Paper-default undersmoothing rule for kernel order

Builds a tuning spec implementing $$s(\lambda) = \lfloor \lambda n^{(k_x+1)/(k_x+2)} \rfloor$$ with finite-sample default \(\lambda = 0.4\) and optional candidate grid.

Usage

tuning_dnn_paper(
  lambda = 0.4,
  lambda_grid = c(0.3, 0.4, 0.5, 0.75),
  rounding = "floor",
  kx_override = NULL
)

Arguments

lambda

Numeric scalar tuning constant.

lambda_grid

Numeric vector of candidate constants for grid mode.

rounding

One of "floor", "ceiling", or "round".

kx_override

Optional positive integer to override runtime kx.

Value

A dnn_tuning spec.

Examples

tun <- tuning_dnn_paper()
tuning_eval(tun, n = 1000, kx = 2, k = 1)
#> $s
#> [1] 71
#> 
#> $meta
#> $meta$lambda
#> [1] 0.4
#> 
#> $meta$formula
#> [1] "lambda * n^((kx + 1)/(kx + 2))"
#> 
#> $meta$rounding
#> [1] "floor"
#> 
#> $meta$name
#> [1] "paper_undersmooth"
#> 
#> $meta$n
#> [1] 1000
#> 
#> $meta$kx
#> [1] 2
#> 
#> 
#> $feasibility
#> $feasibility$k_le_s
#> [1] TRUE
#> 
#> $feasibility$s_lt_n
#> [1] TRUE
#> 
#>