Efficient Exploration of Multi-Modal Posterior Distributions

Abstract

The Markov Chain Monte Carlo (MCMC) algorithm is a widely recognised as an efficient method for sampling a specified posterior distribution. However, when the posterior is multi-modal, conventional MCMC algorithms either tend to become stuck in one local mode, become non-Markovian or require an excessively long time to explore the global properties of the distribution. We propose a novel variant of MCMC, mixed MCMC, which exploits a specially designed proposal density to allow the generation candidate points from any of a number of different modes. This new method is efficient by design, and is strictly Markovian. We present our method and apply it to a toy model inference problem to demonstrate its validity.