On mixing models

The Context-Tree-Weighting algorithm is an example of an computationally efficient way to compute a Bayesian mixture of context-tree models. In this talk I will present several classes of models for which efficient mixing procedures exist. After briefly touching the binary CTW method I will discuss an efficient way to determine the MAP context tree model given an observed data sequence. Then I discuss the way to extend this method to non-binary alphabets. The decision tree problem requires a different model class and I will briefly review a possible method to find these trees using mixtures. Finally I will discuss a problem of Bayesian classification using partially dependent features. This problem cannot be formulated efficiently in a context tree setting and I will discuss another efficient mixture approach