Title: Mixing efficiency of trans-model MCMC algorithms in Bayesian phylogenetics
Authors: Xiyun Jiao - Southern University of Science and Technology (China) [presenting]
Ziheng Yang - University College London (United Kingdom)
Tomas Flouri - University College London (United Kingdom)
Abstract: Bayesian models are commonly used for phylogenetic reconstruction. However, the corresponding Markov chain Monte Carlo (MCMC) methods, especially the trans-model ones, typically suffer from poor mixing properties. We have explored the relationship between the efficiency of trans-model MCMC and the proposal distributions it uses. Based on the findings, we have proposed some methods to design the proposal distributions so that the performance of trans-model MCMC algorithms can be enhanced. We used two toy examples to illustrate our findings and deployed both theoretical arguments and real data analyses in Phylogenetics to show the effects of our methods.