ABSTRACT

A variant of the Metropolis algorithm is proposed that allows parallel processing. Rather than generating a single candidate point, as in the Metropolis algorithm, for each chain iteration a number of candidates are generated. Energy calculations for each of these candidates can be carried out in parallel. This algorithm would be advantageous in fitting model parameters to data in a Bayesian context, where the forward model calculations (the analog of the energy calculations), are time consuming.

Keywords:

Metropolis algorithm, Parallel processing, Markov Chain Monte Carlo (MCMC), Bayesian statistics, Data analysis.