Sequential MCMC With The Discrete Bouncy Particle Sampler

Sequential MCMC (SMCMC) methods are a useful alternative to particle filters for performing sequential inference in a Bayesian framework in nonlinear and non-Gaussian state-space models. The weight degeneracy phenomenon which impacts the performance of even the most advanced particle filters in higher dimensions is avoided. In this paper, we explore the applicability of the discrete bouncy particle sampler, which is based on constructing a guided random walk and performing delayed rejection, to perform more effective sampling within SMCMC. We perform numerical simulations to examine when the proposed method offers advantages compared to state-of-the-art SMCMC techniques.