Feedback Loop¶
The feedback agent guides improvements between iterations.
How It Works¶
After each iteration completes:
- Results Collected - Metrics, model outputs, and analysis from all steps
- Feedback Generated - A separate LLM analyzes what worked and what didn't
- Instructions Created - Specific guidance for each step in the next iteration
- Next Iteration Starts - Agents receive feedback as part of their context
Feedback Structure¶
The feedback agent produces instructions for each step:
IterationInstructions
├── data_exploration_instructions
├── data_split_instructions
├── data_representation_instructions
├── model_architecture_instructions
├── model_training_instructions
├── inference_instructions
├── prediction_exploration_instructions
└── other_instructions
What Feedback Considers¶
Performance Metrics¶
- Current vs. previous validation metrics
- Best iteration so far
- Metric trends across iterations
Issues Detected¶
- Overfitting (train >> validation performance)
- Underfitting (poor performance on both)
- Class imbalance effects
- Feature problems
Model Behavior¶
- Training convergence
- Prediction distribution
- Error patterns
Time Constraints¶
- Remaining time (if timeout set)
- Training duration of previous models
Feedback Examples¶
Overfitting Detected¶
Model Architecture Instructions: The previous model overfit significantly (train ACC 0.98, val ACC 0.72). Try stronger regularization: increase dropout to 0.5, add L2 regularization, or use a simpler architecture with fewer layers.
Underfitting Detected¶
Model Architecture Instructions: The model underfits (train ACC 0.65, val ACC 0.63). Consider a more complex model: deeper network, more trees, or additional features in the representation.
Good Progress¶
Model Training Instructions: Good improvement from 0.75 to 0.82 AUROC. Continue with similar architecture but try fine-tuning learning rate (try 1e-4) and increasing epochs.
Exploration vs. Optimization¶
The feedback agent manages two phases:
Exploration Phase¶
Early iterations (controlled by --exploration-iterations):
- Try diverse approaches
- Test different model types
- Explore data representations
- Build baseline understanding
Optimization Phase¶
Later iterations:
- Focus on best-performing approaches
- Fine-tune hyperparameters
- Improve on winning strategy
- Polish the solution
Combination Tracking¶
The feedback agent tracks which combinations have been tried:
Representation × Architecture Combinations:
✓ StandardScaler + LogisticRegression → 0.72 ACC
✓ StandardScaler + RandomForest → 0.78 ACC
✓ PCA + RandomForest → 0.75 ACC
✗ MinMaxScaler + GradientBoosting → Not tried yet
This prevents repeating failed approaches and encourages exploration.
Split Management¶
For early iterations, the agent may suggest split changes:
Data Split Instructions: Validation set may be too small for reliable metrics. Consider increasing validation size to 30% for better estimates.
After --split-allowed-iterations, splits are frozen to ensure fair comparison.
Feedback Agent Model¶
The feedback agent uses the same LLM as the main agent.
Viewing Feedback¶
Feedback is included in iteration reports:
Each report shows: - What feedback was received - How the agent responded - Results of the iteration
Configuration¶
| Parameter | Default | Description |
|---|---|---|
--exploration-iterations |
4 | Iterations for broad exploration |
--split-allowed-iterations |
1 | Iterations that can modify splits |
Next Steps¶
- Agent Architecture - Full iteration loop
- Evaluation - How metrics are calculated