Caissa v1.3

This release greatly improves playing strength in ultra-fast time controls.

Other changes:

• NN trainer improvements
• New neural networks
• Endgame neural network is now used for <= 6 pieces positions
• Search tweaks and improvements
• Fix time management when using X/Y+Z time controls

TC=10+0.5

Score of Caissa 1.3 vs Caissa 1.2 (AVX2/BMI2): 138 - 109 - 147  [0.537] 394
...      Caissa 1.3 playing White: 115 - 19 - 64  [0.742] 198
...      Caissa 1.3 playing Black: 23 - 90 - 83  [0.329] 196
...      White vs Black: 205 - 42 - 147  [0.707] 394
Elo difference: 25.6 +/- 27.2, LOS: 96.7 %, DrawRatio: 37.3 %

TC=1+0

Score of Caissa 1.3 vs Caissa 1.2 (AVX2/BMI2): 585 - 399 - 296  [0.573] 1280
...      Caissa 1.3 playing White: 375 - 125 - 141  [0.695] 641
...      Caissa 1.3 playing Black: 210 - 274 - 155  [0.450] 639
...      White vs Black: 649 - 335 - 296  [0.623] 1280
Elo difference: 50.8 +/- 16.8, LOS: 100.0 %, DrawRatio: 23.1 %

TC=20/1+0

Score of Caissa 1.3 vs Caissa 1.2 (AVX2/BMI2): 484 - 365 - 426  [0.547] 1275
...      Caissa 1.3 playing White: 358 - 84 - 196  [0.715] 638
...      Caissa 1.3 playing Black: 126 - 281 - 230  [0.378] 637
...      White vs Black: 639 - 210 - 426  [0.668] 1275
Elo difference: 32.5 +/- 15.6, LOS: 100.0 %, DrawRatio: 33.4 %

FRC, TC=5+0.1

Score of Caissa 1.3 vs Caissa 1.2 (AVX2/BMI2): 1103 - 981 - 1861  [0.515] 3945
...      Caissa 1.3 playing White: 596 - 461 - 915  [0.534] 1972
...      Caissa 1.3 playing Black: 507 - 520 - 946  [0.497] 1973
...      White vs Black: 1116 - 968 - 1861  [0.519] 3945
Elo difference: 10.7 +/- 7.9, LOS: 99.6 %, DrawRatio: 47.2 %

Caissa v1.2

• Switch regular PSQT to king-relative PSQT
• Neural network trainer optimization and improvements
• PSQT trainer refactor
• Add razoring and alpha pruning
• Tweak SEE pruning, late move prunning, etc
• Update neural network to 704x512x16x32 architecture and retrain on more data
• Fix castling move printing in non-960 mode

Regular chess improvement:

...      Caissa 1.2 (AVX2/BMI2) playing White: 204 - 26 - 144  [0.738] 374
...      Caissa 1.2 (AVX2/BMI2) playing Black: 85 - 140 - 145  [0.426] 370
...      White vs Black: 344 - 111 - 289  [0.657] 744
Elo difference: 58.0 +/- 19.6, LOS: 100.0 %, DrawRatio: 38.8 %

Chess960 improvement:

...      Caissa 1.2 (AVX2/BMI2) playing White: 154 - 51 - 121  [0.658] 326
...      Caissa 1.2 (AVX2/BMI2) playing Black: 121 - 81 - 124  [0.561] 326
...      White vs Black: 235 - 172 - 245  [0.548] 652
Elo difference: 77.5 +/- 21.3, LOS: 100.0 %, DrawRatio: 37.6 %

Caissa v1.1

• Implemented support for Chess960
• Minor bugfixes
• Improved endgame evaluation

Caissa 1.0

• Added support for Gaviota tablebases (for now used only for root moves and during training)
• Introduced endgame neural network (used for positions with 4 and 5 pieces)
• Incrementally updated PSQT [Optimization]
• Neural network tuner tweaks
• New neural network trained on more data
• More hardcoded endgame knowledge
• Aspiration windows tweaks

Score of Caissa 1.0 (AVX2/BMI2) vs Caissa 0.9.0 (AVX2/BMI2): 350 - 184 - 262  [0.604] 796
...      Caissa 1.0 (AVX2/BMI2) playing White: 252 - 34 - 111  [0.775] 397
...      Caissa 1.0 (AVX2/BMI2) playing Black: 98 - 150 - 151  [0.435] 399
...      White vs Black: 402 - 132 - 262  [0.670] 796
Elo difference: 73.5 +/- 20.0, LOS: 100.0 %, DrawRatio: 32.9 %

Full Changelog: 0.9...1.0

Caissa 0.9

• Neural network tuner improvements and optimization
• Retrained neural network with different parameters and picked the best net
• Minor search tweaks
• Minor bug fixes
• Implemented staged move generation

Score of Caissa 0.9.0 (AVX2/BMI2) vs Caissa 0.8.0 (AVX2/BMI2): 303 - 170 - 198  [0.599] 671
...      Caissa 0.9.0 (AVX2/BMI2) playing White: 220 - 42 - 76  [0.763] 338
...      Caissa 0.9.0 (AVX2/BMI2) playing Black: 83 - 128 - 122  [0.432] 333
...      White vs Black: 348 - 125 - 198  [0.666] 671
Elo difference: 69.8 +/- 22.3, LOS: 100.0 %, DrawRatio: 29.5 %

Caissa 0.8

• Changed neural net architecture
• Improved trainer and trained the net on new data
• Fixed stackoverflow crash
• Fix bug in SSE4 version
• Optimize neural net evaluation (both at runtime and during training)

Playing strength improvement compared to previous release: ~70 Elo

Score of Caissa 0.8.0 (AVX2/BMI2) vs Caissa 0.7.0 (AVX2/BMI2): 247 - 126 - 170  [0.611] 543
...      Caissa 0.8.0 (AVX2/BMI2) playing White: 176 - 22 - 75  [0.782] 273
...      Caissa 0.8.0 (AVX2/BMI2) playing Black: 71 - 104 - 95  [0.439] 270
...      White vs Black: 280 - 93 - 170  [0.672] 543
Elo difference: 78.7 +/- 24.6, LOS: 100.0 %, DrawRatio: 31.3 %

Full Changelog: 0.7...0.8

Caissa 0.7

• Implemented custom neural network and trainer for it. It means Caissa is now fully independent engine and does use external NNUE code and networks.
• Tuned PSQT values based on self-play games
• Numerous improvements in the search and move ordering
• Fixing search explosions in weird positions
• Implemented more endgame knowledge

Caissa 0.5

Estimated strength gain compared to previous version: +200 Elo

Changelog:

• Multithreaded search strength improvement
• History heuristics improvements
• Fixed max hash size reported to UCI
• Time management tuned, notably fixed time allocation in games without increments
• Fixed bug in Null Move Prunning
• Fixed bug in time management when doing multithreaded search leading to incorrect move being reported
• Improved Syzygy tablebases support
• Improved move ordering
• Fixed FEN parsing
• Quiesence search improvements and optimizations
• Fixed bug in Static Exchange Evaluation
• Fixed search explosions in certain positions
• Late Move Reductions and Prunning tweaks
• Added more endgame knowledge
• Optimized output prining
• Numerous search improvements and tweaks

Caissa 0.4

• Fixed bug in Null Move Prunning
• Added singularity search for root node
• Introduced separate binaries for instruction sets:
  • x64 BMI2 - for modern CPUs supporting BMI2 and AVX2
  • x64 POPCNT - for older CPUs supporting SSE4 and POPCNT
  • x64 - for older legacy CPUs supporting only SSE2

Caissa 0.3

• Fixed multi-threading crash
• Search improvements
• Endgame evaluation improvements
• Transposition table improvements