AdaptiveLogDirMRW: AdaptiveLogDirMRW

In sourceR: Fits a Non-Parametric Bayesian Source Attribution Model

Description

This class implements an Adaptive Multi-site logarithmic Metropolis-Hastings random walk algorithm, constrained so the parameter vector sums to 1.

Format

Object of R6Class with methods for updating a DirichletNode instance.

Details

An adaptive multivariate log-Gaussian proposal is used for $d-1$ elements of a $d$-dimensional parameter vector contained in node, with the $d$th element updated to ensure that the vector sums to 1. This makes the updater useful for Dirichlet distributed random variables, improving on AdaptiveDirMRW by ensuring proposals do not go negative.

For details of the adaptive scheme, see Roberts and Rosenthal (2012) Examples of Adaptive MCMC. Journal of Computational and Graphical Statistics. 18:349–367.

Please note that no checks are performed as to the suitability of this algorithm for a particular StochasticNode. It is up to the user to use the correct update algorithm for the appropriate nodes.

Value

Object of AdaptiveLogDirMRW

Fields

cov

the current covariance

burnin

the number of updates to burn in

tune

the current tuning matrix

naccept

the number of accepted proposals

ncalls

the number of times update has been called

node

the node to which the updater is attached

Methods

new(node, toupdate = function() 1:length(node$getData()), tune = rep(0.1, length(node$getData())), burning = 100)

constructor takes an instance of a StochasticNode node, function to choose the indices of the elements to update (by default all elements), initial tuning vector (diagonal of adaptive tuning matrix), and number of calls between adaptations.

update()

when called, updates node

acceptance()

return the acceptance rate

sourceR documentation built on Aug. 31, 2020, 5:06 p.m.