Handle zero-probability decision contexts

pycid/core/causal_bayesian_network.py

@@ -7,7 +7,7 @@
 import networkx as nx
 import numpy as np
 from pgmpy.factors.discrete import TabularCPD
-from pgmpy.inference.ExactInference import BeliefPropagation
+from pgmpy.inference.ExactInference import BeliefPropagation, VariableElimination
 from pgmpy.models import BayesianNetwork
 from pgmpy.sampling import BayesianModelSampling
 
@@ -194,11 +194,29 @@ def intervene(self, intervention: Dict[str, Outcome]) -> None:
             new_dist[intervention[variable]] = 1
             self.model[variable] = new_dist
 
+    def _context_has_zero_probability(
+        self, context: Dict[str, Outcome], intervention: Optional[Dict[str, Outcome]] = None
+    ) -> bool:
+        """Return whether the conditioning context has probability zero."""
+        if not context:
+            return False
+
+        cbn = self.copy()
+        if intervention:
+            cbn.intervene(intervention)
+
+        # Test event probability without conditioning on the event. Decision
+        # domains still provide valid CPDs for this diagnostic check.
+        factor = VariableElimination(cbn).query(list(context), show_progress=False)
+        index = tuple(factor.state_names[variable].index(context[variable]) for variable in factor.variables)
+        return bool(factor.values[index] == 0)
+
     def expected_value(
         self,
         variables: Iterable[str],
         context: Dict[str, Outcome],
         intervention: Optional[Dict[str, Outcome]] = None,
+        zero_prob_value: Optional[float] = None,
     ) -> List[float]:
         """Compute the expected value of a real-valued variable for a given context,
         under an optional intervention
@@ -210,8 +228,16 @@ def expected_value(
         context: Node values to condition upon. A dictionary mapping of node => value.
 
         intervention: Interventions to apply. A dictionary mapping node => value.
+
+        zero_prob_value: Value to return for each variable if context has zero probability.
         """
         factor = self.query(variables, context, intervention=intervention)
+        # Conditional expectations on zero-probability contexts are undefined.
+        # Confirm that cause directly, instead of treating every NaN as this case.
+        if np.isnan(factor.values).all() and self._context_has_zero_probability(context, intervention):
+            if zero_prob_value is not None:
+                return [zero_prob_value for _ in factor.variables]
+            raise RuntimeError(f"query {variables} | {context} conditions on a zero-probability event")
         factor.normalize()  # make probs add to one
 
         ev = np.array([0.0 for _ in factor.variables])

pycid/core/macid_base.py

@@ -394,7 +394,9 @@ def opt_policy(**parent_values: Outcome) -> Outcome:
             eu = {}
             for d in domain:
                 parent_values[decision] = d
-                eu[d] = sum(copy.expected_value(descendant_utility_nodes, parent_values))
+                # Unreachable decision contexts do not affect expected utility, so
+                # use a neutral fallback while constructing the decision rule.
+                eu[d] = sum(copy.expected_value(descendant_utility_nodes, parent_values, zero_prob_value=0.0))
             return max(eu, key=eu.get)  # type: ignore
 
         self.add_cpds(StochasticFunctionCPD(decision, opt_policy, self, domain=domain, label="opt"))