EMglmnet: Penalized Mixtures of Experts
In schiffner/EMnet: Train Penalized Mixtures of Experts by EM Algorithm

Fits a penalized mixture of experts model using the EM algorithm. In the EM steps penalized logistic models are fitted using function glmnet from glmnet is used.

EMglmnet(y, X, colsGating = 1:ncol(X), colsExperts = 1:ncol(X), J = 2,
  lambda, alpha = 1, standardize = FALSE,
  type.multinomial = c("ungrouped", "grouped"), calculate.path = TRUE,
  interceptGating = TRUE, interceptExperts = TRUE, offsetGating = NULL,
  offsetExperts = NULL, iter.max = 100, tol = 1e-06,
  init.weights = NULL, minprior = 0.05, sd.tol = 1e-04)

`y`	A `factor` specifying the class membership for each observation.
`X`	A `matrix` or `data.frame` or `Matrix` containing the explanatory variables (must not contain an intercept column).
`colsGating`	Names or indices of columns in `X` to be used for the gating model. Default is all columns.
`colsExperts`	Names or indices of columns in `X` to be used for the expert models. Default is all columns.
`J`	The number of experts / mixture components. Defaults to 2.
`lambda`	Penalty parameter. Can be a scalar or a vector of length `J+1` with different components for the `J` experts and the gating model. All components must be >= 0.
`alpha`	Mixing parameter for the elastic net penalty. Can be a scalar or a vector of length `J+1` with different components for the `J` experts and the gating model. All components must be in [0,1]. Defaults to 1.
`standardize`	Logical. Should the columns of `X` be standardized prior to fitting the model?
`type.multinomial`	If `"grouped"` and the number of mixture components `J` and/or the number of classes is larger than two, a group Lasso penalty will be used on the multinomial coefficients for a variable. Defaults to `"ungrouped"`.
`calculate.path`	Logical. Should paths for the gating and expert models be calculated after the EM algorithm terminated? These are based on the weights calculated in the last E step.
`interceptGating`	Logical. Does the gating model include an intercept? Defaults to `TRUE`.
`interceptExperts`	Logical. Does the expert model include an intercept? Defaults to `TRUE`.
`offsetGating`	Offset term for the gating model.
`offsetExperts`	Offset term for the expert model.
`iter.max`	Maximum number of EM iterations.
`tol`	Small positive value to detect convergence of the EM algorithm.
`init.weights`	An `nrow(X)` times `J` matrix containing initial group membership values for the EM algorithm. The row sums must be 1. Defaults to `NULL`, which means that initial weights will be obtained by random segmentation.
`minprior`	Minimal relative size of experts. Defaults to 0.05. Experts which are smaller are removed.
`sd.tol`	Tolerance for detecting constant gating models, i.e., gating models not depending on the inputs and thus making constant predictions.

An object of class EMglmnet. A list with entries:

`gating`	The gating model. An object of class `glmnet`.
`experts`	A `list` of `J` expert models. These are objects of class `glmnet`.
`weights`	The final group membership values.
`iters`	The number of iterations in the EM algorithm.
`J,K, colsGating, colsExperts, lambda, offsetGating, offsetExperts`	See arguments.
`lev1`	Class labels present in the data.
`lev`	Class labels.
`gatingPath`	Gating model fitted for a sequence of penalty parameters. An object of class `glmnet`.
`expertsPath`	A `list` of `J` expert models fitted for a sequence of penalty parameters. These are objects of class `glmnet`.