Experiment 2
Summary
Use 16 PSQT buckets and layer stacks for SmallNet.
Description
Setup
Continued from Experiment 1.
Use (pos.count<ALL_PIECES>() - 1) / 2. Edit model/lightning_module.py, model/model.py and training_data_loader.cpp accordingly.
Set max epoch to 400.
Local Testing
10,000 fixed games
Openings: UHO_Lichess_4852_v1
Time control: 10+0.1, nodestime=69 (simulated 1+0.01)
Engine options: Threads=1, Hash=16
Epoch 100
Results of Small16Bucket-ep100 vs master (10+0.1, 1t, 16MB, UHO_Lichess_4852_v1.epd):
Elo: -2.02 +/- 4.74, nElo: -2.90 +/- 6.81
LOS: 20.21 %, DrawRatio: 41.26 %, PairsRatio: 0.98
Games: 10000, Wins: 3149, Losses: 3207, Draws: 3644, Points: 4971.0 (49.71 %)
Ptnml(0-2): [332, 1150, 2063, 1154, 301], WL/DD Ratio: 2.08Epoch 150
Epoch 200
Epoch 250
Epoch 300
Epoch 350
Epoch 400
fishtest
Epoch 250 (nn-e1e21c0c34ff.nnue)
Result: FAILED
Epoch 350 (nn-c98a3dc5ad4b.nnue)
Result: FAILED
Last updated