Learn R Programming

lmmlasso (version 0.1-2)

lmmlassoControl: Options for the lmmlasso Algorithm

Description

Definition of various kinds of options in the algorithm.

Usage

lmmlassoControl(tol = 10^(-4), trace = 1, maxIter = 1000, maxArmijo = 20, number = 5, a_init = 1, delta = 0.1, rho = 0.001, gamma = 0, lower = 10^(-6), upper = 10^8, seed = 532, VarInt = c(0, 10), CovInt = c(-5, 5), thres = 10^(-4))

Arguments

tol
convergence tolerance
trace
integer. 1 prints no output, 2 prints warnings, 3 prints the current function values and warnings (not recommended)
maxIter
maximum number of (outer) iterations
maxArmijo
maximum number of steps to be chosen in the Armijo Rule. If the maximum is reached, the algorithm continues with optimizing the next coordinate.
number
integer. Determines the active set algorithm. The zero fixed-effects coefficients are only updated each number iteration. It may be that a smaller number increases the speed of the algorithm. Use $0 \le number \le 5$.
a_init
$\alpha_{init}$ in the Armijo step. See Schelldorfer et. al. (2010).
delta
$\delta$ in the Armijo step. See Schelldorfer et. al. (2010)
rho
$\rho$ in the Armijo step. See Schelldorfer et. al. (2010)
gamma
$\gamma$ in the Armijo step. See Schelldorfer et. al. (2010)
lower
lower bound for the Hessian
upper
upper bound for the Hessian
seed
set.seed for calculating the starting value, which performs a 10-fold cross-validation.
VarInt
Only for opt="optimize". The interval for the variance parameters used in "optimize". See help("optimize")
CovInt
Only for opt="optimize". The interval for the covariance parameters used in "optimize". See help("optimize")
thres
If a variance or covariance parameter has smaller absolute value than thres, the parameter is set to exactly zero.

Details

For the Armijo step parameters, see Bertsekas (2003)

References

Dimitri P. Bertsekas (2003) Nonlinear Programming, Athena Scientific. J. Schelldorfer, P. B\"uhlmann and S. van de Geer (2011), Estimation for High-Dimensional Linear Mixed-Effects Models Using $\ell_1$-penalization, arXiv preprint 1002.3784v2