I typically use naive random or grid search to choose them, but I'd like to ask if there are built-in or paclet-based methods that can help perform more principled general hyperparameter searches?

For example, in Python's scikit one can wrap the fitting function within a search, eg. GridSearchCV(..., scoring='neg_mean_absolute_error').

I understand this is partially automated within Predict and Classify, but I'd like to use it as a general tool.

Many ml-methods have a range of hyperparameters and they may interact in nonlinear ways, any guidance on (ExternalEvaluated) async tooling here would be helpful as well.



Related Old Questions:

  • $\begingroup$ The BayesianMinimization documentation Applications section provides examples of using this framework to do ML hyper parameter searches. $\endgroup$ Feb 1 at 22:44
  • $\begingroup$ Was this updated recently? Those functions were not very usaful last time I checked. $\endgroup$
    – user5601
    Feb 6 at 0:34
  • $\begingroup$ Documentation lists it as unchanged since 11.2 (2017), so no? But it seemed to work OK last night. Obvious limitation is that it is serial (rather than parallel like Optuna or Keras Tuner) $\endgroup$ Feb 6 at 1:56
  • $\begingroup$ Yes that's a huge limitation. I see the new docs (not refpage but guide) say it was updated in the v14 summary of recent features. $\endgroup$
    – user5601
    Feb 7 at 18:28


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.