10/22

[diazrenata]

• Leave-one-out is extremely slow
  • HPG troubleshooting: qs for Ethan, Hao
    • any insight re: why only a few processes run simultaneously?
  • Low nit to get more seeds: Juniper?
    • Does running a bunch of (simple) models with ~100 iterations, with many different seeds, seems like a reliable way to explore the LDA + TS space simultaneously?
  • Fewer chunks
    • Alternatively, we could pull out fewer years, or pull out larger chunks. Is there a preference from the TS culture?
• Fun results to look at
  • BBS data: What if there really are no changepoints detectable, and the best strategy is to fit the most-efficient-possible mean for everyone?
  • Directional changepoint: 🙃
• Visualization note
  • Right now to visualize any instance of a model configuration, you have to choose a year (to be left out) and a sim. Does using the mean abundance predicted across all the models + all the sims seem reasonable to capture the variation?

[diazrenata]

[diazrenata]