GPU acceleration of the particle filter: the Metropolis resampler

TL;DR

This paper explores GPU-accelerated particle filtering focusing on a Metropolis resampler that offers potential speed advantages by requiring only pair-wise weight ratios, suitable for real-time and performance-critical applications.

Contribution

It introduces a GPU-friendly Metropolis resampler for particle filters that reduces synchronization needs and can outperform traditional resamplers under certain conditions.

Findings

Metropolis resampler requires only pair-wise weight ratios.

It can be faster than traditional resamplers when importance weights have low variance.

Suitable for real-time particle filtering applications.

Abstract

We consider deployment of the particle filter on modern massively parallel hardware architectures, such as Graphics Processing Units (GPUs), with a focus on the resampling stage. While standard multinomial and stratified resamplers require a sum of importance weights computed collectively between threads, a Metropolis resampler favourably requires only pair-wise ratios between weights, computed independently by threads, and can be further tuned for performance by adjusting its number of iterations. While achieving respectable results for the stratified and multinomial resamplers, we demonstrate that a Metropolis resampler can be faster where the variance in importance weights is modest, and so is worth considering in a performance-critical context, such as particle Markov chain Monte Carlo and real-time applications.