Minimum Message Length

Minimum message length, or MML, is a Bayesian, information-theoretic method for comparing and selecting between statistical models. It restates Occam's razor formally: among models that fit the observed data equally well, the one whose combined description, the model itself plus the data given that model, can be encoded in the fewest bits is judged more likely to be correct. Minimum message length was invented by Chris Wallace, first appearing in his paper An Information Measure for Classification, and unlike the related idea of Kolmogorov complexity, it does not require a Turing complete language to describe a model, which is meant to make it usable as a practical technique rather than only a theoretical one.