search

cboettig / nimbios-transients

Example data that may or may not contain transients
MIT License
2 stars 0 forks source link

Conditions for overconfident learning? #4

Closed henryrscharf closed 5 years ago

henryrscharf commented 5 years ago

@henryrscharf Conditions for overconfident learning? (posterior is much narrower than it should be)

henryrscharf commented 5 years ago

This no longer seems like a concern based on the corrected log likelihood surfaces (see ghost_nimble.pdf). The previous log likelihood calculations were incorrect because they ignored the truncation of the innovation process (noise) that keeps the population non-negative. It remains an interesting idea, but I'm closing the issue for now since we don't have any empirical evidence of this actually happening.