[Biojava-l] Mixing order of models

Thomas Down td2@sanger.ac.uk
Thu, 11 Oct 2001 08:37:22 +0100

On Thu, Oct 11, 2001 at 01:01:20PM +1300, Mark Schreiber wrote:
> Hi -
> A while ago I asked a question about mixing the order of Markov Models.
> The reason I want to do this is that some regions contain enough sequence
> to make 2nd order models other regions can easily make 5th order models. I
> seem to remember someone suggesting the use of redundant hexamers to mimic
> trimers (or something), how would this work??

Well, it's certainly possible to mix orders of models, by using
distributions which are high-order, but where some (or all)
of the conditioning symbols are ignored.  I guess that's equivalent
to what you say about redundant hexamers.

Training such a distribution might be a little more interesting,
but it should be possible to write a DistributionTrainer which
picks an appropriate order of distribution, based on the amount
of supporting evidence.  In a nice, Bayesian world, this should
just be a question of using a sensible prior.  (Matthew: am I
missing something here).