Model Selection using the Minimum Description Length Principle

Abstract

Abstract The minimum description length (MDL) principle articulated in the last decade by Rissanen and his co-workers yields new criteria for statistical model selection. MDL criteria permit data-based choices from among alternative statistical descriptions of data without necessarily assuming that the data were sampled randomly. This article explains the MDL principle informally, indicates the criteria it yields in the common cases of multinomial distributions and Gaussian regression, and illustrates MDL's use with numerical examples. We hope thereby to stimulate experimentation and debate about the pedagogical and practical implications of the MDL approach.