Releases: TheBlackPlague/StockNemo
Releases · TheBlackPlague/StockNemo
Stingray (5.7.0.0)
What's Changed
- Decrease learning rate and drop learning rate during Neural Network Training. by @TheBlackPlague in #87
- Decrement History Heuristic Scores for non-best moves. by @TheBlackPlague in #88
NNUE:
IN ACCUMULATOR HIDDEN OUT
______________ _______ ______________________________________________________________ _____
| WHITE: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) \ | | |
| | | | | \ | | |
| | | | | CONCATENATE(ColorToMove): (512) | -> | (1) |
| | | | | / | | |
| BLACK: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) / | | |
-------------- ------- -------------------------------------------------------------- -----
TAG: EXP-0006
Data: 1.7B FEN --- DEPTH: 9 (1.2B) + NODES: 5K (500M)
Batch Size: 500000
Epochs: 80
LR: 5e-3
LR Drop: 30
WDL: 0.4
Full Changelog: 5.6.0.0...5.7.0.0
ELO Difference
Using UHO_XXL_+0.90_+1.19.epd:
TC: 10s + 0.1s (STC)
ELO | 73.55 +- 5.43 (95%)
CONF | 10.0+0.10s Threads=1 Hash=16MB
GAMES | N: 10000 W: 4156 L: 2070 D: 3774
TC: 60s + 0.6s (LTC)
ELO | 70.94 +- 5.12 (95%)
CONF | 60.0+0.60s Threads=1 Hash=256MB
GAMES | N: 10000 W: 3802 L: 1788 D: 4410
Contributors
TheBlackPlague
Assets 13
- 27.1 MB
2022-09-28T12:53:20Z - 26.5 MB
2022-09-28T12:53:14Z - 28.1 MB
2022-09-28T12:53:08Z - 28.1 MB
2022-09-28T12:53:02Z - 28.1 MB
2022-09-28T12:52:56Z - 28.1 MB
2022-09-28T12:52:49Z - 28.1 MB
2022-09-28T12:52:43Z - 26.7 MB
2022-09-28T12:52:37Z - 27.2 MB
2022-09-28T12:52:25Z - 28.4 MB
2022-09-28T12:52:31Z -
2022-09-28T11:34:35Z -
2022-09-28T11:34:35Z - Loading
Stingray (5.6.0.0)
What's Changed
- Templated Make Move for Normal, Classical, and NN Updates. by @TheBlackPlague in #83
- Update History Heuristics on Beta Cutoffs. by @TheBlackPlague in #84
- Increase neural network training batch size & reliance on game result. by @TheBlackPlague in #85
NNUE:
IN ACCUMULATOR HIDDEN OUT
______________ _______ ______________________________________________________________ _____
| WHITE: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) \ | | |
| | | | | \ | | |
| | | | | CONCATENATE(ColorToMove): (512) | -> | (1) |
| | | | | / | | |
| BLACK: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) / | | |
-------------- ------- -------------------------------------------------------------- -----
Data: 1.7B FEN --- DEPTH: 9 (1.2B) + NODES: 5K (500M)
Epochs: 30
WDL: 0.4
Batch Size: 500K
Full Changelog: 5.5.0.0...5.6.0.0
ELO Difference
Using UHO_XXL_+0.90_+1.19.epd:
TC: 10s + 0.1s (STC)
ELO | 107.31 +- 5.72 (95%)
CONF | 10.0+0.10s Threads=1 Hash=16MB
GAMES | N: 10000 W: 4845 L: 1851 D: 3304
TC: 60s + 0.6s (LTC)
ELO | 110.14 +- 7.68 (95%)
CONF | 60.0+0.60s Threads=1 Hash=256MB
GAMES | N: 5000 W: 2294 L: 760 D: 1946
Contributors
TheBlackPlague
Assets 13
Stingray (5.5.0.0)
What's Changed
- Search Stack to allow better API for recursive search. by @TheBlackPlague in #71
- Simplify Transposition Table Lookup in QSearch. by @TheBlackPlague in #72
- Increase search benchmark sample size. by @TheBlackPlague in #73
- Improve Transposition Table Replacement Scheme. by @PGG106 in #74
- Implement Internal Iterative Reductions. by @TheBlackPlague in #75
- Improve Neural Network Operation Cache Locality. by @TheBlackPlague in #78
- Increase Accumulator Stack Size. by @TheBlackPlague in #79
- Implement Hybrid Data Neural Network. by @TheBlackPlague in #80
- Implement Transposition Table Entry Priority. by @TheBlackPlague in #81
New Contributors
NNUE:
IN ACCUMULATOR HIDDEN OUT
______________ _______ ______________________________________________________________ _____
| WHITE: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) \ | | |
| | | | | \ | | |
| | | | | CONCATENATE(ColorToMove): (512) | -> | (1) |
| | | | | / | | |
| BLACK: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) / | | |
-------------- ------- -------------------------------------------------------------- -----
Data: 1.7B FEN --- DEPTH: 9 (1.2B) + NODES: 5K (500M)
Epochs: 30
Full Changelog: 5.4.0.0...5.5.0.0
ELO Difference
Using UHO_XXL_+0.90_+1.19.epd:
TC: 10s + 0.1s (STC)
ELO | 65.13 +- 5.51 (95%)
CONF | 10.0+0.10s Threads=1 Hash=16MB
GAMES | N: 10000 W: 4136 L: 2283 D: 3581
TC: 60s + 0.6s (LTC)
ELO | 50.88 +- 5.25 (95%)
CONF | 60.0+0.60s Threads=1 Hash=256MB
GAMES | N: 10000 W: 3679 L: 2225 D: 4096
Contributors
TheBlackPlague and PGG106
Assets 13
Stingray (5.4.0.0)
What's Changed
- Introduce a templated-generic Move Search. by @TheBlackPlague in #67
- Add Selective Depth to Engine Output. by @TheBlackPlague in #68
- Use Transposition Table in QSearch. by @TheBlackPlague in #69
- Implement Approximated Static Exchange Evaluation in QSearch. by @TheBlackPlague in #70
NNUE:
IN ACCUMULATOR HIDDEN OUT
______________ _______ ______________________________________________________________ _____
| WHITE: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) \ | | |
| | | | | \ | | |
| | | | | CONCATENATE(ColorToMove): (512) | -> | (1) |
| | | | | / | | |
| BLACK: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) / | | |
-------------- ------- -------------------------------------------------------------- -----
Data: 1.2B FEN --- DEPTH: 9
Epochs: 40
Full Changelog: 5.3.0.0...5.4.0.0
ELO Difference
Using UHO_XXL_+0.90_+1.19.epd:
TC: 10s + 0.1s (STC)
ELO | 32.44 +- 3.57 (95%)
CONF | 10.0+0.10s Threads=1 Hash=16MB
GAMES | N: 20000 W: 6431 L: 4569 D: 9000
TC: 60s + 0.6s (LTC)
ELO | 31.49 +- 6.89 (95%)
CONF | 60.0+0.60s Threads=1 Hash=256MB
GAMES | N: 5000 W: 1505 L: 1053 D: 2442
Contributors
TheBlackPlague
Assets 13
Stingray (5.3.0.0)
What's Changed
- Refactor LMR into an improved conditionally aggressive version. by @TheBlackPlague in #65
- Implement Depth 9 data-based Neural Network. by @TheBlackPlague in #66
NNUE:
IN ACCUMULATOR HIDDEN OUT
______________ _______ ______________________________________________________________ _____
| WHITE: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) \ | | |
| | | | | \ | | |
| | | | | CONCATENATE(ColorToMove): (512) | -> | (1) |
| | | | | / | | |
| BLACK: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) / | | |
-------------- ------- -------------------------------------------------------------- -----
Data: 1.2B FEN --- DEPTH: 9
Epochs: 40
Full Changelog: 5.2.0.0...5.3.0.0
ELO Difference
Using UHO_XXL_+0.90_+1.19.epd:
TC: 10s + 0.1s (STC)
ELO | 87.30 +- 5.75 (95%)
CONF | 10.0+0.10s Threads=1 Hash=16MB
GAMES | N: 10000 W: 4660 L: 2199 D: 3141
TC: 60s + 0.6s (LTC)
ELO | 85.78 +- 17.89 (95%)
CONF | 60.0+0.60s Threads=1 Hash=256MB
GAMES | N: 1000 W: 449 L: 207 D: 344
Contributors
TheBlackPlague
Assets 13
Stingray (5.2.0.0)
What's Changed
- Speed up Neural Network Inference. by @TheBlackPlague in #60
- Implement Futility Pruning. by @TheBlackPlague in #61
- Train the Neural Network for 40 epochs. by @TheBlackPlague in #62
- Make vectorized matrix multiplication more efficient. by @TheBlackPlague in #63
- Make LMR more aggressive when we are improving. by @TheBlackPlague in #64
NNUE:
- Architecture:
IN ACCUMULATOR HIDDEN OUT
______________ _______ ______________________________________________________________ _____
| WHITE: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) \ | | |
| | | | | \ | | |
| | | | | CONCATENATE(ColorToMove): (512) | -> | (1) |
| | | | | / | | |
| BLACK: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) / | | |
-------------- ------- -------------------------------------------------------------- -----
- Training Data: ~1.2B FENs
Full Changelog: 5.1.0.0...5.2.0.0
ELO Difference
Using UHO_XXL_+0.90_+1.19.epd:
TC: 10s + 0.1s (STC)
ELO | 64.88 +- 3.90 (95%)
CONF | 10.0+0.10s Threads=1 Hash=16MB
GAMES | N: 20000 W: 8285 L: 4593 D: 7122
TC: 60s + 0.6s (LTC)
ELO | 59.29 +- 11.95 (95%)
CONF | 60.0+0.60s Threads=1 Hash=256MB
GAMES | N: 2000 W: 773 L: 435 D: 792
Contributors
TheBlackPlague
Assets 13
Stingray (5.1.0.0)
What's Changed
- Refactor PieceColor API. by @TheBlackPlague in #57
- Fix crash at high plies due to check extensions. by @TheBlackPlague in #59
NNUE:
- Architecture:
IN ACCUMULATOR HIDDEN OUT
______________ _______ ______________________________________________________________ _____
| WHITE: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) \ | | |
| | | | | \ | | |
| | | | | CONCATENATE(ColorToMove): (512) | -> | (1) |
| | | | | / | | |
| BLACK: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) / | | |
-------------- ------- -------------------------------------------------------------- -----
- Training Data: ~1.2B FENs
Full Changelog: 5.0.0.1...5.1.0.0
Contributors
TheBlackPlague
Assets 13
Stingray (5.0.0.1)
What's Changed
- Update Best Move in Aspiration Search for Fail-High. by @TheBlackPlague in #54
- Scale reduction depth for null move pruning linearly. by @TheBlackPlague in #55
- Upgrade to an NNUE trained on Berserk Data. by @TheBlackPlague in #56
NNUE:
- Architecture:
IN ACCUMULATOR HIDDEN OUT
______________ _______ ______________________________________________________________ _____
| WHITE: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) \ | | |
| | | | | \ | | |
| | | | | CONCATENATE(ColorToMove): (512) | -> | (1) |
| | | | | / | | |
| BLACK: (768) | -> | (256) | -> | ClippedReLU(0, 1) -> (256) / | | |
-------------- ------- -------------------------------------------------------------- -----
- Training Data: ~1.2B FENs
Full Changelog: 5.0.0.0...5.0.0.1
ELO Difference
Using UHO_XXL_+0.90_+1.19.epd:
TC: 10s + 0.1s (STC)
ELO | 76.17 +- 5.76 (95%)
CONF | 10.0+0.10s Threads=1 Hash=16MB
GAMES | N: 10000 W: 4549 L: 2391 D: 3060
Contributors
TheBlackPlague
Assets 13
Stingray (5.0.0.0)
What's Changed
- Implement UCI Go Command: Moves To Go. by @TheBlackPlague in #51
- Implement Check Extensions. by @TheBlackPlague in #52
- Implement an Efficiently Updatable Neural Network (NNUE). by @TheBlackPlague in #53
NNUE:
Architecture: BASIC [ (768 -> 256)x2 -> 1 ]
Training Data Size: 112,500 FENs
Full Changelog: 4.0.0.0...5.0.0.0
ELO Difference
Using UHO_XXL_+0.90_+1.19.epd:
TC: 10s + 0.1s (STC)
ELO | 321.45 +- 6.17 (95%)
CONF | 10.0+0.10s Threads=1 Hash=16MB
GAMES | N: 20000 W: 16126 L: 1559 D: 2315
TC: 120s + 1s (LTC)
ELO | 358.37 +- 21.34 (95%)
CONF | 120.0+1.00s Threads=1 Hash=16MB
GAMES | N: 2000 W: 1661 L: 112 D: 227
Contributors
TheBlackPlague
Assets 13
Hammerhead (4.0.0.1)
What's Changed
- Implement UCI Go Command: Moves To Go. by @TheBlackPlague in #51
Full Changelog: 4.0.0.0...4.0.0.1
This release is a UCI command improvement. Thus, does not require ELO Difference. You may notice positive ELO Difference on repeating TCs, though, as it targets that.
Contributors
TheBlackPlague