8ce7c52586
- Fixed feature dimension mismatch in evaluate_dqn_main_orchestrator.rs - Updated all 5 occurrences: state_dim, input comments, feature vector type - Aligned with Wave 16D training (128 features: 125 market + 3 portfolio) Issue: Validation backtest reveals 100% HOLD action collapse - requires reward system investigation and redesign per latest RL research.
90 lines
2.7 KiB
Plaintext
90 lines
2.7 KiB
Plaintext
AGENT 36: WAVE 14-15 FIX INTEGRATION CHECKLIST
|
|
==============================================
|
|
|
|
MISSION: Wire ALL Wave 14-15 fixes into hyperopt pipeline
|
|
|
|
FILES TO MODIFY:
|
|
----------------
|
|
1. ml/examples/hyperopt_dqn_demo.rs
|
|
- Add 6 CLI flags:
|
|
* --enable-preprocessing (bool)
|
|
* --preprocess-window (usize, default 50)
|
|
* --preprocess-clip-sigma (f32, default 5.0)
|
|
* --enable-polyak (bool)
|
|
* --tau (f32, default 0.001)
|
|
* --feature-count (usize, default 225)
|
|
|
|
2. ml/src/hyperopt/adapters/dqn.rs
|
|
- Add fields to DQNTrainer struct
|
|
- Pass preprocessing config to InternalDQNTrainer
|
|
- Pass Polyak config to DQN model
|
|
- Wire feature_count through
|
|
|
|
3. ml/src/trainers/dqn.rs
|
|
- Call preprocessing::apply_stationarity() if enabled
|
|
- Pass tau to target_update::polyak_update() if enabled
|
|
- Use reward::apply_q_constraints() (Agent 27)
|
|
- Filter features if feature_count < 225
|
|
|
|
4. ml/src/dqn/dqn.rs (model)
|
|
- Accept tau parameter in constructor
|
|
- Use polyak_update() instead of hard copy if tau provided
|
|
|
|
VALIDATION TESTS:
|
|
-----------------
|
|
Run 3-trial campaign with ALL fixes enabled:
|
|
|
|
cargo run --release -p ml --example hyperopt_dqn_demo --features cuda -- \
|
|
--parquet-file test_data/ES_FUT_180d.parquet \
|
|
--trials 3 \
|
|
--epochs 5 \
|
|
--enable-preprocessing \
|
|
--preprocess-window 50 \
|
|
--preprocess-clip-sigma 5.0 \
|
|
--enable-polyak \
|
|
--tau 0.001 \
|
|
--feature-count 125 \
|
|
2>&1 | tee /tmp/agent36_validation.log
|
|
|
|
SUCCESS CRITERIA:
|
|
-----------------
|
|
✅ At least 1/3 trials complete (not pruned)
|
|
✅ Gradient norms < 200 (ideally < 100)
|
|
✅ Log shows "Applying preprocessing" messages
|
|
✅ Log shows "Using Polyak averaging (tau=0.001)"
|
|
✅ Log shows "125 dimensions" not 225
|
|
✅ At least 1 trial has Sharpe > 0.5
|
|
|
|
FAIL CONDITIONS:
|
|
----------------
|
|
❌ All 3 trials pruned (gradient explosion)
|
|
❌ Gradient norms still > 1000
|
|
❌ No preprocessing logs (not wired correctly)
|
|
❌ Still showing 225 dimensions
|
|
|
|
DEPENDENCIES:
|
|
-------------
|
|
Agent 27: ✅ Q-value constraints (ml/src/dqn/reward.rs)
|
|
Agent 28: ✅ Preprocessing (ml/src/preprocessing.rs)
|
|
Agent 29: ❌ Feature reduction (NOT IMPLEMENTED - need Agent 37)
|
|
Agent 30: ✅ Polyak averaging (ml/src/dqn/target_update.rs)
|
|
Agent 34: ✅ Backtesting (ml/src/evaluation/backtesting.rs)
|
|
|
|
NEXT AGENT:
|
|
-----------
|
|
Agent 37: Implement feature reduction (225→125)
|
|
- After Agent 36 completes wiring
|
|
- ETA: 4-6 hours
|
|
|
|
ESTIMATED TIME:
|
|
---------------
|
|
Integration: 8-12 hours
|
|
Validation: 30 minutes
|
|
Total: 8.5-12.5 hours
|
|
|
|
REFERENCE DOCS:
|
|
---------------
|
|
- AGENT_35_VALIDATION_CAMPAIGN.md (full report)
|
|
- WAVE15_CRITICAL_FAILURE_SUMMARY.txt (quick summary)
|
|
- Wave 14-15 agent reports (individual fix details)
|