Considerations on the Weighted Minimum Description Length Criterion.
Nobuhito Matsushiro · The Journal of The Institute of Image Information and Television Engineers · 1996
A weighted minimum description length (WMDL) criterion which is employed in adaptive model selection for an observed symbol series has been derived. The criterion has been derived from a weighted log likelihood instead of the log likelihood which is the basis of the minimum description length (MDL) criterion. In this paper, characteristics of estimated parameters of models and characteristics of the predictive WMDL criterion against the WMDL criterion are proved. Further the accuracy of model selection by the predictive WMDL criterion depending on weighting coefficients is discussed using computer simulations.