Discipline
Computational statistics; Bayesian inference; applied mathematics; bioinformatics
Methodology
Hamiltonian Monte Carlo; gradient-based inference; Markov chain Monte Carlo; continuous relaxation; automatic differentiation; algorithm development; simulation and computational benchmarking
Date
30 Jul 2026
Description
Derivative-based methods such as Hamiltonian Monte Carlo can provide highly efficient Bayesian inference in continuous parameter spaces, but they cannot be applied directly when important model components are discrete or combinatorial. This project will investigate methods for extending gradient-based inference to models containing discrete variables or structures. Potential approaches include continuous relaxations, marginalisation, differentiable approximations and hybrid algorithms that combine derivative-based updates for continuous parameters with specialised proposals for discrete components. Phylogenetic inference will provide a major application, with the resulting methods evaluated for sampling trees and epidemiological parameters in BEAST.
Keywords
Derivative-based inference;discrete variables;Hamiltonian Monte Carlo;Bayesian computation
Would Suit Applicants Who
Have strong interests in computational statistics, applied mathematics, computer science or phylogenetics and experience with scientific programming