Google Summer of Code 2015: Improving mlr's hyperparameter and tuning system for efficient model selection

Basic data for this project

Description

mlr is a powerful package for general-purpose machine learning in R. It already contains a very powerful parameter description system for machine learning methods which is based on the ParamHelpers package. The latter is used for the definition of learner parameters as well as parameter tuning. However, the current implementation also has options for strong improvements as the following list indicates: The defaults of the learner parameters can only be static constants. Allowing for data-dependent functions would be much more flexible. For instance, the default value of the parameter "gamma" within an SVM should rather depend on the number of features Learners should encode whether a parameter can likely affect the predictive performance, e.g. the runtime, or is just "technical", e.g. a "verbose" parameter. Parameters should know their default box-constraints for optimization. Currently the user has to set these manually for a tuning run. Parameters should encode whether it is preferable to optimize them with a transformation. For instance, for the SVM again, one typically optimizes "C" and "gamma" on a log2 scale from -15 to 15.