On model order estimation for partially observed Markov chains (Q1105545)

Model order estimation for the partially observed Markov chain (POMC), also called the hidden Markov model, is addressed. Model order is defined as the number of independent model parameters. Simulation is used to investigate the small sample performance of a number of model order criteria, as applied to POMC models with low state dimension and binary observations. Consistency of these criteria in the context of POMC models with discrete state and observation space is also investigated. A criterion due to Akaike is shown not to be consistent. Simulation results indicate that Akaike's criterion can provide more accurate estimation in the small sample case.