A comparison of reversible jump MCMC algorithms for DNA sequence segmentation using hidden Markov models

摘要

This paper describes a Bayesian approach to determining the number of hid- den states in a hidden Markov model (HMM) via reversible jump Markov chain Monte Carlo (MCMC) methods. Acceptance rates for these algorithms can be quite low, resulting in slow exploration of the posterior distribution. We consider a variety of reversible jump strategies which allow inferences to be made in discretely observed HMMs, with particular emphasis placed on the comparison of the competing strategies in terms of computational expense, algebraic complexity and performance. The methods are illustrated with an application to the segmentation of DNA sequences into compositionally ho- mogeneous regions.