 |
Center for Nonlinear Studies |
Sequential Monte Carlo algorithms enable computation of the likelihood function for general partially observed Markov process (POMP) models. A POMP model consists of a latent Markov process observed via a collection of noisy measurements. A collection of independent POMP models with some shared parameters is called a PanelPOMP model. A POMP model for which the Markov process has a tree-valued structure appropriate for disease transmission modeling, with measurements of both tree-related quantities (i.e., genetic sequences) and population quantities, is called a GenPOMP. A high-dimensional POMP model comprised of many coupled units is called a SpatPOMP since each unit may correspond to a spatial location. We discuss advances in the theory and practice of inference for POMP, PanelPOMP, GenPOMP and SpatPOMP models via sequential Monte Carlo. From a data analysis perspective, we demonstrate software for working with these model classes. From a methodological perspective, we discuss the principles underlying our algorithms.
Host: Ethan Obie Romero-Severson