ML LERR: Difference between revisions

Revision as of 14:10, 7 December 2022

ML_LERR = [logical]
Default: ML_LERR = .TRUE.

Description: Decides whether the Bayesian error estimates are calculated and written out or not.

The Bayesian error estimation is always required for training (ML_ISTART=0,1,3) but may be omitted for prediction runs, e.g., MD simulations with ML_ISTART=2. In this case, setting ML_LERR=.FALSE. can significantly speed up the run because the error estimation is comparable in computational cost to the actual prediction.

Warning: Combine ML_LERR=.FALSE. only with ML_ISTART=2! Any other choice, e.g., ML_ISTART=1 with NSW=0 and ML_IALGO_LINREG=3, is untested and has no computational advantage.

@@ Line 4: / Line 4: @@
 Description: Decides whether the Bayesian error estimates are calculated and written out or not.
 ----
-If the Bayesian error estimates are not required in the prediction runs ({{TAG|ML_ISTART}}=2), it is recommended to set {{TAG|ML_LERR}}=.FALSE., since the calculation time of the error estimates is of the same order of magnitude as the prediction step.
+The Bayesian error estimation is always required for training ({{TAG|ML_ISTART}}=0,1,3) but may be omitted for prediction runs, e.g., MD simulations with {{TAG|ML_ISTART}}=2. In this case, setting {{TAG|ML_LERR}}=.FALSE. can significantly speed up the run because the error estimation is comparable in computational cost to the actual prediction. {{NB|warning|Combine {{TAG|ML_LERR}}{{=}}.FALSE. only with {{TAG|ML_ISTART}}{{=}}2! Any other choice, e.g., {{TAG|ML_ISTART}}{{=}}1 with {{TAG|NSW}}{{=}}0 and {{TAG|ML_IALGO_LINREG}}{{=}}3, is untested and has no computational advantage.}}
 == Related tags and articles ==