refactor(qav3): add tunables for knn based trials selection

Signed-off-by: Jérôme Benoit <jerome.benoit@piment-noir.org>

diff --git a/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py b/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py
index 8980184448cd2bc7be19f56b06583f58906efa02..40d50945887eb101f32de21b7b1d8967b5cef2dc 100644 (file)
--- a/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py
+++ b/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py
@@ -437,8 +437,10 @@ class QuickAdapterRegressorV3(BaseRegressionModel):
             "harmonic_mean",
             "power_mean",
             "weighted_sum",
-            "d1",
-            "d2",
+            "knn-d1",
+            "knn-d2-mean",
+            "knn-d2-median",
+            "knn-d2-max",
         }
         if label_metric not in metrics:
             raise ValueError(
@@ -470,6 +472,10 @@ class QuickAdapterRegressorV3(BaseRegressionModel):
             )
             if np_weights.size != normalized_matrix.shape[1]:
                 raise ValueError("label_weights length must match number of objectives")
+            label_knn_metric = self.ft_params.get("label_knn_metric", "euclidean")
+            knn_kwargs = {}
+            if label_knn_metric == "minkowski" and isinstance(p_order, float):
+                knn_kwargs["p"] = p_order
             ideal_point = np.ones(normalized_matrix.shape[1])
 
             if metric in {
@@ -533,23 +539,34 @@ class QuickAdapterRegressorV3(BaseRegressionModel):
                 ) - sp.stats.pmean(normalized_matrix, p=p, weights=np_weights, axis=1)
             elif metric == "weighted_sum":
                 return np.sum(np_weights * (ideal_point - normalized_matrix), axis=1)
-            elif metric == "d1":
+            elif metric == "knn-d1":
                 if normalized_matrix.shape[0] < 2:
                     return np.full(normalized_matrix.shape[0], np.inf)
-                nbrs = sklearn.neighbors.NearestNeighbors(n_neighbors=2).fit(
-                    normalized_matrix
-                )
+                nbrs = sklearn.neighbors.NearestNeighbors(
+                    n_neighbors=2, metric=label_knn_metric, **knn_kwargs
+                ).fit(normalized_matrix)
                 distances, _ = nbrs.kneighbors(normalized_matrix)
                 return distances[:, 1]
-            elif metric == "d2":
+            elif metric in {"knn-d2-mean", "knn-d2-median", "knn-d2-max"}:
                 if normalized_matrix.shape[0] < 2:
                     return np.full(normalized_matrix.shape[0], np.inf)
-                k = min(4, normalized_matrix.shape[0] - 1) + 1
-                nbrs = sklearn.neighbors.NearestNeighbors(n_neighbors=k).fit(
-                    normalized_matrix
+                n_neighbors = (
+                    min(
+                        int(self.ft_params.get("label_knn_d2_min_n_neighbors", "4")),
+                        normalized_matrix.shape[0] - 1,
+                    )
+                    + 1
                 )
+                nbrs = sklearn.neighbors.NearestNeighbors(
+                    n_neighbors=n_neighbors, metric=label_knn_metric, **knn_kwargs
+                ).fit(normalized_matrix)
                 distances, _ = nbrs.kneighbors(normalized_matrix)
-                return np.mean(distances[:, 1:], axis=1)
+                if metric == "knn-d2-mean":
+                    return np.mean(distances[:, 1:], axis=1)
+                elif metric == "knn-d2-median":
+                    return np.median(distances[:, 1:], axis=1)
+                elif metric == "knn-d2-max":
+                    return np.max(distances[:, 1:], axis=1)
             else:
                 raise ValueError(f"Unsupported distance metric: {metric}")