nlr (version 0.1-3)

nl.form-class: Class "nl.form"

Description

It is a class of nonlinear regression model function, can embed expression formula and function. It is a more general object such that the heterogeneous variance model and robust loss function in nlr is saved as nl.form object.

Objects from the Class

Objects can be created by calls of the form:

new("nl.form",formula,fnc,formtype,p, inv=NULL,name=name,par,arguments=list(...),dependent, independent,origin,selfStart) ,

or constructor:

nl.form(form,p=NULL,inv=NULL,name,par=NULL,dependent=NULL,independent=NULL,origin=NULL,selfStart=NULL,...)

Slots

formula:

Object of class "call or NULL" it can be 1- a two sided formula with response (or a function of response) in left of ~ and nonlinear function model is a function of predictors and parameters, or 2- a one sided formula with ~nonlinear model in right, is again a function of predictors and unknown parameters. If Null then the nonlinear model is a R function stored in fnc slot as bellow. Each of right side or left side formula can return "gradient" and "hessian" as attributed value.

fnc:

Object of class "function or NULL" is nonlinear model stored as R function.

dependent:

Object of class "characterorNULL" of predictor variable, null value means the formula slot is one sided.

independent:

Object of class "characterorNULL" caharacter vector name of predictor variables, which can be more than one predictor.

formtype:

Object of class "character" character type of nl.form. Do not insert this slot, it will be set automatically by creater, it use internaly for evaluation.

p:

Object of class "numericorNULL" number of parameters.

inv:

Object of class "callorNULL" if nonlinear model is function of one predictor the inverse function define in this slot. Still not functioning, designed for feature extention, so it can be ignored at the moment, but it is better to define.

name:

Object of class "character" a character name for the nonlinear function model.

par:

Object of class "list" of parameters, assigned value to parameters will be used in worst case that initial values can not be computed.

arguments:

Object of class "list" list of extra arguments use in formula or fnc slot.

origin:

Object of class "callorNULL" is original one sided or two sided expression of nonlinear model withought gradient and hessian.

selfStart:

Object of class "functionorNULL" selfstart function defined for initial values guiss. Is same as nls or nlme::nlme functions.

Extends

Class "nl.formorNULL", directly.

Methods

$

signature(x = "nl.form"): return slots.

all.vars

signature(expr = "nl.form"): Return a character vector containing all the names which occur in "formula" or or "fnc".

eval

signature(expr = "nl.form"): evaluate nl.form object in the environment include parameters and predictor variables.

evald

signature(expr = "nl.form"): same functionality as eval but extended for compatibility. eval.nl.form has same functionality created for compatibility purpose.

getInitial

signature(object = "nl.form"): get initial value from selfstart or par slot.

selfStart

signature(model = "nl.form"): self computing initial value, if not given "getInitial" method return initial values from environment or "par" slot.

eval.nl.form

signature(expr = "eval.nl.form"): eval.nl.form has same functionality as eval, created for compatibility purpose.

References

Riazoshams H, Midi H, and Ghilagaber G, 2018,. Robust Nonlinear Regression, with Application using R, Joh Wiley and Sons.

See Also

nl.robfuncs

Examples

# NOT RUN {
# define hampel robust loss as a function
hampel <- nl.form(
	  form = function(t,
	                  a = 1.345,
	                  k0 = 3.73677,
	                  k1 = 4,
	                  maxrho5 = 1.345,
	                  ...) {
	    U <- abs(t)
	    Ugrta <- (U > abs(a))
	    .rho <- .grad <- .hess <- .weight <- NULL
	    .rho[Ugrta] <- 2. * abs(a) * U[Ugrta] - a * a
	    .rho[!Ugrta] <- t[!Ugrta] ^ 2
	    .grad[Ugrta] <- 2. * abs(a) * sign(t[Ugrta])
	    .grad[!Ugrta] <- 2. * t[!Ugrta]
	    .hess[Ugrta] <- 0.
	    .hess[!Ugrta] <- 2.
	    .weight[Ugrta] <- 2. * abs(a) / U[Ugrta]
	    .weight[!Ugrta] <- 2.
	    attr(.rho, "gradient") <- .grad
	    attr(.rho, "hessian") <- .hess
	    attr(.rho, "weight") <- .weight
	    return(.rho)
	  },
	  name = "huber",
	  independent = "t",
	  a = 1.345,
	  k0 = 3.73677,
	  k1 = 4,
	  maxrho5 = 1.345
	)
# }