optimStepSizeFactor: Coarse line search for optimum step-size modification factor

In GAMBoost: Generalized linear and additive models by likelihood based boosting

Description

This routine helps in finding an optimum step-size modification factor for GAMBoost, i.e., that results in an optimum in terms of cross-validated log-likelihood.

Usage

optimStepSizeFactor(x=NULL,y,x.linear=NULL,
                    direction=c("down","up","both"),start.stepsize=0.1,
                    iter.max=10,constant.cv.res=NULL,parallel=FALSE,
                    trace=FALSE,...) 

Arguments

x	n * p matrix of covariates with potentially non-linear influence. If this is not given (and argument x.linear is employed), a generalized linear model is fitted.
y	response vector of length n.
x.linear	optional n * q matrix of covariates with linear influence.
direction	direction of line search for an optimal step-size modification factor (starting from value 1).
start.stepsize	step size used for the line search. A final step is performed using half this size.
iter.max	maximum number of search iterations.
constant.cv.res	result of cv.GAMBoost (with just.criterion=TRUE) for stepsize.factor.linear=1, that can be provided for saving computing time, if it already is available.
parallel	logical value indicating whether evaluation of cross-validation folds should be performed in parallel on a compute cluster. This requires library snowfall.
trace	logical value indicating whether information on progress should be printed.
...	miscellaneous parameters for cv.GAMBoost.

Details

A coarse line search is performed for finding the best parameter stepsize.factor.linear for GAMBoost. If an pendistmat.linear argument is provided (which is passed on to GAMBoost), a search for factors smaller than 1 is sensible (corresponding to direction="down"). If no connection information is provided, it is reasonable to employ direction="both", for avoiding restrictions without subject matter knowledge.

Value

List with the following components:

factor.list	array with the evaluated step-size modification factors.
critmat	matrix with the mean log-likelihood for each step-size modification factor in the course of the boosting steps.
optimal.factor.index	index of the optimal step-size modification factor.
optimal.factor	optimal step-size modification factor.
optimal.step	optimal boosting step number, i.e., with minimum mean log-likelihood, for step-size modification factor optimal.factor.

Author(s)

Written by Harald Binder [e-mail].

References

Binder, H. and Schumacher, M. (2009). Incorporating pathway information into boosting estimation of high-dimensional risk prediction models. BMC Bioinformatics. 10:18.

See Also

GAMBoost, cv.GAMBoost

Examples

## Not run: 
##  Generate some data 
n <- 100; p <- 10

#   covariates with non-linear (smooth) effects
x <- matrix(runif(n*p,min=-1,max=1),n,p)             
eta <- -0.5 + 2*x[,1] + 2*x[,3]^2 + x[,9]-.5
y <- rbinom(n,1,binomial()$linkinv(eta))

#  Determine step-size modification factor for a generalize linear model
#  As there is no connection matrix,  perform search into both directions 

optim.res <- optimStepSizeFactor(direction="both",
                                y=y,x.linear=x,family=binomial(),
                                penalty.linear=200,
                                trace=TRUE)

#   Fit with obtained step-size modification parameter and optimal number of boosting
#   steps obtained by cross-validation

gb1 <- GAMBoost(x=NULL,y=y,x.linear=x,family=binomial(),penalty.linear=200,
                stepno=optim.res$optimal.step,
                stepsize.factor.linear=optim.res$optimal.factor) 

summary(gb1)


## End(Not run)

GAMBoost documentation built on May 2, 2019, 12:40 p.m.