How naroaipar/vrp-frop works
Project Type and Purpose
• This is an academic/industrial-style ML system implementing a Reinforcement Learning solution for the Stochastic Vehicle Routing Problem (SVRP).
• The goal is to train a neural‐policy with attention and memory embeddings to solve routing under demand and travel‐cost uncertainty.
File‐Structure Analysis
• Top-Level:
• main.py – Orchestrator for training and testing
• config.py – Global hyperparameters and CLI flags
• checkpoints/ – Serialized model states per epoch
• results/ – Generated CSVs and figures
• models/ – Policy network (embeddings, attention, output)
• env/ – Custom SVRP environment (state generation, weather simulation)
• training/ – RL algorithms (REINFORCE, baseline) and training utilities
• inference/ – Deployment strategies (greedy, random sampling, beam search)
• Build/Deployment Files: requirements.txt, venv/
• Data Resources: env/*.txt (demands, positions, weather), results/
Key README Insights
• Architecture components are already separated by folder: environment, model, training, inference.
• Core algorithm: policy gradient (REINFORCE) with learnable baseline and attention mechanism.
• Multiple inference modes for evaluation.
Final Advice
• Both control flow (orchestration) and data flow (state/action/reward, checkpoint I/O, result plots).
• Keep it high-level enough to avoid overwhelming detail, but detailed enough to illustrate every folder/module’s role and interaction.