e-lexicase will ignore cases where all models got np.inf as error

cavalab · Dec 26, 2023 · b1094c2 · b1094c2
1 parent e7de8a5
commit b1094c2
Show file tree

Hide file tree

Showing 2 changed files with 12 additions and 5 deletions.
diff --git a/src/brush/deap_api/utils.py b/src/brush/deap_api/utils.py
@@ -25,10 +25,14 @@ def e_lexicase(individuals, k):
 
         while len(cases) > 0 and len(candidates) > 1:
             errors_on_case = np.array([x.errors[cases[0]] for x in candidates])
-
+            
             MAD = np.median(np.abs(errors_on_case - np.median(errors_on_case)))
-
             best_on_case = np.min(errors_on_case)
+
+            # Skip if this case is np.inf for all individuals
+            if not np.isfinite(best_on_case+MAD):
+                continue
+
             candidates = [x for x in candidates
                           if x.errors[cases[0]] <= best_on_case + MAD]
 

diff --git a/src/brush/estimator.py b/src/brush/estimator.py
@@ -172,15 +172,18 @@ def _fitness_validation(self, ind, data: _brush.Dataset):
             "complexity": ind.prg.complexity()
         }
 
-        # Setting individual errors for lexicase
-        ind.errors = np.power( data.y - ind.prg.predict(data), 2 )
-
         return [ ind_objectives[obj] for obj in self.objectives ]
 
 
     def _fitness_function(self, ind, data: _brush.Dataset):
+        # fit the expression, then evaluate.
+
         ind.prg.fit(data)
 
+        # Setting individual errors for lexicase
+        ind.errors = np.nan_to_num(
+            np.abs( data.y - ind.prg.predict(data) ), nan=np.inf)
+
         return self._fitness_validation(ind, data)