Fix spurious warning in select_by_pct_loss() when losses tie for best

NEWS.md

@@ -12,6 +12,8 @@
 
 * `min_grid()` now returns rows in the order of the non-submodel parameters rather than putting rows with submodels first (#987).
 
+* `select_by_pct_loss()` no longer emits a spurious warning when multiple tuning parameter combinations tie for the best loss (#1195)
+
 # tune 2.1.0
 
 * Model tuning has been enabled for quantile regression models. (#1125)

R/select_best.R

@@ -202,7 +202,8 @@ select_by_pct_loss.tune_results <- function(
 
   # discard models more complex than the best and
   # remove models with greater increase in loss than the limit
-  best_index <- which(summary_res$.loss == 0)
+  # (use max() since multiple configurations can tie for the best loss)
+  best_index <- max(which(summary_res$.loss == 0))
   summary_res |>
     dplyr::slice(1:best_index) |>
     dplyr::filter(.loss < limit) |>

tests/testthat/_snaps/select_best.md

@@ -213,6 +213,11 @@
       Error in `select_by_pct_loss()`:
       ! Could not sort results by `weight_funk` and `desc(K)`.
 
+# select_by_pct_loss() handles ties for the best loss (#1195)
+
+    Code
+      res <- select_by_pct_loss(knn_results, metric = "accuracy", limit = 5, K)
+
 # show_best with survival models
 
     Code

tests/testthat/test-select_best.R

@@ -225,6 +225,28 @@ test_that("percent loss", {
   )
 })
 
+test_that("select_by_pct_loss() handles ties for the best loss (#1195)", {
+  skip_if_not_installed("dials", minimum_version = "1.4.0")
+
+  knn_results <- readRDS(test_path("data", "knn_results.rds"))
+
+  tied_metrics <- tibble::tibble(
+    K = c(1, 4, 12, 14, 17),
+    weight_func = "rectangular",
+    exponent = 1,
+    .config = paste0("cfg", 1:5),
+    mean = c(0.70, 0.75, 0.90, 0.90, 0.80)
+  )
+  local_mocked_bindings(
+    .filter_perf_metrics = function(...) tied_metrics
+  )
+
+  expect_snapshot(
+    res <- select_by_pct_loss(knn_results, metric = "accuracy", limit = 5, K)
+  )
+  expect_equal(res$K, 12)
+})
+
 test_that("select_by_* can handle metrics with direction == 'zero'", {
   skip_on_cran()
   skip_if_not_installed("kknn")