An Honest Look at the Weighted Particle Filter.

• Author(s) / Creator(s)

  • Kouritzin, Michael

• The classical particle filter, introduced in 1993, approximates the normalized filter directly. It has two defiencies, over resampling and the inability to distinguish models, the former of which was overcome but the later is fundamental. Conversely, the weighted particle filter, motivated by the unnormalized filter development, does not employ resampling and facilitates Bayes’ factor model selection but often suffers particle spread, where the majority of particles do not track the underlying signal. Still, resampling introduces noise and there are situations where the weighted particle filter does perform well. Herein, the weighted particle filter is analyzed in a simple discrete-time setting and rate-of-convergence baseline results are established that can be compared to results for other particle filters. Moreover, an example illustrating failure of the weighted particle filter is given.

• Date created

  2014-06-05

• Subjects / Keywords

  • Law of large numbers
  • Large deviations principle
  • Particle filters
  • Unnormalized filter
  • Functional law of the iterated logarithms
  • Functional central limit theorem

• Type of Item

  Article (Draft / Submitted)

• DOI

  https://doi.org/10.7939/R3KW57M4W

• License

  Attribution-NonCommercial-NoDerivatives 3.0 International