Allow hyper-parameter tuning for immutable models. #174
Description
Some context: JuliaML/TableTransforms.jl#67
I don't think this would be too bad, and useful preparation for making the MLJ model interface more flexible later.
The MLJTuning API doesn't really touch on this point. A tuning strategy needs to implement a models method to generate models to evaluate, but doesn't say how the models are generated. They needn't be mutations of a single object. However, the MLJ model interface currently states that models must be mutable, so some tuning strategies do use mutation to generate their models.
TODO:
- To see if the change would be breaking, update this table:
| tuning strategy | assumes model types are mutable | pkg providing strategy |
|---|---|---|
Grid |
yes | MLJTuning |
RandomSearch |
yes | MLJTuning |
LatinHypercube |
yes | MLJTuning.jl |
MLJTreeParzenTuning() |
? | TreeParzen.jl |
ParticleSwarm |
? | MLJParticleSwarmOptimization.jl |
AdaptiveParticleSwarm |
? | MLJParticleSwarmOptimization.jl |
Explicit() |
no | MLJTuning.jl |
cc @juliohm