    Minimum Description Length Principle and Model Selection

                         Bin Yu
            Statistics Department, UC Berkeley


By viewing statistical modeling as a means of generating 
descriptions of observed data, the Minimum Description Length 
(MDL) framework discriminates between competing models based on the 
complexity of each description. This approach began with Kolmogorov's
theories of algorithmic complexity, matured in the literature on
information theory, and has recently received renewed interest within
the statistics community. 

In this talk, we review the basic ideas in MDL for problems of model 
selection.  Because model selection in linear regression is an extremely
common problem that arises in many applications, we present 
in this context derivations of two mixture form MDL criteria and
discuss their properties through a genetics 
example and a simulation study. In particular,
we show that these mixture forms of MDL
are adaptive, that is, they behave as AIC or BIC depending
on which one is more desirable.
