BeliefPropagation

---

⚠️ This package is heavily work in progress, some breaking changes should be expected.

---

This package implements a generic version of the Belief Propagation (BP) algorithm for the approximation of probability distributions factorized on a graph

$$\begin{equation} p(x_1,x_2,\ldots,x_n) \propto \prod_{a\in F} \psi_a(\underline{x}_a) \prod_{i\in V} \phi_i(x_i) \end{equation}$$

where $F$ is the set of factors, $V$ the set of variables, and $\underline{x}_a$ is the set of variables involved in factor $a$.

Installation

import Pkg; Pkg.add("BeliefPropagation")

Quickstart

Check out the examples folder.

Overview

The goal of this package is to provide a simple, flexible, and ready-to-use interface to the BP algorithm. It is enough for the user to provide the factor graph (encoded in an adjacency matrix or as a Graphs.jl graph) and the factors, everything else is taken care of.

At the same time, the idea is that refinements can be made to improve performance on a case-by-case basis. For example, messages are stored as Vectors by default, but when working with binary variables, one real number is enough, allowing for considerable speed-ups (see the Ising example). Also, a version of BP for continuous variables such as Gaussian BP can be introduced in the framework, although it is not yet implemented.

See also

• BeliefPropagation.jl: implements BP for the Ising model and the matching problem.
• FactorGraph.jl: implements Gaussian BP and other message-passing algorithms.
• ITensorNetworks.jl: implements BP as a technique for approximate tensor network contraction.
• ReactiveMP.jl: allows to solve Bayesian inference problems using message-passing.
• CodingTheory: has a specialized implementation of BP for the coding problem