v35.0 (2023-10-16)

Regression test

LTC:

Elo   | 53.65 +- 2.66 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 20000 W: 4686 L: 1622 D: 13692

v34.0 (2022-11-03)

Regression test

LTC:

ELO   | 52.97 +- 2.89 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 20000 W: 5175 L: 2149 D: 12676

v33.0 (2022-05-12)

Regression test

LTC:

ELO   | 51.85 +- 2.90 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 20000 W: 5171 L: 2208 D: 12621

v32.0 (2021-12-02)

Regression test

LTC:

ELO   | 65.02 +- 4.42 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 10000 W: 3059 L: 1209 D: 5732

Fixed (2 changes)

Fixed 32-bit detection for printing 64-bit variables.

Fixed KPK table probing for KBPsK endgames when the Pawns are on the Kingside.

ELO   | 0.66 +- 4.47 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=16MB
GAMES | N: 10000 W: 2170 L: 2151 D: 5679

Performance (13 changes)

Reduced the margin for Reverse Futility Pruning when position is improving.

ELO   | 2.92 +- 2.32 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 37512 W: 8312 L: 7997 D: 21203

ELO   | 2.43 +- 1.93 (95%)
SPRT  | 40.0+0.40s Threads=1 Hash=64MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 43008 W: 7581 L: 7280 D: 28147

Disabled singular extensions on positions with known win/loss.

ELO   | 5.73 +- 4.85 (95%)
SPRT  | 40.0+0.40s Threads=1 Hash=64MB
LLR   | 2.99 (-2.94, 2.94) [-4.00, 1.00]
GAMES | N: 6608 W: 1164 L: 1055 D: 4389

Decreased Delta Pruning margin in qsearch.

ELO   | 3.09 +- 2.45 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.99 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 33800 W: 7580 L: 7279 D: 18941

ELO   | 3.27 +- 2.58 (95%)
SPRT  | 40.0+0.40s Threads=1 Hash=64MB
LLR   | 3.03 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 23880 W: 4214 L: 3989 D: 15677

Changed the way SEE pruning triggers in qsearch. Instead of recomputing SEE for each attempt to prune a move, we directly use information from the move ordering structure which already tells us if, at this stage, captures lose material. It also avoids pruning TT moves when they have a bad SEE.
```
ELO   | 4.40 +- 3.45 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 16744 W: 3707 L: 3495 D: 9542
```
Added more draw rules for endgames with Pawns on one side. Fixed eval for King + Bishop + Pawns vs lone King, and King + Pawns vs lone King when the endgame is clearly drawn.
```
ELO   | 1.10 +- 3.07 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [-4.00, 1.00]
GAMES | N: 21472 W: 4736 L: 4668 D: 12068
```

Reduced history entry sizes from int32_t to int16_t, as the resolution used is only 14 bits wide. Small speedup at search initialization.

ELO   | 0.85 +- 2.88 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [-4.00, 1.00]
GAMES | N: 24512 W: 5410 L: 5350 D: 13752

Switched to Adam optimizer for tuning the eval. Retuned eval for testing, using a dataset of 4.5M positions coming from selfplay games at 1+0.01 time control.
```
ELO   | 1.02 +- 3.15 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [-4.00, 1.00]
GAMES | N: 22504 W: 5469 L: 5403 D: 11632
```

Included King Proximity in the Passed Pawn evaluation. Also reduced batch size used in tuning from 32768 to 2048.

ELO   | 22.27 +- 9.86 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 2296 W: 626 L: 479 D: 1191

Removed the King square from the Mobility zone. Allowed King Safety code to trigger with one attacker when the attacking side still has a Queen on the board.
```
ELO   | 10.86 +- 6.37 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.97 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 5024 W: 1184 L: 1027 D: 2813
```

Added capture histories for better move ordering.

ELO   | 4.58 +- 3.52 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 15712 W: 3404 L: 3197 D: 9111

Limited history bonus uniformly at 2000 units. Also reduced the history bonus quadratic coefficient from 16 to 14.

ELO   | 5.45 +- 4.04 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 11912 W: 2592 L: 2405 D: 6915

Decreased Late Move Reductions of killer moves and countermoves by one ply.

ELO   | 3.50 +- 2.78 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 3.03 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 25296 W: 5479 L: 5224 D: 14593

Made Late Move Pruning much more aggressive.

ELO   | 24.76 +- 10.03 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
GAMES | N: 1968 W: 490 L: 350 D: 1128

Changed (8 changes)

Simplified the square to bitboard conversion by removing the SquareBits table.
Removed the validation loss from the tuning code, since the eval doesn't overfit for datasets > 500k positions.
Fixed core usage in parallel loops during tuning, where one more thread would be briefly used due to how workload size was computed.
Fixed gradient updating, now using a mutex to avoid concurrent writes to the gradient array during tuning.
Fixed dataset generation, now simply filtering non-quiet positions, no playouts applied from the engine's search.
Added FEN printing when a 'd' command is sent.
Small naming changes in qsearch.
Made the Transposition Table reset multithreaded. Enforced memory zeroing to avoid page faults slowing down the program on the first search of a game.

v31.0 (2021-06-26)

Regression test

LTC:

ELO   | 61.64 +- 7.50 (95%)
SPRT  | 40.0+0.40s Threads=1 Hash=64MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 1.00]
Games | N: 3736 W: 1176 L: 520 D: 2040

Fixed (3 changes)

Fixed issues with benchmark containing commas for CI.
Fixed some odd behaviors on fixed depth/infinite search with SMP + mate pruning on root nodes. (This is technically a functional change, but since it only triggers in mate-in-one positions, no regression tests were made.)
Fixed a dead store of the variable 'piece' in undo_move() function.

Performance (11 changes)

Added AdaGrad tuning to the code base and retuned all eval parameters. To use Stash in tuning mode, compile with 'CFLAGS="-DTUNE -fopenmp" make ...'. You then can do ./stash-bot dataset_file to launch the tuning session. The expected format for the dataset is FEN OUTCOME, with the outcome being 1.0 for White wins, 0.5 for draws, and 0.0 for Black wins. (Note that it is possible to use more fine-grained values like 0.7 and such, but it hasn't been tested yet.)
```
ELO   | 9.47 +- 6.20 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 6496 W: 1840 L: 1663 D: 2993
```
The tuner was later changed to include a validation loss and have a better output of tuned values.
Rewrote the King Safety code from scratch, now including the total attack count and the potential safe checks on the enemy King. Also changed the KS scaling to be quadratic in middlegame, and linear in endgame.
```
ELO   | 4.24 +- 3.38 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 19736 W: 4941 L: 4700 D: 10095
```
Weak and safe squares were redefined in the next test, pushing the total gain to +13 Elo.
```
ELO   | 9.69 +- 5.90 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.08 (-2.94, 2.94) [0.00, 5.00]
Games | N: 6064 W: 1466 L: 1297 D: 3301
```

Allowed capture/promotion moves to be reduced in LMR by one ply.

ELO   | 4.88 +- 3.77 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [0.00, 5.00]
Games | N: 15174 W: 3640 L: 3427 D: 8107

Forced LMR depth to be > 0 to avoid dropping into qsearch.

ELO   | 3.14 +- 2.51 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 33842 W: 7925 L: 7619 D: 18298

Added ponder support to the codebase.

Raised SEE Pruning max depth to 5.

ELO   | 7.18 +- 4.95 (95%)
SPRT  | 10.0+0.10s Threads=1 Hash=16MB
LLR   | 2.99 (-2.94, 2.94) [0.00, 5.00]
Games | N: 8612 W: 2055 L: 1877 D: 4680

Disabled all pruning code when in check and removed eval computing. The logic behind that is that King Safety values get pretty wild for positions with check, because if the piece attacking the King is a slider, it may have a lot of other possible safe checks along the King line. This test was done in two commits, the first one removing computations from search:
```
ELO   | 2.36 +- 1.66 (95%)
SPRT  | 10.0+0.10s Threads=1 Hash=16MB
LLR   | 3.01 (-2.94, 2.94) [0.00, 5.00]
Games | N: 75936 W: 17338 L: 16822 D: 41776
```
and the second one, done later, removing computations from qsearch:
```
ELO   | 4.25 +- 3.35 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 17976 W: 4038 L: 3818 D: 10120
```

Allowed entries with lower depth to conditionally replace other ones in TT. Credits to Jay Honnold (Berserk's author) for this patch.

ELO   | 7.49 +- 5.03 (95%)
SPRT  | 10.0+0.10s Threads=1 Hash=16MB
LLR   | 3.03 (-2.94, 2.94) [0.00, 5.00]
Games | N: 8160 W: 1909 L: 1733 D: 4518

ELO   | 28.87 +- 11.52 (95%)
SPRT  | 5.0+0.05s Threads=8 Hash=64MB
LLR   | 2.95 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 1520 W: 393 L: 267 D: 860

A later patch that failed yellow was pushed to make Stash's behavior much more stable when reaching positions with lots of fail-highs (such as mate positions).

ELO   | 0.76 +- 1.58 (95%)
SPRT  | 40.0+0.40s Threads=1 Hash=64MB
LLR   | -2.97 (-2.94, 2.94) [0.00, 5.00]
Games | N: 64400 W: 11243 L: 11102 D: 42055

Changed Razoring method so that it only triggers at depth 1.

ELO   | 8.05 +- 5.24 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 3.04 (-2.94, 2.94) [0.00, 5.00]
Games | N: 7424 W: 1721 L: 1549 D: 4154

Add draw specifications for 5-man endgames with an extra pawn on one side, such as KRPvKR, KBPvKN and such. (KQPvKQ isn't supported yet.)

ELO   | 3.29 +- 2.65 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 28640 W: 6349 L: 6078 D: 16213

Added bonuses for connected Pawns (Phalanx and Defender) in Pawn eval.

ELO   | 25.38 +- 10.40 (95%)
SPRT  | 8.0+0.08s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [0.00, 5.00]
Games | N: 2002 W: 540 L: 394 D: 1068

ELO   | 18.57 +- 8.45 (95%)
SPRT  | 40.0+0.40s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 2640 W: 608 L: 467 D: 1565

Changed (3 changes)

Allowed qsearch to differentiate between PV and non-PV nodes. This also permits querying the PV from qsearch.

ELO   | 1.29 +- 3.16 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 20200 W: 4433 L: 4358 D: 11409

Removed the condition that used to disable aspiration windows with high scores. (Latest TT patches made Stash much more competitive at finding mates without this condition. This is technically a functional change, but no tests were run because the value is above the resign threshold used for test games.)
Added a script for generating opening books in PGN format to the utils/ folder, and changed the FEN extractor to do a shallow search before adding FENs to the dataset.

v30.0 (2021-04-30)

Regression test

LTC:

ELO   | 51.14 +- 4.40 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 0.50]
Games | N: 8472 W: 2144 L: 906 D: 5422

Fixed (1 change)

Fixed continuation history indexing (was looking for the piece on the 'to' square instead of the 'from' square).

ELO   | 6.33 +- 5.18 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.97 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 6040 W: 1115 L: 1005 D: 3920

Removed (2 changes)

Removed Knight pair and Rook pair bonuses from the evaluation function.

ELO   | 2.08 +- 3.73 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 14867 W: 3370 L: 3281 D: 8216

Removed Internal Iterative Reductions from the search.

ELO   | 0.31 +- 2.48 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 34064 W: 7711 L: 7681 D: 18672

ELO   | 3.53 +- 4.08 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 9856 W: 1795 L: 1695 D: 6366

Performance (14 changes)

Made the continuation history additionally use our previous move, and changed the butterfly history indexing from piece to color-only.

ELO   | 3.58 +- 2.85 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.02 (-2.94, 2.94) [0.00, 5.00]
Games | N: 25904 W: 6041 L: 5774 D: 14089

ELO   | 3.04 +- 2.44 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.97 (-2.94, 2.94) [0.00, 5.00]
Games | N: 27000 W: 4796 L: 4560 D: 17644

Disabled aspiration windows for very high scores to allow Stash to find mates faster.

ELO   | 1.57 +- 3.91 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.97 (-2.94, 2.94) [-5.00, 0.00]
Games | N: 13924 W: 3228 L: 3165 D: 7531

Made the Late Move Reductions start from move 4 instead of 5.

ELO   | 2.59 +- 1.95 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.97 (-2.94, 2.94) [0.00, 5.00]
Games | N: 54686 W: 12535 L: 12127 D: 30024

ELO   | 2.55 +- 2.04 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 38850 W: 6912 L: 6627 D: 25311

Changed later to start from move 5 at the root, and 3 otherwise.

ELO   | 6.19 +- 4.50 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 10448 W: 2482 L: 2296 D: 5670

ELO   | 2.45 +- 1.93 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.99 (-2.94, 2.94) [0.00, 5.00]
Games | N: 43024 W: 7607 L: 7304 D: 28113

Made quiescence search use the TT search score in place of the static eval if the TT bounds suggest a more accurate result.

ELO   | 8.99 +- 5.73 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 6336 W: 1507 L: 1343 D: 3486

Changed Killer move array to update as a queue instead of a stack (meaning old killer moves get replaced by newer ones).

ELO   | 5.71 +- 4.23 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 11552 W: 2677 L: 2487 D: 6388

Changed piece mobility detection to account for pins.

ELO   | 6.15 +- 4.50 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 10623 W: 2569 L: 2381 D: 5673

Added Futility Pruning on parent nodes (meaning checking if a move is unlikely to improve alpha before playing it).

ELO   | 2.08 +- 3.73 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 14867 W: 3370 L: 3281 D: 8216

ELO   | 6.70 +- 4.41 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 8198 W: 1495 L: 1337 D: 5366

Added endgame functions analyzing very specific material configurations (like KQvKP or KRvKN).

ELO   | 3.76 +- 3.00 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [0.00, 5.00]
Games | N: 23448 W: 5474 L: 5220 D: 12754

ELO   | 3.88 +- 3.11 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 21848 W: 5115 L: 4871 D: 11862

Re-enabled check extensions that were disabled after a messed up scope.

ELO   | 9.34 +- 6.83 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 4504 W: 1084 L: 963 D: 2457

Replaced square-to-bitboard conversion by using a lookup table instead of always computing the shift.

ELO   | 4.98 +- 3.79 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.02 (-2.94, 2.94) [0.00, 5.00]
Games | N: 14660 W: 3439 L: 3229 D: 7992

Made the KXK specialization more tolerant, changed some endgame scaling factors based on pawn count.

ELO   | 9.09 +- 5.73 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [0.00, 5.00]
Games | N: 6308 W: 1495 L: 1330 D: 3483

Made Late Move Pruning more aggressive for positions where the eval doesn't improve.

ELO   | 16.20 +- 8.08 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.99 (-2.94, 2.94) [0.00, 5.00]
Games | N: 3240 W: 815 L: 664 D: 1761

Increased history score granularity (from 1024 to 8192).

ELO   | 5.63 +- 4.17 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [0.00, 5.00]
Games | N: 11848 W: 2735 L: 2543 D: 6570

Increased Late Move Pruning max depth (from 3 to 5).

ELO   | 6.73 +- 4.71 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [0.00, 5.00]
Games | N: 9184 W: 2111 L: 1933 D: 5140

Changed (1 change)

Very large refactor of the code, including extensive variable renaming. As a side result the compilation time halved on some systems.

ELO   | 4.48 +- 8.52 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.01 (-2.94, 2.94) [-10.00, 0.00]
Games | N: 2944 W: 698 L: 660 D: 1586

v29.0 (2021-03-05)

Regression test

LTC:

ELO   | 54.19 +- 10.27 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.97 (-2.94, 2.94) [45.00, 50.00]
Games | N: 1590 W: 414 L: 168 D: 1008

Fixed (2 changes)

Fixed memory leaks when using score pairs in UCI options.
Fixed some potential timeouts during CI due to singular extensions on the 'depth 20' test by adding a 'movetime' limit.

Performance (4 changes)

Added staged move generation to the movepicker, and placed bad captures to the end of the move list.

ELO   | 17.86 +- 8.64 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.04 (-2.94, 2.94) [0.00, 5.00]
Games | N: 3018 W: 810 L: 655 D: 1553

Replaced insertion in place_top_move() by swapping, effectively moving less elements when retrieving the next move to analyze.

ELO   | 9.18 +- 5.87 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [0.00, 5.00]
Games | N: 6436 W: 1628 L: 1458 D: 3350

Added Static Exchange Evaluation Pruning, and stopped pruning capture moves on Late Move Pruning.

ELO   | 9.94 +- 6.21 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.97 (-2.94, 2.94) [0.00, 5.00]
Games | N: 5874 W: 1523 L: 1355 D: 2996

ELO   | 3.95 +- 3.07 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 3.00 (-2.94, 2.94) [0.00, 5.00]
Games | N: 18098 W: 3442 L: 3236 D: 11420

Replaced Late Move Pruning threshold (from quiet move count to move count).

ELO   | 18.75 +- 8.76 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.01 (-2.94, 2.94) [0.00, 5.00]
Games | N: 2782 W: 716 L: 566 D: 1500

ELO   | 24.52 +- 9.38 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 3.01 (-2.94, 2.94) [0.00, 5.00]
Games | N: 1930 W: 422 L: 286 D: 1222

Changed (2 changes)

Added some helper macros to facilitate UCI option structure usage when running SPSA tests.
Removed 'mate score' handling in time management to allow mates to be correctly solved during search.

v28.0 (2021-02-23)

Regression test

LTC:

ELO   | 55.80 +- 10.33 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.98 (-2.94, 2.94) [45.00, 50.00]
Games | N: 1972 W: 604 L: 290 D: 1078

Fixed (3 changes)

Fixed the multi-job support for the command make re -j.
Fixed the compilation issues for Windows.

Fixed the center files' bonus detection for Knight Outposts.

ELO   | 0.95 +- 3.09 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 23019 W: 5513 L: 5450 D: 12056

Performance (12 changes)

Simplified TT saving by removing redundant checks.

ELO   | 5.30 +- 5.48 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 7472 W: 1869 L: 1755 D: 3848

Simplified material value updates, added a condition to trigger King Safety code only when needed, simplified OCB check.

ELO   | 3.50 +- 2.77 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.97 (-2.94, 2.94) [0.00, 4.00]
Games | N: 28704 W: 6976 L: 6687 D: 15041

Added Knight Outposts and Shielded minor pieces as eval bonuses.

ELO   | 10.16 +- 6.34 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.02 (-2.94, 2.94) [0.00, 5.00]
Games | N: 5984 W: 1644 L: 1469 D: 2871

Enlarged the Pawn table cache, resetting it only when 'ucinewgame' is sent, and saved the Pawn attack span in it for quickly finding outposts.

ELO   | 8.26 +- 6.57 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.97 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 5008 W: 1230 L: 1111 D: 2667

Removed the bonus for Far Outposts in the eval.

ELO   | 2.36 +- 1.60 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.03 (-2.94, 2.94) [0.00, 5.00]
Games | N: 85712 W: 20551 L: 19968 D: 45193

Increased the history scores' granularity.

ELO   | 3.73 +- 3.00 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.97 (-2.94, 2.94) [0.00, 5.00]
Games | N: 24192 W: 5810 L: 5550 D: 12832

Tuned the Initiative term, along with the Shielded and Outposts terms.

ELO   | 8.14 +- 5.33 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.12 (-2.94, 2.94) [0.00, 5.00]
Games | N: 7856 W: 1983 L: 1799 D: 4074

Added back candidate passers to the pawn evaluation.

ELO   | 6.87 +- 4.85 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [0.00, 5.00]
Games | N: 9459 W: 2365 L: 2178 D: 4916

Changed the condition to trigger Late Move Reductions from "not in check" to "not a capture or promotion".

ELO   | 6.89 +- 4.88 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.97 (-2.94, 2.94) [0.00, 5.00]
Games | N: 9280 W: 2304 L: 2120 D: 4856

ELO   | 4.99 +- 3.74 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 12944 W: 2631 L: 2445 D: 7868

Added a function for KXK endgames (including the infamous KBNK situation), and disabled Null Move Pruning when the side to move has no pieces left.

ELO   | 2.61 +- 1.94 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 57580 W: 13719 L: 13287 D: 30574

ELO   | 4.04 +- 3.16 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 17824 W: 3528 L: 3321 D: 10975

Adjusted Late Move Reductions' value based on the move history.

ELO   | 3.43 +- 2.76 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 28864 W: 7010 L: 6725 D: 15129

ELO   | 4.72 +- 3.55 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.97 (-2.94, 2.94) [0.00, 5.00]
Games | N: 13846 W: 2707 L: 2519 D: 8620

Added Singular Extensions to the search.

ELO   | 5.91 +- 4.38 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 11280 W: 2732 L: 2540 D: 6008

ELO   | 13.50 +- 6.87 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.98 (-2.94, 2.94) [0.00, 5.00]
Games | N: 3708 W: 773 L: 629 D: 2306

Changed (5 changes)

Changed seeding to accelerate magic bb generation.
Changed a lot of function prototypes to make the code cleaner.
Replaced the color-specialized functions in movegen by generic ones.
Moved the code used to print the PV string to a new function.

Merged the root search function with the standard search.

ELO   | -2.38 +- 2.45 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | -2.97 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 36119 W: 8354 L: 8601 D: 19164

v27.0 (2021-01-14)

Regression test

LTC:

ELO   | 76.51 +- 14.31 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 3.01 (-2.94, 2.94) [45.00, 50.00]
Games | N: 992 W: 325 L: 110 D: 557

Removed (1 change)

Removed scaling for KPK and pawnful endgames.

ELO   | -1.73 +- 1.73 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | -2.96 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 76000 W: 18423 L: 18802 D: 38775

Performance (4 changes)

Improved the mobility zone by excluding rammed pawns and pawns on low ranks.

ELO   | 19.95 +- 9.63 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 2720 W: 817 L: 661 D: 1242

Added a countermove history.

ELO   | 6.07 +- 4.51 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [0.00, 5.00]
Games | N: 11504 W: 3004 L: 2803 D: 5697

Added a continuation history.

ELO   | 6.06 +- 4.51 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 11232 W: 2873 L: 2677 D: 5682

Stopped searching bad captures in Quiescence Search, except if best_value suggests that we're getting mated. (negative-SEE pruning)

ELO   | 67.39 +- 18.03 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [0.00, 5.00]
Games | N: 736 W: 257 L: 116 D: 363

Changed (4 changes)

Merged some files with very small functions (like the UCI command handlers).

Merged the piece functions in movegen to a single generic one.

ELO   | 1.33 +- 3.44 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.95 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 19360 W: 4831 L: 4757 D: 9772

Added support for large transposition tables, currently up to 32 TB of memory.

ELO   | -0.63 +- 1.54 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.98 (-2.94, 2.94) [-4.00, 1.00]
Games | N: 96928 W: 24036 L: 24211 D: 48681

Changed Makefile to include x86-64-modern compilation flags by default.

v26.0

Regression test

LTC:

ELO   | 68.36 +- 12.97 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [45.00, 50.00]
Games | N: 1184 W: 370 L: 140 D: 674

Fixed (2 changes)

Fixed a stack overflow on some systems when the search got too deep. (As a result each worker uses a minimum of 8 MB for its stack.)
Fixed an array overflow in King Safety when the opponent has exactly 8 pieces attacking the King.

Performance (4 changes)

Changed some eval terms' positions (like the piece pair bonus/malus in the corresponding piece eval function) to gain some speed.

ELO   | 2.59 +- 1.72 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.99 (-2.94, 2.94) [0.00, 5.00]
Games | N: 88480 W: 25262 L: 24602 D: 38616

Changed the Initiative static value to a dynamic one, based on all pieces that are threatened by lower-valued ones. The final score is given like this:

stm_iv = !in_check + stm_threats;
opp_iv = opp_threats;
bonus = Initiative * (stm_iv * stm_iv - opp_iv * opp_iv);

so that having attacks on the opponent's pieces is rewarded even more if we are the side to move.

ELO   | 10.13 +- 6.50 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.99 (-2.94, 2.94) [0.00, 5.00]
Games | N: 6208 W: 1849 L: 1668 D: 2691

ELO   | 7.77 +- 5.23 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 7600 W: 1790 L: 1620 D: 4190

Created a "real" time managenement system, using the type of bestmove found (capture with very large SEE, checking move, only-move, etc.), the eval direction (the more it raises, the quicker the move is played), and the bestmove stability (stable bestmoves being played quicker). Increased max time usage to time / sqrt(movestogo) so that iterations have a high chance of being finished before sending the bestmove.
```
ELO   | 30.96 +- 12.32 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [0.00, 5.00]
Games | N: 1744 W: 572 L: 417 D: 755

ELO   | 29.02 +- 11.82 (95%)
SPRT  | 40/10.0s Threads=1 Hash=16MB
LLR   | 3.02 (-2.94, 2.94) [0.00, 5.00]
Games | N: 1872 W: 603 L: 447 D: 822
```

Moved stand-pat evaluation after TT probing in Quiescence Search.

ELO   | 35.68 +- 13.22 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 1456 W: 471 L: 322 D: 663

ELO   | 28.46 +- 10.64 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 1664 W: 406 L: 270 D: 988

Added a scaling function for endgames. (Note: this patch has been done via two different methods, the first one was using a material key and mapping endgames to specialized evals, the second only used the material quantity to evaluate the "drawishness" of the eval. The second one looked slightly worse, but allowed me to remove about 550+ lines of code with little to no effect. Here's the result for the initial version:
```
ELO   | 3.94 +- 3.09 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 3.03 (-2.94, 2.94) [0.00, 5.00]
Games | N: 20192 W: 4330 L: 4101 D: 11761
```
And here's the result for the massive simplification relative to the initial version:
```
ELO   | -1.65 +- 4.12 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | -2.95 (-2.94, 2.94) [0.00, 5.00]
Games | N: 13472 W: 3295 L: 3359 D: 6818
```
So the overall gain is likely about 2-3 Elo.)

Increased LMR by one ply for non-PV nodes.

ELO   | 5.09 +- 3.92 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=16MB
LLR   | 3.00 (-2.94, 2.94) [0.00, 5.00]
Games | N: 14944 W: 3819 L: 3600 D: 7525

ELO   | 4.57 +- 3.48 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 3.01 (-2.94, 2.94) [0.00, 5.00]
Games | N: 15354 W: 3190 L: 2988 D: 9176

Changed (1 change)

Added a new option structure to easily add UCI options for tuning sessions, with any framework like SPSA, chess-tuning-tools or such.

Files

CHANGELOG.md

Latest commit

History

CHANGELOG.md

File metadata and controls

v35.0 (2023-10-16)

Regression test

v34.0 (2022-11-03)

Regression test

v33.0 (2022-05-12)

Regression test

v32.0 (2021-12-02)

Regression test

Fixed (2 changes)

Performance (13 changes)

Changed (8 changes)

v31.0 (2021-06-26)

Regression test

Fixed (3 changes)

Performance (11 changes)

Changed (3 changes)

v30.0 (2021-04-30)

Regression test

Fixed (1 change)

Removed (2 changes)

Performance (14 changes)

Changed (1 change)

v29.0 (2021-03-05)

Regression test

Fixed (2 changes)

Performance (4 changes)

Changed (2 changes)

v28.0 (2021-02-23)

Regression test

Fixed (3 changes)

Performance (12 changes)

Changed (5 changes)

v27.0 (2021-01-14)

Regression test

Removed (1 change)

Performance (4 changes)

Changed (4 changes)

v26.0

Regression test

Fixed (2 changes)

Performance (4 changes)

Changed (1 change)