Certain errors from the function or optimizer evaluation should be caught

[danielinteractive]

Motivation: For small data sets e.g. certain optimizers might fail with error messages. However, other optimizers might still succeed.

Problem: Currently fit_single_optimizer will return any errors from optimizers directly as errors, without filtering.

Solution idea:
When calling h_record_all_output() we could add certain errors in the remove list or similar mechanism, such that convergence is signalled as failed, however the user would not directly get the error but the search for other optimizers to converge could proceed.

Examples:

• L-BFGS-B needs finite values of 'fn'
• NA/NaN Hessian evaluation

[Dominic7227]

Hi Daniel, I also encountered the same issue. Basically the error message reported by one certain optimizer, such as L-BFGS-B, ends up the computing without trying other optimizers, even when I specified multiple optimizer using mmrm_control. Do you happen to know how to tackle with this? Many thanks!

[danielinteractive]

Hi @Dominic7227 , yeah we are working on the fix right now, see the corresponding PR which is in flight

[Dominic7227]

That is so much appreciated!

BTW, the current solution that works for me is to put L-BFGS-B (the error reporting optimizer) at the end of optimizer list when specified into mmrm_control.