摘要:It was with great interest that I read Jain and Neal's paper. In the paper, they address a
tough problem, namely how to improve the mixing/convergence of Markov chain Monte
Carlo (MCMC) algorithms for an important class of models. The models are those
involving mixtures of Dirichlet processes, ranging from a fairly straightforward mixture
of Dirichlet processes model to the more complex models that are springing up in a wide
variety of applications. The algorithms are in the split-merge vein, allowing a di
erent
kind of step than incremental Gibbs samplers. The extension of the split-merge technology
with targeted proposals to conditionally conjugate models is a welcome addition
to the collection of transitions available for tting models that include the Dirichlet
process as a component.