Hidemaro Suwa, Synge Todo
The Markov chain Monte Carlo update method to construct an irreversible kernel has been reviewed and extended to general state spaces. The several convergence conditions of the Markov chain were discussed. The alternative methods to the Gibbs sampler and the Metropolis-Hastings algorithm were proposed and assessed in some models. The distribution convergence and the sampling efficiency are significantly improved in the Potts model, the bivariate Gaussian model, and so on. This approach using the irreversible kernel can be applied to any Markov chain Monte Carlo sampling and it is expected to improve the efficiency in general.
View original:
http://arxiv.org/abs/1207.0258
No comments:
Post a Comment