pmcmc_chains: pMCMC with manual chain scheduling

In mrc-ide/mcstate: Monte Carlo Methods for State Space Models

pMCMC with manual chain scheduling

Description

Run a pMCMC, with sensible random number behaviour, but schedule execution of the chains yourself. Use this if you want to distribute chains over (say) the nodes of an HPC system.

Usage

pmcmc_chains_prepare(path, pars, filter, control, initial = NULL)

pmcmc_chains_run(chain_id, path, n_threads = NULL)

pmcmc_chains_collect(path)

pmcmc_chains_cleanup(path)

Arguments

path	The path to use to exchange inputs and results. You can use a temporary directory or a different path (relative or absolute). Several rds files will be created. It is strongly recommended not to use .
pars	A pmcmc_parameters object containing information about parameters (ranges, priors, proposal kernel, translation functions for use with the particle filter).
filter	A particle_filter object
control	A pmcmc_control object which will control how the MCMC runs, including the number of steps etc.
initial	Optional initial starting point. If given, it must be compatible with the parameters given in pars, and must be valid against your prior. You can use this to override the initial conditions saved in your pars object. You can provide either a vector of initial conditions, or a matrix with n_chains columns to use a different starting point for each chain.
chain_id	The integer identifier of the chain to run
n_threads	Optional thread count, overriding the number set in the control. This will be useful where preparing the threads on a machine with one level of resource and running it on another.

Details

Basic usage will look like

path <- mcstate::pmcmc_chains_prepare(tempfile(), pars, filter, control)
for (i in seq_len(control$n_chains)) {
  mcstate::pmcmc_chains_run(i, path)
}
samples <- mcstate::pmcmc_chains_collect(path)
mcstate::pmcmc_chains_cleanup(path)

You can safely parallelise (or not) however you like at the point where the loop is (even across other machines) and get the same outputs regardless.

mrc-ide/mcstate documentation built on July 3, 2024, 1:34 p.m.