| name | run-experiment |
| description | Run a KANDy experiment on a dynamical system to discover its governing equations. Use when the user wants to train a KAN model, fit equations, or run the full KANDy pipeline on a system. |
| argument-hint | ["system-name"] |
Run KANDy Experiment
Run a KANDy experiment on the $ARGUMENTS dynamical system.
Available Systems
!uv run kandy --list 2>/dev/null
Current Results
!ls results/ 2>/dev/null || echo "No results yet"
Instructions
Use the kandy-researcher agent to:
- Run
uv run kandy $ARGUMENTS (or uv run python examples/<system>_example.py if the system name doesn't match exactly)
- Monitor the output for training loss convergence, discovered equations, and rollout metrics
- After the experiment completes, use the model-quality-assessor agent to evaluate the generated plots in
results/
- Document findings in
notes/ with full experiment metadata
- Report back with:
- Discovered equations in clean symbolic format
- Key metrics (training loss, one-step RMSE, rollout RMSE)
- Quality assessment of the results
- Any issues or recommendations
If no system name is provided, list available systems and ask the user which one to run.
Critical Reminders
- KAN width is always
[m, n] (single layer)
- Cross-terms must be in the lift, not learned by the KAN
- All results go to
results/<SystemName>/
- Use
kandy.plotting.use_pub_style() for publication-quality figures
