Seminar Announcement

Adaptive Squential Posterior Simulators for Massively Parallel Computing Environments

John Geweke, University of Technology, Sydney

Monday, October 22, 2012

4:00pm, room 223 Weber Building

 

Abstract

Massively parallel desktop computing capabilities now well within the reach of individual academics
modify the environment for posterior simulation in fundamental and potentially quite advantageous
ways. This paper shows that the new environment provides a generic method for evaluating simulation
error. Only single-instruction multiple-data algorithms can fully exploit massively parallel computing
environments, and sequential Monte Carlo comes very close to this ideal whereas other approaches
like Markov chain Monte Carlo do not. For such inference they must adapt to posterior distributions.
This creates complex structures of dependence that have impeded the development of the corresponding
distribution theory for evaluating simulation error. The paper provides a solution for this problem that is
practical in a massively parallel computing environment. It then introduces a specific sequential posterior
simulator that requires only code for evaluation of prior and data densities for application to a particular
model. It is robust to multimodal and other pathological posterior distributions, and provides accurate
marginal likelihood simulation approximations. The paper concludes with an application chosen to
illustrate important practical considerations in using sequential posterior simulators for applied Bayesian
inference.