arfc_control() collects numerical and adaptive-weight controls used by
arfc() and arfc_select(). The defaults reproduce the computational
convention used in the accompanying ARFC study. No training/validation
split is created by these controls.
Usage
arfc_control(
max_iter = 120L,
tol = 1e-08,
norm_floor = 1e-08,
relative_cutoff_multiplier = 1e-04,
cutoff_floor = 1e-08,
refit = TRUE,
trace = FALSE,
keep_matrices = FALSE,
adaptive_alpha = 1,
adaptive_epsilon_fraction = 0.05,
adaptive_clip = c(0.25, 4),
zero_tolerance = 1e-08,
tie_relative_tolerance = 1e-12,
strict = TRUE
)Arguments
- max_iter
Positive integer giving the maximum LQA iterations for each tuning candidate.
- tol
Positive relative coefficient convergence tolerance.
- norm_floor
Positive stabilizer for regional norms near zero.
- relative_cutoff_multiplier
Nonnegative multiplier applied to the largest absolute smooth-pilot coefficient. The resulting threshold is used for exact coefficient thresholding.
- cutoff_floor
Positive lower bound for the absolute threshold.
- refit
Logical; refit the roughness-penalized model on retained basis coefficients after thresholding.
- trace
Logical; retain exact-objective traces for tuning candidates.
- keep_matrices
Logical; retain design and regional matrices in fitted objects. This can substantially increase object size.
- adaptive_alpha
Positive exponent in the pilot-adaptive regional weights.
- adaptive_epsilon_fraction
Positive fraction of the median positive pilot norm used in the stabilizing offset.
- adaptive_clip
Two positive increasing numbers bounding normalized adaptive weights.
- zero_tolerance
Nonnegative tolerance used only to display estimated zero regions in plots.
- tie_relative_tolerance
Nonnegative relative tolerance for validation loss ties in
arfc_select().- strict
Logical. If
TRUE, only successful, converged, successfully refitted candidates with finite coefficients and BIC are eligible.