re-starts at test time

[davidwhogg]

Right now we do a single optimization at test time. This can't be right, since the optimization is non-convex. Also, Anna Ho finds that for LAMOST there are wrong answers caused by local minima. We should define a set of K initializations and start from all K, and choose the best answer (best in a likelihood sense). This, unfortunately, is high priority...

[davidwhogg]

ps. Rix requested it by phone today, and rightly so.

[andycasey]

K need to be decided before analysing all 150,000 spectra again with the preferred regularized model. So I propose this, which is probably overkill:

1. The median values of all training set labels.
2. The median Teff and logg, with the 25th percentile of all abundances.
3. The median Teff and logg, with the 75th percentile of all abundances.
4. The 25th percentile in Teff and logg, with the median of all training set abundances.
5. The 25th percentile in Teff and logg, with the 25th percentile of all abundances.
6. The 25th percentile in Teff and logg, with the 75th percentile of all abundances.
7. The 75th percentile in Teff and logg, with the median of all training set abundances.
8. The 75th percentile in Teff and logg, with the 25th percentile of all abundances.
9. The 75th percentile in Teff and logg, with the 75th percentile of all abundances.

[davidwhogg]

Happy!!