Sequential Monte Carlo Methods for Dynamic Systems：（动态系统的序贯蒙特卡罗方法）.pdfVIP

Sequential Monte Carlo Metho ds for Dynamic Systems Jun S Liu and Rong Chen 1 Abstract A general framework for using Monte Carlo metho ds in dynamic systems is provided and its wide applications indicated Under this framework several currently available techniques are studied and generalized to accommo date more complex features All of these metho ds are partial combinations of three ingredients imp ortance sampling and resampling rejection sampling and Markov chain iterations We deliver a guideline on how they should b e used and under what circumstance each metho d is most suitable Through the analysis of dier ences and connections we consolidate these metho ds into a generic algorithm by combining desirable features In addition we prop ose a general use of RaoBlackwellization to improve p erformances Examples from econometrics and engineering are presented to demonstrate the imp ortance of RaoBlackwellization and to compare dierent Monte Carlo pro cedures Keywords Blind deconvolution Bo otstrap lter Gibbs sampling Hidden Markov mo del Kalman lter Markov chain Monte Carlo Particle lter Sequential imputation State space mo del Target tracking 1 Jun S Liu is an assistant professor of Statistics Department of Statistics Stanford University Stanford CA Rong Chen is an asso ciate professor of Statistics Department of Statistics Texas AM University College Station TX Lius research is partly supp orted by NSF grants DMS and DMS and the Terman fellowship from Stanford University Chens research is partly supp orted by NSF grant DMS