Improving estimation accuracy of particle filter by efficient interpolation based on crossover

This paper presents a new particle filter (PF) that efficiently works with the smaller number of particles than the original one. PF is one of the most promising sequential data assimilation methods because it can be applied to nonlinear and/or non-Gaussian systems. However, PF has a problem in that its estimation accuracy deteriorates when the number of particles in the ensemble is small. We believe that the deterioration is caused because PF is hardly able to create new particles in high-likelihood regions if no particles are created in the regions in initialization. In order to remedy the problem of PF, we propose a new PF named the particle filter with interpolation by crossover (PF-IC) that is able to create new particles in the high-likelihood regions during estimation. PF-IC generates new particles by means of a real-coded crossover operator developed for real-coded genetic algorithms. In order to investigate the effectiveness of PF-IC, we applied PF-IC and the original PF to a benchmark problem. As the result, we confirmed that PF-IC showed 20% better performance than the original PF in terms of the root mean squared error.