Erik Lindström

A Delayed Propagation Filter for Partially Observed Diffusion Processes

Abstract

We develop a sequential Monte Carlo filtering algorithm for non-linear,
partially observed  diffusion processes. The filter behaves advantageous
compared to the ordinary particle filters (i.e. the Bootstrap filter)
when the measurement noise is small, as the approximation does not
degenerate.

The performance of the filter can be improved by applying importance
samplers. Several importance samplers for partially observed diffusions
are derived, including a near optimal importance sampler. The samplers
can be used in ordinary particle filters as well as in the proposed
filter. The performance of the near optimal filter is competitive over a
large range of different parameter values.

This talk is based on joint work with Hans R. Künsch, ETHZ