David Hastie
Title: Towards automatic reversible jump Markov chain Monte Carlo
Abstract:
Since its introduction by Green (1995), reversible jump MCMC has been
recognised as a powerful tool for making posterior inference about a
wide range of statistical problems. Although the method has been
applied to many problems within a variety of disciplines, its
popularity has been tempered by the common perception that reversible
jump samplers can be difficult to implement.
This talk presents steps towards the design of an automatic reversible
jump sampler, with the aim of taking the method outside the domain of
the MCMC expert. Generalising the first steps taken by Green (2003),
the generic AutoMix sampler is introduced. Requiring minimal user
input, the AutoMix sampler uses adaptive techniques to perform
self-tuning and calibration, allowing the consideration of many
trans-dimensional statistical problems.
The main features of the sampler are detailed, demonstrating the
sampler's broad applicability. We present examples of the AutoMix
sampler in practice, indicating performance that is at least as good as
problem-specific samplers, designed and tuned by hand. Finally, we
discuss potential areas for future research.