Adding default scale to Hypervolume log

Assets/StreamingAssets/BOData/BayesianOptimization/meta_mobo_runtime.py

@@ -489,15 +489,20 @@ def rewrite_pareto_flags(y_all_norm):
     df.to_csv(obs_csv, sep=';', index=False)
 
 
-def save_hypervolume_to_file(hvs, iteration):
+def save_hypervolume_to_file(hvs, iteration, ref_point):
     hv_csv = os.path.join(PROJECT_PATH, "HypervolumePerEvaluation.csv")
     os.makedirs(os.path.dirname(hv_csv), exist_ok=True)
     write_header = not os.path.exists(hv_csv) or os.path.getsize(hv_csv) == 0
     with open(hv_csv, 'a', newline='') as f:
         w = csv.writer(f, delimiter=';')
         if write_header:
-            w.writerow(["Hypervolume", "Run"])
-        w.writerow([hvs[-1], iteration])
+            w.writerow(["Hypervolume", "Run", "Scale", "ReferencePoint"])
+        w.writerow([
+            hvs[-1],
+            iteration,
+            "normalized maximize-space [-1,1] per objective",
+            "[" + ",".join(str(float(value)) for value in ref_point) + "]",
+        ])
 
 
 def append_meta_weights_row(weights_csv, iteration, optimizer, source_names):
@@ -595,22 +600,23 @@ def meta_execute(conn, seed, iterations, initial_samples):
     print("Initial Sobol X in [0,1]:", x_init, flush=True)
 
     y_rows = []
+    hvs = []
     for i in range(initial_samples):
         print(f"---- Initial Sample {i+1}", flush=True)
         y_row = objective_function(conn, x_init[i])
         y_rows.append(y_row)
         append_observation_row(i + 1, 'sampling', y_row, x_init[i])
         send_json_line(conn, {"type": "tempCoverage", "value": float(i + 1) / float(max(1, initial_samples))})
+        y_so_far = np.vstack(y_rows)
+        hvs.append(float(compute_hypervolume(y_so_far, np.asarray(ref_point, dtype=np.float64))))
+        save_hypervolume_to_file(hvs, i + 1, ref_point)
+        send_json_line(conn, {"type": "coverage", "value": float(hvs[-1])})
 
     y_all = np.vstack(y_rows)
     x_all = np.asarray(x_init, dtype=np.float64)
     rewrite_pareto_flags(y_all)
     optimizer.observe(x_all, y_all)
 
-    hvs = [float(compute_hypervolume(y_all, np.asarray(ref_point, dtype=np.float64)))]
-    save_hypervolume_to_file(hvs, 0)
-    send_json_line(conn, {"type": "coverage", "value": float(hvs[-1])})
-
     # ---- optimization phase (ask/tell against openbo) ----
     for it in range(1, iterations + 1):
         t0 = time.time()
@@ -628,7 +634,7 @@ def meta_execute(conn, seed, iterations, initial_samples):
         append_observation_row(initial_samples + it, 'optimization', y_row, x_next[0])
         rewrite_pareto_flags(y_all)
         hvs.append(float(compute_hypervolume(y_all, np.asarray(ref_point, dtype=np.float64))))
-        save_hypervolume_to_file(hvs, it)
+        save_hypervolume_to_file(hvs, initial_samples + it, ref_point)
         send_json_line(conn, {"type": "coverage", "value": float(hvs[-1])})
 
     send_json_line(conn, {"type": "optimization_finished"})

Assets/StreamingAssets/BOData/BayesianOptimization/mobo.py

@@ -412,7 +412,7 @@ def objective_function(conn, x_tensor):
     return torch.tensor(fs, dtype=torch.double)
 
 # -------------------- data IO --------------------
-def generate_initial_data(conn, n_samples):
+def generate_initial_data(conn, n_samples, hv_util=None, hvs=None):
     global PROJECT_PATH
     if n_samples < 1:
         raise ValueError("n_samples must be >= 1 for non-warm-start runs.")
@@ -442,6 +442,12 @@ def generate_initial_data(conn, n_samples):
         with open(obs_csv, 'a', newline='') as f:
             csv.writer(f, delimiter=';').writerow(row)
         send_json_line(conn, {"type": "tempCoverage", "value": float(i+1)/float(max(1,n_samples))})
+        if hv_util is not None and hvs is not None:
+            y_so_far = torch.stack(train_obj, dim=0).to(dtype=torch.double)
+            volume = hv_util.compute(y_so_far[is_non_dominated(y_so_far)])
+            hvs.append(volume)
+            save_hypervolume_to_file(hvs, i + 1)
+            send_json_line(conn, {"type": "coverage", "value": float(volume)})
 
     Y = torch.stack(train_obj, dim=0).to(dtype=torch.double)
     # Ensure sampling-only runs (N_ITERATIONS=0) have globally-correct IsPareto flags.
@@ -616,8 +622,13 @@ def save_hypervolume_to_file(hvs, iteration):
     with open(hv_csv, 'a', newline='') as f:
         w = csv.writer(f, delimiter=';')
         if write_header:
-            w.writerow(["Hypervolume", "Run"])
-        w.writerow([hvs[-1], iteration])
+            w.writerow(["Hypervolume", "Run", "Scale", "ReferencePoint"])
+        w.writerow([
+            hvs[-1],
+            iteration,
+            "normalized maximize-space [-1,1] per objective",
+            "[" + ",".join(str(float(value)) for value in as_numpy_array(ref_point)) + "]",
+        ])
 
 # -------------------- main loop --------------------
 def mobo_execute(conn, seed, iterations, initial_samples):
@@ -638,7 +649,9 @@ def mobo_execute(conn, seed, iterations, initial_samples):
     if WARM_START:
         train_x, train_y = load_data()
     else:
-        train_x, train_y = generate_initial_data(conn, n_samples=initial_samples)
+        train_x, train_y = generate_initial_data(
+            conn, n_samples=initial_samples, hv_util=hv_util, hvs=hvs
+        )
 
     expected_x_dim = PROBLEM_DIM + (1 if CONTEXT_SETUP is not None else 0)
     if train_x.shape[0] != train_y.shape[0]:
@@ -650,10 +663,11 @@ def mobo_execute(conn, seed, iterations, initial_samples):
 
     mll, model = initialize_model(train_x, train_y)
 
-    volume = current_context_hypervolume(hv_util, train_x, train_y)
-    hvs.append(volume)
-    save_hypervolume_to_file(hvs, 0)
-    send_json_line(conn, {"type": "coverage", "value": float(volume)})
+    if WARM_START:
+        volume = current_context_hypervolume(hv_util, train_x, train_y)
+        hvs.append(volume)
+        save_hypervolume_to_file(hvs, 0)
+        send_json_line(conn, {"type": "coverage", "value": float(volume)})
 
     for it in range(1, iterations + 1):
         t0 = time.time()
@@ -673,7 +687,7 @@ def mobo_execute(conn, seed, iterations, initial_samples):
         volume = current_context_hypervolume(hv_util, train_x, train_y)
         hvs.append(volume)
         save_xy(train_x, train_y, it)
-        save_hypervolume_to_file(hvs, it)
+        save_hypervolume_to_file(hvs, initial_samples + it)
         send_json_line(conn, {"type": "coverage", "value": float(volume)})
         mll, model = initialize_model(train_x, train_y)
 

README.md

@@ -914,7 +914,7 @@ Common files:
 
 MOBO (`mobo.py`, `m >= 2`):
 * `ObservationsPerEvaluation.csv` uses `IsPareto`.
-* `HypervolumePerEvaluation.csv` stores hypervolume per iteration.
+* `HypervolumePerEvaluation.csv` stores hypervolume after every sampling and optimization evaluation. Its `Scale` column records that objectives are normalized to maximize-space `[-1,1]`, and `ReferencePoint` records the default `[-1,...,-1]` reference point.
 * Unity `coverage` corresponds to current hypervolume.
 
 Single-objective BO (`bo.py`, `m = 1`):

tests/test_meta_mobo_runtime.py

@@ -230,7 +230,7 @@ def test_full_protocol_run_with_sources(self):
             types = [m["type"] for m in sent]
             self.assertEqual(types.count("parameters"), 4)
             self.assertEqual(types.count("tempCoverage"), 2)
-            self.assertEqual(types.count("coverage"), 3)  # after sampling + 2 iterations
+            self.assertEqual(types.count("coverage"), 4)  # after every evaluation
             self.assertEqual(types.count("optimization_finished"), 1)
             self.assertEqual(types[-1], "optimization_finished")
 
@@ -270,8 +270,10 @@ def test_full_protocol_run_with_sources(self):
 
             with open(run_dir / "HypervolumePerEvaluation.csv", newline="") as f:
                 hv_rows = list(csv.reader(f, delimiter=";"))
-            self.assertEqual(hv_rows[0], ["Hypervolume", "Run"])
-            self.assertEqual([r[1] for r in hv_rows[1:]], ["0", "1", "2"])
+            self.assertEqual(hv_rows[0], ["Hypervolume", "Run", "Scale", "ReferencePoint"])
+            self.assertEqual([r[1] for r in hv_rows[1:]], ["1", "2", "3", "4"])
+            self.assertTrue(all(r[2] == "normalized maximize-space [-1,1] per objective" for r in hv_rows[1:]))
+            self.assertTrue(all(r[3] == "[-1.0,-1.0]" for r in hv_rows[1:]))
 
             with open(run_dir / "ExecutionTimes.csv", newline="") as f:
                 exec_rows = list(csv.reader(f, delimiter=";"))

tests/test_mobo.py

@@ -637,16 +637,17 @@ def test_save_hypervolume_to_file_writes_header_once(self):
         mobo = load_mobo_module()
         with tempfile.TemporaryDirectory() as tmp:
             mobo.PROJECT_PATH = tmp
+            mobo.ref_point = FakeTensor([-1.0])
             mobo.save_hypervolume_to_file([0.1], iteration=0)
             mobo.save_hypervolume_to_file([0.2], iteration=1)
             hv_csv = pathlib.Path(tmp) / "HypervolumePerEvaluation.csv"
             with hv_csv.open() as f:
                 rows = list(csv.reader(f, delimiter=";"))
 
-        self.assertEqual(rows[0], ["Hypervolume", "Run"])
+        self.assertEqual(rows[0], ["Hypervolume", "Run", "Scale", "ReferencePoint"])
         self.assertEqual(len(rows), 3)
-        self.assertEqual(rows[1], ["0.1", "0"])
-        self.assertEqual(rows[2], ["0.2", "1"])
+        self.assertEqual(rows[1], ["0.1", "0", "normalized maximize-space [-1,1] per objective", "[-1.0]"])
+        self.assertEqual(rows[2], ["0.2", "1", "normalized maximize-space [-1,1] per objective", "[-1.0]"])
 
     def test_mobo_execute_with_simulated_unity_objective_stream(self):
         mobo = load_mobo_module()
@@ -710,8 +711,8 @@ def test_mobo_execute_with_simulated_unity_objective_stream(self):
             np.array([[0.6, 0.6], [-0.6, -0.6], [0.0, 0.0]]),
             atol=1e-12,
         )
-        # Hypervolume logs: iteration 0 + iteration 1
-        self.assertEqual(len(hvs), 2)
+        # Hypervolume logs cover both sampling evaluations and the optimization step.
+        self.assertEqual(len(hvs), 3)
         # Ensure loop sent completion signal.
         self.assertTrue(any(m.get("type") == "optimization_finished" for m in out_msgs))