UC Berkeley

The super learner is a general loss-based learning method designed to find the optimal combination of a set of learners. The super learner framework is built on the theory of cross-validation and allows for a general class of algorithms to be considered for the ensemble. The oracle results for the cross-validation selector are extended to the super learner. Due to the established oracle results for the cross-validation selector, the super learner is proven to represent an asymptotically optimal framework for learning. We discuss the super learner algorithm and demonstrate the method on a series of data analysis problems. The super learner framework is extended to cover censored data by applying an appropriate observed data loss function. The final chapter presents an R package for implementing the super learner based on a general library of learners.