Source code for pgmpy.factors.discrete.NoisyOR

from itertools import product

import numpy as np

from pgmpy.factors.discrete import TabularCPD




class NoisyORCPD(TabularCPD):
    """
    Initializes the NoisyORCPD class.

    The NoisyOR CPD is a special case of TabularCPD for binary variables
    where a given variable can be activated by each of the parent variables
    with a specified probability. This activation probability is defined
    in the `prob_values` argument.

    Parameters
    ----------
    variable: str
        The variable for which the CPD is to be defined.

    prob_values: iterable
        A list of probabilities values for each `evidence` variable
        to activate `variable`.

    evidence: list
        List of evidence variables, i.e., conditional variables.

    Examples
    --------
    >>> from pgmpy.factors.discrete import NoisyORCPD
    >>> cpd = NoisyORCPD(variable="Y", prob_values=[0.3, 0.5], evidence=["X1", "X2"])
    >>> from pgmpy.models import DiscreteBayesianNetwork
    >>> model = DiscreteBayesianNetwork(
    ...     [("A", "B")]
    ... )  # With nodes with no parents, we can not define a NoisyORCPD.
    >>> cpd_a = TabularCPD(
    ...     variable="A",
    ...     variable_card=2,
    ...     values=[[0.2], [0.8]],
    ...     state_names={"A": ["True", "False"]},
    ... )
    >>> cpd_b = NoisyORCPD(variable="B", prob_values=[0.8], evidence=["A"])
    >>> model.add_cpds(cpd_a, cpd_b)
    """

    def __init__(self, variable, prob_values, evidence):
        if len(prob_values) != len(evidence):
            raise ValueError("Number of prob_values should be equal to number of evidence variables")

        self.prob_values = np.array(prob_values)
        if any(self.prob_values > 1) or any(self.prob_values < 0):
            raise ValueError("Values in prob_values should be between 0 and 1")

        inv_prob_values = 1 - self.prob_values
        tabular_values = np.zeros(2 ** len(evidence))
        state_comb = product([True, False], repeat=len(evidence))
        for i, states in enumerate(state_comb):
            tabular_values[i] = 1 - np.prod(inv_prob_values[np.array(states)])

        tabular_values = np.stack((tabular_values, 1 - tabular_values))

        state_names = {variable: ["True", "False"]}
        state_names.update({var: ["True", "False"] for var in evidence})

        super().__init__(
            variable=variable,
            variable_card=2,
            values=tabular_values,
            evidence=evidence,
            evidence_card=[2] * len(evidence),
            state_names=state_names,
        )