clarkkitchen22's picture
Initial GeoBot Forecasting Framework commit
484e3bc
"""
Do-Calculus Module - Intervention Reasoning
Implements Pearl's do-calculus for counterfactual analysis and policy simulation.
Instead of just forecasting "what will happen," this module enables:
- "What if the U.S. sanctions X?"
- "What if China mobilizes?"
- "What if NATO deploys troops?"
- "What if an election is rigged?"
This is the foundation for counterfactual geopolitics.
"""
import numpy as np
import pandas as pd
from typing import Dict, List, Set, Optional, Tuple, Any
from ..models.causal_graph import CausalGraph, StructuralCausalModel
class DoCalculus:
"""
Implement Pearl's do-calculus for causal inference.
The do-calculus provides rules for transforming interventional
distributions into observational ones, enabling causal effect
estimation from observational data.
"""
def __init__(self, causal_graph: CausalGraph):
"""
Initialize do-calculus engine.
Parameters
----------
causal_graph : CausalGraph
Causal graph structure
"""
self.graph = causal_graph
def is_identifiable(
self,
treatment: str,
outcome: str,
confounders: Optional[Set[str]] = None
) -> bool:
"""
Check if causal effect is identifiable.
Parameters
----------
treatment : str
Treatment variable
outcome : str
Outcome variable
confounders : Set[str], optional
Known confounders
Returns
-------
bool
True if effect is identifiable
"""
# Basic check: are treatment and outcome d-separated after intervention?
# This is a simplified version
# Get all backdoor paths
backdoor_paths = self._get_backdoor_paths(treatment, outcome)
if len(backdoor_paths) == 0:
# No backdoor paths, effect is identifiable
return True
if confounders is not None:
# Check if confounders block all backdoor paths
return self._blocks_backdoor_paths(backdoor_paths, confounders)
return False
def _get_backdoor_paths(self, treatment: str, outcome: str) -> List[List[str]]:
"""
Get all backdoor paths from treatment to outcome.
A backdoor path is a path from treatment to outcome that
starts with an arrow into the treatment.
Parameters
----------
treatment : str
Treatment variable
outcome : str
Outcome variable
Returns
-------
List[List[str]]
List of backdoor paths
"""
import networkx as nx
backdoor_paths = []
# Get all simple paths from treatment to outcome
try:
all_paths = list(nx.all_simple_paths(
self.graph.graph.to_undirected(),
treatment,
outcome
))
except nx.NetworkXNoPath:
return []
# Filter for backdoor paths
for path in all_paths:
if len(path) > 2: # Must have intermediate nodes
# Check if first edge goes into treatment
second_node = path[1]
if self.graph.graph.has_edge(second_node, treatment):
backdoor_paths.append(path)
return backdoor_paths
def _blocks_backdoor_paths(
self,
paths: List[List[str]],
conditioning_set: Set[str]
) -> bool:
"""
Check if conditioning set blocks all backdoor paths.
Parameters
----------
paths : List[List[str]]
Backdoor paths
conditioning_set : Set[str]
Variables to condition on
Returns
-------
bool
True if all paths are blocked
"""
for path in paths:
if not self._is_path_blocked(path, conditioning_set):
return False
return True
def _is_path_blocked(self, path: List[str], conditioning_set: Set[str]) -> bool:
"""
Check if a path is blocked by conditioning set.
Parameters
----------
path : List[str]
Path to check
conditioning_set : Set[str]
Conditioning set
Returns
-------
bool
True if path is blocked
"""
# Simplified version: check if any non-collider in path is in conditioning set
for node in path[1:-1]: # Exclude endpoints
if node in conditioning_set:
# Check if it's a collider
idx = path.index(node)
prev_node = path[idx - 1]
next_node = path[idx + 1]
# It's a collider if both edges point to it
is_collider = (
self.graph.graph.has_edge(prev_node, node) and
self.graph.graph.has_edge(next_node, node)
)
if not is_collider:
return True
return False
def find_adjustment_set(
self,
treatment: str,
outcome: str,
method: str = 'backdoor'
) -> Set[str]:
"""
Find valid adjustment set for identifying causal effect.
Parameters
----------
treatment : str
Treatment variable
outcome : str
Outcome variable
method : str
Method to use ('backdoor', 'minimal')
Returns
-------
Set[str]
Valid adjustment set
"""
if method == 'backdoor':
return self._backdoor_adjustment_set(treatment, outcome)
elif method == 'minimal':
return self._minimal_adjustment_set(treatment, outcome)
else:
raise ValueError(f"Unknown method: {method}")
def _backdoor_adjustment_set(self, treatment: str, outcome: str) -> Set[str]:
"""
Find backdoor adjustment set.
Parameters
----------
treatment : str
Treatment variable
outcome : str
Outcome variable
Returns
-------
Set[str]
Backdoor adjustment set
"""
# Get all parents of treatment (excluding outcome's descendants)
parents = set(self.graph.get_parents(treatment))
# Remove outcome and its descendants
outcome_descendants = self.graph.get_descendants(outcome)
adjustment_set = parents - outcome_descendants - {outcome}
return adjustment_set
def _minimal_adjustment_set(self, treatment: str, outcome: str) -> Set[str]:
"""
Find minimal adjustment set.
Parameters
----------
treatment : str
Treatment variable
outcome : str
Outcome variable
Returns
-------
Set[str]
Minimal adjustment set
"""
# Start with backdoor set
backdoor_set = self._backdoor_adjustment_set(treatment, outcome)
# Try removing variables one by one
minimal_set = backdoor_set.copy()
for var in backdoor_set:
candidate_set = minimal_set - {var}
backdoor_paths = self._get_backdoor_paths(treatment, outcome)
if self._blocks_backdoor_paths(backdoor_paths, candidate_set):
minimal_set = candidate_set
return minimal_set
def compute_ate(
self,
data: pd.DataFrame,
treatment: str,
outcome: str,
adjustment_set: Optional[Set[str]] = None
) -> float:
"""
Compute Average Treatment Effect (ATE).
ATE = E[Y | do(X=1)] - E[Y | do(X=0)]
Parameters
----------
data : pd.DataFrame
Observational data
treatment : str
Treatment variable
outcome : str
Outcome variable
adjustment_set : Set[str], optional
Variables to adjust for
Returns
-------
float
Average Treatment Effect
"""
if adjustment_set is None:
adjustment_set = self.find_adjustment_set(treatment, outcome)
# Stratification estimator
if len(adjustment_set) == 0:
# No confounding
treated = data[data[treatment] == 1][outcome].mean()
control = data[data[treatment] == 0][outcome].mean()
return treated - control
# With adjustment
# Group by adjustment variables
adjustment_vars = list(adjustment_set)
ate = 0.0
for strata, group in data.groupby(adjustment_vars):
if len(group) > 0:
# Compute effect in this stratum
treated = group[group[treatment] == 1][outcome].mean()
control = group[group[treatment] == 0][outcome].mean()
if not np.isnan(treated) and not np.isnan(control):
strata_effect = treated - control
strata_weight = len(group) / len(data)
ate += strata_effect * strata_weight
return ate
class InterventionSimulator:
"""
Simulate policy interventions using structural causal models.
This class provides high-level interface for testing
"what if" scenarios in geopolitical contexts.
"""
def __init__(self, scm: StructuralCausalModel):
"""
Initialize intervention simulator.
Parameters
----------
scm : StructuralCausalModel
Structural causal model
"""
self.scm = scm
self.do_calculus = DoCalculus(scm.graph)
def simulate_intervention(
self,
intervention: Dict[str, float],
n_samples: int = 1000,
outcomes: Optional[List[str]] = None
) -> Dict[str, np.ndarray]:
"""
Simulate an intervention.
Parameters
----------
intervention : dict
Intervention specification {variable: value}
n_samples : int
Number of Monte Carlo samples
outcomes : List[str], optional
Outcome variables to track
Returns
-------
dict
Simulated outcomes
"""
# Sample from intervened distribution
samples = self.scm.sample(n_samples=n_samples, interventions=intervention)
if outcomes is not None:
samples = {k: v for k, v in samples.items() if k in outcomes}
return samples
def compare_interventions(
self,
interventions: List[Dict[str, float]],
outcome: str,
n_samples: int = 1000
) -> Dict[str, Dict[str, float]]:
"""
Compare multiple interventions.
Parameters
----------
interventions : List[dict]
List of interventions to compare
outcome : str
Outcome variable to compare
n_samples : int
Number of samples per intervention
Returns
-------
dict
Comparison results
"""
results = {}
for i, intervention in enumerate(interventions):
samples = self.simulate_intervention(intervention, n_samples, [outcome])
outcome_samples = samples[outcome]
results[f"intervention_{i}"] = {
'intervention': intervention,
'mean': np.mean(outcome_samples),
'std': np.std(outcome_samples),
'median': np.median(outcome_samples),
'q25': np.percentile(outcome_samples, 25),
'q75': np.percentile(outcome_samples, 75)
}
return results
def optimal_intervention(
self,
target_var: str,
intervention_vars: List[str],
intervention_ranges: Dict[str, Tuple[float, float]],
objective: str = 'maximize',
n_trials: int = 100,
n_samples: int = 1000
) -> Dict[str, Any]:
"""
Find optimal intervention to achieve target.
Parameters
----------
target_var : str
Target variable to optimize
intervention_vars : List[str]
Variables that can be intervened on
intervention_ranges : dict
Ranges for each intervention variable
objective : str
'maximize' or 'minimize'
n_trials : int
Number of random trials
n_samples : int
Samples per trial
Returns
-------
dict
Optimal intervention and results
"""
best_intervention = None
best_value = float('-inf') if objective == 'maximize' else float('inf')
for _ in range(n_trials):
# Sample random intervention
intervention = {}
for var in intervention_vars:
low, high = intervention_ranges[var]
intervention[var] = np.random.uniform(low, high)
# Simulate
samples = self.simulate_intervention(intervention, n_samples, [target_var])
mean_value = np.mean(samples[target_var])
# Update best
if objective == 'maximize':
if mean_value > best_value:
best_value = mean_value
best_intervention = intervention
else:
if mean_value < best_value:
best_value = mean_value
best_intervention = intervention
return {
'optimal_intervention': best_intervention,
'optimal_value': best_value,
'objective': objective
}
def counterfactual_analysis(
self,
observed: Dict[str, float],
intervention: Dict[str, float],
outcome: str
) -> Dict[str, float]:
"""
Perform counterfactual analysis.
"Given that we observed X, what would have happened if we had done Y?"
Parameters
----------
observed : dict
Observed values
intervention : dict
Counterfactual intervention
outcome : str
Outcome variable
Returns
-------
dict
Counterfactual results
"""
counterfactual = self.scm.compute_counterfactual(observed, intervention)
return {
'observed_outcome': observed.get(outcome, None),
'counterfactual_outcome': counterfactual.get(outcome, None),
'effect': counterfactual.get(outcome, 0) - observed.get(outcome, 0)
}