stepPenal (version 0.2)

optimPenaLik: Variable selection based on the combined penalty CL= (1-w)L0 + wL1

Description

Methods to use for optimization include Hooke-Jeeves derivative-free minimization algorithm (hjk), or the BFGS method (modified Quasi-Newton). This method does variable selection by shrinking the coefficients towards zero using the combined penalty (CL= (1-w)L0 + wL1).

Usage

optimPenaLik(Data, lamda, w, standardize = TRUE, algorithms = c("QN",
  "hjk"))

Arguments

Data

should have the following structure: the first column must be the response variable y

lamda

tuning penalty parameter

w

the weight parameter for the sum (1-w)L0+ wL1

standardize

standardize Logical flag for x variable standardization, prior to fitting the model sequence. The coefficients are always returned on the original scale. Default is standardize=TRUE

algorithms

select between BFGS ('QN') or Hooke-Jeeves (hjk) algorithm.

Value

a list with the shrinked coefficients and the names of the selected variables, i.e those variables with estimated coefficient different from zero.

Details

it is recommended to use the tuneParam function to tune parameters lamda and w prior using the optimPenaLik function.

See Also

optim

Examples

# NOT RUN {
# use the optimPenaLik function on a simulated dataset, with given lamda and w.
# }
# NOT RUN {
set.seed(14)
beta    <- c(3, 2, -1.6, -1)
noise   <- 5
simData <- SimData(N=100, beta=beta, noise=noise, corr=TRUE)

# use BFGS

before  <- Sys.time()
PenalQN <- optimPenaLik(Data=simData, lamda=1.5, w=0.7,
                     algorithms=c("QN"))
(tot <- Sys.time()-before)
PenalQN


# use Hooke-Jeeves algorithm

before  <- Sys.time()
Penalhjk <- optimPenaLik(Data=simData, lamda=1.5, w=0.7,
                       algorithms=c("hjk"))
(totRun  <- Sys.time() - before)
# total run of approx 0.25sec

Penalhjk
# }
# NOT RUN {
# }