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.
Introduction
In critical applications like autonomous robotics and aviation, a state estimate is only as useful as the confidence measure attached to it. While the classical Kalman Filter (KF) provides an optimal uncertainty metric (error covariance) given a perfect system model, it often fails in real-world scenarios where models are mismatched or partially unknown. In such cases, the KF becomes "overconfident," predicting low error while its actual performance degrades.
Conversely, deep learning (DL) methods can learn complex dynamics from data but typically lack the rigorous, interpretable uncertainty quantification of model-based filters. This paper explores KalmanNet, a hybrid architecture designed to bridge this gap by offering the flexibility of AI with the rigorous uncertainty tracking of classical control theory.
Methods: The KalmanNet Architecture
KalmanNet preserves the structural flow of the classical Kalman Filter but replaces the computation of the Kalman Gain (KG) with a Recurrent Neural Network (RNN). This allows the system to learn optimal gain parameters directly from data without requiring explicit knowledge of noise statistics ($Q$ and $R$ matrices).
The researchers' key technical contribution is the mathematical derivation of a method to extract the error covariance matrix directly from the RNN’s internal features. By proving that the learned Kalman Gain can be used to compute a valid uncertainty measure, the authors transform a "black box" neural network into a self-aware tracking system. This interpretability ensures that the RNN isn't just guessing; it is learning a specific, physically meaningful control parameter.
Experimental Evaluation
The approach was tested using two primary simulations:
- Linear Gaussian State-Space Models: Used to establish a baseline against the theoretically optimal (but "all-knowing") KF.
- The Lorenz Attractor: A chaotic, non-linear system used to evaluate performance under complex dynamics and discretization errors (the gap between continuous physics and discrete computer timestamps).
Key Results: Robustness and Self-Awareness
The study reveals that when the system model is perfectly known, KalmanNet matches the performance and uncertainty estimation of the optimal KF. However, its true strength lies in scenarios with model mismatch—for instance, when the filter assumes a different evolution speed than the actual system.
As shown in the results, KalmanNet remains "honest" about its performance. While the standard KF underestimates its error (leading to potential system failures), KalmanNet’s predicted uncertainty closely tracks its actual empirical performance. On the chaotic Lorenz attractor, KalmanNet not only achieved a lower Mean Squared Error (MSE) than the Extended Kalman Filter (EKF) but also provided a far more realistic measure of uncertainty, avoiding the EKF's tendency toward over-pessimism.
Conclusion and Future Work
KalmanNet demonstrates that interpretability in AI is a powerful feature, not a constraint. By constraining the neural network within a model-based framework, the authors achieved a system that is both flexible enough to learn from data and rigorous enough to quantify its own uncertainty.
The current derivation relies on the observation model ($H$) being known and having "full column rank" (having at least as many independent sensor measurements as there are hidden states). Future work will focus on handling cases where the observation model is non-linear or partially unknown.
Summary updated