Data-Driven Model Set Design for Model Averaged Particle Filter

TL;DR

This paper introduces a generic, data-driven approach for designing model sets in Bayesian model averaged particle filters, which does not require prior knowledge of true model parameters and adapts using Bayesian optimization on pre-obtained noisy observations.

Contribution

It proposes a novel, practical method for model set design in BMAPF that leverages Bayesian optimization without prior parameter knowledge, ensuring performance and diversity.

Findings

Effective model set design demonstrated through simulations

Improved filtering performance over existing methods

No prior model parameter knowledge required

Abstract

This paper is concerned with sequential state filtering in the presence of nonlinearity, non-Gaussianity and model uncertainty. For this problem, the Bayesian model averaged particle filter (BMAPF) is perhaps one of the most efficient solutions. Major advances of BMAPF have been made, while it still lacks a generic and practical approach to design the model set. This paper fills in this gap by proposing a generic data-driven method for BMAPF model set design. Unlike existent methods, the proposed solution does not require any prior knowledge on the parameter value of the true model; it only assumes that a small number of noisy observations are pre-obtained. The Bayesian optimization (BO) method is adapted to search the model components, each of which is associated with a specific segment of the pre-obtained dataset.The average performance of these model components is guaranteed since each one's parameter value is elaborately tuned via BO to maximize the marginal likelihood. The diversity in the model components is also ensured, as different components match the different segments of the pre-obtained dataset, respectively. Computer simulations are used to demonstrate the effectiveness of the proposed method.