fmin: Minimize a function f using a Quasi-Newton method

In r-glennie/fmin: Minimize a multivariate function

Description

Performs a quasi-Newton line search optimization with BFGS updates of the inverse hessian. See Nocedal and Wright (2006), Chapter 3 for details.

Usage

fmin(
  obj,
  theta,
  gobj = NULL,
  funit = 1,
  units = NULL,
  maxit = 1000,
  tol = 1e-10,
  stepmax = 1,
  maxsubsteps = 10,
  save = FALSE,
  verbose = FALSE,
  digits = 4,
  ...
)

Arguments

obj	function to be minimised which takes parameter vector as first argument
gobj	(optional) gradient function of f, uses finite differencing otherwise
funit	scaling for function values, try to pick this so that f values vary between -1 and 1
units	a vector of same length as start giving coordinate units, try to pick units such that values of parameter in that coordinate direction vary between -1 and 1
maxit	maximum number of iterations to try
tol	tolerance for stopping criteria
stepmax	maximum relative step length, default of 1 means no steps longer than full Newton step
maxsubsteps	maximum number of substeps in line search
save	if TRUE then parameter values, function values, and gradients are saved for every iteration
verbose	if TRUE, function value and then parameter value are printed for each iteration
digits	number of digits to print when verbose = TRUE
...	other named arguments to f and gobj (if given)
start	vector of starting values

Value

a list with elements:

• estimate - a vector of the optimal estimates

• value - value of the objective function at the optimum

• g - gradient vector at the optimum

• H - hessian at the optimum computed by finite difference

• conv - is TRUE if convergence was satisfied, otherwise may not have converged on optimum

• niter: number of iterations taken

r-glennie/fmin documentation built on Dec. 22, 2021, 11:53 a.m.