fit.fmmstil.parallel: This function finds the penalised quasi likelihood estiamtes...

In henrylobster/mstil: 'Multivariate Skew t Distribution with Independent Logistic Skewing Function (MSTIL)'

Description

This function finds the penalised quasi likelihood estiamtes for finite mixture of mstil via EM in parallel.

Usage

fit.fmmstil.parallel(
  x,
  K,
  ncore = 1,
  param = NULL,
  init.cluster,
  init.param.method,
  show.progress = TRUE,
  control = list()
)

Arguments

x	matrix of quantiles of size n x k. Each row is taken as a quantile.
K	number of clusters.
ncore	number of core to be used.
param	list of lists of inital parameters, containing omega, lambda, delta, Ainv, and nu.
init.cluster	(optional) initial clusters used to find initial parameters.
init.param.method	(optional) method to obtain initial parameters. It needs to be a function of x and K, and return a list of list of parameters.
show.progress	a logical value. If TRUE, progress of the algorithm will be printed in console. By default TRUE.
control	list of control variables, see 'details'.

Details

The control argument is a list that accepts the following components.

numLikSample

a positive integer, represents the number of samples used to estimate the density and log-likelihood functions. By default 1e6.

conLevel

a value between 0.5 and 1, represents the confidence level of the log-likelihood to be calculated. By default 0.95.

cvgN

a positive integer. The algorithm stops when the estimated log-likelihood is not improved in cvgN iterations. By default 5.

lambdaPenalty

a positive value, represents the L2 penalty coefficient for lambda. By default 1e-6.

ainvPenalty

a positive value, represents the L2 penalty coefficient for Ainv. By default 1e-6.

maxit

a positive integer, represents the maximum number of EM iterations allowed. By default 1e3.

maxitOptim

a positive integer, represents the maximum number of iterations in optim allowed within each M-step. By default 10.

numGradSample

a positive integer, represents the number of samples used to estimate the gradient. By default 1e4.

finDiffStep

a positive value, represents the step size to estimate gradient w.r.t. nu. By default 1e-5.

stepSgd

a positive value, represents the step size to be used in the stochastic gradient step to optimise nu. By default 1e-2.

iterSgd

a positive integer, represents the number of iterations to be used in the stochastic gradient step to optimise nu. By default 1e2.

dimRateSgd

a positive value, represents the diminishing rate for step size to be used in the stochastic gradient step to optimise nu. By default 1e-2.

batchSize

a positive integer, represents the batch sample size. By default n.

Value

a list with components:

logLik	a vector of the estimated log-likelihood after each itereation.
par	a list of lists of lists of fitted parameters after each iteration.
time	a vector recorded the time elapsed after each iteration.

Examples

# Not run:
# data(RiverFlow)
# fit.fmmstil(as.matrix(log(RiverFlow)), 2)

henrylobster/mstil documentation built on Sept. 25, 2020, 3:48 p.m.