Wiki
Clone wikicosmosis / multinest
Multinest
Overview
Authors: Farhan Feroz, Mike Hobson, Michael Bridges
Website: http://ccpforge.cse.rl.ac.uk/gf/project/multinest/
Paper: http://arxiv.org/abs/0809.3437
Multinest is an implementation of an advanced form of nested sampling, in which an ensemble of live points is gradually shrunk towards the peak by proposing into a set of ellipses fit to the ensemble and accepting only new points with greater likelihood.
Multinest can be run in parallel.
Compiling and running
If a fortran MPI compiler is found during the build then it will be used to compile an MPI version of Multinest. Otherwise a serial version will be built. If you look closely in the output of "make" it tells you which.
You can run multinest in serial like this:
#!bash
cosmosis params.ini
or with MPI like this, for e.g. 4 processes. Note that this is different to the other parallel samplers!:
#!bash mpirun -n 4 cosmosis params.ini
Options
Multinest is highly configurable and has a large number of options. The basic ones are described in this example ini file section; some others are described in the multinest readme.
#!ini [runtime] sampler=multinest [multinest] ; These are the only required parameters ; max_iterations is the total maximum number of iterations before finishing ; this value is enough for typical problems max_iterations = 50000 ; live points is the number of points in the ensemble. A few hundred is typical; more points means slower convergence but higher accuracy live_points = 400 ; The remainder of these parameters have default values, which are shown here ;Whether to print output during sampling feedback = True ; Get outputs in multinest's own format as well as from cosmosis using this filename base. ; Also enables restarting if the run is interrupted multinest_outfile_root = ;Resume from a pre-exising run. Requires a set (and unchanged from previous run) multinest_outfile_root resume = F ;Whether to allow multiple ellipses to cover the samples ;Useful for banana-shaped spaces and other oddities ;and max number of such modes mode_separation = False max_modes = 100 ; Random seed deciding sampling. If -1 set from time. random_seed = -1 ; Target error on Bayesian evidence tolerance = 0.1 ;Iterations before output update_interval = 200 ; See the link above for descriptions of these importance = True efficiency = 1.0 log_zero = 1e-6 const_efficiency = False cluster_dimensions = -1 mode_ztolerance = 0.5
Updated