jgrusewski 86b950df10 fix(ml): DQN hyperopt overhaul — 7 root cause fixes (B1-B3, C1-C4) ...

B1: Eval mode — disable noisy layer noise, use softmax action selection
    (was greedy argmax, causing train/eval policy mismatch)
B3: Per-bar portfolio state sync in eval (was frozen within 1024-bar chunks)
C1: Extrinsic-only replay buffer — curiosity reward no longer stored
    (was corrupting Q-values to learn novelty instead of trading P&L)
C2: Single exploration mechanism — noisy nets only. Removed count bonus
    from Q-values and epsilon floor (was triple-stacking exploration)
C3: Neutral hold reward (0.0) — removed hold_penalty_weight from 31D→30D
    search space (was biasing Q-values toward excessive trading)
C4: Re-enabled early stopping with adaptive plateau_window = epochs/2

2720 tests pass, 0 failures, 0 clippy warnings.

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

2026-03-06 10:23:51 +01:00

..

baseline_common

fix(ml): wire spread_cost_bps into RL training — commission + spread = total tx cost

2026-03-01 22:10:43 +01:00

cuda_test.rs

chore: clean up examples, update ML binaries and risk tests

2026-02-27 01:33:18 +01:00

download_baseline.rs

chore: clean up examples, update ML binaries and risk tests

2026-02-27 01:33:18 +01:00

evaluate_baseline.rs

fix(ml): DQN hyperopt overhaul — 7 root cause fixes (B1-B3, C1-C4)

2026-03-06 10:23:51 +01:00

evaluate_supervised.rs

fix(metrics): push training metrics to pushgateway before pod exit

2026-03-06 00:32:09 +01:00

hyperopt_baseline_rl.rs

feat(ml): multi-window backtest objective + top-K ensemble training

2026-03-06 08:42:03 +01:00

hyperopt_baseline_supervised.rs

fix(metrics): push training metrics to pushgateway before pod exit

2026-03-06 00:32:09 +01:00

train_baseline_rl.rs

feat(ml): multi-window backtest objective + top-K ensemble training

2026-03-06 08:42:03 +01:00

train_baseline_supervised.rs

fix(metrics): push training metrics to pushgateway before pod exit

2026-03-06 00:32:09 +01:00