NesterovInitialize: Nesterov Initialization

In nchenderson/nidaarem: Nesterov Initialized Damped Anderson Acceleration with Epsilon Monotonicity

Description

Performs Nesterov initialization where an initial parameter value is found by performing a preliminary number of Nesterov steps.

Usage

NesterovInitialize(par, fixptfn, objfn = NULL, test = "monotone", ..., 
                   control = list(maxiter = 500))

Arguments

par	The initial parameter vector
fixptfn	The fixed point function which performs the basic parameter update.
objfn	The objective function of interest.
test	The test which determines termination of the Nesterov iterations. Can be either "monotone" or "gradient".
...	Additional arguments that may be needed by fixptfn or objfn
control	A list of additional control parameters

Value

A list with the following components

par	Initial value of the parameter vector
value.obj	value of objective function at termination
objfn.val	values of the objective function at each iteration
num.iter	number of fixed-point updates performed

Author(s)

Nicholas Henderson

Examples

n <- 1000
p <- 100

beta.coef <- rt(p, df=2)
XX <- matrix(rnorm(n*p), nrow=n, ncol=p)
yy <- XX%*%beta.coef + rnorm(n) 
VV <- eigen(crossprod(XX))
LL <- max(VV$values)

ne.mon <- NesterovInitialize(par=rep(0, p), fixptfn=GDLassoStep, objfn = LassoObjFn, 
                               test = "monotone", X=XX, y=yy, lambda=.0001, 
                               stplngth=1/LL)

ne.grad <- NesterovInitialize(par=rep(0, p), fixptfn=GDLassoStep, objfn = LassoObjFn, 
                                test = "gradient", X=XX, y=yy, lambda=.0001, 
                                stplngth=1/LL)

ne.mon$num.iter
ne.grad$num.iter

nchenderson/nidaarem documentation built on Feb. 19, 2020, 12:45 p.m.