Core Concepts
DANSE provides a closed-form posterior of the state of a model-free process, given linear measurements, without requiring knowledge of the process dynamics or supervised learning.
Abstract
The article proposes DANSE, a Data-driven Nonlinear State Estimation method, to address Bayesian state estimation and forecasting for a model-free process in an unsupervised learning setup.
Key highlights:
DANSE does not require any a-priori knowledge of the process dynamics or the state space model. It is designed for complex, model-free processes.
DANSE uses a data-driven recurrent neural network (RNN) to provide the parameters of a prior distribution of the state, which depends on past measurements.
The closed-form posterior of the state is then computed using the current measurement and the RNN-based prior.
DANSE is trained in an unsupervised manner using only a dataset of noisy measurement trajectories, without access to the true state trajectories.
Experiments on linear and nonlinear process models (Lorenz attractor, Chen attractor) show that DANSE provides competitive performance against model-driven methods like Kalman filter, extended Kalman filter, unscented Kalman filter, as well as data-driven methods like deep Markov model and hybrid methods like KalmanNet.
DANSE is also shown to work for high-dimensional state estimation tasks.
Stats
The article does not provide any specific numerical data or statistics to support the key logics. The performance comparisons are shown through plots of normalized mean-squared error (NMSE) versus signal-to-measurement noise ratio (SMNR).
Quotes
There are no direct quotes from the content that support the key logics.