Observed arrival · 2026-11-01
Policy Manifolds: edit an RL agent one curve at a time
An educational research prototype for inspecting and editing KAN reinforcement-learning policies, with a CartPole interface for comparing live and original actions.
- For
- Reinforcement-learning researchers and curious KAN learners
- Worth noticing
- Edits stay in the browser session, preserve the checkpoint, and can be compared with an original-policy ghost from the same state.
Field notes
The interface reports a 4 → 2 → 2 KAN actor with 12 learned edge curves and 114 parameters, and explains that each edge combines a cubic B-spline with a weighted SiLU term. Users can adjust coefficients, undo or reset edits, and compare the altered actor with an original-policy ghost from the same state. The page says edits remain in the browser session and do not overwrite the checkpoint; it also says the trained model is not in the local app and asks users to load exported actor JSON.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue