batch.estimation: Two stage estimation, a pilot estimate of mixing alpha and a...

In thq80/Owen_2000_Safe-and-Effective-IS: Optimal Mixture Weights in Multiple Importance Sampling

Description

Two stage estimation, a pilot estimate of mixing alpha and a following importance sampling, with or without control variates

Usage

batch.estimation(seed, batch.size, mixture.param,
  eps = rep(0.1/mixture.param$J, mixture.param$J), fname = "f",
  rpname = "rp", rqname = "rq", dpname = "dp", dqname = "dq",
  cv = TRUE, opt.info = FALSE, opt.param = list(reltol = 0.001, relerr =
  0.001, rho0 = 1, maxin = 20, maxout = 30))

Arguments

seed	seed for sampling
batch.size	length two vector of batch sizes
mixture.param	mixture.param = list(p, J, ...), where p is the dimension of the sample, and J is the number of mixture components, including the defensive one. mixture.param should be compatible with user defined functions f(n, j, mixture.param), rp(n, mixture.param), rq(n, j, mixture.param), dp(xmat, mixture.param), dq(xmat, j, mixture.param)
eps	the lower bound for optimizing alpha. if eps is of length 1, it is expanded to rep(eps, J), default to be rep(0.1/J, J)
fname	name of user defined function fname(xmat, j, mixture.param). xmat is an n \times p matrix of n samples with p dimensions. fname returns a vector of function values for each row in xmat. fname is defined for j = 1, \cdots, J. j = 1, \cdots, J - 1 corresponds to different proposal mixture components, and j = J corresponds to the defensive mixture component.
rpname	name of user definded function rpname(n, mixture.param). It generates n random samples from target distribution pname. Parameters can be specified in mixture.param. rpname returns an n \times p matrix.
rqname	name of user defined function rqname(n, j, mixture.param). It generate n random samples from the j^{th} mixture component of proposal mixture distribution. rqname returns an n \times p matrix. rqname is defined for j = 1, \cdots, J - 1.
dpname	name of user defined function dpname(xmat, mixture.param). It returns the density of xmat from the target distribution pname as a vector of length nrow(xmat). Note that only the ratio between dpname and dqname matters. So dpname and dqname can return values of C \timesdpname and C \timesdqname respectively.
dqname	name of user defined function dqname(xmat, j, mixture.param). It returns the density of xmat from the proposal distribution q_j as a vector of length nrow(xmat). dqname is defined for j = 1, \cdots, J - 1. Note that only the ratio between dpname and dqname matters. So dpname and dqname can return values of C \timesdpname and C \timesdqname respectively.
cv	TRUE indicates optimizing alpha and beta at the same time, and estimate with the formula \hat{μ}_{α_{**},β} = \frac{1}{n_2}∑\limits_{i = 1}^{n_2}\frac{f(x_{i})p(x_{i}) - {β}^{\mathsf{T}}\bigl(\boldsymbol{q}(x_{i}) - p(x_{i})\boldsymbol{1}\bigr)}{q_{α_{**}}(x_{i})}. FALSE indicates optimizing alpha only, and estimate with the formula \hat{μ}_{α_{*}}= \frac{1}{n_2}∑\limits_{i = 1}^{n_2}\frac{f(x_{i})p(x_{i})}{q_{α_{*}}(x_{i})}.
opt.info	logical value indicating whether to save the returned value of the optimization procedure. See penoptpersp and penoptpersp.alpha.only for the returned value.
opt.param	a list of paramters for the damped Newton method with backtracking line search reltol relative tolerence in dampednewton step, default to be 10^{-2} relerr Newton step stop when within (1+relerr) of minimum variance, default to be 10^{-3} rho0 initial value for rho, default to be 1 Only need to supply part of the list to change the default value.

Details

Estimate E_p f with a two step importance sample procedure. See He & Owen(2014) for details.

Value

a list of

mu.hat

the estimate for mu

sd.hat

estimated sd of mu.hat

alpha.opt

the estimated optimal alpha

beta.opt

if cv = TRUE, the estimated optimal beta

opt.info

if opt.info = TRUE, also return the list(x=x, y=y, z=z, alpha=alpha,beta=beta,rho=rho,f=f

rhopen=rhopen, outer=log(rho0/rho,mu), relerr = relerr, alphasum = sum(alpha))

from the optimization after batch 1.

References

Boyd, S. and Vandeberghe, L. (2004). Convex optimization. Cambridge University Press, Cambridge

Examples

library(optismixture)
seed <- 1
p <- 5
rho <- 1/2
gamma <- 2.4
sigma.dvar <- function(rho, p){
   sigma <- matrix(0, p, p)
   for(i in 1:(p-1)){
       for(j in (i+1):p){
           sigma[i,j] <- rho^(abs(i-j))
       }
   }
   sigma <- sigma + t(sigma)
   diag(sigma) <- 1
   return(sigma)
}
sigma <- sigma.dvar(rho, p)
batch.size <- c(10^4, 1002)
j.vec <- 2^-(seq(1,5,1))
eps.status <- 1
eps.safe <- 0.1
## initialization and construct derived parameters
mu <- rep(0, p)
x0 <- matrix(1, 1, p)
x0.mat <- rbind(rep(1,p), rep(-1, p))
j.mat <- data.frame(centerid = rep(1:dim(x0.mat)[1], each = length(j.vec)),
                    variance = rep(j.vec, 2))
J <- dim(j.mat)[1] + 1
eps <- rep(0.1/J, J)
mixture.param <- list(x0 = x0, x0.mat = x0.mat, p = p,
sigma = sigma, gamma = gamma, j.mat = j.mat, J = J)
f <- function(x, j, mixture.param){
    f1 <- function(x, mixture.param){
        x0 <- mixture.param$x0
        gamma <- mixture.param$gamma
        return(sum((x - x0)^2)^(-gamma/2))
    }
    return(apply(x, 1, f1, mixture.param))
}
dq <- function(x, j, mixture.param){
    centerid <- mixture.param$j.mat[j, 1]
    j.param <- mixture.param$j.mat[j, 2]
    return(mvtnorm::dmvnorm(x, mixture.param$x0.mat[centerid,], j.param*diag(mixture.param$p)))
}
dp <- function(x, mixture.param){
    return(mvtnorm::dmvnorm(x, rep(0, mixture.param$p), mixture.param$sigma))
}
rq <- function(n, j, mixture.param){
    centerid <- mixture.param$j.mat[j, 1]
    j.param <- mixture.param$j.mat[j,2]
    return(mvtnorm::rmvnorm(n, mixture.param$x0.mat[centerid, ], j.param*diag(mixture.param$p)))
}
rp <- function(n, mixture.param){
    mu <- rep(0, mixture.param$p)
    sigma <- mixture.param$sigma
    return(mvtnorm::rmvnorm(n, mu, sigma))
}
a <- batch.estimation(seed, batch.size, mixture.param, eps, cv = FALSE,
fname = "f", rpname = "rp", rqname = "rq", dpname = "dp", dqname = "dq")

thq80/Owen_2000_Safe-and-Effective-IS documentation built on May 10, 2019, 9:53 a.m.