Adaptive parallel/serial sampling mechanisms for particle filtering in dynamic bayesian networks

Abstract

Monitoring the variables of real world dynamical systems is a difficult task due to their inherent complexity and uncertainty. Particle Filters (PF) perform that task, yielding probability distribution over the unobserved variables. However, they suffer from the curse of dimensionality problem: the necessary number of particles grows exponentially with the dimensionality of the hidden state space. The problem is aggravated when the initial distribution of the variables is not well known, as happens in global localization problems. In this paper we present two new adaptive sampling mechanisms for PFs for systems whose variable dependencies can be factored into a Dynamic Bayesian Network. The novel PFs, developed over the proposed sampling mechanisms, exploit the strengths of other existing PFs. Their adaptive mechanisms 1) modify or establish probabilistic links among the subspaces of hidden variables that are independently explored to build particles consistent with the current measurements and past history, and 2) tune the performance of the new PFs toward the behaviors of several existing PFs. We demonstrate their performance on some complex dynamical system estimation problems, showing that our methods successfully localize and track hidden states, and outperform some of the existing PFs.