Rao-Blackwellised Particle Filtering for Dynamic Bayesian Networks

Particle filtering in high dimensional state-spaces can be inefficient because a large number of samples is needed to represent the posterior. A standard technique to increase the efficiency of sampling techniques is to reduce the size of the state space by marginalizing out some of the variables analytically; this is called Rao-Blackwellisation (Casella and Robert 1996). Combining these two techniques results in Rao-Blackwellised particle filtering (RBPF) (Doucet 1998, Doucet, de Freitas, Murphy and Russell 2000). In this chapter, we explain RBPF, discuss when it can be used, and give a detailed example of its application to the problem of map learning for a mobile robot, which has a very large (~ 2100) discrete state space.