Siddiqi et al 2009 Reduced-Rank Hidden Markov Models

Online version

[1]

Summary

This paper introduces Reduced-Rank Hidden Markov Models (RR-HMMs). A RR-HMM is similar to standard HMM, except the rank of the transition matrix is less than the number of hidden states. The dynamics evolve in a subspace of the hidden state probability space.

Method

Let $x_{t}$ be the observed output of the RR-HMM and let:

The learning algorithm uses a singular value decomposition (SVD) of the correlation matrix between past and future observations. The algorithm is borrowed from Hsu et al 2009, with no change for the reduced-rank case.

${\hat {b}}_{1}$ is the initial state distribution, ${\hat {b}}_{\infty }$ is the final state distribution, and ${\hat {B}}_{x}$ is the transition matrix when x is observed. $m$ is the rank of the reduced state space. Note that $X^{+}$ denotes the Moore-Penrose pseudo-inverse of the matrix $X$.

Inference can be performed using the model parameters:

Experimental Result

Related Papers

In progress by User:Jmflanig

Comment

This stuff is super cool but a little tricky to wrap one's head around. I wrote up these notes about the older Hsu paper, and Siddiqi too, a while ago and they may not be totally correct but here we go: http://brenocon.com/matrix_hmm.pdf --Brendan 23:21, 13 October 2011 (UTC)