Likelihood consensus: Principles and application to distributed particle filtering

Hlinka, Ondrej; Slučiak, Ondrej; Hlawatsch, Franz; Djurić, Petar M.; Rupp, Markus

doi:10.1109/acssc.2010.5757533

Cited by 24 publications

(39 citation statements)

References 5 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The distributed particle filtering schemes [5,6] use parametric approximations of the local posterior (particle based) densities and employ gossip on these parameters. In [7], the local likelihoods are parametrized and gossip is used to compute the parameters of the global likelihood. In both these cases, parametrization is used to reduce the communication overhead, but the estimation accuracy also decreases.…”

Section: Discussionmentioning

confidence: 99%

“…The SM and the Top-m schemes gossip on the high dimensional particle weights which increases their communication overhead. Since the SM scheme performs oversampling of the particles it can [7] 0.24 ± 0.02 0.24 ± 0.02 0.84 ± 0.14 SM [8] 0.29 ± 0.10 0.45 ± 0.02 0.49 ± 0.17(1) Top-m [9] 0.24 ± 0.02 --UKF [78] 0.24 ± 0.02 0.37 ± 0.20 -UKF* [78] 0. still work at lower communication costs by using a small sample size. The GA scheme is more balanced in terms of the data dimensionality and the demand of gossip.…”

Section: Performance Comparison Of Distributed Filtersmentioning

confidence: 99%

“…In [7], Hlinka et al propose a scheme that is based on distributed computation of the global likelihood function. This approach identifies a set of sufficient statistics that are exchanged between the sensor nodes using gossip algorithm and an approximation to the global likelihood function is computed locally.…”

Section: Distributed Particle Filteringmentioning

confidence: 99%

“…However, these methods have a poor performance for problems with non-linear state dynamics and/or non-linear measurement models. Distributed particle filtering methods [5,6,7,8,9] are attractive for non-linear estimation problems because they have the ability to approximate more general distributions. However, they have a much higher computational and communication cost.…”

Section: Motivationmentioning

confidence: 99%

“…In future work, one could investigate how different accelerated gossip procedures affect the proposed and comparative schemes. We consider the following methods for comparison: the Gaussian approximation (GA) distributed particle filter [6], the sequential importance resampling (SIR) version of the likelihood consensus (LC) distributed particle filter [7], the set membership (SM) distributed particle filter [8], the top-m selective gossip (Top-m) distributed particle filter [9], and the distributed unscented Kalman filter (UKF) [78]. The distributed filtering schemes [81,82] could not be included in this comparative study because these methods are developed specifically for the bearings-only measurement model and do not have the ability to incorporate other measurement models.…”

Section: Performance Comparison Of Distributed Filtersmentioning

confidence: 99%

See 4 more Smart Citations

Distributed ensemble Kalman filtering

Shahid

Üstebay

Coates

2014

2014 IEEE 8th Sensor Array and Multichannel Signal Processing Workshop (SAM)

View full text Add to dashboard Cite

Distributed estimation in a wireless sensor network has many advantages. It eliminates the need of the centralized knowledge of the measurement model parameters and does not have a single point of failure. Also, any sensor node can be queried to retrieve the state estimate. Furthermore, for high dimensional measurements, local processing of information also results in a significant reduction in the communication overhead.The existing distributed Kalman filtering schemes have good computational and communication efficiency, but do not work well for non-linear and non-Gaussian problems. On the other hand, the distributed particle filtering schemes handle the non-linearities well, but have a much higher computational and communication cost.In this thesis, we propose novel distributed filtering schemes based on the ensemble Kalman filter (EnKF). We consider three forms of the EnKF and express their update equations in an alternative information form. This allows us to use the randomized gossip algorithm to reach consensus on the sufficient statistics and perform local updates. The simulation results show that all three forms of the EnKF have a considerably lower computational cost compared to the particle filters. The results suggest that the proposed distributed schemes outperform the existing state-of-the-art distributed filtering schemes in two scenarios, a) linear measurement model with non-linear state dynamics, and b) high dimensional measurements (model parameters known everywhere in the network) with nonlinear measurement model and state dynamics. In both of these scenarios, the proposed schemes achieve an estimation accuracy comparable to the existing state-of-the-art schemes while significantly reducing the communication cost.ii Résumé L'estimation distribuée dans un réseau de capteurs sans fil possède plusieurs avantages. Ellé elimine le besoin d'une connaissance centralisée des paramtres du modèle de mesure et n'a pas un point de défaillance unique. Aussi, n'importe quel agent-capteur peut-être consulté pour obtenir une approximation de l'état général. De plus, pour des mesures de hautes dimensions, le calcul local d'informations résulte en une réduction significative des coûts de communication.Les implémentations courantes du filtre de Kalman sont efficaces sur les plans de la charge de calcul et de la communication mais ne le sont pas pour les problmes non-linéaires et non-gaussiens. D'un autre côté, les techniques distribuées de filtrage particulaire gèrent avec succès les cas non-linéaires mais sont coûteuses sur les plans de la charge de calcul et de la communication.Dans cette thèse, nous proposons des techniques de filtrage distribué basées sur le filtre de Kalman d'ensemble (FKEn). Nous considérons trois formes du FKEn et exprimons leurséquations de changement sous une forme alternative. Cela nous permet d'utiliser un algorithme de gossip aléatoire afin d'atteindre un consensus sur les statistiques suffisantes et calculer les changements locaux. Les résultats des simulations montrent que ...

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Performance Comparison Of Distributed Filtersmentioning

confidence: 99%

Section: Distributed Particle Filteringmentioning

confidence: 99%

Section: Motivationmentioning

confidence: 99%

Section: Performance Comparison Of Distributed Filtersmentioning

confidence: 99%

See 3 more Smart Citations

Distributed ensemble Kalman filtering

Shahid

Üstebay

Coates

2014

2014 IEEE 8th Sensor Array and Multichannel Signal Processing Workshop (SAM)

View full text Add to dashboard Cite

show abstract

Distributed estimation for a class of nonlinear systems with application to Navier‐Stokes flow estimation

Park

Guay

2015

Can J Chem Eng

View full text Add to dashboard Cite

This paper presents a distributed estimation for a class of nonlinear systems operating over a sensor network. The paper investigates alternative consensus filters based on a nonlinear filter. This nonlinear filter arises in the design of observers for the estimation of Navier-Stokes flow based on a proper orthogonal decomposition method. Using point measurements operating over a sensor network, it is shown that the proposed distributed filter can be used to generate a consensus estimate of the overall Navier-Stokes flow field at each sensor location. A two-dimensional room case-study is considered to demonstrate the effectiveness of the proposed approach.

show abstract