You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Within the perfect model calibration tests in the ice_nucleation_calibration.jl file under test, each test is being run with different, arbitrary tolerances. Calibration might not be working as we want it because some values wouldn't pass the test if we set a reasonable relative tolerance.
For example:
ABDINM calibrated c coefficient is tested with an arbitrary absolute tolerance of 3. This is because the true value is around ~2, but because of the noise and standard deviation given to the priors, the calibrated value could be as high as 4. This is double the true value and requires around rtol = 0.9 which seems wrong. However, if we run the perfect_calib.jl file, we see that the resulting parcel output for ICNC matches fairly well with parcel outputs using the true coefficients.
Need to check with Ollie to see if this is okay. If not, find a way for these unit tests to pass with a relative tolerance of ~0.3.
The text was updated successfully, but these errors were encountered:
All ice calibration tests are now running with 0.3 relative tolerance on the final number concentration. This results in similar tolerance for all free parameters but for the second parameter in ABDINM experiment. Is that accuracy acceptable, or should we expect better from the algorithm?
Within the perfect model calibration tests in the
ice_nucleation_calibration.jl
file undertest
, each test is being run with different, arbitrary tolerances. Calibration might not be working as we want it because some values wouldn't pass the test if we set a reasonable relative tolerance.For example:
ABDINM calibrated c coefficient is tested with an arbitrary absolute tolerance of 3. This is because the true value is around ~2, but because of the noise and standard deviation given to the priors, the calibrated value could be as high as 4. This is double the true value and requires around rtol = 0.9 which seems wrong. However, if we run the
perfect_calib.jl
file, we see that the resulting parcel output for ICNC matches fairly well with parcel outputs using the true coefficients.Need to check with Ollie to see if this is okay. If not, find a way for these unit tests to pass with a relative tolerance of ~0.3.
The text was updated successfully, but these errors were encountered: