Canada Spends Logo

Explorez notre analyse du budget 2026

Grant: Discovery Grants Program - Individual

Université de Montréal

$14,000.00

Efficiency of Markov chain Monte Carlo methods

EF 2013-14

Résumé
Random walk Metropolis (RWM) algorithms, an important class of Markov chain Monte Carlo (MCMC) algorithms, are methods to generate data from highly complex probability distributions (the target distribution). Researchers and practitioners in various fields of applications such as biostatistics, computer science, physics, finance, and applied statistics extensively use such methods. In applying RWM algorithms, it is necessary to choose a proposal density; due to its accessibility, the normal density is the most popular choice. The variance of this distribution must be selected carefully in order to have some level of optimality in the performance of the algorithm; this is called the optimal scaling problem. Results about this issue are available in the literature for high-dimensional target distributions with independent components.My goal is to provide users with more efficient algorithms for less restrictive models (i.e., target models actually arising in practice). One way of achieving this is to study the optimal scaling problem for target distributions whose components admit a nontrivial correlation structure; specifically, I propose to study hierarchical models, which are omnipresent in Bayesian statistics and as such often arise in biostatistics. There also exist many variations of the RWM algorithm that work well in practice, but that have not been studied theoretically. One of them is the Multiple-Try Metropolis algorithm, which generates several ""trial"" proposed moves at every step. I would like to study the weak convergence and optimal scaling theory of such variations. Finally, I intend to propose an algorithm specifically adapted for sampling from the usually problematic bimodal target densities.I shall also study some applications of MCMC algorithms, notably in engineering (in a problem related to air traffic control).
Programme
Discovery Grants Program - Individual
Emplacement
Quebec, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data

Tel que publié

Application id
346215-2008
Name nom
Bédard, Mylène
Department département
Mathématiques et de statistique
Organization id
63
Institution établissement
Université de Montréal
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2013
Competition year année de concours
2008
Award amount
14000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
Discovery grants programs
Group fr
Programmes de subventions à la découverte
Committee code
14
Committee name en
Statistical Sciences
Committee name fr
Statistique
Area of application code
405
Area of application group en
Environment
Area of application group fr
Environnement
Area of application en
Modelling and mathematical simulation of natural processes
Area of application fr
Modélisation et simulation mathématique des processus naturels
Research subject code
3000
Research subject group en
Statistics and probability
Research subject group fr
Statistique et probabilité
Research subject en
Statistics and probability
Research subject fr
Statistique et probabilité
Partie
5
Nb partie
5
Application title
Efficiency of Markov chain Monte Carlo methods
Keyword
NSERC
Application summary
Random walk Metropolis (RWM) algorithms, an important class of Markov chain Monte Carlo (MCMC) algorithms, are methods to generate data from highly complex probability distributions (the target distribution). Researchers and practitioners in various fields of applications such as biostatistics, computer science, physics, finance, and applied statistics extensively use such methods. In applying RWM algorithms, it is necessary to choose a proposal density; due to its accessibility, the normal density is the most popular choice. The variance of this distribution must be selected carefully in order to have some level of optimality in the performance of the algorithm; this is called the optimal scaling problem. Results about this issue are available in the literature for high-dimensional target distributions with independent components.My goal is to provide users with more efficient algorithms for less restrictive models (i.e., target models actually arising in practice). One way of achieving this is to study the optimal scaling problem for target distributions whose components admit a nontrivial correlation structure; specifically, I propose to study hierarchical models, which are omnipresent in Bayesian statistics and as such often arise in biostatistics. There also exist many variations of the RWM algorithm that work well in practice, but that have not been studied theoretically. One of them is the Multiple-Try Metropolis algorithm, which generates several ""trial"" proposed moves at every step. I would like to study the weak convergence and optimal scaling theory of such variations. Finally, I intend to propose an algorithm specifically adapted for sampling from the usually problematic bimodal target densities.I shall also study some applications of MCMC algorithms, notably in engineering (in a problem related to air traffic control).