Applied Mathematical and Computational Sciences vol:3 issue:3 pages:261-274
Particle filters have proven their value in many sensor-based robotics applications, and experience is gained about their computational properties such as cost, convergence robustness and the need for careful configuration of the resampling step. This paper investigates the appropriateness, in the robotics domain, of quasi-Monte Carlo techniques (QMC) which have shown a number of promising computational properties. Most notably: a better uniform distribution compared to pseudo-random samples and faster converging integral approximations.
The following quasi-random sequences will be tested in this paper: Sobol, Halton, reverse Halton, Niederreiter and several lattice based generators, which posses the completely uniform distributed (CUD) property. Their performance in terms of the convergence of the particle filter is evaluated on the ''robot maze'' example, which was carefully chosen to be as simple as possible but still representative for real-world particle filter applications.
No single quasi-random method is found to perform best and no clear advantage of the more uniform quasi-random samples is observed in the ''robot maze'' application. This is supporting the slumbering belief that, although QMC proved very efficient in many different areas, it might not be in a particle filter context.