[ENH] for `gfo` allow `numpy` coercibles and/or `sklearn` parameter grids as search spaces

[fkiraly]

The current gfo implementation allows to pass grid-like search spaces, similar to sklearn.

However, currently we cannot pass - but it would be desirable to:

• values that are not np.ndarray-s, but coercible, i.e., 1D array-likes. If passed (e.g., lists), they get coerced to np.ndarray internally
• union grids, such as in sklearn

More generally, it should not be too complicated to allow the full sklearn syntax to work - although coercion is probably much easier than supporting union grids.

[amitsubhashchejara]

Hi @fkiraly
I would like to take on this issue. Can you elaborate a bit on this?

[DCchoudhury15]

Hi @fkiraly , I'd like to work on this issue.

Proposed Approach

1. Input Coercion

   • Update the internal handling of gfo inputs so that any array-like structure (e.g., Python lists, tuples) is automatically converted into a numpy.ndarray.
   • This ensures consistent downstream processing and aligns with the examples provided in the issue where a list currently causes unexpected behavior.

2. Parameter Grid Handling

   • Add support for common scikit-learn grid formats ({"param": [values...]}) when compatible with the existing optimizer structure.
   • In cases where full scikit-learn union grids are not compatible, provide a clear and informative error message explaining the limitation.
   • Ensure that the behavior remains deterministic and compatible with existing grid search logic in Hyperactive.
     If this approach looks good, I would be happy to proceed with the implementation.