s-news
[Top] [All Lists]

Hidden Markov Model and Underflow problem

To: "'s-news@lists.biostat.wustl.edu'" <s-news@lists.biostat.wustl.edu>
Subject: Hidden Markov Model and Underflow problem
From: Yingfu Xie <Yingfu.Xie@sekon.slu.se>
Date: Mon, 12 Jul 2004 18:01:50 +0200
Hello, all,

Does anybody know any function or package in S-plus or R for estimation of a
Hidden Markov Model(HMM)? 

I have tried to write one by myself. First I compute the likelihood function
recursively, then use nlminb or optim to maximize it. However, my algorithm
suffers a serious underflow problem, i.e., the value of the likelihood of
the model at some stage of the algorithm is too small to be distinguishable
from zero. Simply multiplying by a large constant or taking algorithm
doesn't work. The scaling method introduced in the monograph by MacDonald
and Zucchini (Hidden Markov and Other Models for Discrete-Valued Time
Series) isn't satisfactory either. Any idea how can I improve it?  
Thanks a lot in advance!

Regards,
Yingfu
###########################################

This message has been scanned by F-Secure 
Anti-Virus for Microsoft Exchange.

###########################################



<Prev in Thread] Current Thread [Next in Thread>