Variations on Bayesian Prediction and Inference
Abstract
The familiar Bayesian framework, where observed data is used to update prior information, via BayesÕs formula, has many desirable features. This project aims to address shortcomings of this Bayesian approach in two essential problems, namely, prediction and inference. First, for the prediction problem, the Monte Carlo computation required to obtain a genuine Bayesian predictive distribution can be too slow for use with streaming data, and a new recursive estimator of the Bayesian predictive distribution is proposed which is both fast to compute and has desirable theoretical properties. Second, for the inference problem, there are cases where a full probability model for all the unknowns is not available and/or is not desirable, so there is a need for "likelihood-free" Bayesian inference. New tools are developed to address various theoretical and computational questions related to the use of so-called Gibbs models for such problems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 12, 2017
- Source ID
- W911NF1510154
Entities
People
- Ryan Martin
Organizations
- Army Contracting Command
- United States Army
- University of Illinois at Chicago