Journal of Statistics: Advances in Theory and ApplicationsAuthor's: J. A. Boguslavskiy
Pages: [1] - [23]
Received Date: August 31, 2010
Submitted by:
Motivation. The Baum-Welch-algorithm usually recommend for an
estimation of parameters HMM. However, it not reliable as supplies the
vector of estimations, which as soon as corresponds to some nearness
to one of local maxima likelihood.
Results. This paper offers for estimation new MPA-algorithm,
which by means of polynomial approximations, the Bayes approach, and a
compression of an information builds approximations to a vector of the
conditional expectation. It is considered example, where HMM for
r=4 observable symbols and for s=5 hidden states
generates multiple sequences 15000 observable characters. The
MPA-algorithm calculates the estimates of the unknown 25 transition
probabilities and 20 emission probabilities by means of polynomial
approximations 3 orders. It has appeared, that from 45 relative errors
of an estimation almost all are less 0.1. Input MPA-algorithm were 16
numbers of the experimental frequencies received by compression of the
primary information.
hidden Markov model, stochastic matrix, probabilistic function, polynomial approximation, estimation, conditional expectations.
