[Biojava-dev] Training HMMs

[Matthew Pocock]

Hi,

I've just found and fixed a bug in training HMMs with
dot (delete) states. The forwards/backwards matricies
were being incorrectly calculated for these
non-emission states, causing the training to fail in
odd ways. In many cases, where the dot states are used
in a constant context, this would have caused the
models to be aproximately correct. In other cases
where a dot state is used in very different contexts
in different parts of the sequence, e.g. profile HMMs,
this would have been catastrophic.

Good luck.

Matthew

________________________________________________________________________
Want to chat instantly with your online friends?  Get the FREE Yahoo!
Messenger http://uk.messenger.yahoo.com/