llama_0.8:
- models computed during cross-validation can be saved by passing save.models to
  the model builders
- various performance improvements, especially in the score computing functions
- introduce functions for result analysis: perfScatterPlot, predTable
- allow to create train/test splits with bootstrapping
- stratification for the train/test split generation functions is now turned off
  by default
- feature selection functionality has been retired
- some of the internal APIs have changed -- your code may break if you rely on
  these

llama_0.7.2:
- take success (if present) into account when determining best algorithm: if
  nothing was successful on an instance, set to NA -- this means that vbs may
  return NA as well
- fix bugs wrt cost calculations
- fix stupid bug that caused the incorrect best algorithm to be determined in
  some cases
- some addtional small bug fixes

llama_0.7.1:
- allow vbs/singleBest to operate on test splits to simplify the interface
- corrected the implementation of contributions() to handle minimisation and
  maximisation of performance values correctly


llama_0.7:
- add regressionPairs model, which predicts the performance difference for each
  pair of algorithms and makes decisions based on that
- use mlr for machine learning algorithms
- use original problem features along with predictions in stacked learners
