jeff-regier
commented
8 years ago For this kind of testing, consider calling the function that computes the log likelihood rather than the log prior.
1.1~1.2 doesn't seem very good if the synthetic asteroid has the same brightness roughly and has only 2% error. Does the real data look different from the synthetic data? Or could a person looking at the pixel values not tell them apart?
On Fri, Nov 13, 2015, 1:40 AM Aaron Meisner notifications@github.com wrote:
I looked into the accuracy of the maximum probability flux for the case of completely synthetic data with the asteroid centered on one of the image pixels. I ran 500 realizations, and found the median flux to be 1.02 times the true flux (histogram embedded below). That's much better agreement than I was finding for the real data (~1.1-1.2 depending on the band). It occurred to me that the flux prior would be able to drag the maximum probability flux around a bit, but it actually looks like the prior mean is lower than the sample asteroid's flux.
[image: histogram of fluxes]
https://camo.githubusercontent.com/81707ab5ae0b84694d882a5864138a88a70597ec/687474703a2f2f6661756e2e72632e6661732e686172766172642e6564752f616d6569736e65722f746573745f666c75785f61636375726163792e706e67
You can view, comment on, or merge this pull request online at:
https://github.com/jeff-regier/KillerAsteroids.jl/pull/34 Commit Summary
- compute log probability for a grid of trial fluxes
File Changes
- A test/test_flux_accuracy.jl https://github.com/jeff-regier/KillerAsteroids.jl/pull/34/files#diff-0 (21)
Patch Links:
- https://github.com/jeff-regier/KillerAsteroids.jl/pull/34.patch
- https://github.com/jeff-regier/KillerAsteroids.jl/pull/34.diff
— Reply to this email directly or view it on GitHub https://github.com/jeff-regier/KillerAsteroids.jl/pull/34.
I looked into the accuracy of the maximum probability flux for the case of completely synthetic data with the asteroid centered on one of the image pixels. I ran 500 realizations, and found the median flux to be 1.02 times the true flux (histogram embedded below). That's much better agreement than I was finding for the real data (~1.1-1.2 depending on the band). It occurred to me that the flux prior would be able to drag the maximum probability flux around a bit, but it actually looks like the prior mean is lower than the sample asteroid's flux.