Paul Fischer

Probabilistic Numerical Methods allow for the propagation of uncertainty across computational abstractions. This is potentially helpful for inference of latent forces driving dynamical systems from observed trajectories of the system. For example, inferring human and natural carbon sources and sinks from atmospheric remote sensing data, or inferring spatiotemporal changes in contact rates from observed infection events in epidemics and pandemics. The project aims to establish the ability of probabilistic numerical methods to solve such tasks for the case where latent forces are general nonparametric functions; to minimize the computational cost of doing so; and to optimize the numerical stability of this process.