Kenji Yamanishi, University of Tokyo, Japan

Title:

MDL-based gradual model change detection

Abstract:

We are concerned with the issue of detecting model changes when the model may changes gradually over time. We introduce the MDL-change statistics and sequential model change detection algorithms based on it. We give theoretical foundations to the algorithms and show the effectiveness in the practical domains. We further introduce the notion of the descriptional dimension of a model class. We attempt to characterize the transitive process of gradual model changes through the discriptional dimension.