gagam: Genetic Algorithm for Generalized Additive Models
In markcus1/gagam: Genetic Algorithm for Generalized Additive Models

Description Usage Arguments Value References Examples

View source: R/gagam.R

Implements the genetic algorithm for simultaneous variable selection and structure discovery in generalized additive models. For a given dependent variable and a set of explanatory variables, the genetic algorithm determines which regressors should be included linearly, which nonparametrically, and which should be excluded from the regression equation. The aim is to minimize the Bayesian Information Criterion value of the model.

gagam(
  y,
  x,
  pop_size = 500,
  Kvar = 15,
  Kint = 0,
  no_gen = 100,
  p_m = 0.05,
  p_int = 0.1,
  p_nonpar = 0.1,
  p_int_nonpar = 0.1,
  multicore = TRUE,
  cores = NULL,
  k = 10,
  bs = "cr",
  family = gaussian(),
  method = "REML",
  optimizer = c("outer", "newton"),
  reduc = NULL,
  always_par = NULL
)

`y`	Vector, matrix, data frame, or factor containing observations of the dependent variable.
`x`	Matrix or data frame containing all considered explanatory variables. If the columns have names those will be used for variable names in the final output.
`pop_size`	Size of the population (needs to be a multiple of 500). Default is 500.
`Kvar`	Maximum number of variables allowed in the final model. Default is 15.
`Kint`	Maximum number of interactions allowed in the final model. Default is 0.
`no_gen`	Number of generations until convergence. Default is 100.
`p_m`	Mutation rate for variables. Default is 0.05.
`p_int`	Mutation rate for interactions of variables. Default is 0.1.
`p_nonpar`	Mutation rate for the linear/nonparametric indicators for variables. Default is 0.1.
`p_int_nonpar`	Mutation rate for the linear/nonparametric indicators for interactions. Default is 0.1.
`multicore`	Whether to use multiple cores in computation. Strongly recommended but may not work on Windows. Default is TRUE.
`cores`	Number of cores to use with multicore. Default (`NULL`) uses all cores.
`k`	Basis dimension for nonparametric terms estimated using splines. Default is 10.
`bs`	Spline basis for nonparametric terms. Specified as a two letter character string. Default is the natural cubic spline, bs="cr". See `smooth.terms` for an overview of what is available.
`family`	Specifies the family for the gam (see `family` and `family.mgcv`). Default is gaussian().
`method`	Specifies the metric for smoothing parameter selection (see `gam`). Default is "REML".
`optimizer`	Specifies the numerical optimization algorithm for the gam (see `gam`). Default is c("outer","newton").
`reduc`	Implements additional variable elimination methods at the end of the run of the genetic algorithm. User can choose between methods 1, 2, and 3. Multiple methods can be chosen. E.g. reduc=c(1) or reduc=c(1,3). See the GAGAM paper for an explanation of the methods. Default is NULL.
`always_par`	Vector of the column numbers (in x) of the variables always estimated parametrically (for noncontinuous predictors).

A list containing: gam object (fitted best model), vector of indexes or names of variables included linearly, vector of indexes or names of variables included nonparametrically (and the same lists for interactions if Kint is greater than 0).

Cus, Mark. 2020. "Simultaneous Variable Selection And Structure Discovery In Generalized Additive Models". https://github.com/markcus1/gagam/blob/master/GAGAMpaper.pdf.

N <- 500
set.seed(123)
xdat <- matrix(rnorm(N*10,0,1),nrow=N,ncol=10)
ydat <- 4*xdat[,1]+5*xdat[,2]+6*xdat[,3]+(xdat[,4])^2 + 4 + rnorm(N,0,0.25)

## Not run: 
example_gagam <- gagam(ydat,xdat,Kvar = 6,no_gen = 50)

## End(Not run)