Skip to content

Latest commit

 

History

History
520 lines (390 loc) · 10.4 KB

CHANGELOG.md

File metadata and controls

520 lines (390 loc) · 10.4 KB
Raw

v2.1.0

1/11/2024

  • Regression

STC 8.0+0.08 | noob_4moves

Score of New vs v2.0.0: 608 - 303 - 1089  [0.576] 2000
Elo difference: 53.4 +/- 10.2, LOS: 100.0 %, DrawRatio: 54.4 %

LTC 40.0+0.40 | noob_4moves

Score of New vs v2.0.0: 249 - 68 - 683  [0.591] 1000
Elo difference: 63.6 +/- 11.9, LOS: 100.0 %, DrawRatio: 68.3 %

1/9/2024

  • More LTC Tune

LTC 40.0+0.40 | Pohl

Score of New vs Master: 112 - 100 - 270  [0.512] 482
Elo difference: 8.7 +/- 20.6, LOS: 79.5 %, DrawRatio: 56.0 %

1/7/2024

  • Regression vs v2.0.0

LTC 40.0+0.40 | 8moves_v3

Score of New vs v2.0.0: 201 - 57 - 542  [0.590] 800
Elo difference: 63.2 +/- 13.4, LOS: 100.0 %, DrawRatio: 67.8 %

SMP 30.0+0.30 | 8moves_v3

Score of New_4T_256 vs v2.0.0_4T_256: 80 - 7 - 113  [0.682] 200
Elo difference: 132.9 +/- 30.6, LOS: 100.0 %, DrawRatio: 56.5 %
Ordo: +134.2

1/6/2024

  • Tune and limit stability in time management

STC 8.0+0.08 | Pohl

Score of New vs Master: 267 - 246 - 500  [0.510] 1013
Elo difference: 7.2 +/- 15.2

1/5/2024

  • Add cutnode conditions

STC 8.0+0.08 | Pohl

Score of New_16 vs Master_16: 345 - 318 - 680  [0.510] 1343
Elo difference: 7.0 +/- 13.0

1/2/2024

  • 30+0.3 5k games tunes

Probably didn't lose elo but whatever

LMR wasn't tuned due to an initial bug

LTC 40.0+0.40 | Pohl

Score of New vs Master: 211 - 205 - 535  [0.503] 951
Elo difference: 2.2 +/- 14.6, LOS: 61.6 %, DrawRatio: 56.3 %
Ordo: +2.4

11/2/2023

  • New bucketed net: bingshan

Positive elo gain, just trust me

10/30/2023

  • Simplify History (formula from Stormphrax)

STC 8.0+0.08 | Pohl

Score of New vs Master: 228 - 199 - 490  [0.516] 917
Elo difference: 11.0 +/- 15.3, LOS: 92.0 %, DrawRatio: 53.4 %
Ordo: +12.0

10/22/2023

  • Continuation History

STC 8.0+0.08 | Pohl

Score of New vs Master: 878 - 841 - 1714  [0.505] 3433
Elo difference: 3.7 +/- 8.2, LOS: 81.4 %, DrawRatio: 49.9 %
Ordo: +4.2

LTC 40.0+0.40 | Pohl

Score of New vs Master: 97 - 79 - 205  [0.524] 381
Elo difference: 16.4 +/- 23.7

10/2/2023

  • Change to standard aspiration window algorithm

STC 8.0+0.08 | Pohl

Score of New vs Master: 388 - 326 - 795  [0.521] 1509
Elo difference: 14.3 +/- 12.0, LOS: 99.0 %, DrawRatio: 52.7 %
Ordo: +15.8

LTC 40.0+0.40 | Pohl

Score of New vs Master: 158 - 142 - 376  [0.512] 676
Ordo: +9.6

v2.0.0

9/23/2023

  • New net: Xuebeng
    • 512 hidden neurons
    • Squared Clipped ReLU

LTC 40.0+0.40 | Pohl

Score of New vs Master: 616 - 497 - 1088  [0.527] 2201
Elo difference: 18.8 +/- 10.3, LOS: 100.0 %, DrawRatio: 49.4 %
SPRT: llr 2.95 (100.1%), lbound -2.94, ubound 2.94 - H1 was accepted
Ordo: +21.6

~+2 in STC

9/20/2023

  • Progress Check/Regression

LTC 40.0+0.40 | Pohl

Score of Dev vs v1.5.0: 274 - 40 - 186  [0.734] 500
Elo difference: 176.3 +/- 25.0, LOS: 100.0 %, DrawRatio: 37.2 %
Ordo: +215.2

Score of Dev vs Defenchess 2.2: 222 - 139 - 139  [0.583] 500
Elo difference: 58.2 +/- 26.1, LOS: 100.0 %, DrawRatio: 27.8 %
Ordo: +67.8

SMP 4CPU 20.0+0.20 | Pohl

Score of Dev vs v1.5.0: 57 - 12 - 31  [0.725] 100
Elo difference: 168.4 +/- 60.5, LOS: 100.0 %, DrawRatio: 31.0 %
Ordo: +190.8

Estimate CCRL 3350 1CPU Blitz, 3350 4CPU 40/15, 3260 1CPU 40/15.

9/16/2023

  • Consider History in LMR

STC 8.0+0.08 | Pohl

Score of New_16 vs Master_16: 326 - 322 - 715  [0.501] 1363
Elo difference: 1.0 +/- 12.7, LOS: 56.2 %, DrawRatio: 52.5 %

9/15/2023

  • Simplify LMR logic for captures

STC 8.0+0.08 | Pohl

Score of New vs Master: 1402 - 1353 - 3346  [0.504] 6101
Elo difference: 2.8 +/- 5.9, LOS: 82.5 %, DrawRatio: 54.8 %
SPRT: llr 2.96 (100.6%), lbound -2.94, ubound 2.94 - H1 was accepted
Ordo: +3.2

LTC 40.0+0.40 | Pohl

Score of New vs Master: 128 - 124 - 296  [0.504] 548
Elo difference: 2.5 +/- 19.7, LOS: 59.9 %, DrawRatio: 54.0 %
Ordo: +2.4

9/14/2023

  • Flip LMR condition for pv

STC 8.0+0.08 | Pohl

Score of New vs Master: 331 - 291 - 748  [0.515] 1370
Elo difference: 10.1 +/- 12.4, LOS: 94.6 %, DrawRatio: 54.6 %
Ordo: +11.0

9/5/2023

  • Fix SEE bug where threshold > pawn

STC 8.0+0.08 | Pohl

Score of New vs Master: 305 - 205 - 617  [0.544] 1127
Elo difference: 30.9 +/- 13.6, LOS: 100.0 %, DrawRatio: 54.7 %
SPRT: llr 2.96 (100.7%), lbound -2.94, ubound 2.94 - H1 was accepted

9/4/2023

  • Scale history down ("Gravity")
    • Suggested by Engine Programming Discord server

STC 8.0+0.08 | Pohl

Score of New vs Master: 730 - 696 - 1705  [0.505] 3131
Elo difference: 3.8 +/- 8.2, LOS: 81.6 %, DrawRatio: 54.5 %
Ordo: +4.2

9/2/2023

  • Age History instead of clearing it after each search
    • Mostly STC gains because LTC recalculates histories

STC 8.0+0.08 | Pohl

Score of New vs Master: 218 - 159 - 430  [0.537] 807
Elo difference: 25.4 +/- 16.4, LOS: 99.9 %, DrawRatio: 53.3 %
SPRT: llr 2.95 (100.1%), lbound -2.94, ubound 2.94 - H1 was accepted
Ordo: +27.8

LTC 40.0+0.40 | Pohl

Score of New vs Master: 69 - 65 - 226  [0.506] 360
Elo difference: 3.9 +/- 21.9, LOS: 63.5 %, DrawRatio: 62.8 %
Ordo: +4.2

8/30/2023

  • Regression Test vs 1.5.0

STC 8.0+0.08 | Pohl

Score of New vs v1.5.0: 161 - 34 - 83  [0.728] 278
Elo difference: 171.4 +/- 36.4, LOS: 100.0 %, DrawRatio: 29.9 %
SPRT: llr 2.96 (100.4%), lbound -2.94, ubound 2.94 - H1 was accepted
Ordo: +188.3

LTC 40.0+0.40 | Pohl

Score of New vs v1.5.0: 162 - 43 - 115  [0.686] 320
Elo difference: 135.7 +/- 31.4, LOS: 100.0 %, DrawRatio: 35.9 %
SPRT: llr 2.95 (100.1%), lbound -2.94, ubound 2.94 - H1 was accepted
Ordo: +154.5

8/29/2023

  • TT Aging

STC 8.0+0.08 | Pohl

Score of New vs Master: 608 - 575 - 1304  [0.507] 2487
Elo difference: 4.6 +/- 9.4, LOS: 83.1 %, DrawRatio: 52.4 %
Ordo: +5.0

LTC 40.0+0.40 | Pohl

Score of New vs Master: 42 - 34 - 82  [0.525] 158
Elo difference: 17.6 +/- 37.7, LOS: 82.1 %, DrawRatio: 51.9 %
Ordo: +21.3

8/26/2023

  • Material Scaling

STC 10.0+0.10 | 8moves_v3

Score of New vs Master: 178 - 163 - 659  [0.507] 1000
Elo difference: 5.2 +/- 12.6, LOS: 79.2 %, DrawRatio: 65.9 %
Ordo: +5.2

LTC 40.0+0.40 | Pohl

Score of New vs Master: 161 - 143 - 361  [0.514] 665
Cutechess output lost
Ordo: +11.0

8/25/2023

  • 50 move count scaling

STC 10.0+0.10 | 8moves_v3

Score of New vs Master: 1152 - 1072 - 4332  [0.506] 6556
Elo difference: 4.2 +/- 4.9, LOS: 95.5 %, DrawRatio: 66.1 %
SPRT: llr 1.39 (47.3%), lbound -2.94, ubound 2.94
Ordo: +4.3

LTC 40.0+0.40 | Pohl

Score of New vs Master: 339 - 325 - 885  [0.505] 1549
Elo difference: 3.1 +/- 11.3, LOS: 70.7 %, DrawRatio: 57.1 %
SPRT: llr 0.0958 (3.3%), lbound -2.94, ubound 2.94
Ordo: +3.6

8/18/2023

  • New net: net008b

STC 15.0+0.10 | 1000 games | 8moves_v3

Score of New vs Master: 284 - 174 - 542  [0.555] 1000
Elo difference: 38.4 +/- 14.5, LOS: 100.0 %, DrawRatio: 54.2 %
Ordo: +39.1

8/16/2023

  • Switch to new NNUE architecture
  • New net: net007b

STC 20.0+0.10 | 1000 games | 8moves_v3

Score of New vs Master: 404 - 162 - 434  [0.621] 1000
Elo difference: 85.8 +/- 16.3, LOS: 100.0 %, DrawRatio: 43.4 %
Ordo: +86.9

LTC 40.0+0.40 | 300 games | Pohl

Score of New vs Master: 127 - 74 - 99  [0.588] 300
Elo difference: 62.0 +/- 32.5, LOS: 100.0 %, DrawRatio: 33.0 %
Ordo: +76.8

8/5/2023

  • Fix NMP bug

STC 20.0+0.10 | 1000 games | 8moves_v3

Score of New vs Master: 170 - 127 - 703  [0.521] 1000
Elo difference: 14.9 +/- 11.7, LOS: 99.4 %, DrawRatio: 70.3 %

LTC 60.0+0.60 | 1000 games | 8moves_v3

Score of New vs Master: 134 - 144 - 722  [0.495] 1000
Elo difference: -3.5 +/- 11.3, LOS: 27.4 %, DrawRatio: 72.2 %

30.0+0.30 VS 1.5.0:

Score of Master vs 1.5.0: 328 - 258 - 1414  [0.517] 2000
Elo difference: 12.2 +/- 8.2, LOS: 99.8 %, DrawRatio: 70.7 %
  • Do more NMP when position is improving

STC 30.0+0.20 | 1000 games | noob_4moves

Score of New vs Master: 140 - 121 - 739  [0.509] 1000
Elo difference: 6.6 +/- 11.0, LOS: 88.0 %, DrawRatio: 73.9 %

8/4/2023

  • New RFP parameters from Viridithas

STC 15.0+0.10 | 2000 games | 8moves_v3

Score of New vs Master: 357 - 337 - 1306  [0.505] 2000
Elo difference: 3.5 +/- 9.0, LOS: 77.6 %, DrawRatio: 65.3 %

LTC 60.0+0.60 | 1000 games | 8moves_v3

Score of New vs Master: 143 - 127 - 730  [0.508] 1000
Elo difference: 5.6 +/- 11.2, LOS: 83.5 %, DrawRatio: 73.0 %

8/3/2023

  • New LMR parameters from Viridithas

STC 14.0+0.10 | 1000 games | 8moves_v3

181 - 173 - 646  [0.504] 1000
Elo difference: 2.8 +/- 12.8, LOS: 66.5 %, DrawRatio: 64.6 %

LTC 50.0+0.50 | 1000 games | 8moves_v3

[0.510] 1000
Elo difference: 7.0
DrawRatio: 76.52 %

v1.5.0

  • Optimizations
  • Search Tuning
  • Stronger Neural Network
  • Trained on over 25 Million depth 8 positions from lichess elite database
  • Trained on 1.5 Million depth 10 endgame positions
  • LazySMP Implementation

v1.4.0

  • Search Improvements
  • Manual Tuning
  • NNUE Optimizations
  • Time Management

v1.3.1

  • Search Improvements
  • Countermove heuristic fix
  • Tuning

v1.3.0

  • Stronger Neural Network trained on 2GB of data
  • Countermove Heuristics
  • Higher bounds for History Heuristics
  • Improved Aspiration Window

v1.2.0

  • Movegen Bug fixes
  • Tuned Search parameters
  • Search Rewrite
  • Better SEE
  • Stronger Neural Network (depth 8, 500 epoch) featuring 8 buckets

v1.1.0

  • NNUE Optimizations
  • Singular Extension / MultiCut
  • More Aggressive Prunings

v1.0.0

  • Faster Movegen: heavily inspired by Surge
  • Complete Core Rewrite
  • 512-neuron NNUE trained on 50 million positions on depth 4

v0.2.2

  • Bug fixes
  • LMR tuning
  • New SEE algorithm
  • Aspiration Windows

v0.2.1

  • Bug fixes
  • UCI options
  • Improvements on Search

v0.2: Search

  • History heuristics, killer heuristics
  • Better LMR
  • Reversed Futility Pruning
  • Null Move Pruning
  • Razoring
  • Time management
  • Better Transposition Table
  • Static Exchange Evaluation
  • Stronger NNUE network: Flake 2
  • Trained on human games on https://database.lichess.org/ and more engine games.
  • Trained on one million endgame positions
  • 728 -> dense -> 512 -> clipped_relu -> 512 -> dense -> 1 + PSQT

v0.1: NNUE, ~1700 ELO

  • Efficiently Updatable Neural Network trained on top-level engine tournaments
  • Current model: 728 -> dense -> 128 -> clipped_relu -> 128 -> dense -> 5 + PSQT
  • Forward Pass
  • Tuned LMR
  • Bishop pair, doubled pawns, etc.

v0.0: Base, ~1400 ELO

  • Bitboard board representation
  • Magic bitboards
  • Negamax Search with Alpha-Beta pruning
  • Quiescence Search with stand-pat pruning
  • MVV_LVA
  • LMR
  • HCE PSQT Evaluation