Abstract
Particle filtering procedures are widely used in Bayesian state estimation of large-scale dynamic systems given their massive datasets. However, they may end up in a situation called degeneracy, where a single particle abruptly possesses significant amount of normalized weights. An efficacious selection of the importance sampling density plays a pivotal role in achieving good performance of the particle filters by preventing the sampling procedure from generating degenerated weights for particles at early stages of the process. Such a selection of a proposal distribution that takes into account both the transition prior and the likelihood becomes especially crucial for better estimation accuracy when the observation data has significant impacts on the posterior states. In this study, we propose a novel importance density selection structure for particle filters based on the minimized relative entropy principle. First, the theoretical derivation of the proposed structure is presented. Then, the proposed scheme is benchmarked against other sampling schemes that exist in the literature for state estimation in terms of their estimation qualities and computational efficiencies through a set of synthetic experiments. Finally, various energy-related estimation problems that match the applicable circumstances of the proposed scheme are briefly discussed. [PUBLICATION ABSTRACT]