Potential's NNUE Progress

Potential's NNUE Progression

ID	Network	Architecture & Concept	Training Data	Results (Elo)
1	8hl	Replaced HCE with a basic 8 Hidden Layer architecture.	psqt/material-only (1M games, ~80M pos, 80 sb)	Fixed: +111.12 STC: +12.62
2	8hl-hm	Implemented Horizontal Mirroring (HM).	psqt/material-only (1M games, ~80M pos, 80 sb)	STC: +53.57
3	32hl	Expanded the hidden layer size to 32.	psqt/material-only (1M games, ~80M pos, 80 sb)	STC: +129.17 LTC: +138.20
4	64hl	Expanded the hidden layer size to 64.	psqt/material-only (1M games, ~80M pos, 160 sb)	Fixed: +55.70 STC: +46.35 LTC: +47.87
5	64hl-2	Fully regenerated data using the 64hl evaluation network.	64hl eval (2M games, ~147M pos, 160 sb)	Fixed: +54.95 STC: +43.96 LTC: +48.85
6	64hl-3	Switched to linear WDL and rescaled the eval.	64hl eval (2M games, ~147M pos, 160 sb)	Fixed: -2.48 STC: +9.78 LTC: +15.35
7	64hl-4	Introduced output buckets.	64hl eval (2M games, ~147M pos, 160 sb)	Fixed: +8.36 STC: +13.42 LTC: +15.78
8	128hl	Expanded the hidden layer size to 128 (First attempt).	64hl eval (2M games, ~147M pos, 160 sb)	Fixed: +75.71 STC: +36.97 LTC: +36.16
9	128hl-2	Fully regenerated data using the first 128hl net.	128hl eval (5M games, ~307M pos, 160 sb)	Fixed: +13.54 STC: +27.81 LTC: +23.08
10	128hl-3	Introduced 4 input buckets.	128hl eval (5M games, ~307M pos, 160 sb)	Fixed: +9.65 STC: +4.96 LTC: +4.97
11	256hl	Expanded the hidden layer size to 256.	128hl eval (5M games, ~307M pos, 160 sb)	Fixed: +39.85 STC: +6.78 LTC: +8.52
12	256hl-2	New fresh games using the first 256hl net.	256hl eval (5M games, 800 warmup + 160 sb)	Fixed: +6.67 STC: +6.40 LTC: +14.51
13	512hl	Expanded the hidden layer size to 512.	1st & 2nd 256hl (10M games, ~700M pos, 160 sb)	Fixed: +53.37 STC: +13.01 LTC: +37.64