| dc.description.abstract | In the preceding two decades, Hidden Markov models have become the method of choice for obtaining novel information from intensive longitudinal data sequences. One of the fundamental problems in hidden Markov modelling pertains to retrieving the structure of a hidden process phenomenon. Retrieving such structures enables researchers to formulate models which best describe unobserved real-world process phenomena. These types of learning problems are typically addressed with the Gibbs sampler. Methodological guidelines on fitting and optimal input specifications for the Gibbs sampler are however sparse. This study seeks to identify the general and optimal relations between the Gibbs sampler and two of its input variables: the number of event types inherent to, and the length of, the event observation sequence. In doing so it seeks to establish specification references for the appropriate and optimal use of the Gibbs sampler in single sequence HMM learning. Results indicate four event types and a sequence input length of 8000 to result in superior Gibbs sampler estimates. This study’s conclusions consequently coincide with, and add to the extant literature.
Future research avenues in regards to extending current work, and incorporating additional observation variables are discussed. | |
