fits semiparametric SDR models (PHD approach)
semi.phd(
x,
y,
d = 5L,
maxit = 100L,
h = NULL,
opt.method = c("lbfgs.x", "bfgs", "lbfgs2", "bfgs.x", "lbfgs", "spg", "ucminf", "CG",
"nlm", "nlminb", "newuoa"),
nn = 0.95,
init.method = c("random", "phd"),
optimize.nn = FALSE,
verbose = TRUE,
n.samples = 100,
degree = 2,
vic = TRUE,
...
)an n x p matrix of covariates, where each row is an observation and each column is a predictor
vector of responses of length n
an integer representing the structural dimension
maximum number of iterations
bandwidth parameter. By default, a reasonable choice is selected automatically
optimization method to use. Available choices are
c("lbfgs2", "lbfgs.x", "bfgs.x", "bfgs", "lbfgs", "spg", "ucminf", "CG", "nlm", "nlminb", "newuoa")
nearest neighbor parameter for locfit.raw
method for parameter initialization. Either "random" for random initialization or "phd"
for a principle Hessian directions initialization approach
should nn be optimized? Not recommended
should results be printed along the way?
number of samples for the random initialization method
degree of kernel to use
logical value of whether or not to compute the VIC criterion for dimension determination
extra arguments passed to locfit.raw
A list with the following elements
beta estimated sufficient dimension reduction matrix
beta.init initial sufficient dimension reduction matrix -- do not use, just for the sake of comparisons
cov variance covariance matric for the covariates
sqrt.inv.cov inverse square root of the variance covariance matrix for the covariates. Used for scaling
solver.obj object returned by the solver/optimization function
vic the penalized VIC value. This is used for dimension selection, with dimension chosen to minimize this penalized vic value that trades off model complexity and model fit