On the role of interaction in sequential Monte Carlo algorithms

Whiteley, Nick; Lee, Anthony; Heine, Kari

doi:10.3150/14-bej666

Cited by 46 publications

(99 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This is the price to pay for sparse communication. Lastly, we observe that running parallel PFs without communication is not stable as stated in [2]. If the nodes do not communicate with each other, eventually one node dominates the others by holding all the weights and ESS decreases below N/M .…”

Section: Effective Sample Sizementioning

confidence: 85%

“…However in the resampling stage, all the particles must interact with each other which causes communication overhead. Although resampling is inevitable for the stability of PFs [2], it is evidently the bottleneck in the distributed implementation of particle filters. Therefore, the main consideration is to minimize the communication between the nodes in designing distributed resampling algorithms, otherwise the communication overhead prevents PFs from speeding up.…”

Section: Introductionmentioning

confidence: 99%

“…Vergé et al [8] propose the double bootstrap filter with a two level resampling method. Heine et al [9] break down the resampling stage into a sequence of constrained interactions of particles, based on αSMC methodology described in Whiteley et al [2].…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Distributed particle filtering under real-time constraints

Bozkurt

Cemgil

2017

2017 25th European Signal Processing Conference (EUSIPCO)

View full text Add to dashboard Cite

Abstract-Particle filters are powerful methods for state estimation in nonlinear/non-Gaussian dynamical systems. However due to the heavy computational requirements, they may not satisfy the real-time constraints in many applications requiring a large number of particles. By means of distributed implementation, real-time particle filtering can be achieved. However, the resampling stage in particle filters requires particle interaction which causes communication overhead. In this work, we propose a distributed resampling algorithm based on Butterfly Resampling previously described in the literature. We describe three interaction schemes (i) the complete interaction, (ii) the pairwise interaction where the nodes are constrained to communicate in pairs and (iii) the partial pairwise interaction in which only one pair is allowed to communicate. The goal is to diminish the communication cost in exchange for negligible loss of effective sample size. We conduct experiments on a cluster environment and compare our methods in terms of execution time, communication time and effective sample size. We find that the sparse interaction schemes show better performance for distributed systems and they keep the effective sample size nearly as high as the complete interaction scheme does.

show abstract

Section: Effective Sample Sizementioning

confidence: 85%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Distributed particle filtering under real-time constraints

Bozkurt

Cemgil

2017

2017 25th European Signal Processing Conference (EUSIPCO)

View full text Add to dashboard Cite

show abstract

“…In this section, we overview relevant aspects of the general methodology proposed in . An HMM with measurable state space

((), sans-serifX, scriptX)

and observation space

((), sans-serifY, scriptY)

is a process

({}, ((), {normalX}_{n}, Y_{n}); n \geq 0)

where

({}, {normalX}_{n}; n \geq 0)

is a Markov chain on

sans-serifX

, the observations

({}, Y_{n}; n \geq 0)

, valued in

sans-serifY

, are conditionally independent given

({}, {normalX}_{n}; n \geq 0)

, and the conditional distribution of each Y n depends on

({}, {normalX}_{n}; n \geq 0)

only through X n .…”

Section: αSmcmentioning

confidence: 99%

Forest resampling for distributed sequential Monte Carlo

Lee

Whiteley

2015

Statistical Analysis

Self Cite

View full text Add to dashboard Cite

This paper brings explicit considerations of distributed computing architectures and data structures into the rigorous design of Sequential Monte Carlo (SMC) methods. A theoretical result established recently by the authors shows that adapting interaction between particles to suitably control the effective sample size (ESS) is sufficient to guarantee stability of SMC algorithms. Our objective is to leverage this result and devise algorithms which are thus guaranteed to work well in a distributed setting. We make three main contributions to achieve this. First, we study mathematical properties of the ESS as a function of matrices and graphs that parameterize the interaction among particles. Secondly, we show how these graphs can be induced by tree data structures which model the logical network topology of an abstract distributed computing environment. Finally, we present efficient distributed algorithms that achieve the desired ESS control, perform resampling and operate on forests associated with these trees. © 2015 Wiley Periodicals, Inc. Statistical Analysis and Data Mining: The ASA Data Science Journal, 2015

show abstract

“…This idea has been indirectly and implicitly employed in different Monte Carlo schemes: parallel particle filters [9], [10], particle island and related methods [11]- [13], tracking and model selection algorithms [14], nested sequential Monte Carlo schemes [15], [16] are some examples.…”

Section: Introductionmentioning

confidence: 99%

Group metropolis sampling

Martino

Elvira

Camps‐Valls

2017

2017 25th European Signal Processing Conference (EUSIPCO)

View full text Add to dashboard Cite

Abstract-Monte Carlo (MC) methods are widely used for Bayesian inference and optimization in statistics, signal processing and machine learning. Two well-known class of MC methods are the Importance Sampling (IS) techniques and the Markov Chain Monte Carlo (MCMC) algorithms. In this work, we introduce the Group Importance Sampling (GIS) framework where different sets of weighted samples are properly summarized with one summary particle and one summary weight. GIS facilitates the design of novel efficient MC techniques. For instance, we present the Group Metropolis Sampling (GMS) algorithm which produces a Markov chain of sets of weighted samples. GMS in general outperforms other multiple try schemes as shown by means of numerical simulations.

show abstract

On the role of interaction in sequential Monte Carlo algorithms

Cited by 46 publications

References 24 publications

Distributed particle filtering under real-time constraints

Distributed particle filtering under real-time constraints

Forest resampling for distributed sequential Monte Carlo

Group metropolis sampling

Contact Info

Product

Resources

About