GA-inspired Particle Filter for Mitigating Severe Sample Impoverishment

Particle filters (PFs) can handle nonlinear and/or non-Gaussian systems and can provide more information than just mean and covariance. However, there are some practical issues in implementing PFs. A well-known problem is sample impoverishment, which might lead to poor state estimation results. Although the main reason for sample impoverishment is the resampling process, the occurrence of abrupt jumps in the state of the system causes a severe loss of particle diversity. In this paper the idea of using generational genetic algorithm (GA) is proposed for mitigating sample impoverishment caused by an unknown and abrupt jump in the state. The proposed PF, the generational GA-PF (GGAPF), modifies the particles with negligible likelihood and then concentrates them near the posterior peaks. We evaluate the validity of this method by applying it into a benchmark target tracking problem. Simulation results demonstrate excellent performance for GGAPF in comparison with two other types of PFs.