bugfix for NParEGO: default scalarization_weights is None (random wei…

…ghts); enable user input scalarization_weights (previous missing in _parse_aq_kwargs
MSDLLCpapers · Aug 19, 2024 · fcf8220 · fcf8220
1 parent d2df76d
commit fcf8220
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 2 deletions.
diff --git a/obsidian/acquisition/config.py b/obsidian/acquisition/config.py
@@ -44,5 +44,5 @@
     'NIPV': {},
     'EHVI': {'ref_point': {'val': None, 'optional': True}},
     'NEHVI': {'ref_point': {'val': None, 'optional': True}},
-    'NParEGO': {'scalarization_weights': {'val': [1], 'optional': True}},
+    'NParEGO': {'scalarization_weights': {'val': None, 'optional': True}},
 }
diff --git a/obsidian/optimizer/bayesian.py b/obsidian/optimizer/bayesian.py
@@ -459,7 +459,7 @@ def _validate_hypers(self,
             if hps.get(key) is None:
                 if not defaults['optional']:
                     raise ValueError(f'Must specify hyperpameter value {key} for {aq_str}')
-                if key in ['scalarization_weights', 'weights']:
+                if key in ['weights']: #['scalarization_weights', 'weights']:
                     aq_hps[key] = defaults['val'] * o_dim
                 else:
                     aq_hps[key] = defaults['val']
@@ -551,6 +551,13 @@ def _parse_aq_kwargs(self,
             qmc_samples = draw_sobol_samples(bounds=X_bounds, n=128, q=m_batch)
             aq_kwargs['mc_points'] = qmc_samples.squeeze(-2)
 
+        if aq == 'NParEGO':
+            w = hps['scalarization_weights']
+            if isinstance(w,list):
+                w = torch.tensor(w)
+                w = w/torch.sum(torch.abs(w))
+            aq_kwargs['scalarization_weights'] = w
+
         return aq_kwargs
 
     def suggest(self,