Version 1.1
-
The Gibbs sampling algorithm for the hidden states has been
improved. The earlier version computed the likelihood
for each sequence of hidden states by traversing
the whole sequence. However, because of the Markov structure
most terms cancel out. The conditional
probability of a given state conditional on all the other states
depends only on its two neighbouring states and the current
observation. This simplifies the Gibbs sampling procedure
considerably; see equation (2.5a) in
Robert et al..
-
To accelerate the equilibration process, it makes
sense to start from an initialisation that results from
some preprocessing - e.g., using
RecPars. Rather than initialising the state sequence at random,
it can now be read in from file
mosaic.in.
Here is a
detailed description of the changes
.
Last modified: 24 May 2001
Back to the beginning of this documentation.