Sequential Monte Carlo Sampling Method And Study In Parallel

Sequential monte carlo methods is applicable to nonlinearnon-gaussian model, in the signal transmission and compression, machineperception and navigation, maneuvering target tracking, human bodyoutline tracking and unusual behavior analysis, complex industry processfault diagnosis, finance and other fields of data analysis was obtained,and the application of success. But, sequential monte carlo methods hasthe greatest flaw particles degeneration phenomenon, along with theincrease of the time, the algorithm in the course of several iterations,importance weights have may focus on to a small number of particles, anda lot of the importance of the corresponding weights particles to zero,and thus easily make sequential importance sampling(SIS) methods appearsdivergence. According to the degeneration of particle, particleresampling is to reduce the impact of this phenomenon is one of theimportant ways, therefore, study and improve sequential monte carloresampling methods to improve efficiency , having very vitalsignificance.In this paper, study the sequential monte carlo parallel resamplingmethod and its application of dynamic Tobit model, main job overviewbelow.Firstly, proposed any state space model of random sample to producesequential monte carlo parallel resampling method. In order to decreasethe execution time of the particle filter, the parallel resamplingparticle filter algorithm is proposed. In the algorithm, firstly allweights of the particles are sorted according to the ascending order.Secondly the particles space is classified into two independent sets.Finally the particles that will be resampled from two sets respectivelyare found parallelly according to the random search method. Analysis ismade to compare parallel resampling with multinomial resampling, residualresampling and stratified resampling. The results show that theaverage performance of parallel resampling is superior to otherResampling. According to the theoretical analysis and the experimentresults, the algorithm can reduce the search space for resampling and canshorten the search time, so it has high efficiency in the implementation. What is more, the algorithm can overcome the blindness of resampling, andcan better embody the basic idea of resampling which is a good weightparticle to be reproduced more, so it has better filter and estimationperformance.Secondly, Tobit model as a partially observed linear Gaussianstate-space model. The value of the dependent variable, it can not be fullyobservation, in some data points, the value of the dependent variable,is missing. Sequential monte carlo parallel resampling method wassimulated in Tobit dynamic model, the method to state estimationperformance is better than standard sequential monte carlo method.In summary, this thesis proposes parallel resampling of sequentialmonte carlo methods, through the simulation research and case analysis,the feasibility of the method is validated.