moeridge: mixture of expert fit with ridge penalization

Description Usage Arguments Details Value Note Author(s) References Examples

View source: R/moe.R

Description

This function fits a two-component mixture of expert model with ridge penalization on parameters of the linear predictor of proportions.

Usage

1
2
3
4
5
moeridge(y, x, lambda.ridge = 0.1 * log(nrow(x)), 
initial = list(alpha = rep(0.5, (ncol(x) + 1)), 
beta1 = rep(0, ncol(x)), beta2 = rep(0, ncol(x)), 
sigma = 0.88), conv.eps = 1e-08, 
maxiter = list(total = 2500, em = 15))

Arguments

y

numeric vector of response to fit.

x

matrix of predictors with samples in rows and covariates in columns.

lambda.ridge

ridge penalization constant.

initial

a list of initial values.

conv.eps

EM convergence threshold.

maxiter

a list of prespecified integer values for maximum number of iterations.

Details

we may fill this later

Value

alpha

the parameters of the linear predictor of proportions under logit transformation.

beta

matrix of parameters of the linear predictors for mean components.

Note

The choice of initial values is crucial. A common variance is assumed. Maximized log likelihood value may be returned, or the estimated common variance. This function only fits Gaussians with two components. A good choice of penalization constant is proved to be of order log of sample size. This function requires installation of GSL libraries on the machine and may not work under Windows platform.

Author(s)

Vahid Partovi Nia and Abbas Khalili

References

Khalili, A. (2010). New Estimation and Feature Selection Methods in Mixture-of-Experts Models. The Canadian Journal of Statistics, 38, 519-539.

Examples

1
2

mixvarselect documentation built on May 2, 2019, 6:10 p.m.