Simplify Adaptive Random Forest

[smastelini]

While reviewing #1309 I realized that the proposed changes would make the existing code problematic due to the existing class hierarchy. For instance, MSTS is intended for classification tasks, but the ARFRegressor ended up being affected to some extent.

Thus, I decided to first simplify the existing code. Next, we need to think about how to better integrate MSTS.

Suggestions:

• Create a new base Class intended to select subsets of classifiers/regressors in forest-based ensembles to generate the responses. This should become a single new parameter to ARF rather than multiple ones (ARF after MSTS ARFClassifier #1309 would have over 30 parameters!). Such implementation must be generic and work separately from the main ARF logic (or any other target ensemble algorithm).
  • In such case, such selector would be passed very much in the same fashion as the attribute splitters.
  • If the selector is supplied, it produces a list with the classifiers/regressors to use when computing the responses.
  • The selector will be updated during learning_one (thus the need for a carefully planned API) and queried in predict_one.
• Alternatively, we can try and find a way to apply MSTS while only affecting ARFClassifier.