Alper Sarıtaş (Alper Saritas), also known as Mehmet Alper Sarıtaş (Mehmet Alper Saritas) and Mehmet Sarıtaş (Mehmet Saritas)

Fusion GMBP

Dec 2025-present

A self-contained software family implementing Gaussian Mixture Belief Propagation (GMBP) — exact and approximate inference over hybrid discrete-continuous probabilistic graphical models whose factors carry Gaussian mixtures rather than single distributions.

Four libraries compose into a complete learn-and-infer pipeline:

• hybn learns hybrid Bayesian networks (structure and parameters) from tabular data with discrete, continuous, and mixed variables.
• gmbp is the inference core, providing the mixture algebra (multiply, condition, marginalize) and the propagation schedules (exact on trees and forests, approximate on cyclic graphs), with a reduction hook at a defined point in the computation.
• gmred reduces Gaussian and hybrid mixtures under stated preservation objectives (modes, moments, cluster structure, linear response, tails), diagnosing the mixture, planning a method sequence, and measuring the error of the result.
• gmkit is an uncertainty-layer SDK that ties them together — an application declares variables and relations, gmkit learns a model, compiles it for inference, and advances beliefs with a streaming data feed while emitting events back to the application, carrying no domain vocabulary of its own so that domain data travels as opaque metadata.

The project is grounded in three papers (Fusion26 for the continuous foundation, SDF26 for hybrid factor graphs, MFI26 for neural-network mixture reduction).