Control parameters for the pattern model estimator

Usage

pattern_control(
  max_rank = 8L,
  x_scale = c("none", "sd"),
  y_scale = c("none", "sd"),
  noise = list(type = "diag_lowrank", rank = "auto", max_rank = 10L, shrink = 0.1),
  lambda_2 = 0,
  max_outer = 50L,
  tol = 1e-08,
  refine_path = FALSE,
  max_inner = 500L,
  tol_inner = 1e-09,
  tol_iterate = 1e-06
)

Arguments

max_rank: Maximum rank considered (capped at the eligible rank: number of classes minus one for categorical targets, the effective number of target dimensions otherwise, and never above the number of features or observations).
x_scale: Feature scaling on training rows: "none" (centre only) or "sd".
y_scale: Continuous-target scaling before whitening: "none" or "sd".
noise: Residual covariance specification: a list with type ("diag_lowrank", "diag", or "identity"), rank ("auto" selects components above the Marchenko-Pastur edge, or an integer), max_rank, and shrink (shrinkage of the diagonal toward its median).
lambda_2: Reserved for the penalized solver; must be 0. A ridge on the patterns turns the A-step into a Sylvester equation under a general residual covariance, so it is rejected rather than solved approximately.
max_outer: Maximum number of alternating (C-step, A-step) updates.
tol: Relative objective change that declares convergence.
refine_path: Logical; when fitting a rank path, run the alternating refinement for every rank (default FALSE: unpenalized path solutions are the exact reduced-rank optima, so refinement changes nothing). Refinement is always used when a spatial penalty is active.
max_inner: Maximum proximal-gradient iterations per penalized A-step.
tol_inner: Relative objective change that stops the A-step solver.
tol_iterate: Relative change in the patterns required alongside tol_inner. The objective is flat near the optimum, so it can settle while the patterns are still moving; both must be small.

Value

A list of class pattern_control.

Examples

pattern_control(max_rank = 3)
#> $max_rank
#> [1] 3
#> 
#> $x_scale
#> [1] "none"
#> 
#> $y_scale
#> [1] "none"
#> 
#> $noise
#> $noise$type
#> [1] "diag_lowrank"
#> 
#> $noise$rank
#> [1] "auto"
#> 
#> $noise$max_rank
#> [1] 10
#> 
#> $noise$shrink
#> [1] 0.1
#> 
#> 
#> $lambda_2
#> [1] 0
#> 
#> $max_outer
#> [1] 50
#> 
#> $tol
#> [1] 1e-08
#> 
#> $refine_path
#> [1] FALSE
#> 
#> $max_inner
#> [1] 500
#> 
#> $tol_inner
#> [1] 1e-09
#> 
#> $tol_iterate
#> [1] 1e-06
#> 
#> attr(,"class")
#> [1] "pattern_control" "list"