|
Abstract:
|
Spatiotemporal systems arising in epidemiology and ecology may be modeled as a collection of coupled, partially observed, nonlinear stochastic processes. To address the computational challenge of carrying out likelihood-based inference for such models, we developed a new sequential Monte Carlo (SMC) algorithm. SMC has formed the basis of successful inference methods for relatively low-dimensional nonlinear systems. However, SMC scales poorly with the spatial dimension, and is therefore said to suffer from a "curse of dimensionality." We introduce an SMC algorithm that does not formally beat the "curse" but nevertheless has some favorable properties. The plug-and-play property (the model is input to the algorithm as a simulator) is practically convenient. Theoretically, we write the Monte Carlo error as the sum of two components. One component has polynomial computational scaling while the other component is controlled using a target function. We demonstrate the methodology by analyzing coupled measles outbreaks for a
|
