
Nested tuning without information leakage
nested-tuning.RmdCandidate parameters must be compared on fixed inner partitions. When
an outer design is requested, ps_tune_interpolation() sees
each outer holdout only after selecting parameters inside that outer
training set. Every failed candidate remains in the manifest. Without
outer validation, scores are tuning performance and should not be
presented as unbiased final performance.
tuned <- ps_tune_interpolation(
points, "IDW", list(idw_power=c(1, 1.5, 2), idw_nmax=c(8, 15)),
inner_design="spatial_block", outer_design="kfold", seed=31
)