Universität Bielefeld - "Graduiertenkolleg Aufgabenorientierte Kommunikation"

Active Learning of the Generalized High-Low-Game

Martina Hasenjäger and Helge Ritter

Abstract

In this paper, we study the performance of active learning with the query algorithm Query by Committee (QBC), which selects a new query such that it approximately maximizes the expected information gain. As target functions, we introduce a generalization of the High-Low-Game, for which we derive a theoretically optimal query sequence. This allows us to compare the performance of a QBC-learner with an information-optimal active learner. Simulations show that an active learner that selects queries with QBC rapidly converges against a learner trained with theoretically optimal queries.

Postscript-File (~32 k)

Anke Weinberger, 1996-12-10, 1997-04-18