A simple modification for improving inference of non-linear dynamical systems

Abstract

Particle and ensemble filters are increasingly utilized for inference, optimization, and forecast; however, both filtering methods use discrete distributions to simulate continuous state space, a drawback that can lead to degraded performance for non-linear dynamical systems. Here we propose a simple modification, applicable to both particle and ensemble filters, that compensates for this problem. The method randomly replaces one or more model variables or parameters within a fraction of simulated trajectories at each filtering cycle. This modification, termed space re-probing, expands the state space covered by the filter through the introduction of outlying trajectories. We apply the space re-probing modification to three particle filters and three ensemble filters, and use these modified filters to model and forecast influenza epidemics. For both filter types, the space re-probing improves simulation of influenza epidemic curves and the prediction of influenza outbreak peak timing. Further, as fewer particles are needed for the particle filters, the proposed modification reduces the computational cost of these filters.