Default configuration should capture errors

[sebffischer]

During a run with mlr3mbo with the default config we got:

INFO  [04:56:35.792] [bbotk]           429.392             <list[1]> ce827262-3214-4d9d-8be7-febba967c238
INFO  [04:56:35.915] [mlr3] Calling train method of fallback 'regr.ranger' on task 'surrogate_task' with 16 observations {learner: <LearnerRegrRanger/LearnerRegr/Learner/R6>}
Error in self$assert(xs, sanitize = TRUE) : 
  Assertion on 'xs' failed: epochs: May not be NA.
Calls: <Anonymous> ... resolve.list -> signalConditionsASAP -> signalConditions
In addition: There were 16 warnings (use warnings() to see them)
Execution halted

It would be nice if the default configuration for the surrogate model / acq-function optimizer would handle such failures.

[sumny]

Thanks for the report!
Is there any code snippet to reproduce it and develop against?
I assume these NA's in xs are domain values (and not codomain values because having NA in objective values would be impossible due to bbotk anyways)?
Is it a search space with dependencies? If yes, weird that the default surrogate does not handle this correctly but I will check for the fallback also.
If not, how can NA's in domain values happen?

Would be good to get some more input on how xs values happened to be NA here.

[sumny]

https://github.com/mlr-org/mlr3torch/pull/294/files#diff-6f7aedef5090f1739e1b63a28000b5e5757c31fb7f633ba0398618c9c51dc21f
benchmarks/rf_use_case/run_benchmark.R