Seminar :Compound Auxiliary Metropolis and Parallel Annealed Chains for Advanced Bayesian Computation

Seminar| Institute of Mathematical Sciences

Time:Friday, May 16th, 2025,10:00-11:00

Abstract:Modern Bayesian inference often requires sophisticated Monte Carlo algorithms capable of effectively navigating complex probability distributions characterized by multimodality, local optima, and high-dimensional parameter spaces. In this talk, I introduce the Compound Auxiliary Metropolis (CAM) algorithm, a novel extension of the Multiple-Try Metropolis (MTM) method. CAM incorporates auxiliary variables into the candidate generation process, significantly enhancing the exploration and traversal of challenging topographies in the target distribution. The proposed CAM algorithm rigorously maintains Markov chain validity while enabling improved performance over traditional MTM approaches, especially in problems with plateaus and complex local structures. Further, we present a new Monte Carlo methodology leveraging CAM within a framework of parallel annealed chains. This innovative combination integrates annealing strategies into a robust and highly parallelizable computational framework. Simulation studies and theoretical validations demonstrate that our methods substantially increase sampling efficiency, reduce the duration of the pre-stationary period, and improve the overall ability to escape local optima. These methods represent powerful tools for advanced Bayesian computation, addressing the computational challenges of contemporary statistical modeling problems.

王亮亮简介：

王亮亮教授，加拿大英属哥伦比亚大学（University of British Columbia）统计学博士，西蒙菲莎大学（Simon Fraser University）副教授。王亮亮教授长期从事贝叶斯统计、机器学习、函数型数据以及生物统计的科学研究；担任多个国际会议的组委会成员；在国际和国内著名的统计和机器学习杂志上发表学术论文近六十篇；主持和参与过多个国家基金项目，如加拿大统计科学研究所（CANSSI）项目、加拿大自然科学与工程研究理事会（NSERC）项目等。