Learning Hybrid Bayesian Networks using Mixtures of Truncated Basis Functions

Learning, manipulation and evaluation of mixtures of truncated basis functions (MoTBFs), which include mixtures of polynomials (MOPs) and mixtures of truncated exponentials (MTEs). MoTBFs are a flexible framework for modelling hybrid Bayesian networks (I. Pérez-Bernabé, A. Salmerón, H. Langseth (2015) ; H. Langseth, T.D. Nielsen, I. Pérez-Bernabé, A. Salmerón (2014) ; I. Pérez-Bernabé, A. Fernández, R. Rumí, A. Salmerón (2016) ). The package provides functionality for learning univariate, multivariate and conditional densities, with the possibility of incorporating prior knowledge. Structural learning of hybrid Bayesian networks is also provided. A set of useful tools is provided, including plotting, printing and likelihood evaluation. This package makes use of S3 objects, with two new classes called 'motbf' and 'jointmotbf'.

MoTBFs

Development version of the MoTBFs package

The aim of the MoTBFs package is to provide a free and accessible implementation of algorithms for learning MoTBFs from data. The package implements learning algorithms for univariate, conditional, and joint MoTBF densities, explained in the Mixtures of Truncated Basis Functions vignette. By extension, functionality is also provided for learning MoTBF-based Bayesian networks by leveraging functionality from the bnlearn package. Furthermore, the MoTBFs package also provides methods for integrating prior domain knowledge in the learning process, thus also supporting data sparse domains. An extensive example included in the Workflow of MoTBFs pacakge vignette illustrates the main functionality of the package.

New functionalities are included to the MoTBFs package, including the implementation of the variable elimination algorithm for inference, and the adaptation of the Chow-Liu algorithm to MoTBF densities to learn the structure of TAN models.