A1239
Title: Identification-driven MCMC
Authors: Yizhou Kuang - University of Manchester (United Kingdom) [presenting]
Toru Kitagawa - Brown University (United States)
Abstract: Sampling methods often face slow or non-convergence with irregular target distributions or in high-dimensional spaces. The purpose is to introduce a novel MCMC approach that leverages the knowledge of observationally equivalent sets of model parameters. It is first shown that this method performs on par with or better than conventional MCMC techniques, particularly in high-dimensional settings. It is also shown in simulation that it compares favorably to other prevalent sampling strategies, such as (adaptive) sequential Monte Carlo, especially as the dimensionality of the variables increases. For application, the performance of the algorithm is illustrated in an SVMA setting, allowing for non-invertibility.
