How matthijsdej/imitation_learning_topology_control works
Main Components
• Configuration Layer
– config.yaml and auxiliary/config.py (loads & validates config for all stages)
• Simulation Stage
– Top‐level script: simulate.py
– Core modules: simulation/simulation.py, simulation/strategy.py (Agent, AgentStrategy subclasses)
– parse_logs.py (log summarization)
– External dependency: Grid2Op environment (rte_case14_realistic)
– Outputs: raw state‐action datasets, simulation logs
• Data Preprocessing Stage
– Top‐level script: preprocess_data.py
– Core modules:
· data_preprocessing_analysis/data_preprocessing.py (batch transformation)
· Jupyter notebook data_preprocessing_analysis/data_analysis.ipynb (EDA)
– Outputs: processed datasets (train/val/test), feature‐stats file, adjacency‐hash table
• Imitation Learning Stage
– Top‐level script: train_network.py
– Core modules under training/:
· dataloader.py (batch loading & ProcessDataPointStrategy)
· models.py (Model ABC → FCNN)
· training.py (Run object, training loop, evaluate_dataset)
· evaluation.py (IncrementalAverageMetrics)
· postprocessing.py (ActSpaceCache)
– Evaluation notebook: evaluation.ipynb
– External service: Weights & Biases for logging
– Outputs: trained model artifacts, W&B logs
Relationships & Data Flows
• That auxiliary modules (generate_action_space.py, grid2op_util.py, util.py) are shared utilities imported by all stages.
• External integrations:
– Grid2Op env (feed into Simulation)
– W&B (sink for Training logs)
Architectural Patterns & Principles
• Layered Pipeline Architecture: clearly separate Simulation, Preprocessing, Training layers.
• Plug‐in Strategy & Model Patterns:
– AgentStrategy ABC with concrete strategy subclasses
– Model ABC with FCNN subclass
– ProcessDataPointStrategy for data‐specific preprocessing in Dataloader
• Modular, file‐based communication (data artifacts on disk as interfaces between stages)
Relevant Technologies & Frameworks
• Python modules and scripts
• Grid2Op environment (deprecated rte_case14_realistic)
• PyTorch (torch.nn.Module)
• Jupyter Notebooks for analysis
• YAML for configuration
• NumPy for scenario splits, data buffers
• Weights & Biases for experiment tracking
Essential Elements to Capture
• Clear, descriptive labels on each component and data store
• Callouts for key patterns (e.g., “AgentStrategy ABC,” “Model ABC”)