The adaptive patched particle filter and its implementation

TL;DR

This paper introduces an adaptive particle filtering method based on the Kusuoka-Lyons-Victoir approach, combining high-order approximation with local recombination and importance sampling to efficiently estimate evolving stochastic systems.

Contribution

It presents a novel adaptive particle filter that integrates high-order KLV approximation, local support recombination, and importance sampling for improved stochastic system state estimation.

Findings

Enhanced accuracy in stochastic system estimation

Reduced particle count through local recombination

Effective high-order adaptive importance sampling

Abstract

There are numerous contexts where one wishes to describe the state of a randomly evolving system. Effective solutions combine models that quantify the underlying uncertainty with available observational data to form relatively optimal estimates for the uncertainty in the system state. Stochastic differential equations are often used to mathematically model the underlying system. The Kusuoka-Lyons-Victoir (KLV) approach is a higher order particle method for approximating the weak solution of a stochastic differential equation that uses a weighted set of scenarios to approximate the evolving probability distribution to a high order of accuracy. The algorithm can be performed by integrating along a number of carefully selected bounded variation paths and the iterated application of the KLV method has a tendency for the number of particles to increase. Together with local dynamic recombination that simplifies the support of discrete measure without harming the accuracy of the approximation, the KLV method becomes eligible to solve the filtering problem for which one has to maintain an accurate description of the ever-evolving conditioned measure. Besides the alternate application of the KLV method and recombination for the entire family of particles, we make use of the smooth nature of likelihood to lead some of the particles immediately to the next observation time and to build an algorithm that is a form of automatic high order adaptive importance sampling.