Computational statistics using the Bayesian Inference Engine

TL;DR

The paper presents the Bayesian Inference Engine, a versatile software platform optimized for Bayesian analysis in astronomy, enabling efficient parameter inference, model comparison, and handling complex, high-dimensional data.

Contribution

It introduces the first dedicated platform for Bayesian computational statistics tailored to astronomical problems, with novel algorithms for marginal likelihood computation and robust sampling of multimodal posteriors.

Findings

Supports high-dimensional Bayesian inference with kernel density estimation.

Includes algorithms for computing marginal likelihood from posterior.

Facilitates collaborative and extensible Bayesian analysis in astronomy.

Abstract

This paper introduces the Bayesian Inference Engine (BIE), a general parallel, optimised software package for parameter inference and model selection. This package is motivated by the analysis needs of modern astronomical surveys and the need to organise and reuse expensive derived data. The BIE is the first platform for computational statistics designed explicitly to enable Bayesian update and model comparison for astronomical problems. Bayesian update is based on the representation of high-dimensional posterior distributions using metric-ball-tree based kernel density estimation. Among its algorithmic offerings, the BIE emphasises hybrid tempered MCMC schemes that robustly sample multimodal posterior distributions in high-dimensional parameter spaces. Moreover, the BIE is implements a full persistence or serialisation system that stores the full byte-level image of the running inference and previously characterised posterior distributions for later use. Two new algorithms to compute the marginal likelihood from the posterior distribution, developed for and implemented in the BIE, enable model comparison for complex models and data sets. Finally, the BIE was designed to be a collaborative platform for applying Bayesian methodology to astronomy. It includes an extensible object-oriented and easily extended framework that implements every aspect of the Bayesian inference. By providing a variety of statistical algorithms for all phases of the inference problem, a scientist may explore a variety of approaches with a single model and data implementation. Additional technical details and download details are available from http://www.astro.umass.edu/bie. The BIE is distributed under the GNU GPL.