Intelligent Genetic Crossover Algorithm for Improving State Estimation in Particle Filtering

Particle filtering (PF) is a Monte Carlo (MC) method widely employed to sequentially approximate the time-varying posterior PDF of states or values of variables of interest conditional to noisy observation data from non-linear/non-Gaussian dynamic systems. In many applications, we also infer (or estimate) the hidden state from the obtained approximated posterior PDF. Particles are randomly drawn samples of state vectors that form the approximation of the posterior PDF. However, representation of the approximated posterior PDF may not always be good because sometimes there can be only a few particles with low likelihoods (or importance weights), while there can be many particles with near-zero (or negligible) weights. New state vectors with high weights then should be introduced to the swarm in order to reshape the approximated posterior PDF. This paper introduces an intelligent genetic crossover algorithm (IGCA) that assists PF by applying crossover schemes employed in genetic algorithms (GAs) to reshape the approximated posterior PDF. Experimental results shown that the proposed algorithm improved the accuracy and performance in nonlinear state estimation.