[Biojava-dev] Query about DP example - Dice.java

[Matthew Pocock]

Hi David,

You are right. The actual distribution object used is
returned by getWeights(). This is needed because a
particular HMM class may use a special distribution
implementation which can't be substituted for by any
old implementation. If you feel keen, update the docs.

If we rewrite the dp interfaces, I would be tempted to
replace the individual probabilitites over states with
a single conditional probability distribution over
pairs of states (source, destination), which will
simplify training of complicated models and make it
more transparent for working with models with
non-probability based parameters (e.g. information
scores from pam matricies). Illegal tranisitons could
be flagged with NaN, or zero or something. Live &
learn.

Matthew

 --- David Huen <smh1008@cus.cam.ac.uk> wrote: > Hi,
> I note that the example shows manipulation of the
> transition probabilities 
> by:-
>       dist =
> casino.getWeights(casino.magicalState());
>       dist.setWeight(fairS, 0.8);
>       dist.setWeight(loadedS, 0.2);
> 
> This appears to suggest that getWeights() must
> return the actual 
> Distribution object it uses: a copy of it is
> unacceptable if this is to 
> work.  If this is to be the expected behaviour with
> all future 
> implementations of this class, we should explicitly
> state it in the 
> javadocs.
> 
> Otherwise, perhaps the above snippet should be
> appended with
>       casino.setWeights(casino.magicalState(),
> dist);
> 
> Would this make sense?
> 
> Regards,
> David Huen
> 
> _______________________________________________
> biojava-dev mailing list
> biojava-dev@biojava.org
> http://biojava.org/mailman/listinfo/biojava-dev 

__________________________________________________
Do You Yahoo!?
Everything you'll ever need on one web page
from News and Sport to Email and Music Charts
http://uk.my.yahoo.com