Основні поняття
Efficiently optimizing particle filters using normalizing flows for improved performance in state-space models.
Анотація
In this content, the authors introduce a differentiable particle filtering framework using normalizing flows to enhance the performance of particle filters in complex state-space models. The framework allows for flexible modeling of dynamics, valid probability densities, and effective proposal distributions. The authors conduct experiments to evaluate the proposed method's performance in various scenarios.
Experiment Setup:
Linear Gaussian state-space model with unknown parameters.
Training objective: ELBO approximation.
Model optimization for 500 iterations with fixed learning rate.
Evaluation metrics: L2-norm of parameter estimation, posterior mean difference, ELBO, and effective sample size.
Experimental Results:
AESMC-bootstrap converges fastest but exhibits higher parameter estimation error.
NF-DPF converges faster than AESMC and PFRNN with improved ELBO and lower parameter estimation error.
PFNet shows poor tracking performance despite high ELBO.
Статистика
Die Parameter θ werden mit einer L2-Norm von ||θ - θ*||2 verglichen.
Die Differenz der geschätzten Posteriormittelwerte wird mit einer L2-Norm von ||χT - χ*T||2 bewertet.
Die Effektivität wird durch den ELBO und die effektive Stichprobengröße bewertet.
Цитати
"The NF-DPF converges faster than the AESMC and the PFRNN has the highest ELBO and the lowest parameter estimation error."
"The AESMC-bootstrap converges the fastest but exhibits slightly larger parameter estimation error than other methods."