Implement Skill Generator Training Loop

[kirubel-Nigussie]

Description

Implements the training pipeline for the 2-head MLP SkillGenerator network. The trained model predicts skill outcomes (payoff, motives) from environment observations in milliseconds, enabling fast skill evaluation without running full environment episodes — critical for efficient CDS/PDS certification at scale.

---

Files Changed

File	Change
generator/losses.py	NEW — Combined MSE loss utility for the 2-head generator
generator/train_generator.py	NEW — Full training script (DataLoader, epoch loop, model save, loss plot)
tests/test_generator_training.py	NEW — 5 automated verification tests
requirements.txt	Added matplotlib>=3.7.0 for loss plot generation
.gitignore	Excluded data/raw/, models/, plots/ (generated artifacts)

---

What Was Implemented

generator/losses.py

• GeneratorLoss class with configurable payoff_weight and motive_weight
• Combined loss: total = MSE(pred_payoff, actual_payoff) + MSE(pred_motives, actual_motives)
• breakdown() method returning individual loss components for logging
• Designed for future extension (regularization, dynamic weighting)

generator/train_generator.py

• SkillDataset — PyTorch Dataset that loads all .npz files from data/raw/
• train_one_epoch() — single epoch training function
• train() — full 50-epoch loop with Adam optimizer (lr=1e-3, batch_size=32)
• Epoch-by-epoch console output (total, payoff, and motive losses)
• Saves trained weights → models/generator.pt
• Saves training curve → plots/generator_training.png

tests/test_generator_training.py

• test_training_runs_without_error — full pipeline completes without raising
• test_loss_decreases_over_epochs — final loss < initial loss
• test_trained_model_beats_random — post-training MSE < untrained MSE on held-out data
• test_model_saves_and_loads — save/load round-trip produces identical predictions
• test_loss_plot_is_generated — plot file exists and is non-empty

---