Training and Validation

[cortner]

@tinatorabi is planning to implement a wrapper around the ASP solver that picks the best model from the model path based on a validation set. More generally I can envision us wanting to provide some limited hyperparameter selection based on a dataset split. This issue is to discuss a possible interface for this. In order to keep the package agnostic about the kind of data, I think there are two options:

solve(A, y, solver, Itrain, Ival)    # with Itrain, Ival vectors of row-indices
solve(Atrain, ytrain, Aval, yval, solver)

Maybe we can simply provide both and make the first option just a wrapper of the second option?

@wcwitt - do you have a view on this?

[cortner]

TODO: add validated TSVD interface

[cortner]

I think we have now settled on

solve( solver,  At, yt) 
solve( solver, At, yt, Av, yv )

and for not enforce the split to be done outside of ACEfit.

[cortner]

was closed by #84 and #83