New ideas for improving the tuning
Any ideas worth trying? Put them here.
- Reducing/estimating systematics and overfitting by "bagging" - averaging over multiple random choices of fit bins.
- Flag any bins with particularly bad fits to the N-d quadratic.
- Should the interpolation error be included in computing goodness of fit?
- How to deal with empty bins? (Do we even know?)
- How to deal with genuinely failed runs? E.g., cases where there is no phase space for the requested event type to run. Can we flag regions of the parameter space as dysfunctional?