Improvement And Application Of Particle Filter Algorithm Based On Genetic Optimiztion

Particle filtering is a filtering method developed on the basis of Monte Carlo method and Bayesian theory.Its basic idea is to approximate the probability distribution in the required state space with a set of particles.The particles in the state space are actually the samples in statistics,which are generated by Monte Carlo sampling technology according to the conditional distribution of the state vector of the system.Generally speaking,the approximation degree of the initial empirical condition distribution corresponding to the particle set to the probability distribution of the required solution is not accurate enough,and the position and weight of the particles need to be adjusted according to continuous measurement.The empirical conditional distribution is modified by new particle information,thus obtaining the solution of the problem close to the real value.The common problems of particle filter algorithm are the weakening of particle diversity and the inevitable particle degradation.Resampling can improve the influence brought by the degradation of particle filter algorithm,but on the other hand,continuous resampling makes particles with high weight be duplicated repeatedly,while particles with low weight may be discarded,thus losing particle diversity.Genetic algorithm is an important intelligent optimization method.Its genetic steps are similar to the resampling steps of particle filter algorithm.Based on the particle filter algorithm of genetic algorithm,this paper proposes a method of constructing cross mutation probability function by Sigmiod function,in order to improve the diversity of particle set.The specific operation steps are as follows: firstly,judging whether the particles need to be crossed and mutated according to the weight value of the particles;If necessary,the particle set is updated by the adaptive crossover probability formula and mutation probability formula constructed in this paper.Finally,the particles with larger weight are retained,and the particles with smaller weight are changed by using larger mutation probability.In this way,the diversity of particle sets is maintained and the particle filter algorithm is optimized.The error problem between particle filter algorithm and adaptive genetic algorithm for prediction estimation of particle distribution state and real distribution is studied.Based on the above methods,this paper uses KL distance(Kullback-Leibler Divergence)to measure the error between the real distribution and the predicted distribution of particles.This method can coordinate the relationship between the number of particles and the prediction error and further reduce the prediction error of the algorithm.To sum up,based on the particle filter algorithm integrated with genetic algorithm,this paper proposes two improved methods for genetic steps: one is Sigmiod function,the other is KL distance method.Based on the cross mutation probability function constructed by Sigmiod function,the prediction error of the algorithm is further optimized by KL distance.In this paper,the improved particle filter algorithm is used to simulate scalar model,four-dimensional bearings-only model and target tracking model.Comparing and analyzing the algorithm with particle filter algorithm and genetic particle filter algorithm,it shows that the algorithm is improved in terms of computation and approximation.