jgrusewski 796fbd01cf fix: set c51_alpha_max=0.5 — prevent C51 gradient convergence on H100 ...

H100 fold 2 showed grad_norm→0 when C51 fully replaced MSE (alpha=1.0).
MSE provides a non-vanishing gradient floor. Default 0.5 = 50/50 blend.
Hyperopt can search [0.3, 0.9].

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-09 20:02:48 +02:00

..

dqn-hyperopt.toml

feat: add c51_alpha_max to cap MSE→C51 blend and prevent gradient starvation

2026-04-09 15:42:41 +02:00

dqn-localdev.toml

fix: all DQN configs — IBKR costs (0.18 bps) + min_hold_bars=10 for volume bars

2026-04-08 21:05:59 +02:00

dqn-production.toml

fix: set c51_alpha_max=0.5 — prevent C51 gradient convergence on H100

2026-04-09 20:02:48 +02:00

dqn-smoketest.toml

fix(critical): d_adv_logits buffer under-allocation — 32 elements short

2026-04-09 17:06:30 +02:00

ppo-production.toml

feat: TOML training profile system — config-driven hyperparameters

2026-03-21 23:35:41 +01:00

ppo-smoketest.toml

feat: TOML training profile system — config-driven hyperparameters

2026-03-21 23:35:41 +01:00

supervised-production.toml

feat: TOML training profile system — config-driven hyperparameters

2026-03-21 23:35:41 +01:00

supervised-smoketest.toml

feat: TOML training profile system — config-driven hyperparameters

2026-03-21 23:35:41 +01:00

walk-forward.toml

feat: TOML training profile system — config-driven hyperparameters

2026-03-21 23:35:41 +01:00