arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2209.04750stat.COmath.PRmath.STstat.TH
收藏

并行马尔可夫链蒙特卡洛算法:理论基础、算法设计与案例研究

Parallel MCMC Algorithms: Theoretical Foundations, Algorithm Design, Case Studies

Nathan E. Glatt-Holtz, Andrew J. Holbrook, Justin A. Krometis, Cecilia F. Mondaini

AI总结:

本文为并行MCMC算法建立测度论基础,推导多提议接受准则,提出多提议pCN算法,并通过并行化与贝叶斯反演案例验证其高效性。

摘要:

并行马尔可夫链蒙特卡洛(pMCMC)算法在每一步生成提议云(proposal cloud),以高效求解目标概率分布。我们为pMCMC算法构建了一个严格的基础框架,将这些方法置于统一的“扩展相空间”测度论形式体系中。借鉴我们最近关于可逆单提议方法的综合理论工作,我们在此推导了多提议接受机制的一般准则,这些准则在一般状态空间上产生遍历链。我们的公式涵盖多种方法论,包括提议云重采样和哈密顿方法,同时为推导新算法提供了基础。特别是,我们为源自“条件独立”提议结构的一类方法获得了自上而下的视角。作为直接应用,我们识别出几种新算法,包括适用于高维和无限维目标测度(这些测度关于高斯基测度绝对连续)的流行预条件Crank-Nicolson(pCN)采样器的多提议版本。为补充我们的理论结果,我们进行了一系列数值案例研究,以评估这些新算法的有效性。首先,注意到pMCMC算法的真正潜力源于其自然并行性,我们使用TensorFlow和图形处理单元进行了一项有限的并行化研究,以扩展在每一步利用多达10万个提议的pMCMC算法。其次,我们使用我们的多提议pCN算法(mpCN)来解决由流体测量驱动的偏微分方程贝叶斯统计反演中的一系列问题。这些例子为mpCN在处理具有复杂几何形状和多模态结构的高维目标分布方面的有效性提供了初步证据。

英文摘要:

Parallel Markov Chain Monte Carlo (pMCMC) algorithms generate clouds of proposals at each step to efficiently resolve a target probability distribution. We build a rigorous foundational framework for pMCMC algorithms that situates these methods within a unified 'extended phase space' measure-theoretic formalism. Drawing on our recent work that provides a comprehensive theory for reversible single proposal methods, we herein derive general criteria for multiproposal acceptance mechanisms which yield ergodic chains on general state spaces. Our formulation encompasses a variety of methodologies, including proposal cloud resampling and Hamiltonian methods, while providing a basis for the derivation of novel algorithms. In particular, we obtain a top-down picture for a class of methods arising from 'conditionally independent' proposal structures. As an immediate application, we identify several new algorithms including a multiproposal version of the popular preconditioned Crank-Nicolson (pCN) sampler suitable for high- and infinite-dimensional target measures which are absolutely continuous with respect to a Gaussian base measure. To supplement our theoretical results, we carry out a selection of numerical case studies that evaluate the efficacy of these novel algorithms. First, noting that the true potential of pMCMC algorithms arises from their natural parallelizability, we provide a limited parallelization study using TensorFlow and a graphics processing unit to scale pMCMC algorithms that leverage as many as 100k proposals at each step. Second, we use our multiproposal pCN algorithm (mpCN) to resolve a selection of problems in Bayesian statistical inversion for partial differential equations motivated by fluid measurement. These examples provide preliminary evidence of the efficacy of mpCN for high-dimensional target distributions featuring complex geometries and multimodal structures.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑