This is a production-ready regression parameter search and prediction using a
bootstrapped ensemble of unpruned majority vote regression trees. Information
gain on branching is calculated as RMSE of mean prediction.

Please read "production-ready" as "yes, we had this running in production".
We no longer do, as the 3GB model turned out to be not worth the RAM it was
using. It could capture some of the problem space complexity, but generally
speaking it was off the mark way too often to be useful.

The implementation is not exactly blazingly fast, and memory consumption was
never really optimized.

Note that the offline parameter search is tailored to accept a very specific
dataset, with each observation consisting of a floating point dependent
variable, followed by four continuous predictors, followed by binary
predictors until the end of the row. You'll need to change the hard-coded
data reader in rg-train.cxx if you want to experiment with some other data
layout.

You'll need a working gearmand and client to experiment with prediction
service, as that's the only platform supported.