Wiki
Clone wikicosmosis / samplers / emcee
The Emcee sampler
Ensemble walker sampling
Name: Emcee
Version: 2.1.0
Author(s): Dan Foreman-Mackey and contributors
Cite: PASP, 125, 925, 306-312
Parallel: parallel
The emcee sampler is a form of Monte-Carlo Markov Chain that uses an ensemble of 'walkers' that explore the parameter space. Each walker chooses another walker at random and proposes along the line connecting the two of them using the Metropolis acceptance rule. The proposal scale is given by the separation of the two walkers.
It is parallel, so multiple processes can be used to speed up the running. It is also affine invariant, so that no covariance matrix or other tuning is required for the proposal.
The burn-in behavior of emcee can sometimes be poor; it is much better to start the chain as near to the maximum posterior as possible. The maxlike or similar samplers can help you find this.
The total number of samples taken is walkers*samples.
Installation
emcee is included in the cosmosis bootstrap, but if you are installing manually you can get emcee using one of the following commands:
pip install emcee # to install centrally, may require sudo # or ... pip install emcee --user # to install just for you
Parameters
These parameters can be set in the sampler's section in the ini parameter file.
If no default is specified then the parameter is required. A listing of "(empty)" means a blank string is the default.
Parameter | Type | Meaning | Default |
---|---|---|---|
walkers | integer | number of walkers in the space | |
samples | integer | number of jumps to attempt per walker | |
nsteps | integer | number of sample steps taken in between writing output | |
random_start | bool | whether to start the walkers at random points in the prior instead of near the start. Usually a bad idea | N |
start_points | string | a file containing starting points for the walkers. If not specified walkers are initialized randomly from the prior distribution. | (empty) |
covmat | string | a file containing a covariance matrix for initializing the walkers. | (empty) |
Updated