Module bayesian_network_inference 

Bayesian Network Inference — variable elimination, belief propagation, and sampling-based inference over discrete Bayesian networks.

Overview

A BayesianNetwork is a directed acyclic graph (DAG) of discrete random variables where each variable has a ConditionalProbabilityTable (CPT) that encodes P(variable | parents). Given observed Evidence, the BayesianNetworkInference engine answers marginal-probability queries using one of three algorithms:

• Variable Elimination — exact inference by successive factor multiplication and marginalization.
• Belief Propagation — sum-product message passing (exact on trees, approximate on loopy graphs).
• Sampling — rejection / likelihood-weighting sampling with a reproducible xorshift64 PRNG (no rand dependency).

Example

use ipfrs_tensorlogic::bayesian_network_inference::{
    BayesianNetwork, BayesianNetworkInference, BniConfig, ConditionalProbabilityTable,
    EliminationOrder, Evidence, Factor, InferenceAlgorithm, InferenceQuery, RandomVariable,
};
use std::collections::HashMap;

// Build a tiny Rain → Wet-Grass network.
let rain = RandomVariable { id: "Rain".into(), states: vec!["T".into(), "F".into()], cardinality: 2 };
let wet  = RandomVariable { id: "Wet".into(),  states: vec!["T".into(), "F".into()], cardinality: 2 };

// P(Rain)
let f_rain = Factor {
    id: "f_rain".into(),
    variables: vec!["Rain".into()],
    values: vec![0.2, 0.8],
    shape: vec![2],
};
let cpt_rain = ConditionalProbabilityTable {
    variable: "Rain".into(),
    parents:  vec![],
    factor:   f_rain,
};

// P(Wet | Rain)  — rows: Rain=T, Rain=F; cols: Wet=T, Wet=F
let f_wet = Factor {
    id: "f_wet".into(),
    variables: vec!["Rain".into(), "Wet".into()],
    values: vec![0.9, 0.1, 0.2, 0.8],
    shape: vec![2, 2],
};
let cpt_wet = ConditionalProbabilityTable {
    variable: "Wet".into(),
    parents:  vec!["Rain".into()],
    factor:   f_wet,
};

let mut variables = HashMap::new();
variables.insert("Rain".into(), rain);
variables.insert("Wet".into(), wet);
let mut adjacency = HashMap::new();
adjacency.insert("Wet".into(), vec!["Rain".into()]);

let net = BayesianNetwork {
    variables,
    cpts: vec![cpt_rain, cpt_wet],
    adjacency,
};

let config = BniConfig::default();
let mut engine = BayesianNetworkInference::new(net, config).expect("example: should succeed in docs");

let query = InferenceQuery {
    query_variables: vec!["Rain".into()],
    evidence: vec![],
    algorithm: InferenceAlgorithm::VariableElimination,
};
let results = engine.query(&query).expect("example: should succeed in docs");
assert_eq!(results.len(), 1);
let rain_dist = &results[0].distribution;
assert!((rain_dist[0].1 - 0.2).abs() < 1e-9);

Structs

BayesianNetwork

A Bayesian network: a DAG of random variables with CPTs.

BayesianNetworkInference

Bayesian network inference engine.

BniConfig

Configuration for BayesianNetworkInference.

BniStats

Runtime statistics for the inference engine.

ConditionalProbabilityTable

A conditional probability table encoding P(variable | parents).

Evidence

A single piece of observed evidence.

Factor

A factor (un-normalised joint distribution table) over a set of variables.

InferenceQuery

A query to the inference engine.

QueryResult

Per-variable result of a Bayesian inference query.

RandomVariable

A discrete random variable with named states.

Enums

BniError

All errors that can arise during Bayesian network construction or inference.

EliminationOrder

Variable elimination ordering heuristic.

InferenceAlgorithm

Describes which inference algorithm to use.

Functions

bni_xorshift64

Xorshift64 PRNG — advance state and return next pseudo-random u64.