Skip to main content

Infer.NET user guide : Running inference : Working with different inference algorithms

Expectation Propagation

Expectation Propagation (EP) is a deterministic approximate inference method. It is a generalization of loopy belief propagation and assumed density filtering. Infer.NET only implements EP with a fully factorised approximating family. This version of EP is essentially an approximation of loopy belief propagation, in which the messages are not arbitrary distributions but are projected onto tractable families (such as Gaussians) by matching expectations. In the case where the BP messages are always tractable (such as linear-Gaussian or discrete networks) then it is precisely loopy belief propagation.

EP has the following properties:

Further reading on EP:

EP and VMP are in fact both part of a more general class of algorithms referred to as Power EP which will be available in a later release, and which greatly increases the class of problems for which there are tractable solutions.