Struct TreeReweightedBP 

pub struct TreeReweightedBP {
    pub max_iterations: usize,
    pub tolerance: f64,
    pub edge_weights: HashMap<(String, String), f64>,
}

Tree-reweighted belief propagation (TRW-BP).

Uses a convex combination of spanning trees to provide an upper bound on the log partition function. More robust than standard BP for loopy graphs.

Messages are reweighted by edge appearance probabilities ρ_e ∈ [0,1].

Fields

max_iterations: usize

Maximum iterations

tolerance: f64

Convergence tolerance

edge_weights: HashMap<(String, String), f64>

Edge appearance probabilities (default: uniform)

Implementations

impl TreeReweightedBP

pub fn new(max_iterations: usize, tolerance: f64) -> Self

Create with custom parameters.

pub fn set_edge_weight(&mut self, var: String, factor: String, weight: f64)

Set edge appearance probability for a variable-factor edge.

pub fn initialize_uniform_weights(&mut self, graph: &FactorGraph)

Initialize uniform edge weights for all edges in graph.

pub fn run( &mut self, graph: &FactorGraph, ) -> Result<HashMap<String, ArrayD<f64>>>

Run tree-reweighted belief propagation.

Returns variable beliefs (marginals) and an upper bound on log Z.

pub fn compute_log_partition_upper_bound( &self, _graph: &FactorGraph, _beliefs: &HashMap<String, ArrayD<f64>>, ) -> Result<f64>

Compute upper bound on log partition function.

log Z ≤ log Z_TRW = Σ_i ρ_i log Z_i

Trait Implementations

impl Default for TreeReweightedBP

fn default() -> Self

Returns the “default value” for a type.