forked from vshcherbyna/igel
-
Notifications
You must be signed in to change notification settings - Fork 0
/
history.txt
1497 lines (1170 loc) · 49.3 KB
/
history.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
Igel - open source chess engine forked from GreKo 2018.01.
(c) 2018-2023 Volodymyr Shcherbyna <[email protected]> (Igel author)
(c) 2002-2018 Vladimir Medvedev <[email protected]> (GreKo author)
Igel 3.5.0
-------------
22-Jun-2023
- new network architecture
- new network c049c117
Igel 3.4.0
-------------
30-Jan-2023
- fixing major bug in accumulator handling for null moves
- avoid heap objects (std::vector) in search function
- switching to pure nnue evaluation
- new network architecture HalfKav2_hm
- new network ign-3-debc71a4
Igel 3.3.0
-------------
15-Jan-2023
- new network architecture HalfKav2
- new network ign-2-51ba2968
Igel 3.2.0
-------------
17-Dec-2022
- smaller margin for prunning quiets
- scale down eval to 1000
- history extensions only on non pv moves
- adjust scale formula
- less reductions in lmp when not improving
- negative extensions
- futility pruning improvements
- implement cutnode
- new network ign-1-9a48854b created by using prev network (ign-1-3e998e78) and re-trained with a different lambda
Igel 3.1.0
-------------
03-Jun-2022
- Scale eval at 1280
- More aggressive prunning of quiets when not improving
- Use 150 cp when razoring
- Avoid tt cutoff for rule50
- Improve statistics in SMP mode
- New network trained using 16B of d5-d8 data from Igel HCE/ign-1-139b702b with d16 validation using ign-1-139b702b
Igel 3.0.5
-------------
28-Apr-2021
- Train a new network using Igel 2.6.0 (HCE), Igel 2.9.0 (NNUE) and Igel 3.0.0 (NNUE) data: ign-1-d593efbd
- Better time management in sd time control
- Skip tt cutoff for null moves in search
- Skip null move prunning on tt hit
- Vote for best thread in SMP mode
- Less aggressive time use in non repeated tc
Igel 3.0.0
-------------
04-Apr-2021
- Train a new network using Igel 2.6.0 (HCE) and Igel 2.9.0 (NNUE) data: ign-1-139b702b
- Fix in check extensions
- Use incbin on Linux to inject network file into the binary
- Tune null move prunning conditions
- Tune probcut prunning conditions
- Tune quiets prunning
- Fix bogus condition for history calculation
- Implement more aggressive null move prunning
- Stop extension of captures
- Prevent explosion of history extensions
- Implement better time management for sudden death time control
- Count not played quiets for LMP prunning
- Remove dummy 100 multiplier for history ordering calculation
- Fix compilation issues with gcc 10 and Android NDK's Clang
- Remove unused counter move table
- Fix out of boundary access for history array
- Fix issue with singular extensions - reported by ChizhovVadim (author of CounterGo chess engine)
- Fix crash when 'ucinewgame' command is not issued - reported by Ipmanchess
- Fix crash when thread position is not initialized - reported by Ipmanchess
Igel 2.9.0
-------------
25-Dec-2020
- Switch to own network based on Igel search and evaluation (ign-0-9b1937cc)
- NNUE dynamic eval scaling
Igel 2.8.0
-------------
23-Sep-2020
- Fix null move prunning bug (54 elo)
- Use 250cp as lazy eval threshould
- Improve time management for sudden death time controls
- Use Dietrich Kappe's Night Nurse improved net (internal version ninu-0.2.5b.nnue)
Igel 2.7.0
-------------
19-Aug-2020
- Switch to NNUE as a main evaluation function
- Use Dietrich Kappe's Night Nurse NNUE net as a default net for Igel 2.7.0 release
Igel 2.6.0
-------------
01-Aug-2020
- Implement piece threats evaluation
- Simplify connected pass pawn evaluation
- Implement backward pawn evaluation
- Apply tempo after interpolation and speed up null move evaluation
- Implement verification qsearch for probcut
- Remove L1 normalisation from tuner and retune all params
- Implement trapped rook evaluation
- Implement razoring in search
- Implement king danger evaluation
- Implement hanging piece evaluation
- Implement weak pawn evaluation
- Implement restricted piece evaluation
- Implement lazy evaluation with a threshould of 300cp
- Implement multicut search
- Implement threat by a protected pawn evaluation
- Improve safe check evaluations
- Bench 7538060
Igel 2.5.0
-------------
15-Jun-2020
- Implement OpenBench support
- Implement BMI2 instruction support for faster search
- Implement prefetch instruction support for faster search
- Implement ponder support
- Implement improved sudden death time control management
- Implement use of hugepages in Linux OS
- Implement more efficient tt management and re-write tt replacement scheme to 'always replace'
- Implement proper handling of chopper moves
- Implement improved king safety evaluation
- Implement improved weights management
- Implement symmetric evaluation function
- Set tt score in qsearch
- Set bestmove as a hashmove by default
- Bugfix for knight outpost
- Bugfix for depth skipping in smp mode
- Implement pst based evaluation for connected passers
- Bugfix for probcut in check
- Factor in 'improving' even when in checks
- Adjust bestscore from tt score
- Bugfix shuffling of moves at root in smp mode
- Bench 8175057
Igel 2.4.0
-------------
22-Mar-2020
- Implement singlular extensions
- Detect drop of score and using time more aggressively
- Remove IID and simplify search
- Do not tt cutoff on pv moves
- Simplification for nullmove pruning
- Simplification for connected pawn evalution
- Basic endgame knowledge to detect draw in eval/search
- Retune everything + decrease number of tunnable parameters from 2600 to 1500
- Apply LMR when at least one legal move has been played
- Decrease reductions in LMR for pv moves
Igel 2.3.1
-------------
14-Jan-2020
- Re-write the root search routine
- Re-write aspiration loop with a smaller aspiration window
- Do not cutoff qsearch on pv moves in qsearch
- Handle repetitions at the qsearch entance
- More strict pruning conditions for quiet moves
- Proper way to pass position to child threads in SMP mode to handle repetitions
- Unified condition to killer moves as non-tactical moves
- Perft tests for six different positions
Igel 2.3.0
-------------
01-Jan-2020
- Fix issue when reductions in LMR could lead to directly qsearch
- Rename uci option 'Level' into 'Skill Level'
- Do not apply razoring when in check
- Do not apply 'improving' factor when in check
- Do not prune quiets at root
- Fetch history only when a quiet move is detected
- Improve branching factor and apply LMR more aggressively
- Fix compilaton issue with clang/gcc9 compiler
Igel 2.2.1
-------------
29-Dec-2019
- Rename uci option 'Level' into 'Skill Level'
- Do not apply razoring when in check
- Do not apply 'improving' factor when in check
- Do not prune quiets at root
- Fetch history only when a quiet move is detected
- Improve branching factor and apply LMR more aggressively
- Fix compilaton issue with clang compiler
Igel 2.2.0
-------------
23-Dec-2019
- Fix futility pruning and use it only after at least one legal move has been played
- Fix 32bit overflow in syzygy probing code that caused bugs when probing 7 piece positions
- Fix syncronisation issue that caused "connection stalled" in Linux OS
- Fix overflow bug that prevented using transposition table bigger than 32 GB
- Improve performance in SMP mode and avoid copying the pawn hash table for each worker
- Improve performance in SMP mode when clearing large hash size (more than 64 Gb)
- Improve performance in SMP mode and solve cache coherency issues when calculating node/tbhits count
- Improve performance of tuner and perform multithreaded texel tuning
- Impove performance and use hardware instruction to count leading zeroes (lsb/poplsb)
- Improve performance in SMP mode and start worker threads before the main thread
- Improve performance in SMP mode and skip random depths by worker threads (Laser/Ethereal method)
- Improve performance in Linux OS and use -flto compiler option (around 10% of increase of NPS)
Igel 2.1.0
-------------
20-Nov-2019
- Implement ponder option
- Implement level uci option: where 1 to 10 is a level suitable for amateurs, level 11 to 19 suitable for professional players and level 20 is suitable for chess engines
- Re-tuning evaluation using licess-quite set and texel tunning
Igel 2.0.0
-------------
05-Oct-2019
- Fix evaluation bug in SMP mode that significantly affects performance (~300 elo)
Igel 1.9.2
-------------
24-Sep-2019
- Tuning of evaluation parameters
Igel 1.9.1
-------------
22-Sep-2019
- Fix time control mode "nodes limit"
- Fix problem when a bestmove may become 0 in case tb search failed
- Fix bug #96: incorrect evaluation of position 4k3/8/8/8/5n2/1R6/6P1/5K2 w KQkq -
- Fix bug #97: incorrect evaluation of king exposed
- Expose evaluation of a position by implementing "eval" command option
Igel 1.9.0
-------------
01-Sep-2019
- Support macOS
- Use different base for evaluation parameters and tune them using coordinate descent
- Implement queen and knight mobility evaluation
- Improve speed when checking time limits
- Fix underrun bug when clearing pieceStack
- Extend recaptures only when searching in PV
- Disable ponder until discovered issues are fixed
Igel 1.8.3
-------------
28-Jul-2019
- Drop into qsearch at depth <= 0 without checking for tt
- Avoid null move prunning if a tt hit has been discovered
- Fix the detection of a root position
- SEE margin improvements
Igel 1.8.2
-------------
17-Jul-2019
- Improve time management of worker threads when number of workers is high
- Improve time management in absolute time controls
Igel 1.8.0
-------------
09-Jul-2019
- Re-write UCI communication functionality
- Implement ponder (aka 'permanent brain')
- Do not store result from interrupted search in tt hash
- Do not return result from interrupted search as 'best move'
- Fix the pv count when returning draw or moves from tt hash
- Move to Visual Studio 2019 cmake syntax
Igel 1.7.0
-------------
17-Jun-2019
- Avoid LMR for moves that make pawn pushes to 5th and 6th ranks
- Apply LMR at shallow depths
- Implement counter move, follow up move and history move pruning
- Implement probcut
- Avoid prunning in IID path
- Enable LMR during endgames
Igel 1.6.0
-------------
26-May-2019
- Re-write killer moves functionality and drop mate killers
- Re-write move ordering and introduce counter move heuristics
- Re-write LMR functionality: take into account 'improving' factor and prune more aggressively
- Increase MAX_PLY to 128 from 64
- Use tt moves in qsearch as a first move
Igel 1.5.1
-------------
09-May-2019
- Improving time management
- Fix old cpu builds by removing popcount instructions
- Optimize lazy smp code
Igel 1.5.0
-------------
05-May-2019
- Improving usage of transposition table in search and qsearch
- New pruning techniques: static null pruning, mate distance pruning
- New eval params + texel tuning + code refactoring of eval function
Igel 1.4.2
-------------
09-Apr-2019
- Texel tunning using Zurichess test positions (725k) taken from http://www.zurichess.xyz/blog/texels-tuning-method/
- Fixing a bug with evaluation of pair of the knights. The buggy code was evaluating the pair of black kings and the pair of white kings
Strength compared to Igel 1.4.1 and other engines:
Igel 1.4.2 64 POPCNT vs 1.4.1 64 POPCNT
tc=all/10+0.1
threads=1
hash=1024
book=up to 12 moves for both engines
EGTB=3-4-5 piece
Score of Igel 1.4.2 64 POPCNT vs Igel 1.4.1 64 POPCNT: 1656 - 1236 - 725 [0.558] 3617
Elo difference: 40.53 +/- 10.16
Igel 1.4.2 64 POPCNT vs 1.4.1 64 POPCNT
tc=all/60+0.6
threads=1
hash=1024
book=up to 12 moves for both engines
EGTB=3-4-5 piece
Score of Igel 1.4.2 64 POPCNT vs Igel 1.4.1 64 POPCNT: 1166 - 887 - 627 [0.552] 2680
Elo difference: 36.30 +/- 11.54
Igel 1.4.1
-------------
25-Mar-2019
- Fix bug with evaluation of a position at a max ply
Igel 1.4
-------------
17-Mar-2019
- Implemented support for syzygy tablebases (uci options SyzygyPath and SyzygyProbeDepth)
- Changed the move ordering during root search
- Changed evaluation of late end-games
- Minor memory optimisations (decrease memory usage by ~2 Mb)
- Cosmetic changes: report tbhits, proper node count from primary & lazy smp threads
Strength compared to Igel 1.2, Igel 1.3 (various configurations) and other engines:
Igel 1.4 64 POPCNT vs 1.3 64 POPCNT
tc=all/60+0.6
threads=11 (for Igel 1.4)
threads=1 (for Igel 1.3)
hash=6144
book=up to 8 moves for both engines
EGTB=3-4-5 piece (used only by Igel 1.4, because Igel 1.3 does not support EGTB)
Score of Igel 1.4 64 POPCNT vs Igel 1.3 64 POPCNT: 118 - 13 - 54 [0.784] 185
Elo difference: 223.72 +/- 46.36
Igel 1.4 64 POPCNT vs Igel 1.2 64 POPCNT
tc=all/10+0.1
threads=1
hash=1024
book=up to 8 moves for both engines
EGTB=3-4-5 piece (used only by Igel 1.4, because Igel 1.2 does not support EGTB)
Score of Igel 1.4 64 POPCNT vs Igel 1.2 64 POPCNT: 883 - 440 - 677 [0.611] 2000
Elo difference: 78.25 +/- 12.52
Igel 1.4 64 POPCNT vs Igel 1.2 64 POPCNT
tc=all/10+0.1
threads=1
hash=1024
book=up to 8 moves for both engines
EGTB=disabled for both engines
Score of Igel 1.4 64 POPCNT vs Igel 1.2 64 POPCNT: 813 - 506 - 686 [0.577] 2005
Elo difference: 53.62 +/- 12.39
Igel 1.4 64 POPCNT vs Igel 1.3 64 POPCNT
tc=all/10+0.1
threads=1
hash=1024
book=up to 8 moves for both engines
EGTB=disabled for both engines
Score of Igel 1.4 64 POPCNT vs Igel 1.3 64 POPCNT: 727 - 536 - 740 [0.548] 2003
Elo difference: 33.23 +/- 12.09
Igel 1.4 64 POPCNT vs Igel 1.3 64 POPCNT
tc=all/10+0.1
threads=1
hash=1024
book=up to 8 moves for both engines
EGTB=3-4-5 piece (used only by Igel 1.4, because Igel 1.3 does not support EGTB)
Score of Igel 1.4 64 POPCNT vs Igel 1.3 64 POPCNT: 789 - 538 - 673 [0.563] 2000
Elo difference: 43.83 +/- 12.44
Igel 1.4 64 POPCNT vs GreKo 2018 (2708)
tc=40/960 (CCRL 40/40)
threads=1
hash=1024
book=up to 8 moves for both engines
EGTB=3-4-5 piece (used only by Igel, because Greko does not support EGTB)
Score of Igel 1.4 64 POPCNT vs GreKo 2018: 10 - 13 - 27 [0.470] 50
Elo difference: -20.87 +/- 65.99
Igel 1.4 64 POPCNT vs WyldChess 1.51 64 POPCNT (2687)
tc=40/960 (CCRL 40/40)
threads=1
hash=1024
book=up to 8 moves for both engines
EGTB=3-4-5 piece (used by Igel and WyldChess)
Score of Igel 1.4 64 POPCNT vs WyldChess: 19 - 18 - 13 [0.510] 50
Elo difference: 6.95 +/- 84.36
Igel 1.4 64 POPCNT vs ECE-X3.5 64 (2646)
tc=40/960 (CCRL 40/40)
threads=1
hash=1024
book=up to 8 moves for both engines
EGTB=3-4-5 piece (used only by Igel, because ECE-X3.5 does not support EGTB)
Score of Igel 1.4 64 POPCNT vs ECE-X3.5 64Bit: 23 - 21 - 6 [0.520] 50
Elo difference: 13.90 +/- 92.40
Igel 1.4 64 POPCNT vs DanaSah 7.3 (2640)
tc=40/960 (CCRL 40/40)
threads=1
hash=1024
book=up to 8 moves for both engines
EGTB=3-4-5 piece (used only by Igel, because DanaSah 7.3 does not support EGTB)
Score of Igel 1.4 64 POPCNT vs DanaSah 7.3: 19 - 15 - 15 [0.541] 49
Elo difference: 28.43 +/- 82.62
Igel 1.4 64 POPCNT vs Fridolin 3.00 (2672)
tc=40/960 (CCRL 40/40)
threads=1
hash=1024
book=up to 8 moves for both engines
EGTB=3-4-5 piece (used by Igel and Fridolin)
Score of Igel 1.4 64 POPCNT vs Fridolin 3.00: 11 - 18 - 19 [0.427] 48
Elo difference: -51.03 +/- 77.93
Igel 1.3
-------------
05-Mar-2019
- New eval terms for pair of rooks, knights, bishops
- Connected pawn evaluation
Strength compared to 1.2
threads=1
hash=1024
tc all/60+0.6
Score of Igel 1.2 64 POPCNT vs Igel 1.3 64 POPCNT: 947 - 1093 - 1360 [0.479] 3400
Elo difference: -14.93 +/- 9.04
threads=1 (for 1.2)
threads=11 (for 1.3)
hash=6GB
tc all/60+0.6
Score of Igel 1.2 64 POPCNT vs Igel 1.3 64 POPCNT: 38 - 138 - 97 [0.317] 273
Elo difference: -133.47 +/- 34.12
Igel 1.2
-------------
06-Aug-2018
- Fixed tt bug that resulted in random blunders
- Fixed NPS bug
- Fixed infinite stop bug
- Implemented POPCNT support
Igel 1.1
-------------
03-Jul-2018
- Remove leftovers of Strength option to allow running in ChessGUI software
Igel 0.8
-------------
30-Jun-2018
First release of igel engine, derivative of GreKo 2018.01
What's new:
- lazy smp
- more compact tt (30% smaller than in GreKo)
- max tt size increased from 1024 Mb to 16384 Mb
- razoring at depth <= 2
- futility check improvements
- variable reduction depth in null move prunning
- smaller, variable aspiration window
- variable time management that takes into account opponent time
- miscellaneous improvements
Removed features:
- xboard
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 <level>
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 <hash size in Mb>
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 <games_file> [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):