How sparkdroidwastaken/dqn-geometry-steering works
This repository is a CartPole DQN interpretability project: it trains and runs an agent, steers its hidden activations, and studies the resulting representation geometry. The documented workflow proceeds from browser-based training through state collection and steering to geometric analysis and visualization. The tree includes analysis scripts and data, but not every artifact named in the README; the graph therefore uses only paths present in the tree. Some relationships are documented in the README rather than visible in the sampled source.
Main entry points: states.zip Trajectory data · GetSteeringVector.py Vector calculation · CartpoleSteering.html Steered inference · GeometryAnalysis.py Geometry findings
Components 14
Agent workflow
- DQN training
Analysis/CartpoleManifold.html - State collection
Analysis/CartpoleStateCollecter.html - Trajectory data
Data/states.zip
Activation steering
- Steered inference
Analysis/CartpoleSteering.html - Vector calculation
Analysis/GetSteeringVector.py - Identity vector
Data/v_linear_idn - ReLU vector
Data/v_linear_relu.json - Tanh vector
Data/v_linear_tanh.json - Sigmoid vector
Data/v_linear_sigmoid.json
Geometry analysis
- Trajectory geometry
Analysis/GeometryAnalysis.py - Steering comparison
Analysis/SteeringAnalysis.py
Results and visualization
- Manifold visualizer
Analysis/Visualisation.py - Geometry findings
Analysis/GeometryAnalysis.py
Other components
- Researcher
Connections 20
- Researcher → DQN training: trains agent
README.md - DQN training → State collection: exports policy
README.md - State collection → Trajectory data: collects states
README.md - Trajectory data → Vector calculation: provides states
Analysis/GetSteeringVector.py - Vector calculation → Identity vector: computes vector
- Vector calculation → ReLU vector: computes vector
Analysis/GetSteeringVector.py - Vector calculation → Tanh vector: computes vector
- Vector calculation → Sigmoid vector: computes vector
- Identity vector → Steered inference: steers activations
- ReLU vector → Steered inference: steers activations
README.md - Tanh vector → Steered inference: steers activations
- Sigmoid vector → Steered inference: steers activations
- Steered inference → Trajectory data: produces trajectories
README.md - Trajectory data → Trajectory geometry: supplies trajectories
Analysis/GeometryAnalysis.py - Trajectory data → Steering comparison: supplies episode data
Analysis/SteeringAnalysis.py - Trajectory geometry → Geometry findings: computes metrics
Analysis/GeometryAnalysis.py - Steering comparison → Geometry findings: compares metrics
README.md - Trajectory data → Manifold visualizer: visualizes trajectories
- Geometry findings → Manifold visualizer: shows metrics
Analysis/Visualisation.py - Researcher → Manifold visualizer: explores manifolds
README.md