jgrusewski 8777288880 feat: v_range computed from reward_scale + gamma — config flows HP → DQNConfig → GpuConfig ...

Task 4: Add `reward_scale` field (default 10.0) to DQNHyperparameters with
`computed_v_min()`/`computed_v_max()` methods. Formula:
v_range = (reward_scale / (1 - gamma) * 1.2).clamp(20, 300).
conservative() now computes v_min/v_max = +-240 (was hardcoded +-50).
Hyperopt adapter uses same formula instead of hardcoded max_abs_reward.

Task 5: Verified DQNConfig receives v_min/v_max from DQNHyperparameters
in constructor.rs (lines 285-286). Chain intact.

Task 6: Verified GpuDqnTrainConfig receives v_min/v_max from DQNConfig
in fused_training.rs (lines 169-170). Chain intact.

All Default impls updated: DQNConfig, GpuDqnTrainConfig,
ExperienceCollectorConfig, DqnBacktestConfig — zero hardcoded v_min/v_max.

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

2026-03-25 10:16:26 +01:00

97 KiB

Raw Blame History

View Raw