EstimateKL: Function to calculate expected KL

In bomeara/KLchecker: Estimates Kullback-Leibler distance between a true and candidate model

Description

Based on Jhwueng et al. 2014, especially equation 7

Usage

EstimateKL(simulator.true.model, true.parameters,
  param.estimator.candidate.model, simulator.candidate.model,
  likelihood.data.with.true.model, likelihoods.data.with.candidate.model, K,
  nrep.outer = 100, nrep.inner = 100)

Arguments

simulator.true.model	Function to create a dataset given vector of true parameters
true.parameters	Vector of parameters to use for simulations
param.estimator.candidate.model	Function to estimate the parameters under a candidate model given an input dataset
simulator.candidate.model	Function to simulate a dataset given vector of estimated parameters
likelihood.data.with.true.model	Function to calculate the log likelihood (not neg log like) of the data under the true model
likelihoods.data.with.candidate.model	Function to calculate the log likelihood (not neg log like) of the data under the candidate model
K	Number of free parameters in candidate model
nrep.outer	Number of outer loops
nrep.inner	Number of inner loops

Value

KL in same space (so multiplied by -2) as AIC

Examples

simulator.true.model <- function(parameters) {
  return(rnorm(n=1000, mean=parameters[1], sd=parameters[2]))
}

true.parameters <- c(1000, 0.12)

param.estimator.candidate.model <- function(x) {
  return(mean(1/x)) #here our candidate is exp
}

simulator.candidate.model <- function(parameters) {
  return(rexp(n=1000, rate=parameters[1]))
}

likelihood.data.with.true.model <- function(data, parameters) {
  return(sum(dnorm(data, mean=parameters[1], sd=parameters[2]), log=TRUE))
}

likelihood.data.with.candidate.model <- function(data) {
  return(sum(dexp(data, rate=param.estimator.candidate.model(data)), log=TRUE))
}

K <- 1

NormVsExp <- EstimateKL(simulator.true.model, true.parameters, param.estimator.candidate.model, simulator.candidate.model, likelihood.data.with.true.model, likelihoods.data.with.candidate.model, K)


simulator.true.model <- function(parameters) {
  return(rnorm(n=1000, mean=parameters[1], sd=parameters[2]))
}

true.parameters <- c(1000, 0.12)

param.estimator.candidate.model <- function(x) {
  return(c(mean(x), sd(x))) #here our candidate is exp
}

simulator.candidate.model <- function(parameters) {
  return(rnorm(n=1000, mean=parameters[1], sd=parameters[2]))
}

likelihood.data.with.true.model <- function(data, parameters) {
  return(sum(dnorm(data, mean=parameters[1], sd=parameters[2]), log=TRUE))
}

likelihood.data.with.candidate.model <- function(data) {
  return(sum(dnorm(data, mean=mean(data), sd=sd(data)), log=TRUE))
}

K <- 2

NormVsNorm <- EstimateKL(simulator.true.model, true.parameters, param.estimator.candidate.model, simulator.candidate.model, likelihood.data.with.true.model, likelihoods.data.with.candidate.model, K)

bomeara/KLchecker documentation built on May 14, 2019, 11:15 a.m.