Skip to contents

Newton-Raphson optimizer

OptimizerNR.Rd

Implements Newton-Raphson updates by combining gradients and Hessians of the objective with optional Armijo backtracking. Falls back to the gradient direction when the Hessian is singular or indefinite.

Super class

vistool::Optimizer -> OptimizerNR

Methods

Public methods

Inherited methods

Method new()

Creates a new Newton-Raphson optimizer.

Usage

OptimizerNR$new(
  objective,
  x_start,
  step_size = 1,
  gamma = 0.99,
  tau = 0.5,
  max_backtracks = 20L,
  fallback = c("gradient", "stop"),
  id = "Newton-Raphson",
  print_trace = TRUE
)

Arguments

objective

(Objective)
The objective to optimize.

objective

(Objective)
The objective to optimize.

x_start

(numeric())
Start value of the optimization. Note, after the first call of $optimize() the last value is used to continue optimization. Get this value with $x.

x_start

(numeric())
Start value of the optimization. Note, after the first call of $optimize() the last value is used to continue optimization. Get this value with $x.

step_size

(numeric(1)) Initial step size used as the starting value for line search. Default 1.

gamma

(numeric(1)) Armijo slope parameter in (0, 1).

tau

(numeric(1)) Multiplicative backtracking factor in (0, 1).

max_backtracks

(integer(1)) Maximum number of Armijo backtracking attempts.

fallback

(character(1)) Strategy when the Hessian cannot be inverted. One of "gradient" or "stop".

id

(character(1))
Id of the object.

id

(character(1))
Id of the object.

print_trace

(logical(1))
Indicator whether to print the status of $optimize().

print_trace

(logical(1))
Indicator whether to print the status of $optimize().

Method optimize()

Optimize the associated objective for a number of steps.

Usage

OptimizerNR$optimize(steps = 1L, step_size_control = NULL, minimize = NULL)

Arguments

steps

(integer(1))
Number of steps/iterations.

step_size_control

(function()) Optional callback with signature (x, u, obj, opt) returning a scalar step size. Defaults to Armijo backtracking that uses the configured gamma, tau, and max_backtracks parameters.

minimize

(logical(1)) Whether to minimize the objective. Defaults to self$objective$minimize. Maximization is not supported and triggers an error.

Method step_size()

Access or update the base step size.

Usage

OptimizerNR$step_size(x)

Arguments

x

(numeric(1)) Replacement value. Omit to retrieve the current setting.

Method gamma()

Access or update the Armijo slope parameter.

Usage

OptimizerNR$gamma(x)

Arguments

x

(numeric(1))

Method tau()

Access or update the Armijo shrink factor.

Usage

OptimizerNR$tau(x)

Arguments

x

(numeric(1))

Method clone()

The objects of this class are cloneable with this method.

Usage

OptimizerNR$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.