compiled-knowledge 4.0.0a20__cp312-cp312-macosx_10_13_x86_64.whl

ck/probability/probability_space.py

@@ -0,0 +1,622 @@
+"""
+An abstract class for object providing probabilities.
+"""
+import math
+from abc import ABC, abstractmethod
+from itertools import chain
+from typing import Sequence, Tuple, Iterable, Callable
+
+import numpy as np
+
+from ck.pgm import Indicator, RandomVariable, rv_instances_as_indicators, number_of_states, rv_instances, Instance
+from ck.utils.iter_extras import combos as _combos
+from ck.utils.map_set import MapSet
+from ck.utils.np_extras import dtype_for_number_of_states, NDArrayFloat64, DTypeStates, NDArrayNumeric
+
+# Type defining a condition.
+Condition = None | Indicator | Iterable[Indicator]
+
+
+class ProbabilitySpace(ABC):
+    """
+    An abstract mixin class for a class providing probabilities over a state space defined by random variables.
+    Each possible world of the state space is referred to as an 'instance'.
+    """
+    __slots__ = ()
+
+    @property
+    @abstractmethod
+    def rvs(self) -> Sequence[RandomVariable]:
+        """
+        Return the random variables that define the state space.
+        Each random variable, rv, has a length len(rv) which
+        is the number of states, and rv[i] is the 'indicator' for
+        the ith state of the random variable. Indicators must
+        be unique across all rvs as rv[i] indicates the
+        condition 'rv == i'.
+        """
+
+    @abstractmethod
+    def wmc(self, *condition: Condition) -> float:
+        """
+        Return the weight of instances matching the given condition.
+
+        If multiple indicators of the same random variable appear in
+        the parameter 'indicators' then they are interpreted as
+        a disjunction, otherwise indicators are interpreted as
+        a conjunction. E.g.:  X=0, Y=1, Y=3  means X=0 and (Y=1 or Y=3)
+
+        Args:
+            condition: zero or more indicators that specify a condition.
+        """
+
+    @property
+    @abstractmethod
+    def z(self) -> float:
+        """
+        Return the summed weight of all instances.
+        This is equivalent to self.wmc(), with no arguments.
+        """
+
+    def probability(self, *indicators: Indicator, condition: Condition = ()) -> float:
+        """
+        Return the joint probability of the given indicators,
+        conditioned on any conditions, and
+        marginalised over any unmentioned random variables.
+
+        If multiple indicators of the same random variable appear in
+        the parameters 'indicators' or 'condition' then they are interpreted as
+        a disjunction, otherwise indicators are interpreted as
+        a conjunction. E.g.:  X=0, Y=1, Y=3  means X=0 and (Y=1 or Y=3).
+
+        Args:
+            indicators: Indicators that specify which set of instances to compute probability.
+            condition: Indicators that specify conditions for a conditional probability.
+        Returns:
+            the probability of the given indicators, conditioned on the given conditions.
+        """
+        condition: Tuple[Indicator, ...] = check_condition(condition)
+
+        if len(condition) == 0:
+            z = self.z
+            if z <= 0:
+                return np.nan
+        else:
+            z = self.wmc(*condition)
+            if z <= 0:
+                return np.nan
+
+            # Combine the indicators with the condition
+            # If a variable is mentioned in both the indicators and condition, then
+            # we need to take the intersection, and check for contradictions.
+            # If a variable is mentioned in the condition but not indicators, then
+            # the rv condition needs to be added to the indicators.
+            indicator_groups: MapSet[int, Indicator] = _group_indicators(indicators)
+            condition_groups: MapSet[int, Indicator] = _group_indicators(condition)
+
+            for rv_idx, indicators in condition_groups.items():
+                indicator_group = indicator_groups.get(rv_idx)
+                if indicator_group is None:
+                    indicator_groups.add_all(rv_idx, indicators)
+                else:
+                    indicator_group.intersection_update(indicators)
+                    if len(indicator_group) == 0:
+                        # A contradiction between the indicators and conditions
+                        return 0.0
+
+            # Collect all the indicators from the updated indicator_groups
+            indicators = chain(*indicator_groups.values())
+
+        return self.wmc(*indicators) / z
+
+    def marginal_distribution(self, *rvs: RandomVariable, condition: Condition = ()) -> NDArrayNumeric:
+        """
+        What is the marginal probability distribution over the states of the given random variables.
+        Assumes that no indicators of rv in rvs appear in the conditions (if supplied).
+
+        When multiple rvs are supplied, the order of instantiations is as per
+        `rv_instances_as_indicators(*rvs)`.
+
+        If multiple indicators of the same random variable appear in
+        the parameter 'condition' then they are interpreted as
+        a disjunction, otherwise indicators are interpreted as
+        a conjunction. E.g.:  X=0, Y=1, Y=3  means X=0 and (Y=1 or Y=3).
+
+        This is not an efficient implementation as it will call self.probability(...)
+        for each possible state of the given random variable. If efficient marginal
+        probability calculations are required, consider using a different method.
+
+        Warning:
+            If the probability of each state of rv (given the condition) is
+            zero, then the marginal distribution is il-defined and the returned probabilities will
+            all be NAN.
+
+        Args:
+            rvs: Random variables to compute the marginal distribution over.
+            condition: Indicators that specify conditions for conditional probability.
+            
+        Returns:
+            marginal probability distribution as an array co-indexed with `rv_instances_as_indicators(*rvs)`.
+        """
+        condition = check_condition(condition)
+
+        # We have to be careful of the situation where indicators of rvs appear in condition.
+        # If an RV has at least 1 indicator in condition then it must match it to have non-zero probability.
+        wmc = self._get_wmc_for_marginals(rvs, condition)
+
+        result: NDArrayFloat64 = np.fromiter(
+            (wmc(indicators) for indicators in rv_instances_as_indicators(*rvs)),
+            count=number_of_states(*rvs),
+            dtype=np.float64
+        )
+        _normalise_marginal(result)
+        return result
+
+    def map(self, *rvs: RandomVariable, condition: Condition = ()) -> Tuple[float, Instance]:
+        """
+        Determine the maximum apriori probability (MAP).
+
+        If there are tied solutions, one solution is returned, which
+        is selected arbitrarily.
+
+        If multiple indicators of the same random variable appear in
+        the parameter 'condition' then they are interpreted as
+        a disjunction, otherwise indicators are interpreted as
+        a conjunction. E.g.: X=0, Y=1, Y=3  means X=0 and (Y=1 or Y=3)
+
+        Warning:
+            This is not an efficient implementation as it will call `self.wmc`
+            for each possible state of the given random variables. If efficient MAP
+            probability calculations are required, consider using a different method.
+
+        Args:
+            rvs: random variables to find the MAP over.
+            condition: any conditioning indicators.
+
+        Returns:
+            (probability, instance) where
+            probability: is the MAP probability
+            instance: is the MAP state (co-indexed with the given rvs).
+        """
+        condition: Sequence[Indicator] = check_condition(condition)
+
+        rv_indexes = set(rv.idx for rv in rvs)
+        assert len(rv_indexes) == len(rvs), 'duplicated random variables not allowed'
+
+        # Group conditioning indicators by random variable.
+        conditions_by_rvs = _group_states(condition)
+
+        # See if any MAP random variable is also conditioned.
+        # Reduce the state space of any conditioned MAP rv.
+        loop_rvs = []
+        reduced_space = False
+        for rv in rvs:
+            states = conditions_by_rvs.get(rv.idx)
+            if states is None:
+                loop_rvs.append(rv)
+            else:
+                loop_rvs.append([rv[i] for i in sorted(states)])
+                reduced_space = True
+
+        # If the random variables we are looping over does not have any conditions
+        # then it is expected to be faster by using computed marginal probabilities.
+        if not reduced_space:
+            prs = self.marginal_distribution(*rvs, condition=condition)
+            best_probability = float('-inf')
+            best_states = None
+            for probability, inst in zip(prs, rv_instances(*rvs)):
+                if probability > best_probability:
+                    best_probability = probability
+                    best_states = inst
+            return best_probability, best_states
+
+        else:
+            # Remove any condition indicators with rv in rvs.
+            new_conditions = tuple(ind for ind in condition if ind.rv_idx not in rv_indexes)
+
+            # Loop over the state space of the 'loop' rvs
+            best_probability = float('-inf')
+            best_states = None
+            indicators: Tuple[Indicator, ...]
+            for indicators in _combos(loop_rvs):
+                probability = self.wmc(*(indicators + new_conditions))
+                if probability > best_probability:
+                    best_probability = probability
+                    best_states = tuple(ind.state_idx for ind in indicators)
+            condition_probability = self.wmc(*condition)
+            return best_probability / condition_probability, best_states
+
+    def correlation(self, indicator1: Indicator, indicator2: Indicator, condition: Condition = ()) -> float:
+        """
+        What is the correlation between the two given indicators, r(indicator1, indicator2).
+        
+        Args:
+            indicator1: a first random variable and its state.
+            indicator2: a second random variable and its state.
+            condition: any conditioning indicators.
+            
+        Returns:
+            correlation between the two given indicators.
+        """
+        condition = check_condition(condition)
+
+        p1 = self.probability(indicator1, condition=condition)
+        p2 = self.probability(indicator2, condition=condition)
+        p12 = self._joint_probability(indicator1, indicator2, condition=condition)
+        d = p1 * (1.0 - p1) * p2 * (1.0 - p2)
+        if d == 0.0:
+            # As any marginal probability approaches zero, correlation approaches zero
+            return 0.0
+        else:
+            return (p12 - p1 * p2) / math.sqrt(d)
+
+    def entropy(self, rv: RandomVariable, condition: Condition = ()) -> float:
+        """
+        Calculate the entropy of the given random variable, H(rv).
+        
+        Args:
+            rv: random variable to calculate the entropy for.
+            condition: any conditioning indicators.
+            
+        Returns:
+            entropy of the given random variable.
+        """
+        condition = check_condition(condition)
+        e = 0.0
+        for ind in rv:
+            p = self.probability(ind, condition=condition)
+            if p > 0.0:
+                e -= p * math.log2(p)
+        return e
+
+    def conditional_entropy(self, rv1: RandomVariable, rv2: RandomVariable, condition: Condition = ()) -> float:
+        """
+        Calculate the conditional entropy, H(rv1 | rv2).
+        
+        Args:
+            rv1: random variable to calculate the entropy for.
+            rv2: the conditioning random variable for entropy calculation.
+            condition: any conditioning indicators to restrict the state space.
+            
+        Returns:
+            entropy of rv1, conditioned on rv2.
+        """
+        condition = check_condition(condition)
+        e = 0.0
+        for ind1 in rv1:
+            for ind2 in rv2:
+                p = self._joint_probability(ind1, ind2, condition=condition)
+                if p > 0.0:
+                    # if p > 0 then p2 > 0, as p <= p2
+                    p2 = self.probability(ind2, condition=condition)
+                    e -= p * math.log2(p / p2)
+        return e
+
+    def joint_entropy(self, rv1: RandomVariable, rv2: RandomVariable, condition: Condition = ()) -> float:
+        """
+        Calculate the joint entropy of the two random variables, H(rv1; rv2).
+        
+        Args:
+            rv1: a first random variable to calculate joint entropy.
+            rv2: a second random variable to calculate joint entropy.
+            condition: any conditioning indicators to restrict the state space.
+        Returns:
+            joint entropy of the given random variables.
+        """
+        condition = check_condition(condition)
+        e = 0.0
+        for ind1 in rv1:
+            for ind2 in rv2:
+                p = self._joint_probability(ind1, ind2, condition=condition)
+                if p > 0.0:
+                    e -= p * math.log2(p)
+        return e
+
+    def mutual_information(self, rv1: RandomVariable, rv2: RandomVariable, condition: Condition = ()) -> float:
+        """
+        Calculate the mutual information between two random variables, I(rv1; rv2).
+
+        Args:
+            rv1: a first random variable
+            rv2: a second random variable
+            condition: indicators to specify a condition restricting the state space.
+        Returns:
+            mutual_information(rv1, rv2) / denominator
+        """
+        condition = check_condition(condition)
+        p1s = self.marginal_distribution(rv1, condition=condition)
+        p2s = self.marginal_distribution(rv2, condition=condition)
+        info = 0.0
+        for ind1, p1 in zip(rv1, p1s):
+            for ind2, p2 in zip(rv2, p2s):
+                p12 = self._joint_probability(ind1, ind2, condition=condition)
+                if p12 > 0.0:
+                    info += p12 * math.log2(p12 / p1 / p2)
+        return info
+
+    def total_correlation(self, rv1: RandomVariable, rv2: RandomVariable, condition: Condition = ()) -> float:
+        """
+        Calculate the 'total correlation' measure.
+        total_correlation = I(rv1; rv2) / min(H(rv1), H(rv2)).
+        This is a normalised mutual information between two random variables.
+        0 => no mutual information.
+        1 => perfect mutual information.
+
+        Args:
+            rv1: a first random variable
+            rv2: a second random variable
+            condition: indicators to specify a condition restricting the state space.
+        Returns:
+            total correlation between the given random variables.
+        """
+        condition = check_condition(condition)
+        denominator = min(self.entropy(rv1), self.entropy(rv2, condition=condition))
+        return self._normalised_mutual_information(rv1, rv2, denominator, condition=condition)
+
+    def uncertainty(self, rv1: RandomVariable, rv2: RandomVariable, condition: Condition = ()) -> float:
+        """
+        Calculate the 'uncertainty' measure, C, between two random variables
+        C(rv1, rv2) = I(rv1; rv2) / H(rv2)
+        This is a normalised mutual information between two random variables.
+        Note that it is not a symmetric measure; in general C(rv1, rv2) does not equal C(rv2, rv1).
+        0 => no mutual information.
+        1 => perfect mutual information.
+
+        Args:
+            rv1: a first random variable
+            rv2: a second random variable
+            condition: indicators to specify a condition restricting the state space.
+        Returns:
+            uncertainty between the given random variables.
+        """
+        condition = check_condition(condition)
+        denominator = self.entropy(rv2, condition=condition)
+        return self._normalised_mutual_information(rv1, rv2, denominator, condition=condition)
+
+    def symmetric_uncertainty(self, rv1: RandomVariable, rv2: RandomVariable, condition: Condition = ()) -> float:
+        """
+        Calculate the 'symmetric uncertainty' measure.
+        symmetric_uncertainty = 2 * I(rv1, rv2) / (H(rv1) + H(rv2)).
+        This is the harmonic mean of the two uncertainty coefficients,
+        C(rv1, rv2) = I(rv1; rv2) / H(rv2) and C(rv2, rv1) = I(rv1; rv2) / H(rv1).
+        This is a normalised mutual information between two random variables.
+        0 => no mutual information.
+        1 => perfect mutual information.
+
+        Args:
+            rv1: a first random variable
+            rv2: a second random variable
+            condition: indicators to specify a condition restricting the state space.
+        Returns:
+            symmetric uncertainty between the given random variables.
+        """
+        condition = check_condition(condition)
+        denominator = self.entropy(rv1) + self.entropy(rv2, condition=condition)
+        return 2.0 * self._normalised_mutual_information(rv1, rv2, denominator, condition=condition)
+
+    def iqr(self, rv1: RandomVariable, rv2: RandomVariable, condition: Condition = ()) -> float:
+        """
+        Calculate the Information Quality Ratio (IQR).
+        IQR = I(rv1; rv2) / H(rv1; rv2).
+        Also known as 'dual total correlation'.
+        This is a normalised mutual information between two random variables.
+        0 => no mutual information.
+        1 => perfect mutual information.
+
+        Args:
+            rv1: a first random variable
+            rv2: a second random variable
+            condition: indicators to specify a condition restricting the state space.
+        Returns:
+            Information Quality Ratio between the given random variables.
+        """
+        condition = check_condition(condition)
+        denominator = self.joint_entropy(rv1, rv2, condition=condition)
+        return self._normalised_mutual_information(rv1, rv2, denominator, condition=condition)
+
+    def covariant_normalised_mutual_information(self, rv1: RandomVariable, rv2: RandomVariable,
+                                                condition: Condition = ()) -> float:
+        """
+        Calculate the covariant normalised mutual information
+        = I(rv1; rv2) / sqrt(H(rv1) * H(rv2)).
+        This is a normalised mutual information between two random variables.
+        0 => no mutual information.
+        1 => perfect mutual information.
+
+        Args:
+            rv1: a first random variable
+            rv2: a second random variable
+            condition: indicators to specify a condition restricting the state space.
+        Returns:
+            covariant normalised mutual information between the given random variables.
+        """
+        condition = check_condition(condition)
+        denominator = math.sqrt(self.entropy(rv1, condition=condition) * self.entropy(rv2, condition=condition))
+        return self._normalised_mutual_information(rv1, rv2, denominator, condition=condition)
+
+    def _normalised_mutual_information(
+            self,
+            rv1: RandomVariable,
+            rv2: RandomVariable,
+            denominator: float,
+            condition: Tuple[Indicator, ...],
+    ) -> float:
+        """
+        Helper function for normalised mutual information calculations.
+
+        Args:
+            rv1: a first random variable
+            rv2: a second random variable
+            denominator: the normalisation factor
+            condition: indicators to specify a condition restricting the state space.
+        Returns:
+            mutual_information(rv1, rv2) / denominator
+        """
+        if denominator == 0.0:
+            return 0.0
+        else:
+            return self.mutual_information(rv1, rv2, condition) / denominator
+
+    def _joint_probability(
+            self,
+            indicator1: Indicator,
+            indicator2: Indicator,
+            condition: Tuple[Indicator, ...],
+    ) -> float:
+        """
+        Helper function to correctly calculate a joint probability even if the two indicators
+        are from the same random variable.
+
+        If the indicators are from the different random variables then
+        probability(indicator1 and indicator2 | condition).
+
+        If the indicators are from the same random variable then
+        probability(indicator1 or indicator2 | condition).
+
+        Args:
+            indicator1: a first Indicator.
+            indicator2: a second Indicator
+            condition: indicators to specify a condition restricting the state space.
+        Returns:
+            joint probability of the two indicators, given the condition.
+        """
+        if indicator1 == indicator2:
+            # Ensure correct behaviour, same random variable and same states
+            return self.probability(indicator1, condition=condition)
+        elif indicator1.rv_idx == indicator2.rv_idx:
+            # Efficiency shortcut, same random variable but different states
+            return 0.0
+        else:
+            # General case, two different random variables
+            return self.probability(indicator1, indicator2, condition=condition)
+
+    def _get_wmc_for_marginals(
+            self,
+            rvs: Sequence[RandomVariable],
+            condition: Tuple[Indicator, ...],
+    ) -> Callable[[Sequence[Indicator]], float]:
+        """
+        Return a wmc function that is suitable for calculating marginal distributions.
+
+        This implementation is careful of the situation where indicators of rvs appear in condition.
+        If an RV has at least 1 indicator in condition then it must match it to have non-zero probability.
+
+        Args:
+            rvs: random variables to calculate marginal distributions for.
+            condition: indicators to specify a condition restricting the state space.
+        Returns:
+            A function from a condition, specified as a sequence of indicators, to a weighted model count.
+        """
+        if len(condition) > 0:
+            check_sets = []
+            overlap_detected = False
+            cond_set = set(condition)
+            for rv in rvs:
+                in_condition = set()
+                for ind in rv:
+                    if ind in cond_set:
+                        in_condition.add(ind)
+                        cond_set.discard(ind)
+                        overlap_detected = True
+                if len(in_condition) == 0:
+                    in_condition.update(rv)
+                check_sets.append(in_condition)
+
+            if overlap_detected:
+                __wmc__condition = tuple(cond_set)
+
+                def wmc(indicators: Sequence[Indicator]) -> float:
+                    for indicator, check_set in zip(indicators, check_sets):
+                        if indicator not in check_set:
+                            return 0.0
+                    full_condition = tuple(indicators) + __wmc__condition
+                    return self.wmc(*full_condition)
+            else:
+                __wmc__condition = tuple(condition)
+
+                def wmc(indicators: Sequence[Indicator]) -> float:
+                    full_condition = tuple(indicators) + __wmc__condition
+                    return self.wmc(*full_condition)
+        else:
+            def wmc(indicators: Sequence[Indicator]) -> float:
+                return self.wmc(*indicators)
+
+        return wmc
+
+
+def check_condition(condition: Condition) -> Tuple[Indicator, ...]:
+    """
+    Make the best effort to interpret the given condition.
+
+    Args:
+        condition: a relaxed specification of a condition.
+    Returns:
+        a formal specification of the condition as a tuple of indicators with no duplicates.
+    """
+    if condition is None:
+        return ()
+    elif isinstance(condition, Indicator):
+        return (condition,)
+    else:
+        return tuple(set(condition))
+
+
+def dtype_for_state_indexes(rvs: Iterable[RandomVariable]) -> DTypeStates:
+    """
+    Infer a numpy dtype to hold any state index from any given random variable.
+
+    Args:
+        rvs: some random variables.
+    Returns:
+        a numpy dtype.
+    """
+    return dtype_for_number_of_states(max((len(rv) for rv in rvs), default=0))
+
+
+def _group_indicators(indicators: Iterable[Indicator]) -> MapSet[int, Indicator]:
+    """
+    Group the given indicators by rv_idx.
+
+    Args:
+        indicators: the indicators to group.
+
+    Returns:
+        A mapping from rv_idx to set of indicators.
+    """
+    groups: MapSet[int, Indicator] = MapSet()
+    for indicator in indicators:
+        groups.add(indicator.rv_idx, indicator)
+    return groups
+
+
+def _group_states(indicators: Iterable[Indicator]) -> MapSet[int, int]:
+    """
+    Group the given indicator states by rv_idx.
+
+    Args:
+        indicators: the indicators to group.
+
+    Returns:
+        A mapping from rv_idx to set of state indexes.
+    """
+    groups: MapSet[int, int] = MapSet()
+    for indicator in indicators:
+        groups.add(indicator.rv_idx, indicator.state_idx)
+    return groups
+
+
+def _normalise_marginal(distribution: NDArrayFloat64) -> None:
+    """
+    Update the values in the given distribution to
+    properly represent a marginal distribution.
+
+    The update is made in-place.
+
+    Args:
+        a 1D numpy array of likelihoods.
+    """
+    total = np.sum(distribution)
+    if total <= 0:
+        distribution[:] = np.nan
+    elif total != 1:
+        distribution /= total

ck/program/__init__.py

@@ -0,0 +1,3 @@
+from .raw_program import RawProgram
+from .program_buffer import ProgramBuffer
+from .program import Program