Time and Place. Tuesday 14 June 2016, 12:15–13:15, ZIB Seminar Room 2006, Zuse Institute Berlin, Takustraße 7, 14195 Berlin
Abstract. In an ongoing push to construct probabilistic extensions of classic ODE solvers for application in statistics and machine learning, two recent papers have provided distinct methods that return probability measures instead of point estimates, based on sampling and filtering respectively. While both approaches leverage classical numerical analysis, by building on well-studied solutions of existing seminal solvers, the different constructions of probability measures strike a divergent balance between a formal quantification of epistemic uncertainty and a low computational overhead.
On the one hand, Conrad et al. proposed to randomise existing non-probabilistic one-step solvers by adding suitably scaled Gaussian noise after every step and thereby inducing a probability measure over the solution space of the ODE which contracts to a Dirac measure on the true unknown solution in the order of convergence of the underlying classic numerical method. But the computational cost of these methods is significantly above that of classic solvers.
On the other hand, Schober et al. recast the estimation of the solution as state estimation by a Gaussian (Kalman) filter and proved that employing a integrated Wiener process prior returns a posterior Gaussian process whose maximum likelihood (ML) estimate matches the solution of classic Runge–Kutta methods. In an attempt to amend this method's rough uncertainty calibration while sustaining its negligible cost overhead, we propose a novel way to quantify uncertainty in this filtering framework by probing the gradient using Bayesian quadrature.