A0171
Title: A merging algorithm for hidden Markov models with unknown number of states
Authors: Chu-Lan Kao - National Yang Ming Chiao Tung University (Taiwan) [presenting]
Cheng-Der Fuh - ()
Yang Chen - University of Michigan (United States)
Abstract: Most inference techniques for hidden Markov models (HMM) require a prespecified number of hidden states. Traditional approaches follow two steps, order selection followed by inference, which potentially suffer from model misspecification. Another related field for time trajectory modeling, the change-point detection, allows us to cut the data into segments. These segments are likely to represent different states. We propose a merging algorithm that uses the change-point detection technique to conduct inference on HMM without requiring a predetermined number of states. The proposed algorithm connects these two highly investigated fields in statistics. Both theoretical and computational performances of the proposed algorithm are given.
