Discussion Lead: Samuel Livingstone
Topic: The Barker proposal: Combining robustness and efficiency in gradient‐based MCMC
Abstract: I will introduce a class of stochastic processes that can be defined on both discrete and continuous state spaces to have a user-specified invariant distribution. I'll give some examples in both cases of how they can be used to design Monte Carlo algorithms, and try to give a selective review of the literature, some of which is my own work and some that of others. I'll try to discuss a bit about each of these and also try to speculate a bit on wider use in both sampling and optimization if time permits.
Reference: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9303935/
Pedagogical intro: https://arxiv.org/pdf/2012.09731.pdf