Comments on: Markov Chain (Model) Builder…

By: Sergey Kulanov

Sergey Kulanov — Thu, 21 Apr 2011 11:44:41 +0000

How to get P(observation|MakrovModel)
this is what we have
result=prod(reshape(A.^o,1,[]))
where
A – transition matrix;
o – is the transition matrix of the observation sequence (which can be produced by
o=full(sparse(observation(1:end-1),observation(2:end),1));)
Of course we have to add initial State o(1) probability:
result*InitProb(o(1))

But the best way is to use logarithm scale
A more elegant way to count Probability P(Observation|model)was found in PMTK by Kevin Murphy:
see
log=markovLogprob(model,data)

By: Sergey Kulanov

Sergey Kulanov — Tue, 25 May 2010 06:05:04 +0000

Yet another approach:
A=full(sparse(sequence(1:end-1),sequence(2:end),1));
T=normalize(A,2);

By: admin

admin — Tue, 25 May 2010 05:28:39 +0000

A new approach was found in markovFit() function from PMTK toolkit, please refer http://pmtk3.googlecode.com/svn/trunk/docs/synopsis/Markov_models.html

By: admin

admin — Mon, 19 Apr 2010 06:33:21 +0000

More elegant way was found from mk_stochastic function of BNT toolbox (Written by Kevin Murphy)

A=full(sparse(sequence(1:end-1),sequence(2:end),1));
Z = sum(A,2);
S = Z + (Z==0);
norm = repmat(S, 1, size(A,2));
A = A ./ norm;