Variational Markov chain Monte Carlo for Bayesian smoothing of non-linear diffusions

Yuan Shen*, Dan Cornford, Manfred Opper, Cédric Archambeau

*Corresponding author for this work

Research output: Contribution to journalArticle

Abstract

In this paper we develop set of novel Markov Chain Monte Carlo algorithms for Bayesian smoothing of partially observed non-linear diffusion processes. The sampling algorithms developed herein use a deterministic approximation to the posterior distribution over paths as the proposal distribution for a mixture of an independence and a random walk sampler. The approximating distribution is sampled by simulating an optimized time-dependent linear diffusion process derived from the recently developed variational Gaussian process approximation method. The novel diffusion bridge proposal derived from the variational approximation allows the use of a flexible blocking strategy that further improves mixing, and thus the efficiency, of the sampling algorithms. The algorithms are tested on two diffusion processes: one with double-well potential drift and another with SINE drift. The new algorithm's accuracy and efficiency is compared with state-of-the-art hybrid Monte Carlo based path sampling. It is shown that in practical, finite sample applications the algorithm is accurate except in the presence of large observation errors and low to a multi-modal structure in the posterior distribution over paths. More importantly, the variational approximation assisted sampling algorithm outperforms hybrid Monte Carlo in terms of computational efficiency, except when the diffusion process is densely observed with small errors in which case both algorithms are equally efficient.

Original languageEnglish
Pages (from-to)149-176
Number of pages28
JournalComputational Statistics
Volume27
Issue number1
DOIs
Publication statusPublished - Mar 2012

Fingerprint

Nonlinear Diffusion
Markov Chain Monte Carlo
Markov processes
Smoothing
Diffusion Process
Hybrid Monte Carlo
Variational Approximation
Sampling
Posterior distribution
Path Sampling
Linear Diffusion
Double-well Potential
Path
Markov Chain Monte Carlo Algorithms
Linear Process
Nonlinear Process
Markov chain Monte Carlo
Gaussian Process
Approximation Methods
Computational Efficiency

Keywords

  • bridge sampling
  • data assimilation
  • stochastic dynamic systems

Cite this

Shen, Yuan ; Cornford, Dan ; Opper, Manfred ; Archambeau, Cédric. / Variational Markov chain Monte Carlo for Bayesian smoothing of non-linear diffusions. In: Computational Statistics. 2012 ; Vol. 27, No. 1. pp. 149-176.
@article{f3eb853e298e4813a1511d01a508c596,
title = "Variational Markov chain Monte Carlo for Bayesian smoothing of non-linear diffusions",
abstract = "In this paper we develop set of novel Markov Chain Monte Carlo algorithms for Bayesian smoothing of partially observed non-linear diffusion processes. The sampling algorithms developed herein use a deterministic approximation to the posterior distribution over paths as the proposal distribution for a mixture of an independence and a random walk sampler. The approximating distribution is sampled by simulating an optimized time-dependent linear diffusion process derived from the recently developed variational Gaussian process approximation method. The novel diffusion bridge proposal derived from the variational approximation allows the use of a flexible blocking strategy that further improves mixing, and thus the efficiency, of the sampling algorithms. The algorithms are tested on two diffusion processes: one with double-well potential drift and another with SINE drift. The new algorithm's accuracy and efficiency is compared with state-of-the-art hybrid Monte Carlo based path sampling. It is shown that in practical, finite sample applications the algorithm is accurate except in the presence of large observation errors and low to a multi-modal structure in the posterior distribution over paths. More importantly, the variational approximation assisted sampling algorithm outperforms hybrid Monte Carlo in terms of computational efficiency, except when the diffusion process is densely observed with small errors in which case both algorithms are equally efficient.",
keywords = "bridge sampling, data assimilation, stochastic dynamic systems",
author = "Yuan Shen and Dan Cornford and Manfred Opper and C{\'e}dric Archambeau",
year = "2012",
month = "3",
doi = "10.1007/s00180-011-0246-4",
language = "English",
volume = "27",
pages = "149--176",
journal = "Computational Statistics",
issn = "0943-4062",
publisher = "Springer",
number = "1",

}

Variational Markov chain Monte Carlo for Bayesian smoothing of non-linear diffusions. / Shen, Yuan; Cornford, Dan; Opper, Manfred; Archambeau, Cédric.

In: Computational Statistics, Vol. 27, No. 1, 03.2012, p. 149-176.

Research output: Contribution to journalArticle

TY - JOUR

T1 - Variational Markov chain Monte Carlo for Bayesian smoothing of non-linear diffusions

AU - Shen, Yuan

AU - Cornford, Dan

AU - Opper, Manfred

AU - Archambeau, Cédric

PY - 2012/3

Y1 - 2012/3

N2 - In this paper we develop set of novel Markov Chain Monte Carlo algorithms for Bayesian smoothing of partially observed non-linear diffusion processes. The sampling algorithms developed herein use a deterministic approximation to the posterior distribution over paths as the proposal distribution for a mixture of an independence and a random walk sampler. The approximating distribution is sampled by simulating an optimized time-dependent linear diffusion process derived from the recently developed variational Gaussian process approximation method. The novel diffusion bridge proposal derived from the variational approximation allows the use of a flexible blocking strategy that further improves mixing, and thus the efficiency, of the sampling algorithms. The algorithms are tested on two diffusion processes: one with double-well potential drift and another with SINE drift. The new algorithm's accuracy and efficiency is compared with state-of-the-art hybrid Monte Carlo based path sampling. It is shown that in practical, finite sample applications the algorithm is accurate except in the presence of large observation errors and low to a multi-modal structure in the posterior distribution over paths. More importantly, the variational approximation assisted sampling algorithm outperforms hybrid Monte Carlo in terms of computational efficiency, except when the diffusion process is densely observed with small errors in which case both algorithms are equally efficient.

AB - In this paper we develop set of novel Markov Chain Monte Carlo algorithms for Bayesian smoothing of partially observed non-linear diffusion processes. The sampling algorithms developed herein use a deterministic approximation to the posterior distribution over paths as the proposal distribution for a mixture of an independence and a random walk sampler. The approximating distribution is sampled by simulating an optimized time-dependent linear diffusion process derived from the recently developed variational Gaussian process approximation method. The novel diffusion bridge proposal derived from the variational approximation allows the use of a flexible blocking strategy that further improves mixing, and thus the efficiency, of the sampling algorithms. The algorithms are tested on two diffusion processes: one with double-well potential drift and another with SINE drift. The new algorithm's accuracy and efficiency is compared with state-of-the-art hybrid Monte Carlo based path sampling. It is shown that in practical, finite sample applications the algorithm is accurate except in the presence of large observation errors and low to a multi-modal structure in the posterior distribution over paths. More importantly, the variational approximation assisted sampling algorithm outperforms hybrid Monte Carlo in terms of computational efficiency, except when the diffusion process is densely observed with small errors in which case both algorithms are equally efficient.

KW - bridge sampling

KW - data assimilation

KW - stochastic dynamic systems

UR - http://www.scopus.com/inward/record.url?scp=84856015685&partnerID=8YFLogxK

UR - http://link.springer.com/article/10.1007%2Fs00180-011-0246-4

U2 - 10.1007/s00180-011-0246-4

DO - 10.1007/s00180-011-0246-4

M3 - Article

AN - SCOPUS:84856015685

VL - 27

SP - 149

EP - 176

JO - Computational Statistics

JF - Computational Statistics

SN - 0943-4062

IS - 1

ER -