[Biojava-l] Parameter Settings in BaumWelchTraining]

[Thomas Down]

On 12 Mar 2004, at 06:00, mark.schreiber at group.novartis.com wrote:

> When you call the train() method of the BaumWelchTrainer you supply it
> with a SequenceDB. The sequences from this DB are used to optimize the
> weights of the model.
>
> However, I have a bad feeling that when you train your model with the
> BaumWelchTrainer your previously set counts will be ignored and
> overwritten. You could check by looking into 
> AbstractModelTrainer.train()
> (which is what the BaumWelchTrainer extends). You could also run some
> tests to see if using a pre-trained model makes any difference to the
> final outcome. Does anyone more expert than me on the DP package (ie 
> most
> people) know if the counts are overwritten?

The Baum-Welch algorithm is actually the Expectation Maximization 
algorithm applied to HMMs.  When you run BaumWelchTrainer, it takes the 
existing model then calculates a matrix which defines the probability 
that each symbol in your training database was emitted by a particular 
state in the HMM.  It then optimizes the transition and emission 
probabilities of the model to maximize the likelihood of the data 
*given that assignment of states to data*.

This means that, although all the model parameters get overwritten 
every cycle, they do still depend on the previous state of the model.  
If you're starting with a model that is a good, but not quite optimal, 
fit to your data, BaumWelchTrainer ought to do what you want.

Another thing you might look at is building a model with a mixture of 
trainable and untrainable distributions.  This allows you to 
specifically optimize the bits of the model which you aren't sure about 
yet, while holding the bits which you already trust constant.

      Thomas.