Well-Calibrated Predictions from Online Compression Models

Author: Vladimir Vovk .

Source: Lecture Notes in Artificial Intelligence Vol. 2842, 2003, 268 - 282.

Abstract. It has been shown recently that Transductive Confidence Machine (TCM) is automatically well-calibrated when used in the on-line mode and provided that the data sequence is generated by an exchangeable distribution. In this paper we strengthen this result by relaxing the assumption of exchangeability of the data-generating distribution to the much weaker assumption that the data agrees with a given ``on-line compression model''.

©Copyright 2003 Springer