Sequential Monte Carlo: Past and Present

Abstract

This talk will review historical developments in polymer simulation, which is closely related to the recently-popular Monte Carlo particle filtering, or more precisely, sequential Monte Carlo (SMC) methods. Two key elements in SMC are sequential importance sampling (SIS) and resampling (or pruning-enrichment). SIS is a generic, but useful strategy for building up the trial distribution for a high-dimensional problem and can be applied naturally to accommodate dynamic systems and to mimic a learning mechanism. Because of SIS's sequential structure, one can monitor its importance sampling weights along with the sequential sampling and make appropriate interference, such as resampling and rejection sampling, to control Monte Carlo variations. SIS together with many interference techniques gives rise to a collection of related methods with the name "sequential Monte Carlo." We show some success stories of the method in energy minimization for protein folding, statistical inference, target tracking, digital telecommunications, etc.