Caissa v1.8
- Refactor training data loading in neural net trainer
- Experiment with Adam optimizer, different activation functions (both experiments failed)
- New neural net and piece-square tables trained on 1.2B positions
- Switch back to 1536 accumulator size
- Increase last layer bucket count from 4 to 8
- Two training sessions, 48 hours in total
- Fix 'hashfull' reporting zero
- Minor search tweaks
TC=60+0.6
Score of Caissa 1.8 vs Caissa 1.7: 512 - 360 - 673 [0.549] 1545
... Caissa 1.8 playing White: 410 - 60 - 303 [0.726] 773
... Caissa 1.8 playing Black: 102 - 300 - 370 [0.372] 772
... White vs Black: 710 - 162 - 673 [0.677] 1545
Elo difference: 34.3 +/- 13.0, LOS: 100.0 %, DrawRatio: 43.6 %
SPRT: llr 2.95 (100.1%), lbound -2.94, ubound 2.94 - H1 was accepted
Full Changelog: 1.7...1.8