|
Abstract:
|
Modern parallel computing devices are increasingly used for statistical inference. They are particularly well-suited for implementing Sequential Monte Carlo (SMC) algorithms; this powerful and versatile class of methods relies on importance sampling combined with a re-sampling step in order to propagate a set of particles forward in time. While the propagation step is straightforward to parallelize, the resampling step is more difficult. We discuss in this talk a variation of a recently proposed class of locally-interacting particle filters. We describe trade-offs between computational and statistical efficiency, as well as long-time stability of these methods. We prove in particular that, in many scenarios, locally-interacting particle filters inherit robust asymptotic properties of more standard fully-interacting particle filters.
|
