Resampled Branching Particle Filters

• Author(s) / Creator(s)

  • Kouritzin, Michael A.

• A large class of discrete-time branching particle filters with Bayesian model selection ca-pabilities and effective resampling is introduced in algorithmic form, shown empirically to outperform the popular bootstrap algorithm and analyzed mathematically. The particles interact weakly in the resampling procedure. The weighted particle filter, which has no resampling, and the fully-resampled branching particle filter are included in the class as extreme points. Each particle filter in the class is coupled to a McKean-Vlasov particle system, corresponding to a reduced, unimplementable particle filter, for which Marcinkiewicz strong laws of large numbers (Mllns) and the central limit theorem (clt) can be written down. Coupling arguments are used to show the reduced system can be used to predict performance of the particle filter and to transfer the Mllns to the original weakly-interacting particle filter. This clt is also shown transferable when extra particles are used.

• Date created

  2015-04-20

• Subjects / Keywords

  • Central Limit Theorem
  • Exchangeable Random Variables
  • Particle filters
  • McKean-Vlasov
  • Bayesian Model Selection
  • Branching Process
  • Coupling

• Type of Item

  Research Material

• DOI

  https://doi.org/10.7939/R39S1KN7W

• License

  © 2015 Michael A. Kouritzin. This version of this article is open access and can be downloaded and shared. The original author and source must be cited.