a7c8f545.nn is a (768->512)x2->1 network using SCReLU activation. It has been trained from scratch, with a completely separate lineage to the master network.

Before I begin, I'd like to thank the contributors that have at some point in the past been responsible for training the best Halogen network to date:

• @Luecx (Add support for new 'half' network architecture #370)
• @AndyGrant (Update default Neural Network [b8a1b7d5.nn] (768x512x1) #312)

Training a network from zero for Halogen purely through self play reinforcement + supervised learning was always been a long term goal of mine. The combined effort to achieve this began in March 2024, and represents the single greatest development effort in Halogen's history.

Training began with a novel implementation of TDLeaf(λ) reinforcement learning¹. The exact lineage of best networks were:

• 11.4.1_td_leaf_learn_v2.0.0 starting from randomly initialized weights trained for 8 hours to produce 768-512x2-1_g917495.nn
• 11.4.1_td_leaf_learn_v2.1.0 starting from 768-512x2-1_g917495.nn trained for 8 hours to produce 768-512x2-1_e2_g1768709.nn
• 11.4.1_td_leaf_learn_v2.2.0 starting from 768-512x2-1_e2_g1768709.nn trained for 13 hours to produce 768-512x2-1_e3_g2065630.nn

Temporal Coherence² was then used to further the training process:

• 11.4.1_td_leaf_learn_v2.3.1a starting from 768-512x2-1_e3_g2065630.nn trained for 16 hours to produce 768-512x2-1_e8_g1726579
• 11.7.0_td_leaf_learn_3.0.0 starting from 768-512x2-1_e8_g1726579 trained for 16 hours to produce 768-512x2-1_r12_g2297813.nn

By playing out Syzygy endgames and allowing the network to learn from those positions, endgame play was improved:

• 11.7.0_td_leaf_learn_3.3.0 starting from 768-512x2-1_r12_g2297813.nn trained for 16 hours to produce 768-512x2-1_r15_g2720489.nn

By adding 5% DFRC, the FRC performance was greatly improved:

• 11.7.0_td_leaf_learn_3.5.0 starting from 768-512x2-1_r15_g2720489.nn trained for 16 hours to produce 768-512x2-1_r17_g2894950.nn

By filtering openings that are wildly unbalanced +/- 500cp, the final TDLeaf(λ) network was trained:

• 11.7.0_td_leaf_learn_3.6.0 starting from 768-512x2-1_r17_g2894950.nn trained for 16 hours to produce 768-512x2-1_r18_g2771316.nn

The final TDLeaf(λ) network was Elo | -182.26 +- 8.45 (95%) to compared to master.

At this point I switched to supervised learning, using the bullet trainer³

• datagen 11.24.0 + 768-512x2-1_r18_g2771316.nn Generated 612m fens at 40k nodes. This was filtered and used to train bullet_r10_768-512x2-1-epoch100.bin:

File Path      : ...
Threads        : 20
WDL Proportion : start 0.3 end 0.3
Max Epochs     : 100
Save Rate      : 10
Batch Size     : 16384
Net Name       : bullet_r10_768-512x2-1
LR Scheduler   : start 0.001 gamma 0.1 drop every 40 epochs
Scale          : 160
Positions      : 505179452

• datagen 11.30.0 + bullet_r10_768-512x2-1-epoch100.bin Generated 739m fens at 40k nodes. This was filtered and combined with the previous 612m fens and used to train the final network bullet_r17_768-512x2-1_e50.nn:

File Path      : ...
Threads        : 20
WDL Proportion : start 0.3 end 0.3
Max Epochs     : 50
Save Rate      : 10
Batch Size     : 16384
Net Name       : bullet_r17_768-512x2-1
LR Scheduler   : start 0.001 gamma 0.95 drop every 1 epochs
Scale          : 160
Positions      : 1116214793

Elo   | 31.56 +- 5.70 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 5000 W: 1530 L: 1077 D: 2393
Penta | [29, 473, 1126, 760, 112]
http://chess.grantnet.us/test/37549/

Elo   | 30.92 +- 6.37 (95%)
Conf  | 8.0+0.08s Threads=1 Hash=8MB
Games | N: 5002 W: 1626 L: 1182 D: 2194
Penta | [73, 504, 997, 760, 167]
http://chess.grantnet.us/test/37548/