posted by Lawrence Murray on 19 May 2014
Four new packages are now available to demonstrate the bridge particle filter included in LibBi 1.1.0. The bridge particle filter is described in this new paper:
Del Moral, P. & Murray, L. M. Sequential Monte Carlo with highly informative Observations, 2014. [arXiv]
It is meant for state-space models where each observation is highly informative on the state process. A special case is that of diffusion bridge sampling, where there is no observation noise—the observation is of the state directly. The filter works by introducing a schedule of intermediate times at which additional weighting and resampling steps are performed.
The bridge particle filter is enabled in LibBi by adding a --filter bridge
command-line option, and a --nbridges n option to set the frequency of
intermediate times. A bridge top-level block is then added to the model,
describing the additional weighting function.
The four packages are:
OrnsteinUhlenbeckBridge a linear–Gaussian Ornstein–Uhlenbeck process with fixed parameters.
FederalFundsRate a linear–Gaussian Ornstein–Uhlenbeck process with parameter estimation for a Federal Funds Rate data set.
PeriodicDriftBridge a nonlinear periodic drift process with fixed parameters.
SIR a multivariate and nonlinear susceptible/infected/recovered compartmental model used in epidemiology, with parameter estimation for an influenza data set.
The existing NPZD package has also been updated to support (and use by default) the new bridge particle filter.
These four new packages, plus the NPZD package, are used in the paper to demonstrate the new method, so further details can be found there.