|
|
|
Enabling Bayesian Inference for the Astronomy Masses
|
|
|
| The wealth of data being acquired by NASA missions promises detailed
tests of scientific theories. Bayesian analysis has the power to
efficiently combine information from these diverse sources, rigorously
establish confidence bounds on theoretical models, and provide
powerful probability-based methods for model selection and complex
hypothesis testing. This statistical approach is superior to those
commonly used in astronomy. However, it is not currently favored
owing to its computational difficulty. To remedy this deficiency, a
multi-disciplinary investigator team from the Departments of Astronomy
and Computer Science at UMass recently developed the Bayesian
Inference Engine (BIE). BIE's success is aided by advances in Markov
Chain Monte Carlo (MCMC) algorithms and the availability of commodity
parallel hardware. Most importantly, the system frees the scientist
from the database mechanics of the ongoing statistical investigation
without sacrificing rigor, while providing high-level interaction with
the inquiry.
To promote wider application and use, we propose key enhancements and
stand-alone applications to allow the astronomical community to take
immediate advantage of these statistical inference tools. First, we
will implement a persistence system. This allows the entire inference
to be stored in a research log to be later recalled and reused. A
persistence system enables what-if exploration, checkpointing, and
collaborative investigations. We will enhance the BIE with advanced
MCMC algorithms, non-parametric prior distributions and new graphical
tools. Second, we will develop three stand-alone "killer" applications
that will demonstrate the power of Bayesian inference as well as
provide an introduction to using BIE and Bayesian methods for more
specialized problems. The applications are: 1) Analysis of Structure
in Catalog Databases--a general set of star and galaxy counting
routines with I/O interfaces to SQL relational databases (e.g 2MASS
and SDSS); 2) Determining Galaxy Components from Images--a galaxy
classification tool based on the popular GIM2D package with a BIE
back-end that will allow sophisticated inference over image ensembles;
and 3) A Rigorous Statistical Basis for Semi-Analytic Models
(SAMS)--these widely used multiparameter phenomenological models for
galaxy formation make predictions for and use constraints from
multiple types of observations. The BIE back-end will enable
methodical exploration of the free parameters and allow sophisticated
hypothesis testing. |
|
|
