Fix/softmax out of bounds

CHANGELOG.md

@@ -50,6 +50,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Fix `simulation_export` function to handle DataFrame output correctly
 - Fix `detailed_pandas_compiler` function to support new numpy versions
 - Fix probability adjustements in `cpm.optimisation.minimise.LogLikelihood` method to ensure correct parameter estimates
+- Fix NaN handling in `cpm.models.decision.Softmax`, and `cpm.models.decision.Sigmoid` due to infinities in the exponential function for out-of-bounds parameters
 
 ### Changed
 

cpm/__version__.py

@@ -1 +1 @@
-__version__ = "0.24.7"
+__version__ = "0.24.8"

cpm/models/decision.py

@@ -55,7 +55,7 @@ class Softmax:
     >>> softmax.compute()
     array([0.30719589, 0.18632372, 0.50648039])
     >>> softmax.choice() # This will randomly choose one of the actions based on the computed probabilities.
-    2  
+    2
     >>> Softmax(temperature=temperature, activations=activations).compute()
     array([0.30719589, 0.18632372, 0.50648039])
     """
@@ -72,6 +72,15 @@ def __init__(self, temperature=None, xi=None, activations=None, **kwargs):
             self.activations = activations.copy()
         else:
             self.activations = np.zeros(1)
+        ## Throw error if activations contain missing values of infities
+        if np.isnan(self.activations).any():
+            raise ValueError(
+                "Activations contain NaN values. Please remove or impute missing values."
+            )
+        if np.isinf(self.activations).any():
+            raise ValueError(
+                "Activations contain infinite values. Please remove or impute infinite values."
+            )
         self.policies = np.zeros(self.activations.shape[0])
         self.shape = self.activations.shape
         if len(self.shape) > 1:
@@ -81,7 +90,6 @@ def __init__(self, temperature=None, xi=None, activations=None, **kwargs):
                 "Activations should be a 1D array, but a 2D array was provided. "
                 "Flattening the activations to a 1D array."
             )
-
 
         self.__run__ = False
 
@@ -97,6 +105,14 @@ def compute(self):
             np.exp(self.activations * self.temperature)
         )
         self.policies = output
+
+        if np.isnan(self.policies).any():
+            self.policies[np.isnan(self.policies)] = 1
+            self.policies /= self.policies.sum()
+            warnings.warn(
+                "NaN values found in policies. Replacing NaN values with 1 and normalising the policies to sum to 1."
+            )
+
         self.__run__ = True
         return self.policies
 
@@ -196,6 +212,7 @@ def __init__(self, temperature=None, activations=None, beta=0, **kwargs):
                 "Activations should be a 1D array, but a 2D array was provided. "
                 "Flattening the activations to a 1D array."
             )
+
         self.__run__ = False
 
     def compute(self):
@@ -207,11 +224,16 @@ def compute(self):
         output: ndarray
             A 2D array of outputs computed using the sigmoid function.
         """
-        output = 1 / (
-            1
-            + np.exp((self.activations - self.beta) * -self.temperature)
-        )
+        output = 1 / (1 + np.exp((self.activations - self.beta) * -self.temperature))
         self.policies = output
+
+        if np.isnan(self.policies).any():
+            self.policies[np.isnan(self.policies)] = 1
+            self.policies /= self.policies.sum()
+            warnings.warn(
+                "NaN values found in policies. Replacing NaN values with 1 and normalising the policies to sum to 1."
+            )
+
         self.__run__ = True
         return output
 

pyproject.toml

@@ -6,7 +6,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "cpm-toolbox"
-version = "0.24.7"
+version = "0.24.8"
 description = "A fundamental scientific toolbox for computational psychiatry and psychology."
 readme = "README.md"
 authors = [

test/applications/test_reinforcement_learning.py

@@ -117,7 +117,7 @@ def test_model_warnings(setup_data):
         assert len(w) > 0, "Warnings not raised"
         assert issubclass(w[-1].category, UserWarning), "Warning is not a UserWarning"
         assert (
-            "NaN in policy with parameters: 0.5, 1000, \nand with policy: [nan  0.]\n"
+            "NaN values found in policies. Replacing NaN values with 1 and normalising the policies to sum to 1."
             in str(w[-1].message)
         ), "Warning message mismatch"
 

test/models/test_decision.py

@@ -34,14 +34,28 @@ def test_softmax_choice():
     activations = np.array([0.1, 0, 0.2])
     softmax = Softmax(temperature=1, activations=activations)
     choice = softmax.choice()
-    assert choice in [0, 1, 2], "The Softmax.choice output is not in the expected range."
+    assert choice in [
+        0,
+        1,
+        2,
+    ], "The Softmax.choice output is not in the expected range."
+
 
 def test_softmax_input_shape():
     activations = np.array([[0.1, 0, 0.2], [0.3, 0.4, 0.5]])
     softmax = Softmax(temperature=1, activations=activations)
     assert len(softmax.shape) == 1, "The Softmax model should flatten 2D input arrays."
 
 
+def test_softmax_nan_handling():
+    activations = np.array([0.1, np.nan, 0.2])
+    softmax = Softmax(temperature=1, activations=activations)
+    policies = softmax.compute()
+    assert not np.isnan(
+        policies
+    ).any(), "The Softmax model should handle NaN values in activations."
+
+
 def test_sigmoid():
     expected = np.array([0.52497919, 0.5, 0.549834])
     activations = np.array([0.1, 0, 0.2])
@@ -58,7 +72,11 @@ def test_sigmoid_choice():
     activations = np.array([0.1, 0, 0.2])
     sigmoid = Sigmoid(temperature=1, activations=activations)
     choice = sigmoid.choice()
-    assert choice in [0, 1, 2], "The Sigmoid.choice output is not in the expected range."
+    assert choice in [
+        0,
+        1,
+        2,
+    ], "The Sigmoid.choice output is not in the expected range."
 
 
 def test_greedy_rule():