This thesis is concerned with recursive Bayesian estimation of non-linear dynamical
systems, which can be modeled as discretely observed stochastic differential
equations. The recursive real-time estimation algorithms for these continuous-
discrete filtering problems are traditionally called optimal filters and the algorithms
for recursively computing the estimates based on batches of observations
are called optimal smoothers. In this thesis, new practical algorithms for approximate
and asymptotically optimal continuous-discrete filtering and smoothing are
presented.
The mathematical approach of this thesis is probabilistic and the estimation
algorithms are formulated in terms of Bayesian inference. This means that the
unknown parameters, the unknown functions and the physical noise processes are
treated as random processes in the same joint probability space. The Bayesian approach
provides a consistent way of computing the optimal filtering and smoothing
estimates, which are optimal given the model assumptions and a consistent
way of analyzing their uncertainties.
The formal equations of the optimal Bayesian continuous-discrete filtering
and smoothing solutions are well known, but the exact analytical solutions are
available only for linear Gaussian models and for a few other restricted special
cases. The main contributions of this thesis are to show how the recently developed
discrete-time unscented Kalman filter, particle filter, and the corresponding
smoothers can be applied in the continuous-discrete setting. The equations for the
continuous-time unscented Kalman-Bucy filter are also derived.
The estimation performance of the new filters and smoothers is tested using
simulated data. Continuous-discrete filtering based solutions are also presented to
the problems of tracking an unknown number of targets, estimating the spread of
an infectious disease and to prediction of an unknown time series.