Skip to contents

Optimize a Sample Design With Random Forest Simulated Annealing

Source: R/rf_optimization.R
ss_rf_optimize.Rd

Starts from a cLHS baseline and iteratively swaps sample points using simulated annealing to minimize cross-validated Random Forest MSE (ss_rf_mse()), following Wadoux et al. (2019).

Usage

ss_rf_optimize(
  covariates,
  n_samples,
  n_iterations = 500,
  seed = NULL,
  temperature = 1000,
  cooling_rate = 0.95,
  target_var = NULL
)

Arguments

covariates

A SpatRaster stack of environmental covariates.

n_samples

Integer, number of samples in the design.

n_iterations

Integer, number of simulated annealing iterations. Default 500.

seed

Optional integer seed. The cLHS baseline uses seed; the annealing loop uses seed + 1, so the two stages stay reproducible but independent.

temperature

Numeric, initial annealing temperature. Default 1000.

cooling_rate

Numeric in (0, 1), multiplicative cooling factor applied each iteration. Default 0.95.

target_var

Character, covariate to use as the RF response. Passed to ss_rf_mse().

Value

A list with:

optimized_samples

Data frame, the best design found.

initial_samples

Data frame, the cLHS baseline design.

initial_mse

Numeric, baseline MSE.

final_mse

Numeric, MSE of optimized_samples.

improvement

Numeric, percent MSE reduction vs. baseline.

See also

ss_clhs_sample(), ss_rf_mse(), ss_rf_size()

Examples

if (FALSE) { # \dontrun{
r <- terra::rast("data/predictors.tif")
res <- ss_rf_optimize(r, n_samples = 100, n_iterations = 500, seed = 123)
res$improvement
} # }