MCSampler is an implementation of the 'Stretch-Move' algorithm for Markvo-Chain Monte Carlo sampling developed by Goodman & Weare (2010), written in a modern C++ template format for arbitrary functional analysis.
This work is heavily inspired by the Python implementation thereof, emcee, by Foreman-Mackey et. al (2013), and also makes use of the autocorrelation function Acor, also written by Goodman (https://www.math.nyu.edu/~goodman/software/acor/).
Parallelisation is supported through the <thread> library Unlike the Foreman-Mackey implementation, detailed balance is maintained on the entire ensemble by partioning the ensemble's history and its current position, allowing for arbitrary parellisation without impacting on the statistical validity of the ensemble.
The basic functionality of the code is as follows:
#include "MCSampler.h"
double logLikelihood(const std::vector<double> & parameters)
{
return //some function of the input parameters
}
int main(int argc, char** argv)
{
int nWalkers = //the number of members of the ensemble - the higher the number the more parallelisation helps you
int dimensions = //the number of parameters required for logLikelihodd
MCMC::Sampler sampler(nWalkers, dimensions;)
sampler.Run(targetFunction, nSamples, initialGuess);
}This works on any function which takes an std::vector<double> as input, and returns a double (including Lambdas). If the function requires more inputs -- such as a dataset which is being evaluated against -- then it is probably best to construct a functor which contains these parameters. If the functor possesses an overload for operator ()(std::vector<double>), then it works identically to the above:
#include "MCSampler.h"
struct inputData {}; //some data structure for holding the internal data
class LikelihoodFunctor
{
inputData MyData
LiklihoodFunctor(inputData data)
{
MyData = data;
}
double operator ()(const std::vector<double> & parameters)
{
return //some function of the input parameters and MyData
}
}
int main(int argc, char** argv)
{
int nWalkers = //the number of members of the ensemble - the higher the number the more parallelisation helps you
int dimensions = //the number of parameters required for logLikelihodd
MCMC::Sampler sampler(nWalkers, dimensions;)
inputData data = ///some data stored in some structure
LikelihoodFunctor targetFunction(data);
sampler.Run(targetFunction, nSamples, initialGuess);
}
This produces an MCMC chain for each walker, computes the autocorrelation time, and then flattens the chains into a single, shuffled array of samples. Drawing samples from the distribution is then equivalent to taking subsequent numbers from this array.
- Plotter (inc. 2D)