Uncertainty in Data-Driven Kalman Filtering for Partially Known State-Space Models

Itzik Klein, Guy Revach, Nir Shlezinger, Jonas E. Mehr, Ruud J. G. van Sloun, Yonina C. Eldar

This paper, titled "Uncertainty in Data-Driven Kalman Filtering for Partially Known State-Space Models," presents a significant step forward in making deep-learning-based tracking systems as reliable and "self-aware" as classical engineering tools.

TL;DR

The authors demonstrate that KalmanNet—a hybrid deep learning/model-based state estimator—can provide reliable uncertainty measures (error covariance) just like a classical Kalman Filter. By exploiting its interpretable architecture, KalmanNet produces accurate uncertainty estimates even when the underlying physical model is mismatched or partially unknown, outperforming traditional filters in real-world scenarios.

What Problem Does This Paper Address?

In fields like autonomous driving, robotics, and aviation, knowing how much to trust a sensor estimate is just as important as the estimate itself. This is called uncertainty quantification.

The gold standard for this has long been the Kalman Filter (KF). If you have a perfect mathematical model of your system, the KF provides the mathematically optimal estimate and a precise measure of its own error (the error covariance). However, real-world models are rarely perfect. If the model is slightly wrong (model mismatch), the KF becomes "overconfident"—it predicts low error while actually performing poorly.

On the other hand, Deep Learning (DL) models are great at learning from data when models are unknown, but they are often "black boxes." They struggle to provide reliable uncertainty metrics and require massive amounts of data. This paper investigates whether KalmanNet, a hybrid approach, can offer the best of both worlds: the flexibility of AI with the rigorous uncertainty tracking of classical control theory.

Methods

The researchers focus on KalmanNet, an architecture that keeps the structural "flow" of a Kalman Filter but replaces the most difficult part—calculating the Kalman Gain (KG)—with a Recurrent Neural Network (RNN).

The key technical contribution is Theorem 1. The authors mathematically prove that if the observation model (how sensors see the state) is known and sufficiently descriptive, the error covariance can be derived directly from the RNN’s internal features (the learned Kalman Gain).

The Process: 1. Filtering: KalmanNet uses its RNN to estimate the state of the system. 2. Uncertainty Extraction: Instead of discarding the RNN's internal "Kalman Gain" output, the system plugs it into a newly derived formula to calculate the predicted error covariance in real-time.

Datasets

The authors evaluated their approach using two primary simulations:

  • Linear Gaussian State-Space Models: Synthetic data used to compare KalmanNet against the theoretically optimal (but "all-knowing") Kalman Filter.
  • The Lorenz Attractor: A famous, highly non-linear chaotic system. This was used to test how the system handles complex dynamics and "discretization error" (the gap between continuous physics and discrete computer timestamps).

Key Insights and Ideas

  • Interpretability is a Feature, Not a Constraint: By forcing the neural network to mimic the structure of a Kalman Filter, the authors made the "hidden" layers of the AI physically meaningful. The RNN isn't just guessing; it's learning a specific control parameter (the Gain).
  • Self-Awareness under Mismatch: The most striking insight is that when the physical model is wrong (e.g., the system thinks an object is moving slower than it actually is), KalmanNet's uncertainty estimate remains "honest." It tracks the actual empirical error, whereas a standard Kalman Filter remains stuck in its (wrong) theoretical assumptions.
  • Data-Driven Statistics: Unlike a standard KF, which requires engineers to manually tune noise statistics ($Q$ and $R$ matrices), KalmanNet learns these implicitly from data, making it much easier to deploy in complex environments.

Key Results

  • Perfect Model Match: When the model is perfectly known, KalmanNet’s uncertainty estimates are identical to the optimal Kalman Filter. It learns the "truth" perfectly.
  • Robustness to Mismatch: In scenarios with a mismatched evolution model, KalmanNet provided a much more accurate estimation of its own error. While the standard KF underestimated its error (leading to potential system failures), KalmanNet’s predicted uncertainty closely tracked its actual performance.
  • Non-Linear Performance: On the chaotic Lorenz attractor, KalmanNet achieved a lower Mean Squared Error (MSE) than the Extended Kalman Filter (EKF) and provided a far more realistic measure of uncertainty, whereas the EKF tended to be overly pessimistic (overestimating error).

Limitations and Future Work

The current derivation relies on the observation model ($H$) being known and having "full column rank." In plain English, this means you need to have at least as many independent sensor measurements as there are hidden states you are trying to track.

The authors note that future work will focus on:

  1. Handling cases where the observation model is not full rank (fewer sensors than states).
  2. Further exploring scenarios where the observation model itself is partially unknown or non-linear.

Summary updated

Abstract

Providing a metric of uncertainty alongside a state estimate is often crucial when tracking a dynamical system. Classic state estimators, such as the Kalman filter (KF), provide a time-dependent uncertainty measure from knowledge of the underlying statistics; however, deep learning based tracking systems struggle to reliably characterize uncertainty. In this paper, we investigate the ability of KalmanNet, a recently proposed; hybrid; model-based; deep state tracking algorithm, to estimate an uncertainty measure. By exploiting the interpretable nature of KalmanNet, we show that the error covariance matrix can be computed based on its internal features, as an uncertainty measure. We demonstrate that when the system dynamics are known, KalmanNet—which learns its mapping from data without access to the statistics—provides uncertainty similar to that provided by the KF; and while in the presence of evolution model-mismatch, KalmanNet provides a more accurate error estimation.
BibTeX
@inproceedings{klein2022uncertainty,
  author = {Itzik Klein and Guy Revach and Nir Shlezinger and Jonas E. Mehr and Ruud J. G. van Sloun and Yonina C. Eldar},
  title = {{Uncertainty in Data-Driven Kalman Filtering for Partially Known State-Space Models}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2022},
  url = {https://openreview.net/forum?id=PSfvbAtHpwh},
}

Loading related content...

Figure