with 27 weight coefficients with the following default values: 1. VAL_P = 100 2. VAL_N = 400 3. VAL_B = 400 4. VAL_R = 600 5. VAL_Q = 1200 6. PawnDoubled = -10 7. PawnIsolated = -19 8. PawnBackwards = -5 9. PawnCenter = 9 10. PawnPassedFreeMax = 128 11. PawnPassedBlockedMax = 101 12. PawnPassedKingDist = 9 13. PawnPassedSquare = 200 14. KnightCenter = 27 15. KnightOutpost = 7 16. KnightMobility = 19 17. BishopPairMidgame = 20 18. BishopPairEndgame = 95 19. BishopCenter = 9 20. BishopMobility = 72 21. Rook7th = 24 22. RookOpen = 17 23. RookMobility = 40 24. QueenKingTropism = 99 25. KingCenterMid = -41 26. KingCenterEnd = 33 27. KingPawnShield = 120 These numbers are stored in file weights.txt. If this file doesn't exist it is created with default values when program starts. To begin learning in console mode please execute the following command: learn [minIndex] [maxIndex] where: games_file - path to PGN file, with or without .pgn extension (for example, for database.pgn you can type "learn database") minIndex - low bound of learning interval (optional, 1...27, default = 6) maxIndex - high bound of learning interval (optional, 1...27, default = 27) If minIndex and maxIndex are not specified learning will affect only positional subset of weights, point values remain unchanged. Tuning weights of pieces usually takes longer time and does not always increase playing strength. On the first execution of learn command for a particular PGN file the program creates FEN file with the same name (database.fen in our example). It contains FEN notation for positions to learn from (not in check, not after capture) plus some additional information (game result, ply, move played). On the next run only this file is used. If you have changed your PGN file and want to repeat learning please delete FEN file first. During learning some information is printed to console (with copy in file learning.log): number of iteration, current values of weight, loss function on training and validation datasets. After learning new weights are in weights.txt, the original set is in weights.old file. For files with ~10 K games learning usually takes 1-3 hours on a standard PC. The function was tested with volumes up to 100 K games. Larger files can cause an out of memory error. Intermediate values of coefficients are always in weights.txt, which can prevent data loss in case of unexpected process termination. Learning algorithm is randomized, on each run the whole dataset from FEN file is randomly divided into training (80%) and validation (20%) subsets. This can lead to slightly different final vectors of weights after different runs. GreKo 2015 ---------- 18-Nov-2015 GreKo 2015 is the final release of the third "major" GreKo series, which started in 2006 with GreKo 5.0. No future changes, improvements or fixes (even critical) are planned on this codebase now. It is essentially GreKo 12.8, the strongest version in CCRL rating list so far, with two bugfixes: (fix) * Fixed possible bug in conditions for KBP-K endgame recognition. (fix) * Mobility table for knights - fixed asymmetry. Version 13.1 ------------ 07-Oct-2015 (fix) * Fixed possible bug in conditions for KBP-K endgame recognition. (new) * Different weights for pawns in opening/middlegame/endgame: 80...120. Version 13.0 ------------ 01-Oct-2015 (new) * Penalty for pawn islands in endgame. (chg) * Default lazy eval margin changed to 250 centipawns. (chg) * Mobility table for knights - fixed asymmetry. Version 12.9 ------------ 01-Aug-2015 (chg) * Better time management for time controls like "N moves in M minutes". (chg) * Separate configurable null-move reduction value for non-PV and PV. Default R=3 and R=0, which means not using null-move on PV. Version 12.8a -------------- 15-July-2015 (chg) * Changed time management for zero-increment time controls. Version 12.8 ------------ 19-May-2015 (new) * Added UCI option QChecks: generate checks in qserch when qply < QChecks. Default value = 2. (chg) * Do not use LMR in late endgame and when move is a killer move. (chg) * Small changes in search extensions logic. Version 12.6 ------------ 25-Apr-2015 (chg) * Default weight for material changed from 60 to 50 percent. (chg) * Default lazy eval margin changed from 150 to 200 centipawns. (chg) * Small optimizations in moves sorting: less swaps, return hash move immediately. Version 12.5 ------------ 21-Dec-2014 (fix) * Case insensitive parsing for UCI "setoption" command, in line with UCI protocol description. (chg) * Time management formula changed, providing much stronger play at superfast controls. (chg) * Replaced std::vector with fixed-syze arrays for storing principal variations in search. (chg) * Recapture extensions added. (chg) * Extensions now are limited with the rule: ply < 2 * (current iteration number). (chg) * LMR if turned off in late endgames. (chg) * Code reformatting here and there. Version 12.1 ------------ 07-Oct-2014 (fix) * Linux build: fixed missed typedef in types.h. (fix) * Linux build: fixed compiler warnings. (fix) * Linux build: fixed time measurement in perft. (chg) * Evaluation: scale bishops pair between midgame and endgame. About +15 Elo in self-play after 600 games. Version 12.0 ------------ 1-Apr-2014 *** JUBILEE EDITION *** Change in the version number marks 12-year anniversary of GreKo :) (new) * Added internal iterative deepening code (IID search). (chg) * Late moves reduction logic changed (count quiet moves only). (chg) * Default evaluation weights modified. Version 10.3.1 -------------- 16-Mar-2014 This is a bugfix release. No changes in playing strength or style expected. (fix) * Typedefs for Windows build fixed. (fix) * Fixed incorrect draw claim in "King + minor vs. King + minor" endgames. (chg) * Evaluation size changed from 32 to 16 bits. (chg) * Hash entry size changed from 24 to 16 bytes. (chg) * Small protocol fix for Xboard (start in UCI mode). (chg) * "console" command (for debug in command line). Version 10.3 ------------ 31-Dec-2013 *** HAPPY NEW YEAR !!! *** (fix) * Engine output in Xboard mode modified. (fix) * Time management improved. Do not leave much unused time in games with increment. (chg) * Lazy eval margin reduced from 400 to 150 centipawns (as before v10.2). (chg) * Non-linear scale for passed pawns, mobility, king safety. (chg) * Different scoring for blocked and free passed pawns. (chg) * New formula for adjustable speed/rating -- thanks to Alexander Schmidt. (chg) * Default UCI_ELO changed to 2800. (chg) * Source code reformatted a bit. Version 10.2 ------------ 30-June-2013 (new) * Draw score parameter added (sort of contempt factor). (chg) * More configurable parameters for search and evaluation. (chg) * Lazy eval margin increased from 150 to 400 centipawns. Version 10.0 ------------ 01-Mar-2013 (fix) * Change in function signature in unix.cpp, to fix Unix build. (old) * Rollback to search/evaluation from v9.7, which is stronger than v9.8. (new) * Nullmove depth condition changed from (depth > 2) to (depth > 1). (chg) * Source code reformatted for better readability. Version 9.8 ----------- 31-Dec-2012 (new) * Randomizing eval UCI/config parameter (just for fun). (chg) * Evaluation: PSQ-tables changed, bonus for moving forward. (chg) * SEE and MVV/LVA code updated, more clean design. (chg) * Small tweaks in search. (chg) * Mate threats extensions. (chg) * Bitboards code reformatted. Version 9.7 ----------- 31-Aug-2012 (chg) * Endgame heuristics for KBP-K and opposite color bishops. (chg) * Depth limits for null-move and LMR modified. (chg) * Hash overwriting rule modified (simpler). Version 9.6 ----------- 05-Aug-2012 (chg) * Return to scheme with one killer move per ply (plays better). (chg) * LMR: history condition restored. (chg) * UCI parameters for evaluation modified. (chg) * Elo/Knps formula modified: play weaker at handicap levels. (chg) * Code cleanup. Return to C++98 Standard. Version 9.5 ----------- 24-June-2012 (fix) * Config file parameter names fixed. (new) * Evaluation weights: config and UCI options. (chg) * LMR conditions modified. Version 9.3.1 ------------- 15-June-2012 (fix) * 'position fen ...' UCI command fixed. Version 9.3 ----------- 12-June-2012 (fix) * Fixed pv output in UCI mode. (fix) * Claim draw after 3rd repetition. (new) * 'uci go depth ...' option added. (chg) * Checks in qsearch code modified. Version 9.2 ----------- 15-May-2012 (chg) * Source code refactoring. No changes in search/eval. (chg) * New formula for Elo/Knps settings. Version 9.0 ----------- 31-Dec-2011 (chg) * MVV/LVA instead of SEE in sorting moves (works much faster). (chg) * Small optimization for sorting procedure in qsearch. (chg) * Conditions for LMR modified. (chg) * Null move on PV granted. (chg) * Using std::vector for game history and undo information. Version 8.2 ----------- 01-Aug-2011 (chg) * Search code refactoring: global functions and variables converted to class members. (chg) * History counters scheme changed. (chg) * Use history condition in LMR. (chg) * New terms in evaluation: - king-pawn distance in endgame, - knight mobility, - trapped bishop. (chg) * Evaluation weights tuned. (chg) * Small optimization in hash table. (chg) * Hash overwriting logic changed. Version 8.1.1 ------------- 08-July-2011 (fix) * Fixed bug in opening book lookup code: added check for moves legality. Version 8.1 ----------- 05-July-2011 (chg) * Futility code optimized. (chg) * Fixed a typo in evaluation. (chg) * Small code refactoring. Version 8.0 ----------- 31-Dec-2010 (new) * Mating moves in sort ordering. (chg) * LMR conditions changed; verification when LMR score exceeds alpha. (chg) * Evaluation function completely rewritten. Version 7.2 ----------- 05-Aug-2010 (chg) * Returned to pseudo-legal move generator. Faster, simpler. (chg) * Single reply extensions removed. (chg) * Late move reduction in a very basic form. No verification re-search. (chg) * Limited strength levels tuned. Version 7.1 ----------- 19-July-2010 (new) * New options: UCI_LimitStrength, UCI_Elo and LimitKNPS. (chg) * Time management tuned. Version 7.0 ----------- 21-June-2010 (chg) * Search modified. Version 6.5 ----------- 01-June-2009 (new) * Fixed nodes and fixed depth modes, under UCI and WB. (new) * Evaluation weights now can be set with UCI 'setoption' command. (chg) * Removed not used bitboard masks. (chg) * Time allocation algorithm changed. (chg) * Command parser reorganized. (chg) * Removed many terms in eval. Modified some other. Added TEMPO term. (chg) * Futility conditions changed. (chg) * Delta pruning conditions changed. (chg) * Nodes counters changed to int64. (chg) * Overall code cleanup and reformatting. (chg) * Makefile simplified. Version 6.4 ----------- 14-May-2009 (fix) * Single reply extensions restricted to PV-nodes only. Prevents from search explosion in some won positions. (new) * GreKo.ini file added, with evaluation weights - sort of "personality settings". Not tuned at all. (new) * Makefile.linux added. (chg) * Command line parameters (-hash and -log) are not used anymore. Moved to GreKo.ini file. (chg) * Code clean-up and reformatting - a lot of obsolete code removed or changed. Version 6.31 ------------ 29-Apr-2009 (chg) * Removed non-used field "flags" in hash entry, effective hash capacity increased. (chg) * Hash allocation simplified. (chg) * Code clean-up and reformatting. Version 6.30 ------------ 05-Feb-2009 A lot of clean-up work without significant changes in core algorithms. (chg) * Small change in CountBits() function. (chg) * Time management scheme simplified. (chg) * Sort values increased, to make less history scalings. (chg) * Adjust mate scores in hash only for real mates. (chg) * Pruning conditions changed for nullmove, futility, LMR and delta pruning. (chg) * Removed some evaluation terms. (chg) * Mobility evaluation changed. (chg) * Passed pawns evaluation changed. (chg) * Weights tuned a little. Version 6.25 ------------ 15-Dec-2008 (fix) * Fixed bug with "..." in UCI mode under Arena. (fix) * Fixed: do not leave search with empty root pv. (chg) * Faster search: do not copy pv in beta nodes. (chg) * Turned off nullmove and LMR on PV. (chg) * Generate checks in qsearch only on qply == 0. (chg) * SEE delta pruning in qsearch on qply > 1. (chg) * Small changes in code formatting. (chg) * Small changes in popcount function. Version 6.1 ----------- 24-Oct-2008 (fix) * Fixed "book import" and "book load" commands to make correct book merges. (fix) * Fixed setup from FEN: correct hash for positions with black to move. (fix) * Fixed short notation recognizer: now accepts "O-O" and "O-O-O". (chg) * Pawn hash is now a separate 32-bit field, not a part of the main 64-bit hash. (chg) * New opening book (not compatible with 6.0 book). (chg) * Time management scheme tuned a little. (chg) * Small changes in search (qchecks, pvs, extensions). (chg) * Small speed-up in evaluation. Version 6.0 ----------- 29-Sept-2008 (fix) * Fixed bugs in move generator (en-passant) and incremental check detector (en-passant, castlings); all of them could lead to making illegal moves. (new) * Opening book: internal binary format and PGN-like text files supported. (chg) * More compact magic lookup tables (nothing special, just conventional variable bits per square approach). (chg) * More reliable search. (chg) * Short notation support in command line. When GreKo starts, it reads book.bin file. If there is no book.bin, it tries to read book.txt. You can create your own book.txt, or import games from PGN and save them as book.bin. Source PGN file should not contain comments or variations. How to create book (example): White(1): book clean White(1): book import file1.pgn 40 w maxPly = 40 white's moves only Games: 1193, nodes: 22730 White(1): book import file2.pgn 40 b maxPly = 40 black's moves only Games: 1157, nodes: 44409 White(1): book import file3.pgn maxPly = 20 Games: 1919, nodes: 52633 White(1): book save book.bin writing book.bin... White(1): bk e4 54%, d4 33%, Nf3 8%, c4 2%, g3 0%, Nc3 0%, b3 0% Version 5.9 ----------- 02-Aug-2008 (fix) * Fix in reporting NPS under UCI (fix) * Check extensions fixed (chg) * Make the move instantly when checkmate is found (chg) * Draw recognizer modified (chg) * Bishop and bishop pair values tuned (chg) * Small changes in search Version 5.8.5 ------------- 16-July-2008 Changes in move generator and in search: (new) * Magic bitboards (can be turned off when MAGIC is not #defined in bitboards.h). (new) * Legal-only move generation on all stages. At last GreKo can cheat in perft by not actually making moves at depth = 1 :) (new) * Simple checks generation on first levels in qsearch. (new) * Static exchange evaluation for captures (with batteries, without pins and promotions on ply > 0). (chg) * Pruning decisions do not rely on pv/non-pv node type, again. (chg) * Futility pruning turned off, again. (chg) * Use SEE score for delta pruning in qsearch instead of VALUE[captured]. (fix) * Fixed incorrect PGN disambiguation code (introduced in 5.8, but does not affect playing games, just in epdtest mode) -- thanks to Dann Corbit for bug report. Tactical abilities seem to increase: WACNew WACNew NewIQ 1 sec. 5 sec. 10 sec. 5.7.1 271 287 243 5.8 270 286 239 (-1...-4 solved positions - due to PGN bug) 5.8.5 290 297 276 This version could be named 5.9 since it has so many changes, but I prefer to save numbering space for some possible improvements. Version 5.8 ----------- 10-July-2008 The main goal of this version is source code refactoring. Core datastructures are now more "C++ style". Position and Move data encapsulated in classes. Got rid of pointers, some #defines etc. Small optimization for all bitscan loops. As a result the source code is 5% more compact than 5.7.1. Perft and search run 20-30% faster (at least in the initial position). Small changes in the move generator (castlings). Search and evaluation algorithms are almost unchanged except for a little fix in the hashing scheme. En-passant and castling rights were not taken into account in the hash code, now they are. Time management scheme changed a little, to prevent possible losses on time in 3+3 games. Thanks to Volker Pittlik for mentioning the bug. Multi-PV UCI mode modified, however it is still slow and non-optimized. Version 5.7.1 ------------- 25-May-2008 (fix) Fixed bug in PVS code. Version 5.7 ----------- 19-May-2008 (fix) Mate scores in hash now handled correctly (fix) Do not skip draw scores in hash (fix) PV update code: replaced "for" statement with "memcpy" call (chg) Time management tuned a little (chg) Source code reformatted (new) MSVC8 project file added Version 5.6 ----------- 12-Apr-2008 (fix) Fix in MVV/LVA sorting (fix) Fix in "stop" command processing - don't report it is incorrect (new) Futility pruning at non-PV nodes, depth = 1 (new) Delta pruning in qsearch (chg) Replaced get_new_depth() function with more clear code (chg) Principal variation search (PVS) instead of plain alpha-beta (chg) Nullmove conditions changed (R = 4, etc.) (chg) LMR conditions changed (chg) Fifty moves draw condition - after checkmate detection (chg) Reply to checks in qsearch (coming back) (chg) Time management scheme changed (use more time in increment games) (chg) Trapped bishop evaluation changed (chg) Repetition handling changed Version 5.5 ----------- 31-Dec-2007 *** HAPPY NEW YEAR!!! *** (chg) Hash lock 32-bit instead of 64. (chg) New hash allocation scheme. (chg) Futility pruning turned off. Version 5.4 24-Jul-2007 ----------- (chg) Minor change in storing generated moves. (chg) Recompiled as C++ code with MSVC 2005 Express (Windows) and g++ 3.3.3 (Linux). Version 5.3 20-May-2007 ----------- (chg) Using hash eval from previous iteration for pruning (chg) Small optimization in attack detector (chg) Small optimization in PV handling (fix) Fixed crash on epd-files ended with empty lines Version 5.2.5 23-Feb-2007 ------------- (chg) Adjust checkmate scores in hash (chg) Minor changes in evaluation (mobility, bishops pair etc.) Version 5.2 12-Feb-2007 ----------- (fix) Main hash table size is not divided by 2 anymore. (new) Added futility pruning, in a very simple and straightforward form. (chg) Search: tuned some conditions for LMR and Nullmove. (chg) Eval: King safety value scaled up a little. Overall improvement in strength is 50...70 points for blitz games. Version 5.1 17-Jan-2007 ----------- (new) New terms in evaluation: weak fields, pieces on strong fields, etc. (chg) Piece-square tables tuned a little. Version 5.05 8-Jan-2007 ------------ (fix) Reply to "isready" command in thinking mode (fix) Fixed bug with undo buffer overflow in UCI at move ~150 (new) Currentmove and nps info under UCI (new) K+B or K+N can't win regardless of opponent's material (new) Far passed pawns bonus (chg) More robust and stable search (chg) New time management scheme: not it does not lose on time in 1+0 games under Arena (chg) Piece values tuned Version 5.0 31-Dec-2006 ----------- Engine completely rewritten in pure C. * Both winboard/xboard and UCI support * Multi-PV mode in UCI * No pondering * No opening book and learning * Strength is about +20 points compared to version 2.817 Command line keys for xboard: -log write all input and output in greko.log file -hash 32 set hash size no more than 32 Mb (for example) Version 3.3a 16-Jul-2006 ----------- (fix) * Few lines in pondering code fixed. Now seems to work corectly (at least, under Winboard and Arena). Version 3.3 14-Jul-2006 ----------- 1. Bugfixes. 2. Search clean-up. (fix) * Fixed bug in UpdateHistory which caused assertion fails from time to time (incorrect constant). (fix) * Now "force" command stops pondering mode. Thanks to Volker Pittlik for bug report. (chg) * Futility pruning thrown away from search. (chg) * Out-of-check extension changed to 0.75 ply - always. (chg) * Nullmove parameter changed from R = (3 or 4) to const R = 3, more safe. (chg) * Bitboard operation PopLSB() inlined. Version 3.2 26-Apr-2006 ----------- 1. Bugfixes and code stability. 2. Evaluation clean-up. (fix) * Wrong comment in MainCommandProcessor() fixed. Multithreading is not mentioned anymore, current GreKo is a single-threaded program. (fix) * A really gross bug with negative depth in UpdateHistory() & RecordHash() fixed. Now the depth is checked to be positive in both functions. (chg) * Evaluation: A lot of alchemy with attacks and mobility removed from StaticEval(). Added very basic evaluation of R, N, B mobility instead. Works faster and better. (chg) * Out-of-board field index constant changed from -1 to NF = 64. Now all field index variables expected to be unsigned: 0...63 for fields, 64 for offboard. (chg) * Mailbox arrays U[64], L[64], UR[64], ... converted to single array STEPS [8][64]. (src) * Begin to re-format source code. Indentation style: all tabs replaced by spaces. Version 3.1 8-Jan-2006 ----------- After one-year meditations GreKo 3.x branch is again in development. This version goals: 1. bugfixes, 2. some changes in search 3. reviewing old code (fix) * Buffer-overrun safety code in user input. Many thanks to Dann Corbit! (fix) * Single-reply extension (1 ply) __really__ turned on (it worked only like simple check extension in 3.0). (fix) * Mate threat detection now works correctly (in 3.0 it was never triggered because of too small search window in nullmove). (new) * "perft explain" command: shows perft values for all subtrees in the root. (chg) * Win32 binary built with MSVC 6.0 + STLPort. Works faster than MSVC's STL version. (chg) * Makefile for gcc (MinGW, Linux, FreeBSD) simplified. Filenames corrected. No more ../bin and ../obj directories. (chg) * Out-of-check and pawn-on-7th extensions restricted for far-behind and far-ahead matearial balance, respectively. (chg) * Window size in root diminished to +/-50 centipawns. (chg) * Recapture extensions diminished from ply to ply/2. (chg) * Default setting: ShowEvalForWhite = 0. Version 3.0 1-Jan-2005 ----------- * Optimizations in speed. * Built-in random generator instead of standard rand() function. * Opening book now is compiler-independent (MSVC, gcc at least). Version 2.85 31-Dec-2004 ------------ * Fixed: incorrect resignation in some drawn endgames (incorrect pruning in QSearch). * New feature: GreKo now can create opening book from PGN for selected player. * A lot of changes in evaluation (more realistic), minor changes in search (extensions, pruning). * Anti-zeitnot code for blitz without increment, now GreKo loses on time more rarely. Version 2.817 01-Nov-2004 * Fixed: bug with small hash size specified in .ini file. * New command: "perft verify" for testing legality of move generator. * Some speedup in move generator and evaluator (less bitboard operations etc.) Version 2.815 01-Sept-2004 * Evaluation rolled back to 2.69 SNG. * Search and other stuff like in 2.812. * 50 Elo stronger compared to 2.812 in my tests. Version 2.812 23-July-2004 * Search: inefficient copying of internal fail high/low pv's deleted. * Evaluation: passed pawn weights diminished. * Time management scheme for fail high/low in the root simplified. * More statistics in "epdtest" routine. This version is approx. 50 Elo points stronger in blitz than ver. 2.81. Version 2.81 01-June-2004 * Overall code clenup for some basic operations (bitboard functions, make/unmake moves etc.), that gives +3...5% performance in various tests. * Some kind of futility pruning implemented in search. * "epdtest" command output modified, time and nodes calculations in test suites corrected. * New default settings: Learning=0, ShowEvalForWhite=0. Version 2.80 15-May-2004 * Fixed bug in QSearch checks generator. Version 2.71 15-Mar-2004 * Hash structure optimized: sizeof(CHashEntry) reduced from 32 to 16 bytes, sizeof(CPawnHashEntry) reduced as well. * New console command: book digest Creates plain text list of all opening lines in pgn file, useful for creation GreKo's repertoire file. Version 2.70 01-Mar-2004 * Internal iterative deepening added. * Opening repertoire file added. Version 2.69 SNG 17-Feb-2004 Private version for participation in CIS Computer Championship. * Evaluation for pawn structure tuned. Now at last it can solve WAC-100. * Weak pawns are penalized less if they are not on the semi-open lines (or opponent does not have rooks). Version 2.68 09-Feb-2004 * First check in row is extended now by 1 halfmove, and second in a row - by 0.75. * New term in evaluation: Queen near weak King (in addition to Q-K tropism). * Display evaluation either for White or for engine (configurable). * Piece-field tables changed (now like in Crafty). Two new tables for K in endgame. Version 2.67 05-Feb-2004 * Move generator changed: under check generate only candidates to check evasion. Significant speedup in tactical positions and test suites. * New evaluation terms: pinned pieces (currently only pinned to King and Queen). * New parameter in GreKo.ini: OfferDraw, 0 or 1. (To avoid pop-up messagebox in games vs. Plywood :)) Version 2.65 31-Jan-2004 * All move sorting (both in QSearch and normal search) use now SEE instead of MVV/LVA. * New time management scheme: add more time after fail high/low in the root. Version 2.64 27-Jan-2004 * Fixed GREAT bug in QSearch pruning (now it does not prune first capture). * Does not prune captures on first ply of QSearch at all. Version 2.63 20-Jan-2003 * Optimized mobility evaluation. * Tweaking extensions. * Added pawn phalangues and blocked on initial field in pawn evaluation. Version 2.62 12-Jan-2004 * Added simple checks on the first ply of QSearch. * Tweaks in evaluation. * Check extension returned back to 1 halfmove. * Yet another bugfix in KBPK endgame (wrong field color tests for black bishop). Version 2.61 7-Jan-2004 * Faster PopCount function (pure arithmetics instead of table lookup). * Tweaks in evaluation. * Fixed output of fail high/low PV's in epdtest routine. Version 2.60 22-Dec-2003 * Fixed bug in KBPK endgame (did not test file=a/h for black pawn). * Changes in QSearch: under check always generate all legal moves, do not return static eval * New opening book format - more compact .bin file and more detailed "bk" command output. * New parameter for configuring book logic: learn weight (see GreKo.ini file). * Binary book and learn files now have headers with version signature. * Fixed (?) bug in making book from big PGN. * SAN parser improved for faster detection of protocol commands which are non-moves. * Adding more time for thinking when fail-low at the root. Version 2.59a 08-Dec-2003 * Check extensions hadnling fixed. Version 2.59 07-Dec-2003 * Some critical fixes in time management for better stability (got rid (?) of sudden wrong mate claiming bugs) * New check extension scheme: the first check is extended by 1 ply, then by 0.75 * BAD_TRADE weight reduced to 0.40 pawn * "result" command parsing in pondering mode - fixed Version 2.58 01-Dec-2003 * g_ThinkPos replaced with CPosition parameter in all AlphaBeta() functions * check extension returned back to 1 halfmove (was 0.75) * extend only at ply <= 2 * g_IterDepth (was g_IterDepth + 11) * QUEEN_AND KNIGHT and BISHOP_AND ROOK returned back to static eval * Knight-to-King tropism only with friendly Queen on the board * no new hash hit after extension, just check the depth field in hash entry Version 2.57 23-Nov-2003 * New statistics format; more correct time calculation * if( !ppv0->empty() ) check before assigning PV to approved PV * do not print PV for unfinished iteration * minor tuning in eval coeffs Version 2.56 7-Nov-2003 * Changes in extensions using SEE score * More valuable passed pawns when opponent has little piece material Version 2.54 * New routine for bit-scan operation (LSB) using magic 0x78291ACF constant. * File History.txt added to source package. Version 2.53 15-Oct-2003 * Source code is now portable to Linux and FreeBSD. Many thanks to Andrey Shytov. * Got rid of all 0x88 stuff in board representation and moves generation. * More accurate evaluation of passed pawns and possible draw in KBP-K endgame. * Static exchange evaluator used in move sorting instead of old MVV/LVA scheme. * Mate threat extensions added. * Check extension value changed from 1.0 to 0.75 halfmove. Version 2.45 1-Aug-2003 * This version was included in DVD disc attached to "Game.EXE" #9/2003 (Russian magazine about computer games) Version 2.32 June 2003 Bitboards added. Hybrid scheme (0x88 + bitboards). Version 2.01 December 2002 After half a year of development GreKo was re-written from scratch to improve data structures and overall design. 0x88 move generator. Version 1.10 July 2002 Stable version with analyze and pondering support. Version 0.41 11-May-2002 * Hashtable implemented (using Zobrist scheme). Version 0.36 1-Apr-2002 * First public release of GreKo. Only basic features.