Motor v0.8.0

Progression Tests:

Tests on UHO_Lichess_4852_v1.epd opening book:

Elo   | 39.00 +- 3.67 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=32MB
Games | N: 10002 W: 3176 L: 2058 D: 4768
Penta | [33, 774, 2342, 1746, 106]

Elo   | 38.35 +- 6.32 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 3002 W: 878 L: 548 D: 1576
Penta | [4, 225, 721, 539, 12]

Test on 8moves_v3.pgn:

Elo   | 18.43 +- 4.42 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 3000 W: 312 L: 153 D: 2535
Penta | [0, 108, 1134, 249, 9]

Changes

• Most changes come from experiments with correction history:
  • Threat structure correction history
  • Minor structure correction history (idea from Sirius)
  • Continuation correction history
  • New weight formula
  • Qsearch correction
• New network with 8 input buckets
• New history table for move ordering indexed by material index

Motor v0.7.0

Progression Tests:

Elo   | 77.52 +- 7.70 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 2000 W: 733 L: 294 D: 973
Penta | [1, 77, 430, 466, 26]

Elo   | 74.18 +- 3.72 (95%)
Conf  | 10.0+0.10s Threads=1 Hash=32MB
Games | N: 10008 W: 3766 L: 1661 D: 4581
Penta | [22, 454, 2152, 2149, 227]

Tests on Pohl.pgn opening book

Changes

• NNUE
  • 3 king input buckets +11
  • MPE 2.5 loss function +24
  • Stricter weight clipping +8
• Static evaluation correction history
  • Pawn structure +6 (idea from Caissa)
  • Material count +3 (idea from Caissa)
  • Material structure +3 (idea from Starzix)
• Probcut +4, +2
• TT reduction in pv nodes +8.5
• QS check evasions +12.5
• Material scaling +6.75
• FP tweaks 4.5, +2.75, +7.8

Extra thanks to Zzzzz for sharing with me his OB instance.

Motor v0.6.0

• New network size: (768 -> 1536) x 2 -> 1
• Horizontal mirroring
• New faster move generator (legal move generator with kindergarten bitboards)
• Node tm
• Threat history
• Capture history
• Improving boolean in LMP

STC:

Score of Motor 0.6.0 vs Motor 0.5.0: 256 - 65 - 679  [0.596] 1000
...      Motor 0.6.0 playing White: 167 - 21 - 313  [0.646] 501
...      Motor 0.6.0 playing Black: 89 - 44 - 366  [0.545] 499
...      White vs Black: 211 - 110 - 679  [0.550] 1000
Elo difference: 67.2 +/- 11.9, LOS: 100.0 %, DrawRatio: 67.9 % 

LTC:

Score of Motor 0.6.0 vs Motor 0.5.0: 123 - 24 - 353  [0.599] 500
...      Motor 0.6.0 playing White: 88 - 9 - 153  [0.658] 250
...      Motor 0.6.0 playing Black: 35 - 15 - 200  [0.540] 250
...      White vs Black: 103 - 44 - 353  [0.559] 500
Elo difference: 69.7 +/- 16.0, LOS: 100.0 %, DrawRatio: 70.6 %

Tests on 8moves_v3.pgn opening book

Use the AVX2 version if possible.

Motor v0.5.0

• New network size: (768 -> 1024) x 2 -> 1
• Multicut
• Futility pruning
• History rework
• Movegen speedup

STC:

Score of Motor 0.5.0 vs Motor 0.4.0: 510 - 152 - 1338  [0.590] 2000
...      Motor 0.5.0 playing White: 289 - 45 - 666  [0.622] 1000
...      Motor 0.5.0 playing Black: 221 - 107 - 672  [0.557] 1000
...      White vs Black: 396 - 266 - 1338  [0.532] 2000
Elo difference: 62.9 +/- 8.6, LOS: 100.0 %, DrawRatio: 66.9 %

LTC:

Score of Motor 0.5.0 vs Motor 0.4.0: 103 - 22 - 275  [0.601] 400
...      Motor 0.5.0 playing White: 65 - 5 - 132  [0.649] 202
...      Motor 0.5.0 playing Black: 38 - 17 - 143  [0.553] 198
...      White vs Black: 82 - 43 - 275  [0.549] 400
Elo difference: 71.3 +/- 18.5, LOS: 100.0 %, DrawRatio: 68.8 %

Tests on 8moves_v3.pgn opening book

Use the AVX2 version if possible.

Motor v0.4.0

• New network size: (768 -> 512) x 2 -> 1
• SCRelu activation function
• Custom AVX2 simd
• SEE pruning
• Singular extension + Double extensions
• Soft-tm

STC:

Score of motor 0.4.0 vs motor 0.3.0: 1880 - 367 - 1653  [0.694] 3900
...      motor 0.4.0 playing White: 1424 - 48 - 479  [0.853] 1951
...      motor 0.4.0 playing Black: 456 - 319 - 1174  [0.535] 1949
...      White vs Black: 1743 - 504 - 1653  [0.659] 3900
Elo difference: 142.2 +/- 8.4, LOS: 100.0 %, DrawRatio: 42.4 %

LTC:

Score of motor 0.4.0 vs motor 0.3.0: 525 - 93 - 532  [0.688] 1150
...      motor 0.4.0 playing White: 412 - 10 - 153  [0.850] 575
...      motor 0.4.0 playing Black: 113 - 83 - 379  [0.526] 575
...      White vs Black: 495 - 123 - 532  [0.662] 1150
Elo difference: 137.2 +/- 14.7, LOS: 100.0 %, DrawRatio: 46.3 %

Use the AVX2 version if possible.

Motor v0.3.0

New network size: (768 -> 256) x 2 -> 1
History gravity + 1-ply conthist
New LMR changes
Improving boolean

Fixed movegen bug (enpassant with double discovery check).
Fixed search in max depth.

Motor v0.2.0

New network (768 -> 64) x 2 - > 1
Mostly changes with tranposition tables. Added tt move ordering and tt into qs.
New aspiration window and lmp formulas.

Elo difference: 278.3 +/- 54.9

Thanks to Jim Ablett for the compilation

Motor v0.1.0

First release of Motor. NNUE (768 -> 32) x 2 -> 1 trained with forge.