Introduction to labeled island particle models and their asymptotic properties

TL;DR

This paper introduces the labeled island particle algorithm, a novel sequential method for jointly estimating a stochastic process and its environment parameters, with proven convergence and error bounds, demonstrated on a filtering problem.

Contribution

The paper presents a new labeled island particle algorithm that approximates both a stochastic process and its environment parameters, with theoretical convergence and error bounds.

Findings

Proves convergence of the labeled island particle algorithm.

Establishes $ ext{L}^p$ bounds and time-uniform error bounds.

Demonstrates effectiveness on a filtering problem with noisy observations.

Abstract

Estimation of stochastic processes evolving in a random environment is of crucial importance for example to predict aircraft trajectories evolving in an unknown atmosphere. For fixed parameter, interacting particle systems are a convenient way to approximate such stochastic process. But the second level of uncertainty provided by the environment parameters leads us to also consider interacting particles on the parameter space. This novel algorithm is described in this paper. It allows to approximate both a random environment and a stochastic process evolving in this environment, given noisy observations of the process. It is a sequential algorithm that generalizes island particle models including a parameter. It is referred by us as labeled island particle algorithm. We prove the convergence of the labeled island particle algorithm and we establish $\mathbb{L}^p$ bound as well as time uniform $\mathbb{L}^p$ bound for the asymptotic error introduced by this double level of approximation. Finally, we illustrate these results on a filtering problem where one learns a dynamical parameter through noisy observations of a stochastic process influenced by the parameter.