A desicion-theoretic generalization of on-line learning and an application to boosting BT - Computational learning theory

Abstract

... More generally, we may allow the learner to select a distribution D t over the space of decisions , in which case it suffers the expected loss of a decision randomly ... The main idea of this improvement is to replace the "hard" {0, 1 }-valued decision used by hf by a "soft" threshold. ...