Combining particle filters for Bayesian inference in multiple complex models
Abstract
Many complex real-world systems can be described through dynamical models that relate observed data to unknown quantities of interest. Fields where these systems can he found include biology, medicine, econometrics, computer science, artificial intelligence, astronomy. physics, chemistry, communications, earth science, among many others These models mathematically represent the connection between the dynamical hidden state of the system and the observed data> and the goal is usually in estimating unknowns, e.g., the evolving hidden state or future observations. In the Bayesian approach, the unknowns are characterized by a posterior probability distribution instead of a point-wise estimate, While there is an increasing interest in the Bayesian approach, the models of the aforementioned disciplines have also augmented their complexity. Unfortunately, the posterior is intractable in those realistic models, and it must be approximated. Sequential Monte Carlo (SMC) methods are the de facto computational tools for inference in dynamical models, approximating the posterior distributions with random samples. The increase of the available computational power has propelled the development of SMC. However, essential challenges remain open, limiting a much wider adoption. In this project, we will develop novel SMC methodologies able to work with several candidate models in parallel. We will investigate the self-assessments of the models in real-time during the inference process- Moreover, we will develop methods for an online identification of the cause of the malfunctioning automatically in simple models. This project constitutes a first keystone for a paradigmatic breakthrough in SMC methods, impacting and expanding its application to a wide range of applied fields.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 09, 2020
- Source ID
- W911NF2010126
Entities
People
- Victor Elvira-arregui
Organizations
- Army Contracting Command
- United States Army
- University of Edinburgh