Skip to contents

Tune interpolation parameters under recorded validation partitions

ps_tune_interpolation.Rd

Compares explicit candidate configurations using the same deterministic inner partitions. With an outer design, tuning occurs only inside each outer training partition and the selected configuration is evaluated on its outer holdout. Without an outer design, reported performance is tuning performance, not an unbiased estimate of final predictive performance.

Usage

ps_tune_interpolation(
  points,
  method,
  candidates,
  inner_design = "spatial_block",
  outer_design = NULL,
  inner_folds = 5,
  outer_folds = 5,
  repeats = 1,
  metric = "rmse",
  minimum_coverage = 0.9,
  template = NULL,
  mask = NULL,
  grid_res = NULL,
  refit = TRUE,
  seed = 1,
  progress = NULL,
  search = c("grid", "random"),
  maximum_runs = 1000
)

Arguments

points

Groundwater-head points.

method

One of "TPS", "IDW", "OK", or "UK".

candidates

Candidate table or named parameter list.

inner_design, outer_design

Inner and optional outer validation designs.

inner_folds, outer_folds, repeats

Fold and repeat counts.

metric

Objective metric minimized during selection.

minimum_coverage

Minimum finite-prediction coverage.

template, mask, grid_res

Fixed mapping controls.

refit

Refit the selected configuration to all observations.

seed

Deterministic seed.

progress

Optional callback.

Exhaustive grid or reproducible row sampling.

maximum_runs

Maximum candidate-by-fold run guard.

Value

A potentiomap_tuning object.

Examples

data("synthetic_wells")
p <- ps_make_points(synthetic_wells[1:12, ], "x", "y", "gw_elevation",
                    "well_id", "EPSG:26916")
tuned <- ps_tune_interpolation(p, "IDW", list(idw_power = c(1.5, 2)),
                               inner_design = "kfold", inner_folds = 3,
                               refit = FALSE, seed = 4)
#> [inverse distance weighted interpolation]
#> [inverse distance weighted interpolation]
#> [inverse distance weighted interpolation]
#> [inverse distance weighted interpolation]
#> [inverse distance weighted interpolation]
#> [inverse distance weighted interpolation]
tuned$candidates[, c("candidate_id", "metric", "coverage", "selected")]
#>     candidate_id   metric coverage selected
#> 1 candidate_0001 1.336077        1    FALSE
#> 2 candidate_0002 1.222580        1     TRUE
# These are tuning scores, not unbiased final performance estimates.