irls: Subsampling IRLS

In jonlachmann/irls.sgd: Stochastic IRLS with SGD

Subsampling IRLS

Description

Subsampling IRLS

Usage

irls(
  X,
  y,
  family,
  ctrl = list(subs = 1, maxit = 100, tol = 1e-07, cooling = c(3, 0.9, 0.95), expl =
    c(3, 1.5, 1))
)

Arguments

X	A matrix containing the covariates (including an intercept if one wants to use one)
y	The dependent variable
family	A glm family for the distribution to use, i.e. "binomial()"
ctrl	A list of control parameters

Details

The control parameters that can be set are

• "subs" The subsample proportion to use for each iteration

• "maxit" The maximum number of iterations to run for

• "tol" The relative tolerance to consider when measuring convergence (not really useful for subsampling)

• "cooling" The cooling schedule to use, 3 numericals. First is number of iterations to stay constant for, the second is the temperature to start the first cooling iteration at and third is the exponential decay.

• "expl" The explosion detection settings, 3 numericals. First is number of iterations to not detect at, the second is the relative change in deviance to consider an explosion and the third is by which factor to change the subsample size at an explosion.

jonlachmann/irls.sgd documentation built on March 11, 2023, 7:42 a.m.