Save a new reward kernel bandwidth inside a learner, and return new reward predictions generated using this bandwidth for the original data used to train the learner.