Initial additions for Random Grid Search for Time Series

[ngupta23]

Describe the changes you've made

Added initial code for Random Grid Search for TS

Type of change

Please delete options that are not relevant.

• New feature (non-breaking change which adds functionality)

How Has This Been Tested?

Unit Tests have been added

Describe if there is any unusual behaviour of your code(Write NA if there isn't)

NA

Checklist:

• My code follows the style guidelines of this project.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have made corresponding changes to the documentation.
• My changes generate no new warnings.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• Any dependent changes have been merged and published in downstream modules.

[TremaMiguel]

the evaluate_candidates is callable?

[ngupta23]

Yes. its defined in the GridSearch class.

[TremaMiguel]

even though it is a "private" method, perhaps we could raise an error if evaluate_candidates is not callable

[ngupta23]

Since we are controlling that, it should be ok. In any case, it is mimicking the sklearn and sktime implementation so we should be ok.

[ngupta23]

https://github.com/scikit-learn/scikit-learn/blob/95119c13af77c76e150b753485c662b7c52a41a2/sklearn/model_selection/_search.py#L1625

[TremaMiguel]

Looks good, just added a pair of comments.

[Yard1]

LGTM, just one thing here - in general distributions should be left for external search libraries, for normal random/grid search just use tune_grid with lists

[Yard1]

It would be best if you used PyCaret's distributions (pycaret/internal/distributions.py) - feel free to expand upon them if necessary

[ngupta23]

Hi @Yard1 , Any particular reason why? I looked at that before implementing this, but it looks like that is needed only because you are interfacing with many external GS libraries and methods. In this case, it is all internal since it needs custom argument ordering (y, X vs X, y), etc so it wont work with something like BayesGridSearchCV from skoptimize unless we implement out own version custom for TS.

Do you see any issues leaving it as is?

[Yard1]

I would want to keep it consistent so that it will be easier when we implement wrappers for search algos in the future. Distributions would be the same, after all.

[ngupta23]

@Yard1 , I have implemented the changes to use PyCaret distributions. Please have a look