Scalable Sequential Monte Carlo Samplers for Numerical Bayesian Inference


Carter, Matthew ORCID: 0000-0002-0368-7042 (2025) Scalable Sequential Monte Carlo Samplers for Numerical Bayesian Inference. PhD thesis, University of Liverpool.

[thumbnail of 201371920_Jun2025.pdf] Text
201371920_Jun2025.pdf - Author Accepted Manuscript
Access to this file is embargoed until 1 August 2026.
Download (3MB)

Abstract

Sequential Monte Carlo (SMC) samplers offer a promising alternative to Markov Chain Monte Carlo (MCMC) methods for inferring a target density associated with a static dataset. Empirical evidence suggests that SMC samplers may require a shorter burn-in period than MCMC. Moreover, SMC samplers can be parallelised, provide estimates of the normalising constant, and support a variety of proposal distributions. However, current state-of-the-art SMC configurations have notable limitations: they suffer from the curse of dimensionality, are computationally inefficient, operate sequentially, and have long run times due to expensive tempering schemes. Recent studies have shown that SMC samplers can effectively utilise clusters of CPUs. Yet, these samplers fail to exploit hardware accelerators and are often restricted to specialist high-performance computing facilities. These challenges highlight the need for SMC samplers that can handle high-dimensional problems, avoid costly tempering schemes, leverage hardware accelerators, and operate on commodity computing resources. This thesis begins by outlining the Bayesian framework and MCMC methods before focusing on the importance sampling framework and SMC methods. A novel SMC sampler is introduced, designed to tackle high-dimensional problems and exhibit strong parallel scaling behaviour. Particle recycling schemes are then explored to enhance the computational efficiency of the sampler. A mechanism for selecting iterations to recycle and forming unbiased recycled estimates of high-order statistics is proposed. The proposed SMC sampler is parallelised on shared and distributed memory architectures, and various parallel computing frameworks are evaluated. Subsequently, a framework for distributing the SMC sampler on an opportunistic computing environment, comprising a heterogeneous collection of commodity computing resources, is presented. This framework offers a cost-effective and efficient way for practitioners to benefit from SMC samplers. A novel SMC sampler is then introduced, providing an unbiased, lower-error alternative to multiple short parallel MCMC chains. The thesis concludes with a summary of contributions and recommendations for future work.

Item Type: Thesis (PhD)
Uncontrolled Keywords: Bayesian Inference, Decision Making, Distributed Computing, High Performance Computing, Sequential Monte Carlo
Divisions: Faculty of Science and Engineering
Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science
Depositing User: Symplectic Admin
Date Deposited: 14 Aug 2025 10:26
Last Modified: 14 Aug 2025 12:12
DOI: 10.17638/03193233
Supervisors:
URI: https://livrepository.liverpool.ac.uk/id/eprint/3193233
Altmetric
CORE (COnnecting REpositories)