Sequential Monte Carlo Methods in Practice
Sequential Monte Carlo Methods in Practice
复制标题
DOI:
10.1198/tech.2003.s23
复制
发表时间:
2003-02
期刊:
影响因子:
2.5
通讯作者:
Pradipta Sarkar
中科院分区:
文献类型:
--
作者:
Pradipta Sarkar
Sequential importance sampling (cid:2)SIS(cid:3) was (cid:4)rst developed in (cid:5)(cid:6)(cid:7)(cid:8)s for molecular simulation(cid:9) Although half a century has passed by(cid:10) the SIS methodology remains one of the most versatile and power(cid:11) ful means for the simulation and optimization of chain polymers(cid:9) In (cid:5)(cid:6)(cid:6)(cid:8)s(cid:10) statisticians reinvented the same methodology in a more gen(cid:11) eral form(cid:10) brought forth a number of enhancements(cid:10) and applied it to a much broader spectrum of problems(cid:9) In this article(cid:10) along with a historical account of the methodology(cid:10) we present a theoretical frame(cid:11) work for the SIS with resampling(cid:9) We emphasize the basic concept of a weighted sample(cid:10) the fundamental idea of sequential build(cid:2)up(cid:10) the important technique of reweighting and resampling(cid:10) and various other methods(cid:10) e(cid:9)g(cid:9)(cid:10) the partial rejection control and marginalization(cid:10) for im(cid:11) proving a sequential importance sampler(cid:9) To illustrate(cid:10) we show how to analyze a state(cid:11)space model with SIS and how to treat the con(cid:11) ditional dynamic linear model with marginalization(cid:10) resampling(cid:10) and rejection control techniques(cid:9) We report some simulation results for two(cid:11)dimensional target tracking in clutter(cid:9)