Exponential Natural Particle Filter

Particle Filter (PF) is the most widely used Bayesian sequential estimation method for obtaining hidden states of nonlinear dynamic systems. However, it still suffers from certain problems such as the loss of particle diversity, the need for large number of particles, and the costly selection of the importance density functions. In this paper, a novel PF called Exponential Natural Particle Filter (xNPF) is introduced to solve the above problems. In this approach, a state transitional probability with the use of natural gradient learning is proposed which balances exploration and exploitation more robustly. PF with the proposed density function does not need a large number of particles and it retains particles’ diversity in a course of run. The proposed system is evaluated in a time-varying parameter estimation problem on a dynamic model of HIV virus immune response. This model is used to show the performance of the xNPF in comparison with several state of the art particle filter variants such as Annealed PF, Bootstrap PF, iterative PF, equivalent weight PF, and intelligent PF. The results show that xNPF converges much closer to the true target states than the other methods.