[Biojava-l] unsupervised training of transition weights

[wendy wong]

Thanks for your advice! I am able to train a subset of transition
probabilities now!

I found something strange, first I changed my emission distributions
to untrainabledistributions and  the trainer didn't seem to be doing
anything, all cycles have the same score. I then changed it back to
SimpleDistribution (still keepting my getWeightImp in my custom
distribution). this time it works and it doesn't seem to be modifying
my emission probabilities. So it works for me - I am just curious if
it is a bug or if I was doing something wrong?

Thanks again!
wendy


On 3/31/06, Matthew Pocock <matthew.pocock at ncl.ac.uk> wrote:
> > The DP code does some caching of probabilities, I don't think there's
> > any way to turn this off without modifying the DP implementations.
> >
> >            Thomas.
>
> My reccolection is that if you did turn this off, the algorithm would run
> very, very much more slowly. Internally to the DP objects, the distribution
> probabilities (in fact, they aren't even probabilities by this stage) are
> stored in a data-structure optimized for the type of lookups performed during
> the dynamic programming recursions.
>
> Matthew
>