Generalized Two-Stage Particle Filter for High Dimensions

ICASSP 2023

Particle filters (PF)

PFs are online algorithms that estimate a hidden process from noisy data. They operate on two main steps per step: i) propose particles from proposal, and ii) compute their weights. Standard PFs do not work well for systems of high dimensions.

Two-Stage PF

A two-stage PF for high dimensions was proposed that modifies the proposal distribution by tempering particles, but could only be applied to systems with observation model that has separable equations, and a manually chosen tempering coefficient.

Generalized Two-Stage PF

We propose a generalized two-stage particle filter (GTPF) that can be applied to all models, and computes the distribution of the tempering coefficient. For thorough details, please refer to the following link.

Reference paper: http://tinyurl.com/ieeexplore-iloska-gtpf

(Reference for original two-stage PF can be found with-in).

Code

1. main_script.py - script to run the proposed filter

   • generates synthetic data according to user settings
   • calls and runs filter
   • plots the estimation of a random state trajectory

2. synthetic_data.py - module that can generate state-space model or linear model data

3. gtpf.py - module that stores the code of the proposed filter

4. posterior_compute.py - module that computes the posterior of the tempering parameter beta