Model comparison for Gibbs random fields using noisy reversible jump Markov chain Monte Carlo

摘要: Abstract The reversible jump Markov chain Monte Carlo (RJMCMC) method offers an across-model simulation approach for Bayesian estimation and model comparison, by exploring the sampling space that consists of several models possibly varying dimensions. A naive implementation RJMCMC to like Gibbs random fields suffers from computational difficulties: posterior distribution each is termed doubly-intractable since computation likelihood function rarely available. Consequently, it simply impossible simulate a transition in presence intractability. variant presented, called noisy RJMCMC, where underlying kernel replaced with approximation based on unbiased estimators. Based previous theoretical developments, convergence guarantees algorithm are provided. experiments show can be much more efficient than other exact methods, provided estimator controlled variance used, fact which agreement analysis.