State-space models and approximate probabilistic inference over networks
Abstract
In this project, we focus on dynamical systems that evolve over time. Several agents acquire (potentially overlapped) observations, that are processed in a distributed manner for estimation, classification, forecasting, and other key statistical tasks that are needed for a fast response data-driven decision-making process. We consider a Bayesian/probabilistic framework, allowing for quantifying the uncertainty, which is key for optimal real-time decisions in evolving scenarios. Unfortunately, Bayesian inference in most realistic models require particle (Monte Carlo) approximations as a proxy of the true probability densities. These approximate methods are known to be asymptotically exact with the number of particles, N, but their computational cost is large and grows to infinity with N. In this project, we focus on the problem where several agents potentially share some of the computational resources, and an allocation of those resources must be performed in order to optimize a given performance metric (e.g., minimizing the MSE in the prediction of future observations). We start by studying the static framework where all observations are available from the beginning, extracting useful insights for the more challenging dynamic setting. Then, we target the ambitious problem of compressing information in particle-based methods, which is known to be an unsolved problem that needs to be tackled for distributed processing in networks. Finally, we will propose novel methodology for the combination of particle approximations, either in the static or in the dynamic framework. Again, this problem has been under-looked in the literature due to its difficulty while it is fundamental for a unified decision-making process that gathers the responses from several intelligent agents.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 08, 2022
- Source ID
- W911NF2210235
Entities
People
- Victor Elvira
Organizations
- Army Contracting Command
- United States Army
- University of Edinburgh