skip to main content
OSTI.GOV title logo U.S. Department of Energy
Office of Scientific and Technical Information

Title: Sythesis of MCMC and Belief Propagation

Abstract

Markov Chain Monte Carlo (MCMC) and Belief Propagation (BP) are the most popular algorithms for computational inference in Graphical Models (GM). In principle, MCMC is an exact probabilistic method which, however, often suffers from exponentially slow mixing. In contrast, BP is a deterministic method, which is typically fast, empirically very successful, however in general lacking control of accuracy over loopy graphs. In this paper, we introduce MCMC algorithms correcting the approximation error of BP, i.e., we provide a way to compensate for BP errors via a consecutive BP-aware MCMC. Our framework is based on the Loop Calculus (LC) approach which allows to express the BP error as a sum of weighted generalized loops. Although the full series is computationally intractable, it is known that a truncated series, summing up all 2-regular loops, is computable in polynomial-time for planar pair-wise binary GMs and it also provides a highly accurate approximation empirically. Motivated by this, we first propose a polynomial-time approximation MCMC scheme for the truncated series of general (non-planar) pair-wise binary models. Our main idea here is to use the Worm algorithm, known to provide fast mixing in other (related) problems, and then design an appropriate rejection scheme to sample 2-regularmore » loops. Furthermore, we also design an efficient rejection-free MCMC scheme for approximating the full series. The main novelty underlying our design is in utilizing the concept of cycle basis, which provides an efficient decomposition of the generalized loops. In essence, the proposed MCMC schemes run on transformed GM built upon the non-trivial BP solution, and our experiments show that this synthesis of BP and MCMC outperforms both direct MCMC and bare BP schemes.« less

Authors:
 [1];  [2];  [3]
  1. Korea Advanced Institute of Science and Technology, Daejeon (South Korea)
  2. Los Alamos National Lab. (LANL), Los Alamos, NM (United States)
  3. Korea Advanced Institute of Science and Technology, Daejeon (South Korea)
Publication Date:
Research Org.:
Los Alamos National Lab. (LANL), Los Alamos, NM (United States)
Sponsoring Org.:
USDOE Office of Electricity Delivery and Energy Reliability (OE)
OSTI Identifier:
1254988
Report Number(s):
LA-UR-16-23749
DOE Contract Number:
AC52-06NA25396
Resource Type:
Technical Report
Country of Publication:
United States
Language:
English
Subject:
97 MATHEMATICS AND COMPUTING; Computer Science

Citation Formats

Ahn, Sungsoo, Chertkov, Michael, and Shin, Jinwoo. Sythesis of MCMC and Belief Propagation. United States: N. p., 2016. Web. doi:10.2172/1254988.
Ahn, Sungsoo, Chertkov, Michael, & Shin, Jinwoo. Sythesis of MCMC and Belief Propagation. United States. doi:10.2172/1254988.
Ahn, Sungsoo, Chertkov, Michael, and Shin, Jinwoo. Fri . "Sythesis of MCMC and Belief Propagation". United States. doi:10.2172/1254988. https://www.osti.gov/servlets/purl/1254988.
@article{osti_1254988,
title = {Sythesis of MCMC and Belief Propagation},
author = {Ahn, Sungsoo and Chertkov, Michael and Shin, Jinwoo},
abstractNote = {Markov Chain Monte Carlo (MCMC) and Belief Propagation (BP) are the most popular algorithms for computational inference in Graphical Models (GM). In principle, MCMC is an exact probabilistic method which, however, often suffers from exponentially slow mixing. In contrast, BP is a deterministic method, which is typically fast, empirically very successful, however in general lacking control of accuracy over loopy graphs. In this paper, we introduce MCMC algorithms correcting the approximation error of BP, i.e., we provide a way to compensate for BP errors via a consecutive BP-aware MCMC. Our framework is based on the Loop Calculus (LC) approach which allows to express the BP error as a sum of weighted generalized loops. Although the full series is computationally intractable, it is known that a truncated series, summing up all 2-regular loops, is computable in polynomial-time for planar pair-wise binary GMs and it also provides a highly accurate approximation empirically. Motivated by this, we first propose a polynomial-time approximation MCMC scheme for the truncated series of general (non-planar) pair-wise binary models. Our main idea here is to use the Worm algorithm, known to provide fast mixing in other (related) problems, and then design an appropriate rejection scheme to sample 2-regular loops. Furthermore, we also design an efficient rejection-free MCMC scheme for approximating the full series. The main novelty underlying our design is in utilizing the concept of cycle basis, which provides an efficient decomposition of the generalized loops. In essence, the proposed MCMC schemes run on transformed GM built upon the non-trivial BP solution, and our experiments show that this synthesis of BP and MCMC outperforms both direct MCMC and bare BP schemes.},
doi = {10.2172/1254988},
journal = {},
number = ,
volume = ,
place = {United States},
year = {Fri May 27 00:00:00 EDT 2016},
month = {Fri May 27 00:00:00 EDT 2016}
}

Technical Report:

Save / Share:
  • Loop calculus introduced in [1], [2] constitutes a new theoretical tool that explicitly expresses symbol Maximum-A-Posteriori (MAP) solution of a general statistical inference problem via a solution of the Belief Propagation (BP) equations. This finding brought a new significance to the BP concept, which in the past was thought of as just a loop-free approximation. In this paper they continue a discussion of the Loop Calculus, partitioning the results into three Sections. In Section 1 they introduce a new formulation of the Loop Calculus in terms of a set of transformations (gauges) that keeping the partition function of the problemmore » invariant. The full expression contains two terms referred to as the 'ground state' and 'excited states' contributions. The BP equations are interpreted as a special (BP) gauge fixing condition that emerges as a special orthogonality constraint between the ground state and excited states, which also selects loop contributions as the only surviving ones among the excited states. In Section 2 they demonstrate how the invariant interpretation of the Loop Calculus, introduced in Section 1, allows a natural extension to the case of a general q-ary alphabet, this is achieved via a loop tower sequential construction. The ground level in the tower is exactly equivalent to assigning one color (out of q available) to the 'ground state' and considering all 'excited' states colored in the remaining (q-1) colors, according to the loop calculus rule. Sequentially, the second level in the tower corresponds to selecting a loop from the previous step, colored in (q-1) colors, and repeating the same ground vs excited states splitting procedure into one and (q-2) colors respectively. The construction proceeds till the full (q-1)-levels deep loop tower (and the corresponding contributions to the partition function) are established. In Section 3 they discuss an ultimate relation between the loop calculus and the Bethe-Free energy variational approach of [3].« less
  • A bench-scale reactor is being used to conduct studies of the conversion of synthesis gas to methanol (MeOH) by a novel process. In previous reports, we provided evidence for a two-step reaction in series, the carbonylation reaction taking place mainly in a non-equilibrium region in the vicinity of the copper chromite surface, and the hydrogenolysis reaction taking place on the surface of the copper chromite. Interaction between the two catalysts enhances the rate of methanol formation. In this quarter, we investigated the effect of pore diffusion on reaction rate and obtained an expression for the rate of reaction for themore » methanol/methyl formate concurrent synthesis.« less
  • The study analyzes a subsample of 523 households from the 1975 Lifestyles and Household Energy Use Survey conducted for the Washington Center for Metropolitan Studies. The study explores the empirical relationship between a set of four Energy-Related Attitude/Belief (ERAB) variables, household electricity and natural gas consumption, and three Energy-Related Discrete Choice (ERDC) variables. Using principal components factor analysis, the ERAB variables were constructed from a portion of the survey responses dealing with what households felt should be done to handle current or future energy shortages. A key finding of the study is that in the context of a conventional econometricmore » specification of electricity and natural gas consumption, ERAB variables are statistically significant, although less significant than conventional explanatory variables for household energy consumption.« less
  • Modeling and simulation is playing an increasing role in supporting tough regulatory decisions, which are typically characterized by variabilities and uncertainties in the scenarios, input conditions, failure criteria, model parameters, and even model form. Variability exists when there is a statistically significant database that is fully relevant to the application. Uncertainty, on the other hand, is characterized by some degree of ignorance. A simple algebraic problem was used to illustrate how various risk methodologies address variability and uncertainty in a regulatory context. These traditional risk methodologies include probabilistic methods (including frequensic and Bayesian perspectives) and second-order methods where variabilities andmore » uncertainties are treated separately. Representing uncertainties with (subjective) probability distributions and using probabilistic methods to propagate subjective distributions can lead to results that are not logically consistent with available knowledge and that may not be conservative. The Method of Belief Scales (MBS) is developed as a means to logically aggregate uncertain input information and to propagate that information through the model to a set of results that are scrutable, easily interpretable by the nonexpert, and logically consistent with the available input information. The MBS, particularly in conjunction with sensitivity analyses, has the potential to be more computationally efficient than other risk methodologies. The regulatory language must be tailored to the specific risk methodology if ambiguity and conflict are to be avoided.« less
  • The thermal challenge problem has been developed at Sandia National Laboratories as a testbed for demonstrating various types of validation approaches and prediction methods. This report discusses one particular methodology to assess the validity of a computational model given experimental data. This methodology is based on Bayesian Belief Networks (BBNs) and can incorporate uncertainty in experimental measurements, in physical quantities, and model uncertainties. The approach uses the prior and posterior distributions of model output to compute a validation metric based on Bayesian hypothesis testing (a Bayes' factor). This report discusses various aspects of the BBN, specifically in the context ofmore » the thermal challenge problem. A BBN is developed for a given set of experimental data in a particular experimental configuration. The development of the BBN and the method for ''solving'' the BBN to develop the posterior distribution of model output through Monte Carlo Markov Chain sampling is discussed in detail. The use of the BBN to compute a Bayes' factor is demonstrated.« less