The Influence of Velocity Refresh in Sequential MCMC with the Invertible Particle Flow and Discrete Bouncy Particle Sampler

Abstract

Sequential Markov Chain Monte Carlo (SMCMC) methods can be applied in the Bayesian inference framework with the nonlinear non-Gaussian state space model. SMCMC can avoid the weight degeneracy which impact the performance of Sequential Monte Carlo (SMC) methods in the high-dimensional state space model. Recently, Discrete Bouncy Particle Sampler (DBPS) is proposed as the refinement step in the Composite Metropolis-Hasting (MH) Kernel of SMCMC framework. Traditional Bouncy Particle Sampler has the reducible problem. In this paper, we explore different velocity refresh method to avoid the reducible problem in the DBPS method and embed the velocity refresh step into the SMCMC framework. We perform experiments to evaluate the proposed methods and the state-of-the-art SMCMC methods.