How jetnew/smawm works
Main Components
a) Data Pipeline
• Dataset Generator
• Data Storage (datasets/expert, random, spurious)
b) Model Components
• World Model (WM)
• Structured Multi-Agent World Model (SMAWM)
• Reinforcement Learning Agents
c) Training Pipeline
• World Model Trainer
• SMAWM Trainer
• RL Agent Trainer
d) Experimentation & Analysis
• Experiment Runner
• Analysis Tools
• Visualization Components
Flow Directions
• Dataset Generator → Datasets → Training Components → Models → Experiments → Analysis
Detailed Subcomponents
For each major component, include relevant submodules:
World Model:
• VAE (vae.tar)
• MDRNN (mdrnn.tar)
• PPO Agent (wm-ppo.zip)
SMAWM:
• Structured Components (swm.tar)
• PPO Agent (swm-ppo.zip)
Specific elements
• Dataset types (expert, random, spurious)
• Model storage locations
• Analysis outputs (figures directory)
• Experiment results storage
Implementation Details
Files and their roles:
• generate_dataset.py
• train_wm.py
• train_swm.py
• train_rl.py
• run_experiments.py
• run_analysis.py
Storage Components
Storage locations:
• datasets/
• models/
• experiments/
• figures/
Metrics Flow
How performance metrics flow:
• Training metrics
• Experimental results
• Analysis outputs
• Visualization generation
• The parallel nature of WM and SMAWM implementations
• Data flow from generation to analysis
• Clear separation of training and evaluation phases
• Integration points between components
• Storage and retrieval patterns
Remember to include version information and any relevant hyperparameters that affect the system's behavior.
• How data flows through the system
• Where models are trained and stored
• How experiments are conducted
• How results are analyzed and visualized
• The relationship between traditional World Models and SMAWM