In the seminal paper "Rao-Blackwellised Particle Filtering for Dynamic Bayesian Networks" by A. Doucet et. al. a sequential monte carlo filter (particle filter) is proposed, which makes use of a linear substructure $x^L_k$ in a markov process $x_k = (x^L_k, x^N_k)$. By marginalization of this linear structure, the filter can be split into two parts: a non-linear part which uses a particle filter, and one linear part which can be handled by a Kalman filter (conditioned on the non-linear part $x^N_k$).

I understand the marginalization part (and sometimes the described filter is also called marginalized filter). My intuition why it is called a Rao-Blackwellized Particle Filter (RBPF) is that the Gaussian parameters are a sufficient statistics for the underlying linear process, and following from the theorem of Rao-Blackwell an estimator conditioned on these parameters performs at least as good as the sampling estimator.

The Rao-Blackwell estimator is defined as $E(\delta(X)|T(X)) = \delta_1(X)$. In this context I would guess that $\delta(X)$ is the monte carlo estimator, $\delta_1(X)$ the RBPF, and $T(X)$ the gaussian parametrization. My problem is that I don't see where this is actually applied in the paper.

So why is this called a Rao-Blackwellized Particle Filter, and where does the Rao-Blackwellization actually happen?


In $\widehat{I^1}$ the Monte Carlo estimate of $\mathbb{E}[f]$ is used. In $\widehat{I^2}$ the expectation is computed exactly. This is the RB-part.

Later in the paper the expectation is computed using Kalman filters.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.