jgrusewski 6ffa38cc04 feat(ml): NaN gradient detection with auto-halt during training ...

Add check_gradients_finite() utility that scans GradStore for NaN/Inf
values. Wire into DQN (after gradient accumulation, before optimizer step)
and PPO (both MLP and LSTM training paths, after backward pass).

Prevents silent model corruption from exploding gradients or numerical
instability — training halts immediately with a descriptive error.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-02-23 11:26:49 +01:00

..

flow_policy

fix(ml): fix PPO flow policy shape dimensions and clean up tests

2026-02-20 23:52:12 +01:00

action_masking.rs

feat(ml): factored 45-action masking, confidence EMA, RAII memory docs

2026-02-23 01:03:44 +01:00

action_space.rs

feat(ppo): gradient accumulation, clip-higher, and WorkingPPO→PPO rename

2026-02-21 01:00:54 +01:00

circuit_breaker.rs

feat(ml): complete all algorithm gaps — TFT, Mamba2, PPO, CPCV, FDR

2026-02-20 19:19:16 +01:00

continuous_action_masking.rs

feat(ml): factored 45-action masking, confidence EMA, RAII memory docs

2026-02-23 01:03:44 +01:00

continuous_demo.rs

feat(wave-d): Complete Wave D Phase 6 with 240+ parallel agents

2025-10-19 09:10:55 +02:00

continuous_policy.rs

Wave 11: Rainbow DQN integration + 23/23 tests passing

2025-11-18 13:53:59 +01:00

continuous_ppo.rs

fix(ml): fix PPO flow policy shape dimensions and clean up tests

2026-02-20 23:52:12 +01:00

continuous_transaction_costs.rs

refactor(ml): consolidate OrderType into common/action.rs

2026-02-22 23:35:42 +01:00

entropy_regularization.rs

Wave 11: Rainbow DQN integration + 23/23 tests passing

2025-11-18 13:53:59 +01:00

factored_action.rs

refactor(ml): consolidate OrderType into common/action.rs

2026-02-22 23:35:42 +01:00

gae.rs

refactor(ml): remove 797 lines of commented-out code and disabled imports

2026-02-20 19:43:47 +01:00

hidden_state_manager.rs

Wave 11: Rainbow DQN integration + 23/23 tests passing

2025-11-18 13:53:59 +01:00

lstm_networks.rs

feat(ppo): add gradient clipping and LSTM checkpoint loading

2026-02-21 00:45:32 +01:00

mod.rs

feat(ppo): gradient accumulation, clip-higher, and WorkingPPO→PPO rename

2026-02-21 01:00:54 +01:00

portfolio_tracker.rs

Wave 11: Rainbow DQN integration + 23/23 tests passing

2025-11-18 13:53:59 +01:00

position_limits.rs

Wave 11: Rainbow DQN integration + 23/23 tests passing

2025-11-18 13:53:59 +01:00

ppo.rs

feat(ml): NaN gradient detection with auto-halt during training

2026-02-23 11:26:49 +01:00

reward_normalizer.rs

feat(dqn): Implement adaptive C51 bounds for two-phase training

2025-11-22 19:21:51 +01:00

stress_testing.rs

Wave 11: Rainbow DQN integration + 23/23 tests passing

2025-11-18 13:53:59 +01:00

trainable_adapter.rs

feat(ppo): gradient accumulation, clip-higher, and WorkingPPO→PPO rename

2026-02-21 01:00:54 +01:00

trajectories.rs

refactor(ml): remove 797 lines of commented-out code and disabled imports

2026-02-20 19:43:47 +01:00

transaction_costs.rs

refactor(ml): consolidate OrderType into common/action.rs

2026-02-22 23:35:42 +01:00

unified_ppo.rs

chore: remove deprecated FeatureVector54 type alias

2026-02-20 13:46:56 +01:00